Assistive workflow: source-bound implementation An agreed outcome becomes a maintained plan, direct or delegated implementation, proportional verification and an honest delivery report. Task context and dependency advice support progress; real user stops and host permissions remain effective. Greenfield and brownfield population use the paired native writer and exact finalization. Exact blob bindings identify inspected source bytes; the linked revision is the committed source base. Implementation, verification, live host qualification and release remain separate claims.
Current source changes Routine artifact instructions now focus on the installed writer, typed plan, bounded task reads, factual prerequisites, actual checks and safe retry. The default leaf is about 7.3 KB and includes an executable stdin bootstrap with optional operation-schema discovery; explicit attestation remains available on demand without burdening ordinary task context. A spec's SMARTS design rationale lives in its approach, decisions and one SEC-SMARTS section. core/surface/includes/artifacts.md Plans preserve agreed criteria, useful tasks, dependency ancestry, bounded shared inputs and observable checks. Stable source identities and typed maintenance keep changes within the selected outcome. A design-drift guard routes an undecided material solution choice back to brainstorming instead of scoring it during planning. core/surface/skills/writing-plans/SKILL.md The pinned native bridge enables task-context as a bounded informational read and does not require future reviewer/approval producers simply to author a structurally valid draft. Context requests route to ordinary context-preview, context-finalize-preview and exact-reference context-apply; historical receipt input stays a separate compatibility path. Installed host qualification is separate. Native Pi writer admission does not establish a complete live Pi workflow. On NO_LOCAL_STORE the client runs init-store once and retries the request once; a refused initialization keeps NO_LOCAL_STORE. On Windows the binary's architecture comes from the kernel (IsWow64Process2), never WMI or PROCESSOR_* environment. Index and outline pages are now fully checked before rows are yielded, including continuation identity and row shape. core/pysrc/_artifactlib.py A new explicit batch review can select a passing rerun of still-reviewed work without replacing the original task review. The selected run must match its own recorded workspace; changed source or supporting evidence still requires reassessment. core/artifacts/internal/evidence/completion.go The user-facing route table points feature and sprint work at useful planned execution and selected context refresh at its existing owner. Existing authorization covers ordinary work; actual host, destructive-action and commit controls remain. core/surface/includes/routing-table.md The skill index connects planning, direct or delegated implementation, proportional verification and selected context maintenance to their owning workflows. core/surface/skills/INDEX.md The sprint caller performs an index-first recorded-intent read, preserves actual owner decisions and ADR conflicts, then proceeds with authorized work without a second artifact approval system. A SMARTS Unknown names its missing observation and cannot bypass a hard gate or manufacture authority. core/surface/SPRINT.md The sprint entry now routes to the assistive implementation body instead of requiring a fresh user approval and two reviews per task. core/surface/commands/sprint.md The feature entry now resumes a maintained spec/plan and task progress without another artifact certification prompt; independent commit and test controls remain. The small lane runs an inline SMARTS scan only for a material solution choice. core/surface/commands/feature.md Source refinement respects recorded deferrals and accepted ADR conflicts without turning each ordinary method adjustment into a separate authorization interview. SMARTS compares approaches only when two or more are materially plausible; a single approach gets one line and no table. core/surface/skills/brainstorming/SKILL.md Attended feature execution now treats checkpoints as progress groups, not compulsory approval stops; it forwards real failures and continues independent work. core/surface/skills/executing-plans/SKILL.md Implementation works directly or delegates a coherent unit when useful. Bounded task context, applicable dependency advice and observed checks support progress while preserving actual user stops and host permissions. Verification checks the change against a recorded SMARTS rationale for material invalidation only. core/surface/skills/subagent-driven-development/SKILL.md Confirmed fixes require regression RED/GREEN. Features use current criterion-focused checks and meaningful negative controls where needed; refactors preserve passing characterization baselines. Coverage, lint and applicable security requirements remain explicit. core/surface/skills/tdd/SKILL.md The commit owner checks authorization, the staged diff, secrets and the verification approach appropriate to the change. Current criterion checks, genuine fix regressions and refactor parity retain their distinct evidence requirements. Reuse applicable observed checks and reviews across phases. Managed v2 maintenance selects affected claims and uses native paired writes; unknown evidence defers that document. core/surface/skills/commit-gate/SKILL.md PR readiness uses the explicit base-to-HEAD subject for current checks, secrets scanning and applicable specialized review. Existing evidence is reused only while its source and inputs match. Merge requires authorization and exact-head hosted CI. core/surface/skills/finishing-a-development-branch/SKILL.md Read-only task-context reports direct dependencies and unresolved ancestors, scoped_changed, scoped_unchanged or unassessed inputs, and actual check observations. Missing broader context remains explicit; the packet does not certify verification. core/artifacts/internal/operations/task_context.go An unanswered legacy reconciliation can be superseded only when route and pending state agree and no durable attempt or matching observed reply exists. Captured replies recover through their original immutable source and receipt; in-flight mutations retain exact replay and real user stops. core/pysrc/_reconciliationlib.py The installed recover command finishes a captured original reply or replays an in-flight mutation without asking for a new user answer. core/pysrc/artifact-reconcile.py Task-check observations now retain directly referenced criterion identity and an optional bounded source-input fingerprint or an explicit unassessed reason, without executing tests or minting approval receipts. core/artifacts/internal/operations/task_check.go New checks bind the selected criterion and applicable constraint bodies, including explicit references and scope applicability. Unrelated constraints remain outside the digest; old observations never gain unassessed constraint coverage. core/artifacts/internal/model/assistive.go Exact and batch symbol retrieval exposes reported check history without an approval ticket. Outline paging now accounts for encoded rows incrementally while retaining response budgets and continuation behavior. core/artifacts/internal/read/read.go Human HTML presents reported command checks in an escaped, explicitly unverified section; implementation state and source validation remain distinct. core/artifacts/internal/render/render.go The closed task-check request accepts bounded command observations matched to plan definitions. Ordinary context-preview and context-finalize-preview omit prompt identity; context-apply accepts an exact preview reference and digest. Invalid requests, changed preimages and stale CAS writes remain rejected. The historical smarts-apply request accepts the 0.2.0 SMARTS decision record. init-store takes a protocol-only request. core/artifacts/internal/operations/protocol.go The unpublished 0.3.2 native schema adds optional shared_inputs and scoped task-check fields without silently relabeling old records. core/artifacts/schemas/common.schema.json The bounded source-fingerprint reader hashes actual declared task/shared paths, directory membership and known build-control files; unsafe, missing and over-limit scopes are explicitly unassessed rather than blocking work or proving a full dependency graph. core/artifacts/internal/operations/input_scope.go The candidate-only runbook distinguishes 0.3.1/0.3.2 mixed-schema readers, exact replay and forward recovery while preserving recorded user stops. The paired prompt retirement scans immutable observed replies before calling a request unanswered; submitted or uncertain decisions are never erased. Older-client downgrade after modern edits remains unsupported. docs/artifacts/ASSISTIVE-UPGRADE.md Ordinary approval and sprint instructions bypass legacy capture without changing pending history or inventing authority. Explicit legacy replies retain exact matching. The paired adapter can retire a genuinely unanswered 0.2 sprint-pair request using its serialized cancellation lock and submission check; submitted decisions must replay, while standalone approval 0.1 refuses automatic retirement. core/pysrc/_approvallib.py Ordinary prerequisite instructions bypass legacy capture. The existing adapter can retire only a validated ARMED pending question under its original per-plan OS lock. Confirmed or captured decisions are not disposable; no authority receipt or new native permission API is created. core/pysrc/_prerequisitelib.py The old paired approval producer publishes a content-addressed observed reply before its durable submission. Cancellation checks the matching observation, content hash, prompt and frozen context under its existing lock, with a strict scan bound. A matching or uncertain pre-submission reply blocks automatic retirement; normal tasks and context reads have no added cost. core/pysrc/_sprintapprovallib.py The optional reference retains strict historical evidence inputs without replacing ordinary work. Explicit draft source refresh preserves progressed plans and history with both CAS bindings; initial approval restrictions remain on historical paired attestation. Its verify command requires literal absolute paths; linked agent subdirectories remain unsupported for strict authority review. These limitations apply to requested attestation, not ordinary task start. core/surface/includes/artifacts-attestation.md The command catalog starts from the agreed outcome and maintained plan; SMARTS is useful for difficult tradeoffs and does not authorize a changed outcome. core/surface/COMMANDS.md Resident instructions preserve existing work authorization and actual owner decisions without another spec approval or a SMARTS record for every routine step. core/surface/arbiter.md Status separates reported DONE from independently ACCEPTED work, shows actual blocked decisions, and does not infer an approval request from draft governance state. core/surface/commands/status.md Delegated authors preserve the caller-selected verification approach: genuine regression RED/GREEN for fixes, current criterion-focused feature checks and passing characterization baselines for refactors. Actual results and pending checks return to the caller; staging requires assigned scope and authorization. core/surface/includes/author-tdd-workflow.md Codex dispatch requires attestation only for explicitly claimed host proof; an unavailable requested farm backend is reported without silently switching provider or model. Ordinary context inspection can proceed inline within actual host permissions; an agent label supplies no containment or semantic verification. core/surface/includes/codex-host-notes.md Historical sprint approval and SMARTS proof instructions remain available only for existing explicit attestation, outside the ordinary execution path. core/surface/includes/sprint-authority.md The public Codex farm boundary matches its caller: report the unavailable backend and use another provider or model only within existing authorization. docs/parity.md The preparation helper binds bounded drafts and retained sources to exact document/provenance preimages. Finalization preparation hashes all thirteen required outputs and retains missing output failures without modifying files. core/pysrc/_contextwriterlib.py The installed context-draft command prepares a paired document preview; context-finalize prepares initial finalization. context-apply commits the exact selected preview and digest. Context modes preserve scaffold and initialized refresh boundaries. core/pysrc/init-codearbiter.py Ordinary context-preview retains an inert proposal and exact preimages without approval or review claims. context-apply revalidates the digest, closed request and preview binding before the native paired transaction; context-finalize-preview uses the same exact-reference path for initial population. No approve-context reply is required. core/artifacts/internal/operations/context_assistive.go The finite context contract permits a whole-document Markdown entry only as the sole entry for one of five named content targets. Explicit typed field entries remain available; the writer does not expose arbitrary repository-file writes. core/artifacts/internal/contextdocument/contract.go Whole-document replacement requires the explicit document selection and exact existing bytes. It preserves CONTEXT frontmatter and initialization state, checks the named document identity and Markdown structure, and retains the code-map size limit. Human content outside the authorized change remains the caller's preservation obligation. core/artifacts/internal/contextdocument/document_selection.go Ordinary create/adopt/update preserves the same source-evidence, selection and compare-and-swap checks as the native transaction. Missing or matching v1 provenance can be adopted; existing v2 ownership can transition through explicit whole-document selection. Original document/provenance preimages remain in transaction history for exact recovery. core/artifacts/internal/contextdocument/operations.go Initial finalization binds thirteen output identities, rechecks supporting source evidence, and changes the initialization marker with its paired provenance digest. Ordinary finalization requires no artifact approval receipt and does not certify semantic correctness. Already initialized repositories refuse a new finalization. core/artifacts/internal/contextdocument/finalization.go The context schema describes both finite field entries and the explicit whole-document entry for the five supported content documents. Its typed structure records evidence references; schema validity does not establish that the claims were independently reviewed. core/artifacts/schemas/context.schema.json Brownfield population covers six logical categories and five content documents through direct bounded inspection or optional scoped agents. The paired writer preserves source bindings and protected content; initial finalization checks all required outputs. core/surface/skills/context-creation/SKILL.md A drift audit stays read-only; an already authorized selected or full refresh proceeds without asking again per document. The installed draft/preview/apply example preserves source bindings and selected content. Matching hashes establish identity only; initialized refresh never reruns scaffolding, clears the marker or repeats finalization. core/surface/skills/context-check/SKILL.md Initial brownfield setup checks the intended repository and actual host permissions, then uses bounded inspection and the installed safe context writer. Ordinary Windows work needs no special scout profile or per-document artifact approval. An existing initialized repository routes to selected refresh instead of another scaffold or population cycle. core/surface/commands/init.md Optional context scouts report only their assigned scope, source identities, coverage, conflicts and unknowns. Six logical categories do not require six child agents. A role label or read-only instruction proves neither sandbox containment nor independent semantic verification; actual host permissions and excluded paths still apply. core/surface/agents/scout.md Release preparation inserts a new version below the entire Unreleased section and above prior versions, preserving pending notes. This guidance correction does not grant publication authority or replace the existing release-PR and hosted qualification boundaries. core/surface/skills/release/SKILL.md The lazy reference map routes provenance drift or requested initialized refresh to context-check after reading selected documents, provenance and the code map. The owner reuses selected scope and existing authorization. core/surface/includes/reference-map.md Authors and reviewers share task-critical reference and constraint checks. Missing or stale provenance remains advisory. Two delivery attempts end with an advisory unknown observation, without another retry or an invented receipt. core/pysrc/_contextselectlib.py Feature actors receive scoped requirements, named critical references and observed evidence. Applicable conflicts or missing required inputs stop dependent work; provenance warnings remain visible and advisory. core/pysrc/_artifactpromptlib.py Confirmed fixes preserve causal evidence and a named regression. Delegated input retains critical requirements; advisory delivery exhaustion ends retries while actual missing inputs and conflicts retain their dependent-work boundary. Distinct remediation strategies are compared with SMARTS only after the root cause is confirmed. core/surface/commands/fix.md A review binds one resolved diff and supplies scoped inputs to each reviewer. Missing critical references or applicable conflicts prevent a substantive verdict; provenance and unobserved delivery diagnostics remain advisory. core/surface/commands/review.md Characterization tests can pass against existing correct behavior. Preserve behavioral assertions against the post-backfill baseline; mechanical import or symbol-reference updates retain the same oracle. Coverage and parity checks remain required. A precise surface table proceeds under existing user authorization; only material expansion or a changed outcome needs another decision. Distinct restructuring strategies are compared with SMARTS only after surface and parity are settled. core/surface/skills/refactor/SKILL.md Greenfield interviews persist under plans/decomposition and retain actual decisions. Five document/v2-provenance pairs use context-draft and context-apply; context-finalize binds all thirteen outputs. ADRs and task rows use their owners, and retained interview evidence survives finalization. core/surface/skills/decompose/SKILL.md Scheduling follows the full dependency ancestry and exposes unresolved ancestor IDs and reasons. Reconciled progress remains advisory; blocked ancestors prevent premature downstream readiness while independent tasks remain available. Response-budget omissions are explicit, retain the global scheduling decision and direct callers to supported detail reads. core/artifacts/internal/operations/assistive.go Explicit draft source refresh retains both source CAS components and same-pair identity without requiring an unstarted plan. Existing progress, approvals and evidence are preserved without asserting current authority. Dependency invalidation preserves an independently BLOCKED task and its reason. Verification summaries disclose conflicting caller reports separately from receipt freshness; later successful verification covers only reports present when its context was issued. core/artifacts/internal/operations/lifecycle.go A checkpoint retains the reviewed commit, UTC review-start time and worktree dirty state, then writes the report through its aggregator. last-checkpoint holds only the override count. core/surface/commands/checkpoint.md Checkpoint reports persist observed source_commit, recorded_at and worktree_dirty values together with every review finding. Missing anchors remain unknown. core/surface/agents/checkpoint-aggregator.md Checkpoint audit windows use the report timestamp and an ancestor commit. Invalid or missing anchors require an explicit ref/date window; dirty review scope remains a stated limitation. core/surface/commands/audit.md Bounded read-time pointers include ordinary draft and ready HTML specs as advisory context. Rejected or withdrawn specs stay excluded; no pointer confers approval. core/pysrc/_readinjectlib.py Routine maintenance selects changed v2 claims intersecting staged sources using a current bounded snapshot. Unknown evidence defers the affected document; native paired writes preserve owned fields and current identities. core/pysrc/_provenancelib.py Ordinary uppercase words such as Approve DATA migration remain ordinary input. All-letter capture codes require a matching retained request; real armed replies, mismatches and user stops preserve their original handling. core/pysrc/_replylib.py Only an actual verifier wrapper executable/script prefix selects authority handling. Inert PowerShell data mentioning the filename is not a verifier call; malformed real verifier invocations still fail closed. Fresh legacy context approval arming is refused; historical pending routes remain readable. core/pysrc/_artifactauthoritylib.py Bounded isolated Python stdin diagnostics may read literal protected paths or git show HEAD objects. Unknown code, mutable calls, shell expansion and extra commands retain the guard checks; ordinary commands avoid AST parsing. core/pysrc/_bashguardlib.py 01 Overview 02 Feature / sprint 03 Context layers 04 Change lanes 05 Review / delivery 06 Knowledge / operations 07 Brownfield lifecycle 08 Debug handoff 09 Tribunal lifecycle codeArbiter / route atlas A command selects an owner. The owner loads context, branches through gates, and returns a bounded result. An agreed outcome becomes a maintained plan, direct or delegated implementation, proportional verification and an honest delivery report. Task context and dependency advice support progress; real user stops and host permissions remain effective. Greenfield and brownfield population use the paired native writer and exact finalization. Exact blob bindings identify inspected source bytes; the linked revision is the committed source base. Implementation, verification, live host qualification and release remain separate claims. Source 9496cff6332fe0b97195b7ebe67b73f62f30e5b3. Select a node for details or use the complete text path. HISTORICAL ROUTES / CODEARBITER codeArbiter / route atlas A command selects an owner. The owner loads context, branches through gates, and returns a bounded result. SOURCE 9496cff6332f / 2026-10-07 01 / ESTABLISH KNOWLEDGE Project initialization feeds later workflows; it does not implement the product. 02 / CREATE A VERIFIED CHANGE Solid arrows are execution. Shared task execution is not a promise that every lane creates a spec artifact. 03 / DELIVER UNDER SEPARATE AUTHORITY A local commit, an open PR, a merge and a release are different outcomes. 04 / INVESTIGATE OR CHALLENGE These routes return findings, not automatic implementation or delivery. 05 / PRESERVE INTENT Decisions govern later work; accepted intent is not implementation evidence. 06 / OPERATE, REPAIR CONTEXT AND FIND HELP Supporting routes below are independent entry points. Their outputs feed a later invocation only when relevant. Execution Execution Execution Execution Context / data: inputs inputs Execution: full full Conditional / authority: small small Execution Execution Execution Execution Execution Execution Execution Execution: fresh lane proof fresh lane proof Execution Execution: authorized adoption / non-behavioral work authorized adoption / non-behavioral work Conditional / authority: request request Conditional / authority: selected selected Conditional / authority: request request Conditional / authority: verified work → commit only when authorized verified work → commit only when authorized /init
--check: report only
Otherwise scaffold, then populate
Sources: S23 /init --check: report only Otherwise scaffold, then populate S23 decompose
Greenfield: six persisted interview layers
Sources: S14 decompose Greenfield: six persisted interview layers S14 context-creation
Qualified scouts → bounded reports
Native paired writes + finalization
Sources: S13, S26 context-creation Qualified scouts → bounded reports Native paired writes + finalization S13 Initialized project knowledge
CONTEXT • stack • standards • security
ADRs • plans • questions • code-map
Provenance supports freshness checks
Sources: S13, S14, S23 Initialized project knowledge CONTEXT • stack • standards • security ADRs • plans • questions • code-map Provenance supports freshness checks S13 Runtime + on-demand context
Startup state; safety-core + active persona
On-demand rules + complete typed task reads
Task-scoped packets delivered to each actual actor
Context is not approval, admission or consumption proof
Sources: S27, S15, S16, S10 CONTEXT / NOT A BULK READ Runtime + on-demand context Startup state; safety-core + active persona On-demand rules + complete typed task reads Task-scoped packets delivered to each actual actor Context is not approval, admission or consumption proof S27 /feature
New behavior; attended workflow
Sources: S06 /feature New behavior; attended workflow S06 /sprint
Coherent goal; approved autonomy
Sources: S07 /sprint Coherent goal; approved autonomy S07 /fix · /refactor
Regression proof / behavior parity
Sources: S19, S21 /fix · /refactor Regression proof / behavior parity S19 /chore · /add-dep
Type-scaled work / vetted adoption
Sources: S20, S04 /chore · /add-dep Type-scaled work / vetted adoption S20 brainstorming → writing-plans
Full: approved spec, then bound plan
Sources: S06, S10 brainstorming → writing-plans Full: approved spec, then bound plan S06 Spec + plan pair approval
One real user reply before execution
Sources: S07, S10 Spec + plan pair approval One real user reply before execution S07 tdd / parity owner
Small feature: confirmed inline mini-spec
Sources: S06, S09, S19, S21 tdd / parity owner Small feature: confirmed inline mini-spec S06 Type / dependency gates
Prose: no TDD; behavior: feature or fix
Sources: S20, S04 Type / dependency gates Prose: no TDD; behavior: feature or fix S20 executing-plans
Scoped batches + human checkpoints
Sources: S06, S21 executing-plans Scoped batches + human checkpoints S06 Full-plan handoff
No attended batch pauses; hard stops stay
Sources: S07 Full-plan handoff No attended batch pauses; hard stops stay S07 Task engine
Fresh actor context → TDD
Spec proof + fresh verification
Combined-scope quality review
Accept only on current evidence
Sources: S08, S27, S09 SUBAGENT-DRIVEN DEVELOPMENT Task engine Fresh actor context → TDD Spec proof + fresh verification Combined-scope quality review Accept only on current evidence S08 Verified change
Current acceptance for typed plans
Regression / parity / type proof
Persistence still needs authority
Sources: S08, S11, S19, S20, S21 Verified change Current acceptance for typed plans Regression / parity / type proof Persistence still needs authority S08 /commit → commit-gate
Explicit permission • nine phases
Focused proof • selective stage
Output: local commit SHA
Sources: S11 /commit → commit-gate Explicit permission • nine phases Focused proof • selective stage Output: local commit SHA S11 /pr → finishing owner
Current-head proof + diff review
Open PR; sprint ends here
Feature retains terminal choice
Sources: S12 /pr → finishing owner Current-head proof + diff review Open PR; sprint ends here Feature retains terminal choice S12 Merge through the PR
Explicit user authority
Exact-head hosted CI + ancestry
No direct-to-default delivery
Sources: S12 Merge through the PR Explicit user authority Exact-head hosted CI + ancestry No direct-to-default delivery S12 /release
Release preparation PR → STOP
Qualified hosted publisher
Separate publication authority
Sources: S33, S34 /release Release preparation PR → STOP Qualified hosted publisher Separate publication authority S33 /watch · /cleanup
/pr --watch / --cleanup
Early branches, not PR creation
No automatic merge or discard
Sources: S02, S04, S12 /watch · /cleanup /pr --watch / --cleanup Early branches, not PR creation No automatic merge or discard S02 /debug · /spike
Debug: five evidenced outcomes
Validated handoff ≠ repair consent
Spike: findings or feature
Sources: S25, S04 /debug · /spike Debug: five evidenced outcomes Validated handoff ≠ repair consent Spike: findings or feature S25 /preview · /review
Predict fleet / review actual diff
/checkpoint persists a sweep
Review verdict is not a commit
Sources: S04, S18, S22 /preview · /review Predict fleet / review actual diff /checkpoint persists a sweep Review verdict is not a commit S04 /threat-model · /tribunal
Distinct opt-in review owners
Tribunal: thirteen lens cards
No routine merge gate
Sources: S29, S04 /threat-model · /tribunal Distinct opt-in review owners Tribunal: thirteen lens cards No routine merge gate S29 /adr · /adr-status
decision-lifecycle
Attributed decision / health report
Sources: S04, S21 /adr · /adr-status decision-lifecycle Attributed decision / health report S04 /reconcile
decision-variance
scout → grader → user ruling
Report-only writes nothing
Sources: S04, S21, S22 /reconcile decision-variance scout → grader → user ruling Report-only writes nothing S04 /conflict
Internal protocol
Contradicting rules: STOP
User resolves the authority gap
Sources: S03, S04 /conflict Internal protocol Contradicting rules: STOP User resolves the authority gap S03 /status · /context-check
v1/v2 state and drift report
Explicit scoped/full refresh selection
Native v2 pairs; preserve history
Sources: S26, S13 /status · /context-check v1/v2 state and drift report Explicit scoped/full refresh selection Native v2 pairs; preserve history S26 /task · /standup
Sanctioned task-board transitions
Hygiene with per-action consent
Sources: S02, S04 /task · /standup Sanctioned task-board transitions Hygiene with per-action consent S02 /audit · /metrics
Dated source-record packet
Windowed governance trends
Sources: S02, S04 /audit · /metrics Dated source-record packet Windowed governance trends S02 /doctor · /statusline
Install / payload diagnostics
Statusline is Claude-specific
Sources: S02, S10 /doctor · /statusline Install / payload diagnostics Statusline is Claude-specific S02 /override · /prune
Permitted, logged exception
Host-specific context compaction
Sources: S02, S03, S04 /override · /prune Permitted, logged exception Host-specific context compaction S02 Watch/cleanup branch before creation preflight. Commit alone does not imply push. Release is not an automatic next step. /commands → grouped catalog /btw → deprecated Q&A /create-context → /init --brownfield /decompose → /init --greenfield Mode tokens are not slash commands: mode --arbiter / --ops / --dangerous. Safety-core remains the floor; gates shown here are arbiter-mode routes. 37 catalog entries accounted for. Canonical, advanced, alias, internal and deprecated visibility is preserved in the selectable HTML explorer. Capability boundaries: brownfield scouts and native writers require qualification; Pi new-HTML admission and HTML farm remain blocked. Read the map
codeArbiter / route atlas A command selects an owner. The owner loads context, branches through gates, and returns a bounded result.
Execution Context / data Conditional / authority Return / repeat Stop / refusal
Choose an entry to highlight its route. Select a node for context, related entries and pinned sources.
Start with a route /feature /fix /sprint /init /review /tribunal /pr /reconcile
Historical route snapshot at 9496cff6332fe0b97195b7ebe67b73f62f30e5b3. Current assistive source bindings identify inspected implementation bytes, with 3549ed56f5d5849116aef8d5e946423d8ee677fc as the committed source base. Live host qualification and release are separate evidence. Curated current-source workflow contracts and selected context-delivery implementation; not a release, installed-host qualification, or exhaustive runtime call graph. Original source links and spatial positions retained; task labels show the selected work owner. Source compatibility revalidated at 25980c292b529fca90c39f5f26d69a1c0674de44. Unittest collector compatibility bound to reviewed blob 76a2998f4bd5b6f441f5fd162bbe810b2de2ee93.
Snapshot limits Historical snapshot: New full-lane feature/sprint defaults to typed HTML only with an admitted installed payload and qualified producers. Main source checkout does not carry the release binary.
Historical snapshot: Pi fails new HTML workflow admission in the pinned artifacts contract. Existing Markdown pairs and inline small features remain separate.
Historical snapshot: HTML --farm dispatch is blocked. Legacy Markdown farm is an off-by-default preview, not the blessed default.
Historical snapshot: Read-time approved-spec pointers index Markdown only in the inspected implementation. Typed HTML execution reads are a distinct engine path.
Historical snapshot: Shared surface generation is not live host parity. Claude/Pi READ notices and Codex’s absence of a native READ event differ.
Historical snapshot: Task quality review is once per current combined scope after per-task spec proof and fresh verification, not a duplicate full fleet for every tiny task.
Historical snapshot: The diagrams describe default arbiter-mode workflow contracts; deterministic mode tokens are separate posture controls and not new slash commands.
Historical snapshot: Snapshots of narrative instructions do not prove every gate is mechanically enforced. No live workflows were executed for this atlas.
Historical snapshot: Brownfield context dispatch needs an actually qualified read-only caller profile; the generic scout charter and generated plugin surfaces alone do not qualify it. Repository-only preparation fixtures are not installed capability.
Historical snapshot: Five content/provenance pairs use independently approved native transactions. Initial finalization has its own 13-identity approval. Initialized refresh retains the marker and never re-enters finalization.
Historical snapshot: Task-context selection is inert. The caller delivers a separate bounded input to each actor; reported delivery, engine tickets and parent session markers are not interchangeable authority or proof.
Historical snapshot: Debug has five dispositions. Validation alone is not evidence truth or repair authority. Diagnosis-only requests stop after the recommendation.
Historical snapshot: Source wording ambiguity: brownfield Phase 3 preserves source-scoped upstream ADR references, while a retained Phase 5 sentence asks for local summaries. Local-summary semantics and authority need clarification; this atlas does not infer automatic local decision authority.
Historical snapshot: Pi approval is currently DEGRADED because the adapter is not implemented; the newer runtime input event means the older HOST-IMPOSSIBLE explanation must not be repeated.
Historical snapshot: Tribunal is opt-in. Its source-bound run store is separate from typed spec/plan authority; report-written does not establish terminal completion.
Historical snapshot: Ten new commits include correctness and waiting-limit repairs, not ten new routes. Their merge descriptions are not independent runtime proof. No latency improvement is asserted.
Historical snapshot: The release PR boundary was under-specified in the previous atlas; making it explicit is a carried presentation correction, not a behavior newly introduced by these ten commits.
Historical snapshot: The native page and offline editions share this source-bound atlas. Website integration does not establish released behavior, installed-host qualification or completed workflow execution.
Complete text routes and handoffs Complete routes and handoffs Historical route snapshot at 9496cff6332fe0b97195b7ebe67b73f62f30e5b3. Current assistive source bindings identify inspected implementation bytes, with 3549ed56f5d5849116aef8d5e946423d8ee677fc as the committed source base. Live host qualification and release are separate evidence. Curated current-source workflow contracts and selected context-delivery implementation; not a release, installed-host qualification, or exhaustive runtime call graph. Original source links and spatial positions retained; task labels show the selected work owner. Source compatibility revalidated at 25980c292b529fca90c39f5f26d69a1c0674de44. Unittest collector compatibility bound to reviewed blob 76a2998f4bd5b6f441f5fd162bbe810b2de2ee93.
Assistive workflow: source-bound implementation An agreed outcome becomes a maintained plan, direct or delegated implementation, proportional verification and an honest delivery report. Task context and dependency advice support progress; real user stops and host permissions remain effective. Greenfield and brownfield population use the paired native writer and exact finalization. Exact blob bindings identify inspected source bytes; the linked revision is the committed source base. Implementation, verification, live host qualification and release remain separate claims.
Current source changes Routine artifact instructions now focus on the installed writer, typed plan, bounded task reads, factual prerequisites, actual checks and safe retry. The default leaf is about 7.3 KB and includes an executable stdin bootstrap with optional operation-schema discovery; explicit attestation remains available on demand without burdening ordinary task context. A spec's SMARTS design rationale lives in its approach, decisions and one SEC-SMARTS section. core/surface/includes/artifacts.md Plans preserve agreed criteria, useful tasks, dependency ancestry, bounded shared inputs and observable checks. Stable source identities and typed maintenance keep changes within the selected outcome. A design-drift guard routes an undecided material solution choice back to brainstorming instead of scoring it during planning. core/surface/skills/writing-plans/SKILL.md The pinned native bridge enables task-context as a bounded informational read and does not require future reviewer/approval producers simply to author a structurally valid draft. Context requests route to ordinary context-preview, context-finalize-preview and exact-reference context-apply; historical receipt input stays a separate compatibility path. Installed host qualification is separate. Native Pi writer admission does not establish a complete live Pi workflow. On NO_LOCAL_STORE the client runs init-store once and retries the request once; a refused initialization keeps NO_LOCAL_STORE. On Windows the binary's architecture comes from the kernel (IsWow64Process2), never WMI or PROCESSOR_* environment. Index and outline pages are now fully checked before rows are yielded, including continuation identity and row shape. core/pysrc/_artifactlib.py A new explicit batch review can select a passing rerun of still-reviewed work without replacing the original task review. The selected run must match its own recorded workspace; changed source or supporting evidence still requires reassessment. core/artifacts/internal/evidence/completion.go The user-facing route table points feature and sprint work at useful planned execution and selected context refresh at its existing owner. Existing authorization covers ordinary work; actual host, destructive-action and commit controls remain. core/surface/includes/routing-table.md The skill index connects planning, direct or delegated implementation, proportional verification and selected context maintenance to their owning workflows. core/surface/skills/INDEX.md The sprint caller performs an index-first recorded-intent read, preserves actual owner decisions and ADR conflicts, then proceeds with authorized work without a second artifact approval system. A SMARTS Unknown names its missing observation and cannot bypass a hard gate or manufacture authority. core/surface/SPRINT.md The sprint entry now routes to the assistive implementation body instead of requiring a fresh user approval and two reviews per task. core/surface/commands/sprint.md The feature entry now resumes a maintained spec/plan and task progress without another artifact certification prompt; independent commit and test controls remain. The small lane runs an inline SMARTS scan only for a material solution choice. core/surface/commands/feature.md Source refinement respects recorded deferrals and accepted ADR conflicts without turning each ordinary method adjustment into a separate authorization interview. SMARTS compares approaches only when two or more are materially plausible; a single approach gets one line and no table. core/surface/skills/brainstorming/SKILL.md Attended feature execution now treats checkpoints as progress groups, not compulsory approval stops; it forwards real failures and continues independent work. core/surface/skills/executing-plans/SKILL.md Implementation works directly or delegates a coherent unit when useful. Bounded task context, applicable dependency advice and observed checks support progress while preserving actual user stops and host permissions. Verification checks the change against a recorded SMARTS rationale for material invalidation only. core/surface/skills/subagent-driven-development/SKILL.md Confirmed fixes require regression RED/GREEN. Features use current criterion-focused checks and meaningful negative controls where needed; refactors preserve passing characterization baselines. Coverage, lint and applicable security requirements remain explicit. core/surface/skills/tdd/SKILL.md The commit owner checks authorization, the staged diff, secrets and the verification approach appropriate to the change. Current criterion checks, genuine fix regressions and refactor parity retain their distinct evidence requirements. Reuse applicable observed checks and reviews across phases. Managed v2 maintenance selects affected claims and uses native paired writes; unknown evidence defers that document. core/surface/skills/commit-gate/SKILL.md PR readiness uses the explicit base-to-HEAD subject for current checks, secrets scanning and applicable specialized review. Existing evidence is reused only while its source and inputs match. Merge requires authorization and exact-head hosted CI. core/surface/skills/finishing-a-development-branch/SKILL.md Read-only task-context reports direct dependencies and unresolved ancestors, scoped_changed, scoped_unchanged or unassessed inputs, and actual check observations. Missing broader context remains explicit; the packet does not certify verification. core/artifacts/internal/operations/task_context.go An unanswered legacy reconciliation can be superseded only when route and pending state agree and no durable attempt or matching observed reply exists. Captured replies recover through their original immutable source and receipt; in-flight mutations retain exact replay and real user stops. core/pysrc/_reconciliationlib.py The installed recover command finishes a captured original reply or replays an in-flight mutation without asking for a new user answer. core/pysrc/artifact-reconcile.py Task-check observations now retain directly referenced criterion identity and an optional bounded source-input fingerprint or an explicit unassessed reason, without executing tests or minting approval receipts. core/artifacts/internal/operations/task_check.go New checks bind the selected criterion and applicable constraint bodies, including explicit references and scope applicability. Unrelated constraints remain outside the digest; old observations never gain unassessed constraint coverage. core/artifacts/internal/model/assistive.go Exact and batch symbol retrieval exposes reported check history without an approval ticket. Outline paging now accounts for encoded rows incrementally while retaining response budgets and continuation behavior. core/artifacts/internal/read/read.go Human HTML presents reported command checks in an escaped, explicitly unverified section; implementation state and source validation remain distinct. core/artifacts/internal/render/render.go The closed task-check request accepts bounded command observations matched to plan definitions. Ordinary context-preview and context-finalize-preview omit prompt identity; context-apply accepts an exact preview reference and digest. Invalid requests, changed preimages and stale CAS writes remain rejected. The historical smarts-apply request accepts the 0.2.0 SMARTS decision record. init-store takes a protocol-only request. core/artifacts/internal/operations/protocol.go The unpublished 0.3.2 native schema adds optional shared_inputs and scoped task-check fields without silently relabeling old records. core/artifacts/schemas/common.schema.json The bounded source-fingerprint reader hashes actual declared task/shared paths, directory membership and known build-control files; unsafe, missing and over-limit scopes are explicitly unassessed rather than blocking work or proving a full dependency graph. core/artifacts/internal/operations/input_scope.go The candidate-only runbook distinguishes 0.3.1/0.3.2 mixed-schema readers, exact replay and forward recovery while preserving recorded user stops. The paired prompt retirement scans immutable observed replies before calling a request unanswered; submitted or uncertain decisions are never erased. Older-client downgrade after modern edits remains unsupported. docs/artifacts/ASSISTIVE-UPGRADE.md Ordinary approval and sprint instructions bypass legacy capture without changing pending history or inventing authority. Explicit legacy replies retain exact matching. The paired adapter can retire a genuinely unanswered 0.2 sprint-pair request using its serialized cancellation lock and submission check; submitted decisions must replay, while standalone approval 0.1 refuses automatic retirement. core/pysrc/_approvallib.py Ordinary prerequisite instructions bypass legacy capture. The existing adapter can retire only a validated ARMED pending question under its original per-plan OS lock. Confirmed or captured decisions are not disposable; no authority receipt or new native permission API is created. core/pysrc/_prerequisitelib.py The old paired approval producer publishes a content-addressed observed reply before its durable submission. Cancellation checks the matching observation, content hash, prompt and frozen context under its existing lock, with a strict scan bound. A matching or uncertain pre-submission reply blocks automatic retirement; normal tasks and context reads have no added cost. core/pysrc/_sprintapprovallib.py The optional reference retains strict historical evidence inputs without replacing ordinary work. Explicit draft source refresh preserves progressed plans and history with both CAS bindings; initial approval restrictions remain on historical paired attestation. Its verify command requires literal absolute paths; linked agent subdirectories remain unsupported for strict authority review. These limitations apply to requested attestation, not ordinary task start. core/surface/includes/artifacts-attestation.md The command catalog starts from the agreed outcome and maintained plan; SMARTS is useful for difficult tradeoffs and does not authorize a changed outcome. core/surface/COMMANDS.md Resident instructions preserve existing work authorization and actual owner decisions without another spec approval or a SMARTS record for every routine step. core/surface/arbiter.md Status separates reported DONE from independently ACCEPTED work, shows actual blocked decisions, and does not infer an approval request from draft governance state. core/surface/commands/status.md Delegated authors preserve the caller-selected verification approach: genuine regression RED/GREEN for fixes, current criterion-focused feature checks and passing characterization baselines for refactors. Actual results and pending checks return to the caller; staging requires assigned scope and authorization. core/surface/includes/author-tdd-workflow.md Codex dispatch requires attestation only for explicitly claimed host proof; an unavailable requested farm backend is reported without silently switching provider or model. Ordinary context inspection can proceed inline within actual host permissions; an agent label supplies no containment or semantic verification. core/surface/includes/codex-host-notes.md Historical sprint approval and SMARTS proof instructions remain available only for existing explicit attestation, outside the ordinary execution path. core/surface/includes/sprint-authority.md The public Codex farm boundary matches its caller: report the unavailable backend and use another provider or model only within existing authorization. docs/parity.md The preparation helper binds bounded drafts and retained sources to exact document/provenance preimages. Finalization preparation hashes all thirteen required outputs and retains missing output failures without modifying files. core/pysrc/_contextwriterlib.py The installed context-draft command prepares a paired document preview; context-finalize prepares initial finalization. context-apply commits the exact selected preview and digest. Context modes preserve scaffold and initialized refresh boundaries. core/pysrc/init-codearbiter.py Ordinary context-preview retains an inert proposal and exact preimages without approval or review claims. context-apply revalidates the digest, closed request and preview binding before the native paired transaction; context-finalize-preview uses the same exact-reference path for initial population. No approve-context reply is required. core/artifacts/internal/operations/context_assistive.go The finite context contract permits a whole-document Markdown entry only as the sole entry for one of five named content targets. Explicit typed field entries remain available; the writer does not expose arbitrary repository-file writes. core/artifacts/internal/contextdocument/contract.go Whole-document replacement requires the explicit document selection and exact existing bytes. It preserves CONTEXT frontmatter and initialization state, checks the named document identity and Markdown structure, and retains the code-map size limit. Human content outside the authorized change remains the caller's preservation obligation. core/artifacts/internal/contextdocument/document_selection.go Ordinary create/adopt/update preserves the same source-evidence, selection and compare-and-swap checks as the native transaction. Missing or matching v1 provenance can be adopted; existing v2 ownership can transition through explicit whole-document selection. Original document/provenance preimages remain in transaction history for exact recovery. core/artifacts/internal/contextdocument/operations.go Initial finalization binds thirteen output identities, rechecks supporting source evidence, and changes the initialization marker with its paired provenance digest. Ordinary finalization requires no artifact approval receipt and does not certify semantic correctness. Already initialized repositories refuse a new finalization. core/artifacts/internal/contextdocument/finalization.go The context schema describes both finite field entries and the explicit whole-document entry for the five supported content documents. Its typed structure records evidence references; schema validity does not establish that the claims were independently reviewed. core/artifacts/schemas/context.schema.json Brownfield population covers six logical categories and five content documents through direct bounded inspection or optional scoped agents. The paired writer preserves source bindings and protected content; initial finalization checks all required outputs. core/surface/skills/context-creation/SKILL.md A drift audit stays read-only; an already authorized selected or full refresh proceeds without asking again per document. The installed draft/preview/apply example preserves source bindings and selected content. Matching hashes establish identity only; initialized refresh never reruns scaffolding, clears the marker or repeats finalization. core/surface/skills/context-check/SKILL.md Initial brownfield setup checks the intended repository and actual host permissions, then uses bounded inspection and the installed safe context writer. Ordinary Windows work needs no special scout profile or per-document artifact approval. An existing initialized repository routes to selected refresh instead of another scaffold or population cycle. core/surface/commands/init.md Optional context scouts report only their assigned scope, source identities, coverage, conflicts and unknowns. Six logical categories do not require six child agents. A role label or read-only instruction proves neither sandbox containment nor independent semantic verification; actual host permissions and excluded paths still apply. core/surface/agents/scout.md Release preparation inserts a new version below the entire Unreleased section and above prior versions, preserving pending notes. This guidance correction does not grant publication authority or replace the existing release-PR and hosted qualification boundaries. core/surface/skills/release/SKILL.md The lazy reference map routes provenance drift or requested initialized refresh to context-check after reading selected documents, provenance and the code map. The owner reuses selected scope and existing authorization. core/surface/includes/reference-map.md Authors and reviewers share task-critical reference and constraint checks. Missing or stale provenance remains advisory. Two delivery attempts end with an advisory unknown observation, without another retry or an invented receipt. core/pysrc/_contextselectlib.py Feature actors receive scoped requirements, named critical references and observed evidence. Applicable conflicts or missing required inputs stop dependent work; provenance warnings remain visible and advisory. core/pysrc/_artifactpromptlib.py Confirmed fixes preserve causal evidence and a named regression. Delegated input retains critical requirements; advisory delivery exhaustion ends retries while actual missing inputs and conflicts retain their dependent-work boundary. Distinct remediation strategies are compared with SMARTS only after the root cause is confirmed. core/surface/commands/fix.md A review binds one resolved diff and supplies scoped inputs to each reviewer. Missing critical references or applicable conflicts prevent a substantive verdict; provenance and unobserved delivery diagnostics remain advisory. core/surface/commands/review.md Characterization tests can pass against existing correct behavior. Preserve behavioral assertions against the post-backfill baseline; mechanical import or symbol-reference updates retain the same oracle. Coverage and parity checks remain required. A precise surface table proceeds under existing user authorization; only material expansion or a changed outcome needs another decision. Distinct restructuring strategies are compared with SMARTS only after surface and parity are settled. core/surface/skills/refactor/SKILL.md Greenfield interviews persist under plans/decomposition and retain actual decisions. Five document/v2-provenance pairs use context-draft and context-apply; context-finalize binds all thirteen outputs. ADRs and task rows use their owners, and retained interview evidence survives finalization. core/surface/skills/decompose/SKILL.md Scheduling follows the full dependency ancestry and exposes unresolved ancestor IDs and reasons. Reconciled progress remains advisory; blocked ancestors prevent premature downstream readiness while independent tasks remain available. Response-budget omissions are explicit, retain the global scheduling decision and direct callers to supported detail reads. core/artifacts/internal/operations/assistive.go Explicit draft source refresh retains both source CAS components and same-pair identity without requiring an unstarted plan. Existing progress, approvals and evidence are preserved without asserting current authority. Dependency invalidation preserves an independently BLOCKED task and its reason. Verification summaries disclose conflicting caller reports separately from receipt freshness; later successful verification covers only reports present when its context was issued. core/artifacts/internal/operations/lifecycle.go A checkpoint retains the reviewed commit, UTC review-start time and worktree dirty state, then writes the report through its aggregator. last-checkpoint holds only the override count. core/surface/commands/checkpoint.md Checkpoint reports persist observed source_commit, recorded_at and worktree_dirty values together with every review finding. Missing anchors remain unknown. core/surface/agents/checkpoint-aggregator.md Checkpoint audit windows use the report timestamp and an ancestor commit. Invalid or missing anchors require an explicit ref/date window; dirty review scope remains a stated limitation. core/surface/commands/audit.md Bounded read-time pointers include ordinary draft and ready HTML specs as advisory context. Rejected or withdrawn specs stay excluded; no pointer confers approval. core/pysrc/_readinjectlib.py Routine maintenance selects changed v2 claims intersecting staged sources using a current bounded snapshot. Unknown evidence defers the affected document; native paired writes preserve owned fields and current identities. core/pysrc/_provenancelib.py Ordinary uppercase words such as Approve DATA migration remain ordinary input. All-letter capture codes require a matching retained request; real armed replies, mismatches and user stops preserve their original handling. core/pysrc/_replylib.py Only an actual verifier wrapper executable/script prefix selects authority handling. Inert PowerShell data mentioning the filename is not a verifier call; malformed real verifier invocations still fail closed. Fresh legacy context approval arming is refused; historical pending routes remain readable. core/pysrc/_artifactauthoritylib.py Bounded isolated Python stdin diagnostics may read literal protected paths or git show HEAD objects. Unknown code, mutable calls, shell expansion and extra commands retain the guard checks; ordinary commands avoid AST parsing. core/pysrc/_bashguardlib.py /add-dep /adr /adr-status /audit /btw /checkpoint /chore /cleanup /commands /commit /conflict /context-check /create-context /debug /decompose /doctor /feature /fix /init /metrics /override /pr /preview /prune /reconcile /refactor /release /review /spike /sprint /standup /status /statusline /task /threat-model /tribunal /watch Operate / alias
/context-check context-check
Canonical entry: status drift
Route to outcome Default status and initial drift report are read-only Assess legacy v1 drift and each enrolled v2 identity/coverage state separately; unknown is not fresh Wait for explicit selected documents and affected scope, or explicit full refresh; defer changes nothing Re-scout only the selected assignments through qualified containment and a bounded current join For v2, approve the exact owned-field document/provenance preview and apply only its native receipt Read back committed pairs; retain the initialized marker, unselected human content and separate task/question/audit history; do not stage or commit Selected context Current bounded source/target snapshots and action-boundary checks
v1 drift separately from v2 current/stale/unknown coverage; identity matching is not semantic proof
Explicit document selection, actual field ownership, qualified scouts and native writer capability
Output Read-only state/drift report, or explicitly selected native context updates with declared full/incremental coverage.
Report alone is not refresh consent. No reinitialization or marker removal. No legacy raw rebaseline fallback for v2, no unsupported human-layout overwrite, no staging or commit.
Host / scope Claude Code / Codex / Pi; admission and runtime capabilities still apply
claude /ca:context-check
codex $ca-context-check
pi /ca-context-checkRead the exact procedure
Related entries Pinned evidence S26 · core/surface/skills/context-check/SKILL.mdInitialized refresh: read-only report, explicit selection, v1/v2 separation, native paired updates and preserved human history. Full owner read. S13 · core/surface/skills/context-creation/SKILL.mdCurrent brownfield owner: qualified containment, bounded reports and join, scoped gaps, native document/provenance writes, separate 13-identity finalization and recovery. S23 · core/surface/commands/init.mdCurrent init owner: activation/scaffold versus initial population; initialized refresh belongs to status drift, not reinitialization. Knowledge / alias
/create-context context-creation
Canonical entry: init --brownfield
Route to outcome Uninitialized project with meaningful source → bounded brownfield initial population Brownfield: qualified read-only scout profile → closed reports → scoped synthesis and relevant gap decisions Approve each native document/provenance preview; receipt-only apply, then separately approved finalization Finalize the 13 required identities through its own native preview/receipt; recover uncertain operations by the same operation ID Selected context Same bounded brownfield source, containment, report, ownership and provenance contract as init --brownfield
Output Five current content/provenance pairs and separately finalized initialized state, or a precise capability/evidence conflict.
Initial population only; existing initialized state routes to status drift. Generic scout role or a source-only preparation fixture does not establish qualified containment or writer admission.
Host / scope Claude Code / Codex / Pi; admission and runtime capabilities still apply
claude /ca:create-context
codex $ca-create-context
pi /ca-create-contextRead the exact procedure
Related entries Pinned evidence S13 · core/surface/skills/context-creation/SKILL.mdCurrent brownfield owner: qualified containment, bounded reports and join, scoped gaps, native document/provenance writes, separate 13-identity finalization and recovery. S23 · core/surface/commands/init.mdCurrent init owner: activation/scaffold versus initial population; initialized refresh belongs to status drift, not reinitialization. S26 · core/surface/skills/context-check/SKILL.mdInitialized refresh: read-only report, explicit selection, v1/v2 separation, native paired updates and preserved human history. Full owner read. Change / advanced
/debug debug
Route to outcome Capture the cited symptom, expectation, reproduction state and actual request scope Test evidence-backed mechanisms using safe discriminators; preserve inconclusive observations and capture limits Choose exactly one disposition: confirmed_code_defect, confirmed_noncode_cause, design_question, no_action or unresolved For transfer, validate the bounded transient debug-handoff/1.0.0 packet with the installed helper Return the diagnosis and truthful continuation; only an already-authorized fix caller can continue a qualifying code-defect handoff Selected context Available project instructions and scoped source evidence, not compulsory new onboarding
Observed versus expected behavior, reproduction status, supported causes and regression oracle
Relevant worktree, source/build/runtime identity, dirty state, limitations and independent caller authority
Output One evidence-backed disposition and a bounded summary; transient validated handoff when available. No default persistent case store, task or ADR.
Diagnosis never authorizes repair. Invalid/unavailable handoff validation stops transfer; recheck current evidence. No-action requires positive closure evidence. A design question is not an automatic ADR.
Host / scope Claude Code / Codex / Pi; admission and runtime capabilities still apply
claude /ca:debug
codex $ca-debug
pi /ca-debugRead the exact procedure
Related entries Pinned evidence S25 · core/surface/skills/debug/SKILL.mdFive-disposition diagnostic owner; transient validated handoff, freshness and authority boundaries. Full owner read. S19 · core/surface/commands/fix.mdCurrent full fix owner: validated handoff, independent caller authority, evidence revalidation, selected actor input and genuine target-red gate. S27 · core/pysrc/_contextselectlib.pyInert scoped selector and delivery implementation: commands/cwd/evidence, native instruction order, actor/worktree/epoch binding, two-attempt limit and reviewer preflight. S04 · core/surface/includes/routing-table.mdAuthoritative intent-to-owner table, scope-touch routes and stop/return semantics. Change / core
/feature brainstorming → writing-plans → executing-plans
Route to outcome Discover an existing exact-format pair and resume before new triage Small: confirm 1–3 criteria and log triage, then TDD Full: approve specification, bind/approve plan, then attended batches Fresh task authors, spec proof, scope-quality acceptance Before each author or independent reviewer acts, compose its own current scoped input and perform bounded actor delivery; a coordinator read is not receipt Current acceptance → authorized commit → selected PR disposition Selected context Project purpose, stage and code-map
Scope-specific rules and recorded ADR intent
Approved criteria and exact plan binding
Task paths, dependencies, verification and complete context ticket
Current actor-scoped selection: map/provenance, command cwd and observed evidence, constraints and host-effective instructions
Actual author and independent reviewer delivery bound to task/worktree/source/host epochs; reselect after resume, compaction or scope change
Use exact engine-returned IDs, not assumed SPEC-/PLAN- prefixes. Approved prerequisites use one visible 24-hour request/reply deadline.
Output Full lane: typed spec + plan, accepted code/tests, evidence and selected PR outcome. Small lane: inline mini-spec, no spec artifact.
No guessed unknowns; no implementation before red; batch checkpoints; no mixed-format resume; installed HTML admission required. Selected context is orientation, not task admission. Small features retain their confirmed inline mini-spec without invented artifact IDs.
Host / scope Claude Code / Codex: admitted installed typed workflow only. Pi 1.0.2 is the current runtime target, but its native approval adapter remains unimplemented (DEGRADED); legacy Markdown pairs and inline small-feature work have separate contracts.
claude /ca:feature
codex $ca-feature
pi /ca-featureRead the exact procedure
Related entries Pinned evidence S06 · core/surface/commands/feature.mdCurrent feature owner: same-format resume and full/small gates; adds actual per-actor context composition. S08 · core/surface/skills/subagent-driven-development/SKILL.mdCurrent implementation-dispatch and independent spec-review sections re-read; remaining scope contract also checked against current skill index. Fresh actor input and bounded delivery are distinct from engine tickets. S09 · core/surface/skills/tdd/SKILL.mdCurrent TDD context preflight and obligation derivation re-read. Six-phase gate sequence also confirmed by current skill index; fix and test-only paths do not fabricate feature artifacts. S10 · core/surface/includes/artifacts.mdInstalled admission, typed reads, context completion, authority receipts and format boundaries. S11 · core/surface/skills/commit-gate/SKILL.mdCurrent Phase 5.5 through terminal commit rules re-read. Legacy heal vocabulary must not be generalized to v2 paired native writes. S12 · core/surface/skills/finishing-a-development-branch/SKILL.mdWatch/cleanup early routing; commit and acceptance preflight; PR and merge authority. S27 · core/pysrc/_contextselectlib.pyInert scoped selector and delivery implementation: commands/cwd/evidence, native instruction order, actor/worktree/epoch binding, two-attempt limit and reviewer preflight. S31 · docs/parity.mdCurrent Pi 1.0.2 target, separate host counts, DEGRADED typed approval and Codex farm exception. S32 · core/pysrc/_artifactauthoritylib.pyNative artifact ID grammar, producer states, Claude join lock, UTF-8/UTF-16-key-order observation binding. Change / core
/fix tdd — bug variant
Route to outcome Establish the confirmed defect from current reproduction or supported source/contract causal evidence If needed, use one bounded internal debug prerequisite and return to this same authorized fix caller Validate any available handoff; independently recheck caller authority and actual worktree/source/runtime applicability Select and deliver bounded current context to the actual fix author Observe the named regression red for the correct defect, then implement the minimum fix and complete TDD Return proof; commit, PR and other follow-ups retain their separate authority Selected context Current expectation, causal evidence, reproduction limits and named regression oracle
Task paths; scoped map/provenance; command cwd and current collector evidence
Applicable constraints and native effective instructions; actor/worktree/source/host epoch binding
Output Minimal authorized correction with genuine failing-before / passing-after regression evidence.
No repair from packet intent alone. Only a validated confirmed_code_defect → fix handoff can resume an authorized repair. Setup/import/unrelated red is not target-red proof. No feature artifact is invented.
Host / scope Claude Code / Codex / Pi; admission and runtime capabilities still apply
claude /ca:fix
codex $ca-fix
pi /ca-fixRead the exact procedure
Related entries Pinned evidence S19 · core/surface/commands/fix.mdCurrent full fix owner: validated handoff, independent caller authority, evidence revalidation, selected actor input and genuine target-red gate. S25 · core/surface/skills/debug/SKILL.mdFive-disposition diagnostic owner; transient validated handoff, freshness and authority boundaries. Full owner read. S27 · core/pysrc/_contextselectlib.pyInert scoped selector and delivery implementation: commands/cwd/evidence, native instruction order, actor/worktree/epoch binding, two-attempt limit and reviewer preflight. S09 · core/surface/skills/tdd/SKILL.mdCurrent TDD context preflight and obligation derivation re-read. Six-phase gate sequence also confirmed by current skill index; fix and test-only paths do not fabricate feature artifacts. Knowledge / core
/init init-codearbiter.py → context-creation / decompose
Route to outcome --check reports only; activation/scaffold is a separate action Uninitialized project with meaningful source → bounded brownfield initial population Brownfield: qualified read-only scout profile → closed reports → scoped synthesis and relevant gap decisions Approve each native document/provenance preview; receipt-only apply, then separately approved finalization No meaningful source → retain the separate six-layer greenfield interview and persisted drafts Already initialized → normal operation; a selected refresh belongs to status drift, not init Selected context Source shape and maturity; existing file ownership and initialization state
Qualified scout reports and current source snapshots for brownfield, persisted interview answers for greenfield
Output Initialized project knowledge after the selected owner completes its real gates; a stub alone is not ready.
Brownfield profile or qualified native writer unavailable → block that dependent step. Five document/provenance pairs have independent approvals; finalization has its own 13-identity preview and receipt. Never clear the marker to refresh.
Host / scope Claude Code / Codex / Pi; admission and runtime capabilities still apply
claude /ca:init
codex $ca-init
pi /ca-initModes --check · --stage N · --brownfield · --greenfield
Read the exact procedure
Related entries Pinned evidence S23 · core/surface/commands/init.mdCurrent init owner: activation/scaffold versus initial population; initialized refresh belongs to status drift, not reinitialization. S13 · core/surface/skills/context-creation/SKILL.mdCurrent brownfield owner: qualified containment, bounded reports and join, scoped gaps, native document/provenance writes, separate 13-identity finalization and recovery. S14 · core/surface/skills/decompose/SKILL.mdCurrent greenfield interview and completion contract re-read. Its persisted six-layer draft flow remains separate from the brownfield writer lifecycle. S26 · core/surface/skills/context-check/SKILL.mdInitialized refresh: read-only report, explicit selection, v1/v2 separation, native paired updates and preserved human history. Full owner read. Operate / advanced
/metrics read-only trend report
Route to outcome Read selected and prior windows Summarize overrides, small lanes and low-confidence choices Return comparison Selected context Override, triage and sprint decision logs
Output Trend report against prior window.
Reporting is not state mutation or workflow approval.
Host / scope Claude Code / Codex / Pi; admission and runtime capabilities still apply
claude Not exposed on this host
codex Not exposed on this host
pi Not exposed on this host Modes --window N
Removed on 2026-10-23; no host exposes this route.
Related entries Pinned evidence S02 · core/surface/COMMANDS.mdCatalog, modes, host-limited commands and purpose of each public route. Ship / core
/release release
Route to outcome Resolve the declared target, version policy, payload and reviewed executable inputs; dry-run reports only and skips writes and execution. Prepare version/changelog and declared release surfaces through the ordinary commit gate on a non-default branch. Report the exact assets and publisher, open the release PR and STOP. Interactive preparation does not compose a local tag. Only after the PR merges may a qualifying hosted publisher continue at the exact merged, fetched and verified default-branch commit. Compose and validate the exact tag/cohort only through that publisher. Apply the existing one-cohort publication authority; CI success alone is not publication permission. Read back the exact non-draft release and declared inventory. A declared provenance closeout uses its own receipt PR; an already-published receipt-only path never publishes again. Selected context Declared target row, exact manifest/changelog/payload ownership
Exact release-surface commit, fetched default branch and current hosted evidence
Existing publication authority for this cohort, exact assets and any declared provenance requirement
Output A release-preparation PR; separately, an authorized hosted publication and its exact readback. Declared receipt closeout remains a separate PR.
Mandatory release-PR stop. No local tag fallback, red/stale-head publication or implicit consent. Dry-run does not fetch, execute checks, commit or publish.
Host / scope Claude Code / Codex / Pi; admission and runtime capabilities still apply
claude /ca:release
codex $ca-release
pi /ca-releaseModes --dry-run
Read the exact procedure
Related entries Pinned evidence S33 · core/surface/skills/release/SKILL.mdInteractive release preparation and mandatory release-PR boundary; qualified hosted tag composition. S34 · site/src/content/docs/concepts/workflow-routes.mdxExisting public route-comparison owner; distinct release/initialization endpoints and source-only explanation boundary. Change / core
/sprint SPRINT.md → subagent-driven-development
Route to outcome Resolve same-format artifacts; retain existing approved intent Prepare spec and draft plan for one real user approval of the pair Execute the full plan without attended batch pauses SMARTS governs delegated in-scope choices; correction must rerun failed gates Before each author or independent reviewer acts, compose its own current scoped input and perform bounded actor delivery; a coordinator read is not receipt Current acceptance → authorized commit → open PR only Selected context Project rules and recorded intent
Approved spec/plan and explicit delegation terms
Task-specific contextual reads
Actual review, verification and authority receipts
Current actor-scoped selection: map/provenance, command cwd and observed evidence, constraints and host-effective instructions
Actual author and independent reviewer delivery bound to task/worktree/source/host epochs; reselect after resume, compaction or scope change
Use exact engine-returned IDs, not assumed SPEC-/PLAN- prefixes. Approved prerequisites use one visible 24-hour request/reply deadline.
Output Accepted plan and implementation, sprint decision log, open PR and summary. No autonomous merge.
Initial approval never self-issued. Security/authority/irreversible-action stops remain. HTML --farm blocked; legacy farm is opt-in preview. Selected context is orientation, not task admission.
Host / scope Claude Code / Codex: admitted installed typed workflow only. Pi 1.0.2 is the current runtime target, but its native approval adapter remains unimplemented (DEGRADED); legacy Markdown pairs and inline small-feature work have separate contracts.
claude /ca:sprint
codex $ca-sprint
pi /ca-sprintModes --farm
Read the exact procedure
Related entries Pinned evidence S07 · core/surface/SPRINT.mdInitial paired approval; full-plan execution; delegated methods; recovery and farm boundaries. S08 · core/surface/skills/subagent-driven-development/SKILL.mdCurrent implementation-dispatch and independent spec-review sections re-read; remaining scope contract also checked against current skill index. Fresh actor input and bounded delivery are distinct from engine tickets. S10 · core/surface/includes/artifacts.mdInstalled admission, typed reads, context completion, authority receipts and format boundaries. S12 · core/surface/skills/finishing-a-development-branch/SKILL.mdWatch/cleanup early routing; commit and acceptance preflight; PR and merge authority. S27 · core/pysrc/_contextselectlib.pyInert scoped selector and delivery implementation: commands/cwd/evidence, native instruction order, actor/worktree/epoch binding, two-attempt limit and reviewer preflight. S09 · core/surface/skills/tdd/SKILL.mdCurrent TDD context preflight and obligation derivation re-read. Six-phase gate sequence also confirmed by current skill index; fix and test-only paths do not fabricate feature artifacts. S31 · docs/parity.mdCurrent Pi 1.0.2 target, separate host counts, DEGRADED typed approval and Codex farm exception. S32 · core/pysrc/_artifactauthoritylib.pyNative artifact ID grammar, producer states, Claude join lock, UTF-8/UTF-16-key-order observation binding. Operate / core
/status status / context-check
Route to outcome Default status and initial drift report are read-only Assess legacy v1 drift and each enrolled v2 identity/coverage state separately; unknown is not fresh Wait for explicit selected documents and affected scope, or explicit full refresh; defer changes nothing Re-scout only the selected assignments through qualified containment and a bounded current join For v2, approve the exact owned-field document/provenance preview and apply only its native receipt Read back committed pairs; retain the initialized marker, unselected human content and separate task/question/audit history; do not stage or commit Selected context Current bounded source/target snapshots and action-boundary checks
v1 drift separately from v2 current/stale/unknown coverage; identity matching is not semantic proof
Explicit document selection, actual field ownership, qualified scouts and native writer capability
Output Read-only state/drift report, or explicitly selected native context updates with declared full/incremental coverage.
Report alone is not refresh consent. No reinitialization or marker removal. No legacy raw rebaseline fallback for v2, no unsupported human-layout overwrite, no staging or commit.
Host / scope Claude Code / Codex / Pi; admission and runtime capabilities still apply
claude /ca:status
codex $ca-status
pi /ca-statusModes drift
Read the exact procedure
Related entries Pinned evidence S26 · core/surface/skills/context-check/SKILL.mdInitialized refresh: read-only report, explicit selection, v1/v2 separation, native paired updates and preserved human history. Full owner read. S13 · core/surface/skills/context-creation/SKILL.mdCurrent brownfield owner: qualified containment, bounded reports and join, scoped gaps, native document/provenance writes, separate 13-identity finalization and recovery. S23 · core/surface/commands/init.mdCurrent init owner: activation/scaffold versus initial population; initialized refresh belongs to status drift, not reinitialization. Operate / advanced
/statusline Claude Code statusline control
Route to outcome Select install, uninstall or status Change or inspect the owned Claude statusline configuration Return status Selected context Claude Code install and statusline settings
Output Statusline configuration or status report.
Claude-specific catalog surface; not a cross-host workflow.
Host / scope Claude Code only
claude /ca:statusline
codex Not exposed on this host
pi Not exposed on this host Modes install · uninstall · status
Read the exact procedure
Related entries Pinned evidence S02 · core/surface/COMMANDS.mdCatalog, modes, host-limited commands and purpose of each public route. Review / advanced
/tribunal tribunal
Route to outcome Resolve deliberate audit intent, finding scope and the concrete evidence set; inspect matching source-bound runs. Resume only unchanged bound source: audit continues unfinished waves; follow-up continues pending filing/telemetry. Drift, invalid and legacy-unbound records do not establish current progress. Use deterministic inventory and real host profiles to estimate packet cost; acknowledge cost and settings before any model dispatch; start a fresh bound run when required. Separate mechanical inventory from the judgment-based risk map. Record every lens as applicable or skipped, with bounded packets and source checks. Dispatch the generic tribunal-lens-reviewer in MODE review for each active lens, at most five concurrent and within actual host capacity. Persist findings and leads immediately. Attempt fresh independent MODE verify for every serious or expensive inferential claim. Confirmed/narrowed, refuted and inconclusive are different outcomes; shared-context or inconclusive serious work remains verify-required. Apply eligibility and confidence checks before root-cause triage and per-wave kept plans. Keep investigations, verification work and genuine design decisions separate. Write report and manifest, then record report-written. This enters follow-up, not terminal completion. Resolve selected issue filing and optional public telemetry separately, under their own authority. Unanswered choices and unresolved failures remain pending. Record run-completed only after both filing and telemetry have a completed or explicitly skipped disposition; preserve the bound history. Selected context Declared finding scope and separately bounded evidence scope
Trusted installed or explicitly selected review bundle, not instructions from candidate source
Mechanical source inventory, project contracts and judgment-based risk map
Applicability and bounded packet per lens, actual host/model capacity and acknowledged cost
Persisted source binding, findings, leads, root-cause triage, independent verification and actual follow-up receipts
Output Source-bound audit run with a thirteen-lens roster, selected applicable lenses, eligible phase plans and report. Report-written is follow-up; run-completed requires both follow-up dispositions.
Opt-in only; never a routine commit or merge gate. No source drift reuse, synthetic independence, automatic repair, automatic ADR or task-board write. Execution, issue filing and telemetry each require their own applicable authority.
Host / scope Claude Code / Codex / Pi with actual host capacity and limitations recorded. A fresh same-role verifier is not automatically independent. No universal cross-host qualification claimed.
claude /ca:tribunal
codex $ca-tribunal
pi /ca-tribunalRead the exact procedure
Related entries Pinned evidence S29 · core/surface/skills/tribunal/SKILL.mdSource-bound opt-in audit, guarded hash-bound stranded-lock recovery, thirteen-card roster, review/verify, durable triage, report and separate follow-up completion. S30 · core/pysrc/_tribunalrunlib.pyImplemented helper API, per-run writer guard, hash-bound stranded-lock archive and audit/follow-up/terminal event vocabulary; cooperative source-bound persistence. S31 · docs/parity.mdCurrent Pi 1.0.2 target, separate host counts, DEGRADED typed approval and Codex farm exception. codeArbiter / route atlas A command selects an owner. The owner loads context, branches through gates, and returns a bounded result.
Selected cmd
/init Assistive workflow: source-bound implementation An agreed outcome becomes a maintained plan, direct or delegated implementation, proportional verification and an honest delivery report. Task context and dependency advice support progress; real user stops and host permissions remain effective. Greenfield and brownfield population use the paired native writer and exact finalization. Exact blob bindings identify inspected source bytes; the linked revision is the committed source base. Implementation, verification, live host qualification and release remain separate claims.
Current source changes The user-facing route table points feature and sprint work at useful planned execution and selected context refresh at its existing owner. Existing authorization covers ordinary work; actual host, destructive-action and commit controls remain. core/surface/includes/routing-table.md The installed context-draft command prepares a paired document preview; context-finalize prepares initial finalization. context-apply commits the exact selected preview and digest. Context modes preserve scaffold and initialized refresh boundaries. core/pysrc/init-codearbiter.py Initial brownfield setup checks the intended repository and actual host permissions, then uses bounded inspection and the installed safe context writer. Ordinary Windows work needs no special scout profile or per-document artifact approval. An existing initialized repository routes to selected refresh instead of another scaffold or population cycle. core/surface/commands/init.md --check: report only Otherwise scaffold, then populate
Related entries Pinned evidence S23 · core/surface/commands/init.mdCurrent init owner: activation/scaffold versus initial population; initialized refresh belongs to status drift, not reinitialization. Selected skill
decompose Assistive workflow: source-bound implementation An agreed outcome becomes a maintained plan, direct or delegated implementation, proportional verification and an honest delivery report. Task context and dependency advice support progress; real user stops and host permissions remain effective. Greenfield and brownfield population use the paired native writer and exact finalization. Exact blob bindings identify inspected source bytes; the linked revision is the committed source base. Implementation, verification, live host qualification and release remain separate claims.
Current source changes Greenfield interviews persist under plans/decomposition and retain actual decisions. Five document/v2-provenance pairs use context-draft and context-apply; context-finalize binds all thirteen outputs. ADRs and task rows use their owners, and retained interview evidence survives finalization. core/surface/skills/decompose/SKILL.md Greenfield: six persisted interview layers
Related entries Pinned evidence S14 · core/surface/skills/decompose/SKILL.mdCurrent greenfield interview and completion contract re-read. Its persisted six-layer draft flow remains separate from the brownfield writer lifecycle. Selected skill
context-creation Assistive workflow: source-bound implementation An agreed outcome becomes a maintained plan, direct or delegated implementation, proportional verification and an honest delivery report. Task context and dependency advice support progress; real user stops and host permissions remain effective. Greenfield and brownfield population use the paired native writer and exact finalization. Exact blob bindings identify inspected source bytes; the linked revision is the committed source base. Implementation, verification, live host qualification and release remain separate claims.
Current source changes The skill index connects planning, direct or delegated implementation, proportional verification and selected context maintenance to their owning workflows. core/surface/skills/INDEX.md Brownfield population covers six logical categories and five content documents through direct bounded inspection or optional scoped agents. The paired writer preserves source bindings and protected content; initial finalization checks all required outputs. core/surface/skills/context-creation/SKILL.md Initial brownfield setup checks the intended repository and actual host permissions, then uses bounded inspection and the installed safe context writer. Ordinary Windows work needs no special scout profile or per-document artifact approval. An existing initialized repository routes to selected refresh instead of another scaffold or population cycle. core/surface/commands/init.md Six full logical assignments require observed read-only containment. Generated plugins alone do not qualify that profile. Each of five content/provenance pairs needs its own native approval, then initialization is a separately approved transaction.
Related entries Pinned evidence S13 · core/surface/skills/context-creation/SKILL.mdCurrent brownfield owner: qualified containment, bounded reports and join, scoped gaps, native document/provenance writes, separate 13-identity finalization and recovery. S26 · core/surface/skills/context-check/SKILL.mdInitialized refresh: read-only report, explicit selection, v1/v2 separation, native paired updates and preserved human history. Full owner read. Selected state
Initialized project knowledge Assistive workflow: source-bound implementation An agreed outcome becomes a maintained plan, direct or delegated implementation, proportional verification and an honest delivery report. Task context and dependency advice support progress; real user stops and host permissions remain effective. Greenfield and brownfield population use the paired native writer and exact finalization. Exact blob bindings identify inspected source bytes; the linked revision is the committed source base. Implementation, verification, live host qualification and release remain separate claims.
Current source changes Initial finalization binds thirteen output identities, rechecks supporting source evidence, and changes the initialization marker with its paired provenance digest. Ordinary finalization requires no artifact approval receipt and does not certify semantic correctness. Already initialized repositories refuse a new finalization. core/artifacts/internal/contextdocument/finalization.go Brownfield population covers six logical categories and five content documents through direct bounded inspection or optional scoped agents. The paired writer preserves source bindings and protected content; initial finalization checks all required outputs. core/surface/skills/context-creation/SKILL.md Greenfield interviews persist under plans/decomposition and retain actual decisions. Five document/v2-provenance pairs use context-draft and context-apply; context-finalize binds all thirteen outputs. ADRs and task rows use their owners, and retained interview evidence survives finalization. core/surface/skills/decompose/SKILL.md CONTEXT • stack • standards • security ADRs • plans • questions • code-map Provenance supports freshness checks
Related entries Pinned evidence S13 · core/surface/skills/context-creation/SKILL.mdCurrent brownfield owner: qualified containment, bounded reports and join, scoped gaps, native document/provenance writes, separate 13-identity finalization and recovery. S14 · core/surface/skills/decompose/SKILL.mdCurrent greenfield interview and completion contract re-read. Its persisted six-layer draft flow remains separate from the brownfield writer lifecycle. S23 · core/surface/commands/init.mdCurrent init owner: activation/scaffold versus initial population; initialized refresh belongs to status drift, not reinitialization. Selected state
Runtime + on-demand context Assistive workflow: source-bound implementation An agreed outcome becomes a maintained plan, direct or delegated implementation, proportional verification and an honest delivery report. Task context and dependency advice support progress; real user stops and host permissions remain effective. Greenfield and brownfield population use the paired native writer and exact finalization. Exact blob bindings identify inspected source bytes; the linked revision is the committed source base. Implementation, verification, live host qualification and release remain separate claims.
Current source changes Routine artifact instructions now focus on the installed writer, typed plan, bounded task reads, factual prerequisites, actual checks and safe retry. The default leaf is about 7.3 KB and includes an executable stdin bootstrap with optional operation-schema discovery; explicit attestation remains available on demand without burdening ordinary task context. A spec's SMARTS design rationale lives in its approach, decisions and one SEC-SMARTS section. core/surface/includes/artifacts.md The user-facing route table points feature and sprint work at useful planned execution and selected context refresh at its existing owner. Existing authorization covers ordinary work; actual host, destructive-action and commit controls remain. core/surface/includes/routing-table.md The commit owner checks authorization, the staged diff, secrets and the verification approach appropriate to the change. Current criterion checks, genuine fix regressions and refactor parity retain their distinct evidence requirements. Reuse applicable observed checks and reviews across phases. Managed v2 maintenance selects affected claims and uses native paired writes; unknown evidence defers that document. core/surface/skills/commit-gate/SKILL.md PR readiness uses the explicit base-to-HEAD subject for current checks, secrets scanning and applicable specialized review. Existing evidence is reused only while its source and inputs match. Merge requires authorization and exact-head hosted CI. core/surface/skills/finishing-a-development-branch/SKILL.md The command catalog starts from the agreed outcome and maintained plan; SMARTS is useful for difficult tradeoffs and does not authorize a changed outcome. core/surface/COMMANDS.md Resident instructions preserve existing work authorization and actual owner decisions without another spec approval or a SMARTS record for every routine step. core/surface/arbiter.md The lazy reference map routes provenance drift or requested initialized refresh to context-check after reading selected documents, provenance and the code map. The owner reuses selected scope and existing authorization. core/surface/includes/reference-map.md Startup state; safety-core + active persona On-demand rules + complete typed task reads Task-scoped packets delivered to each actual actor Context is not approval, admission or consumption proof
Related entries Pinned evidence S27 · core/pysrc/_contextselectlib.pyInert scoped selector and delivery implementation: commands/cwd/evidence, native instruction order, actor/worktree/epoch binding, two-attempt limit and reviewer preflight. S15 · core/pysrc/prompt-submit.pyCurrent persona composition and session/mode/compaction injection comments read; not actor-delivery proof. S16 · core/pysrc/_readinjectlib.pyCurrent file-read pointer and dedup implementation interface read; advisory hints remain separate from task-time selected input. S10 · core/surface/includes/artifacts.mdInstalled admission, typed reads, context completion, authority receipts and format boundaries. Selected skill
brainstorming → writing-plans Assistive workflow: source-bound implementation An agreed outcome becomes a maintained plan, direct or delegated implementation, proportional verification and an honest delivery report. Task context and dependency advice support progress; real user stops and host permissions remain effective. Greenfield and brownfield population use the paired native writer and exact finalization. Exact blob bindings identify inspected source bytes; the linked revision is the committed source base. Implementation, verification, live host qualification and release remain separate claims.
Current source changes Routine artifact instructions now focus on the installed writer, typed plan, bounded task reads, factual prerequisites, actual checks and safe retry. The default leaf is about 7.3 KB and includes an executable stdin bootstrap with optional operation-schema discovery; explicit attestation remains available on demand without burdening ordinary task context. A spec's SMARTS design rationale lives in its approach, decisions and one SEC-SMARTS section. core/surface/includes/artifacts.md Plans preserve agreed criteria, useful tasks, dependency ancestry, bounded shared inputs and observable checks. Stable source identities and typed maintenance keep changes within the selected outcome. A design-drift guard routes an undecided material solution choice back to brainstorming instead of scoring it during planning. core/surface/skills/writing-plans/SKILL.md The feature entry now resumes a maintained spec/plan and task progress without another artifact certification prompt; independent commit and test controls remain. The small lane runs an inline SMARTS scan only for a material solution choice. core/surface/commands/feature.md Source refinement respects recorded deferrals and accepted ADR conflicts without turning each ordinary method adjustment into a separate authorization interview. SMARTS compares approaches only when two or more are materially plausible; a single approach gets one line and no table. core/surface/skills/brainstorming/SKILL.md Full: approved spec, then bound plan
Related entries Pinned evidence S06 · core/surface/commands/feature.mdCurrent feature owner: same-format resume and full/small gates; adds actual per-actor context composition. S10 · core/surface/includes/artifacts.mdInstalled admission, typed reads, context completion, authority receipts and format boundaries. Selected gate
Spec + plan pair approval Assistive workflow: source-bound implementation An agreed outcome becomes a maintained plan, direct or delegated implementation, proportional verification and an honest delivery report. Task context and dependency advice support progress; real user stops and host permissions remain effective. Greenfield and brownfield population use the paired native writer and exact finalization. Exact blob bindings identify inspected source bytes; the linked revision is the committed source base. Implementation, verification, live host qualification and release remain separate claims.
Current source changes Routine artifact instructions now focus on the installed writer, typed plan, bounded task reads, factual prerequisites, actual checks and safe retry. The default leaf is about 7.3 KB and includes an executable stdin bootstrap with optional operation-schema discovery; explicit attestation remains available on demand without burdening ordinary task context. A spec's SMARTS design rationale lives in its approach, decisions and one SEC-SMARTS section. core/surface/includes/artifacts.md The sprint caller performs an index-first recorded-intent read, preserves actual owner decisions and ADR conflicts, then proceeds with authorized work without a second artifact approval system. A SMARTS Unknown names its missing observation and cannot bypass a hard gate or manufacture authority. core/surface/SPRINT.md The sprint entry now routes to the assistive implementation body instead of requiring a fresh user approval and two reviews per task. core/surface/commands/sprint.md Source refinement respects recorded deferrals and accepted ADR conflicts without turning each ordinary method adjustment into a separate authorization interview. SMARTS compares approaches only when two or more are materially plausible; a single approach gets one line and no table. core/surface/skills/brainstorming/SKILL.md One real user reply before execution
Related entries Pinned evidence S07 · core/surface/SPRINT.mdInitial paired approval; full-plan execution; delegated methods; recovery and farm boundaries. S10 · core/surface/includes/artifacts.mdInstalled admission, typed reads, context completion, authority receipts and format boundaries. Selected cmd
/status · /context-check Assistive workflow: source-bound implementation An agreed outcome becomes a maintained plan, direct or delegated implementation, proportional verification and an honest delivery report. Task context and dependency advice support progress; real user stops and host permissions remain effective. Greenfield and brownfield population use the paired native writer and exact finalization. Exact blob bindings identify inspected source bytes; the linked revision is the committed source base. Implementation, verification, live host qualification and release remain separate claims.
Current source changes The user-facing route table points feature and sprint work at useful planned execution and selected context refresh at its existing owner. Existing authorization covers ordinary work; actual host, destructive-action and commit controls remain. core/surface/includes/routing-table.md A drift audit stays read-only; an already authorized selected or full refresh proceeds without asking again per document. The installed draft/preview/apply example preserves source bindings and selected content. Matching hashes establish identity only; initialized refresh never reruns scaffolding, clears the marker or repeats finalization. core/surface/skills/context-check/SKILL.md v1/v2 state and drift report Explicit scoped/full refresh selection Native v2 pairs; preserve history
Related entries Pinned evidence S26 · core/surface/skills/context-check/SKILL.mdInitialized refresh: read-only report, explicit selection, v1/v2 separation, native paired updates and preserved human history. Full owner read. S13 · core/surface/skills/context-creation/SKILL.mdCurrent brownfield owner: qualified containment, bounded reports and join, scoped gaps, native document/provenance writes, separate 13-identity finalization and recovery. Handoffs, branches and returns Feature + sprint / execution paths Follow the split at approval, the join at task execution, and the split again when control returns to the caller.
Selected cmd
/feature or /sprint Assistive workflow: source-bound implementation An agreed outcome becomes a maintained plan, direct or delegated implementation, proportional verification and an honest delivery report. Task context and dependency advice support progress; real user stops and host permissions remain effective. Greenfield and brownfield population use the paired native writer and exact finalization. Exact blob bindings identify inspected source bytes; the linked revision is the committed source base. Implementation, verification, live host qualification and release remain separate claims.
Current source changes Routine artifact instructions now focus on the installed writer, typed plan, bounded task reads, factual prerequisites, actual checks and safe retry. The default leaf is about 7.3 KB and includes an executable stdin bootstrap with optional operation-schema discovery; explicit attestation remains available on demand without burdening ordinary task context. A spec's SMARTS design rationale lives in its approach, decisions and one SEC-SMARTS section. core/surface/includes/artifacts.md The pinned native bridge enables task-context as a bounded informational read and does not require future reviewer/approval producers simply to author a structurally valid draft. Context requests route to ordinary context-preview, context-finalize-preview and exact-reference context-apply; historical receipt input stays a separate compatibility path. Installed host qualification is separate. Native Pi writer admission does not establish a complete live Pi workflow. On NO_LOCAL_STORE the client runs init-store once and retries the request once; a refused initialization keeps NO_LOCAL_STORE. On Windows the binary's architecture comes from the kernel (IsWow64Process2), never WMI or PROCESSOR_* environment. Index and outline pages are now fully checked before rows are yielded, including continuation identity and row shape. core/pysrc/_artifactlib.py The user-facing route table points feature and sprint work at useful planned execution and selected context refresh at its existing owner. Existing authorization covers ordinary work; actual host, destructive-action and commit controls remain. core/surface/includes/routing-table.md The sprint entry now routes to the assistive implementation body instead of requiring a fresh user approval and two reviews per task. core/surface/commands/sprint.md The feature entry now resumes a maintained spec/plan and task progress without another artifact certification prompt; independent commit and test controls remain. The small lane runs an inline SMARTS scan only for a material solution choice. core/surface/commands/feature.md The command catalog starts from the agreed outcome and maintained plan; SMARTS is useful for difficult tradeoffs and does not authorize a changed outcome. core/surface/COMMANDS.md Resolve same-slug artifacts Mixed / duplicate matches: STOP Exact pair or lone spec keeps format
Related entries Pinned evidence S06 · core/surface/commands/feature.mdCurrent feature owner: same-format resume and full/small gates; adds actual per-actor context composition. S07 · core/surface/SPRINT.mdInitial paired approval; full-plan execution; delegated methods; recovery and farm boundaries. S10 · core/surface/includes/artifacts.mdInstalled admission, typed reads, context completion, authority receipts and format boundaries. Selected skill
brainstorming Assistive workflow: source-bound implementation An agreed outcome becomes a maintained plan, direct or delegated implementation, proportional verification and an honest delivery report. Task context and dependency advice support progress; real user stops and host permissions remain effective. Greenfield and brownfield population use the paired native writer and exact finalization. Exact blob bindings identify inspected source bytes; the linked revision is the committed source base. Implementation, verification, live host qualification and release remain separate claims.
Current source changes Routine artifact instructions now focus on the installed writer, typed plan, bounded task reads, factual prerequisites, actual checks and safe retry. The default leaf is about 7.3 KB and includes an executable stdin bootstrap with optional operation-schema discovery; explicit attestation remains available on demand without burdening ordinary task context. A spec's SMARTS design rationale lives in its approach, decisions and one SEC-SMARTS section. core/surface/includes/artifacts.md The skill index connects planning, direct or delegated implementation, proportional verification and selected context maintenance to their owning workflows. core/surface/skills/INDEX.md The feature entry now resumes a maintained spec/plan and task progress without another artifact certification prompt; independent commit and test controls remain. The small lane runs an inline SMARTS scan only for a material solution choice. core/surface/commands/feature.md Source refinement respects recorded deferrals and accepted ADR conflicts without turning each ordinary method adjustment into a separate authorization interview. SMARTS compares approaches only when two or more are materially plausible; a single approach gets one line and no table. core/surface/skills/brainstorming/SKILL.md Confirmed fixes require regression RED/GREEN. Features use current criterion-focused checks and meaningful negative controls where needed; refactors preserve passing characterization baselines. Coverage, lint and applicable security requirements remain explicit. core/surface/skills/tdd/SKILL.md Recorded intent + user criteria Unknowns → open-questions No guessed product decisions
Related entries Pinned evidence S06 · core/surface/commands/feature.mdCurrent feature owner: same-format resume and full/small gates; adds actual per-actor context composition. S07 · core/surface/SPRINT.mdInitial paired approval; full-plan execution; delegated methods; recovery and farm boundaries. S10 · core/surface/includes/artifacts.mdInstalled admission, typed reads, context completion, authority receipts and format boundaries. Selected skill
writing-plans Assistive workflow: source-bound implementation An agreed outcome becomes a maintained plan, direct or delegated implementation, proportional verification and an honest delivery report. Task context and dependency advice support progress; real user stops and host permissions remain effective. Greenfield and brownfield population use the paired native writer and exact finalization. Exact blob bindings identify inspected source bytes; the linked revision is the committed source base. Implementation, verification, live host qualification and release remain separate claims.
Current source changes Routine artifact instructions now focus on the installed writer, typed plan, bounded task reads, factual prerequisites, actual checks and safe retry. The default leaf is about 7.3 KB and includes an executable stdin bootstrap with optional operation-schema discovery; explicit attestation remains available on demand without burdening ordinary task context. A spec's SMARTS design rationale lives in its approach, decisions and one SEC-SMARTS section. core/surface/includes/artifacts.md Plans preserve agreed criteria, useful tasks, dependency ancestry, bounded shared inputs and observable checks. Stable source identities and typed maintenance keep changes within the selected outcome. A design-drift guard routes an undecided material solution choice back to brainstorming instead of scoring it during planning. core/surface/skills/writing-plans/SKILL.md The pinned native bridge enables task-context as a bounded informational read and does not require future reviewer/approval producers simply to author a structurally valid draft. Context requests route to ordinary context-preview, context-finalize-preview and exact-reference context-apply; historical receipt input stays a separate compatibility path. Installed host qualification is separate. Native Pi writer admission does not establish a complete live Pi workflow. On NO_LOCAL_STORE the client runs init-store once and retries the request once; a refused initialization keeps NO_LOCAL_STORE. On Windows the binary's architecture comes from the kernel (IsWow64Process2), never WMI or PROCESSOR_* environment. Index and outline pages are now fully checked before rows are yielded, including continuation identity and row shape. core/pysrc/_artifactlib.py The skill index connects planning, direct or delegated implementation, proportional verification and selected context maintenance to their owning workflows. core/surface/skills/INDEX.md The feature entry now resumes a maintained spec/plan and task progress without another artifact certification prompt; independent commit and test controls remain. The small lane runs an inline SMARTS scan only for a material solution choice. core/surface/commands/feature.md Ordinary approval and sprint instructions bypass legacy capture without changing pending history or inventing authority. Explicit legacy replies retain exact matching. The paired adapter can retire a genuinely unanswered 0.2 sprint-pair request using its serialized cancellation lock and submission check; submitted decisions must replay, while standalone approval 0.1 refuses automatic retirement. core/pysrc/_approvallib.py Ordinary prerequisite instructions bypass legacy capture. The existing adapter can retire only a validated ARMED pending question under its original per-plan OS lock. Confirmed or captured decisions are not disposable; no authority receipt or new native permission API is created. core/pysrc/_prerequisitelib.py Exact criterion IDs → tasks Paths • tests • dependencies Checkpoints • prerequisites
Related entries Pinned evidence S06 · core/surface/commands/feature.mdCurrent feature owner: same-format resume and full/small gates; adds actual per-actor context composition. S07 · core/surface/SPRINT.mdInitial paired approval; full-plan execution; delegated methods; recovery and farm boundaries. S10 · core/surface/includes/artifacts.mdInstalled admission, typed reads, context completion, authority receipts and format boundaries. S21 · core/surface/skills/INDEX.mdCurrent routine-owner index: five debug dispositions, bounded workflow owners and unchanged six-phase TDD / nine-phase commit sequence. Selected gate
Approval is observed Assistive workflow: source-bound implementation An agreed outcome becomes a maintained plan, direct or delegated implementation, proportional verification and an honest delivery report. Task context and dependency advice support progress; real user stops and host permissions remain effective. Greenfield and brownfield population use the paired native writer and exact finalization. Exact blob bindings identify inspected source bytes; the linked revision is the committed source base. Implementation, verification, live host qualification and release remain separate claims.
Current source changes Routine artifact instructions now focus on the installed writer, typed plan, bounded task reads, factual prerequisites, actual checks and safe retry. The default leaf is about 7.3 KB and includes an executable stdin bootstrap with optional operation-schema discovery; explicit attestation remains available on demand without burdening ordinary task context. A spec's SMARTS design rationale lives in its approach, decisions and one SEC-SMARTS section. core/surface/includes/artifacts.md The pinned native bridge enables task-context as a bounded informational read and does not require future reviewer/approval producers simply to author a structurally valid draft. Context requests route to ordinary context-preview, context-finalize-preview and exact-reference context-apply; historical receipt input stays a separate compatibility path. Installed host qualification is separate. Native Pi writer admission does not establish a complete live Pi workflow. On NO_LOCAL_STORE the client runs init-store once and retries the request once; a refused initialization keeps NO_LOCAL_STORE. On Windows the binary's architecture comes from the kernel (IsWow64Process2), never WMI or PROCESSOR_* environment. Index and outline pages are now fully checked before rows are yielded, including continuation identity and row shape. core/pysrc/_artifactlib.py The sprint caller performs an index-first recorded-intent read, preserves actual owner decisions and ADR conflicts, then proceeds with authorized work without a second artifact approval system. A SMARTS Unknown names its missing observation and cannot bypass a hard gate or manufacture authority. core/surface/SPRINT.md Source refinement respects recorded deferrals and accepted ADR conflicts without turning each ordinary method adjustment into a separate authorization interview. SMARTS compares approaches only when two or more are materially plausible; a single approach gets one line and no table. core/surface/skills/brainstorming/SKILL.md Ordinary approval and sprint instructions bypass legacy capture without changing pending history or inventing authority. Explicit legacy replies retain exact matching. The paired adapter can retire a genuinely unanswered 0.2 sprint-pair request using its serialized cancellation lock and submission check; submitted decisions must replay, while standalone approval 0.1 refuses automatic retirement. core/pysrc/_approvallib.py The old paired approval producer publishes a content-addressed observed reply before its durable submission. Cancellation checks the matching observation, content hash, prompt and frozen context under its existing lock, with a strict scan bound. A matching or uncertain pre-submission reply blocks automatic retirement; normal tasks and context reads have no added cost. core/pysrc/_sprintapprovallib.py The optional reference retains strict historical evidence inputs without replacing ordinary work. Explicit draft source refresh preserves progressed plans and history with both CAS bindings; initial approval restrictions remain on historical paired attestation. Its verify command requires literal absolute paths; linked agent subdirectories remain unsupported for strict authority review. These limitations apply to requested attestation, not ordinary task start. core/surface/includes/artifacts-attestation.md Historical sprint approval and SMARTS proof instructions remain available only for existing explicit attestation, outside the ordinary execution path. core/surface/includes/sprint-authority.md Ordinary uppercase words such as Approve DATA migration remain ordinary input. All-letter capture codes require a matching retained request; real armed replies, mismatches and user stops preserve their original handling. core/pysrc/_replylib.py Only actual host-observed approval over the current armed identity confers the corresponding authority. Pi 1.0.2 has the runtime input seam but codeArbiter has no implemented Pi-native producer; typed admission remains blocked. For prerequisites, show the packet digest and the shared 24-hour request/reply deadline. Cancellation does not create approval.
Related entries Pinned evidence S10 · core/surface/includes/artifacts.mdInstalled admission, typed reads, context completion, authority receipts and format boundaries. S31 · docs/parity.mdCurrent Pi 1.0.2 target, separate host counts, DEGRADED typed approval and Codex farm exception. Selected gate
Installed admission is required Assistive workflow: source-bound implementation An agreed outcome becomes a maintained plan, direct or delegated implementation, proportional verification and an honest delivery report. Task context and dependency advice support progress; real user stops and host permissions remain effective. Greenfield and brownfield population use the paired native writer and exact finalization. Exact blob bindings identify inspected source bytes; the linked revision is the committed source base. Implementation, verification, live host qualification and release remain separate claims.
Current source changes Routine artifact instructions now focus on the installed writer, typed plan, bounded task reads, factual prerequisites, actual checks and safe retry. The default leaf is about 7.3 KB and includes an executable stdin bootstrap with optional operation-schema discovery; explicit attestation remains available on demand without burdening ordinary task context. A spec's SMARTS design rationale lives in its approach, decisions and one SEC-SMARTS section. core/surface/includes/artifacts.md The pinned native bridge enables task-context as a bounded informational read and does not require future reviewer/approval producers simply to author a structurally valid draft. Context requests route to ordinary context-preview, context-finalize-preview and exact-reference context-apply; historical receipt input stays a separate compatibility path. Installed host qualification is separate. Native Pi writer admission does not establish a complete live Pi workflow. On NO_LOCAL_STORE the client runs init-store once and retries the request once; a refused initialization keeps NO_LOCAL_STORE. On Windows the binary's architecture comes from the kernel (IsWow64Process2), never WMI or PROCESSOR_* environment. Index and outline pages are now fully checked before rows are yielded, including continuation identity and row shape. core/pysrc/_artifactlib.py Pi HTML admission unavailable; legacy Markdown and inline small lane stay distinct. HTML farm remains blocked.
Related entries Pinned evidence S10 · core/surface/includes/artifacts.mdInstalled admission, typed reads, context completion, authority receipts and format boundaries. Selected skill
subagent-driven-development Assistive workflow: source-bound implementation An agreed outcome becomes a maintained plan, direct or delegated implementation, proportional verification and an honest delivery report. Task context and dependency advice support progress; real user stops and host permissions remain effective. Greenfield and brownfield population use the paired native writer and exact finalization. Exact blob bindings identify inspected source bytes; the linked revision is the committed source base. Implementation, verification, live host qualification and release remain separate claims.
Current source changes Routine artifact instructions now focus on the installed writer, typed plan, bounded task reads, factual prerequisites, actual checks and safe retry. The default leaf is about 7.3 KB and includes an executable stdin bootstrap with optional operation-schema discovery; explicit attestation remains available on demand without burdening ordinary task context. A spec's SMARTS design rationale lives in its approach, decisions and one SEC-SMARTS section. core/surface/includes/artifacts.md The user-facing route table points feature and sprint work at useful planned execution and selected context refresh at its existing owner. Existing authorization covers ordinary work; actual host, destructive-action and commit controls remain. core/surface/includes/routing-table.md The skill index connects planning, direct or delegated implementation, proportional verification and selected context maintenance to their owning workflows. core/surface/skills/INDEX.md The sprint caller performs an index-first recorded-intent read, preserves actual owner decisions and ADR conflicts, then proceeds with authorized work without a second artifact approval system. A SMARTS Unknown names its missing observation and cannot bypass a hard gate or manufacture authority. core/surface/SPRINT.md The sprint entry now routes to the assistive implementation body instead of requiring a fresh user approval and two reviews per task. core/surface/commands/sprint.md The feature entry now resumes a maintained spec/plan and task progress without another artifact certification prompt; independent commit and test controls remain. The small lane runs an inline SMARTS scan only for a material solution choice. core/surface/commands/feature.md Attended feature execution now treats checkpoints as progress groups, not compulsory approval stops; it forwards real failures and continues independent work. core/surface/skills/executing-plans/SKILL.md Implementation works directly or delegates a coherent unit when useful. Bounded task context, applicable dependency advice and observed checks support progress while preserving actual user stops and host permissions. Verification checks the change against a recorded SMARTS rationale for material invalidation only. core/surface/skills/subagent-driven-development/SKILL.md Read-only task-context reports direct dependencies and unresolved ancestors, scoped_changed, scoped_unchanged or unassessed inputs, and actual check observations. Missing broader context remains explicit; the packet does not certify verification. core/artifacts/internal/operations/task_context.go Exact and batch symbol retrieval exposes reported check history without an approval ticket. Outline paging now accounts for encoded rows incrementally while retaining response budgets and continuation behavior. core/artifacts/internal/read/read.go Ordinary prerequisite instructions bypass legacy capture. The existing adapter can retire only a validated ARMED pending question under its original per-plan OS lock. Confirmed or captured decisions are not disposable; no authority receipt or new native permission API is created. core/pysrc/_prerequisitelib.py Resident instructions preserve existing work authorization and actual owner decisions without another spec approval or a SMARTS record for every routine step. core/surface/arbiter.md Codex dispatch requires attestation only for explicitly claimed host proof; an unavailable requested farm backend is reported without silently switching provider or model. Ordinary context inspection can proceed inline within actual host permissions; an agent label supplies no containment or semantic verification. core/surface/includes/codex-host-notes.md The public Codex farm boundary matches its caller: report the unavailable backend and use another provider or model only within existing authorization. docs/parity.md Authors and reviewers share task-critical reference and constraint checks. Missing or stale provenance remains advisory. Two delivery attempts end with an advisory unknown observation, without another retry or an invented receipt. core/pysrc/_contextselectlib.py Feature actors receive scoped requirements, named critical references and observed evidence. Applicable conflicts or missing required inputs stop dependent work; provenance warnings remain visible and advisory. core/pysrc/_artifactpromptlib.py Scheduling follows the full dependency ancestry and exposes unresolved ancestor IDs and reasons. Reconciled progress remains advisory; blocked ancestors prevent premature downstream readiness while independent tasks remain available. Response-budget omissions are explicit, retain the global scheduling decision and direct callers to supported detail reads. core/artifacts/internal/operations/assistive.go Resolve current identities → eligible tasks only HTML: complete contextual read ticket → task-start Then separately compose and deliver each actor’s scoped input Legacy Markdown keeps its own dependency/state rules
Related entries Pinned evidence S08 · core/surface/skills/subagent-driven-development/SKILL.mdCurrent implementation-dispatch and independent spec-review sections re-read; remaining scope contract also checked against current skill index. Fresh actor input and bounded delivery are distinct from engine tickets. S27 · core/pysrc/_contextselectlib.pyInert scoped selector and delivery implementation: commands/cwd/evidence, native instruction order, actor/worktree/epoch binding, two-attempt limit and reviewer preflight. S10 · core/surface/includes/artifacts.mdInstalled admission, typed reads, context completion, authority receipts and format boundaries. Selected agent
Spec-compliance review Assistive workflow: source-bound implementation An agreed outcome becomes a maintained plan, direct or delegated implementation, proportional verification and an honest delivery report. Task context and dependency advice support progress; real user stops and host permissions remain effective. Greenfield and brownfield population use the paired native writer and exact finalization. Exact blob bindings identify inspected source bytes; the linked revision is the committed source base. Implementation, verification, live host qualification and release remain separate claims.
Current source changes A new explicit batch review can select a passing rerun of still-reviewed work without replacing the original task review. The selected run must match its own recorded workspace; changed source or supporting evidence still requires reassessment. core/artifacts/internal/evidence/completion.go Implementation works directly or delegates a coherent unit when useful. Bounded task context, applicable dependency advice and observed checks support progress while preserving actual user stops and host permissions. Verification checks the change against a recorded SMARTS rationale for material invalidation only. core/surface/skills/subagent-driven-development/SKILL.md Read-only task-context reports direct dependencies and unresolved ancestors, scoped_changed, scoped_unchanged or unassessed inputs, and actual check observations. Missing broader context remains explicit; the packet does not certify verification. core/artifacts/internal/operations/task_context.go Independent task-spec review and actual source-bound observations still precede acceptance. The current source repairs Claude launch/child correlation and native Unicode observation hashing; neither repair substitutes an author’s report for a reviewer receipt. Re-arm requests made stale by the applicable engine upgrade rather than relabeling old evidence.
Related entries Pinned evidence S08 · core/surface/skills/subagent-driven-development/SKILL.mdCurrent implementation-dispatch and independent spec-review sections re-read; remaining scope contract also checked against current skill index. Fresh actor input and bounded delivery are distinct from engine tickets. S32 · core/pysrc/_artifactauthoritylib.pyNative artifact ID grammar, producer states, Claude join lock, UTF-8/UTF-16-key-order observation binding. S40 · CHANGELOG.mdMerge-scope record for review-join, marker, ID/reorder, Unicode, prerequisite and engine-wait repairs. Selected gate
Fresh task verification Assistive workflow: source-bound implementation An agreed outcome becomes a maintained plan, direct or delegated implementation, proportional verification and an honest delivery report. Task context and dependency advice support progress; real user stops and host permissions remain effective. Greenfield and brownfield population use the paired native writer and exact finalization. Exact blob bindings identify inspected source bytes; the linked revision is the committed source base. Implementation, verification, live host qualification and release remain separate claims.
Current source changes Routine artifact instructions now focus on the installed writer, typed plan, bounded task reads, factual prerequisites, actual checks and safe retry. The default leaf is about 7.3 KB and includes an executable stdin bootstrap with optional operation-schema discovery; explicit attestation remains available on demand without burdening ordinary task context. A spec's SMARTS design rationale lives in its approach, decisions and one SEC-SMARTS section. core/surface/includes/artifacts.md A new explicit batch review can select a passing rerun of still-reviewed work without replacing the original task review. The selected run must match its own recorded workspace; changed source or supporting evidence still requires reassessment. core/artifacts/internal/evidence/completion.go The skill index connects planning, direct or delegated implementation, proportional verification and selected context maintenance to their owning workflows. core/surface/skills/INDEX.md Implementation works directly or delegates a coherent unit when useful. Bounded task context, applicable dependency advice and observed checks support progress while preserving actual user stops and host permissions. Verification checks the change against a recorded SMARTS rationale for material invalidation only. core/surface/skills/subagent-driven-development/SKILL.md Confirmed fixes require regression RED/GREEN. Features use current criterion-focused checks and meaningful negative controls where needed; refactors preserve passing characterization baselines. Coverage, lint and applicable security requirements remain explicit. core/surface/skills/tdd/SKILL.md Read-only task-context reports direct dependencies and unresolved ancestors, scoped_changed, scoped_unchanged or unassessed inputs, and actual check observations. Missing broader context remains explicit; the packet does not certify verification. core/artifacts/internal/operations/task_context.go Task-check observations now retain directly referenced criterion identity and an optional bounded source-input fingerprint or an explicit unassessed reason, without executing tests or minting approval receipts. core/artifacts/internal/operations/task_check.go New checks bind the selected criterion and applicable constraint bodies, including explicit references and scope applicability. Unrelated constraints remain outside the digest; old observations never gain unassessed constraint coverage. core/artifacts/internal/model/assistive.go The closed task-check request accepts bounded command observations matched to plan definitions. Ordinary context-preview and context-finalize-preview omit prompt identity; context-apply accepts an exact preview reference and digest. Invalid requests, changed preimages and stale CAS writes remain rejected. The historical smarts-apply request accepts the 0.2.0 SMARTS decision record. init-store takes a protocol-only request. core/artifacts/internal/operations/protocol.go The unpublished 0.3.2 native schema adds optional shared_inputs and scoped task-check fields without silently relabeling old records. core/artifacts/schemas/common.schema.json The bounded source-fingerprint reader hashes actual declared task/shared paths, directory membership and known build-control files; unsafe, missing and over-limit scopes are explicitly unassessed rather than blocking work or proving a full dependency graph. core/artifacts/internal/operations/input_scope.go Delegated authors preserve the caller-selected verification approach: genuine regression RED/GREEN for fixes, current criterion-focused feature checks and passing characterization baselines for refactors. Actual results and pending checks return to the caller; staging requires assigned scope and authorization. core/surface/includes/author-tdd-workflow.md Characterization tests can pass against existing correct behavior. Preserve behavioral assertions against the post-backfill baseline; mechanical import or symbol-reference updates retain the same oracle. Coverage and parity checks remain required. A precise surface table proceeds under existing user authorization; only material expansion or a changed outcome needs another decision. Distinct restructuring strategies are compared with SMARTS only after surface and parity are settled. core/surface/skills/refactor/SKILL.md Only an actual verifier wrapper executable/script prefix selects authority handling. Inert PowerShell data mentioning the filename is not a verifier call; malformed real verifier invocations still fail closed. Fresh legacy context approval arming is refused; historical pending routes remain readable. core/pysrc/_artifactauthoritylib.py Run the plan’s proving command again Read output and exit code HTML: real producer-owned receipts
Related entries Pinned evidence S08 · core/surface/skills/subagent-driven-development/SKILL.mdCurrent implementation-dispatch and independent spec-review sections re-read; remaining scope contract also checked against current skill index. Fresh actor input and bounded delivery are distinct from engine tickets. S10 · core/surface/includes/artifacts.mdInstalled admission, typed reads, context completion, authority receipts and format boundaries. Selected gate
Correction is not permission to waive a gate Assistive workflow: source-bound implementation An agreed outcome becomes a maintained plan, direct or delegated implementation, proportional verification and an honest delivery report. Task context and dependency advice support progress; real user stops and host permissions remain effective. Greenfield and brownfield population use the paired native writer and exact finalization. Exact blob bindings identify inspected source bytes; the linked revision is the committed source base. Implementation, verification, live host qualification and release remain separate claims.
Current source changes Routine artifact instructions now focus on the installed writer, typed plan, bounded task reads, factual prerequisites, actual checks and safe retry. The default leaf is about 7.3 KB and includes an executable stdin bootstrap with optional operation-schema discovery; explicit attestation remains available on demand without burdening ordinary task context. A spec's SMARTS design rationale lives in its approach, decisions and one SEC-SMARTS section. core/surface/includes/artifacts.md The sprint caller performs an index-first recorded-intent read, preserves actual owner decisions and ADR conflicts, then proceeds with authorized work without a second artifact approval system. A SMARTS Unknown names its missing observation and cannot bypass a hard gate or manufacture authority. core/surface/SPRINT.md Attended feature execution now treats checkpoints as progress groups, not compulsory approval stops; it forwards real failures and continues independent work. core/surface/skills/executing-plans/SKILL.md Implementation works directly or delegates a coherent unit when useful. Bounded task context, applicable dependency advice and observed checks support progress while preserving actual user stops and host permissions. Verification checks the change against a recorded SMARTS rationale for material invalidation only. core/surface/skills/subagent-driven-development/SKILL.md An unanswered legacy reconciliation can be superseded only when route and pending state agree and no durable attempt or matching observed reply exists. Captured replies recover through their original immutable source and receipt; in-flight mutations retain exact replay and real user stops. core/pysrc/_reconciliationlib.py The installed recover command finishes a captured original reply or replays an in-flight mutation without asking for a new user answer. core/pysrc/artifact-reconcile.py The closed task-check request accepts bounded command observations matched to plan definitions. Ordinary context-preview and context-finalize-preview omit prompt identity; context-apply accepts an exact preview reference and digest. Invalid requests, changed preimages and stale CAS writes remain rejected. The historical smarts-apply request accepts the 0.2.0 SMARTS decision record. init-store takes a protocol-only request. core/artifacts/internal/operations/protocol.go The candidate-only runbook distinguishes 0.3.1/0.3.2 mixed-schema readers, exact replay and forward recovery while preserving recorded user stops. The paired prompt retirement scans immutable observed replies before calling a request unanswered; submitted or uncertain decisions are never erased. Older-client downgrade after modern edits remains unsupported. docs/artifacts/ASSISTIVE-UPGRADE.md Ordinary approval and sprint instructions bypass legacy capture without changing pending history or inventing authority. Explicit legacy replies retain exact matching. The paired adapter can retire a genuinely unanswered 0.2 sprint-pair request using its serialized cancellation lock and submission check; submitted decisions must replay, while standalone approval 0.1 refuses automatic retirement. core/pysrc/_approvallib.py Ordinary prerequisite instructions bypass legacy capture. The existing adapter can retire only a validated ARMED pending question under its original per-plan OS lock. Confirmed or captured decisions are not disposable; no authority receipt or new native permission API is created. core/pysrc/_prerequisitelib.py The old paired approval producer publishes a content-addressed observed reply before its durable submission. Cancellation checks the matching observation, content hash, prompt and frozen context under its existing lock, with a strict scan bound. A matching or uncertain pre-submission reply blocks automatic retirement; normal tasks and context reads have no added cost. core/pysrc/_sprintapprovallib.py The optional reference retains strict historical evidence inputs without replacing ordinary work. Explicit draft source refresh preserves progressed plans and history with both CAS bindings; initial approval restrictions remain on historical paired attestation. Its verify command requires literal absolute paths; linked agent subdirectories remain unsupported for strict authority review. These limitations apply to requested attestation, not ordinary task start. core/surface/includes/artifacts-attestation.md Scheduling follows the full dependency ancestry and exposes unresolved ancestor IDs and reasons. Reconciled progress remains advisory; blocked ancestors prevent premature downstream readiness while independent tasks remain available. Response-budget omissions are explicit, retain the global scheduling decision and direct callers to supported detail reads. core/artifacts/internal/operations/assistive.go Explicit draft source refresh retains both source CAS components and same-pair identity without requiring an unstarted plan. Existing progress, approvals and evidence are preserved without asserting current authority. Dependency invalidation preserves an independently BLOCKED task and its reason. Verification summaries disclose conflicting caller reports separately from receipt freshness; later successful verification covers only reports present when its context was issued. core/artifacts/internal/operations/lifecycle.go Ordinary uppercase words such as Approve DATA migration remain ordinary input. All-letter capture codes require a matching retained request; real armed replies, mismatches and user stops preserve their original handling. core/pysrc/_replylib.py Only an actual verifier wrapper executable/script prefix selects authority handling. Inert PowerShell data mentioning the filename is not a verifier call; malformed real verifier invocations still fail closed. Fresh legacy context approval arming is refused; historical pending routes remain readable. core/pysrc/_artifactauthoritylib.py Bounded isolated Python stdin diagnostics may read literal protected paths or git show HEAD objects. Unknown code, mutable calls, shell expansion and extra commands retain the guard checks; ordinary commands avoid AST parsing. core/pysrc/_bashguardlib.py After interruption, resume through the existing exact identity and durable operation outcome. Correct stale context or missing authority at its own boundary. The newer engine wait defaults to 120 seconds, but this is not a latency objective or an extension of outer host deadlines. This atlas does not benchmark runtime responsiveness.
Related entries Pinned evidence S08 · core/surface/skills/subagent-driven-development/SKILL.mdCurrent implementation-dispatch and independent spec-review sections re-read; remaining scope contract also checked against current skill index. Fresh actor input and bounded delivery are distinct from engine tickets. S10 · core/surface/includes/artifacts.mdInstalled admission, typed reads, context completion, authority receipts and format boundaries. S40 · CHANGELOG.mdMerge-scope record for review-join, marker, ID/reorder, Unicode, prerequisite and engine-wait repairs. Selected agent
Combined quality review Assistive workflow: source-bound implementation An agreed outcome becomes a maintained plan, direct or delegated implementation, proportional verification and an honest delivery report. Task context and dependency advice support progress; real user stops and host permissions remain effective. Greenfield and brownfield population use the paired native writer and exact finalization. Exact blob bindings identify inspected source bytes; the linked revision is the committed source base. Implementation, verification, live host qualification and release remain separate claims.
Current source changes A new explicit batch review can select a passing rerun of still-reviewed work without replacing the original task review. The selected run must match its own recorded workspace; changed source or supporting evidence still requires reassessment. core/artifacts/internal/evidence/completion.go Implementation works directly or delegates a coherent unit when useful. Bounded task context, applicable dependency advice and observed checks support progress while preserving actual user stops and host permissions. Verification checks the change against a recorded SMARTS rationale for material invalidation only. core/surface/skills/subagent-driven-development/SKILL.md Task-check observations now retain directly referenced criterion identity and an optional bounded source-input fingerprint or an explicit unassessed reason, without executing tests or minting approval receipts. core/artifacts/internal/operations/task_check.go Human HTML presents reported command checks in an escaped, explicitly unverified section; implementation state and source validation remain distinct. core/artifacts/internal/render/render.go The bounded source-fingerprint reader hashes actual declared task/shared paths, directory membership and known build-control files; unsafe, missing and over-limit scopes are explicitly unassessed rather than blocking work or proving a full dependency graph. core/artifacts/internal/operations/input_scope.go The optional reference retains strict historical evidence inputs without replacing ordinary work. Explicit draft source refresh preserves progressed plans and history with both CAS bindings; initial approval restrictions remain on historical paired attestation. Its verify command requires literal absolute paths; linked agent subdirectories remain unsupported for strict authority review. These limitations apply to requested attestation, not ordinary task start. core/surface/includes/artifacts-attestation.md Once per authored scope / checkpoint Actual combined diff + fresh reviewer inputs No CRITICAL/HIGH; rerun after corrections Scope of one uses the same rule
Related entries Pinned evidence S08 · core/surface/skills/subagent-driven-development/SKILL.mdCurrent implementation-dispatch and independent spec-review sections re-read; remaining scope contract also checked against current skill index. Fresh actor input and bounded delivery are distinct from engine tickets. S27 · core/pysrc/_contextselectlib.pyInert scoped selector and delivery implementation: commands/cwd/evidence, native instruction order, actor/worktree/epoch binding, two-attempt limit and reviewer preflight. S18 · core/surface/includes/review-matrix.mdShared diff-path reviewer selection for review/checkpoint/preview; callers retain their own gates. Selected state
accept-scope / current proof Assistive workflow: source-bound implementation An agreed outcome becomes a maintained plan, direct or delegated implementation, proportional verification and an honest delivery report. Task context and dependency advice support progress; real user stops and host permissions remain effective. Greenfield and brownfield population use the paired native writer and exact finalization. Exact blob bindings identify inspected source bytes; the linked revision is the committed source base. Implementation, verification, live host qualification and release remain separate claims.
Current source changes Routine artifact instructions now focus on the installed writer, typed plan, bounded task reads, factual prerequisites, actual checks and safe retry. The default leaf is about 7.3 KB and includes an executable stdin bootstrap with optional operation-schema discovery; explicit attestation remains available on demand without burdening ordinary task context. A spec's SMARTS design rationale lives in its approach, decisions and one SEC-SMARTS section. core/surface/includes/artifacts.md A new explicit batch review can select a passing rerun of still-reviewed work without replacing the original task review. The selected run must match its own recorded workspace; changed source or supporting evidence still requires reassessment. core/artifacts/internal/evidence/completion.go The sprint caller performs an index-first recorded-intent read, preserves actual owner decisions and ADR conflicts, then proceeds with authorized work without a second artifact approval system. A SMARTS Unknown names its missing observation and cannot bypass a hard gate or manufacture authority. core/surface/SPRINT.md Attended feature execution now treats checkpoints as progress groups, not compulsory approval stops; it forwards real failures and continues independent work. core/surface/skills/executing-plans/SKILL.md Implementation works directly or delegates a coherent unit when useful. Bounded task context, applicable dependency advice and observed checks support progress while preserving actual user stops and host permissions. Verification checks the change against a recorded SMARTS rationale for material invalidation only. core/surface/skills/subagent-driven-development/SKILL.md The commit owner checks authorization, the staged diff, secrets and the verification approach appropriate to the change. Current criterion checks, genuine fix regressions and refactor parity retain their distinct evidence requirements. Reuse applicable observed checks and reviews across phases. Managed v2 maintenance selects affected claims and uses native paired writes; unknown evidence defers that document. core/surface/skills/commit-gate/SKILL.md PR readiness uses the explicit base-to-HEAD subject for current checks, secrets scanning and applicable specialized review. Existing evidence is reused only while its source and inputs match. Merge requires authorization and exact-head hosted CI. core/surface/skills/finishing-a-development-branch/SKILL.md An unanswered legacy reconciliation can be superseded only when route and pending state agree and no durable attempt or matching observed reply exists. Captured replies recover through their original immutable source and receipt; in-flight mutations retain exact replay and real user stops. core/pysrc/_reconciliationlib.py The bounded source-fingerprint reader hashes actual declared task/shared paths, directory membership and known build-control files; unsafe, missing and over-limit scopes are explicitly unassessed rather than blocking work or proving a full dependency graph. core/artifacts/internal/operations/input_scope.go The candidate-only runbook distinguishes 0.3.1/0.3.2 mixed-schema readers, exact replay and forward recovery while preserving recorded user stops. The paired prompt retirement scans immutable observed replies before calling a request unanswered; submitted or uncertain decisions are never erased. Older-client downgrade after modern edits remains unsupported. docs/artifacts/ASSISTIVE-UPGRADE.md The optional reference retains strict historical evidence inputs without replacing ordinary work. Explicit draft source refresh preserves progressed plans and history with both CAS bindings; initial approval restrictions remain on historical paired attestation. Its verify command requires literal absolute paths; linked agent subdirectories remain unsupported for strict authority review. These limitations apply to requested attestation, not ordinary task start. core/surface/includes/artifacts-attestation.md Status separates reported DONE from independently ACCEPTED work, shows actual blocked decisions, and does not infer an approval request from draft governance state. core/surface/commands/status.md HTML: one whole-scope operation All required current receipts must agree Markdown: authoritative ledger transition No rendered labels as authority
Related entries Pinned evidence S08 · core/surface/skills/subagent-driven-development/SKILL.mdCurrent implementation-dispatch and independent spec-review sections re-read; remaining scope contract also checked against current skill index. Fresh actor input and bounded delivery are distinct from engine tickets. S10 · core/surface/includes/artifacts.mdInstalled admission, typed reads, context completion, authority receipts and format boundaries. Handoffs, branches and returns Context / how the layers are assembled Durable knowledge → selective task input → actual actor delivery → independent proof. These are different boundaries.
Selected skill
Greenfield / decompose Assistive workflow: source-bound implementation An agreed outcome becomes a maintained plan, direct or delegated implementation, proportional verification and an honest delivery report. Task context and dependency advice support progress; real user stops and host permissions remain effective. Greenfield and brownfield population use the paired native writer and exact finalization. Exact blob bindings identify inspected source bytes; the linked revision is the committed source base. Implementation, verification, live host qualification and release remain separate claims.
Current source changes Greenfield interviews persist under plans/decomposition and retain actual decisions. Five document/v2-provenance pairs use context-draft and context-apply; context-finalize binds all thirteen outputs. ADRs and task rows use their owners, and retained interview evidence survives finalization. core/surface/skills/decompose/SKILL.md Vision → users → scope → technical shape Integrations → risks, each persisted Architecture / phased plan / backlog User-reviewed decisions and context
Related entries Pinned evidence S14 · core/surface/skills/decompose/SKILL.mdCurrent greenfield interview and completion contract re-read. Its persisted six-layer draft flow remains separate from the brownfield writer lifecycle. Selected skill
Brownfield / context-creation Assistive workflow: source-bound implementation An agreed outcome becomes a maintained plan, direct or delegated implementation, proportional verification and an honest delivery report. Task context and dependency advice support progress; real user stops and host permissions remain effective. Greenfield and brownfield population use the paired native writer and exact finalization. Exact blob bindings identify inspected source bytes; the linked revision is the committed source base. Implementation, verification, live host qualification and release remain separate claims.
Current source changes The skill index connects planning, direct or delegated implementation, proportional verification and selected context maintenance to their owning workflows. core/surface/skills/INDEX.md Codex dispatch requires attestation only for explicitly claimed host proof; an unavailable requested farm backend is reported without silently switching provider or model. Ordinary context inspection can proceed inline within actual host permissions; an agent label supplies no containment or semantic verification. core/surface/includes/codex-host-notes.md Brownfield population covers six logical categories and five content documents through direct bounded inspection or optional scoped agents. The paired writer preserves source bindings and protected content; initial finalization checks all required outputs. core/surface/skills/context-creation/SKILL.md Optional context scouts report only their assigned scope, source identities, coverage, conflicts and unknowns. Six logical categories do not require six child agents. A role label or read-only instruction proves neither sandbox containment nor independent semantic verification; actual host permissions and excluded paths still apply. core/surface/agents/scout.md Qualified read-only scout profile first Six full / affected scoped assignments Bounded report-only synthesis Native content + provenance transactions
Related entries Pinned evidence S13 · core/surface/skills/context-creation/SKILL.mdCurrent brownfield owner: qualified containment, bounded reports and join, scoped gaps, native document/provenance writes, separate 13-identity finalization and recovery. S26 · core/surface/skills/context-check/SKILL.mdInitialized refresh: read-only report, explicit selection, v1/v2 separation, native paired updates and preserved human history. Full owner read. Selected state
.codearbiter/ project knowledge Assistive workflow: source-bound implementation An agreed outcome becomes a maintained plan, direct or delegated implementation, proportional verification and an honest delivery report. Task context and dependency advice support progress; real user stops and host permissions remain effective. Greenfield and brownfield population use the paired native writer and exact finalization. Exact blob bindings identify inspected source bytes; the linked revision is the committed source base. Implementation, verification, live host qualification and release remain separate claims.
Current source changes The finite context contract permits a whole-document Markdown entry only as the sole entry for one of five named content targets. Explicit typed field entries remain available; the writer does not expose arbitrary repository-file writes. core/artifacts/internal/contextdocument/contract.go Whole-document replacement requires the explicit document selection and exact existing bytes. It preserves CONTEXT frontmatter and initialization state, checks the named document identity and Markdown structure, and retains the code-map size limit. Human content outside the authorized change remains the caller's preservation obligation. core/artifacts/internal/contextdocument/document_selection.go Initial finalization binds thirteen output identities, rechecks supporting source evidence, and changes the initialization marker with its paired provenance digest. Ordinary finalization requires no artifact approval receipt and does not certify semantic correctness. Already initialized repositories refuse a new finalization. core/artifacts/internal/contextdocument/finalization.go The context schema describes both finite field entries and the explicit whole-document entry for the five supported content documents. Its typed structure records evidence references; schema validity does not establish that the claims were independently reviewed. core/artifacts/schemas/context.schema.json Greenfield interviews persist under plans/decomposition and retain actual decisions. Five document/v2-provenance pairs use context-draft and context-apply; context-finalize binds all thirteen outputs. ADRs and task rows use their owners, and retained interview evidence survives finalization. core/surface/skills/decompose/SKILL.md Routine maintenance selects changed v2 claims intersecting staged sources using a current bounded snapshot. Unknown evidence defers the affected document; native paired writes preserve owned fields and current identities. core/pysrc/_provenancelib.py CONTEXT: purpose, vocabulary, stage and limits Stack: scoped test/build commands and author mapping Standards + security: rules with preserved human ownership Brownfield: five content docs + their v2 provenance pairs ADRs, boards and greenfield plans retain separate owners
Related entries Pinned evidence S13 · core/surface/skills/context-creation/SKILL.mdCurrent brownfield owner: qualified containment, bounded reports and join, scoped gaps, native document/provenance writes, separate 13-identity finalization and recovery. S14 · core/surface/skills/decompose/SKILL.mdCurrent greenfield interview and completion contract re-read. Its persisted six-layer draft flow remains separate from the brownfield writer lifecycle. S26 · core/surface/skills/context-check/SKILL.mdInitialized refresh: read-only report, explicit selection, v1/v2 separation, native paired updates and preserved human history. Full owner read. Selected state
L1 / startup state Enabled repo: stage, questions, tasks, overrides and hygiene briefing SessionStart carries live state, not the composed persona
Related entries Pinned evidence S17 · docs/hooks.mdDocumented host event differences and startup live-state injection; not blanket enforcement proof. Selected state
L3 / owner + scope rules Assistive workflow: source-bound implementation An agreed outcome becomes a maintained plan, direct or delegated implementation, proportional verification and an honest delivery report. Task context and dependency advice support progress; real user stops and host permissions remain effective. Greenfield and brownfield population use the paired native writer and exact finalization. Exact blob bindings identify inspected source bytes; the linked revision is the committed source base. Implementation, verification, live host qualification and release remain separate claims.
Current source changes The lazy reference map routes provenance drift or requested initialized refresh to context-check after reading selected documents, provenance and the code map. The owner reuses selected scope and existing authorization. core/surface/includes/reference-map.md Read the invoked command/skill body Scope touch → reference-map docs Read relevant ADR/deferral bodies Indexes first; no bulk directory reads
Related entries Pinned evidence S03 · core/surface/arbiter.mdIntent-to-owner routing, on-demand loading, caller extent and mode boundary. S04 · core/surface/includes/routing-table.mdAuthoritative intent-to-owner table, scope-touch routes and stop/return semantics. S05 · core/surface/includes/reference-map.mdScope-to-governing-document and specialist mapping. Selected state
L3a / read-time hints Assistive workflow: source-bound implementation An agreed outcome becomes a maintained plan, direct or delegated implementation, proportional verification and an honest delivery report. Task context and dependency advice support progress; real user stops and host permissions remain effective. Greenfield and brownfield population use the paired native writer and exact finalization. Exact blob bindings identify inspected source bytes; the linked revision is the committed source base. Implementation, verification, live host qualification and release remain separate claims.
Current source changes Bounded read-time pointers include ordinary draft and ready HTML specs as advisory context. Rejected or withdrawn specs stay excluded; no pointer confers approval. core/pysrc/_readinjectlib.py A bounded pointer, not a full document security > ADRs > specs > standards Freshness and dedup are advisory Host event surfaces differ
Related entries Pinned evidence S16 · core/pysrc/_readinjectlib.pyCurrent file-read pointer and dedup implementation interface read; advisory hints remain separate from task-time selected input. S17 · docs/hooks.mdDocumented host event differences and startup live-state injection; not blanket enforcement proof. S24 · core/pysrc/_readinjectlib.pyCurrent approved_spec_index implementation still filters *.md; do not infer typed-HTML read-hint parity. Selected state
L4 / approved specification Assistive workflow: source-bound implementation An agreed outcome becomes a maintained plan, direct or delegated implementation, proportional verification and an honest delivery report. Task context and dependency advice support progress; real user stops and host permissions remain effective. Greenfield and brownfield population use the paired native writer and exact finalization. Exact blob bindings identify inspected source bytes; the linked revision is the committed source base. Implementation, verification, live host qualification and release remain separate claims.
Current source changes Routine artifact instructions now focus on the installed writer, typed plan, bounded task reads, factual prerequisites, actual checks and safe retry. The default leaf is about 7.3 KB and includes an executable stdin bootstrap with optional operation-schema discovery; explicit attestation remains available on demand without burdening ordinary task context. A spec's SMARTS design rationale lives in its approach, decisions and one SEC-SMARTS section. core/surface/includes/artifacts.md The pinned native bridge enables task-context as a bounded informational read and does not require future reviewer/approval producers simply to author a structurally valid draft. Context requests route to ordinary context-preview, context-finalize-preview and exact-reference context-apply; historical receipt input stays a separate compatibility path. Installed host qualification is separate. Native Pi writer admission does not establish a complete live Pi workflow. On NO_LOCAL_STORE the client runs init-store once and retries the request once; a refused initialization keeps NO_LOCAL_STORE. On Windows the binary's architecture comes from the kernel (IsWow64Process2), never WMI or PROCESSOR_* environment. Index and outline pages are now fully checked before rows are yielded, including continuation identity and row shape. core/pysrc/_artifactlib.py Bounded read-time pointers include ordinary draft and ready HTML specs as advisory context. Rejected or withdrawn specs stay excluded; no pointer confers approval. core/pysrc/_readinjectlib.py Stable artifact and criterion IDs Behavior • boundaries • non-goals Contracts and accepted intent Normative change invalidates approval
Related entries Pinned evidence S10 · core/surface/includes/artifacts.mdInstalled admission, typed reads, context completion, authority receipts and format boundaries. S06 · core/surface/commands/feature.mdCurrent feature owner: same-format resume and full/small gates; adds actual per-actor context composition. S07 · core/surface/SPRINT.mdInitial paired approval; full-plan execution; delegated methods; recovery and farm boundaries. Selected state
L5 / exactly bound plan Assistive workflow: source-bound implementation An agreed outcome becomes a maintained plan, direct or delegated implementation, proportional verification and an honest delivery report. Task context and dependency advice support progress; real user stops and host permissions remain effective. Greenfield and brownfield population use the paired native writer and exact finalization. Exact blob bindings identify inspected source bytes; the linked revision is the committed source base. Implementation, verification, live host qualification and release remain separate claims.
Current source changes Routine artifact instructions now focus on the installed writer, typed plan, bounded task reads, factual prerequisites, actual checks and safe retry. The default leaf is about 7.3 KB and includes an executable stdin bootstrap with optional operation-schema discovery; explicit attestation remains available on demand without burdening ordinary task context. A spec's SMARTS design rationale lives in its approach, decisions and one SEC-SMARTS section. core/surface/includes/artifacts.md The pinned native bridge enables task-context as a bounded informational read and does not require future reviewer/approval producers simply to author a structurally valid draft. Context requests route to ordinary context-preview, context-finalize-preview and exact-reference context-apply; historical receipt input stays a separate compatibility path. Installed host qualification is separate. Native Pi writer admission does not establish a complete live Pi workflow. On NO_LOCAL_STORE the client runs init-store once and retries the request once; a refused initialization keeps NO_LOCAL_STORE. On Windows the binary's architecture comes from the kernel (IsWow64Process2), never WMI or PROCESSOR_* environment. Index and outline pages are now fully checked before rows are yielded, including continuation identity and row shape. core/pysrc/_artifactlib.py Keep spec criteria, plan tasks, prerequisites and exact engine-returned IDs. Valid native IDs need not start with SPEC- or PLAN-. A typed exact-permutation record reorder keeps identities/history but is a normative edit that invalidates approval; it does not silently authorize reexecution.
Related entries Pinned evidence S10 · core/surface/includes/artifacts.mdInstalled admission, typed reads, context completion, authority receipts and format boundaries. S32 · core/pysrc/_artifactauthoritylib.pyNative artifact ID grammar, producer states, Claude join lock, UTF-8/UTF-16-key-order observation binding. S40 · CHANGELOG.mdMerge-scope record for review-join, marker, ID/reorder, Unicode, prerequisite and engine-wait repairs. Selected gate
L6 / complete task-context read Assistive workflow: source-bound implementation An agreed outcome becomes a maintained plan, direct or delegated implementation, proportional verification and an honest delivery report. Task context and dependency advice support progress; real user stops and host permissions remain effective. Greenfield and brownfield population use the paired native writer and exact finalization. Exact blob bindings identify inspected source bytes; the linked revision is the committed source base. Implementation, verification, live host qualification and release remain separate claims.
Current source changes Routine artifact instructions now focus on the installed writer, typed plan, bounded task reads, factual prerequisites, actual checks and safe retry. The default leaf is about 7.3 KB and includes an executable stdin bootstrap with optional operation-schema discovery; explicit attestation remains available on demand without burdening ordinary task context. A spec's SMARTS design rationale lives in its approach, decisions and one SEC-SMARTS section. core/surface/includes/artifacts.md Read-only task-context reports direct dependencies and unresolved ancestors, scoped_changed, scoped_unchanged or unassessed inputs, and actual check observations. Missing broader context remains explicit; the packet does not certify verification. core/artifacts/internal/operations/task_context.go Task-check observations now retain directly referenced criterion identity and an optional bounded source-input fingerprint or an explicit unassessed reason, without executing tests or minting approval receipts. core/artifacts/internal/operations/task_check.go New checks bind the selected criterion and applicable constraint bodies, including explicit references and scope applicability. Unrelated constraints remain outside the digest; old observations never gain unassessed constraint coverage. core/artifacts/internal/model/assistive.go Exact and batch symbol retrieval exposes reported check history without an approval ticket. Outline paging now accounts for encoded rows incrementally while retaining response budgets and continuation behavior. core/artifacts/internal/read/read.go Human HTML presents reported command checks in an escaped, explicitly unverified section; implementation state and source validation remain distinct. core/artifacts/internal/render/render.go The unpublished 0.3.2 native schema adds optional shared_inputs and scoped task-check fields without silently relabeling old records. core/artifacts/schemas/common.schema.json The bounded source-fingerprint reader hashes actual declared task/shared paths, directory membership and known build-control files; unsafe, missing and over-limit scopes are explicitly unassessed rather than blocking work or proving a full dependency graph. core/artifacts/internal/operations/input_scope.go Scheduling follows the full dependency ancestry and exposes unresolved ancestor IDs and reasons. Reconciled progress remains advisory; blocked ancestors prevent premature downstream readiness while independent tasks remain available. Response-budget omissions are explicit, retain the global scheduling decision and direct callers to supported detail reads. core/artifacts/internal/operations/assistive.go Explicit draft source refresh retains both source CAS components and same-pair identity without requiring an unstarted plan. Existing progress, approvals and evidence are preserved without asserting current authority. Dependency invalidation preserves an independently BLOCKED task and its reason. Verification summaries disclose conflicting caller reports separately from receipt freshness; later successful verification covers only reports present when its context was issued. core/artifacts/internal/operations/lifecycle.go Bounded isolated Python stdin diagnostics may read literal protected paths or git show HEAD objects. Unknown code, mutable calls, shell expansion and extra commands retain the guard checks; ordinary commands avoid AST parsing. core/pysrc/_bashguardlib.py identity → index → outline → scoped read Follow contextual pages to completion Retain the final engine context ticket Exact read alone is context-incomplete Stale cursor or ticket → read again
Related entries Pinned evidence S08 · core/surface/skills/subagent-driven-development/SKILL.mdCurrent implementation-dispatch and independent spec-review sections re-read; remaining scope contract also checked against current skill index. Fresh actor input and bounded delivery are distinct from engine tickets. S10 · core/surface/includes/artifacts.mdInstalled admission, typed reads, context completion, authority receipts and format boundaries. Selected state
A context ticket is neither a model memory nor a signature of semantic truth Assistive workflow: source-bound implementation An agreed outcome becomes a maintained plan, direct or delegated implementation, proportional verification and an honest delivery report. Task context and dependency advice support progress; real user stops and host permissions remain effective. Greenfield and brownfield population use the paired native writer and exact finalization. Exact blob bindings identify inspected source bytes; the linked revision is the committed source base. Implementation, verification, live host qualification and release remain separate claims.
Current source changes Routine artifact instructions now focus on the installed writer, typed plan, bounded task reads, factual prerequisites, actual checks and safe retry. The default leaf is about 7.3 KB and includes an executable stdin bootstrap with optional operation-schema discovery; explicit attestation remains available on demand without burdening ordinary task context. A spec's SMARTS design rationale lives in its approach, decisions and one SEC-SMARTS section. core/surface/includes/artifacts.md Typed engine context tickets bind complete task reads. They do not establish actual actor delivery, semantic truth or fresh runtime evidence.
Related entries Pinned evidence S10 · core/surface/includes/artifacts.mdInstalled admission, typed reads, context completion, authority receipts and format boundaries. S27 · core/pysrc/_contextselectlib.pyInert scoped selector and delivery implementation: commands/cwd/evidence, native instruction order, actor/worktree/epoch binding, two-attempt limit and reviewer preflight. Selected agent
Author context Assistive workflow: source-bound implementation An agreed outcome becomes a maintained plan, direct or delegated implementation, proportional verification and an honest delivery report. Task context and dependency advice support progress; real user stops and host permissions remain effective. Greenfield and brownfield population use the paired native writer and exact finalization. Exact blob bindings identify inspected source bytes; the linked revision is the committed source base. Implementation, verification, live host qualification and release remain separate claims.
Current source changes Authors and reviewers share task-critical reference and constraint checks. Missing or stale provenance remains advisory. Two delivery attempts end with an advisory unknown observation, without another retry or an invented receipt. core/pysrc/_contextselectlib.py Feature actors receive scoped requirements, named critical references and observed evidence. Applicable conflicts or missing required inputs stop dependent work; provenance warnings remain visible and advisory. core/pysrc/_artifactpromptlib.py Fresh task obligation + delivered scoped input Check worktree, source and constraints TDD retains the original caller’s gates
Related entries Pinned evidence S08 · core/surface/skills/subagent-driven-development/SKILL.mdCurrent implementation-dispatch and independent spec-review sections re-read; remaining scope contract also checked against current skill index. Fresh actor input and bounded delivery are distinct from engine tickets. S09 · core/surface/skills/tdd/SKILL.mdCurrent TDD context preflight and obligation derivation re-read. Six-phase gate sequence also confirmed by current skill index; fix and test-only paths do not fabricate feature artifacts. S27 · core/pysrc/_contextselectlib.pyInert scoped selector and delivery implementation: commands/cwd/evidence, native instruction order, actor/worktree/epoch binding, two-attempt limit and reviewer preflight. Selected agent
Reviewer context Assistive workflow: source-bound implementation An agreed outcome becomes a maintained plan, direct or delegated implementation, proportional verification and an honest delivery report. Task context and dependency advice support progress; real user stops and host permissions remain effective. Greenfield and brownfield population use the paired native writer and exact finalization. Exact blob bindings identify inspected source bytes; the linked revision is the committed source base. Implementation, verification, live host qualification and release remain separate claims.
Current source changes Authors and reviewers share task-critical reference and constraint checks. Missing or stale provenance remains advisory. Two delivery attempts end with an advisory unknown observation, without another retry or an invented receipt. core/pysrc/_contextselectlib.py Feature actors receive scoped requirements, named critical references and observed evidence. Applicable conflicts or missing required inputs stop dependent work; provenance warnings remain visible and advisory. core/pysrc/_artifactpromptlib.py A review binds one resolved diff and supplies scoped inputs to each reviewer. Missing critical references or applicable conflicts prevent a substantive verdict; provenance and unobserved delivery diagnostics remain advisory. core/surface/commands/review.md Own fresh packet; never reuse author input Required refs + independent source checks Reviewer delivery is separately tracked
Related entries Pinned evidence S08 · core/surface/skills/subagent-driven-development/SKILL.mdCurrent implementation-dispatch and independent spec-review sections re-read; remaining scope contract also checked against current skill index. Fresh actor input and bounded delivery are distinct from engine tickets. S28 · core/surface/commands/review.mdCurrent full review owner: stable diff resolution, selected context delivered independently to each reviewer, local verdict before separately authorized posting. S27 · core/pysrc/_contextselectlib.pyInert scoped selector and delivery implementation: commands/cwd/evidence, native instruction order, actor/worktree/epoch binding, two-attempt limit and reviewer preflight. Selected gate
Runner + authority producers Assistive workflow: source-bound implementation An agreed outcome becomes a maintained plan, direct or delegated implementation, proportional verification and an honest delivery report. Task context and dependency advice support progress; real user stops and host permissions remain effective. Greenfield and brownfield population use the paired native writer and exact finalization. Exact blob bindings identify inspected source bytes; the linked revision is the committed source base. Implementation, verification, live host qualification and release remain separate claims.
Current source changes Routine artifact instructions now focus on the installed writer, typed plan, bounded task reads, factual prerequisites, actual checks and safe retry. The default leaf is about 7.3 KB and includes an executable stdin bootstrap with optional operation-schema discovery; explicit attestation remains available on demand without burdening ordinary task context. A spec's SMARTS design rationale lives in its approach, decisions and one SEC-SMARTS section. core/surface/includes/artifacts.md A new explicit batch review can select a passing rerun of still-reviewed work without replacing the original task review. The selected run must match its own recorded workspace; changed source or supporting evidence still requires reassessment. core/artifacts/internal/evidence/completion.go Read-only task-context reports direct dependencies and unresolved ancestors, scoped_changed, scoped_unchanged or unassessed inputs, and actual check observations. Missing broader context remains explicit; the packet does not certify verification. core/artifacts/internal/operations/task_context.go An unanswered legacy reconciliation can be superseded only when route and pending state agree and no durable attempt or matching observed reply exists. Captured replies recover through their original immutable source and receipt; in-flight mutations retain exact replay and real user stops. core/pysrc/_reconciliationlib.py Task-check observations now retain directly referenced criterion identity and an optional bounded source-input fingerprint or an explicit unassessed reason, without executing tests or minting approval receipts. core/artifacts/internal/operations/task_check.go The closed task-check request accepts bounded command observations matched to plan definitions. Ordinary context-preview and context-finalize-preview omit prompt identity; context-apply accepts an exact preview reference and digest. Invalid requests, changed preimages and stale CAS writes remain rejected. The historical smarts-apply request accepts the 0.2.0 SMARTS decision record. init-store takes a protocol-only request. core/artifacts/internal/operations/protocol.go The bounded source-fingerprint reader hashes actual declared task/shared paths, directory membership and known build-control files; unsafe, missing and over-limit scopes are explicitly unassessed rather than blocking work or proving a full dependency graph. core/artifacts/internal/operations/input_scope.go The candidate-only runbook distinguishes 0.3.1/0.3.2 mixed-schema readers, exact replay and forward recovery while preserving recorded user stops. The paired prompt retirement scans immutable observed replies before calling a request unanswered; submitted or uncertain decisions are never erased. Older-client downgrade after modern edits remains unsupported. docs/artifacts/ASSISTIVE-UPGRADE.md Ordinary approval and sprint instructions bypass legacy capture without changing pending history or inventing authority. Explicit legacy replies retain exact matching. The paired adapter can retire a genuinely unanswered 0.2 sprint-pair request using its serialized cancellation lock and submission check; submitted decisions must replay, while standalone approval 0.1 refuses automatic retirement. core/pysrc/_approvallib.py The old paired approval producer publishes a content-addressed observed reply before its durable submission. Cancellation checks the matching observation, content hash, prompt and frozen context under its existing lock, with a strict scan bound. A matching or uncertain pre-submission reply blocks automatic retirement; normal tasks and context reads have no added cost. core/pysrc/_sprintapprovallib.py The optional reference retains strict historical evidence inputs without replacing ordinary work. Explicit draft source refresh preserves progressed plans and history with both CAS bindings; initial approval restrictions remain on historical paired attestation. Its verify command requires literal absolute paths; linked agent subdirectories remain unsupported for strict authority review. These limitations apply to requested attestation, not ordinary task start. core/surface/includes/artifacts-attestation.md Codex dispatch requires attestation only for explicitly claimed host proof; an unavailable requested farm backend is reported without silently switching provider or model. Ordinary context inspection can proceed inline within actual host permissions; an agent label supplies no containment or semantic verification. core/surface/includes/codex-host-notes.md Historical sprint approval and SMARTS proof instructions remain available only for existing explicit attestation, outside the ordinary execution path. core/surface/includes/sprint-authority.md Ordinary uppercase words such as Approve DATA migration remain ordinary input. All-letter capture codes require a matching retained request; real armed replies, mismatches and user stops preserve their original handling. core/pysrc/_replylib.py Only an actual verifier wrapper executable/script prefix selects authority handling. Inert PowerShell data mentioning the filename is not a verifier call; malformed real verifier invocations still fail closed. Fresh legacy context approval arming is refused; historical pending routes remain readable. core/pysrc/_artifactauthoritylib.py Bounded isolated Python stdin diagnostics may read literal protected paths or git show HEAD objects. Unknown code, mutable calls, shell expansion and extra commands retain the guard checks; ordinary commands avoid AST parsing. core/pysrc/_bashguardlib.py Receipts and current source identity are distinct from displayed status and semantic correctness. Current source uses native canonical Unicode hashing for observation binding. Host read-dedup markers are runtime state, not plan-authorized verification inputs; the marker repair must not be generalized into arbitrary .gitignore exclusions.
Related entries Pinned evidence S10 · core/surface/includes/artifacts.mdInstalled admission, typed reads, context completion, authority receipts and format boundaries. S32 · core/pysrc/_artifactauthoritylib.pyNative artifact ID grammar, producer states, Claude join lock, UTF-8/UTF-16-key-order observation binding. S40 · CHANGELOG.mdMerge-scope record for review-join, marker, ID/reorder, Unicode, prerequisite and engine-wait repairs. Selected state
L7 / acceptance evidence Assistive workflow: source-bound implementation An agreed outcome becomes a maintained plan, direct or delegated implementation, proportional verification and an honest delivery report. Task context and dependency advice support progress; real user stops and host permissions remain effective. Greenfield and brownfield population use the paired native writer and exact finalization. Exact blob bindings identify inspected source bytes; the linked revision is the committed source base. Implementation, verification, live host qualification and release remain separate claims.
Current source changes Routine artifact instructions now focus on the installed writer, typed plan, bounded task reads, factual prerequisites, actual checks and safe retry. The default leaf is about 7.3 KB and includes an executable stdin bootstrap with optional operation-schema discovery; explicit attestation remains available on demand without burdening ordinary task context. A spec's SMARTS design rationale lives in its approach, decisions and one SEC-SMARTS section. core/surface/includes/artifacts.md A new explicit batch review can select a passing rerun of still-reviewed work without replacing the original task review. The selected run must match its own recorded workspace; changed source or supporting evidence still requires reassessment. core/artifacts/internal/evidence/completion.go Status separates reported DONE from independently ACCEPTED work, shows actual blocked decisions, and does not infer an approval request from draft governance state. core/surface/commands/status.md task-review → combined quality accept-scope → current acceptance Commit and finish query proof again
Related entries Pinned evidence S08 · core/surface/skills/subagent-driven-development/SKILL.mdCurrent implementation-dispatch and independent spec-review sections re-read; remaining scope contract also checked against current skill index. Fresh actor input and bounded delivery are distinct from engine tickets. S10 · core/surface/includes/artifacts.mdInstalled admission, typed reads, context completion, authority receipts and format boundaries. S11 · core/surface/skills/commit-gate/SKILL.mdCurrent Phase 5.5 through terminal commit rules re-read. Legacy heal vocabulary must not be generalized to v2 paired native writes. S12 · core/surface/skills/finishing-a-development-branch/SKILL.mdWatch/cleanup early routing; commit and acceptance preflight; PR and merge authority. Selected skill
Legacy v1 heal / native v2 refresh Assistive workflow: source-bound implementation An agreed outcome becomes a maintained plan, direct or delegated implementation, proportional verification and an honest delivery report. Task context and dependency advice support progress; real user stops and host permissions remain effective. Greenfield and brownfield population use the paired native writer and exact finalization. Exact blob bindings identify inspected source bytes; the linked revision is the committed source base. Implementation, verification, live host qualification and release remain separate claims.
Current source changes Ordinary create/adopt/update preserves the same source-evidence, selection and compare-and-swap checks as the native transaction. Missing or matching v1 provenance can be adopted; existing v2 ownership can transition through explicit whole-document selection. Original document/provenance preimages remain in transaction history for exact recovery. core/artifacts/internal/contextdocument/operations.go A drift audit stays read-only; an already authorized selected or full refresh proceeds without asking again per document. The installed draft/preview/apply example preserves source bindings and selected content. Matching hashes establish identity only; initialized refresh never reruns scaffolding, clears the marker or repeats finalization. core/surface/skills/context-check/SKILL.md The lazy reference map routes provenance drift or requested initialized refresh to context-check after reading selected documents, provenance and the code map. The owner reuses selected scope and existing authorization. core/surface/includes/reference-map.md v1: conditional staged-source auto-heal v2: selected owned-field preview + receipt Identity match is not semantic verification No raw v2 rebaseline fallback
Related entries Pinned evidence S11 · core/surface/skills/commit-gate/SKILL.mdCurrent Phase 5.5 through terminal commit rules re-read. Legacy heal vocabulary must not be generalized to v2 paired native writes. S26 · core/surface/skills/context-check/SKILL.mdInitialized refresh: read-only report, explicit selection, v1/v2 separation, native paired updates and preserved human history. Full owner read. Selected skill
Retained state / bounded recovery Assistive workflow: source-bound implementation An agreed outcome becomes a maintained plan, direct or delegated implementation, proportional verification and an honest delivery report. Task context and dependency advice support progress; real user stops and host permissions remain effective. Greenfield and brownfield population use the paired native writer and exact finalization. Exact blob bindings identify inspected source bytes; the linked revision is the committed source base. Implementation, verification, live host qualification and release remain separate claims.
Current source changes Ordinary create/adopt/update preserves the same source-evidence, selection and compare-and-swap checks as the native transaction. Missing or matching v1 provenance can be adopted; existing v2 ownership can transition through explicit whole-document selection. Original document/provenance preimages remain in transaction history for exact recovery. core/artifacts/internal/contextdocument/operations.go A drift audit stays read-only; an already authorized selected or full refresh proceeds without asking again per document. The installed draft/preview/apply example preserves source bindings and selected content. Matching hashes establish identity only; initialized refresh never reruns scaffolding, clears the marker or repeats finalization. core/surface/skills/context-check/SKILL.md Scheduling follows the full dependency ancestry and exposes unresolved ancestor IDs and reasons. Reconciled progress remains advisory; blocked ancestors prevent premature downstream readiness while independent tasks remain available. Response-budget omissions are explicit, retain the global scheduling decision and direct callers to supported detail reads. core/artifacts/internal/operations/assistive.go Explicit draft source refresh retains both source CAS components and same-pair identity without requiring an unstarted plan. Existing progress, approvals and evidence are preserved without asserting current authority. Dependency invalidation preserves an independently BLOCKED task and its reason. Verification summaries disclose conflicting caller reports separately from receipt freshness; later successful verification covers only reports present when its context was issued. core/artifacts/internal/operations/lifecycle.go Refresh keeps initialization and human history Recover native operations by the same ID Committed results are read back, not rewritten Decisions remain with their actual owner
Related entries Pinned evidence S13 · core/surface/skills/context-creation/SKILL.mdCurrent brownfield owner: qualified containment, bounded reports and join, scoped gaps, native document/provenance writes, separate 13-identity finalization and recovery. S26 · core/surface/skills/context-check/SKILL.mdInitialized refresh: read-only report, explicit selection, v1/v2 separation, native paired updates and preserved human history. Full owner read. Selected state
Select task-relevant context Assistive workflow: source-bound implementation An agreed outcome becomes a maintained plan, direct or delegated implementation, proportional verification and an honest delivery report. Task context and dependency advice support progress; real user stops and host permissions remain effective. Greenfield and brownfield population use the paired native writer and exact finalization. Exact blob bindings identify inspected source bytes; the linked revision is the committed source base. Implementation, verification, live host qualification and release remain separate claims.
Current source changes Authors and reviewers share task-critical reference and constraint checks. Missing or stale provenance remains advisory. Two delivery attempts end with an advisory unknown observation, without another retry or an invented receipt. core/pysrc/_contextselectlib.py select_task_context is inert. The caller acquires current inputs and native instruction precedence. Copied upstream rules, source-scoped ADRs and declared commands do not gain authority or current verification from packet inclusion.
Related entries Pinned evidence S27 · core/pysrc/_contextselectlib.pyInert scoped selector and delivery implementation: commands/cwd/evidence, native instruction order, actor/worktree/epoch binding, two-attempt limit and reviewer preflight. Selected skill
Compose for this lane and actor Assistive workflow: source-bound implementation An agreed outcome becomes a maintained plan, direct or delegated implementation, proportional verification and an honest delivery report. Task context and dependency advice support progress; real user stops and host permissions remain effective. Greenfield and brownfield population use the paired native writer and exact finalization. Exact blob bindings identify inspected source bytes; the linked revision is the committed source base. Implementation, verification, live host qualification and release remain separate claims.
Current source changes Authors and reviewers share task-critical reference and constraint checks. Missing or stale provenance remains advisory. Two delivery attempts end with an advisory unknown observation, without another retry or an invented receipt. core/pysrc/_contextselectlib.py Feature actors receive scoped requirements, named critical references and observed evidence. Applicable conflicts or missing required inputs stop dependent work; provenance warnings remain visible and advisory. core/pysrc/_artifactpromptlib.py Confirmed fixes preserve causal evidence and a named regression. Delegated input retains critical requirements; advisory delivery exhaustion ends retries while actual missing inputs and conflicts retain their dependent-work boundary. Distinct remediation strategies are compared with SMARTS only after the root cause is confirmed. core/surface/commands/fix.md compose_feature_actor_input is named by the current feature/TDD/SDD contracts. compose_fix_or_test_actor_input preserves bug origin and the existing route. prepare_review_input blocks missing critical references before a substantive verdict.
Related entries Pinned evidence S27 · core/pysrc/_contextselectlib.pyInert scoped selector and delivery implementation: commands/cwd/evidence, native instruction order, actor/worktree/epoch binding, two-attempt limit and reviewer preflight. S09 · core/surface/skills/tdd/SKILL.mdCurrent TDD context preflight and obligation derivation re-read. Six-phase gate sequence also confirmed by current skill index; fix and test-only paths do not fabricate feature artifacts. S08 · core/surface/skills/subagent-driven-development/SKILL.mdCurrent implementation-dispatch and independent spec-review sections re-read; remaining scope contract also checked against current skill index. Fresh actor input and bounded delivery are distinct from engine tickets. S28 · core/surface/commands/review.mdCurrent full review owner: stable diff resolution, selected context delivered independently to each reviewer, local verdict before separately authorized posting. Selected gate
Bound actual actor delivery Assistive workflow: source-bound implementation An agreed outcome becomes a maintained plan, direct or delegated implementation, proportional verification and an honest delivery report. Task context and dependency advice support progress; real user stops and host permissions remain effective. Greenfield and brownfield population use the paired native writer and exact finalization. Exact blob bindings identify inspected source bytes; the linked revision is the committed source base. Implementation, verification, live host qualification and release remain separate claims.
Current source changes Authors and reviewers share task-critical reference and constraint checks. Missing or stale provenance remains advisory. Two delivery attempts end with an advisory unknown observation, without another retry or an invented receipt. core/pysrc/_contextselectlib.py Confirmed fixes preserve causal evidence and a named regression. Delegated input retains critical requirements; advisory delivery exhaustion ends retries while actual missing inputs and conflicts retain their dependent-work boundary. Distinct remediation strategies are compared with SMARTS only after the root cause is confirmed. core/surface/commands/fix.md A review binds one resolved diff and supplies scoped inputs to each reviewer. Missing critical references or applicable conflicts prevent a substantive verdict; provenance and unobserved delivery diagnostics remain advisory. core/surface/commands/review.md prepare_actor_delivery binds the selected packet digest, actor/task, actual worktree and observable epochs. The caller sends delivery_text on deliver and retains the counted attempt. A reported host-load-event or child-input-capture is not a native task-admission certificate.
Related entries Pinned evidence S27 · core/pysrc/_contextselectlib.pyInert scoped selector and delivery implementation: commands/cwd/evidence, native instruction order, actor/worktree/epoch binding, two-attempt limit and reviewer preflight. S08 · core/surface/skills/subagent-driven-development/SKILL.mdCurrent implementation-dispatch and independent spec-review sections re-read; remaining scope contract also checked against current skill index. Fresh actor input and bounded delivery are distinct from engine tickets. Selected warn
Reselect or stop Assistive workflow: source-bound implementation An agreed outcome becomes a maintained plan, direct or delegated implementation, proportional verification and an honest delivery report. Task context and dependency advice support progress; real user stops and host permissions remain effective. Greenfield and brownfield population use the paired native writer and exact finalization. Exact blob bindings identify inspected source bytes; the linked revision is the committed source base. Implementation, verification, live host qualification and release remain separate claims.
Current source changes Authors and reviewers share task-critical reference and constraint checks. Missing or stale provenance remains advisory. Two delivery attempts end with an advisory unknown observation, without another retry or an invented receipt. core/pysrc/_contextselectlib.py Two counted attempts apply to the same actor/task/worktree/epoch/packet binding. Unknown epochs cannot make a saved receipt current. No shared session marker or author receipt substitutes for a reviewer’s input delivery.
Related entries Pinned evidence S27 · core/pysrc/_contextselectlib.pyInert scoped selector and delivery implementation: commands/cwd/evidence, native instruction order, actor/worktree/epoch binding, two-attempt limit and reviewer preflight. S08 · core/surface/skills/subagent-driven-development/SKILL.mdCurrent implementation-dispatch and independent spec-review sections re-read; remaining scope contract also checked against current skill index. Fresh actor input and bounded delivery are distinct from engine tickets. Handoffs, branches and returns Change lanes / alternate entries and exits Different starting evidence creates different obligations. A fix, a refactor, a docs edit and an experiment do not share identical ceremony.
Handoffs, branches and returns Review + delivery / evidence and authority Review, commit and branch disposition retain separate authority; release preparation stops at its PR.
Selected cmd
/review · /checkpoint Assistive workflow: source-bound implementation An agreed outcome becomes a maintained plan, direct or delegated implementation, proportional verification and an honest delivery report. Task context and dependency advice support progress; real user stops and host permissions remain effective. Greenfield and brownfield population use the paired native writer and exact finalization. Exact blob bindings identify inspected source bytes; the linked revision is the committed source base. Implementation, verification, live host qualification and release remain separate claims.
Current source changes A review binds one resolved diff and supplies scoped inputs to each reviewer. Missing critical references or applicable conflicts prevent a substantive verdict; provenance and unobserved delivery diagnostics remain advisory. core/surface/commands/review.md A checkpoint retains the reviewed commit, UTC review-start time and worktree dirty state, then writes the report through its aggregator. last-checkpoint holds only the override count. core/surface/commands/checkpoint.md Checkpoint reports persist observed source_commit, recorded_at and worktree_dirty values together with every review finding. Missing anchors remain unknown. core/surface/agents/checkpoint-aggregator.md /review: current diff /checkpoint: periodic codebase sweep /preview predicts only; no verdict /tribunal and /threat-model are opt-in
Related entries Pinned evidence S04 · core/surface/includes/routing-table.mdAuthoritative intent-to-owner table, scope-touch routes and stop/return semantics. S18 · core/surface/includes/review-matrix.mdShared diff-path reviewer selection for review/checkpoint/preview; callers retain their own gates. S21 · core/surface/skills/INDEX.mdCurrent routine-owner index: five debug dispositions, bounded workflow owners and unchanged six-phase TDD / nine-phase commit sequence. Selected agent
Fresh inputs → scoped reviewer fleet Assistive workflow: source-bound implementation An agreed outcome becomes a maintained plan, direct or delegated implementation, proportional verification and an honest delivery report. Task context and dependency advice support progress; real user stops and host permissions remain effective. Greenfield and brownfield population use the paired native writer and exact finalization. Exact blob bindings identify inspected source bytes; the linked revision is the committed source base. Implementation, verification, live host qualification and release remain separate claims.
Current source changes Authors and reviewers share task-critical reference and constraint checks. Missing or stale provenance remains advisory. Two delivery attempts end with an advisory unknown observation, without another retry or an invented receipt. core/pysrc/_contextselectlib.py A review binds one resolved diff and supplies scoped inputs to each reviewer. Missing critical references or applicable conflicts prevent a substantive verdict; provenance and unobserved delivery diagnostics remain advisory. core/surface/commands/review.md Select current context for each reviewer Required refs + native instruction order Deliver to that actual child, not its parent Then dispatch by the actual diff paths Preserve capability and evidence gaps
Related entries Pinned evidence S28 · core/surface/commands/review.mdCurrent full review owner: stable diff resolution, selected context delivered independently to each reviewer, local verdict before separately authorized posting. S27 · core/pysrc/_contextselectlib.pyInert scoped selector and delivery implementation: commands/cwd/evidence, native instruction order, actor/worktree/epoch binding, two-attempt limit and reviewer preflight. S18 · core/surface/includes/review-matrix.mdShared diff-path reviewer selection for review/checkpoint/preview; callers retain their own gates. Selected cmd
/commit Assistive workflow: source-bound implementation An agreed outcome becomes a maintained plan, direct or delegated implementation, proportional verification and an honest delivery report. Task context and dependency advice support progress; real user stops and host permissions remain effective. Greenfield and brownfield population use the paired native writer and exact finalization. Exact blob bindings identify inspected source bytes; the linked revision is the committed source base. Implementation, verification, live host qualification and release remain separate claims.
Current source changes Routine artifact instructions now focus on the installed writer, typed plan, bounded task reads, factual prerequisites, actual checks and safe retry. The default leaf is about 7.3 KB and includes an executable stdin bootstrap with optional operation-schema discovery; explicit attestation remains available on demand without burdening ordinary task context. A spec's SMARTS design rationale lives in its approach, decisions and one SEC-SMARTS section. core/surface/includes/artifacts.md Current staged scope TDD complete where applicable Typed plans: current acceptance No status-label substitutes
Related entries Pinned evidence S11 · core/surface/skills/commit-gate/SKILL.mdCurrent Phase 5.5 through terminal commit rules re-read. Legacy heal vocabulary must not be generalized to v2 paired native writes. S10 · core/surface/includes/artifacts.mdInstalled admission, typed reads, context completion, authority receipts and format boundaries. Selected gate
Phases 4–6 / prove + inspect Assistive workflow: source-bound implementation An agreed outcome becomes a maintained plan, direct or delegated implementation, proportional verification and an honest delivery report. Task context and dependency advice support progress; real user stops and host permissions remain effective. Greenfield and brownfield population use the paired native writer and exact finalization. Exact blob bindings identify inspected source bytes; the linked revision is the committed source base. Implementation, verification, live host qualification and release remain separate claims.
Current source changes A drift audit stays read-only; an already authorized selected or full refresh proceeds without asking again per document. The installed draft/preview/apply example preserves source bindings and selected content. Matching hashes establish identity only; initialized refresh never reruns scaffolding, clears the marker or repeats finalization. core/surface/skills/context-check/SKILL.md 4 Focused tests, lint, secrets, scope gates 5 Fresh behavioral proof 5.5 Legacy v1 staged-source auto-heal 6 Complete diff review v2 pairs require native selected updates
Related entries Pinned evidence S11 · core/surface/skills/commit-gate/SKILL.mdCurrent Phase 5.5 through terminal commit rules re-read. Legacy heal vocabulary must not be generalized to v2 paired native writes. S26 · core/surface/skills/context-check/SKILL.mdInitialized refresh: read-only report, explicit selection, v1/v2 separation, native paired updates and preserved human history. Full owner read. Selected cmd
/release · selected target Assistive workflow: source-bound implementation An agreed outcome becomes a maintained plan, direct or delegated implementation, proportional verification and an honest delivery report. Task context and dependency advice support progress; real user stops and host permissions remain effective. Greenfield and brownfield population use the paired native writer and exact finalization. Exact blob bindings identify inspected source bytes; the linked revision is the committed source base. Implementation, verification, live host qualification and release remain separate claims.
Current source changes Release preparation inserts a new version below the entire Unreleased section and above prior versions, preserving pending notes. This guidance correction does not grant publication authority or replace the existing release-PR and hosted qualification boundaries. core/surface/skills/release/SKILL.md Resolve declared target + policy Prepare reviewed release surfaces Dry-run stops at its report
Related entries Pinned evidence S33 · core/surface/skills/release/SKILL.mdInteractive release preparation and mandatory release-PR boundary; qualified hosted tag composition. Handoffs, branches and returns Knowledge + operations / supporting routes These entries have bounded outputs: decisions, reports, board transitions, repair or runtime posture. They do not implicitly enter implementation.
Selected skill
context-check / selected refresh Assistive workflow: source-bound implementation An agreed outcome becomes a maintained plan, direct or delegated implementation, proportional verification and an honest delivery report. Task context and dependency advice support progress; real user stops and host permissions remain effective. Greenfield and brownfield population use the paired native writer and exact finalization. Exact blob bindings identify inspected source bytes; the linked revision is the committed source base. Implementation, verification, live host qualification and release remain separate claims.
Current source changes The user-facing route table points feature and sprint work at useful planned execution and selected context refresh at its existing owner. Existing authorization covers ordinary work; actual host, destructive-action and commit controls remain. core/surface/includes/routing-table.md The skill index connects planning, direct or delegated implementation, proportional verification and selected context maintenance to their owning workflows. core/surface/skills/INDEX.md The preparation helper binds bounded drafts and retained sources to exact document/provenance preimages. Finalization preparation hashes all thirteen required outputs and retains missing output failures without modifying files. core/pysrc/_contextwriterlib.py A drift audit stays read-only; an already authorized selected or full refresh proceeds without asking again per document. The installed draft/preview/apply example preserves source bindings and selected content. Matching hashes establish identity only; initialized refresh never reruns scaffolding, clears the marker or repeats finalization. core/surface/skills/context-check/SKILL.md Report v1 and v2 states separately Select affected docs or explicit full refresh Qualified scouts + native v2 pair writes
Related entries Pinned evidence S26 · core/surface/skills/context-check/SKILL.mdInitialized refresh: read-only report, explicit selection, v1/v2 separation, native paired updates and preserved human history. Full owner read. S13 · core/surface/skills/context-creation/SKILL.mdCurrent brownfield owner: qualified containment, bounded reports and join, scoped gaps, native document/provenance writes, separate 13-identity finalization and recovery. Selected out
Status / drift report Assistive workflow: source-bound implementation An agreed outcome becomes a maintained plan, direct or delegated implementation, proportional verification and an honest delivery report. Task context and dependency advice support progress; real user stops and host permissions remain effective. Greenfield and brownfield population use the paired native writer and exact finalization. Exact blob bindings identify inspected source bytes; the linked revision is the committed source base. Implementation, verification, live host qualification and release remain separate claims.
Current source changes A drift audit stays read-only; an already authorized selected or full refresh proceeds without asking again per document. The installed draft/preview/apply example preserves source bindings and selected content. Matching hashes establish identity only; initialized refresh never reruns scaffolding, clears the marker or repeats finalization. core/surface/skills/context-check/SKILL.md Read-only report, or selected native update Keep marker, unselected files and history No staging, commit or reinitialization
Related entries Pinned evidence S26 · core/surface/skills/context-check/SKILL.mdInitialized refresh: read-only report, explicit selection, v1/v2 separation, native paired updates and preserved human history. Full owner read. S13 · core/surface/skills/context-creation/SKILL.mdCurrent brownfield owner: qualified containment, bounded reports and join, scoped gaps, native document/provenance writes, separate 13-identity finalization and recovery. Selected cmd
/audit → dated governance packet Assistive workflow: source-bound implementation An agreed outcome becomes a maintained plan, direct or delegated implementation, proportional verification and an honest delivery report. Task context and dependency advice support progress; real user stops and host permissions remain effective. Greenfield and brownfield population use the paired native writer and exact finalization. Exact blob bindings identify inspected source bytes; the linked revision is the committed source base. Implementation, verification, live host qualification and release remain separate claims.
Current source changes A checkpoint retains the reviewed commit, UTC review-start time and worktree dirty state, then writes the report through its aggregator. last-checkpoint holds only the override count. core/surface/commands/checkpoint.md Checkpoint reports persist observed source_commit, recorded_at and worktree_dirty values together with every review finding. Missing anchors remain unknown. core/surface/agents/checkpoint-aggregator.md Checkpoint audit windows use the report timestamp and an ancestor commit. Invalid or missing anchors require an explicit ref/date window; dirty review scope remains a stated limitation. core/surface/commands/audit.md Read source records for a selected window. Preserve audit lines. Write a new packet only; never overwrite the source or an older packet.
Related entries Pinned evidence S02 · core/surface/COMMANDS.mdCatalog, modes, host-limited commands and purpose of each public route. S04 · core/surface/includes/routing-table.mdAuthoritative intent-to-owner table, scope-touch routes and stop/return semantics. Selected cmd
/decompose · /create-context Assistive workflow: source-bound implementation An agreed outcome becomes a maintained plan, direct or delegated implementation, proportional verification and an honest delivery report. Task context and dependency advice support progress; real user stops and host permissions remain effective. Greenfield and brownfield population use the paired native writer and exact finalization. Exact blob bindings identify inspected source bytes; the linked revision is the committed source base. Implementation, verification, live host qualification and release remain separate claims.
Current source changes The installed context-draft command prepares a paired document preview; context-finalize prepares initial finalization. context-apply commits the exact selected preview and digest. Context modes preserve scaffold and initialized refresh boundaries. core/pysrc/init-codearbiter.py Initial brownfield setup checks the intended repository and actual host permissions, then uses bounded inspection and the installed safe context writer. Ordinary Windows work needs no special scout profile or per-document artifact approval. An existing initialized repository routes to selected refresh instead of another scaffold or population cycle. core/surface/commands/init.md Greenfield interviews persist under plans/decomposition and retain actual decisions. Five document/v2-provenance pairs use context-draft and context-apply; context-finalize binds all thirteen outputs. ADRs and task rows use their owners, and retained interview evidence survives finalization. core/surface/skills/decompose/SKILL.md Initial population only: /init --greenfield / --brownfield. Initialized refresh stays with /status drift and explicit selection.
Related entries Pinned evidence S23 · core/surface/commands/init.mdCurrent init owner: activation/scaffold versus initial population; initialized refresh belongs to status drift, not reinitialization. S26 · core/surface/skills/context-check/SKILL.mdInitialized refresh: read-only report, explicit selection, v1/v2 separation, native paired updates and preserved human history. Full owner read. Handoffs, branches and returns Brownfield context · create, refresh, recover One content/provenance pair per native transaction; initial finalization is separate. Existing human records retain their owners.
Selected cmd
Initial population Assistive workflow: source-bound implementation An agreed outcome becomes a maintained plan, direct or delegated implementation, proportional verification and an honest delivery report. Task context and dependency advice support progress; real user stops and host permissions remain effective. Greenfield and brownfield population use the paired native writer and exact finalization. Exact blob bindings identify inspected source bytes; the linked revision is the committed source base. Implementation, verification, live host qualification and release remain separate claims.
Current source changes The installed context-draft command prepares a paired document preview; context-finalize prepares initial finalization. context-apply commits the exact selected preview and digest. Context modes preserve scaffold and initialized refresh boundaries. core/pysrc/init-codearbiter.py Brownfield population covers six logical categories and five content documents through direct bounded inspection or optional scoped agents. The paired writer preserves source bindings and protected content; initial finalization checks all required outputs. core/surface/skills/context-creation/SKILL.md Initial brownfield setup checks the intended repository and actual host permissions, then uses bounded inspection and the installed safe context writer. Ordinary Windows work needs no special scout profile or per-document artifact approval. An existing initialized repository routes to selected refresh instead of another scaffold or population cycle. core/surface/commands/init.md /init --brownfield → context-creation Meaningful source; initialized marker absent A stub is not a completed project context
Related entries Pinned evidence S13 · core/surface/skills/context-creation/SKILL.mdCurrent brownfield owner: qualified containment, bounded reports and join, scoped gaps, native document/provenance writes, separate 13-identity finalization and recovery. S23 · core/surface/commands/init.mdCurrent init owner: activation/scaffold versus initial population; initialized refresh belongs to status drift, not reinitialization. Selected cmd
Initialized refresh Assistive workflow: source-bound implementation An agreed outcome becomes a maintained plan, direct or delegated implementation, proportional verification and an honest delivery report. Task context and dependency advice support progress; real user stops and host permissions remain effective. Greenfield and brownfield population use the paired native writer and exact finalization. Exact blob bindings identify inspected source bytes; the linked revision is the committed source base. Implementation, verification, live host qualification and release remain separate claims.
Current source changes The preparation helper binds bounded drafts and retained sources to exact document/provenance preimages. Finalization preparation hashes all thirteen required outputs and retains missing output failures without modifying files. core/pysrc/_contextwriterlib.py The installed context-draft command prepares a paired document preview; context-finalize prepares initial finalization. context-apply commits the exact selected preview and digest. Context modes preserve scaffold and initialized refresh boundaries. core/pysrc/init-codearbiter.py A drift audit stays read-only; an already authorized selected or full refresh proceeds without asking again per document. The installed draft/preview/apply example preserves source bindings and selected content. Matching hashes establish identity only; initialized refresh never reruns scaffolding, clears the marker or repeats finalization. core/surface/skills/context-check/SKILL.md The lazy reference map routes provenance drift or requested initialized refresh to context-check after reading selected documents, provenance and the code map. The owner reuses selected scope and existing authorization. core/surface/includes/reference-map.md Routine maintenance selects changed v2 claims intersecting staged sources using a current bounded snapshot. Unknown evidence defers the affected document; native paired writes preserve owned fields and current identities. core/pysrc/_provenancelib.py /status drift → context-check Read-only report, then explicit scoped/full choice Do not clear the marker or rerun initialization
Related entries Pinned evidence S26 · core/surface/skills/context-check/SKILL.mdInitialized refresh: read-only report, explicit selection, v1/v2 separation, native paired updates and preserved human history. Full owner read. Selected cmd
Interrupted operation Assistive workflow: source-bound implementation An agreed outcome becomes a maintained plan, direct or delegated implementation, proportional verification and an honest delivery report. Task context and dependency advice support progress; real user stops and host permissions remain effective. Greenfield and brownfield population use the paired native writer and exact finalization. Exact blob bindings identify inspected source bytes; the linked revision is the committed source base. Implementation, verification, live host qualification and release remain separate claims.
Current source changes Ordinary context-preview retains an inert proposal and exact preimages without approval or review claims. context-apply revalidates the digest, closed request and preview binding before the native paired transaction; context-finalize-preview uses the same exact-reference path for initial population. No approve-context reply is required. core/artifacts/internal/operations/context_assistive.go Brownfield population covers six logical categories and five content documents through direct bounded inspection or optional scoped agents. The paired writer preserves source bindings and protected content; initial finalization checks all required outputs. core/surface/skills/context-creation/SKILL.md Inventory actual files, provenance and outcome Inspect the same operation ID and native journal Do not blindly repeat a possibly committed write
Related entries Pinned evidence S13 · core/surface/skills/context-creation/SKILL.mdCurrent brownfield owner: qualified containment, bounded reports and join, scoped gaps, native document/provenance writes, separate 13-identity finalization and recovery. Selected gate
Qualified containment first Assistive workflow: source-bound implementation An agreed outcome becomes a maintained plan, direct or delegated implementation, proportional verification and an honest delivery report. Task context and dependency advice support progress; real user stops and host permissions remain effective. Greenfield and brownfield population use the paired native writer and exact finalization. Exact blob bindings identify inspected source bytes; the linked revision is the committed source base. Implementation, verification, live host qualification and release remain separate claims.
Current source changes Codex dispatch requires attestation only for explicitly claimed host proof; an unavailable requested farm backend is reported without silently switching provider or model. Ordinary context inspection can proceed inline within actual host permissions; an agent label supplies no containment or semantic verification. core/surface/includes/codex-host-notes.md Brownfield population covers six logical categories and five content documents through direct bounded inspection or optional scoped agents. The paired writer preserves source bindings and protected content; initial finalization checks all required outputs. core/surface/skills/context-creation/SKILL.md Initial brownfield setup checks the intended repository and actual host permissions, then uses bounded inspection and the installed safe context writer. Ordinary Windows work needs no special scout profile or per-document artifact approval. An existing initialized repository routes to selected refresh instead of another scaffold or population cycle. core/surface/commands/init.md Optional context scouts report only their assigned scope, source identities, coverage, conflicts and unknowns. Six logical categories do not require six child agents. A role label or read-only instruction proves neither sandbox containment nor independent semantic verification; actual host permissions and excluded paths still apply. core/surface/agents/scout.md The optional context-profile preparation helper is a repository-only CI fixture, not installed plugin capability. Caller-owned OS/tool containment, child binding, launch budgets and actual observations must be qualified before dispatch. Unknown host behavior keeps this step unavailable.
Related entries Pinned evidence S13 · core/surface/skills/context-creation/SKILL.mdCurrent brownfield owner: qualified containment, bounded reports and join, scoped gaps, native document/provenance writes, separate 13-identity finalization and recovery. Selected agent
Bounded logical assignments Assistive workflow: source-bound implementation An agreed outcome becomes a maintained plan, direct or delegated implementation, proportional verification and an honest delivery report. Task context and dependency advice support progress; real user stops and host permissions remain effective. Greenfield and brownfield population use the paired native writer and exact finalization. Exact blob bindings identify inspected source bytes; the linked revision is the committed source base. Implementation, verification, live host qualification and release remain separate claims.
Current source changes Codex dispatch requires attestation only for explicitly claimed host proof; an unavailable requested farm backend is reported without silently switching provider or model. Ordinary context inspection can proceed inline within actual host permissions; an agent label supplies no containment or semantic verification. core/surface/includes/codex-host-notes.md Brownfield population covers six logical categories and five content documents through direct bounded inspection or optional scoped agents. The paired writer preserves source bindings and protected content; initial finalization checks all required outputs. core/surface/skills/context-creation/SKILL.md Optional context scouts report only their assigned scope, source identities, coverage, conflicts and unknowns. Six logical categories do not require six child agents. A role label or read-only instruction proves neither sandbox containment nor independent semantic verification; actual host permissions and excluded paths still apply. core/surface/agents/scout.md A–F cover stack, infrastructure, architecture, security, tests and data. Scope/subdivision/retries retain one run ledger and frozen source boundary. No missing or failed transport can be counted as complete coverage.
Related entries Pinned evidence S13 · core/surface/skills/context-creation/SKILL.mdCurrent brownfield owner: qualified containment, bounded reports and join, scoped gaps, native document/provenance writes, separate 13-identity finalization and recovery. S26 · core/surface/skills/context-check/SKILL.mdInitialized refresh: read-only report, explicit selection, v1/v2 separation, native paired updates and preserved human history. Full owner read. Selected skill
Validate and join reports Assistive workflow: source-bound implementation An agreed outcome becomes a maintained plan, direct or delegated implementation, proportional verification and an honest delivery report. Task context and dependency advice support progress; real user stops and host permissions remain effective. Greenfield and brownfield population use the paired native writer and exact finalization. Exact blob bindings identify inspected source bytes; the linked revision is the committed source base. Implementation, verification, live host qualification and release remain separate claims.
Current source changes Brownfield population covers six logical categories and five content documents through direct bounded inspection or optional scoped agents. The paired writer preserves source bindings and protected content; initial finalization checks all required outputs. core/surface/skills/context-creation/SKILL.md Optional context scouts report only their assigned scope, source identities, coverage, conflicts and unknowns. Six logical categories do not require six child agents. A role label or read-only instruction proves neither sandbox containment nor independent semantic verification; actual host permissions and excluded paths still apply. core/surface/agents/scout.md Closed assignment/snapshot binding 48 KiB normalized join limit Missing / failed transport → block Recheck the source snapshot
Related entries Pinned evidence S13 · core/surface/skills/context-creation/SKILL.mdCurrent brownfield owner: qualified containment, bounded reports and join, scoped gaps, native document/provenance writes, separate 13-identity finalization and recovery. Selected state
Five content targets Assistive workflow: source-bound implementation An agreed outcome becomes a maintained plan, direct or delegated implementation, proportional verification and an honest delivery report. Task context and dependency advice support progress; real user stops and host permissions remain effective. Greenfield and brownfield population use the paired native writer and exact finalization. Exact blob bindings identify inspected source bytes; the linked revision is the committed source base. Implementation, verification, live host qualification and release remain separate claims.
Current source changes The finite context contract permits a whole-document Markdown entry only as the sole entry for one of five named content targets. Explicit typed field entries remain available; the writer does not expose arbitrary repository-file writes. core/artifacts/internal/contextdocument/contract.go Whole-document replacement requires the explicit document selection and exact existing bytes. It preserves CONTEXT frontmatter and initialization state, checks the named document identity and Markdown structure, and retains the code-map size limit. Human content outside the authorized change remains the caller's preservation obligation. core/artifacts/internal/contextdocument/document_selection.go The context schema describes both finite field entries and the explicit whole-document entry for the five supported content documents. Its typed structure records evidence references; schema validity does not establish that the claims were independently reviewed. core/artifacts/schemas/context.schema.json Brownfield population covers six logical categories and five content documents through direct bounded inspection or optional scoped agents. The paired writer preserves source bindings and protected content; initial finalization checks all required outputs. core/surface/skills/context-creation/SKILL.md CONTEXT.md • tech-stack.md coding-standards.md security-controls.md • code-map.md Each with .provenance/<doc>.json
Related entries Pinned evidence S13 · core/surface/skills/context-creation/SKILL.mdCurrent brownfield owner: qualified containment, bounded reports and join, scoped gaps, native document/provenance writes, separate 13-identity finalization and recovery. S26 · core/surface/skills/context-check/SKILL.mdInitialized refresh: read-only report, explicit selection, v1/v2 separation, native paired updates and preserved human history. Full owner read. Selected skill
Finite typed preview Assistive workflow: source-bound implementation An agreed outcome becomes a maintained plan, direct or delegated implementation, proportional verification and an honest delivery report. Task context and dependency advice support progress; real user stops and host permissions remain effective. Greenfield and brownfield population use the paired native writer and exact finalization. Exact blob bindings identify inspected source bytes; the linked revision is the committed source base. Implementation, verification, live host qualification and release remain separate claims.
Current source changes The pinned native bridge enables task-context as a bounded informational read and does not require future reviewer/approval producers simply to author a structurally valid draft. Context requests route to ordinary context-preview, context-finalize-preview and exact-reference context-apply; historical receipt input stays a separate compatibility path. Installed host qualification is separate. Native Pi writer admission does not establish a complete live Pi workflow. On NO_LOCAL_STORE the client runs init-store once and retries the request once; a refused initialization keeps NO_LOCAL_STORE. On Windows the binary's architecture comes from the kernel (IsWow64Process2), never WMI or PROCESSOR_* environment. Index and outline pages are now fully checked before rows are yielded, including continuation identity and row shape. core/pysrc/_artifactlib.py The closed task-check request accepts bounded command observations matched to plan definitions. Ordinary context-preview and context-finalize-preview omit prompt identity; context-apply accepts an exact preview reference and digest. Invalid requests, changed preimages and stale CAS writes remain rejected. The historical smarts-apply request accepts the 0.2.0 SMARTS decision record. init-store takes a protocol-only request. core/artifacts/internal/operations/protocol.go The preparation helper binds bounded drafts and retained sources to exact document/provenance preimages. Finalization preparation hashes all thirteen required outputs and retains missing output failures without modifying files. core/pysrc/_contextwriterlib.py The installed context-draft command prepares a paired document preview; context-finalize prepares initial finalization. context-apply commits the exact selected preview and digest. Context modes preserve scaffold and initialized refresh boundaries. core/pysrc/init-codearbiter.py Ordinary context-preview retains an inert proposal and exact preimages without approval or review claims. context-apply revalidates the digest, closed request and preview binding before the native paired transaction; context-finalize-preview uses the same exact-reference path for initial population. No approve-context reply is required. core/artifacts/internal/operations/context_assistive.go The finite context contract permits a whole-document Markdown entry only as the sole entry for one of five named content targets. Explicit typed field entries remain available; the writer does not expose arbitrary repository-file writes. core/artifacts/internal/contextdocument/contract.go Whole-document replacement requires the explicit document selection and exact existing bytes. It preserves CONTEXT frontmatter and initialization state, checks the named document identity and Markdown structure, and retains the code-map size limit. Human content outside the authorized change remains the caller's preservation obligation. core/artifacts/internal/contextdocument/document_selection.go The context schema describes both finite field entries and the explicit whole-document entry for the five supported content documents. Its typed structure records evidence references; schema validity does not establish that the claims were independently reviewed. core/artifacts/schemas/context.schema.json A drift audit stays read-only; an already authorized selected or full refresh proceeds without asking again per document. The installed draft/preview/apply example preserves source bindings and selected content. Matching hashes establish identity only; initialized refresh never reruns scaffolding, clears the marker or repeats finalization. core/surface/skills/context-check/SKILL.md The lazy reference map routes provenance drift or requested initialized refresh to context-check after reading selected documents, provenance and the code map. The owner reuses selected scope and existing authorization. core/surface/includes/reference-map.md New document or selected owned field Current source / target identities Exact proposed content + provenance Unsupported human layout → refuse
Related entries Pinned evidence S13 · core/surface/skills/context-creation/SKILL.mdCurrent brownfield owner: qualified containment, bounded reports and join, scoped gaps, native document/provenance writes, separate 13-identity finalization and recovery. S26 · core/surface/skills/context-check/SKILL.mdInitialized refresh: read-only report, explicit selection, v1/v2 separation, native paired updates and preserved human history. Full owner read. Selected gate
Native approve-context reply Assistive workflow: source-bound implementation An agreed outcome becomes a maintained plan, direct or delegated implementation, proportional verification and an honest delivery report. Task context and dependency advice support progress; real user stops and host permissions remain effective. Greenfield and brownfield population use the paired native writer and exact finalization. Exact blob bindings identify inspected source bytes; the linked revision is the committed source base. Implementation, verification, live host qualification and release remain separate claims.
Current source changes The closed task-check request accepts bounded command observations matched to plan definitions. Ordinary context-preview and context-finalize-preview omit prompt identity; context-apply accepts an exact preview reference and digest. Invalid requests, changed preimages and stale CAS writes remain rejected. The historical smarts-apply request accepts the 0.2.0 SMARTS decision record. init-store takes a protocol-only request. core/artifacts/internal/operations/protocol.go The optional reference retains strict historical evidence inputs without replacing ordinary work. Explicit draft source refresh preserves progressed plans and history with both CAS bindings; initial approval restrictions remain on historical paired attestation. Its verify command requires literal absolute paths; linked agent subdirectories remain unsupported for strict authority review. These limitations apply to requested attestation, not ordinary task start. core/surface/includes/artifacts-attestation.md Ordinary context-preview retains an inert proposal and exact preimages without approval or review claims. context-apply revalidates the digest, closed request and preview binding before the native paired transaction; context-finalize-preview uses the same exact-reference path for initial population. No approve-context reply is required. core/artifacts/internal/operations/context_assistive.go Brownfield population covers six logical categories and five content documents through direct bounded inspection or optional scoped agents. The paired writer preserves source bindings and protected content; initial finalization checks all required outputs. core/surface/skills/context-creation/SKILL.md A drift audit stays read-only; an already authorized selected or full refresh proceeds without asking again per document. The installed draft/preview/apply example preserves source bindings and selected content. Matching hashes establish identity only; initialized refresh never reruns scaffolding, clears the marker or repeats finalization. core/surface/skills/context-check/SKILL.md Review this exact proposed pair Capture actual host-observed consent No reused approval, hash-only consent or caller-supplied authority
Related entries Pinned evidence S13 · core/surface/skills/context-creation/SKILL.mdCurrent brownfield owner: qualified containment, bounded reports and join, scoped gaps, native document/provenance writes, separate 13-identity finalization and recovery. S26 · core/surface/skills/context-check/SKILL.mdInitialized refresh: read-only report, explicit selection, v1/v2 separation, native paired updates and preserved human history. Full owner read. Selected out
Receipt-only native apply Assistive workflow: source-bound implementation An agreed outcome becomes a maintained plan, direct or delegated implementation, proportional verification and an honest delivery report. Task context and dependency advice support progress; real user stops and host permissions remain effective. Greenfield and brownfield population use the paired native writer and exact finalization. Exact blob bindings identify inspected source bytes; the linked revision is the committed source base. Implementation, verification, live host qualification and release remain separate claims.
Current source changes The pinned native bridge enables task-context as a bounded informational read and does not require future reviewer/approval producers simply to author a structurally valid draft. Context requests route to ordinary context-preview, context-finalize-preview and exact-reference context-apply; historical receipt input stays a separate compatibility path. Installed host qualification is separate. Native Pi writer admission does not establish a complete live Pi workflow. On NO_LOCAL_STORE the client runs init-store once and retries the request once; a refused initialization keeps NO_LOCAL_STORE. On Windows the binary's architecture comes from the kernel (IsWow64Process2), never WMI or PROCESSOR_* environment. Index and outline pages are now fully checked before rows are yielded, including continuation identity and row shape. core/pysrc/_artifactlib.py The closed task-check request accepts bounded command observations matched to plan definitions. Ordinary context-preview and context-finalize-preview omit prompt identity; context-apply accepts an exact preview reference and digest. Invalid requests, changed preimages and stale CAS writes remain rejected. The historical smarts-apply request accepts the 0.2.0 SMARTS decision record. init-store takes a protocol-only request. core/artifacts/internal/operations/protocol.go The installed context-draft command prepares a paired document preview; context-finalize prepares initial finalization. context-apply commits the exact selected preview and digest. Context modes preserve scaffold and initialized refresh boundaries. core/pysrc/init-codearbiter.py Ordinary context-preview retains an inert proposal and exact preimages without approval or review claims. context-apply revalidates the digest, closed request and preview binding before the native paired transaction; context-finalize-preview uses the same exact-reference path for initial population. No approve-context reply is required. core/artifacts/internal/operations/context_assistive.go Ordinary create/adopt/update preserves the same source-evidence, selection and compare-and-swap checks as the native transaction. Missing or matching v1 provenance can be adopted; existing v2 ownership can transition through explicit whole-document selection. Original document/provenance preimages remain in transaction history for exact recovery. core/artifacts/internal/contextdocument/operations.go The lazy reference map routes provenance drift or requested initialized refresh to context-check after reading selected documents, provenance and the code map. The owner reuses selected scope and existing authorization. core/surface/includes/reference-map.md The context-request bridge invokes the qualified installed writer. Source-only builds do not authorize native context writes. Paired per-document atomicity does not mean all documents become visible atomically as one set.
Related entries Pinned evidence S13 · core/surface/skills/context-creation/SKILL.mdCurrent brownfield owner: qualified containment, bounded reports and join, scoped gaps, native document/provenance writes, separate 13-identity finalization and recovery. S26 · core/surface/skills/context-check/SKILL.mdInitialized refresh: read-only report, explicit selection, v1/v2 separation, native paired updates and preserved human history. Full owner read. Selected warn
Preserve other owners and human bytes Assistive workflow: source-bound implementation An agreed outcome becomes a maintained plan, direct or delegated implementation, proportional verification and an honest delivery report. Task context and dependency advice support progress; real user stops and host permissions remain effective. Greenfield and brownfield population use the paired native writer and exact finalization. Exact blob bindings identify inspected source bytes; the linked revision is the committed source base. Implementation, verification, live host qualification and release remain separate claims.
Current source changes The preparation helper binds bounded drafts and retained sources to exact document/provenance preimages. Finalization preparation hashes all thirteen required outputs and retains missing output failures without modifying files. core/pysrc/_contextwriterlib.py The finite context contract permits a whole-document Markdown entry only as the sole entry for one of five named content targets. Explicit typed field entries remain available; the writer does not expose arbitrary repository-file writes. core/artifacts/internal/contextdocument/contract.go Whole-document replacement requires the explicit document selection and exact existing bytes. It preserves CONTEXT frontmatter and initialization state, checks the named document identity and Markdown structure, and retains the code-map size limit. Human content outside the authorized change remains the caller's preservation obligation. core/artifacts/internal/contextdocument/document_selection.go Ordinary create/adopt/update preserves the same source-evidence, selection and compare-and-swap checks as the native transaction. Missing or matching v1 provenance can be adopted; existing v2 ownership can transition through explicit whole-document selection. Original document/provenance preimages remain in transaction history for exact recovery. core/artifacts/internal/contextdocument/operations.go A drift audit stays read-only; an already authorized selected or full refresh proceeds without asking again per document. The installed draft/preview/apply example preserves source bindings and selected content. Matching hashes establish identity only; initialized refresh never reruns scaffolding, clears the marker or repeats finalization. core/surface/skills/context-check/SKILL.md Task / question boards, append-only audit history, release targets and ADRs keep their own writers. No raw Write/Edit fallback on refusal.
Related entries Pinned evidence S13 · core/surface/skills/context-creation/SKILL.mdCurrent brownfield owner: qualified containment, bounded reports and join, scoped gaps, native document/provenance writes, separate 13-identity finalization and recovery. S26 · core/surface/skills/context-check/SKILL.mdInitialized refresh: read-only report, explicit selection, v1/v2 separation, native paired updates and preserved human history. Full owner read. Selected out
Initial finalize: separate approval Assistive workflow: source-bound implementation An agreed outcome becomes a maintained plan, direct or delegated implementation, proportional verification and an honest delivery report. Task context and dependency advice support progress; real user stops and host permissions remain effective. Greenfield and brownfield population use the paired native writer and exact finalization. Exact blob bindings identify inspected source bytes; the linked revision is the committed source base. Implementation, verification, live host qualification and release remain separate claims.
Current source changes The pinned native bridge enables task-context as a bounded informational read and does not require future reviewer/approval producers simply to author a structurally valid draft. Context requests route to ordinary context-preview, context-finalize-preview and exact-reference context-apply; historical receipt input stays a separate compatibility path. Installed host qualification is separate. Native Pi writer admission does not establish a complete live Pi workflow. On NO_LOCAL_STORE the client runs init-store once and retries the request once; a refused initialization keeps NO_LOCAL_STORE. On Windows the binary's architecture comes from the kernel (IsWow64Process2), never WMI or PROCESSOR_* environment. Index and outline pages are now fully checked before rows are yielded, including continuation identity and row shape. core/pysrc/_artifactlib.py The closed task-check request accepts bounded command observations matched to plan definitions. Ordinary context-preview and context-finalize-preview omit prompt identity; context-apply accepts an exact preview reference and digest. Invalid requests, changed preimages and stale CAS writes remain rejected. The historical smarts-apply request accepts the 0.2.0 SMARTS decision record. init-store takes a protocol-only request. core/artifacts/internal/operations/protocol.go Ordinary context-preview retains an inert proposal and exact preimages without approval or review claims. context-apply revalidates the digest, closed request and preview binding before the native paired transaction; context-finalize-preview uses the same exact-reference path for initial population. No approve-context reply is required. core/artifacts/internal/operations/context_assistive.go Initial finalization binds thirteen output identities, rechecks supporting source evidence, and changes the initialization marker with its paired provenance digest. Ordinary finalization requires no artifact approval receipt and does not certify semantic correctness. Already initialized repositories refuse a new finalization. core/artifacts/internal/contextdocument/finalization.go Brownfield population covers six logical categories and five content documents through direct bounded inspection or optional scoped agents. The paired writer preserves source bindings and protected content; initial finalization checks all required outputs. core/surface/skills/context-creation/SKILL.md The finalized set includes exactly the five content documents, their five v2 provenance companions, open-tasks.md, open-questions.md and overrides.log. Valid empty boards and a zero-event audit header satisfy their actual contracts. A document-write receipt never substitutes for finalization.
Related entries Pinned evidence S13 · core/surface/skills/context-creation/SKILL.mdCurrent brownfield owner: qualified containment, bounded reports and join, scoped gaps, native document/provenance writes, separate 13-identity finalization and recovery. Selected out
Refresh: retain initialized state Assistive workflow: source-bound implementation An agreed outcome becomes a maintained plan, direct or delegated implementation, proportional verification and an honest delivery report. Task context and dependency advice support progress; real user stops and host permissions remain effective. Greenfield and brownfield population use the paired native writer and exact finalization. Exact blob bindings identify inspected source bytes; the linked revision is the committed source base. Implementation, verification, live host qualification and release remain separate claims.
Current source changes The installed context-draft command prepares a paired document preview; context-finalize prepares initial finalization. context-apply commits the exact selected preview and digest. Context modes preserve scaffold and initialized refresh boundaries. core/pysrc/init-codearbiter.py Whole-document replacement requires the explicit document selection and exact existing bytes. It preserves CONTEXT frontmatter and initialization state, checks the named document identity and Markdown structure, and retains the code-map size limit. Human content outside the authorized change remains the caller's preservation obligation. core/artifacts/internal/contextdocument/document_selection.go Initial finalization binds thirteen output identities, rechecks supporting source evidence, and changes the initialization marker with its paired provenance digest. Ordinary finalization requires no artifact approval receipt and does not certify semantic correctness. Already initialized repositories refuse a new finalization. core/artifacts/internal/contextdocument/finalization.go A drift audit stays read-only; an already authorized selected or full refresh proceeds without asking again per document. The installed draft/preview/apply example preserves source bindings and selected content. Matching hashes establish identity only; initialized refresh never reruns scaffolding, clears the marker or repeats finalization. core/surface/skills/context-check/SKILL.md Initial brownfield setup checks the intended repository and actual host permissions, then uses bounded inspection and the installed safe context writer. Ordinary Windows work needs no special scout profile or per-document artifact approval. An existing initialized repository routes to selected refresh instead of another scaffold or population cycle. core/surface/commands/init.md Routine maintenance selects changed v2 claims intersecting staged sources using a current bounded snapshot. Unknown evidence defers the affected document; native paired writes preserve owned fields and current identities. core/pysrc/_provenancelib.py Only selected pairs change Declare scoped versus full coverage Keep marker, human content and history No finalization, staging or commit here
Related entries Pinned evidence S26 · core/surface/skills/context-check/SKILL.mdInitialized refresh: read-only report, explicit selection, v1/v2 separation, native paired updates and preserved human history. Full owner read. Selected skill
Recover by the same operation ID Assistive workflow: source-bound implementation An agreed outcome becomes a maintained plan, direct or delegated implementation, proportional verification and an honest delivery report. Task context and dependency advice support progress; real user stops and host permissions remain effective. Greenfield and brownfield population use the paired native writer and exact finalization. Exact blob bindings identify inspected source bytes; the linked revision is the committed source base. Implementation, verification, live host qualification and release remain separate claims.
Current source changes Ordinary context-preview retains an inert proposal and exact preimages without approval or review claims. context-apply revalidates the digest, closed request and preview binding before the native paired transaction; context-finalize-preview uses the same exact-reference path for initial population. No approve-context reply is required. core/artifacts/internal/operations/context_assistive.go Ordinary create/adopt/update preserves the same source-evidence, selection and compare-and-swap checks as the native transaction. Missing or matching v1 provenance can be adopted; existing v2 ownership can transition through explicit whole-document selection. Original document/provenance preimages remain in transaction history for exact recovery. core/artifacts/internal/contextdocument/operations.go Initial finalization binds thirteen output identities, rechecks supporting source evidence, and changes the initialization marker with its paired provenance digest. Ordinary finalization requires no artifact approval receipt and does not certify semantic correctness. Already initialized repositories refuse a new finalization. core/artifacts/internal/contextdocument/finalization.go Do not reset the project directory, clear the marker, overwrite a human file, or retry an already committed write. Finalization recovery is checked separately. A genuinely absent output needs a fresh typed preview and its own approval.
Related entries Pinned evidence S13 · core/surface/skills/context-creation/SKILL.mdCurrent brownfield owner: qualified containment, bounded reports and join, scoped gaps, native document/provenance writes, separate 13-identity finalization and recovery. Handoffs, branches and returns Debug · evidence, dispositions and safe handoff Investigation is read-only. A validated packet carries evidence; the original caller still owns permission to act.
Handoffs, branches and returns Tribunal: source-bound audit to completed follow-up An opt-in review route, not a required gate. Candidate findings, verification, report and completion are distinct.
Handoffs, branches and returns
Pinned source ledger Close Historical route snapshot at 9496cff6332fe0b97195b7ebe67b73f62f30e5b3. Current assistive source bindings identify inspected implementation bytes, with 3549ed56f5d5849116aef8d5e946423d8ee677fc as the committed source base. Live host qualification and release are separate evidence. Curated current-source workflow contracts and selected context-delivery implementation; not a release, installed-host qualification, or exhaustive runtime call graph. Original source links and spatial positions retained; task labels show the selected work owner. Source compatibility revalidated at 25980c292b529fca90c39f5f26d69a1c0674de44. Unittest collector compatibility bound to reviewed blob 76a2998f4bd5b6f441f5fd162bbe810b2de2ee93.
codeArbiter / route atlas A command selects an owner. The owner loads context, branches through gates, and returns a bounded result. An agreed outcome becomes a maintained plan, direct or delegated implementation, proportional verification and an honest delivery report. Task context and dependency advice support progress; real user stops and host permissions remain effective. Greenfield and brownfield population use the paired native writer and exact finalization. Exact blob bindings identify inspected source bytes; the linked revision is the committed source base. Implementation, verification, live host qualification and release remain separate claims. Source 9496cff6332fe0b97195b7ebe67b73f62f30e5b3. Select a node for details or use the complete text path. HISTORICAL ROUTES / CODEARBITER codeArbiter / route atlas A command selects an owner. The owner loads context, branches through gates, and returns a bounded result. SOURCE 9496cff6332f / 2026-10-07 01 / ESTABLISH KNOWLEDGE Project initialization feeds later workflows; it does not implement the product. 02 / CREATE A VERIFIED CHANGE Solid arrows are execution. Shared task execution is not a promise that every lane creates a spec artifact. 03 / DELIVER UNDER SEPARATE AUTHORITY A local commit, an open PR, a merge and a release are different outcomes. 04 / INVESTIGATE OR CHALLENGE These routes return findings, not automatic implementation or delivery. 05 / PRESERVE INTENT Decisions govern later work; accepted intent is not implementation evidence. 06 / OPERATE, REPAIR CONTEXT AND FIND HELP Supporting routes below are independent entry points. Their outputs feed a later invocation only when relevant. Execution Execution Execution Execution Context / data: inputs inputs Execution: full full Conditional / authority: small small Execution Execution Execution Execution Execution Execution Execution Execution: fresh lane proof fresh lane proof Execution Execution: authorized adoption / non-behavioral work authorized adoption / non-behavioral work Conditional / authority: request request Conditional / authority: selected selected Conditional / authority: request request Conditional / authority: verified work → commit only when authorized verified work → commit only when authorized /init
--check: report only
Otherwise scaffold, then populate
Sources: S23 /init --check: report only Otherwise scaffold, then populate S23 decompose
Greenfield: six persisted interview layers
Sources: S14 decompose Greenfield: six persisted interview layers S14 context-creation
Qualified scouts → bounded reports
Native paired writes + finalization
Sources: S13, S26 context-creation Qualified scouts → bounded reports Native paired writes + finalization S13 Initialized project knowledge
CONTEXT • stack • standards • security
ADRs • plans • questions • code-map
Provenance supports freshness checks
Sources: S13, S14, S23 Initialized project knowledge CONTEXT • stack • standards • security ADRs • plans • questions • code-map Provenance supports freshness checks S13 Runtime + on-demand context
Startup state; safety-core + active persona
On-demand rules + complete typed task reads
Task-scoped packets delivered to each actual actor
Context is not approval, admission or consumption proof
Sources: S27, S15, S16, S10 CONTEXT / NOT A BULK READ Runtime + on-demand context Startup state; safety-core + active persona On-demand rules + complete typed task reads Task-scoped packets delivered to each actual actor Context is not approval, admission or consumption proof S27 /feature
New behavior; attended workflow
Sources: S06 /feature New behavior; attended workflow S06 /sprint
Coherent goal; approved autonomy
Sources: S07 /sprint Coherent goal; approved autonomy S07 /fix · /refactor
Regression proof / behavior parity
Sources: S19, S21 /fix · /refactor Regression proof / behavior parity S19 /chore · /add-dep
Type-scaled work / vetted adoption
Sources: S20, S04 /chore · /add-dep Type-scaled work / vetted adoption S20 brainstorming → writing-plans
Full: approved spec, then bound plan
Sources: S06, S10 brainstorming → writing-plans Full: approved spec, then bound plan S06 Spec + plan pair approval
One real user reply before execution
Sources: S07, S10 Spec + plan pair approval One real user reply before execution S07 tdd / parity owner
Small feature: confirmed inline mini-spec
Sources: S06, S09, S19, S21 tdd / parity owner Small feature: confirmed inline mini-spec S06 Type / dependency gates
Prose: no TDD; behavior: feature or fix
Sources: S20, S04 Type / dependency gates Prose: no TDD; behavior: feature or fix S20 executing-plans
Scoped batches + human checkpoints
Sources: S06, S21 executing-plans Scoped batches + human checkpoints S06 Full-plan handoff
No attended batch pauses; hard stops stay
Sources: S07 Full-plan handoff No attended batch pauses; hard stops stay S07 Task engine
Fresh actor context → TDD
Spec proof + fresh verification
Combined-scope quality review
Accept only on current evidence
Sources: S08, S27, S09 SUBAGENT-DRIVEN DEVELOPMENT Task engine Fresh actor context → TDD Spec proof + fresh verification Combined-scope quality review Accept only on current evidence S08 Verified change
Current acceptance for typed plans
Regression / parity / type proof
Persistence still needs authority
Sources: S08, S11, S19, S20, S21 Verified change Current acceptance for typed plans Regression / parity / type proof Persistence still needs authority S08 /commit → commit-gate
Explicit permission • nine phases
Focused proof • selective stage
Output: local commit SHA
Sources: S11 /commit → commit-gate Explicit permission • nine phases Focused proof • selective stage Output: local commit SHA S11 /pr → finishing owner
Current-head proof + diff review
Open PR; sprint ends here
Feature retains terminal choice
Sources: S12 /pr → finishing owner Current-head proof + diff review Open PR; sprint ends here Feature retains terminal choice S12 Merge through the PR
Explicit user authority
Exact-head hosted CI + ancestry
No direct-to-default delivery
Sources: S12 Merge through the PR Explicit user authority Exact-head hosted CI + ancestry No direct-to-default delivery S12 /release
Release preparation PR → STOP
Qualified hosted publisher
Separate publication authority
Sources: S33, S34 /release Release preparation PR → STOP Qualified hosted publisher Separate publication authority S33 /watch · /cleanup
/pr --watch / --cleanup
Early branches, not PR creation
No automatic merge or discard
Sources: S02, S04, S12 /watch · /cleanup /pr --watch / --cleanup Early branches, not PR creation No automatic merge or discard S02 /debug · /spike
Debug: five evidenced outcomes
Validated handoff ≠ repair consent
Spike: findings or feature
Sources: S25, S04 /debug · /spike Debug: five evidenced outcomes Validated handoff ≠ repair consent Spike: findings or feature S25 /preview · /review
Predict fleet / review actual diff
/checkpoint persists a sweep
Review verdict is not a commit
Sources: S04, S18, S22 /preview · /review Predict fleet / review actual diff /checkpoint persists a sweep Review verdict is not a commit S04 /threat-model · /tribunal
Distinct opt-in review owners
Tribunal: thirteen lens cards
No routine merge gate
Sources: S29, S04 /threat-model · /tribunal Distinct opt-in review owners Tribunal: thirteen lens cards No routine merge gate S29 /adr · /adr-status
decision-lifecycle
Attributed decision / health report
Sources: S04, S21 /adr · /adr-status decision-lifecycle Attributed decision / health report S04 /reconcile
decision-variance
scout → grader → user ruling
Report-only writes nothing
Sources: S04, S21, S22 /reconcile decision-variance scout → grader → user ruling Report-only writes nothing S04 /conflict
Internal protocol
Contradicting rules: STOP
User resolves the authority gap
Sources: S03, S04 /conflict Internal protocol Contradicting rules: STOP User resolves the authority gap S03 /status · /context-check
v1/v2 state and drift report
Explicit scoped/full refresh selection
Native v2 pairs; preserve history
Sources: S26, S13 /status · /context-check v1/v2 state and drift report Explicit scoped/full refresh selection Native v2 pairs; preserve history S26 /task · /standup
Sanctioned task-board transitions
Hygiene with per-action consent
Sources: S02, S04 /task · /standup Sanctioned task-board transitions Hygiene with per-action consent S02 /audit · /metrics
Dated source-record packet
Windowed governance trends
Sources: S02, S04 /audit · /metrics Dated source-record packet Windowed governance trends S02 /doctor · /statusline
Install / payload diagnostics
Statusline is Claude-specific
Sources: S02, S10 /doctor · /statusline Install / payload diagnostics Statusline is Claude-specific S02 /override · /prune
Permitted, logged exception
Host-specific context compaction
Sources: S02, S03, S04 /override · /prune Permitted, logged exception Host-specific context compaction S02 Watch/cleanup branch before creation preflight. Commit alone does not imply push. Release is not an automatic next step. /commands → grouped catalog /btw → deprecated Q&A /create-context → /init --brownfield /decompose → /init --greenfield Mode tokens are not slash commands: mode --arbiter / --ops / --dangerous. Safety-core remains the floor; gates shown here are arbiter-mode routes. 37 catalog entries accounted for. Canonical, advanced, alias, internal and deprecated visibility is preserved in the selectable HTML explorer. Capability boundaries: brownfield scouts and native writers require qualification; Pi new-HTML admission and HTML farm remain blocked. Read the map
codeArbiter / route atlas A command selects an owner. The owner loads context, branches through gates, and returns a bounded result.
Execution Context / data Conditional / authority Return / repeat Stop / refusal
Choose an entry to highlight its route. Select a node for context, related entries and pinned sources.
Start with a route /feature /fix /sprint /init /review /tribunal /pr /reconcile
Historical route snapshot at 9496cff6332fe0b97195b7ebe67b73f62f30e5b3. Current assistive source bindings identify inspected implementation bytes, with 3549ed56f5d5849116aef8d5e946423d8ee677fc as the committed source base. Live host qualification and release are separate evidence. Curated current-source workflow contracts and selected context-delivery implementation; not a release, installed-host qualification, or exhaustive runtime call graph. Original source links and spatial positions retained; task labels show the selected work owner. Source compatibility revalidated at 25980c292b529fca90c39f5f26d69a1c0674de44. Unittest collector compatibility bound to reviewed blob 76a2998f4bd5b6f441f5fd162bbe810b2de2ee93.
Snapshot limits Historical snapshot: New full-lane feature/sprint defaults to typed HTML only with an admitted installed payload and qualified producers. Main source checkout does not carry the release binary.
Historical snapshot: Pi fails new HTML workflow admission in the pinned artifacts contract. Existing Markdown pairs and inline small features remain separate.
Historical snapshot: HTML --farm dispatch is blocked. Legacy Markdown farm is an off-by-default preview, not the blessed default.
Historical snapshot: Read-time approved-spec pointers index Markdown only in the inspected implementation. Typed HTML execution reads are a distinct engine path.
Historical snapshot: Shared surface generation is not live host parity. Claude/Pi READ notices and Codex’s absence of a native READ event differ.
Historical snapshot: Task quality review is once per current combined scope after per-task spec proof and fresh verification, not a duplicate full fleet for every tiny task.
Historical snapshot: The diagrams describe default arbiter-mode workflow contracts; deterministic mode tokens are separate posture controls and not new slash commands.
Historical snapshot: Snapshots of narrative instructions do not prove every gate is mechanically enforced. No live workflows were executed for this atlas.
Historical snapshot: Brownfield context dispatch needs an actually qualified read-only caller profile; the generic scout charter and generated plugin surfaces alone do not qualify it. Repository-only preparation fixtures are not installed capability.
Historical snapshot: Five content/provenance pairs use independently approved native transactions. Initial finalization has its own 13-identity approval. Initialized refresh retains the marker and never re-enters finalization.
Historical snapshot: Task-context selection is inert. The caller delivers a separate bounded input to each actor; reported delivery, engine tickets and parent session markers are not interchangeable authority or proof.
Historical snapshot: Debug has five dispositions. Validation alone is not evidence truth or repair authority. Diagnosis-only requests stop after the recommendation.
Historical snapshot: Source wording ambiguity: brownfield Phase 3 preserves source-scoped upstream ADR references, while a retained Phase 5 sentence asks for local summaries. Local-summary semantics and authority need clarification; this atlas does not infer automatic local decision authority.
Historical snapshot: Pi approval is currently DEGRADED because the adapter is not implemented; the newer runtime input event means the older HOST-IMPOSSIBLE explanation must not be repeated.
Historical snapshot: Tribunal is opt-in. Its source-bound run store is separate from typed spec/plan authority; report-written does not establish terminal completion.
Historical snapshot: Ten new commits include correctness and waiting-limit repairs, not ten new routes. Their merge descriptions are not independent runtime proof. No latency improvement is asserted.
Historical snapshot: The release PR boundary was under-specified in the previous atlas; making it explicit is a carried presentation correction, not a behavior newly introduced by these ten commits.
Historical snapshot: The native page and offline editions share this source-bound atlas. Website integration does not establish released behavior, installed-host qualification or completed workflow execution.
Feature + sprint / execution paths Follow the split at approval, the join at task execution, and the split again when control returns to the caller. An agreed outcome becomes a maintained plan, direct or delegated implementation, proportional verification and an honest delivery report. Task context and dependency advice support progress; real user stops and host permissions remain effective. Greenfield and brownfield population use the paired native writer and exact finalization. Exact blob bindings identify inspected source bytes; the linked revision is the committed source base. Implementation, verification, live host qualification and release remain separate claims. Source 9496cff6332fe0b97195b7ebe67b73f62f30e5b3. Select a node for details or use the complete text path. HISTORICAL ROUTES / CODEARBITER Feature + sprint / execution paths Follow the split at approval, the join at task execution, and the split again when control returns to the caller. SOURCE 9496cff6332f / 2026-10-07 01 / SELECT AUTHORITY BEFORE EXECUTION An existing pipeline resumes from its exact format. New full-lane work defaults to HTML, never a silent Markdown fallback. 02 / CHOOSE THE COORDINATOR The approved caller, not the author agent, controls scope and pacing. 03 / ONE TASK, ONE FRESH AUTHOR CONTEXT The actual order is per-task spec review + fresh verification, then combined quality once the current scope is ready. 04 / COMBINED-SCOPE ACCEPTANCE AND CALLER RETURN Per-task progress cannot bypass this barrier. Direct fix/refactor and small-feature lanes keep their own TDD close. Execution: existing existing Conditional / authority: new / draft new / draft Execution Execution Conditional / authority: new small feature new small feature Execution Execution Execution Execution Execution Execution: record task proof, not final acceptance record task proof, not final acceptance Execution Return / repeat: repeat repeat Execution Return / repeat: scoped scoped Execution: plan done plan done Execution Execution: approved attended scope approved attended scope Execution: approved full plan approved full plan Execution: complete task brief + task-start complete task brief + task-start Conditional / authority: all tasks in current scope have spec proof + fresh verification all tasks in current scope have spec proof + fresh verification Conditional / authority: failed proof failed proof /feature or /sprint
Resolve same-slug artifacts
Mixed / duplicate matches: STOP
Exact pair or lone spec keeps format
Sources: S06, S07, S10 /feature or /sprint Resolve same-slug artifacts Mixed / duplicate matches: STOP Exact pair or lone spec keeps format S06 Resume checkpoint
Draft → approval
Approved spec → planning
Remaining tasks → execution
All accepted → current-proof gate
Sources: S06, S07 Resume checkpoint Draft → approval Approved spec → planning Remaining tasks → execution All accepted → current-proof gate S06 brainstorming
Recorded intent + user criteria
Unknowns → open-questions
No guessed product decisions
Sources: S06, S07, S10 brainstorming Recorded intent + user criteria Unknowns → open-questions No guessed product decisions S06 writing-plans
Exact criterion IDs → tasks
Paths • tests • dependencies
Checkpoints • prerequisites
Sources: S06, S07, S10, S21 writing-plans Exact criterion IDs → tasks Paths • tests • dependencies Checkpoints • prerequisites S06 Approval is observed
Feature: sequential spec → plan
Sprint: one initial pair reply
No execution from draft previews
Sources: S10, S31 Approval is observed Feature: sequential spec → plan Sprint: one initial pair reply No execution from draft previews S10 Small /feature branch
All small-lane conditions → confirm inline criteria → log triage → direct TDD. No spec artifact.
Sources: S06 Small /feature branch All small-lane conditions → confirm inline criteria → log triage → direct TDD. No spec artifact. S06 Installed admission is required
Pi HTML admission unavailable; legacy Markdown and inline small lane stay distinct. HTML farm remains blocked.
Sources: S10 Installed admission is required Pi HTML admission unavailable; legacy Markdown and inline small lane stay distinct. HTML farm remains blocked. S10 executing-plans / attended
/feature: authored HTML checkpoints or bounded legacy batches
Pass selected task IDs; wait for the user between batches
Sources: S06, S21, S08 executing-plans / attended /feature: authored HTML checkpoints or bounded legacy batches Pass selected task IDs; wait for the user between batches S06 SPRINT.md / approved autonomy
/sprint: pass the full plan; no attended batch pauses
SMARTS logs choices; delegated methods cannot widen scope
Sources: S07 SPRINT.md / approved autonomy /sprint: pass the full plan; no attended batch pauses SMARTS logs choices; delegated methods cannot widen scope S07 subagent-driven-development
Resolve current identities → eligible tasks only
HTML: complete contextual read ticket → task-start
Then separately compose and deliver each actor’s scoped input
Legacy Markdown keeps its own dependency/state rules
Sources: S08, S27, S10 subagent-driven-development Resolve current identities → eligible tasks only HTML: complete contextual read ticket → task-start Then separately compose and deliver each actor’s scoped input Legacy Markdown keeps its own dependency/state rules S08 Fresh scoped author input
Map + commands + constraints
Native instructions; actual actor delivery
No inherited prior-task narrative
Sources: S08, S27 Fresh scoped author input Map + commands + constraints Native instructions; actual actor delivery No inherited prior-task narrative S08 tdd / six gates
1 Obligations → 2 Red → 3 Green
4 Obligation verify → 5 Coverage → 6 Lint
Missing proof loops back to the failing gate
Coverage-auditor where contract/security critical
Sources: S09, S21 tdd / six gates 1 Obligations → 2 Red → 3 Green 4 Obligation verify → 5 Coverage → 6 Lint Missing proof loops back to the failing gate Coverage-auditor where contract/security critical S09 Spec-compliance review
Fresh independent reviewer input
Exact task obligation + current source
Failure → corrective author brief
Sources: S08, S32, S40 Spec-compliance review Fresh independent reviewer input Exact task obligation + current source Failure → corrective author brief S08 Fresh task verification
Run the plan’s proving command again
Read output and exit code
HTML: real producer-owned receipts
Sources: S08, S10 Fresh task verification Run the plan’s proving command again Read output and exit code HTML: real producer-owned receipts S08 Task proof / provisional review
HTML task-review binds spec-review + verification receipts. REVIEW is not ACCEPTED.
Sources: S08 Task proof / provisional review HTML task-review binds spec-review + verification receipts. REVIEW is not ACCEPTED. S08 Next eligible task in the same scope
Repeat fresh author → spec proof → fresh verification. No context accumulation across task authors.
Sources: S08 Next eligible task in the same scope Repeat fresh author → spec proof → fresh verification. No context accumulation across task authors. S08 Correction is not permission to waive a gate
Approved sprint: diagnose, reconcile, retry with fresh proof.
Same failure twice without new evidence: stop that strategy.
Authority / security / CONFIRM gaps still stop affected work.
Sources: S08, S10, S40 Correction is not permission to waive a gate Approved sprint: diagnose, reconcile, retry with fresh proof. Same failure twice without new evidence: stop that strategy. Authority / security / CONFIRM gaps still stop affected work. S08 Combined quality review
Once per authored scope / checkpoint
Actual combined diff + fresh reviewer inputs
No CRITICAL/HIGH; rerun after corrections
Scope of one uses the same rule
Sources: S08, S27, S18 Combined quality review Once per authored scope / checkpoint Actual combined diff + fresh reviewer inputs No CRITICAL/HIGH; rerun after corrections Scope of one uses the same rule S08 accept-scope / current proof
HTML: one whole-scope operation
All required current receipts must agree
Markdown: authoritative ledger transition
No rendered labels as authority
Sources: S08, S10 accept-scope / current proof HTML: one whole-scope operation All required current receipts must agree Markdown: authoritative ledger transition No rendered labels as authority S08 Attended return
Batch complete → executing-plans
User checkpoint before next batch
Sources: S08, S21 Attended return Batch complete → executing-plans User checkpoint before next batch S08 Full plan complete
Current acceptance preflight
Commit owner retains permission gate
Sources: S08, S11 Full plan complete Current acceptance preflight Commit owner retains permission gate S08 Authorized delivery
commit-gate → finishing owner
Sprint: open PR only
Feature: selected terminal disposition
No automatic merge
Sources: S11, S12 Authorized delivery commit-gate → finishing owner Sprint: open PR only Feature: selected terminal disposition No automatic merge S11 Optional branches: using-git-worktrees isolates opted-in units; legacy --farm replaces authoring, never review. Neither is the default. Reselect actor context after resume, compaction or worktree/scope change. An engine ticket or parent persona marker is not a child delivery receipt. Read the map
Feature + sprint / execution paths Follow the split at approval, the join at task execution, and the split again when control returns to the caller.
Execution Context / data Conditional / authority Return / repeat Stop / refusal
Choose an entry to highlight its route. Select a node for context, related entries and pinned sources.
Start with a route /feature /fix /sprint /init /review /tribunal /pr /reconcile
Historical route snapshot at 9496cff6332fe0b97195b7ebe67b73f62f30e5b3. Current assistive source bindings identify inspected implementation bytes, with 3549ed56f5d5849116aef8d5e946423d8ee677fc as the committed source base. Live host qualification and release are separate evidence. Curated current-source workflow contracts and selected context-delivery implementation; not a release, installed-host qualification, or exhaustive runtime call graph. Original source links and spatial positions retained; task labels show the selected work owner. Source compatibility revalidated at 25980c292b529fca90c39f5f26d69a1c0674de44. Unittest collector compatibility bound to reviewed blob 76a2998f4bd5b6f441f5fd162bbe810b2de2ee93.
Snapshot limits Historical snapshot: New full-lane feature/sprint defaults to typed HTML only with an admitted installed payload and qualified producers. Main source checkout does not carry the release binary.
Historical snapshot: Pi fails new HTML workflow admission in the pinned artifacts contract. Existing Markdown pairs and inline small features remain separate.
Historical snapshot: HTML --farm dispatch is blocked. Legacy Markdown farm is an off-by-default preview, not the blessed default.
Historical snapshot: Read-time approved-spec pointers index Markdown only in the inspected implementation. Typed HTML execution reads are a distinct engine path.
Historical snapshot: Shared surface generation is not live host parity. Claude/Pi READ notices and Codex’s absence of a native READ event differ.
Historical snapshot: Task quality review is once per current combined scope after per-task spec proof and fresh verification, not a duplicate full fleet for every tiny task.
Historical snapshot: The diagrams describe default arbiter-mode workflow contracts; deterministic mode tokens are separate posture controls and not new slash commands.
Historical snapshot: Snapshots of narrative instructions do not prove every gate is mechanically enforced. No live workflows were executed for this atlas.
Historical snapshot: Brownfield context dispatch needs an actually qualified read-only caller profile; the generic scout charter and generated plugin surfaces alone do not qualify it. Repository-only preparation fixtures are not installed capability.
Historical snapshot: Five content/provenance pairs use independently approved native transactions. Initial finalization has its own 13-identity approval. Initialized refresh retains the marker and never re-enters finalization.
Historical snapshot: Task-context selection is inert. The caller delivers a separate bounded input to each actor; reported delivery, engine tickets and parent session markers are not interchangeable authority or proof.
Historical snapshot: Debug has five dispositions. Validation alone is not evidence truth or repair authority. Diagnosis-only requests stop after the recommendation.
Historical snapshot: Source wording ambiguity: brownfield Phase 3 preserves source-scoped upstream ADR references, while a retained Phase 5 sentence asks for local summaries. Local-summary semantics and authority need clarification; this atlas does not infer automatic local decision authority.
Historical snapshot: Pi approval is currently DEGRADED because the adapter is not implemented; the newer runtime input event means the older HOST-IMPOSSIBLE explanation must not be repeated.
Historical snapshot: Tribunal is opt-in. Its source-bound run store is separate from typed spec/plan authority; report-written does not establish terminal completion.
Historical snapshot: Ten new commits include correctness and waiting-limit repairs, not ten new routes. Their merge descriptions are not independent runtime proof. No latency improvement is asserted.
Historical snapshot: The release PR boundary was under-specified in the previous atlas; making it explicit is a carried presentation correction, not a behavior newly introduced by these ten commits.
Historical snapshot: The native page and offline editions share this source-bound atlas. Website integration does not establish released behavior, installed-host qualification or completed workflow execution.
Context / how the layers are assembled Durable knowledge → selective task input → actual actor delivery → independent proof. These are different boundaries. An agreed outcome becomes a maintained plan, direct or delegated implementation, proportional verification and an honest delivery report. Task context and dependency advice support progress; real user stops and host permissions remain effective. Greenfield and brownfield population use the paired native writer and exact finalization. Exact blob bindings identify inspected source bytes; the linked revision is the committed source base. Implementation, verification, live host qualification and release remain separate claims. Source 9496cff6332fe0b97195b7ebe67b73f62f30e5b3. Select a node for details or use the complete text path. HISTORICAL ROUTES / CODEARBITER Context / how the layers are assembled Durable knowledge → selective task input → actual actor delivery → independent proof. These are different boundaries. SOURCE 9496cff6332f / 2026-10-07 01 / PRODUCE DURABLE PROJECT KNOWLEDGE The durable project store survives a new host, a fresh author context or a compacted conversation. 02 / RESIDENT STATE + SELECTIVE POLICY Only the selected mechanism contributes; no universal concatenation of every document is implied. 03 / BIND APPROVED INTENT TO THE EXACT TASK The HTML engine controls artifact identity, scope and contextual-read completeness; it is not a new public command. 04 / SELECT AND DELIVER TO THE ACTOR Actor-specific delivery boundary: pure selection prepares bounded input; the caller must actually deliver it. Neither step grants authority. 05 / KEEP ROLE CONTEXTS SEPARATE Only the bounded result crosses a role boundary. No author’s narrative can stand in for independent proof. 06 / FEEDBACK INTO DURABLE KNOWLEDGE Context changes only through its owner. A later read selects what is relevant again. Context / data: approved synthesis approved synthesis Context / data: approved approved Context / data: binding binding Context / data Context / data: task-start task-start Execution: result result Context / data: evidence evidence Execution Context / data: live project state live project state Context / data: governing records governing records Context / data Context / data: selected rules + recorded intent inform specification selected rules + recorded intent inform specification Context / data Execution Conditional / authority Return / repeat: changed binding → select again changed binding → select again Context / data Context / data Context / data Context / data Greenfield / decompose
Vision → users → scope → technical shape
Integrations → risks, each persisted
Architecture / phased plan / backlog
User-reviewed decisions and context
Sources: S14 Greenfield / decompose Vision → users → scope → technical shape Integrations → risks, each persisted Architecture / phased plan / backlog User-reviewed decisions and context S14 Brownfield / context-creation
Qualified read-only scout profile first
Six full / affected scoped assignments
Bounded report-only synthesis
Native content + provenance transactions
Sources: S13, S26 Brownfield / context-creation Qualified read-only scout profile first Six full / affected scoped assignments Bounded report-only synthesis Native content + provenance transactions S13 .codearbiter/ project knowledge
CONTEXT: purpose, vocabulary, stage and limits
Stack: scoped test/build commands and author mapping
Standards + security: rules with preserved human ownership
Brownfield: five content docs + their v2 provenance pairs
ADRs, boards and greenfield plans retain separate owners
Sources: S13, S14, S26 .codearbiter/ project knowledge CONTEXT: purpose, vocabulary, stage and limits Stack: scoped test/build commands and author mapping Standards + security: rules with preserved human ownership Brownfield: five content docs + their v2 provenance pairs ADRs, boards and greenfield plans retain separate owners S13 L1 / startup state
Enabled repo: stage, questions, tasks,
overrides and hygiene briefing
SessionStart carries live state,
not the composed persona
Sources: S17 L1 / startup state Enabled repo: stage, questions, tasks, overrides and hygiene briefing SessionStart carries live state, not the composed persona S17 L2 / composed persona
safety-core + current mode body
Arbiter selects existing route owners
Dedup: session × mode × compaction
Mode tokens are a separate control path
Sources: S03, S15 L2 / composed persona safety-core + current mode body Arbiter selects existing route owners Dedup: session × mode × compaction Mode tokens are a separate control path S03 L3 / owner + scope rules
Read the invoked command/skill body
Scope touch → reference-map docs
Read relevant ADR/deferral bodies
Indexes first; no bulk directory reads
Sources: S03, S04, S05 L3 / owner + scope rules Read the invoked command/skill body Scope touch → reference-map docs Read relevant ADR/deferral bodies Indexes first; no bulk directory reads S03 L3a / read-time hints
A bounded pointer, not a full document
security > ADRs > specs > standards
Freshness and dedup are advisory
Host event surfaces differ
Sources: S16, S17, S24 L3a / read-time hints A bounded pointer, not a full document security > ADRs > specs > standards Freshness and dedup are advisory Host event surfaces differ S16 L4 / approved specification
Stable artifact and criterion IDs
Behavior • boundaries • non-goals
Contracts and accepted intent
Normative change invalidates approval
Sources: S10, S06, S07 L4 / approved specification Stable artifact and criterion IDs Behavior • boundaries • non-goals Contracts and accepted intent Normative change invalidates approval S10 L5 / exactly bound plan
Same-format spec binding
Task paths, tests, dependencies
Prerequisites and checkpoint membership
No shadow status ledger
Sources: S10, S32, S40 L5 / exactly bound plan Same-format spec binding Task paths, tests, dependencies Prerequisites and checkpoint membership No shadow status ledger S10 L6 / complete task-context read
identity → index → outline → scoped read
Follow contextual pages to completion
Retain the final engine context ticket
Exact read alone is context-incomplete
Stale cursor or ticket → read again
Sources: S08, S10 L6 / complete task-context read identity → index → outline → scoped read Follow contextual pages to completion Retain the final engine context ticket Exact read alone is context-incomplete Stale cursor or ticket → read again S08 Lane-specific execution brief
Feature: exact spec/plan/task identities
Small feature: confirmed inline criteria
Fix: cited defect + named regression
Test-only: retain the existing caller
Sources: S09, S19, S27 Lane-specific execution brief Feature: exact spec/plan/task identities Small feature: confirmed inline criteria Fix: cited defect + named regression Test-only: retain the existing caller S09 Not every lane needs this artifact stack
Fix uses supported bug-origin evidence; refactor uses existing parity obligations. Neither fabricates a feature spec/plan to get context.
Sources: S09, S19 Not every lane needs this artifact stack Fix uses supported bug-origin evidence; refactor uses existing parity obligations. Neither fabricates a feature spec/plan to get context. S09 A context ticket is neither a model memory nor a signature of semantic truth
Typed engine context tickets bind complete task reads. They do not establish actual actor delivery, semantic truth or fresh runtime evidence.
Sources: S10, S27 A context ticket is neither a model memory nor a signature of semantic truth Typed engine context tickets bind complete task reads. They do not establish actual actor delivery, semantic truth or fresh runtime evidence. S10 Author context
Fresh task obligation + delivered scoped input
Check worktree, source and constraints
TDD retains the original caller’s gates
Sources: S08, S09, S27 Author context Fresh task obligation + delivered scoped input Check worktree, source and constraints TDD retains the original caller’s gates S08 Reviewer context
Own fresh packet; never reuse author input
Required refs + independent source checks
Reviewer delivery is separately tracked
Sources: S08, S28, S27 Reviewer context Own fresh packet; never reuse author input Required refs + independent source checks Reviewer delivery is separately tracked S08 Runner + authority producers
Fresh command result / observed reply
Bound to exact current target
Engine-written receipts, not pasted claims
Sources: S10, S32, S40 Runner + authority producers Fresh command result / observed reply Bound to exact current target Engine-written receipts, not pasted claims S10 L7 / acceptance evidence
task-review → combined quality
accept-scope → current acceptance
Commit and finish query proof again
Sources: S08, S10, S11, S12 L7 / acceptance evidence task-review → combined quality accept-scope → current acceptance Commit and finish query proof again S08 Legacy v1 heal / native v2 refresh
v1: conditional staged-source auto-heal
v2: selected owned-field preview + receipt
Identity match is not semantic verification
No raw v2 rebaseline fallback
Sources: S11, S26 Legacy v1 heal / native v2 refresh v1: conditional staged-source auto-heal v2: selected owned-field preview + receipt Identity match is not semantic verification No raw v2 rebaseline fallback S11 Follow-up harvest / terminal boundary
[NEEDS-TRIAGE] → task writer for work
Unresolved decisions → open-questions.md
Interactive: confirm; sprint: permitted SMARTS logging
A task-board item is not a typed plan acceptance.
Sources: S04, S11 Follow-up harvest / terminal boundary [NEEDS-TRIAGE] → task writer for work Unresolved decisions → open-questions.md Interactive: confirm; sprint: permitted SMARTS logging A task-board item is not a typed plan acceptance. S04 Retained state / bounded recovery
Refresh keeps initialization and human history
Recover native operations by the same ID
Committed results are read back, not rewritten
Decisions remain with their actual owner
Sources: S13, S26 Retained state / bounded recovery Refresh keeps initialization and human history Recover native operations by the same ID Committed results are read back, not rewritten Decisions remain with their actual owner S13 Select task-relevant context
Task paths + optional coarse map
Assessed provenance and unknowns
Command cwd + observed verification
Scoped constraints and native instructions
Sources: S27 Select task-relevant context Task paths + optional coarse map Assessed provenance and unknowns Command cwd + observed verification Scoped constraints and native instructions S27 Compose for this lane and actor
Full feature: exact artifact IDs
Small feature: inline mini-spec
Fix / test-only: existing caller
Reviewer: named critical references
Sources: S27, S09, S08, S28 Compose for this lane and actor Full feature: exact artifact IDs Small feature: inline mini-spec Fix / test-only: existing caller Reviewer: named critical references S27 Bound actual actor delivery
Actor + task + actual worktree
Source / host / instruction epochs
deliver → send bounded child input
observed is reported, not authority
Sources: S27, S08 Bound actual actor delivery Actor + task + actual worktree Source / host / instruction epochs deliver → send bounded child input observed is reported, not authority S27 Reselect or stop
Resume / compaction / scope change
Worktree or evidence epoch change
At most two unobserved deliveries
Same binding still unobserved → block
Sources: S27, S08 Reselect or stop Resume / compaction / scope change Worktree or evidence epoch change At most two unobserved deliveries Same binding still unobserved → block S27 These are parallel context inputs. They are not four mandatory serial workflow steps; explicit scope reads remain the owner’s obligation. Reviewer charter detail is host-specific: Claude’s authority-reviewer serves frozen typed targets; the specialist fleet is not a substitute for that authority seam. Outputs feed the next relevant invocation through the project store. No new always-loaded context tool, automatic corpus merge or future dashboard is asserted. Snapshot boundary: the read-hint approved_spec_index scans *.md. Typed HTML task-context reads use the separate artifact engine path. Claude/Pi expose read-time notices; Codex lacks a native READ event in the documented host matrix. Do not equate generated surface parity with runtime parity. New HTML admission is blocked under Pi at this commit. Existing Markdown and inline small-feature lanes are not automatically converted. No packet invents a feature spec, overrides native instruction precedence, refreshes evidence, or proves the child consumed it. Read the map
Context / how the layers are assembled Durable knowledge → selective task input → actual actor delivery → independent proof. These are different boundaries.
Execution Context / data Conditional / authority Return / repeat Stop / refusal
Choose an entry to highlight its route. Select a node for context, related entries and pinned sources.
Start with a route /feature /fix /sprint /init /review /tribunal /pr /reconcile
Historical route snapshot at 9496cff6332fe0b97195b7ebe67b73f62f30e5b3. Current assistive source bindings identify inspected implementation bytes, with 3549ed56f5d5849116aef8d5e946423d8ee677fc as the committed source base. Live host qualification and release are separate evidence. Curated current-source workflow contracts and selected context-delivery implementation; not a release, installed-host qualification, or exhaustive runtime call graph. Original source links and spatial positions retained; task labels show the selected work owner. Source compatibility revalidated at 25980c292b529fca90c39f5f26d69a1c0674de44. Unittest collector compatibility bound to reviewed blob 76a2998f4bd5b6f441f5fd162bbe810b2de2ee93.
Snapshot limits Historical snapshot: New full-lane feature/sprint defaults to typed HTML only with an admitted installed payload and qualified producers. Main source checkout does not carry the release binary.
Historical snapshot: Pi fails new HTML workflow admission in the pinned artifacts contract. Existing Markdown pairs and inline small features remain separate.
Historical snapshot: HTML --farm dispatch is blocked. Legacy Markdown farm is an off-by-default preview, not the blessed default.
Historical snapshot: Read-time approved-spec pointers index Markdown only in the inspected implementation. Typed HTML execution reads are a distinct engine path.
Historical snapshot: Shared surface generation is not live host parity. Claude/Pi READ notices and Codex’s absence of a native READ event differ.
Historical snapshot: Task quality review is once per current combined scope after per-task spec proof and fresh verification, not a duplicate full fleet for every tiny task.
Historical snapshot: The diagrams describe default arbiter-mode workflow contracts; deterministic mode tokens are separate posture controls and not new slash commands.
Historical snapshot: Snapshots of narrative instructions do not prove every gate is mechanically enforced. No live workflows were executed for this atlas.
Historical snapshot: Brownfield context dispatch needs an actually qualified read-only caller profile; the generic scout charter and generated plugin surfaces alone do not qualify it. Repository-only preparation fixtures are not installed capability.
Historical snapshot: Five content/provenance pairs use independently approved native transactions. Initial finalization has its own 13-identity approval. Initialized refresh retains the marker and never re-enters finalization.
Historical snapshot: Task-context selection is inert. The caller delivers a separate bounded input to each actor; reported delivery, engine tickets and parent session markers are not interchangeable authority or proof.
Historical snapshot: Debug has five dispositions. Validation alone is not evidence truth or repair authority. Diagnosis-only requests stop after the recommendation.
Historical snapshot: Source wording ambiguity: brownfield Phase 3 preserves source-scoped upstream ADR references, while a retained Phase 5 sentence asks for local summaries. Local-summary semantics and authority need clarification; this atlas does not infer automatic local decision authority.
Historical snapshot: Pi approval is currently DEGRADED because the adapter is not implemented; the newer runtime input event means the older HOST-IMPOSSIBLE explanation must not be repeated.
Historical snapshot: Tribunal is opt-in. Its source-bound run store is separate from typed spec/plan authority; report-written does not establish terminal completion.
Historical snapshot: Ten new commits include correctness and waiting-limit repairs, not ten new routes. Their merge descriptions are not independent runtime proof. No latency improvement is asserted.
Historical snapshot: The release PR boundary was under-specified in the previous atlas; making it explicit is a carried presentation correction, not a behavior newly introduced by these ten commits.
Historical snapshot: The native page and offline editions share this source-bound atlas. Website integration does not establish released behavior, installed-host qualification or completed workflow execution.
Change lanes / alternate entries and exits Different starting evidence creates different obligations. A fix, a refactor, a docs edit and an experiment do not share identical ceremony. An agreed outcome becomes a maintained plan, direct or delegated implementation, proportional verification and an honest delivery report. Task context and dependency advice support progress; real user stops and host permissions remain effective. Greenfield and brownfield population use the paired native writer and exact finalization. Exact blob bindings identify inspected source bytes; the linked revision is the committed source base. Implementation, verification, live host qualification and release remain separate claims. Source 9496cff6332fe0b97195b7ebe67b73f62f30e5b3. Select a node for details or use the complete text path. HISTORICAL ROUTES / CODEARBITER Change lanes / alternate entries and exits Different starting evidence creates different obligations. A fix, a refactor, a docs edit and an experiment do not share identical ceremony. SOURCE 9496cff6332f / 2026-10-07 01 / INVESTIGATE FIRST WHEN THE CAUSE OR ANSWER IS UNKNOWN Diagnosis and exploration produce evidence. Neither silently turns into production implementation. 02 / PROVE THE RIGHT OBLIGATION BEFORE CHANGING CODE The code-map is optional orientation. Declared commands and applicable governing rules remain required. 03 / SCALE THE GATES TO NON-BEHAVIORAL WORK A chore is not a back door for changed behavior. 04 / DISTINGUISH DEPENDENCY ADOPTION FROM A ONE-OFF TOOL /add-dep is also a prerequisite branch that returns to the owner that requested it. Execution Execution Execution Conditional / authority: build build Execution Execution Execution Execution Execution Execution Execution Execution Conditional / authority: revert revert Execution Execution Execution Execution Conditional / authority Execution Execution Execution Execution /debug
Diagnosis-only or one internal
prerequisite from an authorized fix
No code modification
Sources: S25, S19 /debug Diagnosis-only or one internal prerequisite from an authorized fix No code modification S25 debug / evidence-led investigation
Cited symptom → real discriminators
Current source/runtime applicability
Exactly one supported disposition
Sources: S25 debug / evidence-led investigation Cited symptom → real discriminators Current source/runtime applicability Exactly one supported disposition S25 /spike
Bounded unknown + timebox
Sources: S02, S04 /spike Bounded unknown + timebox S02 Disposable spike/<slug>
Explore the question in isolation
Exemption cannot land production work
Never merges or opens a PR
Sources: S04 Disposable spike/<slug> Explore the question in isolation Exemption cannot land production work Never merges or opens a PR S04 Findings note
What was learned; uncertainty left
No requirement to build anything
Sources: S04 Findings note What was learned; uncertainty left No requirement to build anything S04 /feature / new production path
A separate approved specification
Experiment is evidence, not approval
Sources: S04, S06 /feature / new production path A separate approved specification Experiment is evidence, not approval S04 /fix
Observed ≠ expected behavior
Confirmed defect; cited evidence
Sources: S19, S27, S25 /fix Observed ≠ expected behavior Confirmed defect; cited evidence S19 Causal evidence → target RED
Cited cause + named regression
Observed target RED for the right reason
No feature spec or unrelated red
Sources: S19, S09, S25 Causal evidence → target RED Cited cause + named regression Observed target RED for the right reason No feature spec or unrelated red S19 Complete tdd
Obligations → red → green
Verify obligations → coverage → lint
Fresh local proof, not self-report
Sources: S09, S21 Complete tdd Obligations → red → green Verify obligations → coverage → lint Fresh local proof, not self-report S09 Verified correction
Scoped code + regression + evidence
Commit only when authorized
Sources: S19, S09, S11 Verified correction Scoped code + regression + evidence Commit only when authorized S19 /refactor
Named restructure
Behavior must stay the same
Sources: S21 /refactor Named restructure Behavior must stay the same S21 Parity-coverage proof
Unchanged pre-existing tests
Coverage backfill / new seams → TDD
No parity proof, no restructure
Sources: S04, S21 Parity-coverage proof Unchanged pre-existing tests Coverage backfill / new seams → TDD No parity proof, no restructure S04 refactor / bounded change
Restructure behind the proven contract
Re-run unchanged behavioral tests
Feature/fix owns behavioral changes
Sources: S04, S21 refactor / bounded change Restructure behind the proven contract Re-run unchanged behavioral tests Feature/fix owns behavioral changes S04 Verified parity
Restructured code + parity evidence
Commit only through commit-gate
Sources: S04, S21, S11 Verified parity Restructured code + parity evidence Commit only through commit-gate S04 /chore
Classify docs / deps / revert
Reject behavior smuggled into the lane
Sources: S20 /chore Classify docs / deps / revert Reject behavior smuggled into the lane S20 docs
Secrets + diff + user-facing copy checks
No TDD required for prose
Sources: S20 docs Secrets + diff + user-facing copy checks No TDD required for prose S20 deps / existing version bump
dependency-reviewer; manifest + lockfile
Affected local tests; code adoption redirects
Sources: S20 deps / existing version bump dependency-reviewer; manifest + lockfile Affected local tests; code adoption redirects S20 revert / named commit
git revert <SHA>; affected tests
Reason and SHA in the commit message
Missed defect? Follow with /fix
Sources: S20 revert / named commit git revert <SHA>; affected tests Reason and SHA in the commit message Missed defect? Follow with /fix S20 Reviewed non-behavioral change
Matching type proof
commit-gate → finishing owner
Exact-head hosted CI before merge
Explicit persistence authority still applies
Sources: S20, S11, S12 Reviewed non-behavioral change Matching type proof commit-gate → finishing owner Exact-head hosted CI before merge Explicit persistence authority still applies S20 /add-dep
Package purpose + identity
Read stack and security policy
Sources: S04, S05 /add-dep Package purpose + identity Read stack and security policy S04 Durable adoption / dependency-reviewer
License • provenance • maintenance • supply chain
Vetting outcome returns to the requesting workflow
Sources: S04, S22 Durable adoption / dependency-reviewer License • provenance • maintenance • supply chain Vetting outcome returns to the requesting workflow S04 One-off branch: pinned tool + registry + confirmation
Sources: S04 One-off branch: pinned tool + registry + confirmation S04 Vetted decision / tool result
Ephemeral: verify no manifest/lock change
Adoption: retain reviewed policy outcome
Sources: S04 Vetted decision / tool result Ephemeral: verify no manifest/lock change Adoption: retain reviewed policy outcome S04 Return to caller
Feature/fix owns behavior adoption
No automatic PR from a tool run
Sources: S03, S04 Return to caller Feature/fix owns behavior adoption No automatic PR from a tool run S03 Five outcomes; no automatic repair or ADR
Code defect → candidate fix handoff | Non-code cause → existing owner
Design question → user decision | No action → positive closure evidence
Unresolved → discriminator + next action + resume condition
Sources: S25, S19 Five outcomes; no automatic repair or ADR Code defect → candidate fix handoff | Non-code cause → existing owner Design question → user decision | No action → positive closure evidence Unresolved → discriminator + next action + resume condition S25 Common persistence boundary: a verified change is eligible to request commit-gate, not permission to commit, push, merge or release. Security, migration and dependency scope touches branch to their existing specialist owners; urgency never changes the selected lane’s gates. Read the map
Change lanes / alternate entries and exits Different starting evidence creates different obligations. A fix, a refactor, a docs edit and an experiment do not share identical ceremony.
Execution Context / data Conditional / authority Return / repeat Stop / refusal
Choose an entry to highlight its route. Select a node for context, related entries and pinned sources.
Start with a route /feature /fix /sprint /init /review /tribunal /pr /reconcile
Historical route snapshot at 9496cff6332fe0b97195b7ebe67b73f62f30e5b3. Current assistive source bindings identify inspected implementation bytes, with 3549ed56f5d5849116aef8d5e946423d8ee677fc as the committed source base. Live host qualification and release are separate evidence. Curated current-source workflow contracts and selected context-delivery implementation; not a release, installed-host qualification, or exhaustive runtime call graph. Original source links and spatial positions retained; task labels show the selected work owner. Source compatibility revalidated at 25980c292b529fca90c39f5f26d69a1c0674de44. Unittest collector compatibility bound to reviewed blob 76a2998f4bd5b6f441f5fd162bbe810b2de2ee93.
Snapshot limits Historical snapshot: New full-lane feature/sprint defaults to typed HTML only with an admitted installed payload and qualified producers. Main source checkout does not carry the release binary.
Historical snapshot: Pi fails new HTML workflow admission in the pinned artifacts contract. Existing Markdown pairs and inline small features remain separate.
Historical snapshot: HTML --farm dispatch is blocked. Legacy Markdown farm is an off-by-default preview, not the blessed default.
Historical snapshot: Read-time approved-spec pointers index Markdown only in the inspected implementation. Typed HTML execution reads are a distinct engine path.
Historical snapshot: Shared surface generation is not live host parity. Claude/Pi READ notices and Codex’s absence of a native READ event differ.
Historical snapshot: Task quality review is once per current combined scope after per-task spec proof and fresh verification, not a duplicate full fleet for every tiny task.
Historical snapshot: The diagrams describe default arbiter-mode workflow contracts; deterministic mode tokens are separate posture controls and not new slash commands.
Historical snapshot: Snapshots of narrative instructions do not prove every gate is mechanically enforced. No live workflows were executed for this atlas.
Historical snapshot: Brownfield context dispatch needs an actually qualified read-only caller profile; the generic scout charter and generated plugin surfaces alone do not qualify it. Repository-only preparation fixtures are not installed capability.
Historical snapshot: Five content/provenance pairs use independently approved native transactions. Initial finalization has its own 13-identity approval. Initialized refresh retains the marker and never re-enters finalization.
Historical snapshot: Task-context selection is inert. The caller delivers a separate bounded input to each actor; reported delivery, engine tickets and parent session markers are not interchangeable authority or proof.
Historical snapshot: Debug has five dispositions. Validation alone is not evidence truth or repair authority. Diagnosis-only requests stop after the recommendation.
Historical snapshot: Source wording ambiguity: brownfield Phase 3 preserves source-scoped upstream ADR references, while a retained Phase 5 sentence asks for local summaries. Local-summary semantics and authority need clarification; this atlas does not infer automatic local decision authority.
Historical snapshot: Pi approval is currently DEGRADED because the adapter is not implemented; the newer runtime input event means the older HOST-IMPOSSIBLE explanation must not be repeated.
Historical snapshot: Tribunal is opt-in. Its source-bound run store is separate from typed spec/plan authority; report-written does not establish terminal completion.
Historical snapshot: Ten new commits include correctness and waiting-limit repairs, not ten new routes. Their merge descriptions are not independent runtime proof. No latency improvement is asserted.
Historical snapshot: The release PR boundary was under-specified in the previous atlas; making it explicit is a carried presentation correction, not a behavior newly introduced by these ten commits.
Historical snapshot: The native page and offline editions share this source-bound atlas. Website integration does not establish released behavior, installed-host qualification or completed workflow execution.
Review + delivery / evidence and authority Review, commit and branch disposition retain separate authority; release preparation stops at its PR. An agreed outcome becomes a maintained plan, direct or delegated implementation, proportional verification and an honest delivery report. Task context and dependency advice support progress; real user stops and host permissions remain effective. Greenfield and brownfield population use the paired native writer and exact finalization. Exact blob bindings identify inspected source bytes; the linked revision is the committed source base. Implementation, verification, live host qualification and release remain separate claims. Source 9496cff6332fe0b97195b7ebe67b73f62f30e5b3. Select a node for details or use the complete text path. HISTORICAL ROUTES / CODEARBITER Review + delivery / evidence and authority Review, commit and branch disposition retain separate authority; release preparation stops at its PR. SOURCE 9496cff6332f / 2026-10-07 01 / REVIEW IS A SCOPED, READ-ONLY EVIDENCE PATH Select by actual scope. Routine reports, change blockers and independent typed-authority reviews have different contracts. 02 / /commit → commit-gate A local commit is the output. Every later action must retain its own requested scope and authority. 03 / /pr → SELECT THE MODE BEFORE CREATION PREFLIGHT Watch and cleanup are early return paths, not steps that must first create a PR. 04 / CONDITIONAL LEGACY CODE-MAP RECOVERY Only inside an authorized integration, after the exact-HEAD resolution is already selected. No extra approval is invented. 05 / /release → DECLARED TARGETS + SEPARATE PUBLICATION AUTHORITY A release is not an automatic continuation of a green PR. A release commit still goes through commit-gate. Execution Execution Execution Execution Execution Execution Conditional / authority Conditional / authority Execution: create / finish create / finish Execution Execution Conditional / authority: chosen merge, not sprint chosen merge, not sprint Execution Execution Conditional / authority: merged merged Conditional / authority: consent consent Conditional / authority Execution Execution /review · /checkpoint
/review: current diff
/checkpoint: periodic codebase sweep
/preview predicts only; no verdict
/tribunal and /threat-model are opt-in
Sources: S04, S18, S21 /review · /checkpoint /review: current diff /checkpoint: periodic codebase sweep /preview predicts only; no verdict /tribunal and /threat-model are opt-in S04 Fresh inputs → scoped reviewer fleet
Select current context for each reviewer
Required refs + native instruction order
Deliver to that actual child, not its parent
Then dispatch by the actual diff paths
Preserve capability and evidence gaps
Sources: S28, S27, S18 Fresh inputs → scoped reviewer fleet Select current context for each reviewer Required refs + native instruction order Deliver to that actual child, not its parent Then dispatch by the actual diff paths Preserve capability and evidence gaps S28 finding-triage
Collect and deduplicate findings
Retain terminal states / gaps
Classify severity and scope
Then pass the complete report onward
Sources: S04, S22 finding-triage Collect and deduplicate findings Retain terminal states / gaps Classify severity and scope Then pass the complete report onward S04 verdict-aggregator
Read-only in-memory result
Own change: CRITICAL/HIGH blocks PR
Inbound PR: local verdict; post separately
Checkpoint: explicit dated persistence
Sources: S28, S04 verdict-aggregator Read-only in-memory result Own change: CRITICAL/HIGH blocks PR Inbound PR: local verdict; post separately Checkpoint: explicit dated persistence S28 /commit
Current staged scope
TDD complete where applicable
Typed plans: current acceptance
No status-label substitutes
Sources: S11, S10 /commit Current staged scope TDD complete where applicable Typed plans: current acceptance No status-label substitutes S11 Phases 1–3 / authorize + scope
1 Explicit permission
2 Non-protected branch
3 Homogeneous change classification
No implied authorization from approval
of a specification or an implementation
Sources: S11 Phases 1–3 / authorize + scope 1 Explicit permission 2 Non-protected branch 3 Homogeneous change classification No implied authorization from approval of a specification or an implementation S11 Phases 4–6 / prove + inspect
4 Focused tests, lint, secrets, scope gates
5 Fresh behavioral proof
5.5 Legacy v1 staged-source auto-heal
6 Complete diff review
v2 pairs require native selected updates
Sources: S11, S26 Phases 4–6 / prove + inspect 4 Focused tests, lint, secrets, scope gates 5 Fresh behavioral proof 5.5 Legacy v1 staged-source auto-heal 6 Complete diff review v2 pairs require native selected updates S11 Phases 7–9 / persist
7 Harvest → explicitly selected paths
8 Conventional Commit message
9 Commit → SHA + clean-tree report
No blanket staging; failed hook returns
to correction, not gate circumvention
Sources: S11 Phases 7–9 / persist 7 Harvest → explicitly selected paths 8 Conventional Commit message 9 Commit → SHA + clean-tree report No blanket staging; failed hook returns to correction, not gate circumvention S11 Conditional scope gates / they do not add new public routes
crypto-compliance + secret-handling → auth-crypto-reviewer; migrations → migration-reviewer; CI/deploy → security-reviewer; dependencies → dependency-reviewer.
Sources: S04, S05, S11 Conditional scope gates / they do not add new public routes crypto-compliance + secret-handling → auth-crypto-reviewer; migrations → migration-reviewer; CI/deploy → security-reviewer; dependencies → dependency-reviewer. S04 /pr
Direct open request or
feature/sprint terminal handoff
Choose mode from user authority
Sources: S12 /pr Direct open request or feature/sprint terminal handoff Choose mode from user authority S12 --watch / retained /watch
Detached hosted-CI watcher
Red → diagnosis; green → offer, never auto-merge
Sources: S04, S12 --watch / retained /watch Detached hosted-CI watcher Red → diagnosis; green → offer, never auto-merge S04 --cleanup / retained /cleanup
post-merge-cleanup
Prove fetched containment, classify residue
Confirm each discard; preserve unique work
Sources: S04, S12, S21 --cleanup / retained /cleanup post-merge-cleanup Prove fetched containment, classify residue Confirm each discard; preserve unique work S04 Creation / current-head preflight
Current acceptance + commit proof
Branch / base / head / diff
Applicable ADR ancestry
Missing or stale proof → owner repair
Sources: S12 Creation / current-head preflight Current acceptance + commit proof Branch / base / head / diff Applicable ADR ancestry Missing or stale proof → owner repair S12 Open-PR procedure in this owner
Applicable reviewers; clear blockers
Compose summary, test plan and ADR links
Push branch, create PR, return URL
Never re-invoke the PR wrapper
Sources: S12 Open-PR procedure in this owner Applicable reviewers; clear blockers Compose summary, test plan and ADR links Push branch, create PR, return URL Never re-invoke the PR wrapper S12 Merge via PR / only if selected
Explicit user authority + exact-head CI
Revalidate ancestry and bound head
Sources: S12 Merge via PR / only if selected Explicit user authority + exact-head CI Revalidate ancestry and bound head S12 Open PR / authorized disposition
Sprint ends at open PR
Feature keeps selected terminal option
Discard requires stated-loss confirmation
Sources: S12 Open PR / authorized disposition Sprint ends at open PR Feature keeps selected terminal option Discard requires stated-loss confirmation S12 /release · selected target
Resolve declared target + policy
Prepare reviewed release surfaces
Dry-run stops at its report
Sources: S33 /release · selected target Resolve declared target + policy Prepare reviewed release surfaces Dry-run stops at its report S33 Release preparation PR
Commit through normal gates
Name assets + declared publisher
Open release PR and STOP
Sources: S33, S34 Release preparation PR Commit through normal gates Name assets + declared publisher Open release PR and STOP S33 Qualified hosted continuation
Only after the release PR merges
Exact fetched head + green evidence
Compose the bound tag and cohort
Sources: S33 Qualified hosted continuation Only after the release PR merges Exact fetched head + green evidence Compose the bound tag and cohort S33 Publication and closeout
Consume applicable cohort authority
Read back release + inventory
Declared receipt closes by separate PR
Sources: S33, S34 Publication and closeout Consume applicable cohort authority Read back release + inventory Declared receipt closes by separate PR S33 Selected legacy resolution
Actual merge + schema-1 provenance
Exact current HEAD chosen
Native or uncertain ownership refuses
Sources: S12 Selected legacy resolution Actual merge + schema-1 provenance Exact current HEAD chosen Native or uncertain ownership refuses S12 Inspect → exact binding
Shipped helper inspects Git + files
Pass the returned binding directly
Changed identities refuse
Sources: S12 Inspect → exact binding Shipped helper inspects Git + files Pass the returned binding directly Changed identities refuse S12 Restore HEAD only
One supported conflict rendering
Exact committed code-map bytes
Provenance and index untouched
Sources: S12 Restore HEAD only One supported conflict rendering Exact committed code-map bytes Provenance and index untouched S12 Review, stage, re-prove
Inspect the resulting diff
Use normal explicit staging
Review + CI bind the integrated head
Sources: S12 Review, stage, re-prove Inspect the resulting diff Use normal explicit staging Review + CI bind the integrated head S12 SDD combined-scope review has its own caller rules. Claude authority-reviewer serves frozen typed targets; no generic reviewer result fabricates its receipt. Actual reviewer delivery is bound to actor/task/worktree/epochs. A source-current identity, parent marker or author packet is not reviewer proof. Local checks are impact-bounded; exhaustive repository and platform proof is exact-head hosted CI before merge. A previously green head is not current evidence. Content/path-bound security-pass behavior is retained from the baseline. New context and debug contracts do not waive existing specialist gates. Informational reports return to their caller. Nothing in the review path grants authority to edit, commit, file issues, merge, publish or spend. Read the map
Review + delivery / evidence and authority Review, commit and branch disposition retain separate authority; release preparation stops at its PR.
Execution Context / data Conditional / authority Return / repeat Stop / refusal
Choose an entry to highlight its route. Select a node for context, related entries and pinned sources.
Start with a route /feature /fix /sprint /init /review /tribunal /pr /reconcile
Historical route snapshot at 9496cff6332fe0b97195b7ebe67b73f62f30e5b3. Current assistive source bindings identify inspected implementation bytes, with 3549ed56f5d5849116aef8d5e946423d8ee677fc as the committed source base. Live host qualification and release are separate evidence. Curated current-source workflow contracts and selected context-delivery implementation; not a release, installed-host qualification, or exhaustive runtime call graph. Original source links and spatial positions retained; task labels show the selected work owner. Source compatibility revalidated at 25980c292b529fca90c39f5f26d69a1c0674de44. Unittest collector compatibility bound to reviewed blob 76a2998f4bd5b6f441f5fd162bbe810b2de2ee93.
Snapshot limits Historical snapshot: New full-lane feature/sprint defaults to typed HTML only with an admitted installed payload and qualified producers. Main source checkout does not carry the release binary.
Historical snapshot: Pi fails new HTML workflow admission in the pinned artifacts contract. Existing Markdown pairs and inline small features remain separate.
Historical snapshot: HTML --farm dispatch is blocked. Legacy Markdown farm is an off-by-default preview, not the blessed default.
Historical snapshot: Read-time approved-spec pointers index Markdown only in the inspected implementation. Typed HTML execution reads are a distinct engine path.
Historical snapshot: Shared surface generation is not live host parity. Claude/Pi READ notices and Codex’s absence of a native READ event differ.
Historical snapshot: Task quality review is once per current combined scope after per-task spec proof and fresh verification, not a duplicate full fleet for every tiny task.
Historical snapshot: The diagrams describe default arbiter-mode workflow contracts; deterministic mode tokens are separate posture controls and not new slash commands.
Historical snapshot: Snapshots of narrative instructions do not prove every gate is mechanically enforced. No live workflows were executed for this atlas.
Historical snapshot: Brownfield context dispatch needs an actually qualified read-only caller profile; the generic scout charter and generated plugin surfaces alone do not qualify it. Repository-only preparation fixtures are not installed capability.
Historical snapshot: Five content/provenance pairs use independently approved native transactions. Initial finalization has its own 13-identity approval. Initialized refresh retains the marker and never re-enters finalization.
Historical snapshot: Task-context selection is inert. The caller delivers a separate bounded input to each actor; reported delivery, engine tickets and parent session markers are not interchangeable authority or proof.
Historical snapshot: Debug has five dispositions. Validation alone is not evidence truth or repair authority. Diagnosis-only requests stop after the recommendation.
Historical snapshot: Source wording ambiguity: brownfield Phase 3 preserves source-scoped upstream ADR references, while a retained Phase 5 sentence asks for local summaries. Local-summary semantics and authority need clarification; this atlas does not infer automatic local decision authority.
Historical snapshot: Pi approval is currently DEGRADED because the adapter is not implemented; the newer runtime input event means the older HOST-IMPOSSIBLE explanation must not be repeated.
Historical snapshot: Tribunal is opt-in. Its source-bound run store is separate from typed spec/plan authority; report-written does not establish terminal completion.
Historical snapshot: Ten new commits include correctness and waiting-limit repairs, not ten new routes. Their merge descriptions are not independent runtime proof. No latency improvement is asserted.
Historical snapshot: The release PR boundary was under-specified in the previous atlas; making it explicit is a carried presentation correction, not a behavior newly introduced by these ten commits.
Historical snapshot: The native page and offline editions share this source-bound atlas. Website integration does not establish released behavior, installed-host qualification or completed workflow execution.
Knowledge + operations / supporting routes These entries have bounded outputs: decisions, reports, board transitions, repair or runtime posture. They do not implicitly enter implementation. An agreed outcome becomes a maintained plan, direct or delegated implementation, proportional verification and an honest delivery report. Task context and dependency advice support progress; real user stops and host permissions remain effective. Greenfield and brownfield population use the paired native writer and exact finalization. Exact blob bindings identify inspected source bytes; the linked revision is the committed source base. Implementation, verification, live host qualification and release remain separate claims. Source 9496cff6332fe0b97195b7ebe67b73f62f30e5b3. Select a node for details or use the complete text path. HISTORICAL ROUTES / CODEARBITER Knowledge + operations / supporting routes These entries have bounded outputs: decisions, reports, board transitions, repair or runtime posture. They do not implicitly enter implementation. SOURCE 9496cff6332f / 2026-10-07 01 / USER-OWNED INTENT AND CONTRADICTION HANDLING Separate a recommendation from a decision, and an accepted decision from delivery evidence. 02 / READ STATE, REPAIR KNOWLEDGE, MANAGE WORK Each writing action uses its sanctioned owner. Read-only entry modes remain read-only. 03 / GOVERNANCE RECORDS AND OBSERVABILITY Source records remain authoritative; a summary neither changes them nor approves a workflow. 04 / HOST POSTURE AND CONTEXT HYGIENE These controls must not be drawn as an unrestricted alternate route around the normal pipeline. 05 / HELP AND RETAINED ENTRY SPELLINGS Aliases preserve the existing owner’s procedure and gates; they are not alternate semantics. Execution Execution Context / data Execution Execution Conditional / authority: conflict conflict Conditional / authority: drift drift Execution Execution Execution /adr · /adr-status
Author or read-only status
Select mode before prerequisites
Sources: S21 /adr · /adr-status Author or read-only status Select mode before prerequisites S21 decision-lifecycle
Author: actual user choice + attribution
Status: age, challenges, supersession
CONFIRM questions are never guessed
Sources: S04, S21 decision-lifecycle Author: actual user choice + attribution Status: age, challenges, supersession CONFIRM questions are never guessed S04 ADR / health report
Numbered, dated decision record
Or read-only lifecycle report
Accepted/Planned ≠ implementation
Sources: S21 ADR / health report Numbered, dated decision record Or read-only lifecycle report Accepted/Planned ≠ implementation S21 Future scope-relevant context
Recorded intent constrains spec authoring
Affected ADRs return through indexes
No automatic product change
Sources: S05, S07, S21 Future scope-relevant context Recorded intent constrains spec authoring Affected ADRs return through indexes No automatic product change S05 /reconcile
Drift / architectural variance
Report-only or explicit reconciliation
Sources: S04, S21 /reconcile Drift / architectural variance Report-only or explicit reconciliation S04 decision-variance
Scoped scout evidence → SMARTS grader
Optional decision-challenger
Grader recommends; user chooses
Sources: S04, S21, S22 decision-variance Scoped scout evidence → SMARTS grader Optional decision-challenger Grader recommends; user chooses S04 Report / attributed ruling
Read-only mode writes nothing
Chosen reconciliation enters decision-log
Separate ADR owner when required
Sources: S04, S21 Report / attributed ruling Read-only mode writes nothing Chosen reconciliation enters decision-log Separate ADR owner when required S04 /conflict / internal STOP
Contradictory governing rules
Stop other work; surface both sides
Resolve authority before resuming
Sources: S03, S04 /conflict / internal STOP Contradictory governing rules Stop other work; surface both sides Resolve authority before resuming S03 /status
Default: read-only state
drift: provenance audit
Alias: /context-check
Sources: S02, S04 /status Default: read-only state drift: provenance audit Alias: /context-check S02 context-check / selected refresh
Report v1 and v2 states separately
Select affected docs or explicit full refresh
Qualified scouts + native v2 pair writes
Sources: S26, S13 context-check / selected refresh Report v1 and v2 states separately Select affected docs or explicit full refresh Qualified scouts + native v2 pair writes S26 Status / drift report
Read-only report, or selected native update
Keep marker, unselected files and history
No staging, commit or reinitialization
Sources: S26, S13 Status / drift report Read-only report, or selected native update Keep marker, unselected files and history No staging, commit or reinitialization S26 /doctor
Install, payload, package ownership
Enforcement config + harmless probe
Output: diagnostic / capability gap
Sources: S02, S10 /doctor Install, payload, package ownership Enforcement config + harmless probe Output: diagnostic / capability gap S02 /task
Authorized add / start / done
Terminal harvest or concrete follow-up
Debug outcome alone writes nothing
Sources: S25, S04 /task Authorized add / start / done Terminal harvest or concrete follow-up Debug outcome alone writes nothing S25 work.py → selected owner
One facade selects native or legacy work
Queued work can complete directly
Checks remain separately evidenced
Sources: S04 work.py → selected owner One facade selects native or legacy work Queued work can complete directly Checks remain separately evidenced S04 Selected work transition
Native HTML or open-tasks.md state
Not a typed-plan acceptance receipt
Later work still selects its own lane
Sources: S04, S08 Selected work transition Native HTML or open-tasks.md state Not a typed-plan acceptance receipt Later work still selects its own lane S04 /standup
Inspect hygiene, branches and worktrees
Confirm each safe cleanup
Never silently lose dirty / unpushed work
Sources: S02, S04 /standup Inspect hygiene, branches and worktrees Confirm each safe cleanup Never silently lose dirty / unpushed work S02 /audit → dated governance packet
Read source records for a selected window. Preserve audit lines.
Write a new packet only; never overwrite the source or an older packet.
Sources: S02, S04 /audit → dated governance packet Read source records for a selected window. Preserve audit lines. Write a new packet only; never overwrite the source or an older packet. S02 /metrics → trends against prior window
Overrides, small-lane classifications and low-confidence sprint choices.
Read-only measurement, not an automatic policy change.
Sources: S02 /metrics → trends against prior window Overrides, small-lane classifications and low-confidence sprint choices. Read-only measurement, not an automatic policy change. S02 /override → attributed exception
Only where the governing hard rule permits; explicit reason and consent.
Append to overrides.log. Not a universal security or authority bypass.
Sources: S03, S04 /override → attributed exception Only where the governing hard rule permits; explicit reason and consent. Append to overrides.log. Not a universal security or authority bypass. S03 mode --arbiter / --ops / --dangerous
Deterministic control tokens, not slash entries.
Safety-core remains composed with the active mode body.
Ops is advisory runtime work; tracked/history changes route normally.
Sources: S03, S04, S15 mode --arbiter / --ops / --dangerous Deterministic control tokens, not slash entries. Safety-core remains composed with the active mode body. Ops is advisory runtime work; tracked/history changes route normally. S03 /prune → host-specific context control
Claude: dry-run by default, never execute against the live transcript.
Pi: native compaction, not active-session rewriting.
No Codex transcript route is asserted.
Sources: S02, S04 /prune → host-specific context control Claude: dry-run by default, never execute against the live transcript. Pi: native compaction, not active-session rewriting. No Codex transcript route is asserted. S02 /statusline → Claude configuration
Install / uninstall / inspect the owned statusline.
This catalog surface is Claude-specific.
It does not establish cross-host runtime parity.
Sources: S02 /statusline → Claude configuration Install / uninstall / inspect the owned statusline. This catalog surface is Claude-specific. It does not establish cross-host runtime parity. S02 /commands · /btw
Grouped catalog / deprecated lightweight Q&A. Direct questions need no slash command.
Sources: S01, S02, S03 /commands · /btw Grouped catalog / deprecated lightweight Q&A. Direct questions need no slash command. S01 /decompose · /create-context
Initial population only: /init --greenfield / --brownfield. Initialized refresh stays with /status drift and explicit selection.
Sources: S23, S26 /decompose · /create-context Initial population only: /init --greenfield / --brownfield. Initialized refresh stays with /status drift and explicit selection. S23 /watch · /cleanup · /context-check
Prefer /pr --watch / --cleanup and /status drift. These modes branch before unrelated creation-time preflights.
Sources: S01, S02, S12 /watch · /cleanup · /context-check Prefer /pr --watch / --cleanup and /status drift. These modes branch before unrelated creation-time preflights. S01 All 37 commands are selectable in the HTML. For auxiliary routes, the atlas records catalog-level owner contracts rather than claiming a line-by-line implementation audit. Read the map
Knowledge + operations / supporting routes These entries have bounded outputs: decisions, reports, board transitions, repair or runtime posture. They do not implicitly enter implementation.
Execution Context / data Conditional / authority Return / repeat Stop / refusal
Choose an entry to highlight its route. Select a node for context, related entries and pinned sources.
Start with a route /feature /fix /sprint /init /review /tribunal /pr /reconcile
Historical route snapshot at 9496cff6332fe0b97195b7ebe67b73f62f30e5b3. Current assistive source bindings identify inspected implementation bytes, with 3549ed56f5d5849116aef8d5e946423d8ee677fc as the committed source base. Live host qualification and release are separate evidence. Curated current-source workflow contracts and selected context-delivery implementation; not a release, installed-host qualification, or exhaustive runtime call graph. Original source links and spatial positions retained; task labels show the selected work owner. Source compatibility revalidated at 25980c292b529fca90c39f5f26d69a1c0674de44. Unittest collector compatibility bound to reviewed blob 76a2998f4bd5b6f441f5fd162bbe810b2de2ee93.
Snapshot limits Historical snapshot: New full-lane feature/sprint defaults to typed HTML only with an admitted installed payload and qualified producers. Main source checkout does not carry the release binary.
Historical snapshot: Pi fails new HTML workflow admission in the pinned artifacts contract. Existing Markdown pairs and inline small features remain separate.
Historical snapshot: HTML --farm dispatch is blocked. Legacy Markdown farm is an off-by-default preview, not the blessed default.
Historical snapshot: Read-time approved-spec pointers index Markdown only in the inspected implementation. Typed HTML execution reads are a distinct engine path.
Historical snapshot: Shared surface generation is not live host parity. Claude/Pi READ notices and Codex’s absence of a native READ event differ.
Historical snapshot: Task quality review is once per current combined scope after per-task spec proof and fresh verification, not a duplicate full fleet for every tiny task.
Historical snapshot: The diagrams describe default arbiter-mode workflow contracts; deterministic mode tokens are separate posture controls and not new slash commands.
Historical snapshot: Snapshots of narrative instructions do not prove every gate is mechanically enforced. No live workflows were executed for this atlas.
Historical snapshot: Brownfield context dispatch needs an actually qualified read-only caller profile; the generic scout charter and generated plugin surfaces alone do not qualify it. Repository-only preparation fixtures are not installed capability.
Historical snapshot: Five content/provenance pairs use independently approved native transactions. Initial finalization has its own 13-identity approval. Initialized refresh retains the marker and never re-enters finalization.
Historical snapshot: Task-context selection is inert. The caller delivers a separate bounded input to each actor; reported delivery, engine tickets and parent session markers are not interchangeable authority or proof.
Historical snapshot: Debug has five dispositions. Validation alone is not evidence truth or repair authority. Diagnosis-only requests stop after the recommendation.
Historical snapshot: Source wording ambiguity: brownfield Phase 3 preserves source-scoped upstream ADR references, while a retained Phase 5 sentence asks for local summaries. Local-summary semantics and authority need clarification; this atlas does not infer automatic local decision authority.
Historical snapshot: Pi approval is currently DEGRADED because the adapter is not implemented; the newer runtime input event means the older HOST-IMPOSSIBLE explanation must not be repeated.
Historical snapshot: Tribunal is opt-in. Its source-bound run store is separate from typed spec/plan authority; report-written does not establish terminal completion.
Historical snapshot: Ten new commits include correctness and waiting-limit repairs, not ten new routes. Their merge descriptions are not independent runtime proof. No latency improvement is asserted.
Historical snapshot: The release PR boundary was under-specified in the previous atlas; making it explicit is a carried presentation correction, not a behavior newly introduced by these ten commits.
Historical snapshot: The native page and offline editions share this source-bound atlas. Website integration does not establish released behavior, installed-host qualification or completed workflow execution.
Brownfield context · create, refresh, recover One content/provenance pair per native transaction; initial finalization is separate. Existing human records retain their owners. An agreed outcome becomes a maintained plan, direct or delegated implementation, proportional verification and an honest delivery report. Task context and dependency advice support progress; real user stops and host permissions remain effective. Greenfield and brownfield population use the paired native writer and exact finalization. Exact blob bindings identify inspected source bytes; the linked revision is the committed source base. Implementation, verification, live host qualification and release remain separate claims. Source 9496cff6332fe0b97195b7ebe67b73f62f30e5b3. Select a node for details or use the complete text path. HISTORICAL ROUTES / CODEARBITER Brownfield context · create, refresh, recover One content/provenance pair per native transaction; initial finalization is separate. Existing human records retain their owners. SOURCE 9496cff6332f / 2026-10-07 01 / PICK THE EXISTING OWNER Initial population, initialized refresh and interrupted-operation recovery are not interchangeable entries. 02 / ACQUIRE BOUNDED EVIDENCE BEFORE SYNTHESIS A missing qualified profile blocks dispatch. Complete returned reports and current snapshots are prerequisites, not optional best effort. 03 / WRITE ONLY THROUGH THE DOCUMENT’S NATIVE CONTRACT Five content documents have paired v2 provenance. An approval for one exact preview does not approve another document or finalization. 04 / COMPLETE THE SELECTED LIFECYCLE, NOT A DIFFERENT ONE Initial finalization and initialized refresh have different completion evidence. Recovery inspects native outcomes before a new write. Execution Conditional / authority: only after explicit selection only after explicit selection Execution Execution Execution Context / data Context / data Execution Execution Conditional / authority: initial population initial population Conditional / authority: selected refresh selected refresh Return / repeat Context / data Context / data Initial population
/init --brownfield → context-creation
Meaningful source; initialized marker absent
A stub is not a completed project context
Sources: S13, S23 Initial population /init --brownfield → context-creation Meaningful source; initialized marker absent A stub is not a completed project context S13 Initialized refresh
/status drift → context-check
Read-only report, then explicit scoped/full choice
Do not clear the marker or rerun initialization
Sources: S26 Initialized refresh /status drift → context-check Read-only report, then explicit scoped/full choice Do not clear the marker or rerun initialization S26 Interrupted operation
Inventory actual files, provenance and outcome
Inspect the same operation ID and native journal
Do not blindly repeat a possibly committed write
Sources: S13 Interrupted operation Inventory actual files, provenance and outcome Inspect the same operation ID and native journal Do not blindly repeat a possibly committed write S13 Qualified containment first
Fresh, bound child identities
Observed host-enforced read-only profile
Generic scout role is not sufficient
Generated plugin alone does not qualify it
Sources: S13 Qualified containment first Fresh, bound child identities Observed host-enforced read-only profile Generic scout role is not sufficient Generated plugin alone does not qualify it S13 Bounded logical assignments
Full run: six domains A–F
Scoped run: affected assignments only
Frozen source, budgets and attempt ledger
Reports only; no inline source fallback
Sources: S13, S26 Bounded logical assignments Full run: six domains A–F Scoped run: affected assignments only Frozen source, budgets and attempt ledger Reports only; no inline source fallback S13 Validate and join reports
Closed assignment/snapshot binding
48 KiB normalized join limit
Missing / failed transport → block
Recheck the source snapshot
Sources: S13 Validate and join reports Closed assignment/snapshot binding 48 KiB normalized join limit Missing / failed transport → block Recheck the source snapshot S13 Synthesize; settle relevant gaps
Use joined reports, not raw source
Preserve conflicting citations
Ask material user-owned choices
Deferral is not approval
Sources: S13 Synthesize; settle relevant gaps Use joined reports, not raw source Preserve conflicting citations Ask material user-owned choices Deferral is not approval S13 Five content targets
CONTEXT.md • tech-stack.md
coding-standards.md
security-controls.md • code-map.md
Each with .provenance/<doc>.json
Sources: S13, S26 Five content targets CONTEXT.md • tech-stack.md coding-standards.md security-controls.md • code-map.md Each with .provenance/<doc>.json S13 Finite typed preview
New document or selected owned field
Current source / target identities
Exact proposed content + provenance
Unsupported human layout → refuse
Sources: S13, S26 Finite typed preview New document or selected owned field Current source / target identities Exact proposed content + provenance Unsupported human layout → refuse S13 Native approve-context reply
Review this exact proposed pair
Capture actual host-observed consent
No reused approval, hash-only consent
or caller-supplied authority
Sources: S13, S26 Native approve-context reply Review this exact proposed pair Capture actual host-observed consent No reused approval, hash-only consent or caller-supplied authority S13 Receipt-only native apply
Qualified installed writer required
Recheck under the transaction lock
Commit document + provenance together
Read back the committed pair
Sources: S13, S26 Receipt-only native apply Qualified installed writer required Recheck under the transaction lock Commit document + provenance together Read back the committed pair S13 Preserve other owners and human bytes
Task / question boards, append-only audit history, release targets and ADRs keep their own writers. No raw Write/Edit fallback on refusal.
Sources: S13, S26 Preserve other owners and human bytes Task / question boards, append-only audit history, release targets and ADRs keep their own writers. No raw Write/Edit fallback on refusal. S13 Initial finalize: separate approval
13 identities: five docs + five provenance
plus task board, question board, audit
Own preview → own approve-context receipt
Marker + CONTEXT provenance commit together
Sources: S13 Initial finalize: separate approval 13 identities: five docs + five provenance plus task board, question board, audit Own preview → own approve-context receipt Marker + CONTEXT provenance commit together S13 Refresh: retain initialized state
Only selected pairs change
Declare scoped versus full coverage
Keep marker, human content and history
No finalization, staging or commit here
Sources: S26 Refresh: retain initialized state Only selected pairs change Declare scoped versus full coverage Keep marker, human content and history No finalization, staging or commit here S26 Recover by the same operation ID
Inspect pending / committed / uncertain result
Native recover: justified complete or rollback
Committed → read back and retain
Divergence / unknown ownership → conflict
Sources: S13 Recover by the same operation ID Inspect pending / committed / uncertain result Native recover: justified complete or rollback Committed → read back and retain Divergence / unknown ownership → conflict S13 Use the result through task-scoped selection
Context knowledge feeds the existing feature, fix, test-change and reviewer callers. They select current relevant inputs and deliver them to each actual actor.
A successful context write or initialized marker is not task authorization, current command proof, or evidence that a child loaded those inputs.
Sources: S27, S09, S28 Use the result through task-scoped selection Context knowledge feeds the existing feature, fix, test-change and reviewer callers. They select current relevant inputs and deliver them to each actual actor. A successful context write or initialized marker is not task authorization, current command proof, or evidence that a child loaded those inputs. S27 A current hash is an identity check, not semantic approval. “Not found in this scope” is not a claim of universal absence. Each pair is transactional; independent successful pairs are not a claim that the entire context set became visible atomically. Source-delivery exception and open live qualification remain distinct from these implemented source contracts. No blanket host-readiness claim. Read the map
Brownfield context · create, refresh, recover One content/provenance pair per native transaction; initial finalization is separate. Existing human records retain their owners.
Execution Context / data Conditional / authority Return / repeat Stop / refusal
Choose an entry to highlight its route. Select a node for context, related entries and pinned sources.
Start with a route /feature /fix /sprint /init /review /tribunal /pr /reconcile
Historical route snapshot at 9496cff6332fe0b97195b7ebe67b73f62f30e5b3. Current assistive source bindings identify inspected implementation bytes, with 3549ed56f5d5849116aef8d5e946423d8ee677fc as the committed source base. Live host qualification and release are separate evidence. Curated current-source workflow contracts and selected context-delivery implementation; not a release, installed-host qualification, or exhaustive runtime call graph. Original source links and spatial positions retained; task labels show the selected work owner. Source compatibility revalidated at 25980c292b529fca90c39f5f26d69a1c0674de44. Unittest collector compatibility bound to reviewed blob 76a2998f4bd5b6f441f5fd162bbe810b2de2ee93.
Snapshot limits Historical snapshot: New full-lane feature/sprint defaults to typed HTML only with an admitted installed payload and qualified producers. Main source checkout does not carry the release binary.
Historical snapshot: Pi fails new HTML workflow admission in the pinned artifacts contract. Existing Markdown pairs and inline small features remain separate.
Historical snapshot: HTML --farm dispatch is blocked. Legacy Markdown farm is an off-by-default preview, not the blessed default.
Historical snapshot: Read-time approved-spec pointers index Markdown only in the inspected implementation. Typed HTML execution reads are a distinct engine path.
Historical snapshot: Shared surface generation is not live host parity. Claude/Pi READ notices and Codex’s absence of a native READ event differ.
Historical snapshot: Task quality review is once per current combined scope after per-task spec proof and fresh verification, not a duplicate full fleet for every tiny task.
Historical snapshot: The diagrams describe default arbiter-mode workflow contracts; deterministic mode tokens are separate posture controls and not new slash commands.
Historical snapshot: Snapshots of narrative instructions do not prove every gate is mechanically enforced. No live workflows were executed for this atlas.
Historical snapshot: Brownfield context dispatch needs an actually qualified read-only caller profile; the generic scout charter and generated plugin surfaces alone do not qualify it. Repository-only preparation fixtures are not installed capability.
Historical snapshot: Five content/provenance pairs use independently approved native transactions. Initial finalization has its own 13-identity approval. Initialized refresh retains the marker and never re-enters finalization.
Historical snapshot: Task-context selection is inert. The caller delivers a separate bounded input to each actor; reported delivery, engine tickets and parent session markers are not interchangeable authority or proof.
Historical snapshot: Debug has five dispositions. Validation alone is not evidence truth or repair authority. Diagnosis-only requests stop after the recommendation.
Historical snapshot: Source wording ambiguity: brownfield Phase 3 preserves source-scoped upstream ADR references, while a retained Phase 5 sentence asks for local summaries. Local-summary semantics and authority need clarification; this atlas does not infer automatic local decision authority.
Historical snapshot: Pi approval is currently DEGRADED because the adapter is not implemented; the newer runtime input event means the older HOST-IMPOSSIBLE explanation must not be repeated.
Historical snapshot: Tribunal is opt-in. Its source-bound run store is separate from typed spec/plan authority; report-written does not establish terminal completion.
Historical snapshot: Ten new commits include correctness and waiting-limit repairs, not ten new routes. Their merge descriptions are not independent runtime proof. No latency improvement is asserted.
Historical snapshot: The release PR boundary was under-specified in the previous atlas; making it explicit is a carried presentation correction, not a behavior newly introduced by these ten commits.
Historical snapshot: The native page and offline editions share this source-bound atlas. Website integration does not establish released behavior, installed-host qualification or completed workflow execution.
Debug · evidence, dispositions and safe handoff Investigation is read-only. A validated packet carries evidence; the original caller still owns permission to act. An agreed outcome becomes a maintained plan, direct or delegated implementation, proportional verification and an honest delivery report. Task context and dependency advice support progress; real user stops and host permissions remain effective. Greenfield and brownfield population use the paired native writer and exact finalization. Exact blob bindings identify inspected source bytes; the linked revision is the committed source base. Implementation, verification, live host qualification and release remain separate claims. Source 9496cff6332fe0b97195b7ebe67b73f62f30e5b3. Select a node for details or use the complete text path. HISTORICAL ROUTES / CODEARBITER Debug · evidence, dispositions and safe handoff Investigation is read-only. A validated packet carries evidence; the original caller still owns permission to act. SOURCE 9496cff6332f / 2026-10-07 01 / INVESTIGATE THE OBSERVED CASE Natural-language diagnosis and /debug select the same owner. An authorized /fix may invoke one bounded internal diagnostic prerequisite. 02 / FIVE EXPLICIT OUTCOMES Exactly one case disposition. Each retains its evidence and limits; none silently creates a task, ADR, repair or operational change. 03 / TRANSFER EVIDENCE ONLY THROUGH A VALID HANDOFF The transient packet is not a persistent case store. Validation checks structure, not truth, freshness, consumed context or permission. 04 / ONLY AN AUTHORIZED, CURRENT CODE-DEFECT HANDOFF ENTERS REPAIR Direct confirmed bugs can enter /fix without public debug. One internal diagnosis returns to the same fix attempt; never recursively retype public routes. 05 / RESUME AND FOLLOW-UP WITHOUT INVENTED STATE The summary is bounded and evidence-linked. The packet is not a memory service or a grant of authority to the next workflow. Execution Execution Execution Conditional / authority Conditional / authority Conditional / authority Conditional / authority Conditional / authority Context / data: candidate handoff candidate handoff Context / data Context / data Context / data Context / data Execution Execution Execution Stop / refusal Stop / refusal Conditional / authority: qualifying result + current evidence + original repair authority qualifying result + current evidence + original repair authority Execution Execution Execution Keep actual request scope
Diagnosis only → recommend and stop
Existing fix authority stays bounded
No code, tests or configuration edits
No compulsory new onboarding
Sources: S25, S19 Keep actual request scope Diagnosis only → recommend and stop Existing fix authority stays bounded No code, tests or configuration edits No compulsory new onboarding S25 Capture symptom + applicability
Observed / expected + cited sources
Known reproduction state and limits
Actual worktree, source/build/runtime IDs
Unknowns stay unknown
Sources: S25 Capture symptom + applicability Observed / expected + cited sources Known reproduction state and limits Actual worktree, source/build/runtime IDs Unknowns stay unknown S25 Check real discriminators
Distinct mechanisms; no fixed hypothesis quota
CONFIRMED / REFUTED / INCONCLUSIVE
Capture limits constrain absent-event claims
New safe discriminator → bounded further check
Sources: S25 Check real discriminators Distinct mechanisms; no fixed hypothesis quota CONFIRMED / REFUTED / INCONCLUSIVE Capture limits constrain absent-event claims New safe discriminator → bounded further check S25 Choose one disposition
Retain multiple supported causes
Keep separate authority blockers
Missing access is not design
Non-reproduction is not closure
Sources: S25 Choose one disposition Retain multiple supported causes Keep separate authority blockers Missing access is not design Non-reproduction is not closure S25 confirmed_code_defect
Supported causal mechanism
Current expected behavior
Concrete regression obligation
Candidate fix, not repair consent
Sources: S25 confirmed_code_defect Supported causal mechanism Current expected behavior Concrete regression obligation Candidate fix, not repair consent S25 confirmed_noncode_cause
Supported external / operational cause
Existing owner, or owner unknown
Next evidence / access action
No forced source patch
Sources: S25 confirmed_noncode_cause Supported external / operational cause Existing owner, or owner unknown Next evidence / access action No forced source patch S25 design_question
Actual competing expectations
Maintainer’s intended-behavior choice
Optional later attributed ADR
No automatic ADR creation
Sources: S25 design_question Actual competing expectations Maintainer’s intended-behavior choice Optional later attributed ADR No automatic ADR creation S25 no_action
Positive applicable closure evidence
No remaining work in this scope
No default task or board note
No state mutation
Sources: S25 no_action Positive applicable closure evidence No remaining work in this scope No default task or board note No state mutation S25 unresolved
Known facts + missing discriminator
Owner and next action when known
Resume condition or prerequisite
No invented cause or no-action verdict
Sources: S25 unresolved Known facts + missing discriminator Owner and next action when known Resume condition or prerequisite No invented cause or no-action verdict S25 Transient debug-handoff/1.0.0
Symptom, expectation, evidence and limits
Source binding, causes and disposition
Regression oracle when a code defect exists
Measure UTF-8 size before first validation
Sources: S25 Transient debug-handoff/1.0.0 Symptom, expectation, evidence and limits Source binding, causes and disposition Regression oracle when a code defect exists Measure UTF-8 size before first validation S25 Installed helper: validate
Python 3 -B, absolute installed helper
Bounded packet bytes on stdin
Exit 0 + valid true + no errors → transfer
Invalid / unavailable → stop, no fallback
Sources: S25, S19 Installed helper: validate Python 3 -B, absolute installed helper Bounded packet bytes on stdin Exit 0 + valid true + no errors → transfer Invalid / unavailable → stop, no fallback S25 Recheck current evidence
Actual repository and worktree
Source/runtime identity and cited evidence
Commit + index + dirty working state
Relevant drift → revalidate or block
Sources: S19, S25 Recheck current evidence Actual repository and worktree Source/runtime identity and cited evidence Commit + index + dirty working state Relevant drift → revalidate or block S19 Check actual caller authority
Only code defect → fix is eligible
Independent repair permission required
Diagnosis-only stops at recommendation
Packet intent cannot authorize work
Sources: S19, S25 Check actual caller authority Only code defect → fix is eligible Independent repair permission required Diagnosis-only stops at recommendation Packet intent cannot authorize work S19 Stop or return the supported result
Invalid packet, unavailable helper, changed evidence, non-code/design/no-action/unresolved result, or absent repair authority: no dependent code repair.
Sources: S25, S19 Stop or return the supported result Invalid packet, unavailable helper, changed evidence, non-code/design/no-action/unresolved result, or absent repair authority: no dependent code repair. S25 Deliver context; prove target RED
Actual fix-author scoped input first
Named regression at the production layer
Fail for the defect’s precise reason
No feature artifact invented
Sources: S19, S27, S09 Deliver context; prove target RED Actual fix-author scoped input first Named regression at the production layer Fail for the defect’s precise reason No feature artifact invented S19 Minimum authorized fix
Cause and scope from current evidence
Change only after valid target RED
Then green, obligations, coverage, lint
No setup/import failure as a substitute
Sources: S19, S09 Minimum authorized fix Cause and scope from current evidence Change only after valid target RED Then green, obligations, coverage, lint No setup/import failure as a substitute S19 Current regression evidence
Read real output and exit codes
Preserve failing-before / passing-after
Original caller’s TDD gates remain
No acceptance by author assertion
Sources: S09, S19 Current regression evidence Read real output and exit codes Preserve failing-before / passing-after Original caller’s TDD gates remain No acceptance by author assertion S09 Separate delivery authority
Verified change is not a commit
/commit retains nine gates
PR, merge and release are separate
No automatic outward publication
Sources: S11, S04 Separate delivery authority Verified change is not a commit /commit retains nine gates PR, merge and release are separate No automatic outward publication S11 Resume only from available evidence
Validate retained bytes; recheck actual worktree, source, runtime and dirty-state applicability.
If context was lost, say so and reconstruct from accessible citations. Do not claim durable memory
or silently replace a malformed packet with an unvalidated repair.
Sources: S25, S19 Resume only from available evidence Validate retained bytes; recheck actual worktree, source, runtime and dirty-state applicability. If context was lost, say so and reconstruct from accessible citations. Do not claim durable memory or silently replace a malformed packet with an unvalidated repair. S25 Concrete follow-up: its own /task authority
A concrete follow-up uses the selected work owner under existing authorization.
The work.py facade routes native and legacy writes without a second board.
A design question may lead to an attributed ADR later, never by automatic conversion.
Sources: S25 Concrete follow-up: its own /task authority A concrete follow-up uses the selected work owner under existing authorization. The work.py facade routes native and legacy writes without a second board. A design question may lead to an attributed ADR later, never by automatic conversion. S25 Human summary: at most 400 words excluding short locators. Packet: at most 65,536 UTF-8 bytes. Neither is a new persistent case store. Read the map
Debug · evidence, dispositions and safe handoff Investigation is read-only. A validated packet carries evidence; the original caller still owns permission to act.
Execution Context / data Conditional / authority Return / repeat Stop / refusal
Choose an entry to highlight its route. Select a node for context, related entries and pinned sources.
Start with a route /feature /fix /sprint /init /review /tribunal /pr /reconcile
Historical route snapshot at 9496cff6332fe0b97195b7ebe67b73f62f30e5b3. Current assistive source bindings identify inspected implementation bytes, with 3549ed56f5d5849116aef8d5e946423d8ee677fc as the committed source base. Live host qualification and release are separate evidence. Curated current-source workflow contracts and selected context-delivery implementation; not a release, installed-host qualification, or exhaustive runtime call graph. Original source links and spatial positions retained; task labels show the selected work owner. Source compatibility revalidated at 25980c292b529fca90c39f5f26d69a1c0674de44. Unittest collector compatibility bound to reviewed blob 76a2998f4bd5b6f441f5fd162bbe810b2de2ee93.
Snapshot limits Historical snapshot: New full-lane feature/sprint defaults to typed HTML only with an admitted installed payload and qualified producers. Main source checkout does not carry the release binary.
Historical snapshot: Pi fails new HTML workflow admission in the pinned artifacts contract. Existing Markdown pairs and inline small features remain separate.
Historical snapshot: HTML --farm dispatch is blocked. Legacy Markdown farm is an off-by-default preview, not the blessed default.
Historical snapshot: Read-time approved-spec pointers index Markdown only in the inspected implementation. Typed HTML execution reads are a distinct engine path.
Historical snapshot: Shared surface generation is not live host parity. Claude/Pi READ notices and Codex’s absence of a native READ event differ.
Historical snapshot: Task quality review is once per current combined scope after per-task spec proof and fresh verification, not a duplicate full fleet for every tiny task.
Historical snapshot: The diagrams describe default arbiter-mode workflow contracts; deterministic mode tokens are separate posture controls and not new slash commands.
Historical snapshot: Snapshots of narrative instructions do not prove every gate is mechanically enforced. No live workflows were executed for this atlas.
Historical snapshot: Brownfield context dispatch needs an actually qualified read-only caller profile; the generic scout charter and generated plugin surfaces alone do not qualify it. Repository-only preparation fixtures are not installed capability.
Historical snapshot: Five content/provenance pairs use independently approved native transactions. Initial finalization has its own 13-identity approval. Initialized refresh retains the marker and never re-enters finalization.
Historical snapshot: Task-context selection is inert. The caller delivers a separate bounded input to each actor; reported delivery, engine tickets and parent session markers are not interchangeable authority or proof.
Historical snapshot: Debug has five dispositions. Validation alone is not evidence truth or repair authority. Diagnosis-only requests stop after the recommendation.
Historical snapshot: Source wording ambiguity: brownfield Phase 3 preserves source-scoped upstream ADR references, while a retained Phase 5 sentence asks for local summaries. Local-summary semantics and authority need clarification; this atlas does not infer automatic local decision authority.
Historical snapshot: Pi approval is currently DEGRADED because the adapter is not implemented; the newer runtime input event means the older HOST-IMPOSSIBLE explanation must not be repeated.
Historical snapshot: Tribunal is opt-in. Its source-bound run store is separate from typed spec/plan authority; report-written does not establish terminal completion.
Historical snapshot: Ten new commits include correctness and waiting-limit repairs, not ten new routes. Their merge descriptions are not independent runtime proof. No latency improvement is asserted.
Historical snapshot: The release PR boundary was under-specified in the previous atlas; making it explicit is a carried presentation correction, not a behavior newly introduced by these ten commits.
Historical snapshot: The native page and offline editions share this source-bound atlas. Website integration does not establish released behavior, installed-host qualification or completed workflow execution.
Tribunal: source-bound audit to completed follow-up An opt-in review route, not a required gate. Candidate findings, verification, report and completion are distinct. An agreed outcome becomes a maintained plan, direct or delegated implementation, proportional verification and an honest delivery report. Task context and dependency advice support progress; real user stops and host permissions remain effective. Greenfield and brownfield population use the paired native writer and exact finalization. Exact blob bindings identify inspected source bytes; the linked revision is the committed source base. Implementation, verification, live host qualification and release remain separate claims. Source 9496cff6332fe0b97195b7ebe67b73f62f30e5b3. Select a node for details or use the complete text path. HISTORICAL ROUTES / CODEARBITER Tribunal: source-bound audit to completed follow-up An opt-in review route, not a required gate. Candidate findings, verification, report and completion are distinct. SOURCE 9496cff6332f / 2026-10-07 01 / BIND THE RUN BEFORE DISPATCH Declared finding scope is not the evidence scope. Source identity, not age, governs reuse. 02 / SELECT EVIDENCE AND REVIEW THE APPLICABLE LENSES The card inventory is the thirteen-lens roster. Skip rationale and coverage limitations remain visible. 03 / VERIFY SERIOUS CLAIMS BEFORE KEEPING WORK A fresh MODE verify assignment is an attempted disproof. Its actual independence and result still matter. 04 / WRITE THE REPORT, THEN RESOLVE FOLLOW-UPS report-written enters follow-up. It neither files issues nor ends the run. 05 / COMPLETE ONLY AFTER BOTH DISPOSITIONS Source checks remain applicable. Explicit skipping is different from an unanswered question or failed action. Execution Conditional / authority: fresh fresh Execution Stop / refusal Conditional / authority: unchanged bound run unchanged bound run Context / data Execution Execution Conditional / authority: verify verify Conditional / authority: supported supported Execution Stop / refusal: refuted refuted Conditional / authority: unproven unproven Conditional / authority: selected selected Conditional / authority: separate separate Execution Execution: fresh work fresh work Return / repeat: resume unfinished audit resume unfinished audit Execution: per recorded wave per recorded wave Execution Conditional / authority: resolved follow-ups resolved follow-ups Conditional / authority: unanswered / failure unanswered / failure Conditional / authority: other candidates other candidates Conditional / authority: genuine design fork genuine design fork Context / data Context / data Context / data Execution: all recorded waves accounted for all recorded waves accounted for /tribunal · deliberate audit
Resolve scope and trusted bundle
Inspect matching run directories
No source text grants execution
Sources: S29 /tribunal · deliberate audit Resolve scope and trusted bundle Inspect matching run directories No source text grants execution S29 Check source-bound status
Same source: audit or follow-up
Drift / unbound / invalid: no reuse
Terminal or new scope: fresh run
Sources: S29 Check source-bound status Same source: audit or follow-up Drift / unbound / invalid: no reuse Terminal or new scope: fresh run S29 Inventory → profiles → cost
Mechanical facts before model work
Bound concrete evidence inputs
User acknowledges cost + settings
Sources: S29 Inventory → profiles → cost Mechanical facts before model work Bound concrete evidence inputs User acknowledges cost + settings S29 Allocate a fresh bound run
start returns run_id + run_dir
Persist source.json before reviews
Record actual budget and capacity
Sources: S29 Allocate a fresh bound run start returns run_id + run_dir Persist source.json before reviews Record actual budget and capacity S29 Retain history; do not relabel
Changed evidence needs a fresh binding.
A deliberate abort records run-aborted.
Sources: S29 Retain history; do not relabel Changed evidence needs a fresh binding. A deliberate abort records run-aborted. S29 Unchanged-source resume
audit → unfinished recorded waves
follow-up → pending dispositions only
Sources: S29 Unchanged-source resume audit → unfinished recorded waves follow-up → pending dispositions only S29 Inventory ≠ risk judgment
inventory.json + readable projection
Risk map adds contract judgment
Optional mappers answer only gaps
No repository scripts run by discovery
Sources: S29 Inventory ≠ risk judgment inventory.json + readable projection Risk map adds contract judgment Optional mappers answer only gaps No repository scripts run by discovery S29 Applicability + bounded packets
Account for every lens card
Declare finding/evidence scopes
Only bound evidence may expand
Recheck source before dispatch
Sources: S29 Applicability + bounded packets Account for every lens card Declare finding/evidence scopes Only bound evidence may expand Recheck source before dispatch S29 Generic reviewer · MODE review
One assignment per active lens
At most 5 concurrent; host-bounded
Own findings / pending-leads files
No child dispatch or shared-log edits
Sources: S29 Generic reviewer · MODE review One assignment per active lens At most 5 concurrent; host-bounded Own findings / pending-leads files No child dispatch or shared-log edits S29 Persist and account for output
Import leads through the helper
Coordinator records actual events
Check source before accepting output
Missing output remains incomplete
Sources: S29 Persist and account for output Import leads through the helper Coordinator records actual events Check source before accepting output Missing output remains incomplete S29 Root-cause triage candidates
Preserve provenance and old keys
Calibrate severity + confidence
Retain serious counterarguments
Sources: S29 Root-cause triage candidates Preserve provenance and old keys Calibrate severity + confidence Retain serious counterarguments S29 Fresh reviewer · MODE verify
All serious / promoted serious claims
Expensive inferential recommendations
Attempt disproof with needed evidence
Sources: S29 Fresh reviewer · MODE verify All serious / promoted serious claims Expensive inferential recommendations Attempt disproof with needed evidence S29 Eligibility + confidence gate
Independent confirmed / narrowed
Keep only the supported surviving claim
Helper refusal cannot be log-edited away
Sources: S29 Eligibility + confidence gate Independent confirmed / narrowed Keep only the supported surviving claim Helper refusal cannot be log-edited away S29 Eligible phase plan
keep / combine only
Preserve narrowed claims
Emit wave-triaged after plan
Sources: S29 Eligible phase plan keep / combine only Preserve narrowed claims Emit wave-triaged after plan S29 Refuted
Preserve result and rationale
Not a confirmed issue
Not kept repair work
Sources: S29 Refuted Preserve result and rationale Not a confirmed issue Not kept repair work S29 Inconclusive / shared context
Record verify-required
Keep investigation work separate
Do not invent independent confirmation
Sources: S29 Inconclusive / shared context Record verify-required Keep investigation work separate Do not invent independent confirmation S29 Genuine design fork
decision-required, not a disguised
factual uncertainty or automatic ADR
Leads need explicit dispositions
Sources: S29 Genuine design fork decision-required, not a disguised factual uncertainty or automatic ADR Leads need explicit dispositions S29 Report + manifest projections
Source binding + actual coverage
Supported work and unresolved claims
Lead and follow-up dispositions
report-written → follow-up
Sources: S29 Report + manifest projections Source binding + actual coverage Supported work and unresolved claims Lead and follow-up dispositions report-written → follow-up S29 Issue filing disposition
Dedup against actual existing / partial receipts
Default: show selected eligible issue commands
Execute only with explicit filing authorization
issues-filed OR explicit filing-skipped
Sources: S29 Issue filing disposition Dedup against actual existing / partial receipts Default: show selected eligible issue commands Execute only with explicit filing authorization issues-filed OR explicit filing-skipped S29 Separate optional public telemetry
Preview the full bounded aggregate payload
No code / paths / source hashes / finding text
Explicit per-run consent before sending
telemetry-sent OR explicit telemetry-skipped
Sources: S29 Separate optional public telemetry Preview the full bounded aggregate payload No code / paths / source hashes / finding text Explicit per-run consent before sending telemetry-sent OR explicit telemetry-skipped S29 Unanswered or failed follow-up
Keep pending; preserve real partial receipts
On resume, check current source again
Do not rerun completed lenses by default
Sources: S29 Unanswered or failed follow-up Keep pending; preserve real partial receipts On resume, check current source again Do not rerun completed lenses by default S29 Both dispositions resolved
Filing done or explicitly skipped
Telemetry done or explicitly skipped
Source-bound follow-up rules still apply
Sources: S29 Both dispositions resolved Filing done or explicitly skipped Telemetry done or explicitly skipped Source-bound follow-up rules still apply S29 run-completed
Terminal signal, not just a written report
Audit history stays durable
No automatic repair, commit, merge or release
Sources: S29 run-completed Terminal signal, not just a written report Audit history stays durable No automatic repair, commit, merge or release S29 Account for every outcome
Eligible work + unsupported claims
Verification / design / investigations
All leads have explicit dispositions
Sources: S29, S30 Account for every outcome Eligible work + unsupported claims Verification / design / investigations All leads have explicit dispositions S29 Resume by source binding; follow-up resumes at panel 04. Long return paths are reading aids, not additional command invocations. Source contract and helper states inspected. No new model-backed audit, installed-host certification or accuracy claim is made here. Read the map
Tribunal: source-bound audit to completed follow-up An opt-in review route, not a required gate. Candidate findings, verification, report and completion are distinct.
Execution Context / data Conditional / authority Return / repeat Stop / refusal
Choose an entry to highlight its route. Select a node for context, related entries and pinned sources.
Start with a route /feature /fix /sprint /init /review /tribunal /pr /reconcile
Historical route snapshot at 9496cff6332fe0b97195b7ebe67b73f62f30e5b3. Current assistive source bindings identify inspected implementation bytes, with 3549ed56f5d5849116aef8d5e946423d8ee677fc as the committed source base. Live host qualification and release are separate evidence. Curated current-source workflow contracts and selected context-delivery implementation; not a release, installed-host qualification, or exhaustive runtime call graph. Original source links and spatial positions retained; task labels show the selected work owner. Source compatibility revalidated at 25980c292b529fca90c39f5f26d69a1c0674de44. Unittest collector compatibility bound to reviewed blob 76a2998f4bd5b6f441f5fd162bbe810b2de2ee93.
Snapshot limits Historical snapshot: New full-lane feature/sprint defaults to typed HTML only with an admitted installed payload and qualified producers. Main source checkout does not carry the release binary.
Historical snapshot: Pi fails new HTML workflow admission in the pinned artifacts contract. Existing Markdown pairs and inline small features remain separate.
Historical snapshot: HTML --farm dispatch is blocked. Legacy Markdown farm is an off-by-default preview, not the blessed default.
Historical snapshot: Read-time approved-spec pointers index Markdown only in the inspected implementation. Typed HTML execution reads are a distinct engine path.
Historical snapshot: Shared surface generation is not live host parity. Claude/Pi READ notices and Codex’s absence of a native READ event differ.
Historical snapshot: Task quality review is once per current combined scope after per-task spec proof and fresh verification, not a duplicate full fleet for every tiny task.
Historical snapshot: The diagrams describe default arbiter-mode workflow contracts; deterministic mode tokens are separate posture controls and not new slash commands.
Historical snapshot: Snapshots of narrative instructions do not prove every gate is mechanically enforced. No live workflows were executed for this atlas.
Historical snapshot: Brownfield context dispatch needs an actually qualified read-only caller profile; the generic scout charter and generated plugin surfaces alone do not qualify it. Repository-only preparation fixtures are not installed capability.
Historical snapshot: Five content/provenance pairs use independently approved native transactions. Initial finalization has its own 13-identity approval. Initialized refresh retains the marker and never re-enters finalization.
Historical snapshot: Task-context selection is inert. The caller delivers a separate bounded input to each actor; reported delivery, engine tickets and parent session markers are not interchangeable authority or proof.
Historical snapshot: Debug has five dispositions. Validation alone is not evidence truth or repair authority. Diagnosis-only requests stop after the recommendation.
Historical snapshot: Source wording ambiguity: brownfield Phase 3 preserves source-scoped upstream ADR references, while a retained Phase 5 sentence asks for local summaries. Local-summary semantics and authority need clarification; this atlas does not infer automatic local decision authority.
Historical snapshot: Pi approval is currently DEGRADED because the adapter is not implemented; the newer runtime input event means the older HOST-IMPOSSIBLE explanation must not be repeated.
Historical snapshot: Tribunal is opt-in. Its source-bound run store is separate from typed spec/plan authority; report-written does not establish terminal completion.
Historical snapshot: Ten new commits include correctness and waiting-limit repairs, not ten new routes. Their merge descriptions are not independent runtime proof. No latency improvement is asserted.
Historical snapshot: The release PR boundary was under-specified in the previous atlas; making it explicit is a carried presentation correction, not a behavior newly introduced by these ten commits.
Historical snapshot: The native page and offline editions share this source-bound atlas. Website integration does not establish released behavior, installed-host qualification or completed workflow execution.
Change / core
/add-dep dependency-reviewer / ephemeral-tool branch
Route to outcome Distinguish durable adoption from one pinned ephemeral run Adoption: review license, provenance, maintenance and supply chain Ephemeral: exact version + approved registry + confirmation, no manifest/lock change Return to requesting workflow; behavior adoption remains feature/fix work Selected context tech-stack.md and security-controls.md
Package identity, purpose and provenance
Output Vetted dependency decision/adoption, or a verified one-off tool result.
Denied license or supply-chain concern blocks. Ephemeral execution must not silently become adoption.
Host / scope Claude Code / Codex / Pi; admission and runtime capabilities still apply
claude /ca:add-dep
codex $ca-add-dep
pi /ca-add-depRead the exact procedure
Related entries /chore /feature /fix
Pinned evidence S02 · core/surface/COMMANDS.mdCatalog, modes, host-limited commands and purpose of each public route. S04 · core/surface/includes/routing-table.mdAuthoritative intent-to-owner table, scope-touch routes and stop/return semantics. S05 · core/surface/includes/reference-map.mdScope-to-governing-document and specialist mapping. Decide / core
/adr decision-lifecycle — author
Route to outcome Read current decision context Present the specific decision and alternatives Obtain explicit user choice and attribution Write numbered ADR through its owner Selected context Project risks, open questions and related ADRs
User decision and evidence
Output Durable, attributed architecture decision; not proof of implementation.
Never guess a CONFIRM answer. Accepted/Planned is not delivery evidence.
Host / scope Claude Code / Codex / Pi; admission and runtime capabilities still apply
claude /ca:adr
codex $ca-adr
pi /ca-adrRead the exact procedure
Related entries /adr-status /reconcile
Pinned evidence S04 · core/surface/includes/routing-table.mdAuthoritative intent-to-owner table, scope-touch routes and stop/return semantics. S21 · core/surface/skills/INDEX.mdCurrent routine-owner index: five debug dispositions, bounded workflow owners and unchanged six-phase TDD / nine-phase commit sequence. Decide / advanced
/adr-status decision-lifecycle — read-only
Route to outcome Select status mode before authoring prerequisites Inspect age, challenges and supersession chains Return ADR health report Selected context Current ADR records and relationships
Output Read-only ADR health/status report.
Status inspection grants no writing authority and is not delivery proof.
Host / scope Claude Code / Codex / Pi; admission and runtime capabilities still apply
claude /ca:adr-status
codex $ca-adr-status
pi /ca-adr-statusModes --adr N
Read the exact procedure
Related entries /adr /reconcile
Pinned evidence S02 · core/surface/COMMANDS.mdCatalog, modes, host-limited commands and purpose of each public route. S21 · core/surface/skills/INDEX.mdCurrent routine-owner index: five debug dispositions, bounded workflow owners and unchanged six-phase TDD / nine-phase commit sequence. Operate / advanced
/audit read-only governance packet
Route to outcome Select requested window Read source governance records Assemble a dated, non-overwriting packet Selected context Source audit logs and decision records
Output Dated governance evidence packet; source records unchanged.
Preserve quoted audit lines; never overwrite an existing packet.
Host / scope Claude Code / Codex / Pi; admission and runtime capabilities still apply
claude /ca:audit
codex $ca-audit
pi /ca-auditModes [range]
Read the exact procedure
Related entries /metrics /checkpoint
Pinned evidence S02 · core/surface/COMMANDS.mdCatalog, modes, host-limited commands and purpose of each public route. S04 · core/surface/includes/routing-table.mdAuthoritative intent-to-owner table, scope-touch routes and stop/return semantics. Help / deprecated
/btw direct lightweight question
Route to outcome Answer the project question without mutation Return to the caller Selected context Only evidence needed for the question
Output Answer only.
Prefer asking directly; a question is not mutation authority.
Host / scope Claude Code / Codex / Pi; admission and runtime capabilities still apply
claude Not exposed on this host
codex Not exposed on this host
pi Not exposed on this host Removed on 2026-10-23; no host exposes this route.
Related entries
Pinned evidence S01 · core/surface/command-routes.jsonAll 37 catalog entries; core, advanced, alias, internal and deprecated visibility. S02 · core/surface/COMMANDS.mdCatalog, modes, host-limited commands and purpose of each public route. S03 · core/surface/arbiter.mdIntent-to-owner routing, on-demand loading, caller extent and mode boundary. Review / advanced
/checkpoint dispatching-parallel-agents → checkpoint-aggregator
Route to outcome Periodic whole-codebase reviewer sweep Read-only findings → triage → verdict Explicit checkpoint-aggregator persistence Selected context Whole-codebase scope and governing ADRs
Complete triaged reviewer reports
Output Dated, non-overwriting .codearbiter/checkpoints/ report.
An informational checkpoint is not a promotion/merge gate.
Host / scope Claude Code / Codex / Pi; admission and runtime capabilities still apply
claude /ca:checkpoint
codex $ca-checkpoint
pi /ca-checkpointRead the exact procedure
Related entries /review /audit
Pinned evidence S04 · core/surface/includes/routing-table.mdAuthoritative intent-to-owner table, scope-touch routes and stop/return semantics. S21 · core/surface/skills/INDEX.mdCurrent routine-owner index: five debug dispositions, bounded workflow owners and unchanged six-phase TDD / nine-phase commit sequence. S22 · core/surface/agents/INDEX.mdRole separation; specialist reviewers; restricted scouts/graders; authority-reviewer host restriction. Change / core
/chore chore — type-scaled command
Route to outcome Choose docs, deps or revert Docs: prose, copy review, secrets and diff checks; no TDD for prose Deps: manifest + lockfile, dependency review, affected tests Revert: named git revert, affected tests and referenced reason Commit-gate → finishing-a-development-branch Selected context Requested non-behavioral scope
Project verification commands
Dependency/security policy when relevant
Output Reviewed prose, dependency bump or named revert, then authorized commit/PR.
Behavioral adoption redirects to feature/fix; no behavioral work hidden inside a chore.
Host / scope Claude Code / Codex / Pi; admission and runtime capabilities still apply
claude /ca:chore
codex $ca-chore
pi /ca-choreModes docs · deps · revert
Read the exact procedure
Related entries /add-dep /feature /fix /commit
Pinned evidence S20 · core/surface/commands/chore.mdDocs/deps/revert branches; behavioral adoption returns to feature/fix. S11 · core/surface/skills/commit-gate/SKILL.mdCurrent Phase 5.5 through terminal commit rules re-read. Legacy heal vocabulary must not be generalized to v2 paired native writes. S12 · core/surface/skills/finishing-a-development-branch/SKILL.mdWatch/cleanup early routing; commit and acceptance preflight; PR and merge authority. Ship / alias
/cleanup post-merge-cleanup
Canonical entry: pr --cleanup
Route to outcome Prove merge containment against fetched default Classify residue Confirm each discard Fast-forward-only transition; preserve unique work and remote branches Selected context Same context contract as pr --cleanup
Output Proven post-merge local cleanup.
Retained compatibility entry. Prefer pr --cleanup. Owner gates remain unchanged.
Host / scope Claude Code / Codex / Pi; admission and runtime capabilities still apply
claude /ca:cleanup
codex $ca-cleanup
pi /ca-cleanupRead the exact procedure
Related entries /pr
Pinned evidence S01 · core/surface/command-routes.jsonAll 37 catalog entries; core, advanced, alias, internal and deprecated visibility. S02 · core/surface/COMMANDS.mdCatalog, modes, host-limited commands and purpose of each public route. S04 · core/surface/includes/routing-table.mdAuthoritative intent-to-owner table, scope-touch routes and stop/return semantics. S21 · core/surface/skills/INDEX.mdCurrent routine-owner index: five debug dispositions, bounded workflow owners and unchanged six-phase TDD / nine-phase commit sequence. Help / advanced
/commands grouped command catalog
Route to outcome Read grouped catalog Show canonical, advanced and retained routes Selected context COMMANDS.md and visibility metadata
Output Command help; no work or state mutation.
Descriptions discover owners; they do not replace full procedures.
Host / scope Claude Code / Codex / Pi; admission and runtime capabilities still apply
claude Not exposed on this host
codex Not exposed on this host
pi Not exposed on this host Removed on 2026-10-23; no host exposes this route.
Related entries
Pinned evidence S01 · core/surface/command-routes.jsonAll 37 catalog entries; core, advanced, alias, internal and deprecated visibility. S02 · core/surface/COMMANDS.mdCatalog, modes, host-limited commands and purpose of each public route. S03 · core/surface/arbiter.mdIntent-to-owner routing, on-demand loading, caller extent and mode boundary. Ship / core
/commit commit-gate
Route to outcome Explicit permission → safe branch → homogeneous classification Focused tests/lint/secrets and specialized gates Fresh behavior proof and current HTML acceptance where applicable Conditional provenance heal → diff review → follow-up harvest Selective staging → Conventional Commit message → local commit Selected context Exact staged diff and declared verification commands
TDD proof and current acceptance receipts
Provenance and scope-specific policy
Output Local commit SHA, explicit paths and gate report. Not automatically a push or PR.
Nine gates; no blanket staging; no protected-branch commit; stale acceptance blocks before staging. Legacy v1 Phase 5.5 heal is not permission to raw-write enrolled v2 document/provenance pairs.
Host / scope Claude Code / Codex / Pi; admission and runtime capabilities still apply
claude /ca:commit
codex $ca-commit
pi /ca-commitRead the exact procedure
Related entries /pr /release
Pinned evidence S11 · core/surface/skills/commit-gate/SKILL.mdCurrent Phase 5.5 through terminal commit rules re-read. Legacy heal vocabulary must not be generalized to v2 paired native writes. S10 · core/surface/includes/artifacts.mdInstalled admission, typed reads, context completion, authority receipts and format boundaries. S26 · core/surface/skills/context-check/SKILL.mdInitialized refresh: read-only report, explicit selection, v1/v2 separation, native paired updates and preserved human history. Full owner read. Decide / internal
/conflict conflict protocol
Route to outcome Detect contradiction between governing rules, documents and code Stop other work Surface both sides and wait for the user Selected context Conflicting authorities and exact evidence
Output Visible stopped state and a specific user decision request.
Ambiguous reconcile-vs-conflict resolves to stopping, not silent continuation.
Host / scope Claude Code / Codex / Pi; admission and runtime capabilities still apply
claude /ca:conflict
codex $ca-conflict
pi /ca-conflictRead the exact procedure
Related entries /reconcile /adr
Pinned evidence S03 · core/surface/arbiter.mdIntent-to-owner routing, on-demand loading, caller extent and mode boundary. S04 · core/surface/includes/routing-table.mdAuthoritative intent-to-owner table, scope-touch routes and stop/return semantics. Operate / alias
/context-check context-check
Canonical entry: status drift
Route to outcome Default status and initial drift report are read-only Assess legacy v1 drift and each enrolled v2 identity/coverage state separately; unknown is not fresh Wait for explicit selected documents and affected scope, or explicit full refresh; defer changes nothing Re-scout only the selected assignments through qualified containment and a bounded current join For v2, approve the exact owned-field document/provenance preview and apply only its native receipt Read back committed pairs; retain the initialized marker, unselected human content and separate task/question/audit history; do not stage or commit Selected context Current bounded source/target snapshots and action-boundary checks
v1 drift separately from v2 current/stale/unknown coverage; identity matching is not semantic proof
Explicit document selection, actual field ownership, qualified scouts and native writer capability
Output Read-only state/drift report, or explicitly selected native context updates with declared full/incremental coverage.
Report alone is not refresh consent. No reinitialization or marker removal. No legacy raw rebaseline fallback for v2, no unsupported human-layout overwrite, no staging or commit.
Host / scope Claude Code / Codex / Pi; admission and runtime capabilities still apply
claude /ca:context-check
codex $ca-context-check
pi /ca-context-checkRead the exact procedure
Related entries /init /status /commit
Pinned evidence S26 · core/surface/skills/context-check/SKILL.mdInitialized refresh: read-only report, explicit selection, v1/v2 separation, native paired updates and preserved human history. Full owner read. S13 · core/surface/skills/context-creation/SKILL.mdCurrent brownfield owner: qualified containment, bounded reports and join, scoped gaps, native document/provenance writes, separate 13-identity finalization and recovery. S23 · core/surface/commands/init.mdCurrent init owner: activation/scaffold versus initial population; initialized refresh belongs to status drift, not reinitialization. Knowledge / alias
/create-context context-creation
Canonical entry: init --brownfield
Route to outcome Uninitialized project with meaningful source → bounded brownfield initial population Brownfield: qualified read-only scout profile → closed reports → scoped synthesis and relevant gap decisions Approve each native document/provenance preview; receipt-only apply, then separately approved finalization Finalize the 13 required identities through its own native preview/receipt; recover uncertain operations by the same operation ID Selected context Same bounded brownfield source, containment, report, ownership and provenance contract as init --brownfield
Output Five current content/provenance pairs and separately finalized initialized state, or a precise capability/evidence conflict.
Initial population only; existing initialized state routes to status drift. Generic scout role or a source-only preparation fixture does not establish qualified containment or writer admission.
Host / scope Claude Code / Codex / Pi; admission and runtime capabilities still apply
claude /ca:create-context
codex $ca-create-context
pi /ca-create-contextRead the exact procedure
Related entries /init /status
Pinned evidence S13 · core/surface/skills/context-creation/SKILL.mdCurrent brownfield owner: qualified containment, bounded reports and join, scoped gaps, native document/provenance writes, separate 13-identity finalization and recovery. S23 · core/surface/commands/init.mdCurrent init owner: activation/scaffold versus initial population; initialized refresh belongs to status drift, not reinitialization. S26 · core/surface/skills/context-check/SKILL.mdInitialized refresh: read-only report, explicit selection, v1/v2 separation, native paired updates and preserved human history. Full owner read. Change / advanced
/debug debug
Route to outcome Capture the cited symptom, expectation, reproduction state and actual request scope Test evidence-backed mechanisms using safe discriminators; preserve inconclusive observations and capture limits Choose exactly one disposition: confirmed_code_defect, confirmed_noncode_cause, design_question, no_action or unresolved For transfer, validate the bounded transient debug-handoff/1.0.0 packet with the installed helper Return the diagnosis and truthful continuation; only an already-authorized fix caller can continue a qualifying code-defect handoff Selected context Available project instructions and scoped source evidence, not compulsory new onboarding
Observed versus expected behavior, reproduction status, supported causes and regression oracle
Relevant worktree, source/build/runtime identity, dirty state, limitations and independent caller authority
Output One evidence-backed disposition and a bounded summary; transient validated handoff when available. No default persistent case store, task or ADR.
Diagnosis never authorizes repair. Invalid/unavailable handoff validation stops transfer; recheck current evidence. No-action requires positive closure evidence. A design question is not an automatic ADR.
Host / scope Claude Code / Codex / Pi; admission and runtime capabilities still apply
claude /ca:debug
codex $ca-debug
pi /ca-debugRead the exact procedure
Related entries /fix /task /adr
Pinned evidence S25 · core/surface/skills/debug/SKILL.mdFive-disposition diagnostic owner; transient validated handoff, freshness and authority boundaries. Full owner read. S19 · core/surface/commands/fix.mdCurrent full fix owner: validated handoff, independent caller authority, evidence revalidation, selected actor input and genuine target-red gate. S27 · core/pysrc/_contextselectlib.pyInert scoped selector and delivery implementation: commands/cwd/evidence, native instruction order, actor/worktree/epoch binding, two-attempt limit and reviewer preflight. S04 · core/surface/includes/routing-table.mdAuthoritative intent-to-owner table, scope-touch routes and stop/return semantics. Knowledge / alias
/decompose decompose
Canonical entry: init --greenfield
Route to outcome Persist six interview layers and draft ADRs Synthesize architecture, phased plan and backlog User reviews the three records Populate project context and lock initialization Selected context Same context contract as init --greenfield
Output Approved greenfield planning records and initialized project knowledge.
Retained compatibility entry. Prefer init --greenfield. Owner gates remain unchanged.
Host / scope Claude Code / Codex / Pi; admission and runtime capabilities still apply
claude /ca:decompose
codex $ca-decompose
pi /ca-decomposeRead the exact procedure
Related entries /init
Pinned evidence S01 · core/surface/command-routes.jsonAll 37 catalog entries; core, advanced, alias, internal and deprecated visibility. S02 · core/surface/COMMANDS.mdCatalog, modes, host-limited commands and purpose of each public route. S04 · core/surface/includes/routing-table.mdAuthoritative intent-to-owner table, scope-touch routes and stop/return semantics. S21 · core/surface/skills/INDEX.mdCurrent routine-owner index: five debug dispositions, bounded workflow owners and unchanged six-phase TDD / nine-phase commit sequence. Operate / core
/doctor installation diagnostics
Route to outcome Inspect active install and payload Verify ownership/enforcement configuration Run harmless live-fire probe Report capability gaps Selected context Installed host and package resources
Output Installation and enforcement diagnostic report.
Diagnostic success is not completed end-to-end workflow qualification.
Host / scope Claude Code / Codex / Pi; admission and runtime capabilities still apply
claude /ca:doctor
codex $ca-doctor
pi /ca-doctorRead the exact procedure
Related entries /init /status
Pinned evidence S02 · core/surface/COMMANDS.mdCatalog, modes, host-limited commands and purpose of each public route. S10 · core/surface/includes/artifacts.mdInstalled admission, typed reads, context completion, authority receipts and format boundaries. Change / core
/feature brainstorming → writing-plans → executing-plans
Route to outcome Discover an existing exact-format pair and resume before new triage Small: confirm 1–3 criteria and log triage, then TDD Full: approve specification, bind/approve plan, then attended batches Fresh task authors, spec proof, scope-quality acceptance Before each author or independent reviewer acts, compose its own current scoped input and perform bounded actor delivery; a coordinator read is not receipt Current acceptance → authorized commit → selected PR disposition Selected context Project purpose, stage and code-map
Scope-specific rules and recorded ADR intent
Approved criteria and exact plan binding
Task paths, dependencies, verification and complete context ticket
Current actor-scoped selection: map/provenance, command cwd and observed evidence, constraints and host-effective instructions
Actual author and independent reviewer delivery bound to task/worktree/source/host epochs; reselect after resume, compaction or scope change
Use exact engine-returned IDs, not assumed SPEC-/PLAN- prefixes. Approved prerequisites use one visible 24-hour request/reply deadline.
Output Full lane: typed spec + plan, accepted code/tests, evidence and selected PR outcome. Small lane: inline mini-spec, no spec artifact.
No guessed unknowns; no implementation before red; batch checkpoints; no mixed-format resume; installed HTML admission required. Selected context is orientation, not task admission. Small features retain their confirmed inline mini-spec without invented artifact IDs.
Host / scope Claude Code / Codex: admitted installed typed workflow only. Pi 1.0.2 is the current runtime target, but its native approval adapter remains unimplemented (DEGRADED); legacy Markdown pairs and inline small-feature work have separate contracts.
claude /ca:feature
codex $ca-feature
pi /ca-featureRead the exact procedure
Related entries /sprint /fix /commit /pr
Pinned evidence S06 · core/surface/commands/feature.mdCurrent feature owner: same-format resume and full/small gates; adds actual per-actor context composition. S08 · core/surface/skills/subagent-driven-development/SKILL.mdCurrent implementation-dispatch and independent spec-review sections re-read; remaining scope contract also checked against current skill index. Fresh actor input and bounded delivery are distinct from engine tickets. S09 · core/surface/skills/tdd/SKILL.mdCurrent TDD context preflight and obligation derivation re-read. Six-phase gate sequence also confirmed by current skill index; fix and test-only paths do not fabricate feature artifacts. S10 · core/surface/includes/artifacts.mdInstalled admission, typed reads, context completion, authority receipts and format boundaries. S11 · core/surface/skills/commit-gate/SKILL.mdCurrent Phase 5.5 through terminal commit rules re-read. Legacy heal vocabulary must not be generalized to v2 paired native writes. S12 · core/surface/skills/finishing-a-development-branch/SKILL.mdWatch/cleanup early routing; commit and acceptance preflight; PR and merge authority. S27 · core/pysrc/_contextselectlib.pyInert scoped selector and delivery implementation: commands/cwd/evidence, native instruction order, actor/worktree/epoch binding, two-attempt limit and reviewer preflight. S31 · docs/parity.mdCurrent Pi 1.0.2 target, separate host counts, DEGRADED typed approval and Codex farm exception. S32 · core/pysrc/_artifactauthoritylib.pyNative artifact ID grammar, producer states, Claude join lock, UTF-8/UTF-16-key-order observation binding. Change / core
/fix tdd — bug variant
Route to outcome Establish the confirmed defect from current reproduction or supported source/contract causal evidence If needed, use one bounded internal debug prerequisite and return to this same authorized fix caller Validate any available handoff; independently recheck caller authority and actual worktree/source/runtime applicability Select and deliver bounded current context to the actual fix author Observe the named regression red for the correct defect, then implement the minimum fix and complete TDD Return proof; commit, PR and other follow-ups retain their separate authority Selected context Current expectation, causal evidence, reproduction limits and named regression oracle
Task paths; scoped map/provenance; command cwd and current collector evidence
Applicable constraints and native effective instructions; actor/worktree/source/host epoch binding
Output Minimal authorized correction with genuine failing-before / passing-after regression evidence.
No repair from packet intent alone. Only a validated confirmed_code_defect → fix handoff can resume an authorized repair. Setup/import/unrelated red is not target-red proof. No feature artifact is invented.
Host / scope Claude Code / Codex / Pi; admission and runtime capabilities still apply
claude /ca:fix
codex $ca-fix
pi /ca-fixRead the exact procedure
Related entries /debug /task /commit
Pinned evidence S19 · core/surface/commands/fix.mdCurrent full fix owner: validated handoff, independent caller authority, evidence revalidation, selected actor input and genuine target-red gate. S25 · core/surface/skills/debug/SKILL.mdFive-disposition diagnostic owner; transient validated handoff, freshness and authority boundaries. Full owner read. S27 · core/pysrc/_contextselectlib.pyInert scoped selector and delivery implementation: commands/cwd/evidence, native instruction order, actor/worktree/epoch binding, two-attempt limit and reviewer preflight. S09 · core/surface/skills/tdd/SKILL.mdCurrent TDD context preflight and obligation derivation re-read. Six-phase gate sequence also confirmed by current skill index; fix and test-only paths do not fabricate feature artifacts. Knowledge / core
/init init-codearbiter.py → context-creation / decompose
Route to outcome --check reports only; activation/scaffold is a separate action Uninitialized project with meaningful source → bounded brownfield initial population Brownfield: qualified read-only scout profile → closed reports → scoped synthesis and relevant gap decisions Approve each native document/provenance preview; receipt-only apply, then separately approved finalization No meaningful source → retain the separate six-layer greenfield interview and persisted drafts Already initialized → normal operation; a selected refresh belongs to status drift, not init Selected context Source shape and maturity; existing file ownership and initialization state
Qualified scout reports and current source snapshots for brownfield, persisted interview answers for greenfield
Output Initialized project knowledge after the selected owner completes its real gates; a stub alone is not ready.
Brownfield profile or qualified native writer unavailable → block that dependent step. Five document/provenance pairs have independent approvals; finalization has its own 13-identity preview and receipt. Never clear the marker to refresh.
Host / scope Claude Code / Codex / Pi; admission and runtime capabilities still apply
claude /ca:init
codex $ca-init
pi /ca-initModes --check · --stage N · --brownfield · --greenfield
Read the exact procedure
Related entries /create-context /decompose /status
Pinned evidence S23 · core/surface/commands/init.mdCurrent init owner: activation/scaffold versus initial population; initialized refresh belongs to status drift, not reinitialization. S13 · core/surface/skills/context-creation/SKILL.mdCurrent brownfield owner: qualified containment, bounded reports and join, scoped gaps, native document/provenance writes, separate 13-identity finalization and recovery. S14 · core/surface/skills/decompose/SKILL.mdCurrent greenfield interview and completion contract re-read. Its persisted six-layer draft flow remains separate from the brownfield writer lifecycle. S26 · core/surface/skills/context-check/SKILL.mdInitialized refresh: read-only report, explicit selection, v1/v2 separation, native paired updates and preserved human history. Full owner read. Operate / advanced
/metrics read-only trend report
Route to outcome Read selected and prior windows Summarize overrides, small lanes and low-confidence choices Return comparison Selected context Override, triage and sprint decision logs
Output Trend report against prior window.
Reporting is not state mutation or workflow approval.
Host / scope Claude Code / Codex / Pi; admission and runtime capabilities still apply
claude Not exposed on this host
codex Not exposed on this host
pi Not exposed on this host Modes --window N
Removed on 2026-10-23; no host exposes this route.
Related entries /audit /sprint
Pinned evidence S02 · core/surface/COMMANDS.mdCatalog, modes, host-limited commands and purpose of each public route. Operate / core
/override permitted logged bypass
Route to outcome Check governing rule permits the specific bypass Obtain explicit confirmation and reason Record user attribution in overrides.log Proceed only within that permitted boundary Selected context Governing rule, requested exception, user identity and reason
Output Attributed override audit entry and bounded continuation.
Not universal bypass authority; forbidden safety boundaries remain forbidden.
Host / scope Claude Code / Codex / Pi; admission and runtime capabilities still apply
claude /ca:override
codex $ca-override
pi /ca-overrideRead the exact procedure
Related entries /audit /metrics
Pinned evidence S03 · core/surface/arbiter.mdIntent-to-owner routing, on-demand loading, caller extent and mode boundary. S04 · core/surface/includes/routing-table.mdAuthoritative intent-to-owner table, scope-touch routes and stop/return semantics. Ship / core
/pr finishing-a-development-branch
Route to outcome Watch/cleanup modes branch before creation prerequisites Creation: current-head commit and current acceptance Applicable diff reviewers → blocker clearance Open-PR procedure in this same owner Merge only when explicitly selected and exact-head hosted CI is green During an authorized integration, an explicitly chosen exact-HEAD legacy code-map conflict resolution may use inspect → bound restore-head. It leaves the index untouched; inspect, stage normally and re-prove the changed head. Selected context Branch, base, head and current diff
Current commit/acceptance proof
Applicable ADR ancestry and hosted CI
Output Open PR; optionally an explicitly authorized merge. Watch and cleanup return their own outputs.
No self-loop back to PR wrapper. Sprint opens only. No inferred merge/discard authority.
Host / scope Claude Code / Codex / Pi; admission and runtime capabilities still apply
claude /ca:pr
codex $ca-pr
pi /ca-prModes --watch · --cleanup
Read the exact procedure
Related entries /watch /cleanup /commit /release
Pinned evidence S12 · core/surface/skills/finishing-a-development-branch/SKILL.mdWatch/cleanup early routing; commit and acceptance preflight; PR and merge authority. S04 · core/surface/includes/routing-table.mdAuthoritative intent-to-owner table, scope-touch routes and stop/return semantics. Review / core
/preview read-only preview command
Route to outcome Inspect current diff without onboarding Predict matching reviewers Run state-free secret scan Return preview only Selected context Current diff and shared review matrix
Output Predicted review fleet and scan report; no project-state mutation.
Prediction is not review evidence. No init, staging or writing required.
Host / scope Claude Code / Codex / Pi; admission and runtime capabilities still apply
claude Not exposed on this host
codex Not exposed on this host
pi Not exposed on this host Removed on 2026-10-23; no host exposes this route.
Related entries /review /init
Pinned evidence S02 · core/surface/COMMANDS.mdCatalog, modes, host-limited commands and purpose of each public route. S04 · core/surface/includes/routing-table.mdAuthoritative intent-to-owner table, scope-touch routes and stop/return semantics. S18 · core/surface/includes/review-matrix.mdShared diff-path reviewer selection for review/checkpoint/preview; callers retain their own gates. Operate / advanced
/prune transcript / native-compaction control
Route to outcome Inspect or select prune policy Claude: dry-run or operate only on a non-live transcript Pi: use native compaction, not active-session rewriting Return inspection, audit or compaction result Selected context Host-specific transcript/compaction state
Output Reduced transcript bulk after resume/compaction or a report.
Never execute pruning against Claude’s live transcript; no supported Codex transcript route is asserted.
Host / scope Claude Code; Pi uses native compaction. Not listed for Codex.
claude /ca:prune
codex Not exposed on this host
pi /ca-pruneModes status · dry · run · audit · on · off
Read the exact procedure
Related entries /statusline
Pinned evidence S02 · core/surface/COMMANDS.mdCatalog, modes, host-limited commands and purpose of each public route. S04 · core/surface/includes/routing-table.mdAuthoritative intent-to-owner table, scope-touch routes and stop/return semantics. Decide / advanced
/reconcile decision-variance
Route to outcome Choose report-only or reconciliation mode Restricted scouts gather cited evidence Grader supplies SMARTS comparison; optional challenger User selects the decision Attributed decision-log write only in reconciliation mode Selected context Three decomposition plans and decision-log
Accepted ADRs, deferrals, scoped source evidence
Output Variance report, or an explicit attributed architectural ruling.
Report-only writes nothing; grader recommends, never decides. Contradictory governing rules use conflict instead.
Host / scope Claude Code / Codex / Pi; admission and runtime capabilities still apply
claude /ca:reconcile
codex $ca-reconcile
pi /ca-reconcileRead the exact procedure
Related entries /adr /conflict
Pinned evidence S04 · core/surface/includes/routing-table.mdAuthoritative intent-to-owner table, scope-touch routes and stop/return semantics. S05 · core/surface/includes/reference-map.mdScope-to-governing-document and specialist mapping. S21 · core/surface/skills/INDEX.mdCurrent routine-owner index: five debug dispositions, bounded workflow owners and unchanged six-phase TDD / nine-phase commit sequence. S22 · core/surface/agents/INDEX.mdRole separation; specialist reviewers; restricted scouts/graders; authority-reviewer host restriction. Change / core
/refactor refactor
Route to outcome Name a behavior-preserving scope Prove parity coverage with pre-existing tests Backfill coverage or new seams through TDD when needed Restructure without changing behavior Verify unchanged contracts; commit only through commit-gate Selected context Existing behavior and unchanged tests
Stage-specific coverage profile
Project standards and affected paths
For targeted test-only work, use the existing caller via compose_fix_or_test_actor_input; do not invent a feature spec merely to supply context.
Output Restructured implementation with behavioral parity evidence.
No parity proof, no refactor; behavior changes belong in feature/fix.
Host / scope Claude Code / Codex / Pi; admission and runtime capabilities still apply
claude /ca:refactor
codex $ca-refactor
pi /ca-refactorRead the exact procedure
Related entries /feature /fix /commit
Pinned evidence S04 · core/surface/includes/routing-table.mdAuthoritative intent-to-owner table, scope-touch routes and stop/return semantics. S21 · core/surface/skills/INDEX.mdCurrent routine-owner index: five debug dispositions, bounded workflow owners and unchanged six-phase TDD / nine-phase commit sequence. S09 · core/surface/skills/tdd/SKILL.mdCurrent TDD context preflight and obligation derivation re-read. Six-phase gate sequence also confirmed by current skill index; fix and test-only paths do not fabricate feature artifacts. S27 · core/pysrc/_contextselectlib.pyInert scoped selector and delivery implementation: commands/cwd/evidence, native instruction order, actor/worktree/epoch binding, two-attempt limit and reviewer preflight. Ship / core
/release release
Route to outcome Resolve the declared target, version policy, payload and reviewed executable inputs; dry-run reports only and skips writes and execution. Prepare version/changelog and declared release surfaces through the ordinary commit gate on a non-default branch. Report the exact assets and publisher, open the release PR and STOP. Interactive preparation does not compose a local tag. Only after the PR merges may a qualifying hosted publisher continue at the exact merged, fetched and verified default-branch commit. Compose and validate the exact tag/cohort only through that publisher. Apply the existing one-cohort publication authority; CI success alone is not publication permission. Read back the exact non-draft release and declared inventory. A declared provenance closeout uses its own receipt PR; an already-published receipt-only path never publishes again. Selected context Declared target row, exact manifest/changelog/payload ownership
Exact release-surface commit, fetched default branch and current hosted evidence
Existing publication authority for this cohort, exact assets and any declared provenance requirement
Output A release-preparation PR; separately, an authorized hosted publication and its exact readback. Declared receipt closeout remains a separate PR.
Mandatory release-PR stop. No local tag fallback, red/stale-head publication or implicit consent. Dry-run does not fetch, execute checks, commit or publish.
Host / scope Claude Code / Codex / Pi; admission and runtime capabilities still apply
claude /ca:release
codex $ca-release
pi /ca-releaseModes --dry-run
Read the exact procedure
Related entries /commit /pr
Pinned evidence S33 · core/surface/skills/release/SKILL.mdInteractive release preparation and mandatory release-PR boundary; qualified hosted tag composition. S34 · site/src/content/docs/concepts/workflow-routes.mdxExisting public route-comparison owner; distinct release/initialization endpoints and source-only explanation boundary. Review / core
/review dispatching-parallel-agents
Route to outcome Select reviewers by actual diff paths Select context separately for every actual reviewer, check named critical references, and deliver its bounded input with current actor/worktree/epoch binding Run independent read-only units Collect reports and terminal states finding-triage → verdict-aggregator Return current-diff verdict Selected context Exact diff and governing project documents
Applicable spec obligations and ADRs
Reviewer-specific scoped reports
Resolved diff paths; source-scoped constraints and native instruction precedence
Reviewer-specific critical references and observed input delivery, not the author’s packet or a parent session marker
Output In-memory triaged verdict with blocking findings separated from lower-severity findings.
CRITICAL/HIGH block the current change. Prediction or missing results cannot substitute for completed review. Missing critical context blocks a substantive verdict; safe bounded orientation may continue. Posting an inbound PR verdict requires separate authority.
Host / scope Claude Code / Codex / Pi; admission and runtime capabilities still apply
claude /ca:review
codex $ca-review
pi /ca-reviewRead the exact procedure
Related entries /checkpoint /pr /preview
Pinned evidence S04 · core/surface/includes/routing-table.mdAuthoritative intent-to-owner table, scope-touch routes and stop/return semantics. S18 · core/surface/includes/review-matrix.mdShared diff-path reviewer selection for review/checkpoint/preview; callers retain their own gates. S22 · core/surface/agents/INDEX.mdRole separation; specialist reviewers; restricted scouts/graders; authority-reviewer host restriction. S28 · core/surface/commands/review.mdCurrent full review owner: stable diff resolution, selected context delivered independently to each reviewer, local verdict before separately authorized posting. S27 · core/pysrc/_contextselectlib.pyInert scoped selector and delivery implementation: commands/cwd/evidence, native instruction order, actor/worktree/epoch binding, two-attempt limit and reviewer preflight. Change / core
/spike self-contained spike command
Route to outcome Frame a question and timebox Explore on disposable spike/<slug> Capture findings Exit to findings or a new feature workflow Selected context Question, uncertainty and experiment bounds
Relevant source and project constraints
Output Findings / throwaway experiment, not production delivery.
Spike branch never merges or opens a PR. Its commit exemption cannot be used to land production code.
Host / scope Claude Code / Codex / Pi; admission and runtime capabilities still apply
claude /ca:spike
codex $ca-spike
pi /ca-spikeRead the exact procedure
Related entries /feature /debug
Pinned evidence S02 · core/surface/COMMANDS.mdCatalog, modes, host-limited commands and purpose of each public route. S04 · core/surface/includes/routing-table.mdAuthoritative intent-to-owner table, scope-touch routes and stop/return semantics. Change / core
/sprint SPRINT.md → subagent-driven-development
Route to outcome Resolve same-format artifacts; retain existing approved intent Prepare spec and draft plan for one real user approval of the pair Execute the full plan without attended batch pauses SMARTS governs delegated in-scope choices; correction must rerun failed gates Before each author or independent reviewer acts, compose its own current scoped input and perform bounded actor delivery; a coordinator read is not receipt Current acceptance → authorized commit → open PR only Selected context Project rules and recorded intent
Approved spec/plan and explicit delegation terms
Task-specific contextual reads
Actual review, verification and authority receipts
Current actor-scoped selection: map/provenance, command cwd and observed evidence, constraints and host-effective instructions
Actual author and independent reviewer delivery bound to task/worktree/source/host epochs; reselect after resume, compaction or scope change
Use exact engine-returned IDs, not assumed SPEC-/PLAN- prefixes. Approved prerequisites use one visible 24-hour request/reply deadline.
Output Accepted plan and implementation, sprint decision log, open PR and summary. No autonomous merge.
Initial approval never self-issued. Security/authority/irreversible-action stops remain. HTML --farm blocked; legacy farm is opt-in preview. Selected context is orientation, not task admission.
Host / scope Claude Code / Codex: admitted installed typed workflow only. Pi 1.0.2 is the current runtime target, but its native approval adapter remains unimplemented (DEGRADED); legacy Markdown pairs and inline small-feature work have separate contracts.
claude /ca:sprint
codex $ca-sprint
pi /ca-sprintModes --farm
Read the exact procedure
Related entries /feature /commit /pr
Pinned evidence S07 · core/surface/SPRINT.mdInitial paired approval; full-plan execution; delegated methods; recovery and farm boundaries. S08 · core/surface/skills/subagent-driven-development/SKILL.mdCurrent implementation-dispatch and independent spec-review sections re-read; remaining scope contract also checked against current skill index. Fresh actor input and bounded delivery are distinct from engine tickets. S10 · core/surface/includes/artifacts.mdInstalled admission, typed reads, context completion, authority receipts and format boundaries. S12 · core/surface/skills/finishing-a-development-branch/SKILL.mdWatch/cleanup early routing; commit and acceptance preflight; PR and merge authority. S27 · core/pysrc/_contextselectlib.pyInert scoped selector and delivery implementation: commands/cwd/evidence, native instruction order, actor/worktree/epoch binding, two-attempt limit and reviewer preflight. S09 · core/surface/skills/tdd/SKILL.mdCurrent TDD context preflight and obligation derivation re-read. Six-phase gate sequence also confirmed by current skill index; fix and test-only paths do not fabricate feature artifacts. S31 · docs/parity.mdCurrent Pi 1.0.2 target, separate host counts, DEGRADED typed approval and Codex farm exception. S32 · core/pysrc/_artifactauthoritylib.pyNative artifact ID grammar, producer states, Claude join lock, UTF-8/UTF-16-key-order observation binding. Operate / advanced
/standup repository hygiene command
Route to outcome Inspect branch/worktree/stash and working-tree state Report unsafe or ambiguous residue Confirm safe cleanup actions individually Apply only permitted fast-forward/cleanup actions Selected context Current Git and repository state
Output Hygiene briefing and individually authorized cleanup results.
Dirty, stashed or unpushed work is not silently discarded.
Host / scope Claude Code / Codex / Pi; admission and runtime capabilities still apply
claude /ca:standup
codex $ca-standup
pi /ca-standupRead the exact procedure
Related entries /status /cleanup
Pinned evidence S02 · core/surface/COMMANDS.mdCatalog, modes, host-limited commands and purpose of each public route. S04 · core/surface/includes/routing-table.mdAuthoritative intent-to-owner table, scope-touch routes and stop/return semantics. Operate / core
/status status / context-check
Route to outcome Default status and initial drift report are read-only Assess legacy v1 drift and each enrolled v2 identity/coverage state separately; unknown is not fresh Wait for explicit selected documents and affected scope, or explicit full refresh; defer changes nothing Re-scout only the selected assignments through qualified containment and a bounded current join For v2, approve the exact owned-field document/provenance preview and apply only its native receipt Read back committed pairs; retain the initialized marker, unselected human content and separate task/question/audit history; do not stage or commit Selected context Current bounded source/target snapshots and action-boundary checks
v1 drift separately from v2 current/stale/unknown coverage; identity matching is not semantic proof
Explicit document selection, actual field ownership, qualified scouts and native writer capability
Output Read-only state/drift report, or explicitly selected native context updates with declared full/incremental coverage.
Report alone is not refresh consent. No reinitialization or marker removal. No legacy raw rebaseline fallback for v2, no unsupported human-layout overwrite, no staging or commit.
Host / scope Claude Code / Codex / Pi; admission and runtime capabilities still apply
claude /ca:status
codex $ca-status
pi /ca-statusModes drift
Read the exact procedure
Related entries /init /context-check /commit
Pinned evidence S26 · core/surface/skills/context-check/SKILL.mdInitialized refresh: read-only report, explicit selection, v1/v2 separation, native paired updates and preserved human history. Full owner read. S13 · core/surface/skills/context-creation/SKILL.mdCurrent brownfield owner: qualified containment, bounded reports and join, scoped gaps, native document/provenance writes, separate 13-identity finalization and recovery. S23 · core/surface/commands/init.mdCurrent init owner: activation/scaffold versus initial population; initialized refresh belongs to status drift, not reinitialization. Operate / advanced
/statusline Claude Code statusline control
Route to outcome Select install, uninstall or status Change or inspect the owned Claude statusline configuration Return status Selected context Claude Code install and statusline settings
Output Statusline configuration or status report.
Claude-specific catalog surface; not a cross-host workflow.
Host / scope Claude Code only
claude /ca:statusline
codex Not exposed on this host
pi Not exposed on this host Modes install · uninstall · status
Read the exact procedure
Related entries /doctor
Pinned evidence S02 · core/surface/COMMANDS.mdCatalog, modes, host-limited commands and purpose of each public route. Operate / core
/task work.py → selected native engine or legacy board writer
Route to outcome Select work by stable ID through the current owner Apply one create, start, complete or reopen request Record local progress without inventing check results Selected context Selected work ID and current owner state
Requested task or harvested work
Output Native work item or legacy board transition, with checks and delivery reported separately.
The facade preserves stable work identity, exact change checks and replay identity. Queued work already performed can complete directly without a fabricated start. Never delete a task to complete it or claim checks passed from local progress.
Host / scope Claude Code / Codex / Pi; admission and runtime capabilities still apply
claude /ca:task
codex $ca-task
pi /ca-taskModes add · start · done
Read the exact procedure
Related entries /standup /feature
Pinned evidence S02 · core/surface/COMMANDS.mdCatalog, modes, host-limited commands and purpose of each public route. S04 · core/surface/includes/routing-table.mdAuthoritative intent-to-owner table, scope-touch routes and stop/return semantics. S25 · core/surface/skills/debug/SKILL.mdFive-disposition diagnostic owner; transient validated handoff, freshness and authority boundaries. Full owner read. Review / advanced
/threat-model security-architecture
Route to outcome Explicitly request sensitive-design analysis Run lightweight STRIDE with applicable specialists Return threats and mitigations to the design owner Selected context Design scope, assets and trust boundaries
security-controls.md
Output Threat/mitigation analysis and constraints.
Opt-in, not a routine change gate; critical unmitigated threat stops this pass.
Host / scope Claude Code / Codex / Pi; admission and runtime capabilities still apply
claude /ca:threat-model
codex $ca-threat-model
pi /ca-threat-modelRead the exact procedure
Related entries /feature /adr
Pinned evidence S04 · core/surface/includes/routing-table.mdAuthoritative intent-to-owner table, scope-touch routes and stop/return semantics. S21 · core/surface/skills/INDEX.mdCurrent routine-owner index: five debug dispositions, bounded workflow owners and unchanged six-phase TDD / nine-phase commit sequence. Review / advanced
/tribunal tribunal
Route to outcome Resolve deliberate audit intent, finding scope and the concrete evidence set; inspect matching source-bound runs. Resume only unchanged bound source: audit continues unfinished waves; follow-up continues pending filing/telemetry. Drift, invalid and legacy-unbound records do not establish current progress. Use deterministic inventory and real host profiles to estimate packet cost; acknowledge cost and settings before any model dispatch; start a fresh bound run when required. Separate mechanical inventory from the judgment-based risk map. Record every lens as applicable or skipped, with bounded packets and source checks. Dispatch the generic tribunal-lens-reviewer in MODE review for each active lens, at most five concurrent and within actual host capacity. Persist findings and leads immediately. Attempt fresh independent MODE verify for every serious or expensive inferential claim. Confirmed/narrowed, refuted and inconclusive are different outcomes; shared-context or inconclusive serious work remains verify-required. Apply eligibility and confidence checks before root-cause triage and per-wave kept plans. Keep investigations, verification work and genuine design decisions separate. Write report and manifest, then record report-written. This enters follow-up, not terminal completion. Resolve selected issue filing and optional public telemetry separately, under their own authority. Unanswered choices and unresolved failures remain pending. Record run-completed only after both filing and telemetry have a completed or explicitly skipped disposition; preserve the bound history. Selected context Declared finding scope and separately bounded evidence scope
Trusted installed or explicitly selected review bundle, not instructions from candidate source
Mechanical source inventory, project contracts and judgment-based risk map
Applicability and bounded packet per lens, actual host/model capacity and acknowledged cost
Persisted source binding, findings, leads, root-cause triage, independent verification and actual follow-up receipts
Output Source-bound audit run with a thirteen-lens roster, selected applicable lenses, eligible phase plans and report. Report-written is follow-up; run-completed requires both follow-up dispositions.
Opt-in only; never a routine commit or merge gate. No source drift reuse, synthetic independence, automatic repair, automatic ADR or task-board write. Execution, issue filing and telemetry each require their own applicable authority.
Host / scope Claude Code / Codex / Pi with actual host capacity and limitations recorded. A fresh same-role verifier is not automatically independent. No universal cross-host qualification claimed.
claude /ca:tribunal
codex $ca-tribunal
pi /ca-tribunalRead the exact procedure
Related entries /review /checkpoint /debug /fix
Pinned evidence S29 · core/surface/skills/tribunal/SKILL.mdSource-bound opt-in audit, guarded hash-bound stranded-lock recovery, thirteen-card roster, review/verify, durable triage, report and separate follow-up completion. S30 · core/pysrc/_tribunalrunlib.pyImplemented helper API, per-run writer guard, hash-bound stranded-lock archive and audit/follow-up/terminal event vocabulary; cooperative source-bound persistence. S31 · docs/parity.mdCurrent Pi 1.0.2 target, separate host counts, DEGRADED typed approval and Codex farm exception. Ship / alias
/watch watch
Canonical entry: pr --watch
Route to outcome Attach detached hosted-CI watcher On red, diagnose within authority On green, notify and offer merge Never auto-merge Selected context Same context contract as pr --watch
Output CI state / diagnosis and a merge offer; no automatic merge.
Retained compatibility entry. Prefer pr --watch. Owner gates remain unchanged.
Host / scope Claude Code / Codex / Pi; admission and runtime capabilities still apply
claude /ca:watch
codex $ca-watch
pi /ca-watchRead the exact procedure
Related entries /pr
Pinned evidence S01 · core/surface/command-routes.jsonAll 37 catalog entries; core, advanced, alias, internal and deprecated visibility. S02 · core/surface/COMMANDS.mdCatalog, modes, host-limited commands and purpose of each public route. S04 · core/surface/includes/routing-table.mdAuthoritative intent-to-owner table, scope-touch routes and stop/return semantics. S21 · core/surface/skills/INDEX.mdCurrent routine-owner index: five debug dispositions, bounded workflow owners and unchanged six-phase TDD / nine-phase commit sequence. Selected cmd
/init Assistive workflow: source-bound implementation An agreed outcome becomes a maintained plan, direct or delegated implementation, proportional verification and an honest delivery report. Task context and dependency advice support progress; real user stops and host permissions remain effective. Greenfield and brownfield population use the paired native writer and exact finalization. Exact blob bindings identify inspected source bytes; the linked revision is the committed source base. Implementation, verification, live host qualification and release remain separate claims.
Current source changes The user-facing route table points feature and sprint work at useful planned execution and selected context refresh at its existing owner. Existing authorization covers ordinary work; actual host, destructive-action and commit controls remain. core/surface/includes/routing-table.md The installed context-draft command prepares a paired document preview; context-finalize prepares initial finalization. context-apply commits the exact selected preview and digest. Context modes preserve scaffold and initialized refresh boundaries. core/pysrc/init-codearbiter.py Initial brownfield setup checks the intended repository and actual host permissions, then uses bounded inspection and the installed safe context writer. Ordinary Windows work needs no special scout profile or per-document artifact approval. An existing initialized repository routes to selected refresh instead of another scaffold or population cycle. core/surface/commands/init.md --check: report only Otherwise scaffold, then populate
Related entries /init /create-context /decompose
Pinned evidence S23 · core/surface/commands/init.mdCurrent init owner: activation/scaffold versus initial population; initialized refresh belongs to status drift, not reinitialization. Selected skill
decompose Assistive workflow: source-bound implementation An agreed outcome becomes a maintained plan, direct or delegated implementation, proportional verification and an honest delivery report. Task context and dependency advice support progress; real user stops and host permissions remain effective. Greenfield and brownfield population use the paired native writer and exact finalization. Exact blob bindings identify inspected source bytes; the linked revision is the committed source base. Implementation, verification, live host qualification and release remain separate claims.
Current source changes Greenfield interviews persist under plans/decomposition and retain actual decisions. Five document/v2-provenance pairs use context-draft and context-apply; context-finalize binds all thirteen outputs. ADRs and task rows use their owners, and retained interview evidence survives finalization. core/surface/skills/decompose/SKILL.md Greenfield: six persisted interview layers
Related entries /init /decompose
Pinned evidence S14 · core/surface/skills/decompose/SKILL.mdCurrent greenfield interview and completion contract re-read. Its persisted six-layer draft flow remains separate from the brownfield writer lifecycle. Selected skill
context-creation Assistive workflow: source-bound implementation An agreed outcome becomes a maintained plan, direct or delegated implementation, proportional verification and an honest delivery report. Task context and dependency advice support progress; real user stops and host permissions remain effective. Greenfield and brownfield population use the paired native writer and exact finalization. Exact blob bindings identify inspected source bytes; the linked revision is the committed source base. Implementation, verification, live host qualification and release remain separate claims.
Current source changes The skill index connects planning, direct or delegated implementation, proportional verification and selected context maintenance to their owning workflows. core/surface/skills/INDEX.md Brownfield population covers six logical categories and five content documents through direct bounded inspection or optional scoped agents. The paired writer preserves source bindings and protected content; initial finalization checks all required outputs. core/surface/skills/context-creation/SKILL.md Initial brownfield setup checks the intended repository and actual host permissions, then uses bounded inspection and the installed safe context writer. Ordinary Windows work needs no special scout profile or per-document artifact approval. An existing initialized repository routes to selected refresh instead of another scaffold or population cycle. core/surface/commands/init.md Six full logical assignments require observed read-only containment. Generated plugins alone do not qualify that profile. Each of five content/provenance pairs needs its own native approval, then initialization is a separately approved transaction.
Related entries /init /create-context
Pinned evidence S13 · core/surface/skills/context-creation/SKILL.mdCurrent brownfield owner: qualified containment, bounded reports and join, scoped gaps, native document/provenance writes, separate 13-identity finalization and recovery. S26 · core/surface/skills/context-check/SKILL.mdInitialized refresh: read-only report, explicit selection, v1/v2 separation, native paired updates and preserved human history. Full owner read. Selected state
Initialized project knowledge Assistive workflow: source-bound implementation An agreed outcome becomes a maintained plan, direct or delegated implementation, proportional verification and an honest delivery report. Task context and dependency advice support progress; real user stops and host permissions remain effective. Greenfield and brownfield population use the paired native writer and exact finalization. Exact blob bindings identify inspected source bytes; the linked revision is the committed source base. Implementation, verification, live host qualification and release remain separate claims.
Current source changes Initial finalization binds thirteen output identities, rechecks supporting source evidence, and changes the initialization marker with its paired provenance digest. Ordinary finalization requires no artifact approval receipt and does not certify semantic correctness. Already initialized repositories refuse a new finalization. core/artifacts/internal/contextdocument/finalization.go Brownfield population covers six logical categories and five content documents through direct bounded inspection or optional scoped agents. The paired writer preserves source bindings and protected content; initial finalization checks all required outputs. core/surface/skills/context-creation/SKILL.md Greenfield interviews persist under plans/decomposition and retain actual decisions. Five document/v2-provenance pairs use context-draft and context-apply; context-finalize binds all thirteen outputs. ADRs and task rows use their owners, and retained interview evidence survives finalization. core/surface/skills/decompose/SKILL.md CONTEXT • stack • standards • security ADRs • plans • questions • code-map Provenance supports freshness checks
Related entries /init /create-context /decompose /feature /sprint /fix /refactor
Pinned evidence S13 · core/surface/skills/context-creation/SKILL.mdCurrent brownfield owner: qualified containment, bounded reports and join, scoped gaps, native document/provenance writes, separate 13-identity finalization and recovery. S14 · core/surface/skills/decompose/SKILL.mdCurrent greenfield interview and completion contract re-read. Its persisted six-layer draft flow remains separate from the brownfield writer lifecycle. S23 · core/surface/commands/init.mdCurrent init owner: activation/scaffold versus initial population; initialized refresh belongs to status drift, not reinitialization. Selected state
Runtime + on-demand context Assistive workflow: source-bound implementation An agreed outcome becomes a maintained plan, direct or delegated implementation, proportional verification and an honest delivery report. Task context and dependency advice support progress; real user stops and host permissions remain effective. Greenfield and brownfield population use the paired native writer and exact finalization. Exact blob bindings identify inspected source bytes; the linked revision is the committed source base. Implementation, verification, live host qualification and release remain separate claims.
Current source changes Routine artifact instructions now focus on the installed writer, typed plan, bounded task reads, factual prerequisites, actual checks and safe retry. The default leaf is about 7.3 KB and includes an executable stdin bootstrap with optional operation-schema discovery; explicit attestation remains available on demand without burdening ordinary task context. A spec's SMARTS design rationale lives in its approach, decisions and one SEC-SMARTS section. core/surface/includes/artifacts.md The user-facing route table points feature and sprint work at useful planned execution and selected context refresh at its existing owner. Existing authorization covers ordinary work; actual host, destructive-action and commit controls remain. core/surface/includes/routing-table.md The commit owner checks authorization, the staged diff, secrets and the verification approach appropriate to the change. Current criterion checks, genuine fix regressions and refactor parity retain their distinct evidence requirements. Reuse applicable observed checks and reviews across phases. Managed v2 maintenance selects affected claims and uses native paired writes; unknown evidence defers that document. core/surface/skills/commit-gate/SKILL.md PR readiness uses the explicit base-to-HEAD subject for current checks, secrets scanning and applicable specialized review. Existing evidence is reused only while its source and inputs match. Merge requires authorization and exact-head hosted CI. core/surface/skills/finishing-a-development-branch/SKILL.md The command catalog starts from the agreed outcome and maintained plan; SMARTS is useful for difficult tradeoffs and does not authorize a changed outcome. core/surface/COMMANDS.md Resident instructions preserve existing work authorization and actual owner decisions without another spec approval or a SMARTS record for every routine step. core/surface/arbiter.md The lazy reference map routes provenance drift or requested initialized refresh to context-check after reading selected documents, provenance and the code map. The owner reuses selected scope and existing authorization. core/surface/includes/reference-map.md Startup state; safety-core + active persona On-demand rules + complete typed task reads Task-scoped packets delivered to each actual actor Context is not approval, admission or consumption proof
Related entries /feature /sprint /fix /refactor /chore /review /commit
Pinned evidence S27 · core/pysrc/_contextselectlib.pyInert scoped selector and delivery implementation: commands/cwd/evidence, native instruction order, actor/worktree/epoch binding, two-attempt limit and reviewer preflight. S15 · core/pysrc/prompt-submit.pyCurrent persona composition and session/mode/compaction injection comments read; not actor-delivery proof. S16 · core/pysrc/_readinjectlib.pyCurrent file-read pointer and dedup implementation interface read; advisory hints remain separate from task-time selected input. S10 · core/surface/includes/artifacts.mdInstalled admission, typed reads, context completion, authority receipts and format boundaries. Selected cmd
/feature New behavior; attended workflow
Related entries /feature
Pinned evidence S06 · core/surface/commands/feature.mdCurrent feature owner: same-format resume and full/small gates; adds actual per-actor context composition. Selected cmd
/sprint Coherent goal; approved autonomy
Related entries /sprint
Pinned evidence S07 · core/surface/SPRINT.mdInitial paired approval; full-plan execution; delegated methods; recovery and farm boundaries. Selected cmd
/fix · /refactor Assistive workflow: source-bound implementation An agreed outcome becomes a maintained plan, direct or delegated implementation, proportional verification and an honest delivery report. Task context and dependency advice support progress; real user stops and host permissions remain effective. Greenfield and brownfield population use the paired native writer and exact finalization. Exact blob bindings identify inspected source bytes; the linked revision is the committed source base. Implementation, verification, live host qualification and release remain separate claims.
Current source changes Confirmed fixes preserve causal evidence and a named regression. Delegated input retains critical requirements; advisory delivery exhaustion ends retries while actual missing inputs and conflicts retain their dependent-work boundary. Distinct remediation strategies are compared with SMARTS only after the root cause is confirmed. core/surface/commands/fix.md Characterization tests can pass against existing correct behavior. Preserve behavioral assertions against the post-backfill baseline; mechanical import or symbol-reference updates retain the same oracle. Coverage and parity checks remain required. A precise surface table proceeds under existing user authorization; only material expansion or a changed outcome needs another decision. Distinct restructuring strategies are compared with SMARTS only after surface and parity are settled. core/surface/skills/refactor/SKILL.md Regression proof / behavior parity
Related entries /fix /refactor
Pinned evidence S19 · core/surface/commands/fix.mdCurrent full fix owner: validated handoff, independent caller authority, evidence revalidation, selected actor input and genuine target-red gate. S21 · core/surface/skills/INDEX.mdCurrent routine-owner index: five debug dispositions, bounded workflow owners and unchanged six-phase TDD / nine-phase commit sequence. Selected cmd
/chore · /add-dep Type-scaled work / vetted adoption
Related entries /chore /add-dep
Pinned evidence S20 · core/surface/commands/chore.mdDocs/deps/revert branches; behavioral adoption returns to feature/fix. S04 · core/surface/includes/routing-table.mdAuthoritative intent-to-owner table, scope-touch routes and stop/return semantics. Selected skill
brainstorming → writing-plans Assistive workflow: source-bound implementation An agreed outcome becomes a maintained plan, direct or delegated implementation, proportional verification and an honest delivery report. Task context and dependency advice support progress; real user stops and host permissions remain effective. Greenfield and brownfield population use the paired native writer and exact finalization. Exact blob bindings identify inspected source bytes; the linked revision is the committed source base. Implementation, verification, live host qualification and release remain separate claims.
Current source changes Routine artifact instructions now focus on the installed writer, typed plan, bounded task reads, factual prerequisites, actual checks and safe retry. The default leaf is about 7.3 KB and includes an executable stdin bootstrap with optional operation-schema discovery; explicit attestation remains available on demand without burdening ordinary task context. A spec's SMARTS design rationale lives in its approach, decisions and one SEC-SMARTS section. core/surface/includes/artifacts.md Plans preserve agreed criteria, useful tasks, dependency ancestry, bounded shared inputs and observable checks. Stable source identities and typed maintenance keep changes within the selected outcome. A design-drift guard routes an undecided material solution choice back to brainstorming instead of scoring it during planning. core/surface/skills/writing-plans/SKILL.md The feature entry now resumes a maintained spec/plan and task progress without another artifact certification prompt; independent commit and test controls remain. The small lane runs an inline SMARTS scan only for a material solution choice. core/surface/commands/feature.md Source refinement respects recorded deferrals and accepted ADR conflicts without turning each ordinary method adjustment into a separate authorization interview. SMARTS compares approaches only when two or more are materially plausible; a single approach gets one line and no table. core/surface/skills/brainstorming/SKILL.md Full: approved spec, then bound plan
Related entries /feature
Pinned evidence S06 · core/surface/commands/feature.mdCurrent feature owner: same-format resume and full/small gates; adds actual per-actor context composition. S10 · core/surface/includes/artifacts.mdInstalled admission, typed reads, context completion, authority receipts and format boundaries. Selected gate
Spec + plan pair approval Assistive workflow: source-bound implementation An agreed outcome becomes a maintained plan, direct or delegated implementation, proportional verification and an honest delivery report. Task context and dependency advice support progress; real user stops and host permissions remain effective. Greenfield and brownfield population use the paired native writer and exact finalization. Exact blob bindings identify inspected source bytes; the linked revision is the committed source base. Implementation, verification, live host qualification and release remain separate claims.
Current source changes Routine artifact instructions now focus on the installed writer, typed plan, bounded task reads, factual prerequisites, actual checks and safe retry. The default leaf is about 7.3 KB and includes an executable stdin bootstrap with optional operation-schema discovery; explicit attestation remains available on demand without burdening ordinary task context. A spec's SMARTS design rationale lives in its approach, decisions and one SEC-SMARTS section. core/surface/includes/artifacts.md The sprint caller performs an index-first recorded-intent read, preserves actual owner decisions and ADR conflicts, then proceeds with authorized work without a second artifact approval system. A SMARTS Unknown names its missing observation and cannot bypass a hard gate or manufacture authority. core/surface/SPRINT.md The sprint entry now routes to the assistive implementation body instead of requiring a fresh user approval and two reviews per task. core/surface/commands/sprint.md Source refinement respects recorded deferrals and accepted ADR conflicts without turning each ordinary method adjustment into a separate authorization interview. SMARTS compares approaches only when two or more are materially plausible; a single approach gets one line and no table. core/surface/skills/brainstorming/SKILL.md One real user reply before execution
Related entries /sprint
Pinned evidence S07 · core/surface/SPRINT.mdInitial paired approval; full-plan execution; delegated methods; recovery and farm boundaries. S10 · core/surface/includes/artifacts.mdInstalled admission, typed reads, context completion, authority receipts and format boundaries. Selected skill
tdd / parity owner Assistive workflow: source-bound implementation An agreed outcome becomes a maintained plan, direct or delegated implementation, proportional verification and an honest delivery report. Task context and dependency advice support progress; real user stops and host permissions remain effective. Greenfield and brownfield population use the paired native writer and exact finalization. Exact blob bindings identify inspected source bytes; the linked revision is the committed source base. Implementation, verification, live host qualification and release remain separate claims.
Current source changes Confirmed fixes preserve causal evidence and a named regression. Delegated input retains critical requirements; advisory delivery exhaustion ends retries while actual missing inputs and conflicts retain their dependent-work boundary. Distinct remediation strategies are compared with SMARTS only after the root cause is confirmed. core/surface/commands/fix.md Characterization tests can pass against existing correct behavior. Preserve behavioral assertions against the post-backfill baseline; mechanical import or symbol-reference updates retain the same oracle. Coverage and parity checks remain required. A precise surface table proceeds under existing user authorization; only material expansion or a changed outcome needs another decision. Distinct restructuring strategies are compared with SMARTS only after surface and parity are settled. core/surface/skills/refactor/SKILL.md Small feature: confirmed inline mini-spec
Related entries /feature /fix /refactor
Pinned evidence S06 · core/surface/commands/feature.mdCurrent feature owner: same-format resume and full/small gates; adds actual per-actor context composition. S09 · core/surface/skills/tdd/SKILL.mdCurrent TDD context preflight and obligation derivation re-read. Six-phase gate sequence also confirmed by current skill index; fix and test-only paths do not fabricate feature artifacts. S19 · core/surface/commands/fix.mdCurrent full fix owner: validated handoff, independent caller authority, evidence revalidation, selected actor input and genuine target-red gate. S21 · core/surface/skills/INDEX.mdCurrent routine-owner index: five debug dispositions, bounded workflow owners and unchanged six-phase TDD / nine-phase commit sequence. Selected skill
Type / dependency gates Prose: no TDD; behavior: feature or fix
Related entries /chore /add-dep
Pinned evidence S20 · core/surface/commands/chore.mdDocs/deps/revert branches; behavioral adoption returns to feature/fix. S04 · core/surface/includes/routing-table.mdAuthoritative intent-to-owner table, scope-touch routes and stop/return semantics. Selected skill
executing-plans Scoped batches + human checkpoints
Related entries /feature
Pinned evidence S06 · core/surface/commands/feature.mdCurrent feature owner: same-format resume and full/small gates; adds actual per-actor context composition. S21 · core/surface/skills/INDEX.mdCurrent routine-owner index: five debug dispositions, bounded workflow owners and unchanged six-phase TDD / nine-phase commit sequence. Selected skill
Full-plan handoff No attended batch pauses; hard stops stay
Related entries /sprint
Pinned evidence S07 · core/surface/SPRINT.mdInitial paired approval; full-plan execution; delegated methods; recovery and farm boundaries. Selected skill
Task engine Fresh actor context → TDD Spec proof + fresh verification Combined-scope quality review Accept only on current evidence
Related entries /feature /sprint
Pinned evidence S08 · core/surface/skills/subagent-driven-development/SKILL.mdCurrent implementation-dispatch and independent spec-review sections re-read; remaining scope contract also checked against current skill index. Fresh actor input and bounded delivery are distinct from engine tickets. S27 · core/pysrc/_contextselectlib.pyInert scoped selector and delivery implementation: commands/cwd/evidence, native instruction order, actor/worktree/epoch binding, two-attempt limit and reviewer preflight. S09 · core/surface/skills/tdd/SKILL.mdCurrent TDD context preflight and obligation derivation re-read. Six-phase gate sequence also confirmed by current skill index; fix and test-only paths do not fabricate feature artifacts. Selected out
Verified change Current acceptance for typed plans Regression / parity / type proof Persistence still needs authority
Related entries /feature /sprint /fix /refactor /chore /add-dep
Pinned evidence S08 · core/surface/skills/subagent-driven-development/SKILL.mdCurrent implementation-dispatch and independent spec-review sections re-read; remaining scope contract also checked against current skill index. Fresh actor input and bounded delivery are distinct from engine tickets. S11 · core/surface/skills/commit-gate/SKILL.mdCurrent Phase 5.5 through terminal commit rules re-read. Legacy heal vocabulary must not be generalized to v2 paired native writes. S19 · core/surface/commands/fix.mdCurrent full fix owner: validated handoff, independent caller authority, evidence revalidation, selected actor input and genuine target-red gate. S20 · core/surface/commands/chore.mdDocs/deps/revert branches; behavioral adoption returns to feature/fix. S21 · core/surface/skills/INDEX.mdCurrent routine-owner index: five debug dispositions, bounded workflow owners and unchanged six-phase TDD / nine-phase commit sequence. Selected cmd
/commit → commit-gate Explicit permission • nine phases Focused proof • selective stage Output: local commit SHA
Related entries /commit /feature /sprint /fix /refactor /chore
Pinned evidence S11 · core/surface/skills/commit-gate/SKILL.mdCurrent Phase 5.5 through terminal commit rules re-read. Legacy heal vocabulary must not be generalized to v2 paired native writes. Selected cmd
/pr → finishing owner Current-head proof + diff review Open PR; sprint ends here Feature retains terminal choice
Related entries /pr /feature /sprint
Pinned evidence S12 · core/surface/skills/finishing-a-development-branch/SKILL.mdWatch/cleanup early routing; commit and acceptance preflight; PR and merge authority. Selected gate
Merge through the PR Explicit user authority Exact-head hosted CI + ancestry No direct-to-default delivery
Related entries /pr
Pinned evidence S12 · core/surface/skills/finishing-a-development-branch/SKILL.mdWatch/cleanup early routing; commit and acceptance preflight; PR and merge authority. Selected cmd
/release Assistive workflow: source-bound implementation An agreed outcome becomes a maintained plan, direct or delegated implementation, proportional verification and an honest delivery report. Task context and dependency advice support progress; real user stops and host permissions remain effective. Greenfield and brownfield population use the paired native writer and exact finalization. Exact blob bindings identify inspected source bytes; the linked revision is the committed source base. Implementation, verification, live host qualification and release remain separate claims.
Current source changes Release preparation inserts a new version below the entire Unreleased section and above prior versions, preserving pending notes. This guidance correction does not grant publication authority or replace the existing release-PR and hosted qualification boundaries. core/surface/skills/release/SKILL.md Release preparation PR → STOP Qualified hosted publisher Separate publication authority
Related entries /release
Pinned evidence S33 · core/surface/skills/release/SKILL.mdInteractive release preparation and mandatory release-PR boundary; qualified hosted tag composition. S34 · site/src/content/docs/concepts/workflow-routes.mdxExisting public route-comparison owner; distinct release/initialization endpoints and source-only explanation boundary. Selected cmd
/watch · /cleanup /pr --watch / --cleanup Early branches, not PR creation No automatic merge or discard
Related entries /watch /cleanup /pr
Pinned evidence S02 · core/surface/COMMANDS.mdCatalog, modes, host-limited commands and purpose of each public route. S04 · core/surface/includes/routing-table.mdAuthoritative intent-to-owner table, scope-touch routes and stop/return semantics. S12 · core/surface/skills/finishing-a-development-branch/SKILL.mdWatch/cleanup early routing; commit and acceptance preflight; PR and merge authority. Selected cmd
/debug · /spike Debug: five evidenced outcomes Validated handoff ≠ repair consent Spike: findings or feature
Related entries /debug /spike
Pinned evidence S25 · core/surface/skills/debug/SKILL.mdFive-disposition diagnostic owner; transient validated handoff, freshness and authority boundaries. Full owner read. S04 · core/surface/includes/routing-table.mdAuthoritative intent-to-owner table, scope-touch routes and stop/return semantics. Selected cmd
/preview · /review Assistive workflow: source-bound implementation An agreed outcome becomes a maintained plan, direct or delegated implementation, proportional verification and an honest delivery report. Task context and dependency advice support progress; real user stops and host permissions remain effective. Greenfield and brownfield population use the paired native writer and exact finalization. Exact blob bindings identify inspected source bytes; the linked revision is the committed source base. Implementation, verification, live host qualification and release remain separate claims.
Current source changes A review binds one resolved diff and supplies scoped inputs to each reviewer. Missing critical references or applicable conflicts prevent a substantive verdict; provenance and unobserved delivery diagnostics remain advisory. core/surface/commands/review.md Predict fleet / review actual diff /checkpoint persists a sweep Review verdict is not a commit
Related entries /preview /review /checkpoint
Pinned evidence S04 · core/surface/includes/routing-table.mdAuthoritative intent-to-owner table, scope-touch routes and stop/return semantics. S18 · core/surface/includes/review-matrix.mdShared diff-path reviewer selection for review/checkpoint/preview; callers retain their own gates. S22 · core/surface/agents/INDEX.mdRole separation; specialist reviewers; restricted scouts/graders; authority-reviewer host restriction. Selected cmd
/threat-model · /tribunal Threat modeling retains its existing specialized owner. Tribunal is a separate opt-in source-bound audit: applicability across thirteen cards, independent verification of serious claims, a report and separately authorized follow-ups. See the Tribunal lifecycle view. Neither is an automatic stage of every feature.
Related entries /threat-model /tribunal
Pinned evidence S29 · core/surface/skills/tribunal/SKILL.mdSource-bound opt-in audit, guarded hash-bound stranded-lock recovery, thirteen-card roster, review/verify, durable triage, report and separate follow-up completion. S04 · core/surface/includes/routing-table.mdAuthoritative intent-to-owner table, scope-touch routes and stop/return semantics. Selected cmd
/adr · /adr-status decision-lifecycle Attributed decision / health report
Related entries /adr /adr-status
Pinned evidence S04 · core/surface/includes/routing-table.mdAuthoritative intent-to-owner table, scope-touch routes and stop/return semantics. S21 · core/surface/skills/INDEX.mdCurrent routine-owner index: five debug dispositions, bounded workflow owners and unchanged six-phase TDD / nine-phase commit sequence. Selected cmd
/reconcile decision-variance scout → grader → user ruling Report-only writes nothing
Related entries /reconcile
Pinned evidence S04 · core/surface/includes/routing-table.mdAuthoritative intent-to-owner table, scope-touch routes and stop/return semantics. S21 · core/surface/skills/INDEX.mdCurrent routine-owner index: five debug dispositions, bounded workflow owners and unchanged six-phase TDD / nine-phase commit sequence. S22 · core/surface/agents/INDEX.mdRole separation; specialist reviewers; restricted scouts/graders; authority-reviewer host restriction. Selected gate
/conflict Internal protocol Contradicting rules: STOP User resolves the authority gap
Related entries /conflict
Pinned evidence S03 · core/surface/arbiter.mdIntent-to-owner routing, on-demand loading, caller extent and mode boundary. S04 · core/surface/includes/routing-table.mdAuthoritative intent-to-owner table, scope-touch routes and stop/return semantics. Selected cmd
/status · /context-check Assistive workflow: source-bound implementation An agreed outcome becomes a maintained plan, direct or delegated implementation, proportional verification and an honest delivery report. Task context and dependency advice support progress; real user stops and host permissions remain effective. Greenfield and brownfield population use the paired native writer and exact finalization. Exact blob bindings identify inspected source bytes; the linked revision is the committed source base. Implementation, verification, live host qualification and release remain separate claims.
Current source changes The user-facing route table points feature and sprint work at useful planned execution and selected context refresh at its existing owner. Existing authorization covers ordinary work; actual host, destructive-action and commit controls remain. core/surface/includes/routing-table.md A drift audit stays read-only; an already authorized selected or full refresh proceeds without asking again per document. The installed draft/preview/apply example preserves source bindings and selected content. Matching hashes establish identity only; initialized refresh never reruns scaffolding, clears the marker or repeats finalization. core/surface/skills/context-check/SKILL.md v1/v2 state and drift report Explicit scoped/full refresh selection Native v2 pairs; preserve history
Related entries /status /context-check
Pinned evidence S26 · core/surface/skills/context-check/SKILL.mdInitialized refresh: read-only report, explicit selection, v1/v2 separation, native paired updates and preserved human history. Full owner read. S13 · core/surface/skills/context-creation/SKILL.mdCurrent brownfield owner: qualified containment, bounded reports and join, scoped gaps, native document/provenance writes, separate 13-identity finalization and recovery. Selected cmd
/task · /standup Sanctioned task-board transitions Hygiene with per-action consent
Related entries /task /standup
Pinned evidence S02 · core/surface/COMMANDS.mdCatalog, modes, host-limited commands and purpose of each public route. S04 · core/surface/includes/routing-table.mdAuthoritative intent-to-owner table, scope-touch routes and stop/return semantics. Selected cmd
/audit · /metrics Assistive workflow: source-bound implementation An agreed outcome becomes a maintained plan, direct or delegated implementation, proportional verification and an honest delivery report. Task context and dependency advice support progress; real user stops and host permissions remain effective. Greenfield and brownfield population use the paired native writer and exact finalization. Exact blob bindings identify inspected source bytes; the linked revision is the committed source base. Implementation, verification, live host qualification and release remain separate claims.
Current source changes A checkpoint retains the reviewed commit, UTC review-start time and worktree dirty state, then writes the report through its aggregator. last-checkpoint holds only the override count. core/surface/commands/checkpoint.md Checkpoint audit windows use the report timestamp and an ancestor commit. Invalid or missing anchors require an explicit ref/date window; dirty review scope remains a stated limitation. core/surface/commands/audit.md Dated source-record packet Windowed governance trends
Related entries /audit /metrics
Pinned evidence S02 · core/surface/COMMANDS.mdCatalog, modes, host-limited commands and purpose of each public route. S04 · core/surface/includes/routing-table.mdAuthoritative intent-to-owner table, scope-touch routes and stop/return semantics. Selected cmd
/doctor · /statusline Install / payload diagnostics Statusline is Claude-specific
Related entries /doctor /statusline
Pinned evidence S02 · core/surface/COMMANDS.mdCatalog, modes, host-limited commands and purpose of each public route. S10 · core/surface/includes/artifacts.mdInstalled admission, typed reads, context completion, authority receipts and format boundaries. Selected cmd
/override · /prune Permitted, logged exception Host-specific context compaction
Related entries /override /prune
Pinned evidence S02 · core/surface/COMMANDS.mdCatalog, modes, host-limited commands and purpose of each public route. S03 · core/surface/arbiter.mdIntent-to-owner routing, on-demand loading, caller extent and mode boundary. S04 · core/surface/includes/routing-table.mdAuthoritative intent-to-owner table, scope-touch routes and stop/return semantics. Selected cmd
/feature or /sprint Assistive workflow: source-bound implementation An agreed outcome becomes a maintained plan, direct or delegated implementation, proportional verification and an honest delivery report. Task context and dependency advice support progress; real user stops and host permissions remain effective. Greenfield and brownfield population use the paired native writer and exact finalization. Exact blob bindings identify inspected source bytes; the linked revision is the committed source base. Implementation, verification, live host qualification and release remain separate claims.
Current source changes Routine artifact instructions now focus on the installed writer, typed plan, bounded task reads, factual prerequisites, actual checks and safe retry. The default leaf is about 7.3 KB and includes an executable stdin bootstrap with optional operation-schema discovery; explicit attestation remains available on demand without burdening ordinary task context. A spec's SMARTS design rationale lives in its approach, decisions and one SEC-SMARTS section. core/surface/includes/artifacts.md The pinned native bridge enables task-context as a bounded informational read and does not require future reviewer/approval producers simply to author a structurally valid draft. Context requests route to ordinary context-preview, context-finalize-preview and exact-reference context-apply; historical receipt input stays a separate compatibility path. Installed host qualification is separate. Native Pi writer admission does not establish a complete live Pi workflow. On NO_LOCAL_STORE the client runs init-store once and retries the request once; a refused initialization keeps NO_LOCAL_STORE. On Windows the binary's architecture comes from the kernel (IsWow64Process2), never WMI or PROCESSOR_* environment. Index and outline pages are now fully checked before rows are yielded, including continuation identity and row shape. core/pysrc/_artifactlib.py The user-facing route table points feature and sprint work at useful planned execution and selected context refresh at its existing owner. Existing authorization covers ordinary work; actual host, destructive-action and commit controls remain. core/surface/includes/routing-table.md The sprint entry now routes to the assistive implementation body instead of requiring a fresh user approval and two reviews per task. core/surface/commands/sprint.md The feature entry now resumes a maintained spec/plan and task progress without another artifact certification prompt; independent commit and test controls remain. The small lane runs an inline SMARTS scan only for a material solution choice. core/surface/commands/feature.md The command catalog starts from the agreed outcome and maintained plan; SMARTS is useful for difficult tradeoffs and does not authorize a changed outcome. core/surface/COMMANDS.md Resolve same-slug artifacts Mixed / duplicate matches: STOP Exact pair or lone spec keeps format
Related entries /feature /sprint
Pinned evidence S06 · core/surface/commands/feature.mdCurrent feature owner: same-format resume and full/small gates; adds actual per-actor context composition. S07 · core/surface/SPRINT.mdInitial paired approval; full-plan execution; delegated methods; recovery and farm boundaries. S10 · core/surface/includes/artifacts.mdInstalled admission, typed reads, context completion, authority receipts and format boundaries. Selected state
Resume checkpoint Draft → approval Approved spec → planning Remaining tasks → execution All accepted → current-proof gate
Related entries /feature /sprint
Pinned evidence S06 · core/surface/commands/feature.mdCurrent feature owner: same-format resume and full/small gates; adds actual per-actor context composition. S07 · core/surface/SPRINT.mdInitial paired approval; full-plan execution; delegated methods; recovery and farm boundaries. Selected skill
brainstorming Assistive workflow: source-bound implementation An agreed outcome becomes a maintained plan, direct or delegated implementation, proportional verification and an honest delivery report. Task context and dependency advice support progress; real user stops and host permissions remain effective. Greenfield and brownfield population use the paired native writer and exact finalization. Exact blob bindings identify inspected source bytes; the linked revision is the committed source base. Implementation, verification, live host qualification and release remain separate claims.
Current source changes Routine artifact instructions now focus on the installed writer, typed plan, bounded task reads, factual prerequisites, actual checks and safe retry. The default leaf is about 7.3 KB and includes an executable stdin bootstrap with optional operation-schema discovery; explicit attestation remains available on demand without burdening ordinary task context. A spec's SMARTS design rationale lives in its approach, decisions and one SEC-SMARTS section. core/surface/includes/artifacts.md The skill index connects planning, direct or delegated implementation, proportional verification and selected context maintenance to their owning workflows. core/surface/skills/INDEX.md The feature entry now resumes a maintained spec/plan and task progress without another artifact certification prompt; independent commit and test controls remain. The small lane runs an inline SMARTS scan only for a material solution choice. core/surface/commands/feature.md Source refinement respects recorded deferrals and accepted ADR conflicts without turning each ordinary method adjustment into a separate authorization interview. SMARTS compares approaches only when two or more are materially plausible; a single approach gets one line and no table. core/surface/skills/brainstorming/SKILL.md Confirmed fixes require regression RED/GREEN. Features use current criterion-focused checks and meaningful negative controls where needed; refactors preserve passing characterization baselines. Coverage, lint and applicable security requirements remain explicit. core/surface/skills/tdd/SKILL.md Recorded intent + user criteria Unknowns → open-questions No guessed product decisions
Related entries /feature /sprint
Pinned evidence S06 · core/surface/commands/feature.mdCurrent feature owner: same-format resume and full/small gates; adds actual per-actor context composition. S07 · core/surface/SPRINT.mdInitial paired approval; full-plan execution; delegated methods; recovery and farm boundaries. S10 · core/surface/includes/artifacts.mdInstalled admission, typed reads, context completion, authority receipts and format boundaries. Selected skill
writing-plans Assistive workflow: source-bound implementation An agreed outcome becomes a maintained plan, direct or delegated implementation, proportional verification and an honest delivery report. Task context and dependency advice support progress; real user stops and host permissions remain effective. Greenfield and brownfield population use the paired native writer and exact finalization. Exact blob bindings identify inspected source bytes; the linked revision is the committed source base. Implementation, verification, live host qualification and release remain separate claims.
Current source changes Routine artifact instructions now focus on the installed writer, typed plan, bounded task reads, factual prerequisites, actual checks and safe retry. The default leaf is about 7.3 KB and includes an executable stdin bootstrap with optional operation-schema discovery; explicit attestation remains available on demand without burdening ordinary task context. A spec's SMARTS design rationale lives in its approach, decisions and one SEC-SMARTS section. core/surface/includes/artifacts.md Plans preserve agreed criteria, useful tasks, dependency ancestry, bounded shared inputs and observable checks. Stable source identities and typed maintenance keep changes within the selected outcome. A design-drift guard routes an undecided material solution choice back to brainstorming instead of scoring it during planning. core/surface/skills/writing-plans/SKILL.md The pinned native bridge enables task-context as a bounded informational read and does not require future reviewer/approval producers simply to author a structurally valid draft. Context requests route to ordinary context-preview, context-finalize-preview and exact-reference context-apply; historical receipt input stays a separate compatibility path. Installed host qualification is separate. Native Pi writer admission does not establish a complete live Pi workflow. On NO_LOCAL_STORE the client runs init-store once and retries the request once; a refused initialization keeps NO_LOCAL_STORE. On Windows the binary's architecture comes from the kernel (IsWow64Process2), never WMI or PROCESSOR_* environment. Index and outline pages are now fully checked before rows are yielded, including continuation identity and row shape. core/pysrc/_artifactlib.py The skill index connects planning, direct or delegated implementation, proportional verification and selected context maintenance to their owning workflows. core/surface/skills/INDEX.md The feature entry now resumes a maintained spec/plan and task progress without another artifact certification prompt; independent commit and test controls remain. The small lane runs an inline SMARTS scan only for a material solution choice. core/surface/commands/feature.md Ordinary approval and sprint instructions bypass legacy capture without changing pending history or inventing authority. Explicit legacy replies retain exact matching. The paired adapter can retire a genuinely unanswered 0.2 sprint-pair request using its serialized cancellation lock and submission check; submitted decisions must replay, while standalone approval 0.1 refuses automatic retirement. core/pysrc/_approvallib.py Ordinary prerequisite instructions bypass legacy capture. The existing adapter can retire only a validated ARMED pending question under its original per-plan OS lock. Confirmed or captured decisions are not disposable; no authority receipt or new native permission API is created. core/pysrc/_prerequisitelib.py Exact criterion IDs → tasks Paths • tests • dependencies Checkpoints • prerequisites
Related entries /feature /sprint
Pinned evidence S06 · core/surface/commands/feature.mdCurrent feature owner: same-format resume and full/small gates; adds actual per-actor context composition. S07 · core/surface/SPRINT.mdInitial paired approval; full-plan execution; delegated methods; recovery and farm boundaries. S10 · core/surface/includes/artifacts.mdInstalled admission, typed reads, context completion, authority receipts and format boundaries. S21 · core/surface/skills/INDEX.mdCurrent routine-owner index: five debug dispositions, bounded workflow owners and unchanged six-phase TDD / nine-phase commit sequence. Selected gate
Approval is observed Assistive workflow: source-bound implementation An agreed outcome becomes a maintained plan, direct or delegated implementation, proportional verification and an honest delivery report. Task context and dependency advice support progress; real user stops and host permissions remain effective. Greenfield and brownfield population use the paired native writer and exact finalization. Exact blob bindings identify inspected source bytes; the linked revision is the committed source base. Implementation, verification, live host qualification and release remain separate claims.
Current source changes Routine artifact instructions now focus on the installed writer, typed plan, bounded task reads, factual prerequisites, actual checks and safe retry. The default leaf is about 7.3 KB and includes an executable stdin bootstrap with optional operation-schema discovery; explicit attestation remains available on demand without burdening ordinary task context. A spec's SMARTS design rationale lives in its approach, decisions and one SEC-SMARTS section. core/surface/includes/artifacts.md The pinned native bridge enables task-context as a bounded informational read and does not require future reviewer/approval producers simply to author a structurally valid draft. Context requests route to ordinary context-preview, context-finalize-preview and exact-reference context-apply; historical receipt input stays a separate compatibility path. Installed host qualification is separate. Native Pi writer admission does not establish a complete live Pi workflow. On NO_LOCAL_STORE the client runs init-store once and retries the request once; a refused initialization keeps NO_LOCAL_STORE. On Windows the binary's architecture comes from the kernel (IsWow64Process2), never WMI or PROCESSOR_* environment. Index and outline pages are now fully checked before rows are yielded, including continuation identity and row shape. core/pysrc/_artifactlib.py The sprint caller performs an index-first recorded-intent read, preserves actual owner decisions and ADR conflicts, then proceeds with authorized work without a second artifact approval system. A SMARTS Unknown names its missing observation and cannot bypass a hard gate or manufacture authority. core/surface/SPRINT.md Source refinement respects recorded deferrals and accepted ADR conflicts without turning each ordinary method adjustment into a separate authorization interview. SMARTS compares approaches only when two or more are materially plausible; a single approach gets one line and no table. core/surface/skills/brainstorming/SKILL.md Ordinary approval and sprint instructions bypass legacy capture without changing pending history or inventing authority. Explicit legacy replies retain exact matching. The paired adapter can retire a genuinely unanswered 0.2 sprint-pair request using its serialized cancellation lock and submission check; submitted decisions must replay, while standalone approval 0.1 refuses automatic retirement. core/pysrc/_approvallib.py The old paired approval producer publishes a content-addressed observed reply before its durable submission. Cancellation checks the matching observation, content hash, prompt and frozen context under its existing lock, with a strict scan bound. A matching or uncertain pre-submission reply blocks automatic retirement; normal tasks and context reads have no added cost. core/pysrc/_sprintapprovallib.py The optional reference retains strict historical evidence inputs without replacing ordinary work. Explicit draft source refresh preserves progressed plans and history with both CAS bindings; initial approval restrictions remain on historical paired attestation. Its verify command requires literal absolute paths; linked agent subdirectories remain unsupported for strict authority review. These limitations apply to requested attestation, not ordinary task start. core/surface/includes/artifacts-attestation.md Historical sprint approval and SMARTS proof instructions remain available only for existing explicit attestation, outside the ordinary execution path. core/surface/includes/sprint-authority.md Ordinary uppercase words such as Approve DATA migration remain ordinary input. All-letter capture codes require a matching retained request; real armed replies, mismatches and user stops preserve their original handling. core/pysrc/_replylib.py Only actual host-observed approval over the current armed identity confers the corresponding authority. Pi 1.0.2 has the runtime input seam but codeArbiter has no implemented Pi-native producer; typed admission remains blocked. For prerequisites, show the packet digest and the shared 24-hour request/reply deadline. Cancellation does not create approval.
Related entries /feature /sprint
Pinned evidence S10 · core/surface/includes/artifacts.mdInstalled admission, typed reads, context completion, authority receipts and format boundaries. S31 · docs/parity.mdCurrent Pi 1.0.2 target, separate host counts, DEGRADED typed approval and Codex farm exception. Selected cmd
Small /feature branch All small-lane conditions → confirm inline criteria → log triage → direct TDD. No spec artifact.
Related entries /feature
Pinned evidence S06 · core/surface/commands/feature.mdCurrent feature owner: same-format resume and full/small gates; adds actual per-actor context composition. Selected gate
Installed admission is required Assistive workflow: source-bound implementation An agreed outcome becomes a maintained plan, direct or delegated implementation, proportional verification and an honest delivery report. Task context and dependency advice support progress; real user stops and host permissions remain effective. Greenfield and brownfield population use the paired native writer and exact finalization. Exact blob bindings identify inspected source bytes; the linked revision is the committed source base. Implementation, verification, live host qualification and release remain separate claims.
Current source changes Routine artifact instructions now focus on the installed writer, typed plan, bounded task reads, factual prerequisites, actual checks and safe retry. The default leaf is about 7.3 KB and includes an executable stdin bootstrap with optional operation-schema discovery; explicit attestation remains available on demand without burdening ordinary task context. A spec's SMARTS design rationale lives in its approach, decisions and one SEC-SMARTS section. core/surface/includes/artifacts.md The pinned native bridge enables task-context as a bounded informational read and does not require future reviewer/approval producers simply to author a structurally valid draft. Context requests route to ordinary context-preview, context-finalize-preview and exact-reference context-apply; historical receipt input stays a separate compatibility path. Installed host qualification is separate. Native Pi writer admission does not establish a complete live Pi workflow. On NO_LOCAL_STORE the client runs init-store once and retries the request once; a refused initialization keeps NO_LOCAL_STORE. On Windows the binary's architecture comes from the kernel (IsWow64Process2), never WMI or PROCESSOR_* environment. Index and outline pages are now fully checked before rows are yielded, including continuation identity and row shape. core/pysrc/_artifactlib.py Pi HTML admission unavailable; legacy Markdown and inline small lane stay distinct. HTML farm remains blocked.
Related entries /feature /sprint
Pinned evidence S10 · core/surface/includes/artifacts.mdInstalled admission, typed reads, context completion, authority receipts and format boundaries. Selected skill
executing-plans / attended /feature: authored HTML checkpoints or bounded legacy batches Pass selected task IDs; wait for the user between batches
Related entries /feature
Pinned evidence S06 · core/surface/commands/feature.mdCurrent feature owner: same-format resume and full/small gates; adds actual per-actor context composition. S21 · core/surface/skills/INDEX.mdCurrent routine-owner index: five debug dispositions, bounded workflow owners and unchanged six-phase TDD / nine-phase commit sequence. S08 · core/surface/skills/subagent-driven-development/SKILL.mdCurrent implementation-dispatch and independent spec-review sections re-read; remaining scope contract also checked against current skill index. Fresh actor input and bounded delivery are distinct from engine tickets. Selected skill
SPRINT.md / approved autonomy /sprint: pass the full plan; no attended batch pauses SMARTS logs choices; delegated methods cannot widen scope
Related entries /sprint
Pinned evidence S07 · core/surface/SPRINT.mdInitial paired approval; full-plan execution; delegated methods; recovery and farm boundaries. Selected skill
subagent-driven-development Assistive workflow: source-bound implementation An agreed outcome becomes a maintained plan, direct or delegated implementation, proportional verification and an honest delivery report. Task context and dependency advice support progress; real user stops and host permissions remain effective. Greenfield and brownfield population use the paired native writer and exact finalization. Exact blob bindings identify inspected source bytes; the linked revision is the committed source base. Implementation, verification, live host qualification and release remain separate claims.
Current source changes Routine artifact instructions now focus on the installed writer, typed plan, bounded task reads, factual prerequisites, actual checks and safe retry. The default leaf is about 7.3 KB and includes an executable stdin bootstrap with optional operation-schema discovery; explicit attestation remains available on demand without burdening ordinary task context. A spec's SMARTS design rationale lives in its approach, decisions and one SEC-SMARTS section. core/surface/includes/artifacts.md The user-facing route table points feature and sprint work at useful planned execution and selected context refresh at its existing owner. Existing authorization covers ordinary work; actual host, destructive-action and commit controls remain. core/surface/includes/routing-table.md The skill index connects planning, direct or delegated implementation, proportional verification and selected context maintenance to their owning workflows. core/surface/skills/INDEX.md The sprint caller performs an index-first recorded-intent read, preserves actual owner decisions and ADR conflicts, then proceeds with authorized work without a second artifact approval system. A SMARTS Unknown names its missing observation and cannot bypass a hard gate or manufacture authority. core/surface/SPRINT.md The sprint entry now routes to the assistive implementation body instead of requiring a fresh user approval and two reviews per task. core/surface/commands/sprint.md The feature entry now resumes a maintained spec/plan and task progress without another artifact certification prompt; independent commit and test controls remain. The small lane runs an inline SMARTS scan only for a material solution choice. core/surface/commands/feature.md Attended feature execution now treats checkpoints as progress groups, not compulsory approval stops; it forwards real failures and continues independent work. core/surface/skills/executing-plans/SKILL.md Implementation works directly or delegates a coherent unit when useful. Bounded task context, applicable dependency advice and observed checks support progress while preserving actual user stops and host permissions. Verification checks the change against a recorded SMARTS rationale for material invalidation only. core/surface/skills/subagent-driven-development/SKILL.md Read-only task-context reports direct dependencies and unresolved ancestors, scoped_changed, scoped_unchanged or unassessed inputs, and actual check observations. Missing broader context remains explicit; the packet does not certify verification. core/artifacts/internal/operations/task_context.go Exact and batch symbol retrieval exposes reported check history without an approval ticket. Outline paging now accounts for encoded rows incrementally while retaining response budgets and continuation behavior. core/artifacts/internal/read/read.go Ordinary prerequisite instructions bypass legacy capture. The existing adapter can retire only a validated ARMED pending question under its original per-plan OS lock. Confirmed or captured decisions are not disposable; no authority receipt or new native permission API is created. core/pysrc/_prerequisitelib.py Resident instructions preserve existing work authorization and actual owner decisions without another spec approval or a SMARTS record for every routine step. core/surface/arbiter.md Codex dispatch requires attestation only for explicitly claimed host proof; an unavailable requested farm backend is reported without silently switching provider or model. Ordinary context inspection can proceed inline within actual host permissions; an agent label supplies no containment or semantic verification. core/surface/includes/codex-host-notes.md The public Codex farm boundary matches its caller: report the unavailable backend and use another provider or model only within existing authorization. docs/parity.md Authors and reviewers share task-critical reference and constraint checks. Missing or stale provenance remains advisory. Two delivery attempts end with an advisory unknown observation, without another retry or an invented receipt. core/pysrc/_contextselectlib.py Feature actors receive scoped requirements, named critical references and observed evidence. Applicable conflicts or missing required inputs stop dependent work; provenance warnings remain visible and advisory. core/pysrc/_artifactpromptlib.py Scheduling follows the full dependency ancestry and exposes unresolved ancestor IDs and reasons. Reconciled progress remains advisory; blocked ancestors prevent premature downstream readiness while independent tasks remain available. Response-budget omissions are explicit, retain the global scheduling decision and direct callers to supported detail reads. core/artifacts/internal/operations/assistive.go Resolve current identities → eligible tasks only HTML: complete contextual read ticket → task-start Then separately compose and deliver each actor’s scoped input Legacy Markdown keeps its own dependency/state rules
Related entries /feature /sprint
Pinned evidence S08 · core/surface/skills/subagent-driven-development/SKILL.mdCurrent implementation-dispatch and independent spec-review sections re-read; remaining scope contract also checked against current skill index. Fresh actor input and bounded delivery are distinct from engine tickets. S27 · core/pysrc/_contextselectlib.pyInert scoped selector and delivery implementation: commands/cwd/evidence, native instruction order, actor/worktree/epoch binding, two-attempt limit and reviewer preflight. S10 · core/surface/includes/artifacts.mdInstalled admission, typed reads, context completion, authority receipts and format boundaries. Selected agent
Fresh scoped author input Map + commands + constraints Native instructions; actual actor delivery No inherited prior-task narrative
Related entries /feature /sprint /fix
Pinned evidence S08 · core/surface/skills/subagent-driven-development/SKILL.mdCurrent implementation-dispatch and independent spec-review sections re-read; remaining scope contract also checked against current skill index. Fresh actor input and bounded delivery are distinct from engine tickets. S27 · core/pysrc/_contextselectlib.pyInert scoped selector and delivery implementation: commands/cwd/evidence, native instruction order, actor/worktree/epoch binding, two-attempt limit and reviewer preflight. Selected skill
tdd / six gates 1 Obligations → 2 Red → 3 Green 4 Obligation verify → 5 Coverage → 6 Lint Missing proof loops back to the failing gate Coverage-auditor where contract/security critical
Related entries /feature /sprint /fix /refactor
Pinned evidence S09 · core/surface/skills/tdd/SKILL.mdCurrent TDD context preflight and obligation derivation re-read. Six-phase gate sequence also confirmed by current skill index; fix and test-only paths do not fabricate feature artifacts. S21 · core/surface/skills/INDEX.mdCurrent routine-owner index: five debug dispositions, bounded workflow owners and unchanged six-phase TDD / nine-phase commit sequence. Selected agent
Spec-compliance review Assistive workflow: source-bound implementation An agreed outcome becomes a maintained plan, direct or delegated implementation, proportional verification and an honest delivery report. Task context and dependency advice support progress; real user stops and host permissions remain effective. Greenfield and brownfield population use the paired native writer and exact finalization. Exact blob bindings identify inspected source bytes; the linked revision is the committed source base. Implementation, verification, live host qualification and release remain separate claims.
Current source changes A new explicit batch review can select a passing rerun of still-reviewed work without replacing the original task review. The selected run must match its own recorded workspace; changed source or supporting evidence still requires reassessment. core/artifacts/internal/evidence/completion.go Implementation works directly or delegates a coherent unit when useful. Bounded task context, applicable dependency advice and observed checks support progress while preserving actual user stops and host permissions. Verification checks the change against a recorded SMARTS rationale for material invalidation only. core/surface/skills/subagent-driven-development/SKILL.md Read-only task-context reports direct dependencies and unresolved ancestors, scoped_changed, scoped_unchanged or unassessed inputs, and actual check observations. Missing broader context remains explicit; the packet does not certify verification. core/artifacts/internal/operations/task_context.go Independent task-spec review and actual source-bound observations still precede acceptance. The current source repairs Claude launch/child correlation and native Unicode observation hashing; neither repair substitutes an author’s report for a reviewer receipt. Re-arm requests made stale by the applicable engine upgrade rather than relabeling old evidence.
Related entries /feature /sprint
Pinned evidence S08 · core/surface/skills/subagent-driven-development/SKILL.mdCurrent implementation-dispatch and independent spec-review sections re-read; remaining scope contract also checked against current skill index. Fresh actor input and bounded delivery are distinct from engine tickets. S32 · core/pysrc/_artifactauthoritylib.pyNative artifact ID grammar, producer states, Claude join lock, UTF-8/UTF-16-key-order observation binding. S40 · CHANGELOG.mdMerge-scope record for review-join, marker, ID/reorder, Unicode, prerequisite and engine-wait repairs. Selected gate
Fresh task verification Assistive workflow: source-bound implementation An agreed outcome becomes a maintained plan, direct or delegated implementation, proportional verification and an honest delivery report. Task context and dependency advice support progress; real user stops and host permissions remain effective. Greenfield and brownfield population use the paired native writer and exact finalization. Exact blob bindings identify inspected source bytes; the linked revision is the committed source base. Implementation, verification, live host qualification and release remain separate claims.
Current source changes Routine artifact instructions now focus on the installed writer, typed plan, bounded task reads, factual prerequisites, actual checks and safe retry. The default leaf is about 7.3 KB and includes an executable stdin bootstrap with optional operation-schema discovery; explicit attestation remains available on demand without burdening ordinary task context. A spec's SMARTS design rationale lives in its approach, decisions and one SEC-SMARTS section. core/surface/includes/artifacts.md A new explicit batch review can select a passing rerun of still-reviewed work without replacing the original task review. The selected run must match its own recorded workspace; changed source or supporting evidence still requires reassessment. core/artifacts/internal/evidence/completion.go The skill index connects planning, direct or delegated implementation, proportional verification and selected context maintenance to their owning workflows. core/surface/skills/INDEX.md Implementation works directly or delegates a coherent unit when useful. Bounded task context, applicable dependency advice and observed checks support progress while preserving actual user stops and host permissions. Verification checks the change against a recorded SMARTS rationale for material invalidation only. core/surface/skills/subagent-driven-development/SKILL.md Confirmed fixes require regression RED/GREEN. Features use current criterion-focused checks and meaningful negative controls where needed; refactors preserve passing characterization baselines. Coverage, lint and applicable security requirements remain explicit. core/surface/skills/tdd/SKILL.md Read-only task-context reports direct dependencies and unresolved ancestors, scoped_changed, scoped_unchanged or unassessed inputs, and actual check observations. Missing broader context remains explicit; the packet does not certify verification. core/artifacts/internal/operations/task_context.go Task-check observations now retain directly referenced criterion identity and an optional bounded source-input fingerprint or an explicit unassessed reason, without executing tests or minting approval receipts. core/artifacts/internal/operations/task_check.go New checks bind the selected criterion and applicable constraint bodies, including explicit references and scope applicability. Unrelated constraints remain outside the digest; old observations never gain unassessed constraint coverage. core/artifacts/internal/model/assistive.go The closed task-check request accepts bounded command observations matched to plan definitions. Ordinary context-preview and context-finalize-preview omit prompt identity; context-apply accepts an exact preview reference and digest. Invalid requests, changed preimages and stale CAS writes remain rejected. The historical smarts-apply request accepts the 0.2.0 SMARTS decision record. init-store takes a protocol-only request. core/artifacts/internal/operations/protocol.go The unpublished 0.3.2 native schema adds optional shared_inputs and scoped task-check fields without silently relabeling old records. core/artifacts/schemas/common.schema.json The bounded source-fingerprint reader hashes actual declared task/shared paths, directory membership and known build-control files; unsafe, missing and over-limit scopes are explicitly unassessed rather than blocking work or proving a full dependency graph. core/artifacts/internal/operations/input_scope.go Delegated authors preserve the caller-selected verification approach: genuine regression RED/GREEN for fixes, current criterion-focused feature checks and passing characterization baselines for refactors. Actual results and pending checks return to the caller; staging requires assigned scope and authorization. core/surface/includes/author-tdd-workflow.md Characterization tests can pass against existing correct behavior. Preserve behavioral assertions against the post-backfill baseline; mechanical import or symbol-reference updates retain the same oracle. Coverage and parity checks remain required. A precise surface table proceeds under existing user authorization; only material expansion or a changed outcome needs another decision. Distinct restructuring strategies are compared with SMARTS only after surface and parity are settled. core/surface/skills/refactor/SKILL.md Only an actual verifier wrapper executable/script prefix selects authority handling. Inert PowerShell data mentioning the filename is not a verifier call; malformed real verifier invocations still fail closed. Fresh legacy context approval arming is refused; historical pending routes remain readable. core/pysrc/_artifactauthoritylib.py Run the plan’s proving command again Read output and exit code HTML: real producer-owned receipts
Related entries /feature /sprint
Pinned evidence S08 · core/surface/skills/subagent-driven-development/SKILL.mdCurrent implementation-dispatch and independent spec-review sections re-read; remaining scope contract also checked against current skill index. Fresh actor input and bounded delivery are distinct from engine tickets. S10 · core/surface/includes/artifacts.mdInstalled admission, typed reads, context completion, authority receipts and format boundaries. Selected state
Task proof / provisional review HTML task-review binds spec-review + verification receipts. REVIEW is not ACCEPTED.
Related entries /feature /sprint
Pinned evidence S08 · core/surface/skills/subagent-driven-development/SKILL.mdCurrent implementation-dispatch and independent spec-review sections re-read; remaining scope contract also checked against current skill index. Fresh actor input and bounded delivery are distinct from engine tickets. Selected skill
Next eligible task in the same scope Repeat fresh author → spec proof → fresh verification. No context accumulation across task authors.
Related entries /feature /sprint
Pinned evidence S08 · core/surface/skills/subagent-driven-development/SKILL.mdCurrent implementation-dispatch and independent spec-review sections re-read; remaining scope contract also checked against current skill index. Fresh actor input and bounded delivery are distinct from engine tickets. Selected gate
Correction is not permission to waive a gate Assistive workflow: source-bound implementation An agreed outcome becomes a maintained plan, direct or delegated implementation, proportional verification and an honest delivery report. Task context and dependency advice support progress; real user stops and host permissions remain effective. Greenfield and brownfield population use the paired native writer and exact finalization. Exact blob bindings identify inspected source bytes; the linked revision is the committed source base. Implementation, verification, live host qualification and release remain separate claims.
Current source changes Routine artifact instructions now focus on the installed writer, typed plan, bounded task reads, factual prerequisites, actual checks and safe retry. The default leaf is about 7.3 KB and includes an executable stdin bootstrap with optional operation-schema discovery; explicit attestation remains available on demand without burdening ordinary task context. A spec's SMARTS design rationale lives in its approach, decisions and one SEC-SMARTS section. core/surface/includes/artifacts.md The sprint caller performs an index-first recorded-intent read, preserves actual owner decisions and ADR conflicts, then proceeds with authorized work without a second artifact approval system. A SMARTS Unknown names its missing observation and cannot bypass a hard gate or manufacture authority. core/surface/SPRINT.md Attended feature execution now treats checkpoints as progress groups, not compulsory approval stops; it forwards real failures and continues independent work. core/surface/skills/executing-plans/SKILL.md Implementation works directly or delegates a coherent unit when useful. Bounded task context, applicable dependency advice and observed checks support progress while preserving actual user stops and host permissions. Verification checks the change against a recorded SMARTS rationale for material invalidation only. core/surface/skills/subagent-driven-development/SKILL.md An unanswered legacy reconciliation can be superseded only when route and pending state agree and no durable attempt or matching observed reply exists. Captured replies recover through their original immutable source and receipt; in-flight mutations retain exact replay and real user stops. core/pysrc/_reconciliationlib.py The installed recover command finishes a captured original reply or replays an in-flight mutation without asking for a new user answer. core/pysrc/artifact-reconcile.py The closed task-check request accepts bounded command observations matched to plan definitions. Ordinary context-preview and context-finalize-preview omit prompt identity; context-apply accepts an exact preview reference and digest. Invalid requests, changed preimages and stale CAS writes remain rejected. The historical smarts-apply request accepts the 0.2.0 SMARTS decision record. init-store takes a protocol-only request. core/artifacts/internal/operations/protocol.go The candidate-only runbook distinguishes 0.3.1/0.3.2 mixed-schema readers, exact replay and forward recovery while preserving recorded user stops. The paired prompt retirement scans immutable observed replies before calling a request unanswered; submitted or uncertain decisions are never erased. Older-client downgrade after modern edits remains unsupported. docs/artifacts/ASSISTIVE-UPGRADE.md Ordinary approval and sprint instructions bypass legacy capture without changing pending history or inventing authority. Explicit legacy replies retain exact matching. The paired adapter can retire a genuinely unanswered 0.2 sprint-pair request using its serialized cancellation lock and submission check; submitted decisions must replay, while standalone approval 0.1 refuses automatic retirement. core/pysrc/_approvallib.py Ordinary prerequisite instructions bypass legacy capture. The existing adapter can retire only a validated ARMED pending question under its original per-plan OS lock. Confirmed or captured decisions are not disposable; no authority receipt or new native permission API is created. core/pysrc/_prerequisitelib.py The old paired approval producer publishes a content-addressed observed reply before its durable submission. Cancellation checks the matching observation, content hash, prompt and frozen context under its existing lock, with a strict scan bound. A matching or uncertain pre-submission reply blocks automatic retirement; normal tasks and context reads have no added cost. core/pysrc/_sprintapprovallib.py The optional reference retains strict historical evidence inputs without replacing ordinary work. Explicit draft source refresh preserves progressed plans and history with both CAS bindings; initial approval restrictions remain on historical paired attestation. Its verify command requires literal absolute paths; linked agent subdirectories remain unsupported for strict authority review. These limitations apply to requested attestation, not ordinary task start. core/surface/includes/artifacts-attestation.md Scheduling follows the full dependency ancestry and exposes unresolved ancestor IDs and reasons. Reconciled progress remains advisory; blocked ancestors prevent premature downstream readiness while independent tasks remain available. Response-budget omissions are explicit, retain the global scheduling decision and direct callers to supported detail reads. core/artifacts/internal/operations/assistive.go Explicit draft source refresh retains both source CAS components and same-pair identity without requiring an unstarted plan. Existing progress, approvals and evidence are preserved without asserting current authority. Dependency invalidation preserves an independently BLOCKED task and its reason. Verification summaries disclose conflicting caller reports separately from receipt freshness; later successful verification covers only reports present when its context was issued. core/artifacts/internal/operations/lifecycle.go Ordinary uppercase words such as Approve DATA migration remain ordinary input. All-letter capture codes require a matching retained request; real armed replies, mismatches and user stops preserve their original handling. core/pysrc/_replylib.py Only an actual verifier wrapper executable/script prefix selects authority handling. Inert PowerShell data mentioning the filename is not a verifier call; malformed real verifier invocations still fail closed. Fresh legacy context approval arming is refused; historical pending routes remain readable. core/pysrc/_artifactauthoritylib.py Bounded isolated Python stdin diagnostics may read literal protected paths or git show HEAD objects. Unknown code, mutable calls, shell expansion and extra commands retain the guard checks; ordinary commands avoid AST parsing. core/pysrc/_bashguardlib.py After interruption, resume through the existing exact identity and durable operation outcome. Correct stale context or missing authority at its own boundary. The newer engine wait defaults to 120 seconds, but this is not a latency objective or an extension of outer host deadlines. This atlas does not benchmark runtime responsiveness.
Related entries /sprint /feature
Pinned evidence S08 · core/surface/skills/subagent-driven-development/SKILL.mdCurrent implementation-dispatch and independent spec-review sections re-read; remaining scope contract also checked against current skill index. Fresh actor input and bounded delivery are distinct from engine tickets. S10 · core/surface/includes/artifacts.mdInstalled admission, typed reads, context completion, authority receipts and format boundaries. S40 · CHANGELOG.mdMerge-scope record for review-join, marker, ID/reorder, Unicode, prerequisite and engine-wait repairs. Selected agent
Combined quality review Assistive workflow: source-bound implementation An agreed outcome becomes a maintained plan, direct or delegated implementation, proportional verification and an honest delivery report. Task context and dependency advice support progress; real user stops and host permissions remain effective. Greenfield and brownfield population use the paired native writer and exact finalization. Exact blob bindings identify inspected source bytes; the linked revision is the committed source base. Implementation, verification, live host qualification and release remain separate claims.
Current source changes A new explicit batch review can select a passing rerun of still-reviewed work without replacing the original task review. The selected run must match its own recorded workspace; changed source or supporting evidence still requires reassessment. core/artifacts/internal/evidence/completion.go Implementation works directly or delegates a coherent unit when useful. Bounded task context, applicable dependency advice and observed checks support progress while preserving actual user stops and host permissions. Verification checks the change against a recorded SMARTS rationale for material invalidation only. core/surface/skills/subagent-driven-development/SKILL.md Task-check observations now retain directly referenced criterion identity and an optional bounded source-input fingerprint or an explicit unassessed reason, without executing tests or minting approval receipts. core/artifacts/internal/operations/task_check.go Human HTML presents reported command checks in an escaped, explicitly unverified section; implementation state and source validation remain distinct. core/artifacts/internal/render/render.go The bounded source-fingerprint reader hashes actual declared task/shared paths, directory membership and known build-control files; unsafe, missing and over-limit scopes are explicitly unassessed rather than blocking work or proving a full dependency graph. core/artifacts/internal/operations/input_scope.go The optional reference retains strict historical evidence inputs without replacing ordinary work. Explicit draft source refresh preserves progressed plans and history with both CAS bindings; initial approval restrictions remain on historical paired attestation. Its verify command requires literal absolute paths; linked agent subdirectories remain unsupported for strict authority review. These limitations apply to requested attestation, not ordinary task start. core/surface/includes/artifacts-attestation.md Once per authored scope / checkpoint Actual combined diff + fresh reviewer inputs No CRITICAL/HIGH; rerun after corrections Scope of one uses the same rule
Related entries /feature /sprint
Pinned evidence S08 · core/surface/skills/subagent-driven-development/SKILL.mdCurrent implementation-dispatch and independent spec-review sections re-read; remaining scope contract also checked against current skill index. Fresh actor input and bounded delivery are distinct from engine tickets. S27 · core/pysrc/_contextselectlib.pyInert scoped selector and delivery implementation: commands/cwd/evidence, native instruction order, actor/worktree/epoch binding, two-attempt limit and reviewer preflight. S18 · core/surface/includes/review-matrix.mdShared diff-path reviewer selection for review/checkpoint/preview; callers retain their own gates. Selected state
accept-scope / current proof Assistive workflow: source-bound implementation An agreed outcome becomes a maintained plan, direct or delegated implementation, proportional verification and an honest delivery report. Task context and dependency advice support progress; real user stops and host permissions remain effective. Greenfield and brownfield population use the paired native writer and exact finalization. Exact blob bindings identify inspected source bytes; the linked revision is the committed source base. Implementation, verification, live host qualification and release remain separate claims.
Current source changes Routine artifact instructions now focus on the installed writer, typed plan, bounded task reads, factual prerequisites, actual checks and safe retry. The default leaf is about 7.3 KB and includes an executable stdin bootstrap with optional operation-schema discovery; explicit attestation remains available on demand without burdening ordinary task context. A spec's SMARTS design rationale lives in its approach, decisions and one SEC-SMARTS section. core/surface/includes/artifacts.md A new explicit batch review can select a passing rerun of still-reviewed work without replacing the original task review. The selected run must match its own recorded workspace; changed source or supporting evidence still requires reassessment. core/artifacts/internal/evidence/completion.go The sprint caller performs an index-first recorded-intent read, preserves actual owner decisions and ADR conflicts, then proceeds with authorized work without a second artifact approval system. A SMARTS Unknown names its missing observation and cannot bypass a hard gate or manufacture authority. core/surface/SPRINT.md Attended feature execution now treats checkpoints as progress groups, not compulsory approval stops; it forwards real failures and continues independent work. core/surface/skills/executing-plans/SKILL.md Implementation works directly or delegates a coherent unit when useful. Bounded task context, applicable dependency advice and observed checks support progress while preserving actual user stops and host permissions. Verification checks the change against a recorded SMARTS rationale for material invalidation only. core/surface/skills/subagent-driven-development/SKILL.md The commit owner checks authorization, the staged diff, secrets and the verification approach appropriate to the change. Current criterion checks, genuine fix regressions and refactor parity retain their distinct evidence requirements. Reuse applicable observed checks and reviews across phases. Managed v2 maintenance selects affected claims and uses native paired writes; unknown evidence defers that document. core/surface/skills/commit-gate/SKILL.md PR readiness uses the explicit base-to-HEAD subject for current checks, secrets scanning and applicable specialized review. Existing evidence is reused only while its source and inputs match. Merge requires authorization and exact-head hosted CI. core/surface/skills/finishing-a-development-branch/SKILL.md An unanswered legacy reconciliation can be superseded only when route and pending state agree and no durable attempt or matching observed reply exists. Captured replies recover through their original immutable source and receipt; in-flight mutations retain exact replay and real user stops. core/pysrc/_reconciliationlib.py The bounded source-fingerprint reader hashes actual declared task/shared paths, directory membership and known build-control files; unsafe, missing and over-limit scopes are explicitly unassessed rather than blocking work or proving a full dependency graph. core/artifacts/internal/operations/input_scope.go The candidate-only runbook distinguishes 0.3.1/0.3.2 mixed-schema readers, exact replay and forward recovery while preserving recorded user stops. The paired prompt retirement scans immutable observed replies before calling a request unanswered; submitted or uncertain decisions are never erased. Older-client downgrade after modern edits remains unsupported. docs/artifacts/ASSISTIVE-UPGRADE.md The optional reference retains strict historical evidence inputs without replacing ordinary work. Explicit draft source refresh preserves progressed plans and history with both CAS bindings; initial approval restrictions remain on historical paired attestation. Its verify command requires literal absolute paths; linked agent subdirectories remain unsupported for strict authority review. These limitations apply to requested attestation, not ordinary task start. core/surface/includes/artifacts-attestation.md Status separates reported DONE from independently ACCEPTED work, shows actual blocked decisions, and does not infer an approval request from draft governance state. core/surface/commands/status.md HTML: one whole-scope operation All required current receipts must agree Markdown: authoritative ledger transition No rendered labels as authority
Related entries /feature /sprint
Pinned evidence S08 · core/surface/skills/subagent-driven-development/SKILL.mdCurrent implementation-dispatch and independent spec-review sections re-read; remaining scope contract also checked against current skill index. Fresh actor input and bounded delivery are distinct from engine tickets. S10 · core/surface/includes/artifacts.mdInstalled admission, typed reads, context completion, authority receipts and format boundaries. Selected skill
Attended return Batch complete → executing-plans User checkpoint before next batch
Related entries /feature
Pinned evidence S08 · core/surface/skills/subagent-driven-development/SKILL.mdCurrent implementation-dispatch and independent spec-review sections re-read; remaining scope contract also checked against current skill index. Fresh actor input and bounded delivery are distinct from engine tickets. S21 · core/surface/skills/INDEX.mdCurrent routine-owner index: five debug dispositions, bounded workflow owners and unchanged six-phase TDD / nine-phase commit sequence. Selected gate
Full plan complete Current acceptance preflight Commit owner retains permission gate
Related entries /feature /sprint
Pinned evidence S08 · core/surface/skills/subagent-driven-development/SKILL.mdCurrent implementation-dispatch and independent spec-review sections re-read; remaining scope contract also checked against current skill index. Fresh actor input and bounded delivery are distinct from engine tickets. S11 · core/surface/skills/commit-gate/SKILL.mdCurrent Phase 5.5 through terminal commit rules re-read. Legacy heal vocabulary must not be generalized to v2 paired native writes. Selected out
Authorized delivery commit-gate → finishing owner Sprint: open PR only Feature: selected terminal disposition No automatic merge
Related entries /feature /sprint
Pinned evidence S11 · core/surface/skills/commit-gate/SKILL.mdCurrent Phase 5.5 through terminal commit rules re-read. Legacy heal vocabulary must not be generalized to v2 paired native writes. S12 · core/surface/skills/finishing-a-development-branch/SKILL.mdWatch/cleanup early routing; commit and acceptance preflight; PR and merge authority. Selected skill
Greenfield / decompose Assistive workflow: source-bound implementation An agreed outcome becomes a maintained plan, direct or delegated implementation, proportional verification and an honest delivery report. Task context and dependency advice support progress; real user stops and host permissions remain effective. Greenfield and brownfield population use the paired native writer and exact finalization. Exact blob bindings identify inspected source bytes; the linked revision is the committed source base. Implementation, verification, live host qualification and release remain separate claims.
Current source changes Greenfield interviews persist under plans/decomposition and retain actual decisions. Five document/v2-provenance pairs use context-draft and context-apply; context-finalize binds all thirteen outputs. ADRs and task rows use their owners, and retained interview evidence survives finalization. core/surface/skills/decompose/SKILL.md Vision → users → scope → technical shape Integrations → risks, each persisted Architecture / phased plan / backlog User-reviewed decisions and context
Related entries /init /decompose
Pinned evidence S14 · core/surface/skills/decompose/SKILL.mdCurrent greenfield interview and completion contract re-read. Its persisted six-layer draft flow remains separate from the brownfield writer lifecycle. Selected skill
Brownfield / context-creation Assistive workflow: source-bound implementation An agreed outcome becomes a maintained plan, direct or delegated implementation, proportional verification and an honest delivery report. Task context and dependency advice support progress; real user stops and host permissions remain effective. Greenfield and brownfield population use the paired native writer and exact finalization. Exact blob bindings identify inspected source bytes; the linked revision is the committed source base. Implementation, verification, live host qualification and release remain separate claims.
Current source changes The skill index connects planning, direct or delegated implementation, proportional verification and selected context maintenance to their owning workflows. core/surface/skills/INDEX.md Codex dispatch requires attestation only for explicitly claimed host proof; an unavailable requested farm backend is reported without silently switching provider or model. Ordinary context inspection can proceed inline within actual host permissions; an agent label supplies no containment or semantic verification. core/surface/includes/codex-host-notes.md Brownfield population covers six logical categories and five content documents through direct bounded inspection or optional scoped agents. The paired writer preserves source bindings and protected content; initial finalization checks all required outputs. core/surface/skills/context-creation/SKILL.md Optional context scouts report only their assigned scope, source identities, coverage, conflicts and unknowns. Six logical categories do not require six child agents. A role label or read-only instruction proves neither sandbox containment nor independent semantic verification; actual host permissions and excluded paths still apply. core/surface/agents/scout.md Qualified read-only scout profile first Six full / affected scoped assignments Bounded report-only synthesis Native content + provenance transactions
Related entries /init /create-context
Pinned evidence S13 · core/surface/skills/context-creation/SKILL.mdCurrent brownfield owner: qualified containment, bounded reports and join, scoped gaps, native document/provenance writes, separate 13-identity finalization and recovery. S26 · core/surface/skills/context-check/SKILL.mdInitialized refresh: read-only report, explicit selection, v1/v2 separation, native paired updates and preserved human history. Full owner read. Selected state
.codearbiter/ project knowledge Assistive workflow: source-bound implementation An agreed outcome becomes a maintained plan, direct or delegated implementation, proportional verification and an honest delivery report. Task context and dependency advice support progress; real user stops and host permissions remain effective. Greenfield and brownfield population use the paired native writer and exact finalization. Exact blob bindings identify inspected source bytes; the linked revision is the committed source base. Implementation, verification, live host qualification and release remain separate claims.
Current source changes The finite context contract permits a whole-document Markdown entry only as the sole entry for one of five named content targets. Explicit typed field entries remain available; the writer does not expose arbitrary repository-file writes. core/artifacts/internal/contextdocument/contract.go Whole-document replacement requires the explicit document selection and exact existing bytes. It preserves CONTEXT frontmatter and initialization state, checks the named document identity and Markdown structure, and retains the code-map size limit. Human content outside the authorized change remains the caller's preservation obligation. core/artifacts/internal/contextdocument/document_selection.go Initial finalization binds thirteen output identities, rechecks supporting source evidence, and changes the initialization marker with its paired provenance digest. Ordinary finalization requires no artifact approval receipt and does not certify semantic correctness. Already initialized repositories refuse a new finalization. core/artifacts/internal/contextdocument/finalization.go The context schema describes both finite field entries and the explicit whole-document entry for the five supported content documents. Its typed structure records evidence references; schema validity does not establish that the claims were independently reviewed. core/artifacts/schemas/context.schema.json Greenfield interviews persist under plans/decomposition and retain actual decisions. Five document/v2-provenance pairs use context-draft and context-apply; context-finalize binds all thirteen outputs. ADRs and task rows use their owners, and retained interview evidence survives finalization. core/surface/skills/decompose/SKILL.md Routine maintenance selects changed v2 claims intersecting staged sources using a current bounded snapshot. Unknown evidence defers the affected document; native paired writes preserve owned fields and current identities. core/pysrc/_provenancelib.py CONTEXT: purpose, vocabulary, stage and limits Stack: scoped test/build commands and author mapping Standards + security: rules with preserved human ownership Brownfield: five content docs + their v2 provenance pairs ADRs, boards and greenfield plans retain separate owners
Related entries /init /decompose /create-context /feature /sprint /fix /refactor
Pinned evidence S13 · core/surface/skills/context-creation/SKILL.mdCurrent brownfield owner: qualified containment, bounded reports and join, scoped gaps, native document/provenance writes, separate 13-identity finalization and recovery. S14 · core/surface/skills/decompose/SKILL.mdCurrent greenfield interview and completion contract re-read. Its persisted six-layer draft flow remains separate from the brownfield writer lifecycle. S26 · core/surface/skills/context-check/SKILL.mdInitialized refresh: read-only report, explicit selection, v1/v2 separation, native paired updates and preserved human history. Full owner read. Selected state
L1 / startup state Enabled repo: stage, questions, tasks, overrides and hygiene briefing SessionStart carries live state, not the composed persona
Related entries /feature /sprint /fix /refactor /review /commit
Pinned evidence S17 · docs/hooks.mdDocumented host event differences and startup live-state injection; not blanket enforcement proof. Selected state
L2 / composed persona safety-core + current mode body Arbiter selects existing route owners Dedup: session × mode × compaction Mode tokens are a separate control path
Related entries /feature /sprint /fix /refactor /review /commit
Pinned evidence S03 · core/surface/arbiter.mdIntent-to-owner routing, on-demand loading, caller extent and mode boundary. S15 · core/pysrc/prompt-submit.pyCurrent persona composition and session/mode/compaction injection comments read; not actor-delivery proof. Selected state
L3 / owner + scope rules Assistive workflow: source-bound implementation An agreed outcome becomes a maintained plan, direct or delegated implementation, proportional verification and an honest delivery report. Task context and dependency advice support progress; real user stops and host permissions remain effective. Greenfield and brownfield population use the paired native writer and exact finalization. Exact blob bindings identify inspected source bytes; the linked revision is the committed source base. Implementation, verification, live host qualification and release remain separate claims.
Current source changes The lazy reference map routes provenance drift or requested initialized refresh to context-check after reading selected documents, provenance and the code map. The owner reuses selected scope and existing authorization. core/surface/includes/reference-map.md Read the invoked command/skill body Scope touch → reference-map docs Read relevant ADR/deferral bodies Indexes first; no bulk directory reads
Related entries /feature /sprint /fix /refactor /review /commit
Pinned evidence S03 · core/surface/arbiter.mdIntent-to-owner routing, on-demand loading, caller extent and mode boundary. S04 · core/surface/includes/routing-table.mdAuthoritative intent-to-owner table, scope-touch routes and stop/return semantics. S05 · core/surface/includes/reference-map.mdScope-to-governing-document and specialist mapping. Selected state
L3a / read-time hints Assistive workflow: source-bound implementation An agreed outcome becomes a maintained plan, direct or delegated implementation, proportional verification and an honest delivery report. Task context and dependency advice support progress; real user stops and host permissions remain effective. Greenfield and brownfield population use the paired native writer and exact finalization. Exact blob bindings identify inspected source bytes; the linked revision is the committed source base. Implementation, verification, live host qualification and release remain separate claims.
Current source changes Bounded read-time pointers include ordinary draft and ready HTML specs as advisory context. Rejected or withdrawn specs stay excluded; no pointer confers approval. core/pysrc/_readinjectlib.py A bounded pointer, not a full document security > ADRs > specs > standards Freshness and dedup are advisory Host event surfaces differ
Related entries /feature /sprint /fix /refactor
Pinned evidence S16 · core/pysrc/_readinjectlib.pyCurrent file-read pointer and dedup implementation interface read; advisory hints remain separate from task-time selected input. S17 · docs/hooks.mdDocumented host event differences and startup live-state injection; not blanket enforcement proof. S24 · core/pysrc/_readinjectlib.pyCurrent approved_spec_index implementation still filters *.md; do not infer typed-HTML read-hint parity. Selected state
L4 / approved specification Assistive workflow: source-bound implementation An agreed outcome becomes a maintained plan, direct or delegated implementation, proportional verification and an honest delivery report. Task context and dependency advice support progress; real user stops and host permissions remain effective. Greenfield and brownfield population use the paired native writer and exact finalization. Exact blob bindings identify inspected source bytes; the linked revision is the committed source base. Implementation, verification, live host qualification and release remain separate claims.
Current source changes Routine artifact instructions now focus on the installed writer, typed plan, bounded task reads, factual prerequisites, actual checks and safe retry. The default leaf is about 7.3 KB and includes an executable stdin bootstrap with optional operation-schema discovery; explicit attestation remains available on demand without burdening ordinary task context. A spec's SMARTS design rationale lives in its approach, decisions and one SEC-SMARTS section. core/surface/includes/artifacts.md The pinned native bridge enables task-context as a bounded informational read and does not require future reviewer/approval producers simply to author a structurally valid draft. Context requests route to ordinary context-preview, context-finalize-preview and exact-reference context-apply; historical receipt input stays a separate compatibility path. Installed host qualification is separate. Native Pi writer admission does not establish a complete live Pi workflow. On NO_LOCAL_STORE the client runs init-store once and retries the request once; a refused initialization keeps NO_LOCAL_STORE. On Windows the binary's architecture comes from the kernel (IsWow64Process2), never WMI or PROCESSOR_* environment. Index and outline pages are now fully checked before rows are yielded, including continuation identity and row shape. core/pysrc/_artifactlib.py Bounded read-time pointers include ordinary draft and ready HTML specs as advisory context. Rejected or withdrawn specs stay excluded; no pointer confers approval. core/pysrc/_readinjectlib.py Stable artifact and criterion IDs Behavior • boundaries • non-goals Contracts and accepted intent Normative change invalidates approval
Related entries /feature /sprint
Pinned evidence S10 · core/surface/includes/artifacts.mdInstalled admission, typed reads, context completion, authority receipts and format boundaries. S06 · core/surface/commands/feature.mdCurrent feature owner: same-format resume and full/small gates; adds actual per-actor context composition. S07 · core/surface/SPRINT.mdInitial paired approval; full-plan execution; delegated methods; recovery and farm boundaries. Selected state
L5 / exactly bound plan Assistive workflow: source-bound implementation An agreed outcome becomes a maintained plan, direct or delegated implementation, proportional verification and an honest delivery report. Task context and dependency advice support progress; real user stops and host permissions remain effective. Greenfield and brownfield population use the paired native writer and exact finalization. Exact blob bindings identify inspected source bytes; the linked revision is the committed source base. Implementation, verification, live host qualification and release remain separate claims.
Current source changes Routine artifact instructions now focus on the installed writer, typed plan, bounded task reads, factual prerequisites, actual checks and safe retry. The default leaf is about 7.3 KB and includes an executable stdin bootstrap with optional operation-schema discovery; explicit attestation remains available on demand without burdening ordinary task context. A spec's SMARTS design rationale lives in its approach, decisions and one SEC-SMARTS section. core/surface/includes/artifacts.md The pinned native bridge enables task-context as a bounded informational read and does not require future reviewer/approval producers simply to author a structurally valid draft. Context requests route to ordinary context-preview, context-finalize-preview and exact-reference context-apply; historical receipt input stays a separate compatibility path. Installed host qualification is separate. Native Pi writer admission does not establish a complete live Pi workflow. On NO_LOCAL_STORE the client runs init-store once and retries the request once; a refused initialization keeps NO_LOCAL_STORE. On Windows the binary's architecture comes from the kernel (IsWow64Process2), never WMI or PROCESSOR_* environment. Index and outline pages are now fully checked before rows are yielded, including continuation identity and row shape. core/pysrc/_artifactlib.py Keep spec criteria, plan tasks, prerequisites and exact engine-returned IDs. Valid native IDs need not start with SPEC- or PLAN-. A typed exact-permutation record reorder keeps identities/history but is a normative edit that invalidates approval; it does not silently authorize reexecution.
Related entries /feature /sprint
Pinned evidence S10 · core/surface/includes/artifacts.mdInstalled admission, typed reads, context completion, authority receipts and format boundaries. S32 · core/pysrc/_artifactauthoritylib.pyNative artifact ID grammar, producer states, Claude join lock, UTF-8/UTF-16-key-order observation binding. S40 · CHANGELOG.mdMerge-scope record for review-join, marker, ID/reorder, Unicode, prerequisite and engine-wait repairs. Selected gate
L6 / complete task-context read Assistive workflow: source-bound implementation An agreed outcome becomes a maintained plan, direct or delegated implementation, proportional verification and an honest delivery report. Task context and dependency advice support progress; real user stops and host permissions remain effective. Greenfield and brownfield population use the paired native writer and exact finalization. Exact blob bindings identify inspected source bytes; the linked revision is the committed source base. Implementation, verification, live host qualification and release remain separate claims.
Current source changes Routine artifact instructions now focus on the installed writer, typed plan, bounded task reads, factual prerequisites, actual checks and safe retry. The default leaf is about 7.3 KB and includes an executable stdin bootstrap with optional operation-schema discovery; explicit attestation remains available on demand without burdening ordinary task context. A spec's SMARTS design rationale lives in its approach, decisions and one SEC-SMARTS section. core/surface/includes/artifacts.md Read-only task-context reports direct dependencies and unresolved ancestors, scoped_changed, scoped_unchanged or unassessed inputs, and actual check observations. Missing broader context remains explicit; the packet does not certify verification. core/artifacts/internal/operations/task_context.go Task-check observations now retain directly referenced criterion identity and an optional bounded source-input fingerprint or an explicit unassessed reason, without executing tests or minting approval receipts. core/artifacts/internal/operations/task_check.go New checks bind the selected criterion and applicable constraint bodies, including explicit references and scope applicability. Unrelated constraints remain outside the digest; old observations never gain unassessed constraint coverage. core/artifacts/internal/model/assistive.go Exact and batch symbol retrieval exposes reported check history without an approval ticket. Outline paging now accounts for encoded rows incrementally while retaining response budgets and continuation behavior. core/artifacts/internal/read/read.go Human HTML presents reported command checks in an escaped, explicitly unverified section; implementation state and source validation remain distinct. core/artifacts/internal/render/render.go The unpublished 0.3.2 native schema adds optional shared_inputs and scoped task-check fields without silently relabeling old records. core/artifacts/schemas/common.schema.json The bounded source-fingerprint reader hashes actual declared task/shared paths, directory membership and known build-control files; unsafe, missing and over-limit scopes are explicitly unassessed rather than blocking work or proving a full dependency graph. core/artifacts/internal/operations/input_scope.go Scheduling follows the full dependency ancestry and exposes unresolved ancestor IDs and reasons. Reconciled progress remains advisory; blocked ancestors prevent premature downstream readiness while independent tasks remain available. Response-budget omissions are explicit, retain the global scheduling decision and direct callers to supported detail reads. core/artifacts/internal/operations/assistive.go Explicit draft source refresh retains both source CAS components and same-pair identity without requiring an unstarted plan. Existing progress, approvals and evidence are preserved without asserting current authority. Dependency invalidation preserves an independently BLOCKED task and its reason. Verification summaries disclose conflicting caller reports separately from receipt freshness; later successful verification covers only reports present when its context was issued. core/artifacts/internal/operations/lifecycle.go Bounded isolated Python stdin diagnostics may read literal protected paths or git show HEAD objects. Unknown code, mutable calls, shell expansion and extra commands retain the guard checks; ordinary commands avoid AST parsing. core/pysrc/_bashguardlib.py identity → index → outline → scoped read Follow contextual pages to completion Retain the final engine context ticket Exact read alone is context-incomplete Stale cursor or ticket → read again
Related entries /feature /sprint
Pinned evidence S08 · core/surface/skills/subagent-driven-development/SKILL.mdCurrent implementation-dispatch and independent spec-review sections re-read; remaining scope contract also checked against current skill index. Fresh actor input and bounded delivery are distinct from engine tickets. S10 · core/surface/includes/artifacts.mdInstalled admission, typed reads, context completion, authority receipts and format boundaries. Selected agent
Lane-specific execution brief Feature: exact spec/plan/task identities Small feature: confirmed inline criteria Fix: cited defect + named regression Test-only: retain the existing caller
Related entries /feature /sprint
Pinned evidence S09 · core/surface/skills/tdd/SKILL.mdCurrent TDD context preflight and obligation derivation re-read. Six-phase gate sequence also confirmed by current skill index; fix and test-only paths do not fabricate feature artifacts. S19 · core/surface/commands/fix.mdCurrent full fix owner: validated handoff, independent caller authority, evidence revalidation, selected actor input and genuine target-red gate. S27 · core/pysrc/_contextselectlib.pyInert scoped selector and delivery implementation: commands/cwd/evidence, native instruction order, actor/worktree/epoch binding, two-attempt limit and reviewer preflight. Selected state
Not every lane needs this artifact stack Fix uses supported bug-origin evidence; refactor uses existing parity obligations. Neither fabricates a feature spec/plan to get context.
Related entries /feature /fix /refactor
Pinned evidence S09 · core/surface/skills/tdd/SKILL.mdCurrent TDD context preflight and obligation derivation re-read. Six-phase gate sequence also confirmed by current skill index; fix and test-only paths do not fabricate feature artifacts. S19 · core/surface/commands/fix.mdCurrent full fix owner: validated handoff, independent caller authority, evidence revalidation, selected actor input and genuine target-red gate. Selected state
A context ticket is neither a model memory nor a signature of semantic truth Assistive workflow: source-bound implementation An agreed outcome becomes a maintained plan, direct or delegated implementation, proportional verification and an honest delivery report. Task context and dependency advice support progress; real user stops and host permissions remain effective. Greenfield and brownfield population use the paired native writer and exact finalization. Exact blob bindings identify inspected source bytes; the linked revision is the committed source base. Implementation, verification, live host qualification and release remain separate claims.
Current source changes Routine artifact instructions now focus on the installed writer, typed plan, bounded task reads, factual prerequisites, actual checks and safe retry. The default leaf is about 7.3 KB and includes an executable stdin bootstrap with optional operation-schema discovery; explicit attestation remains available on demand without burdening ordinary task context. A spec's SMARTS design rationale lives in its approach, decisions and one SEC-SMARTS section. core/surface/includes/artifacts.md Typed engine context tickets bind complete task reads. They do not establish actual actor delivery, semantic truth or fresh runtime evidence.
Related entries /feature /sprint
Pinned evidence S10 · core/surface/includes/artifacts.mdInstalled admission, typed reads, context completion, authority receipts and format boundaries. S27 · core/pysrc/_contextselectlib.pyInert scoped selector and delivery implementation: commands/cwd/evidence, native instruction order, actor/worktree/epoch binding, two-attempt limit and reviewer preflight. Selected agent
Author context Assistive workflow: source-bound implementation An agreed outcome becomes a maintained plan, direct or delegated implementation, proportional verification and an honest delivery report. Task context and dependency advice support progress; real user stops and host permissions remain effective. Greenfield and brownfield population use the paired native writer and exact finalization. Exact blob bindings identify inspected source bytes; the linked revision is the committed source base. Implementation, verification, live host qualification and release remain separate claims.
Current source changes Authors and reviewers share task-critical reference and constraint checks. Missing or stale provenance remains advisory. Two delivery attempts end with an advisory unknown observation, without another retry or an invented receipt. core/pysrc/_contextselectlib.py Feature actors receive scoped requirements, named critical references and observed evidence. Applicable conflicts or missing required inputs stop dependent work; provenance warnings remain visible and advisory. core/pysrc/_artifactpromptlib.py Fresh task obligation + delivered scoped input Check worktree, source and constraints TDD retains the original caller’s gates
Related entries /feature /sprint /fix
Pinned evidence S08 · core/surface/skills/subagent-driven-development/SKILL.mdCurrent implementation-dispatch and independent spec-review sections re-read; remaining scope contract also checked against current skill index. Fresh actor input and bounded delivery are distinct from engine tickets. S09 · core/surface/skills/tdd/SKILL.mdCurrent TDD context preflight and obligation derivation re-read. Six-phase gate sequence also confirmed by current skill index; fix and test-only paths do not fabricate feature artifacts. S27 · core/pysrc/_contextselectlib.pyInert scoped selector and delivery implementation: commands/cwd/evidence, native instruction order, actor/worktree/epoch binding, two-attempt limit and reviewer preflight. Selected agent
Reviewer context Assistive workflow: source-bound implementation An agreed outcome becomes a maintained plan, direct or delegated implementation, proportional verification and an honest delivery report. Task context and dependency advice support progress; real user stops and host permissions remain effective. Greenfield and brownfield population use the paired native writer and exact finalization. Exact blob bindings identify inspected source bytes; the linked revision is the committed source base. Implementation, verification, live host qualification and release remain separate claims.
Current source changes Authors and reviewers share task-critical reference and constraint checks. Missing or stale provenance remains advisory. Two delivery attempts end with an advisory unknown observation, without another retry or an invented receipt. core/pysrc/_contextselectlib.py Feature actors receive scoped requirements, named critical references and observed evidence. Applicable conflicts or missing required inputs stop dependent work; provenance warnings remain visible and advisory. core/pysrc/_artifactpromptlib.py A review binds one resolved diff and supplies scoped inputs to each reviewer. Missing critical references or applicable conflicts prevent a substantive verdict; provenance and unobserved delivery diagnostics remain advisory. core/surface/commands/review.md Own fresh packet; never reuse author input Required refs + independent source checks Reviewer delivery is separately tracked
Related entries /feature /sprint /review /pr
Pinned evidence S08 · core/surface/skills/subagent-driven-development/SKILL.mdCurrent implementation-dispatch and independent spec-review sections re-read; remaining scope contract also checked against current skill index. Fresh actor input and bounded delivery are distinct from engine tickets. S28 · core/surface/commands/review.mdCurrent full review owner: stable diff resolution, selected context delivered independently to each reviewer, local verdict before separately authorized posting. S27 · core/pysrc/_contextselectlib.pyInert scoped selector and delivery implementation: commands/cwd/evidence, native instruction order, actor/worktree/epoch binding, two-attempt limit and reviewer preflight. Selected gate
Runner + authority producers Assistive workflow: source-bound implementation An agreed outcome becomes a maintained plan, direct or delegated implementation, proportional verification and an honest delivery report. Task context and dependency advice support progress; real user stops and host permissions remain effective. Greenfield and brownfield population use the paired native writer and exact finalization. Exact blob bindings identify inspected source bytes; the linked revision is the committed source base. Implementation, verification, live host qualification and release remain separate claims.
Current source changes Routine artifact instructions now focus on the installed writer, typed plan, bounded task reads, factual prerequisites, actual checks and safe retry. The default leaf is about 7.3 KB and includes an executable stdin bootstrap with optional operation-schema discovery; explicit attestation remains available on demand without burdening ordinary task context. A spec's SMARTS design rationale lives in its approach, decisions and one SEC-SMARTS section. core/surface/includes/artifacts.md A new explicit batch review can select a passing rerun of still-reviewed work without replacing the original task review. The selected run must match its own recorded workspace; changed source or supporting evidence still requires reassessment. core/artifacts/internal/evidence/completion.go Read-only task-context reports direct dependencies and unresolved ancestors, scoped_changed, scoped_unchanged or unassessed inputs, and actual check observations. Missing broader context remains explicit; the packet does not certify verification. core/artifacts/internal/operations/task_context.go An unanswered legacy reconciliation can be superseded only when route and pending state agree and no durable attempt or matching observed reply exists. Captured replies recover through their original immutable source and receipt; in-flight mutations retain exact replay and real user stops. core/pysrc/_reconciliationlib.py Task-check observations now retain directly referenced criterion identity and an optional bounded source-input fingerprint or an explicit unassessed reason, without executing tests or minting approval receipts. core/artifacts/internal/operations/task_check.go The closed task-check request accepts bounded command observations matched to plan definitions. Ordinary context-preview and context-finalize-preview omit prompt identity; context-apply accepts an exact preview reference and digest. Invalid requests, changed preimages and stale CAS writes remain rejected. The historical smarts-apply request accepts the 0.2.0 SMARTS decision record. init-store takes a protocol-only request. core/artifacts/internal/operations/protocol.go The bounded source-fingerprint reader hashes actual declared task/shared paths, directory membership and known build-control files; unsafe, missing and over-limit scopes are explicitly unassessed rather than blocking work or proving a full dependency graph. core/artifacts/internal/operations/input_scope.go The candidate-only runbook distinguishes 0.3.1/0.3.2 mixed-schema readers, exact replay and forward recovery while preserving recorded user stops. The paired prompt retirement scans immutable observed replies before calling a request unanswered; submitted or uncertain decisions are never erased. Older-client downgrade after modern edits remains unsupported. docs/artifacts/ASSISTIVE-UPGRADE.md Ordinary approval and sprint instructions bypass legacy capture without changing pending history or inventing authority. Explicit legacy replies retain exact matching. The paired adapter can retire a genuinely unanswered 0.2 sprint-pair request using its serialized cancellation lock and submission check; submitted decisions must replay, while standalone approval 0.1 refuses automatic retirement. core/pysrc/_approvallib.py The old paired approval producer publishes a content-addressed observed reply before its durable submission. Cancellation checks the matching observation, content hash, prompt and frozen context under its existing lock, with a strict scan bound. A matching or uncertain pre-submission reply blocks automatic retirement; normal tasks and context reads have no added cost. core/pysrc/_sprintapprovallib.py The optional reference retains strict historical evidence inputs without replacing ordinary work. Explicit draft source refresh preserves progressed plans and history with both CAS bindings; initial approval restrictions remain on historical paired attestation. Its verify command requires literal absolute paths; linked agent subdirectories remain unsupported for strict authority review. These limitations apply to requested attestation, not ordinary task start. core/surface/includes/artifacts-attestation.md Codex dispatch requires attestation only for explicitly claimed host proof; an unavailable requested farm backend is reported without silently switching provider or model. Ordinary context inspection can proceed inline within actual host permissions; an agent label supplies no containment or semantic verification. core/surface/includes/codex-host-notes.md Historical sprint approval and SMARTS proof instructions remain available only for existing explicit attestation, outside the ordinary execution path. core/surface/includes/sprint-authority.md Ordinary uppercase words such as Approve DATA migration remain ordinary input. All-letter capture codes require a matching retained request; real armed replies, mismatches and user stops preserve their original handling. core/pysrc/_replylib.py Only an actual verifier wrapper executable/script prefix selects authority handling. Inert PowerShell data mentioning the filename is not a verifier call; malformed real verifier invocations still fail closed. Fresh legacy context approval arming is refused; historical pending routes remain readable. core/pysrc/_artifactauthoritylib.py Bounded isolated Python stdin diagnostics may read literal protected paths or git show HEAD objects. Unknown code, mutable calls, shell expansion and extra commands retain the guard checks; ordinary commands avoid AST parsing. core/pysrc/_bashguardlib.py Receipts and current source identity are distinct from displayed status and semantic correctness. Current source uses native canonical Unicode hashing for observation binding. Host read-dedup markers are runtime state, not plan-authorized verification inputs; the marker repair must not be generalized into arbitrary .gitignore exclusions.
Related entries /feature /sprint
Pinned evidence S10 · core/surface/includes/artifacts.mdInstalled admission, typed reads, context completion, authority receipts and format boundaries. S32 · core/pysrc/_artifactauthoritylib.pyNative artifact ID grammar, producer states, Claude join lock, UTF-8/UTF-16-key-order observation binding. S40 · CHANGELOG.mdMerge-scope record for review-join, marker, ID/reorder, Unicode, prerequisite and engine-wait repairs. Selected state
L7 / acceptance evidence Assistive workflow: source-bound implementation An agreed outcome becomes a maintained plan, direct or delegated implementation, proportional verification and an honest delivery report. Task context and dependency advice support progress; real user stops and host permissions remain effective. Greenfield and brownfield population use the paired native writer and exact finalization. Exact blob bindings identify inspected source bytes; the linked revision is the committed source base. Implementation, verification, live host qualification and release remain separate claims.
Current source changes Routine artifact instructions now focus on the installed writer, typed plan, bounded task reads, factual prerequisites, actual checks and safe retry. The default leaf is about 7.3 KB and includes an executable stdin bootstrap with optional operation-schema discovery; explicit attestation remains available on demand without burdening ordinary task context. A spec's SMARTS design rationale lives in its approach, decisions and one SEC-SMARTS section. core/surface/includes/artifacts.md A new explicit batch review can select a passing rerun of still-reviewed work without replacing the original task review. The selected run must match its own recorded workspace; changed source or supporting evidence still requires reassessment. core/artifacts/internal/evidence/completion.go Status separates reported DONE from independently ACCEPTED work, shows actual blocked decisions, and does not infer an approval request from draft governance state. core/surface/commands/status.md task-review → combined quality accept-scope → current acceptance Commit and finish query proof again
Related entries /feature /sprint /commit /pr
Pinned evidence S08 · core/surface/skills/subagent-driven-development/SKILL.mdCurrent implementation-dispatch and independent spec-review sections re-read; remaining scope contract also checked against current skill index. Fresh actor input and bounded delivery are distinct from engine tickets. S10 · core/surface/includes/artifacts.mdInstalled admission, typed reads, context completion, authority receipts and format boundaries. S11 · core/surface/skills/commit-gate/SKILL.mdCurrent Phase 5.5 through terminal commit rules re-read. Legacy heal vocabulary must not be generalized to v2 paired native writes. S12 · core/surface/skills/finishing-a-development-branch/SKILL.mdWatch/cleanup early routing; commit and acceptance preflight; PR and merge authority. Selected skill
Legacy v1 heal / native v2 refresh Assistive workflow: source-bound implementation An agreed outcome becomes a maintained plan, direct or delegated implementation, proportional verification and an honest delivery report. Task context and dependency advice support progress; real user stops and host permissions remain effective. Greenfield and brownfield population use the paired native writer and exact finalization. Exact blob bindings identify inspected source bytes; the linked revision is the committed source base. Implementation, verification, live host qualification and release remain separate claims.
Current source changes Ordinary create/adopt/update preserves the same source-evidence, selection and compare-and-swap checks as the native transaction. Missing or matching v1 provenance can be adopted; existing v2 ownership can transition through explicit whole-document selection. Original document/provenance preimages remain in transaction history for exact recovery. core/artifacts/internal/contextdocument/operations.go A drift audit stays read-only; an already authorized selected or full refresh proceeds without asking again per document. The installed draft/preview/apply example preserves source bindings and selected content. Matching hashes establish identity only; initialized refresh never reruns scaffolding, clears the marker or repeats finalization. core/surface/skills/context-check/SKILL.md The lazy reference map routes provenance drift or requested initialized refresh to context-check after reading selected documents, provenance and the code map. The owner reuses selected scope and existing authorization. core/surface/includes/reference-map.md v1: conditional staged-source auto-heal v2: selected owned-field preview + receipt Identity match is not semantic verification No raw v2 rebaseline fallback
Related entries /commit /feature /sprint
Pinned evidence S11 · core/surface/skills/commit-gate/SKILL.mdCurrent Phase 5.5 through terminal commit rules re-read. Legacy heal vocabulary must not be generalized to v2 paired native writes. S26 · core/surface/skills/context-check/SKILL.mdInitialized refresh: read-only report, explicit selection, v1/v2 separation, native paired updates and preserved human history. Full owner read. Selected skill
Follow-up harvest / terminal boundary [NEEDS-TRIAGE] → task writer for work Unresolved decisions → open-questions.md Interactive: confirm; sprint: permitted SMARTS logging A task-board item is not a typed plan acceptance.
Related entries /task /feature /sprint /commit
Pinned evidence S04 · core/surface/includes/routing-table.mdAuthoritative intent-to-owner table, scope-touch routes and stop/return semantics. S11 · core/surface/skills/commit-gate/SKILL.mdCurrent Phase 5.5 through terminal commit rules re-read. Legacy heal vocabulary must not be generalized to v2 paired native writes. Selected skill
Retained state / bounded recovery Assistive workflow: source-bound implementation An agreed outcome becomes a maintained plan, direct or delegated implementation, proportional verification and an honest delivery report. Task context and dependency advice support progress; real user stops and host permissions remain effective. Greenfield and brownfield population use the paired native writer and exact finalization. Exact blob bindings identify inspected source bytes; the linked revision is the committed source base. Implementation, verification, live host qualification and release remain separate claims.
Current source changes Ordinary create/adopt/update preserves the same source-evidence, selection and compare-and-swap checks as the native transaction. Missing or matching v1 provenance can be adopted; existing v2 ownership can transition through explicit whole-document selection. Original document/provenance preimages remain in transaction history for exact recovery. core/artifacts/internal/contextdocument/operations.go A drift audit stays read-only; an already authorized selected or full refresh proceeds without asking again per document. The installed draft/preview/apply example preserves source bindings and selected content. Matching hashes establish identity only; initialized refresh never reruns scaffolding, clears the marker or repeats finalization. core/surface/skills/context-check/SKILL.md Scheduling follows the full dependency ancestry and exposes unresolved ancestor IDs and reasons. Reconciled progress remains advisory; blocked ancestors prevent premature downstream readiness while independent tasks remain available. Response-budget omissions are explicit, retain the global scheduling decision and direct callers to supported detail reads. core/artifacts/internal/operations/assistive.go Explicit draft source refresh retains both source CAS components and same-pair identity without requiring an unstarted plan. Existing progress, approvals and evidence are preserved without asserting current authority. Dependency invalidation preserves an independently BLOCKED task and its reason. Verification summaries disclose conflicting caller reports separately from receipt freshness; later successful verification covers only reports present when its context was issued. core/artifacts/internal/operations/lifecycle.go Refresh keeps initialization and human history Recover native operations by the same ID Committed results are read back, not rewritten Decisions remain with their actual owner
Related entries /adr /reconcile /status /context-check /feature /sprint
Pinned evidence S13 · core/surface/skills/context-creation/SKILL.mdCurrent brownfield owner: qualified containment, bounded reports and join, scoped gaps, native document/provenance writes, separate 13-identity finalization and recovery. S26 · core/surface/skills/context-check/SKILL.mdInitialized refresh: read-only report, explicit selection, v1/v2 separation, native paired updates and preserved human history. Full owner read. Selected state
Select task-relevant context Assistive workflow: source-bound implementation An agreed outcome becomes a maintained plan, direct or delegated implementation, proportional verification and an honest delivery report. Task context and dependency advice support progress; real user stops and host permissions remain effective. Greenfield and brownfield population use the paired native writer and exact finalization. Exact blob bindings identify inspected source bytes; the linked revision is the committed source base. Implementation, verification, live host qualification and release remain separate claims.
Current source changes Authors and reviewers share task-critical reference and constraint checks. Missing or stale provenance remains advisory. Two delivery attempts end with an advisory unknown observation, without another retry or an invented receipt. core/pysrc/_contextselectlib.py select_task_context is inert. The caller acquires current inputs and native instruction precedence. Copied upstream rules, source-scoped ADRs and declared commands do not gain authority or current verification from packet inclusion.
Related entries /feature /sprint /fix /refactor /review /init /create-context /status /context-check
Pinned evidence S27 · core/pysrc/_contextselectlib.pyInert scoped selector and delivery implementation: commands/cwd/evidence, native instruction order, actor/worktree/epoch binding, two-attempt limit and reviewer preflight. Selected skill
Compose for this lane and actor Assistive workflow: source-bound implementation An agreed outcome becomes a maintained plan, direct or delegated implementation, proportional verification and an honest delivery report. Task context and dependency advice support progress; real user stops and host permissions remain effective. Greenfield and brownfield population use the paired native writer and exact finalization. Exact blob bindings identify inspected source bytes; the linked revision is the committed source base. Implementation, verification, live host qualification and release remain separate claims.
Current source changes Authors and reviewers share task-critical reference and constraint checks. Missing or stale provenance remains advisory. Two delivery attempts end with an advisory unknown observation, without another retry or an invented receipt. core/pysrc/_contextselectlib.py Feature actors receive scoped requirements, named critical references and observed evidence. Applicable conflicts or missing required inputs stop dependent work; provenance warnings remain visible and advisory. core/pysrc/_artifactpromptlib.py Confirmed fixes preserve causal evidence and a named regression. Delegated input retains critical requirements; advisory delivery exhaustion ends retries while actual missing inputs and conflicts retain their dependent-work boundary. Distinct remediation strategies are compared with SMARTS only after the root cause is confirmed. core/surface/commands/fix.md compose_feature_actor_input is named by the current feature/TDD/SDD contracts. compose_fix_or_test_actor_input preserves bug origin and the existing route. prepare_review_input blocks missing critical references before a substantive verdict.
Related entries /feature /sprint /fix /refactor /review /init /create-context /status /context-check
Pinned evidence S27 · core/pysrc/_contextselectlib.pyInert scoped selector and delivery implementation: commands/cwd/evidence, native instruction order, actor/worktree/epoch binding, two-attempt limit and reviewer preflight. S09 · core/surface/skills/tdd/SKILL.mdCurrent TDD context preflight and obligation derivation re-read. Six-phase gate sequence also confirmed by current skill index; fix and test-only paths do not fabricate feature artifacts. S08 · core/surface/skills/subagent-driven-development/SKILL.mdCurrent implementation-dispatch and independent spec-review sections re-read; remaining scope contract also checked against current skill index. Fresh actor input and bounded delivery are distinct from engine tickets. S28 · core/surface/commands/review.mdCurrent full review owner: stable diff resolution, selected context delivered independently to each reviewer, local verdict before separately authorized posting. Selected gate
Bound actual actor delivery Assistive workflow: source-bound implementation An agreed outcome becomes a maintained plan, direct or delegated implementation, proportional verification and an honest delivery report. Task context and dependency advice support progress; real user stops and host permissions remain effective. Greenfield and brownfield population use the paired native writer and exact finalization. Exact blob bindings identify inspected source bytes; the linked revision is the committed source base. Implementation, verification, live host qualification and release remain separate claims.
Current source changes Authors and reviewers share task-critical reference and constraint checks. Missing or stale provenance remains advisory. Two delivery attempts end with an advisory unknown observation, without another retry or an invented receipt. core/pysrc/_contextselectlib.py Confirmed fixes preserve causal evidence and a named regression. Delegated input retains critical requirements; advisory delivery exhaustion ends retries while actual missing inputs and conflicts retain their dependent-work boundary. Distinct remediation strategies are compared with SMARTS only after the root cause is confirmed. core/surface/commands/fix.md A review binds one resolved diff and supplies scoped inputs to each reviewer. Missing critical references or applicable conflicts prevent a substantive verdict; provenance and unobserved delivery diagnostics remain advisory. core/surface/commands/review.md prepare_actor_delivery binds the selected packet digest, actor/task, actual worktree and observable epochs. The caller sends delivery_text on deliver and retains the counted attempt. A reported host-load-event or child-input-capture is not a native task-admission certificate.
Related entries /feature /sprint /fix /refactor /review /init /create-context /status /context-check
Pinned evidence S27 · core/pysrc/_contextselectlib.pyInert scoped selector and delivery implementation: commands/cwd/evidence, native instruction order, actor/worktree/epoch binding, two-attempt limit and reviewer preflight. S08 · core/surface/skills/subagent-driven-development/SKILL.mdCurrent implementation-dispatch and independent spec-review sections re-read; remaining scope contract also checked against current skill index. Fresh actor input and bounded delivery are distinct from engine tickets. Selected warn
Reselect or stop Assistive workflow: source-bound implementation An agreed outcome becomes a maintained plan, direct or delegated implementation, proportional verification and an honest delivery report. Task context and dependency advice support progress; real user stops and host permissions remain effective. Greenfield and brownfield population use the paired native writer and exact finalization. Exact blob bindings identify inspected source bytes; the linked revision is the committed source base. Implementation, verification, live host qualification and release remain separate claims.
Current source changes Authors and reviewers share task-critical reference and constraint checks. Missing or stale provenance remains advisory. Two delivery attempts end with an advisory unknown observation, without another retry or an invented receipt. core/pysrc/_contextselectlib.py Two counted attempts apply to the same actor/task/worktree/epoch/packet binding. Unknown epochs cannot make a saved receipt current. No shared session marker or author receipt substitutes for a reviewer’s input delivery.
Related entries /feature /sprint /fix /refactor /review /init /create-context /status /context-check
Pinned evidence S27 · core/pysrc/_contextselectlib.pyInert scoped selector and delivery implementation: commands/cwd/evidence, native instruction order, actor/worktree/epoch binding, two-attempt limit and reviewer preflight. S08 · core/surface/skills/subagent-driven-development/SKILL.mdCurrent implementation-dispatch and independent spec-review sections re-read; remaining scope contract also checked against current skill index. Fresh actor input and bounded delivery are distinct from engine tickets. Selected cmd
/debug Diagnosis-only or one internal prerequisite from an authorized fix No code modification
Related entries /debug
Pinned evidence S25 · core/surface/skills/debug/SKILL.mdFive-disposition diagnostic owner; transient validated handoff, freshness and authority boundaries. Full owner read. S19 · core/surface/commands/fix.mdCurrent full fix owner: validated handoff, independent caller authority, evidence revalidation, selected actor input and genuine target-red gate. Selected skill
debug / evidence-led investigation Cited symptom → real discriminators Current source/runtime applicability Exactly one supported disposition
Related entries /debug
Pinned evidence S25 · core/surface/skills/debug/SKILL.mdFive-disposition diagnostic owner; transient validated handoff, freshness and authority boundaries. Full owner read. Selected cmd
/spike Bounded unknown + timebox
Related entries /spike
Pinned evidence S02 · core/surface/COMMANDS.mdCatalog, modes, host-limited commands and purpose of each public route. S04 · core/surface/includes/routing-table.mdAuthoritative intent-to-owner table, scope-touch routes and stop/return semantics. Selected skill
Disposable spike/<slug> Explore the question in isolation Exemption cannot land production work Never merges or opens a PR
Related entries /spike
Pinned evidence S04 · core/surface/includes/routing-table.mdAuthoritative intent-to-owner table, scope-touch routes and stop/return semantics. Selected out
Findings note What was learned; uncertainty left No requirement to build anything
Related entries /spike
Pinned evidence S04 · core/surface/includes/routing-table.mdAuthoritative intent-to-owner table, scope-touch routes and stop/return semantics. Selected cmd
/feature / new production path A separate approved specification Experiment is evidence, not approval
Related entries /spike /feature
Pinned evidence S04 · core/surface/includes/routing-table.mdAuthoritative intent-to-owner table, scope-touch routes and stop/return semantics. S06 · core/surface/commands/feature.mdCurrent feature owner: same-format resume and full/small gates; adds actual per-actor context composition. Selected cmd
/fix One bounded internal diagnostic return is permitted if required, without public fix/debug recursion. Before the author acts, select and deliver current context; independently revalidate a retained handoff and caller authority.
Related entries /fix
Pinned evidence S19 · core/surface/commands/fix.mdCurrent full fix owner: validated handoff, independent caller authority, evidence revalidation, selected actor input and genuine target-red gate. S27 · core/pysrc/_contextselectlib.pyInert scoped selector and delivery implementation: commands/cwd/evidence, native instruction order, actor/worktree/epoch binding, two-attempt limit and reviewer preflight. S25 · core/surface/skills/debug/SKILL.mdFive-disposition diagnostic owner; transient validated handoff, freshness and authority boundaries. Full owner read. Selected skill
Causal evidence → target RED Cited cause + named regression Observed target RED for the right reason No feature spec or unrelated red
Related entries /fix
Pinned evidence S19 · core/surface/commands/fix.mdCurrent full fix owner: validated handoff, independent caller authority, evidence revalidation, selected actor input and genuine target-red gate. S09 · core/surface/skills/tdd/SKILL.mdCurrent TDD context preflight and obligation derivation re-read. Six-phase gate sequence also confirmed by current skill index; fix and test-only paths do not fabricate feature artifacts. S25 · core/surface/skills/debug/SKILL.mdFive-disposition diagnostic owner; transient validated handoff, freshness and authority boundaries. Full owner read. Selected skill
Complete tdd Obligations → red → green Verify obligations → coverage → lint Fresh local proof, not self-report
Related entries /fix
Pinned evidence S09 · core/surface/skills/tdd/SKILL.mdCurrent TDD context preflight and obligation derivation re-read. Six-phase gate sequence also confirmed by current skill index; fix and test-only paths do not fabricate feature artifacts. S21 · core/surface/skills/INDEX.mdCurrent routine-owner index: five debug dispositions, bounded workflow owners and unchanged six-phase TDD / nine-phase commit sequence. Selected out
Verified correction Scoped code + regression + evidence Commit only when authorized
Related entries /fix /commit
Pinned evidence S19 · core/surface/commands/fix.mdCurrent full fix owner: validated handoff, independent caller authority, evidence revalidation, selected actor input and genuine target-red gate. S09 · core/surface/skills/tdd/SKILL.mdCurrent TDD context preflight and obligation derivation re-read. Six-phase gate sequence also confirmed by current skill index; fix and test-only paths do not fabricate feature artifacts. S11 · core/surface/skills/commit-gate/SKILL.mdCurrent Phase 5.5 through terminal commit rules re-read. Legacy heal vocabulary must not be generalized to v2 paired native writes. Selected cmd
/refactor Named restructure Behavior must stay the same
Related entries /refactor
Pinned evidence S21 · core/surface/skills/INDEX.mdCurrent routine-owner index: five debug dispositions, bounded workflow owners and unchanged six-phase TDD / nine-phase commit sequence. Selected gate
Parity-coverage proof Unchanged pre-existing tests Coverage backfill / new seams → TDD No parity proof, no restructure
Related entries /refactor
Pinned evidence S04 · core/surface/includes/routing-table.mdAuthoritative intent-to-owner table, scope-touch routes and stop/return semantics. S21 · core/surface/skills/INDEX.mdCurrent routine-owner index: five debug dispositions, bounded workflow owners and unchanged six-phase TDD / nine-phase commit sequence. Selected skill
refactor / bounded change Restructure behind the proven contract Re-run unchanged behavioral tests Feature/fix owns behavioral changes
Related entries /refactor
Pinned evidence S04 · core/surface/includes/routing-table.mdAuthoritative intent-to-owner table, scope-touch routes and stop/return semantics. S21 · core/surface/skills/INDEX.mdCurrent routine-owner index: five debug dispositions, bounded workflow owners and unchanged six-phase TDD / nine-phase commit sequence. Selected out
Verified parity Restructured code + parity evidence Commit only through commit-gate
Related entries /refactor /commit
Pinned evidence S04 · core/surface/includes/routing-table.mdAuthoritative intent-to-owner table, scope-touch routes and stop/return semantics. S21 · core/surface/skills/INDEX.mdCurrent routine-owner index: five debug dispositions, bounded workflow owners and unchanged six-phase TDD / nine-phase commit sequence. S11 · core/surface/skills/commit-gate/SKILL.mdCurrent Phase 5.5 through terminal commit rules re-read. Legacy heal vocabulary must not be generalized to v2 paired native writes. Selected cmd
/chore Classify docs / deps / revert Reject behavior smuggled into the lane
Related entries /chore
Pinned evidence S20 · core/surface/commands/chore.mdDocs/deps/revert branches; behavioral adoption returns to feature/fix. Selected skill
docs Secrets + diff + user-facing copy checks No TDD required for prose
Related entries /chore
Pinned evidence S20 · core/surface/commands/chore.mdDocs/deps/revert branches; behavioral adoption returns to feature/fix. Selected skill
deps / existing version bump dependency-reviewer; manifest + lockfile Affected local tests; code adoption redirects
Related entries /chore
Pinned evidence S20 · core/surface/commands/chore.mdDocs/deps/revert branches; behavioral adoption returns to feature/fix. Selected skill
revert / named commit git revert <SHA>; affected tests Reason and SHA in the commit message Missed defect? Follow with /fix
Related entries /chore
Pinned evidence S20 · core/surface/commands/chore.mdDocs/deps/revert branches; behavioral adoption returns to feature/fix. Selected out
Reviewed non-behavioral change Matching type proof commit-gate → finishing owner Exact-head hosted CI before merge Explicit persistence authority still applies
Related entries /chore /commit /pr
Pinned evidence S20 · core/surface/commands/chore.mdDocs/deps/revert branches; behavioral adoption returns to feature/fix. S11 · core/surface/skills/commit-gate/SKILL.mdCurrent Phase 5.5 through terminal commit rules re-read. Legacy heal vocabulary must not be generalized to v2 paired native writes. S12 · core/surface/skills/finishing-a-development-branch/SKILL.mdWatch/cleanup early routing; commit and acceptance preflight; PR and merge authority. Selected cmd
/add-dep Package purpose + identity Read stack and security policy
Related entries /add-dep
Pinned evidence S04 · core/surface/includes/routing-table.mdAuthoritative intent-to-owner table, scope-touch routes and stop/return semantics. S05 · core/surface/includes/reference-map.mdScope-to-governing-document and specialist mapping. Selected agent
Durable adoption / dependency-reviewer License • provenance • maintenance • supply chain Vetting outcome returns to the requesting workflow
Related entries /add-dep
Pinned evidence S04 · core/surface/includes/routing-table.mdAuthoritative intent-to-owner table, scope-touch routes and stop/return semantics. S22 · core/surface/agents/INDEX.mdRole separation; specialist reviewers; restricted scouts/graders; authority-reviewer host restriction. Selected skill
One-off branch: pinned tool + registry + confirmation
Related entries /add-dep
Pinned evidence S04 · core/surface/includes/routing-table.mdAuthoritative intent-to-owner table, scope-touch routes and stop/return semantics. Selected out
Vetted decision / tool result Ephemeral: verify no manifest/lock change Adoption: retain reviewed policy outcome
Related entries /add-dep
Pinned evidence S04 · core/surface/includes/routing-table.mdAuthoritative intent-to-owner table, scope-touch routes and stop/return semantics. Selected skill
Return to caller Feature/fix owns behavior adoption No automatic PR from a tool run
Related entries /add-dep /feature /fix
Pinned evidence S03 · core/surface/arbiter.mdIntent-to-owner routing, on-demand loading, caller extent and mode boundary. S04 · core/surface/includes/routing-table.mdAuthoritative intent-to-owner table, scope-touch routes and stop/return semantics. Selected out
Five outcomes; no automatic repair or ADR The dedicated Debug handoff map expands the five dispositions. Only confirmed_code_defect plus next_step.target fix, validated packet, current evidence and actual caller repair authority can continue to TDD. A design question does not automatically create an ADR; no-action and unresolved do not automatically create tasks.
Related entries /debug /fix
Pinned evidence S25 · core/surface/skills/debug/SKILL.mdFive-disposition diagnostic owner; transient validated handoff, freshness and authority boundaries. Full owner read. S19 · core/surface/commands/fix.mdCurrent full fix owner: validated handoff, independent caller authority, evidence revalidation, selected actor input and genuine target-red gate. Selected cmd
/review · /checkpoint Assistive workflow: source-bound implementation An agreed outcome becomes a maintained plan, direct or delegated implementation, proportional verification and an honest delivery report. Task context and dependency advice support progress; real user stops and host permissions remain effective. Greenfield and brownfield population use the paired native writer and exact finalization. Exact blob bindings identify inspected source bytes; the linked revision is the committed source base. Implementation, verification, live host qualification and release remain separate claims.
Current source changes A review binds one resolved diff and supplies scoped inputs to each reviewer. Missing critical references or applicable conflicts prevent a substantive verdict; provenance and unobserved delivery diagnostics remain advisory. core/surface/commands/review.md A checkpoint retains the reviewed commit, UTC review-start time and worktree dirty state, then writes the report through its aggregator. last-checkpoint holds only the override count. core/surface/commands/checkpoint.md Checkpoint reports persist observed source_commit, recorded_at and worktree_dirty values together with every review finding. Missing anchors remain unknown. core/surface/agents/checkpoint-aggregator.md /review: current diff /checkpoint: periodic codebase sweep /preview predicts only; no verdict /tribunal and /threat-model are opt-in
Related entries /review /checkpoint /preview /tribunal /threat-model
Pinned evidence S04 · core/surface/includes/routing-table.mdAuthoritative intent-to-owner table, scope-touch routes and stop/return semantics. S18 · core/surface/includes/review-matrix.mdShared diff-path reviewer selection for review/checkpoint/preview; callers retain their own gates. S21 · core/surface/skills/INDEX.mdCurrent routine-owner index: five debug dispositions, bounded workflow owners and unchanged six-phase TDD / nine-phase commit sequence. Selected agent
Fresh inputs → scoped reviewer fleet Assistive workflow: source-bound implementation An agreed outcome becomes a maintained plan, direct or delegated implementation, proportional verification and an honest delivery report. Task context and dependency advice support progress; real user stops and host permissions remain effective. Greenfield and brownfield population use the paired native writer and exact finalization. Exact blob bindings identify inspected source bytes; the linked revision is the committed source base. Implementation, verification, live host qualification and release remain separate claims.
Current source changes Authors and reviewers share task-critical reference and constraint checks. Missing or stale provenance remains advisory. Two delivery attempts end with an advisory unknown observation, without another retry or an invented receipt. core/pysrc/_contextselectlib.py A review binds one resolved diff and supplies scoped inputs to each reviewer. Missing critical references or applicable conflicts prevent a substantive verdict; provenance and unobserved delivery diagnostics remain advisory. core/surface/commands/review.md Select current context for each reviewer Required refs + native instruction order Deliver to that actual child, not its parent Then dispatch by the actual diff paths Preserve capability and evidence gaps
Related entries /review /checkpoint /pr
Pinned evidence S28 · core/surface/commands/review.mdCurrent full review owner: stable diff resolution, selected context delivered independently to each reviewer, local verdict before separately authorized posting. S27 · core/pysrc/_contextselectlib.pyInert scoped selector and delivery implementation: commands/cwd/evidence, native instruction order, actor/worktree/epoch binding, two-attempt limit and reviewer preflight. S18 · core/surface/includes/review-matrix.mdShared diff-path reviewer selection for review/checkpoint/preview; callers retain their own gates. Selected agent
finding-triage Collect and deduplicate findings Retain terminal states / gaps Classify severity and scope Then pass the complete report onward
Related entries /review /checkpoint /pr
Pinned evidence S04 · core/surface/includes/routing-table.mdAuthoritative intent-to-owner table, scope-touch routes and stop/return semantics. S22 · core/surface/agents/INDEX.mdRole separation; specialist reviewers; restricted scouts/graders; authority-reviewer host restriction. Selected agent
verdict-aggregator Read-only in-memory result Own change: CRITICAL/HIGH blocks PR Inbound PR: local verdict; post separately Checkpoint: explicit dated persistence
Related entries /review /checkpoint /pr
Pinned evidence S28 · core/surface/commands/review.mdCurrent full review owner: stable diff resolution, selected context delivered independently to each reviewer, local verdict before separately authorized posting. S04 · core/surface/includes/routing-table.mdAuthoritative intent-to-owner table, scope-touch routes and stop/return semantics. Selected cmd
/commit Assistive workflow: source-bound implementation An agreed outcome becomes a maintained plan, direct or delegated implementation, proportional verification and an honest delivery report. Task context and dependency advice support progress; real user stops and host permissions remain effective. Greenfield and brownfield population use the paired native writer and exact finalization. Exact blob bindings identify inspected source bytes; the linked revision is the committed source base. Implementation, verification, live host qualification and release remain separate claims.
Current source changes Routine artifact instructions now focus on the installed writer, typed plan, bounded task reads, factual prerequisites, actual checks and safe retry. The default leaf is about 7.3 KB and includes an executable stdin bootstrap with optional operation-schema discovery; explicit attestation remains available on demand without burdening ordinary task context. A spec's SMARTS design rationale lives in its approach, decisions and one SEC-SMARTS section. core/surface/includes/artifacts.md Current staged scope TDD complete where applicable Typed plans: current acceptance No status-label substitutes
Related entries /commit /feature /sprint /fix /refactor /chore
Pinned evidence S11 · core/surface/skills/commit-gate/SKILL.mdCurrent Phase 5.5 through terminal commit rules re-read. Legacy heal vocabulary must not be generalized to v2 paired native writes. S10 · core/surface/includes/artifacts.mdInstalled admission, typed reads, context completion, authority receipts and format boundaries. Selected gate
Phases 1–3 / authorize + scope 1 Explicit permission 2 Non-protected branch 3 Homogeneous change classification No implied authorization from approval of a specification or an implementation
Related entries /commit /feature /sprint /fix /refactor /chore
Pinned evidence S11 · core/surface/skills/commit-gate/SKILL.mdCurrent Phase 5.5 through terminal commit rules re-read. Legacy heal vocabulary must not be generalized to v2 paired native writes. Selected gate
Phases 4–6 / prove + inspect Assistive workflow: source-bound implementation An agreed outcome becomes a maintained plan, direct or delegated implementation, proportional verification and an honest delivery report. Task context and dependency advice support progress; real user stops and host permissions remain effective. Greenfield and brownfield population use the paired native writer and exact finalization. Exact blob bindings identify inspected source bytes; the linked revision is the committed source base. Implementation, verification, live host qualification and release remain separate claims.
Current source changes A drift audit stays read-only; an already authorized selected or full refresh proceeds without asking again per document. The installed draft/preview/apply example preserves source bindings and selected content. Matching hashes establish identity only; initialized refresh never reruns scaffolding, clears the marker or repeats finalization. core/surface/skills/context-check/SKILL.md 4 Focused tests, lint, secrets, scope gates 5 Fresh behavioral proof 5.5 Legacy v1 staged-source auto-heal 6 Complete diff review v2 pairs require native selected updates
Related entries /commit /feature /sprint /fix /refactor /chore
Pinned evidence S11 · core/surface/skills/commit-gate/SKILL.mdCurrent Phase 5.5 through terminal commit rules re-read. Legacy heal vocabulary must not be generalized to v2 paired native writes. S26 · core/surface/skills/context-check/SKILL.mdInitialized refresh: read-only report, explicit selection, v1/v2 separation, native paired updates and preserved human history. Full owner read. Selected gate
Phases 7–9 / persist 7 Harvest → explicitly selected paths 8 Conventional Commit message 9 Commit → SHA + clean-tree report No blanket staging; failed hook returns to correction, not gate circumvention
Related entries /commit /feature /sprint /fix /refactor /chore
Pinned evidence S11 · core/surface/skills/commit-gate/SKILL.mdCurrent Phase 5.5 through terminal commit rules re-read. Legacy heal vocabulary must not be generalized to v2 paired native writes. Selected skill
Conditional scope gates / they do not add new public routes crypto-compliance + secret-handling → auth-crypto-reviewer; migrations → migration-reviewer; CI/deploy → security-reviewer; dependencies → dependency-reviewer.
Related entries /commit /feature /sprint /fix /refactor /chore /add-dep
Pinned evidence S04 · core/surface/includes/routing-table.mdAuthoritative intent-to-owner table, scope-touch routes and stop/return semantics. S05 · core/surface/includes/reference-map.mdScope-to-governing-document and specialist mapping. S11 · core/surface/skills/commit-gate/SKILL.mdCurrent Phase 5.5 through terminal commit rules re-read. Legacy heal vocabulary must not be generalized to v2 paired native writes. Selected cmd
/pr Direct open request or feature/sprint terminal handoff Choose mode from user authority
Related entries /pr /feature /sprint
Pinned evidence S12 · core/surface/skills/finishing-a-development-branch/SKILL.mdWatch/cleanup early routing; commit and acceptance preflight; PR and merge authority. Selected skill
--watch / retained /watch Detached hosted-CI watcher Red → diagnosis; green → offer, never auto-merge
Related entries /pr /watch
Pinned evidence S04 · core/surface/includes/routing-table.mdAuthoritative intent-to-owner table, scope-touch routes and stop/return semantics. S12 · core/surface/skills/finishing-a-development-branch/SKILL.mdWatch/cleanup early routing; commit and acceptance preflight; PR and merge authority. Selected skill
--cleanup / retained /cleanup post-merge-cleanup Prove fetched containment, classify residue Confirm each discard; preserve unique work
Related entries /pr /cleanup
Pinned evidence S04 · core/surface/includes/routing-table.mdAuthoritative intent-to-owner table, scope-touch routes and stop/return semantics. S12 · core/surface/skills/finishing-a-development-branch/SKILL.mdWatch/cleanup early routing; commit and acceptance preflight; PR and merge authority. S21 · core/surface/skills/INDEX.mdCurrent routine-owner index: five debug dispositions, bounded workflow owners and unchanged six-phase TDD / nine-phase commit sequence. Selected gate
Creation / current-head preflight Current acceptance + commit proof Branch / base / head / diff Applicable ADR ancestry Missing or stale proof → owner repair
Related entries /pr /feature /sprint
Pinned evidence S12 · core/surface/skills/finishing-a-development-branch/SKILL.mdWatch/cleanup early routing; commit and acceptance preflight; PR and merge authority. Selected skill
Open-PR procedure in this owner Applicable reviewers; clear blockers Compose summary, test plan and ADR links Push branch, create PR, return URL Never re-invoke the PR wrapper
Related entries /pr /feature /sprint
Pinned evidence S12 · core/surface/skills/finishing-a-development-branch/SKILL.mdWatch/cleanup early routing; commit and acceptance preflight; PR and merge authority. Selected gate
Merge via PR / only if selected Explicit user authority + exact-head CI Revalidate ancestry and bound head
Related entries /pr
Pinned evidence S12 · core/surface/skills/finishing-a-development-branch/SKILL.mdWatch/cleanup early routing; commit and acceptance preflight; PR and merge authority. Selected out
Open PR / authorized disposition Sprint ends at open PR Feature keeps selected terminal option Discard requires stated-loss confirmation
Related entries /pr /feature /sprint /cleanup
Pinned evidence S12 · core/surface/skills/finishing-a-development-branch/SKILL.mdWatch/cleanup early routing; commit and acceptance preflight; PR and merge authority. Selected cmd
/release · selected target Assistive workflow: source-bound implementation An agreed outcome becomes a maintained plan, direct or delegated implementation, proportional verification and an honest delivery report. Task context and dependency advice support progress; real user stops and host permissions remain effective. Greenfield and brownfield population use the paired native writer and exact finalization. Exact blob bindings identify inspected source bytes; the linked revision is the committed source base. Implementation, verification, live host qualification and release remain separate claims.
Current source changes Release preparation inserts a new version below the entire Unreleased section and above prior versions, preserving pending notes. This guidance correction does not grant publication authority or replace the existing release-PR and hosted qualification boundaries. core/surface/skills/release/SKILL.md Resolve declared target + policy Prepare reviewed release surfaces Dry-run stops at its report
Related entries /release
Pinned evidence S33 · core/surface/skills/release/SKILL.mdInteractive release preparation and mandatory release-PR boundary; qualified hosted tag composition. Selected gate
Release preparation PR Assistive workflow: source-bound implementation An agreed outcome becomes a maintained plan, direct or delegated implementation, proportional verification and an honest delivery report. Task context and dependency advice support progress; real user stops and host permissions remain effective. Greenfield and brownfield population use the paired native writer and exact finalization. Exact blob bindings identify inspected source bytes; the linked revision is the committed source base. Implementation, verification, live host qualification and release remain separate claims.
Current source changes Release preparation inserts a new version below the entire Unreleased section and above prior versions, preserving pending notes. This guidance correction does not grant publication authority or replace the existing release-PR and hosted qualification boundaries. core/surface/skills/release/SKILL.md Commit through normal gates Name assets + declared publisher Open release PR and STOP
Related entries /release /commit
Pinned evidence S33 · core/surface/skills/release/SKILL.mdInteractive release preparation and mandatory release-PR boundary; qualified hosted tag composition. S34 · site/src/content/docs/concepts/workflow-routes.mdxExisting public route-comparison owner; distinct release/initialization endpoints and source-only explanation boundary. Selected gate
Qualified hosted continuation Only after the release PR merges Exact fetched head + green evidence Compose the bound tag and cohort
Related entries /release
Pinned evidence S33 · core/surface/skills/release/SKILL.mdInteractive release preparation and mandatory release-PR boundary; qualified hosted tag composition. Selected out
Publication and closeout Assistive workflow: source-bound implementation An agreed outcome becomes a maintained plan, direct or delegated implementation, proportional verification and an honest delivery report. Task context and dependency advice support progress; real user stops and host permissions remain effective. Greenfield and brownfield population use the paired native writer and exact finalization. Exact blob bindings identify inspected source bytes; the linked revision is the committed source base. Implementation, verification, live host qualification and release remain separate claims.
Current source changes Release preparation inserts a new version below the entire Unreleased section and above prior versions, preserving pending notes. This guidance correction does not grant publication authority or replace the existing release-PR and hosted qualification boundaries. core/surface/skills/release/SKILL.md Consume applicable cohort authority Read back release + inventory Declared receipt closes by separate PR
Related entries /release
Pinned evidence S33 · core/surface/skills/release/SKILL.mdInteractive release preparation and mandatory release-PR boundary; qualified hosted tag composition. S34 · site/src/content/docs/concepts/workflow-routes.mdxExisting public route-comparison owner; distinct release/initialization endpoints and source-only explanation boundary. Selected gate
Selected legacy resolution Actual merge + schema-1 provenance Exact current HEAD chosen Native or uncertain ownership refuses
Related entries /pr
Pinned evidence S12 · core/surface/skills/finishing-a-development-branch/SKILL.mdWatch/cleanup early routing; commit and acceptance preflight; PR and merge authority. Selected skill
Inspect → exact binding Shipped helper inspects Git + files Pass the returned binding directly Changed identities refuse
Related entries /pr
Pinned evidence S12 · core/surface/skills/finishing-a-development-branch/SKILL.mdWatch/cleanup early routing; commit and acceptance preflight; PR and merge authority. Selected skill
Restore HEAD only One supported conflict rendering Exact committed code-map bytes Provenance and index untouched
Related entries /pr
Pinned evidence S12 · core/surface/skills/finishing-a-development-branch/SKILL.mdWatch/cleanup early routing; commit and acceptance preflight; PR and merge authority. Selected out
Review, stage, re-prove Inspect the resulting diff Use normal explicit staging Review + CI bind the integrated head
Related entries /pr
Pinned evidence S12 · core/surface/skills/finishing-a-development-branch/SKILL.mdWatch/cleanup early routing; commit and acceptance preflight; PR and merge authority. Selected cmd
/adr · /adr-status Author or read-only status Select mode before prerequisites
Related entries /adr /adr-status
Pinned evidence S21 · core/surface/skills/INDEX.mdCurrent routine-owner index: five debug dispositions, bounded workflow owners and unchanged six-phase TDD / nine-phase commit sequence. Selected skill
decision-lifecycle Author: actual user choice + attribution Status: age, challenges, supersession CONFIRM questions are never guessed
Related entries /adr /adr-status
Pinned evidence S04 · core/surface/includes/routing-table.mdAuthoritative intent-to-owner table, scope-touch routes and stop/return semantics. S21 · core/surface/skills/INDEX.mdCurrent routine-owner index: five debug dispositions, bounded workflow owners and unchanged six-phase TDD / nine-phase commit sequence. Selected out
ADR / health report Numbered, dated decision record Or read-only lifecycle report Accepted/Planned ≠ implementation
Related entries /adr /adr-status
Pinned evidence S21 · core/surface/skills/INDEX.mdCurrent routine-owner index: five debug dispositions, bounded workflow owners and unchanged six-phase TDD / nine-phase commit sequence. Selected state
Future scope-relevant context Recorded intent constrains spec authoring Affected ADRs return through indexes No automatic product change
Related entries /adr /adr-status /feature /sprint
Pinned evidence S05 · core/surface/includes/reference-map.mdScope-to-governing-document and specialist mapping. S07 · core/surface/SPRINT.mdInitial paired approval; full-plan execution; delegated methods; recovery and farm boundaries. S21 · core/surface/skills/INDEX.mdCurrent routine-owner index: five debug dispositions, bounded workflow owners and unchanged six-phase TDD / nine-phase commit sequence. Selected cmd
/reconcile Drift / architectural variance Report-only or explicit reconciliation
Related entries /reconcile
Pinned evidence S04 · core/surface/includes/routing-table.mdAuthoritative intent-to-owner table, scope-touch routes and stop/return semantics. S21 · core/surface/skills/INDEX.mdCurrent routine-owner index: five debug dispositions, bounded workflow owners and unchanged six-phase TDD / nine-phase commit sequence. Selected skill
decision-variance Scoped scout evidence → SMARTS grader Optional decision-challenger Grader recommends; user chooses
Related entries /reconcile
Pinned evidence S04 · core/surface/includes/routing-table.mdAuthoritative intent-to-owner table, scope-touch routes and stop/return semantics. S21 · core/surface/skills/INDEX.mdCurrent routine-owner index: five debug dispositions, bounded workflow owners and unchanged six-phase TDD / nine-phase commit sequence. S22 · core/surface/agents/INDEX.mdRole separation; specialist reviewers; restricted scouts/graders; authority-reviewer host restriction. Selected out
Report / attributed ruling Read-only mode writes nothing Chosen reconciliation enters decision-log Separate ADR owner when required
Related entries /reconcile
Pinned evidence S04 · core/surface/includes/routing-table.mdAuthoritative intent-to-owner table, scope-touch routes and stop/return semantics. S21 · core/surface/skills/INDEX.mdCurrent routine-owner index: five debug dispositions, bounded workflow owners and unchanged six-phase TDD / nine-phase commit sequence. Selected gate
/conflict / internal STOP Contradictory governing rules Stop other work; surface both sides Resolve authority before resuming
Related entries /conflict
Pinned evidence S03 · core/surface/arbiter.mdIntent-to-owner routing, on-demand loading, caller extent and mode boundary. S04 · core/surface/includes/routing-table.mdAuthoritative intent-to-owner table, scope-touch routes and stop/return semantics. Selected cmd
/status Default: read-only state drift: provenance audit Alias: /context-check
Related entries /status /context-check
Pinned evidence S02 · core/surface/COMMANDS.mdCatalog, modes, host-limited commands and purpose of each public route. S04 · core/surface/includes/routing-table.mdAuthoritative intent-to-owner table, scope-touch routes and stop/return semantics. Selected skill
context-check / selected refresh Assistive workflow: source-bound implementation An agreed outcome becomes a maintained plan, direct or delegated implementation, proportional verification and an honest delivery report. Task context and dependency advice support progress; real user stops and host permissions remain effective. Greenfield and brownfield population use the paired native writer and exact finalization. Exact blob bindings identify inspected source bytes; the linked revision is the committed source base. Implementation, verification, live host qualification and release remain separate claims.
Current source changes The user-facing route table points feature and sprint work at useful planned execution and selected context refresh at its existing owner. Existing authorization covers ordinary work; actual host, destructive-action and commit controls remain. core/surface/includes/routing-table.md The skill index connects planning, direct or delegated implementation, proportional verification and selected context maintenance to their owning workflows. core/surface/skills/INDEX.md The preparation helper binds bounded drafts and retained sources to exact document/provenance preimages. Finalization preparation hashes all thirteen required outputs and retains missing output failures without modifying files. core/pysrc/_contextwriterlib.py A drift audit stays read-only; an already authorized selected or full refresh proceeds without asking again per document. The installed draft/preview/apply example preserves source bindings and selected content. Matching hashes establish identity only; initialized refresh never reruns scaffolding, clears the marker or repeats finalization. core/surface/skills/context-check/SKILL.md Report v1 and v2 states separately Select affected docs or explicit full refresh Qualified scouts + native v2 pair writes
Related entries /status /context-check
Pinned evidence S26 · core/surface/skills/context-check/SKILL.mdInitialized refresh: read-only report, explicit selection, v1/v2 separation, native paired updates and preserved human history. Full owner read. S13 · core/surface/skills/context-creation/SKILL.mdCurrent brownfield owner: qualified containment, bounded reports and join, scoped gaps, native document/provenance writes, separate 13-identity finalization and recovery. Selected out
Status / drift report Assistive workflow: source-bound implementation An agreed outcome becomes a maintained plan, direct or delegated implementation, proportional verification and an honest delivery report. Task context and dependency advice support progress; real user stops and host permissions remain effective. Greenfield and brownfield population use the paired native writer and exact finalization. Exact blob bindings identify inspected source bytes; the linked revision is the committed source base. Implementation, verification, live host qualification and release remain separate claims.
Current source changes A drift audit stays read-only; an already authorized selected or full refresh proceeds without asking again per document. The installed draft/preview/apply example preserves source bindings and selected content. Matching hashes establish identity only; initialized refresh never reruns scaffolding, clears the marker or repeats finalization. core/surface/skills/context-check/SKILL.md Read-only report, or selected native update Keep marker, unselected files and history No staging, commit or reinitialization
Related entries /status /context-check
Pinned evidence S26 · core/surface/skills/context-check/SKILL.mdInitialized refresh: read-only report, explicit selection, v1/v2 separation, native paired updates and preserved human history. Full owner read. S13 · core/surface/skills/context-creation/SKILL.mdCurrent brownfield owner: qualified containment, bounded reports and join, scoped gaps, native document/provenance writes, separate 13-identity finalization and recovery. Selected cmd
/doctor Install, payload, package ownership Enforcement config + harmless probe Output: diagnostic / capability gap
Related entries /doctor
Pinned evidence S02 · core/surface/COMMANDS.mdCatalog, modes, host-limited commands and purpose of each public route. S10 · core/surface/includes/artifacts.mdInstalled admission, typed reads, context completion, authority receipts and format boundaries. Selected cmd
/task Authorized add / start / done Terminal harvest or concrete follow-up Debug outcome alone writes nothing
Related entries /task
Pinned evidence S25 · core/surface/skills/debug/SKILL.mdFive-disposition diagnostic owner; transient validated handoff, freshness and authority boundaries. Full owner read. S04 · core/surface/includes/routing-table.mdAuthoritative intent-to-owner table, scope-touch routes and stop/return semantics. Selected skill
work.py → selected owner One facade selects native or legacy work Queued work can complete directly Checks remain separately evidenced
Related entries /task
Pinned evidence S04 · core/surface/includes/routing-table.mdAuthoritative intent-to-owner table, scope-touch routes and stop/return semantics. Selected state
Selected work transition Native HTML or open-tasks.md state Not a typed-plan acceptance receipt Later work still selects its own lane
Related entries /task
Pinned evidence S04 · core/surface/includes/routing-table.mdAuthoritative intent-to-owner table, scope-touch routes and stop/return semantics. S08 · core/surface/skills/subagent-driven-development/SKILL.mdCurrent implementation-dispatch and independent spec-review sections re-read; remaining scope contract also checked against current skill index. Fresh actor input and bounded delivery are distinct from engine tickets. Selected cmd
/standup Inspect hygiene, branches and worktrees Confirm each safe cleanup Never silently lose dirty / unpushed work
Related entries /standup
Pinned evidence S02 · core/surface/COMMANDS.mdCatalog, modes, host-limited commands and purpose of each public route. S04 · core/surface/includes/routing-table.mdAuthoritative intent-to-owner table, scope-touch routes and stop/return semantics. Selected cmd
/audit → dated governance packet Assistive workflow: source-bound implementation An agreed outcome becomes a maintained plan, direct or delegated implementation, proportional verification and an honest delivery report. Task context and dependency advice support progress; real user stops and host permissions remain effective. Greenfield and brownfield population use the paired native writer and exact finalization. Exact blob bindings identify inspected source bytes; the linked revision is the committed source base. Implementation, verification, live host qualification and release remain separate claims.
Current source changes A checkpoint retains the reviewed commit, UTC review-start time and worktree dirty state, then writes the report through its aggregator. last-checkpoint holds only the override count. core/surface/commands/checkpoint.md Checkpoint reports persist observed source_commit, recorded_at and worktree_dirty values together with every review finding. Missing anchors remain unknown. core/surface/agents/checkpoint-aggregator.md Checkpoint audit windows use the report timestamp and an ancestor commit. Invalid or missing anchors require an explicit ref/date window; dirty review scope remains a stated limitation. core/surface/commands/audit.md Read source records for a selected window. Preserve audit lines. Write a new packet only; never overwrite the source or an older packet.
Related entries /audit
Pinned evidence S02 · core/surface/COMMANDS.mdCatalog, modes, host-limited commands and purpose of each public route. S04 · core/surface/includes/routing-table.mdAuthoritative intent-to-owner table, scope-touch routes and stop/return semantics. Selected cmd
/metrics → trends against prior window Overrides, small-lane classifications and low-confidence sprint choices. Read-only measurement, not an automatic policy change.
Related entries /metrics
Pinned evidence S02 · core/surface/COMMANDS.mdCatalog, modes, host-limited commands and purpose of each public route. Selected cmd
/override → attributed exception Only where the governing hard rule permits; explicit reason and consent. Append to overrides.log. Not a universal security or authority bypass.
Related entries /override
Pinned evidence S03 · core/surface/arbiter.mdIntent-to-owner routing, on-demand loading, caller extent and mode boundary. S04 · core/surface/includes/routing-table.mdAuthoritative intent-to-owner table, scope-touch routes and stop/return semantics. Selected state
mode --arbiter / --ops / --dangerous Deterministic control tokens, not slash entries. Safety-core remains composed with the active mode body. Ops is advisory runtime work; tracked/history changes route normally.
Related entries
Pinned evidence S03 · core/surface/arbiter.mdIntent-to-owner routing, on-demand loading, caller extent and mode boundary. S04 · core/surface/includes/routing-table.mdAuthoritative intent-to-owner table, scope-touch routes and stop/return semantics. S15 · core/pysrc/prompt-submit.pyCurrent persona composition and session/mode/compaction injection comments read; not actor-delivery proof. Selected cmd
/prune → host-specific context control Claude: dry-run by default, never execute against the live transcript. Pi: native compaction, not active-session rewriting. No Codex transcript route is asserted.
Related entries /prune
Pinned evidence S02 · core/surface/COMMANDS.mdCatalog, modes, host-limited commands and purpose of each public route. S04 · core/surface/includes/routing-table.mdAuthoritative intent-to-owner table, scope-touch routes and stop/return semantics. Selected cmd
/statusline → Claude configuration Install / uninstall / inspect the owned statusline. This catalog surface is Claude-specific. It does not establish cross-host runtime parity.
Related entries /statusline
Pinned evidence S02 · core/surface/COMMANDS.mdCatalog, modes, host-limited commands and purpose of each public route. Selected cmd
/commands · /btw Grouped catalog / deprecated lightweight Q&A. Direct questions need no slash command.
Related entries /commands /btw
Pinned evidence S01 · core/surface/command-routes.jsonAll 37 catalog entries; core, advanced, alias, internal and deprecated visibility. S02 · core/surface/COMMANDS.mdCatalog, modes, host-limited commands and purpose of each public route. S03 · core/surface/arbiter.mdIntent-to-owner routing, on-demand loading, caller extent and mode boundary. Selected cmd
/decompose · /create-context Assistive workflow: source-bound implementation An agreed outcome becomes a maintained plan, direct or delegated implementation, proportional verification and an honest delivery report. Task context and dependency advice support progress; real user stops and host permissions remain effective. Greenfield and brownfield population use the paired native writer and exact finalization. Exact blob bindings identify inspected source bytes; the linked revision is the committed source base. Implementation, verification, live host qualification and release remain separate claims.
Current source changes The installed context-draft command prepares a paired document preview; context-finalize prepares initial finalization. context-apply commits the exact selected preview and digest. Context modes preserve scaffold and initialized refresh boundaries. core/pysrc/init-codearbiter.py Initial brownfield setup checks the intended repository and actual host permissions, then uses bounded inspection and the installed safe context writer. Ordinary Windows work needs no special scout profile or per-document artifact approval. An existing initialized repository routes to selected refresh instead of another scaffold or population cycle. core/surface/commands/init.md Greenfield interviews persist under plans/decomposition and retain actual decisions. Five document/v2-provenance pairs use context-draft and context-apply; context-finalize binds all thirteen outputs. ADRs and task rows use their owners, and retained interview evidence survives finalization. core/surface/skills/decompose/SKILL.md Initial population only: /init --greenfield / --brownfield. Initialized refresh stays with /status drift and explicit selection.
Related entries /init /decompose /create-context
Pinned evidence S23 · core/surface/commands/init.mdCurrent init owner: activation/scaffold versus initial population; initialized refresh belongs to status drift, not reinitialization. S26 · core/surface/skills/context-check/SKILL.mdInitialized refresh: read-only report, explicit selection, v1/v2 separation, native paired updates and preserved human history. Full owner read. Selected cmd
/watch · /cleanup · /context-check Prefer /pr --watch / --cleanup and /status drift. These modes branch before unrelated creation-time preflights.
Related entries /pr /watch /cleanup /status /context-check
Pinned evidence S01 · core/surface/command-routes.jsonAll 37 catalog entries; core, advanced, alias, internal and deprecated visibility. S02 · core/surface/COMMANDS.mdCatalog, modes, host-limited commands and purpose of each public route. S12 · core/surface/skills/finishing-a-development-branch/SKILL.mdWatch/cleanup early routing; commit and acceptance preflight; PR and merge authority. Selected cmd
Initial population Assistive workflow: source-bound implementation An agreed outcome becomes a maintained plan, direct or delegated implementation, proportional verification and an honest delivery report. Task context and dependency advice support progress; real user stops and host permissions remain effective. Greenfield and brownfield population use the paired native writer and exact finalization. Exact blob bindings identify inspected source bytes; the linked revision is the committed source base. Implementation, verification, live host qualification and release remain separate claims.
Current source changes The installed context-draft command prepares a paired document preview; context-finalize prepares initial finalization. context-apply commits the exact selected preview and digest. Context modes preserve scaffold and initialized refresh boundaries. core/pysrc/init-codearbiter.py Brownfield population covers six logical categories and five content documents through direct bounded inspection or optional scoped agents. The paired writer preserves source bindings and protected content; initial finalization checks all required outputs. core/surface/skills/context-creation/SKILL.md Initial brownfield setup checks the intended repository and actual host permissions, then uses bounded inspection and the installed safe context writer. Ordinary Windows work needs no special scout profile or per-document artifact approval. An existing initialized repository routes to selected refresh instead of another scaffold or population cycle. core/surface/commands/init.md /init --brownfield → context-creation Meaningful source; initialized marker absent A stub is not a completed project context
Related entries /init /create-context
Pinned evidence S13 · core/surface/skills/context-creation/SKILL.mdCurrent brownfield owner: qualified containment, bounded reports and join, scoped gaps, native document/provenance writes, separate 13-identity finalization and recovery. S23 · core/surface/commands/init.mdCurrent init owner: activation/scaffold versus initial population; initialized refresh belongs to status drift, not reinitialization. Selected cmd
Initialized refresh Assistive workflow: source-bound implementation An agreed outcome becomes a maintained plan, direct or delegated implementation, proportional verification and an honest delivery report. Task context and dependency advice support progress; real user stops and host permissions remain effective. Greenfield and brownfield population use the paired native writer and exact finalization. Exact blob bindings identify inspected source bytes; the linked revision is the committed source base. Implementation, verification, live host qualification and release remain separate claims.
Current source changes The preparation helper binds bounded drafts and retained sources to exact document/provenance preimages. Finalization preparation hashes all thirteen required outputs and retains missing output failures without modifying files. core/pysrc/_contextwriterlib.py The installed context-draft command prepares a paired document preview; context-finalize prepares initial finalization. context-apply commits the exact selected preview and digest. Context modes preserve scaffold and initialized refresh boundaries. core/pysrc/init-codearbiter.py A drift audit stays read-only; an already authorized selected or full refresh proceeds without asking again per document. The installed draft/preview/apply example preserves source bindings and selected content. Matching hashes establish identity only; initialized refresh never reruns scaffolding, clears the marker or repeats finalization. core/surface/skills/context-check/SKILL.md The lazy reference map routes provenance drift or requested initialized refresh to context-check after reading selected documents, provenance and the code map. The owner reuses selected scope and existing authorization. core/surface/includes/reference-map.md Routine maintenance selects changed v2 claims intersecting staged sources using a current bounded snapshot. Unknown evidence defers the affected document; native paired writes preserve owned fields and current identities. core/pysrc/_provenancelib.py /status drift → context-check Read-only report, then explicit scoped/full choice Do not clear the marker or rerun initialization
Related entries /status /context-check
Pinned evidence S26 · core/surface/skills/context-check/SKILL.mdInitialized refresh: read-only report, explicit selection, v1/v2 separation, native paired updates and preserved human history. Full owner read. Selected cmd
Interrupted operation Assistive workflow: source-bound implementation An agreed outcome becomes a maintained plan, direct or delegated implementation, proportional verification and an honest delivery report. Task context and dependency advice support progress; real user stops and host permissions remain effective. Greenfield and brownfield population use the paired native writer and exact finalization. Exact blob bindings identify inspected source bytes; the linked revision is the committed source base. Implementation, verification, live host qualification and release remain separate claims.
Current source changes Ordinary context-preview retains an inert proposal and exact preimages without approval or review claims. context-apply revalidates the digest, closed request and preview binding before the native paired transaction; context-finalize-preview uses the same exact-reference path for initial population. No approve-context reply is required. core/artifacts/internal/operations/context_assistive.go Brownfield population covers six logical categories and five content documents through direct bounded inspection or optional scoped agents. The paired writer preserves source bindings and protected content; initial finalization checks all required outputs. core/surface/skills/context-creation/SKILL.md Inventory actual files, provenance and outcome Inspect the same operation ID and native journal Do not blindly repeat a possibly committed write
Related entries /init /create-context /status /context-check
Pinned evidence S13 · core/surface/skills/context-creation/SKILL.mdCurrent brownfield owner: qualified containment, bounded reports and join, scoped gaps, native document/provenance writes, separate 13-identity finalization and recovery. Selected gate
Qualified containment first Assistive workflow: source-bound implementation An agreed outcome becomes a maintained plan, direct or delegated implementation, proportional verification and an honest delivery report. Task context and dependency advice support progress; real user stops and host permissions remain effective. Greenfield and brownfield population use the paired native writer and exact finalization. Exact blob bindings identify inspected source bytes; the linked revision is the committed source base. Implementation, verification, live host qualification and release remain separate claims.
Current source changes Codex dispatch requires attestation only for explicitly claimed host proof; an unavailable requested farm backend is reported without silently switching provider or model. Ordinary context inspection can proceed inline within actual host permissions; an agent label supplies no containment or semantic verification. core/surface/includes/codex-host-notes.md Brownfield population covers six logical categories and five content documents through direct bounded inspection or optional scoped agents. The paired writer preserves source bindings and protected content; initial finalization checks all required outputs. core/surface/skills/context-creation/SKILL.md Initial brownfield setup checks the intended repository and actual host permissions, then uses bounded inspection and the installed safe context writer. Ordinary Windows work needs no special scout profile or per-document artifact approval. An existing initialized repository routes to selected refresh instead of another scaffold or population cycle. core/surface/commands/init.md Optional context scouts report only their assigned scope, source identities, coverage, conflicts and unknowns. Six logical categories do not require six child agents. A role label or read-only instruction proves neither sandbox containment nor independent semantic verification; actual host permissions and excluded paths still apply. core/surface/agents/scout.md The optional context-profile preparation helper is a repository-only CI fixture, not installed plugin capability. Caller-owned OS/tool containment, child binding, launch budgets and actual observations must be qualified before dispatch. Unknown host behavior keeps this step unavailable.
Related entries /init /create-context /status /context-check
Pinned evidence S13 · core/surface/skills/context-creation/SKILL.mdCurrent brownfield owner: qualified containment, bounded reports and join, scoped gaps, native document/provenance writes, separate 13-identity finalization and recovery. Selected agent
Bounded logical assignments Assistive workflow: source-bound implementation An agreed outcome becomes a maintained plan, direct or delegated implementation, proportional verification and an honest delivery report. Task context and dependency advice support progress; real user stops and host permissions remain effective. Greenfield and brownfield population use the paired native writer and exact finalization. Exact blob bindings identify inspected source bytes; the linked revision is the committed source base. Implementation, verification, live host qualification and release remain separate claims.
Current source changes Codex dispatch requires attestation only for explicitly claimed host proof; an unavailable requested farm backend is reported without silently switching provider or model. Ordinary context inspection can proceed inline within actual host permissions; an agent label supplies no containment or semantic verification. core/surface/includes/codex-host-notes.md Brownfield population covers six logical categories and five content documents through direct bounded inspection or optional scoped agents. The paired writer preserves source bindings and protected content; initial finalization checks all required outputs. core/surface/skills/context-creation/SKILL.md Optional context scouts report only their assigned scope, source identities, coverage, conflicts and unknowns. Six logical categories do not require six child agents. A role label or read-only instruction proves neither sandbox containment nor independent semantic verification; actual host permissions and excluded paths still apply. core/surface/agents/scout.md A–F cover stack, infrastructure, architecture, security, tests and data. Scope/subdivision/retries retain one run ledger and frozen source boundary. No missing or failed transport can be counted as complete coverage.
Related entries /init /create-context /status /context-check
Pinned evidence S13 · core/surface/skills/context-creation/SKILL.mdCurrent brownfield owner: qualified containment, bounded reports and join, scoped gaps, native document/provenance writes, separate 13-identity finalization and recovery. S26 · core/surface/skills/context-check/SKILL.mdInitialized refresh: read-only report, explicit selection, v1/v2 separation, native paired updates and preserved human history. Full owner read. Selected skill
Validate and join reports Assistive workflow: source-bound implementation An agreed outcome becomes a maintained plan, direct or delegated implementation, proportional verification and an honest delivery report. Task context and dependency advice support progress; real user stops and host permissions remain effective. Greenfield and brownfield population use the paired native writer and exact finalization. Exact blob bindings identify inspected source bytes; the linked revision is the committed source base. Implementation, verification, live host qualification and release remain separate claims.
Current source changes Brownfield population covers six logical categories and five content documents through direct bounded inspection or optional scoped agents. The paired writer preserves source bindings and protected content; initial finalization checks all required outputs. core/surface/skills/context-creation/SKILL.md Optional context scouts report only their assigned scope, source identities, coverage, conflicts and unknowns. Six logical categories do not require six child agents. A role label or read-only instruction proves neither sandbox containment nor independent semantic verification; actual host permissions and excluded paths still apply. core/surface/agents/scout.md Closed assignment/snapshot binding 48 KiB normalized join limit Missing / failed transport → block Recheck the source snapshot
Related entries /init /create-context /status /context-check
Pinned evidence S13 · core/surface/skills/context-creation/SKILL.mdCurrent brownfield owner: qualified containment, bounded reports and join, scoped gaps, native document/provenance writes, separate 13-identity finalization and recovery. Selected skill
Synthesize; settle relevant gaps Assistive workflow: source-bound implementation An agreed outcome becomes a maintained plan, direct or delegated implementation, proportional verification and an honest delivery report. Task context and dependency advice support progress; real user stops and host permissions remain effective. Greenfield and brownfield population use the paired native writer and exact finalization. Exact blob bindings identify inspected source bytes; the linked revision is the committed source base. Implementation, verification, live host qualification and release remain separate claims.
Current source changes Brownfield population covers six logical categories and five content documents through direct bounded inspection or optional scoped agents. The paired writer preserves source bindings and protected content; initial finalization checks all required outputs. core/surface/skills/context-creation/SKILL.md Use joined reports, not raw source Preserve conflicting citations Ask material user-owned choices Deferral is not approval
Related entries /init /create-context /status /context-check
Pinned evidence S13 · core/surface/skills/context-creation/SKILL.mdCurrent brownfield owner: qualified containment, bounded reports and join, scoped gaps, native document/provenance writes, separate 13-identity finalization and recovery. Selected state
Five content targets Assistive workflow: source-bound implementation An agreed outcome becomes a maintained plan, direct or delegated implementation, proportional verification and an honest delivery report. Task context and dependency advice support progress; real user stops and host permissions remain effective. Greenfield and brownfield population use the paired native writer and exact finalization. Exact blob bindings identify inspected source bytes; the linked revision is the committed source base. Implementation, verification, live host qualification and release remain separate claims.
Current source changes The finite context contract permits a whole-document Markdown entry only as the sole entry for one of five named content targets. Explicit typed field entries remain available; the writer does not expose arbitrary repository-file writes. core/artifacts/internal/contextdocument/contract.go Whole-document replacement requires the explicit document selection and exact existing bytes. It preserves CONTEXT frontmatter and initialization state, checks the named document identity and Markdown structure, and retains the code-map size limit. Human content outside the authorized change remains the caller's preservation obligation. core/artifacts/internal/contextdocument/document_selection.go The context schema describes both finite field entries and the explicit whole-document entry for the five supported content documents. Its typed structure records evidence references; schema validity does not establish that the claims were independently reviewed. core/artifacts/schemas/context.schema.json Brownfield population covers six logical categories and five content documents through direct bounded inspection or optional scoped agents. The paired writer preserves source bindings and protected content; initial finalization checks all required outputs. core/surface/skills/context-creation/SKILL.md CONTEXT.md • tech-stack.md coding-standards.md security-controls.md • code-map.md Each with .provenance/<doc>.json
Related entries /init /create-context /status /context-check
Pinned evidence S13 · core/surface/skills/context-creation/SKILL.mdCurrent brownfield owner: qualified containment, bounded reports and join, scoped gaps, native document/provenance writes, separate 13-identity finalization and recovery. S26 · core/surface/skills/context-check/SKILL.mdInitialized refresh: read-only report, explicit selection, v1/v2 separation, native paired updates and preserved human history. Full owner read. Selected skill
Finite typed preview Assistive workflow: source-bound implementation An agreed outcome becomes a maintained plan, direct or delegated implementation, proportional verification and an honest delivery report. Task context and dependency advice support progress; real user stops and host permissions remain effective. Greenfield and brownfield population use the paired native writer and exact finalization. Exact blob bindings identify inspected source bytes; the linked revision is the committed source base. Implementation, verification, live host qualification and release remain separate claims.
Current source changes The pinned native bridge enables task-context as a bounded informational read and does not require future reviewer/approval producers simply to author a structurally valid draft. Context requests route to ordinary context-preview, context-finalize-preview and exact-reference context-apply; historical receipt input stays a separate compatibility path. Installed host qualification is separate. Native Pi writer admission does not establish a complete live Pi workflow. On NO_LOCAL_STORE the client runs init-store once and retries the request once; a refused initialization keeps NO_LOCAL_STORE. On Windows the binary's architecture comes from the kernel (IsWow64Process2), never WMI or PROCESSOR_* environment. Index and outline pages are now fully checked before rows are yielded, including continuation identity and row shape. core/pysrc/_artifactlib.py The closed task-check request accepts bounded command observations matched to plan definitions. Ordinary context-preview and context-finalize-preview omit prompt identity; context-apply accepts an exact preview reference and digest. Invalid requests, changed preimages and stale CAS writes remain rejected. The historical smarts-apply request accepts the 0.2.0 SMARTS decision record. init-store takes a protocol-only request. core/artifacts/internal/operations/protocol.go The preparation helper binds bounded drafts and retained sources to exact document/provenance preimages. Finalization preparation hashes all thirteen required outputs and retains missing output failures without modifying files. core/pysrc/_contextwriterlib.py The installed context-draft command prepares a paired document preview; context-finalize prepares initial finalization. context-apply commits the exact selected preview and digest. Context modes preserve scaffold and initialized refresh boundaries. core/pysrc/init-codearbiter.py Ordinary context-preview retains an inert proposal and exact preimages without approval or review claims. context-apply revalidates the digest, closed request and preview binding before the native paired transaction; context-finalize-preview uses the same exact-reference path for initial population. No approve-context reply is required. core/artifacts/internal/operations/context_assistive.go The finite context contract permits a whole-document Markdown entry only as the sole entry for one of five named content targets. Explicit typed field entries remain available; the writer does not expose arbitrary repository-file writes. core/artifacts/internal/contextdocument/contract.go Whole-document replacement requires the explicit document selection and exact existing bytes. It preserves CONTEXT frontmatter and initialization state, checks the named document identity and Markdown structure, and retains the code-map size limit. Human content outside the authorized change remains the caller's preservation obligation. core/artifacts/internal/contextdocument/document_selection.go The context schema describes both finite field entries and the explicit whole-document entry for the five supported content documents. Its typed structure records evidence references; schema validity does not establish that the claims were independently reviewed. core/artifacts/schemas/context.schema.json A drift audit stays read-only; an already authorized selected or full refresh proceeds without asking again per document. The installed draft/preview/apply example preserves source bindings and selected content. Matching hashes establish identity only; initialized refresh never reruns scaffolding, clears the marker or repeats finalization. core/surface/skills/context-check/SKILL.md The lazy reference map routes provenance drift or requested initialized refresh to context-check after reading selected documents, provenance and the code map. The owner reuses selected scope and existing authorization. core/surface/includes/reference-map.md New document or selected owned field Current source / target identities Exact proposed content + provenance Unsupported human layout → refuse
Related entries /init /create-context /status /context-check
Pinned evidence S13 · core/surface/skills/context-creation/SKILL.mdCurrent brownfield owner: qualified containment, bounded reports and join, scoped gaps, native document/provenance writes, separate 13-identity finalization and recovery. S26 · core/surface/skills/context-check/SKILL.mdInitialized refresh: read-only report, explicit selection, v1/v2 separation, native paired updates and preserved human history. Full owner read. Selected gate
Native approve-context reply Assistive workflow: source-bound implementation An agreed outcome becomes a maintained plan, direct or delegated implementation, proportional verification and an honest delivery report. Task context and dependency advice support progress; real user stops and host permissions remain effective. Greenfield and brownfield population use the paired native writer and exact finalization. Exact blob bindings identify inspected source bytes; the linked revision is the committed source base. Implementation, verification, live host qualification and release remain separate claims.
Current source changes The closed task-check request accepts bounded command observations matched to plan definitions. Ordinary context-preview and context-finalize-preview omit prompt identity; context-apply accepts an exact preview reference and digest. Invalid requests, changed preimages and stale CAS writes remain rejected. The historical smarts-apply request accepts the 0.2.0 SMARTS decision record. init-store takes a protocol-only request. core/artifacts/internal/operations/protocol.go The optional reference retains strict historical evidence inputs without replacing ordinary work. Explicit draft source refresh preserves progressed plans and history with both CAS bindings; initial approval restrictions remain on historical paired attestation. Its verify command requires literal absolute paths; linked agent subdirectories remain unsupported for strict authority review. These limitations apply to requested attestation, not ordinary task start. core/surface/includes/artifacts-attestation.md Ordinary context-preview retains an inert proposal and exact preimages without approval or review claims. context-apply revalidates the digest, closed request and preview binding before the native paired transaction; context-finalize-preview uses the same exact-reference path for initial population. No approve-context reply is required. core/artifacts/internal/operations/context_assistive.go Brownfield population covers six logical categories and five content documents through direct bounded inspection or optional scoped agents. The paired writer preserves source bindings and protected content; initial finalization checks all required outputs. core/surface/skills/context-creation/SKILL.md A drift audit stays read-only; an already authorized selected or full refresh proceeds without asking again per document. The installed draft/preview/apply example preserves source bindings and selected content. Matching hashes establish identity only; initialized refresh never reruns scaffolding, clears the marker or repeats finalization. core/surface/skills/context-check/SKILL.md Review this exact proposed pair Capture actual host-observed consent No reused approval, hash-only consent or caller-supplied authority
Related entries /init /create-context /status /context-check
Pinned evidence S13 · core/surface/skills/context-creation/SKILL.mdCurrent brownfield owner: qualified containment, bounded reports and join, scoped gaps, native document/provenance writes, separate 13-identity finalization and recovery. S26 · core/surface/skills/context-check/SKILL.mdInitialized refresh: read-only report, explicit selection, v1/v2 separation, native paired updates and preserved human history. Full owner read. Selected out
Receipt-only native apply Assistive workflow: source-bound implementation An agreed outcome becomes a maintained plan, direct or delegated implementation, proportional verification and an honest delivery report. Task context and dependency advice support progress; real user stops and host permissions remain effective. Greenfield and brownfield population use the paired native writer and exact finalization. Exact blob bindings identify inspected source bytes; the linked revision is the committed source base. Implementation, verification, live host qualification and release remain separate claims.
Current source changes The pinned native bridge enables task-context as a bounded informational read and does not require future reviewer/approval producers simply to author a structurally valid draft. Context requests route to ordinary context-preview, context-finalize-preview and exact-reference context-apply; historical receipt input stays a separate compatibility path. Installed host qualification is separate. Native Pi writer admission does not establish a complete live Pi workflow. On NO_LOCAL_STORE the client runs init-store once and retries the request once; a refused initialization keeps NO_LOCAL_STORE. On Windows the binary's architecture comes from the kernel (IsWow64Process2), never WMI or PROCESSOR_* environment. Index and outline pages are now fully checked before rows are yielded, including continuation identity and row shape. core/pysrc/_artifactlib.py The closed task-check request accepts bounded command observations matched to plan definitions. Ordinary context-preview and context-finalize-preview omit prompt identity; context-apply accepts an exact preview reference and digest. Invalid requests, changed preimages and stale CAS writes remain rejected. The historical smarts-apply request accepts the 0.2.0 SMARTS decision record. init-store takes a protocol-only request. core/artifacts/internal/operations/protocol.go The installed context-draft command prepares a paired document preview; context-finalize prepares initial finalization. context-apply commits the exact selected preview and digest. Context modes preserve scaffold and initialized refresh boundaries. core/pysrc/init-codearbiter.py Ordinary context-preview retains an inert proposal and exact preimages without approval or review claims. context-apply revalidates the digest, closed request and preview binding before the native paired transaction; context-finalize-preview uses the same exact-reference path for initial population. No approve-context reply is required. core/artifacts/internal/operations/context_assistive.go Ordinary create/adopt/update preserves the same source-evidence, selection and compare-and-swap checks as the native transaction. Missing or matching v1 provenance can be adopted; existing v2 ownership can transition through explicit whole-document selection. Original document/provenance preimages remain in transaction history for exact recovery. core/artifacts/internal/contextdocument/operations.go The lazy reference map routes provenance drift or requested initialized refresh to context-check after reading selected documents, provenance and the code map. The owner reuses selected scope and existing authorization. core/surface/includes/reference-map.md The context-request bridge invokes the qualified installed writer. Source-only builds do not authorize native context writes. Paired per-document atomicity does not mean all documents become visible atomically as one set.
Related entries /init /create-context /status /context-check
Pinned evidence S13 · core/surface/skills/context-creation/SKILL.mdCurrent brownfield owner: qualified containment, bounded reports and join, scoped gaps, native document/provenance writes, separate 13-identity finalization and recovery. S26 · core/surface/skills/context-check/SKILL.mdInitialized refresh: read-only report, explicit selection, v1/v2 separation, native paired updates and preserved human history. Full owner read. Selected warn
Preserve other owners and human bytes Assistive workflow: source-bound implementation An agreed outcome becomes a maintained plan, direct or delegated implementation, proportional verification and an honest delivery report. Task context and dependency advice support progress; real user stops and host permissions remain effective. Greenfield and brownfield population use the paired native writer and exact finalization. Exact blob bindings identify inspected source bytes; the linked revision is the committed source base. Implementation, verification, live host qualification and release remain separate claims.
Current source changes The preparation helper binds bounded drafts and retained sources to exact document/provenance preimages. Finalization preparation hashes all thirteen required outputs and retains missing output failures without modifying files. core/pysrc/_contextwriterlib.py The finite context contract permits a whole-document Markdown entry only as the sole entry for one of five named content targets. Explicit typed field entries remain available; the writer does not expose arbitrary repository-file writes. core/artifacts/internal/contextdocument/contract.go Whole-document replacement requires the explicit document selection and exact existing bytes. It preserves CONTEXT frontmatter and initialization state, checks the named document identity and Markdown structure, and retains the code-map size limit. Human content outside the authorized change remains the caller's preservation obligation. core/artifacts/internal/contextdocument/document_selection.go Ordinary create/adopt/update preserves the same source-evidence, selection and compare-and-swap checks as the native transaction. Missing or matching v1 provenance can be adopted; existing v2 ownership can transition through explicit whole-document selection. Original document/provenance preimages remain in transaction history for exact recovery. core/artifacts/internal/contextdocument/operations.go A drift audit stays read-only; an already authorized selected or full refresh proceeds without asking again per document. The installed draft/preview/apply example preserves source bindings and selected content. Matching hashes establish identity only; initialized refresh never reruns scaffolding, clears the marker or repeats finalization. core/surface/skills/context-check/SKILL.md Task / question boards, append-only audit history, release targets and ADRs keep their own writers. No raw Write/Edit fallback on refusal.
Related entries /init /create-context /status /context-check
Pinned evidence S13 · core/surface/skills/context-creation/SKILL.mdCurrent brownfield owner: qualified containment, bounded reports and join, scoped gaps, native document/provenance writes, separate 13-identity finalization and recovery. S26 · core/surface/skills/context-check/SKILL.mdInitialized refresh: read-only report, explicit selection, v1/v2 separation, native paired updates and preserved human history. Full owner read. Selected out
Initial finalize: separate approval Assistive workflow: source-bound implementation An agreed outcome becomes a maintained plan, direct or delegated implementation, proportional verification and an honest delivery report. Task context and dependency advice support progress; real user stops and host permissions remain effective. Greenfield and brownfield population use the paired native writer and exact finalization. Exact blob bindings identify inspected source bytes; the linked revision is the committed source base. Implementation, verification, live host qualification and release remain separate claims.
Current source changes The pinned native bridge enables task-context as a bounded informational read and does not require future reviewer/approval producers simply to author a structurally valid draft. Context requests route to ordinary context-preview, context-finalize-preview and exact-reference context-apply; historical receipt input stays a separate compatibility path. Installed host qualification is separate. Native Pi writer admission does not establish a complete live Pi workflow. On NO_LOCAL_STORE the client runs init-store once and retries the request once; a refused initialization keeps NO_LOCAL_STORE. On Windows the binary's architecture comes from the kernel (IsWow64Process2), never WMI or PROCESSOR_* environment. Index and outline pages are now fully checked before rows are yielded, including continuation identity and row shape. core/pysrc/_artifactlib.py The closed task-check request accepts bounded command observations matched to plan definitions. Ordinary context-preview and context-finalize-preview omit prompt identity; context-apply accepts an exact preview reference and digest. Invalid requests, changed preimages and stale CAS writes remain rejected. The historical smarts-apply request accepts the 0.2.0 SMARTS decision record. init-store takes a protocol-only request. core/artifacts/internal/operations/protocol.go Ordinary context-preview retains an inert proposal and exact preimages without approval or review claims. context-apply revalidates the digest, closed request and preview binding before the native paired transaction; context-finalize-preview uses the same exact-reference path for initial population. No approve-context reply is required. core/artifacts/internal/operations/context_assistive.go Initial finalization binds thirteen output identities, rechecks supporting source evidence, and changes the initialization marker with its paired provenance digest. Ordinary finalization requires no artifact approval receipt and does not certify semantic correctness. Already initialized repositories refuse a new finalization. core/artifacts/internal/contextdocument/finalization.go Brownfield population covers six logical categories and five content documents through direct bounded inspection or optional scoped agents. The paired writer preserves source bindings and protected content; initial finalization checks all required outputs. core/surface/skills/context-creation/SKILL.md The finalized set includes exactly the five content documents, their five v2 provenance companions, open-tasks.md, open-questions.md and overrides.log. Valid empty boards and a zero-event audit header satisfy their actual contracts. A document-write receipt never substitutes for finalization.
Related entries /init /create-context
Pinned evidence S13 · core/surface/skills/context-creation/SKILL.mdCurrent brownfield owner: qualified containment, bounded reports and join, scoped gaps, native document/provenance writes, separate 13-identity finalization and recovery. Selected out
Refresh: retain initialized state Assistive workflow: source-bound implementation An agreed outcome becomes a maintained plan, direct or delegated implementation, proportional verification and an honest delivery report. Task context and dependency advice support progress; real user stops and host permissions remain effective. Greenfield and brownfield population use the paired native writer and exact finalization. Exact blob bindings identify inspected source bytes; the linked revision is the committed source base. Implementation, verification, live host qualification and release remain separate claims.
Current source changes The installed context-draft command prepares a paired document preview; context-finalize prepares initial finalization. context-apply commits the exact selected preview and digest. Context modes preserve scaffold and initialized refresh boundaries. core/pysrc/init-codearbiter.py Whole-document replacement requires the explicit document selection and exact existing bytes. It preserves CONTEXT frontmatter and initialization state, checks the named document identity and Markdown structure, and retains the code-map size limit. Human content outside the authorized change remains the caller's preservation obligation. core/artifacts/internal/contextdocument/document_selection.go Initial finalization binds thirteen output identities, rechecks supporting source evidence, and changes the initialization marker with its paired provenance digest. Ordinary finalization requires no artifact approval receipt and does not certify semantic correctness. Already initialized repositories refuse a new finalization. core/artifacts/internal/contextdocument/finalization.go A drift audit stays read-only; an already authorized selected or full refresh proceeds without asking again per document. The installed draft/preview/apply example preserves source bindings and selected content. Matching hashes establish identity only; initialized refresh never reruns scaffolding, clears the marker or repeats finalization. core/surface/skills/context-check/SKILL.md Initial brownfield setup checks the intended repository and actual host permissions, then uses bounded inspection and the installed safe context writer. Ordinary Windows work needs no special scout profile or per-document artifact approval. An existing initialized repository routes to selected refresh instead of another scaffold or population cycle. core/surface/commands/init.md Routine maintenance selects changed v2 claims intersecting staged sources using a current bounded snapshot. Unknown evidence defers the affected document; native paired writes preserve owned fields and current identities. core/pysrc/_provenancelib.py Only selected pairs change Declare scoped versus full coverage Keep marker, human content and history No finalization, staging or commit here
Related entries /status /context-check
Pinned evidence S26 · core/surface/skills/context-check/SKILL.mdInitialized refresh: read-only report, explicit selection, v1/v2 separation, native paired updates and preserved human history. Full owner read. Selected skill
Recover by the same operation ID Assistive workflow: source-bound implementation An agreed outcome becomes a maintained plan, direct or delegated implementation, proportional verification and an honest delivery report. Task context and dependency advice support progress; real user stops and host permissions remain effective. Greenfield and brownfield population use the paired native writer and exact finalization. Exact blob bindings identify inspected source bytes; the linked revision is the committed source base. Implementation, verification, live host qualification and release remain separate claims.
Current source changes Ordinary context-preview retains an inert proposal and exact preimages without approval or review claims. context-apply revalidates the digest, closed request and preview binding before the native paired transaction; context-finalize-preview uses the same exact-reference path for initial population. No approve-context reply is required. core/artifacts/internal/operations/context_assistive.go Ordinary create/adopt/update preserves the same source-evidence, selection and compare-and-swap checks as the native transaction. Missing or matching v1 provenance can be adopted; existing v2 ownership can transition through explicit whole-document selection. Original document/provenance preimages remain in transaction history for exact recovery. core/artifacts/internal/contextdocument/operations.go Initial finalization binds thirteen output identities, rechecks supporting source evidence, and changes the initialization marker with its paired provenance digest. Ordinary finalization requires no artifact approval receipt and does not certify semantic correctness. Already initialized repositories refuse a new finalization. core/artifacts/internal/contextdocument/finalization.go Do not reset the project directory, clear the marker, overwrite a human file, or retry an already committed write. Finalization recovery is checked separately. A genuinely absent output needs a fresh typed preview and its own approval.
Related entries /init /create-context /status /context-check
Pinned evidence S13 · core/surface/skills/context-creation/SKILL.mdCurrent brownfield owner: qualified containment, bounded reports and join, scoped gaps, native document/provenance writes, separate 13-identity finalization and recovery. Selected state
Use the result through task-scoped selection Context knowledge feeds the existing feature, fix, test-change and reviewer callers. They select current relevant inputs and deliver them to each actual actor. A successful context write or initialized marker is not task authorization, current command proof, or evidence that a child loaded those inputs.
Related entries /init /create-context /status /context-check /feature /sprint /fix /refactor /review
Pinned evidence S27 · core/pysrc/_contextselectlib.pyInert scoped selector and delivery implementation: commands/cwd/evidence, native instruction order, actor/worktree/epoch binding, two-attempt limit and reviewer preflight. S09 · core/surface/skills/tdd/SKILL.mdCurrent TDD context preflight and obligation derivation re-read. Six-phase gate sequence also confirmed by current skill index; fix and test-only paths do not fabricate feature artifacts. S28 · core/surface/commands/review.mdCurrent full review owner: stable diff resolution, selected context delivered independently to each reviewer, local verdict before separately authorized posting. Selected cmd
Keep actual request scope Diagnosis only → recommend and stop Existing fix authority stays bounded No code, tests or configuration edits No compulsory new onboarding
Related entries /debug /fix
Pinned evidence S25 · core/surface/skills/debug/SKILL.mdFive-disposition diagnostic owner; transient validated handoff, freshness and authority boundaries. Full owner read. S19 · core/surface/commands/fix.mdCurrent full fix owner: validated handoff, independent caller authority, evidence revalidation, selected actor input and genuine target-red gate. Selected state
Capture symptom + applicability Observed / expected + cited sources Known reproduction state and limits Actual worktree, source/build/runtime IDs Unknowns stay unknown
Related entries /debug /fix
Pinned evidence S25 · core/surface/skills/debug/SKILL.mdFive-disposition diagnostic owner; transient validated handoff, freshness and authority boundaries. Full owner read. Selected skill
Check real discriminators Distinct mechanisms; no fixed hypothesis quota CONFIRMED / REFUTED / INCONCLUSIVE Capture limits constrain absent-event claims New safe discriminator → bounded further check
Related entries /debug /fix
Pinned evidence S25 · core/surface/skills/debug/SKILL.mdFive-disposition diagnostic owner; transient validated handoff, freshness and authority boundaries. Full owner read. Selected gate
Choose one disposition Retain multiple supported causes Keep separate authority blockers Missing access is not design Non-reproduction is not closure
Related entries /debug /fix
Pinned evidence S25 · core/surface/skills/debug/SKILL.mdFive-disposition diagnostic owner; transient validated handoff, freshness and authority boundaries. Full owner read. Selected out
confirmed_code_defect Supported causal mechanism Current expected behavior Concrete regression obligation Candidate fix, not repair consent
Related entries /debug /fix
Pinned evidence S25 · core/surface/skills/debug/SKILL.mdFive-disposition diagnostic owner; transient validated handoff, freshness and authority boundaries. Full owner read. Selected out
confirmed_noncode_cause Supported external / operational cause Existing owner, or owner unknown Next evidence / access action No forced source patch
Related entries /debug /fix
Pinned evidence S25 · core/surface/skills/debug/SKILL.mdFive-disposition diagnostic owner; transient validated handoff, freshness and authority boundaries. Full owner read. Selected out
design_question Actual competing expectations Maintainer’s intended-behavior choice Optional later attributed ADR No automatic ADR creation
Related entries /debug /fix
Pinned evidence S25 · core/surface/skills/debug/SKILL.mdFive-disposition diagnostic owner; transient validated handoff, freshness and authority boundaries. Full owner read. Selected out
no_action Positive applicable closure evidence No remaining work in this scope No default task or board note No state mutation
Related entries /debug /fix
Pinned evidence S25 · core/surface/skills/debug/SKILL.mdFive-disposition diagnostic owner; transient validated handoff, freshness and authority boundaries. Full owner read. Selected out
unresolved Known facts + missing discriminator Owner and next action when known Resume condition or prerequisite No invented cause or no-action verdict
Related entries /debug /fix
Pinned evidence S25 · core/surface/skills/debug/SKILL.mdFive-disposition diagnostic owner; transient validated handoff, freshness and authority boundaries. Full owner read. Selected state
Transient debug-handoff/1.0.0 Exact schema identifier is codearbiter.debug-handoff/1.0.0. Keep the packet in the active session. Input is at most 65,536 bytes; only explicit evidence-preserving compaction before first validation is allowed. No new automatic per-turn injection or persistent case directory.
Related entries /debug /fix
Pinned evidence S25 · core/surface/skills/debug/SKILL.mdFive-disposition diagnostic owner; transient validated handoff, freshness and authority boundaries. Full owner read. Selected gate
Installed helper: validate Resolve debug-handoff.py from the trusted loaded package. Exit 2 is invalid input; exit 3 or an unavailable helper blocks validated transfer. Do not silently repair a rejected packet, retry another interpreter, or continue an unvalidated repair.
Related entries /debug /fix
Pinned evidence S25 · core/surface/skills/debug/SKILL.mdFive-disposition diagnostic owner; transient validated handoff, freshness and authority boundaries. Full owner read. S19 · core/surface/commands/fix.mdCurrent full fix owner: validated handoff, independent caller authority, evidence revalidation, selected actor input and genuine target-red gate. Selected gate
Recheck current evidence Actual repository and worktree Source/runtime identity and cited evidence Commit + index + dirty working state Relevant drift → revalidate or block
Related entries /debug /fix
Pinned evidence S19 · core/surface/commands/fix.mdCurrent full fix owner: validated handoff, independent caller authority, evidence revalidation, selected actor input and genuine target-red gate. S25 · core/surface/skills/debug/SKILL.mdFive-disposition diagnostic owner; transient validated handoff, freshness and authority boundaries. Full owner read. Selected gate
Check actual caller authority Only code defect → fix is eligible Independent repair permission required Diagnosis-only stops at recommendation Packet intent cannot authorize work
Related entries /debug /fix
Pinned evidence S19 · core/surface/commands/fix.mdCurrent full fix owner: validated handoff, independent caller authority, evidence revalidation, selected actor input and genuine target-red gate. S25 · core/surface/skills/debug/SKILL.mdFive-disposition diagnostic owner; transient validated handoff, freshness and authority boundaries. Full owner read. Selected warn
Stop or return the supported result Invalid packet, unavailable helper, changed evidence, non-code/design/no-action/unresolved result, or absent repair authority: no dependent code repair.
Related entries /debug /fix
Pinned evidence S25 · core/surface/skills/debug/SKILL.mdFive-disposition diagnostic owner; transient validated handoff, freshness and authority boundaries. Full owner read. S19 · core/surface/commands/fix.mdCurrent full fix owner: validated handoff, independent caller authority, evidence revalidation, selected actor input and genuine target-red gate. Selected gate
Deliver context; prove target RED Actual fix-author scoped input first Named regression at the production layer Fail for the defect’s precise reason No feature artifact invented
Related entries /fix
Pinned evidence S19 · core/surface/commands/fix.mdCurrent full fix owner: validated handoff, independent caller authority, evidence revalidation, selected actor input and genuine target-red gate. S27 · core/pysrc/_contextselectlib.pyInert scoped selector and delivery implementation: commands/cwd/evidence, native instruction order, actor/worktree/epoch binding, two-attempt limit and reviewer preflight. S09 · core/surface/skills/tdd/SKILL.mdCurrent TDD context preflight and obligation derivation re-read. Six-phase gate sequence also confirmed by current skill index; fix and test-only paths do not fabricate feature artifacts. Selected agent
Minimum authorized fix Cause and scope from current evidence Change only after valid target RED Then green, obligations, coverage, lint No setup/import failure as a substitute
Related entries /fix
Pinned evidence S19 · core/surface/commands/fix.mdCurrent full fix owner: validated handoff, independent caller authority, evidence revalidation, selected actor input and genuine target-red gate. S09 · core/surface/skills/tdd/SKILL.mdCurrent TDD context preflight and obligation derivation re-read. Six-phase gate sequence also confirmed by current skill index; fix and test-only paths do not fabricate feature artifacts. Selected out
Current regression evidence Read real output and exit codes Preserve failing-before / passing-after Original caller’s TDD gates remain No acceptance by author assertion
Related entries /fix
Pinned evidence S09 · core/surface/skills/tdd/SKILL.mdCurrent TDD context preflight and obligation derivation re-read. Six-phase gate sequence also confirmed by current skill index; fix and test-only paths do not fabricate feature artifacts. S19 · core/surface/commands/fix.mdCurrent full fix owner: validated handoff, independent caller authority, evidence revalidation, selected actor input and genuine target-red gate. Selected out
Separate delivery authority Verified change is not a commit /commit retains nine gates PR, merge and release are separate No automatic outward publication
Related entries /fix /commit
Pinned evidence S11 · core/surface/skills/commit-gate/SKILL.mdCurrent Phase 5.5 through terminal commit rules re-read. Legacy heal vocabulary must not be generalized to v2 paired native writes. S04 · core/surface/includes/routing-table.mdAuthoritative intent-to-owner table, scope-touch routes and stop/return semantics. Selected state
Resume only from available evidence Validate retained bytes; recheck actual worktree, source, runtime and dirty-state applicability. If context was lost, say so and reconstruct from accessible citations. Do not claim durable memory or silently replace a malformed packet with an unvalidated repair.
Related entries /debug /fix
Pinned evidence S25 · core/surface/skills/debug/SKILL.mdFive-disposition diagnostic owner; transient validated handoff, freshness and authority boundaries. Full owner read. S19 · core/surface/commands/fix.mdCurrent full fix owner: validated handoff, independent caller authority, evidence revalidation, selected actor input and genuine target-red gate. Selected state
Concrete follow-up: its own /task authority A concrete follow-up uses the selected work owner under existing authorization. The work.py facade routes native and legacy writes without a second board. A design question may lead to an attributed ADR later, never by automatic conversion.
Related entries /debug /fix /task /adr
Pinned evidence S25 · core/surface/skills/debug/SKILL.mdFive-disposition diagnostic owner; transient validated handoff, freshness and authority boundaries. Full owner read. Selected cmd
/tribunal · deliberate audit Resolve scope and trusted bundle Inspect matching run directories No source text grants execution
Related entries /tribunal
Pinned evidence S29 · core/surface/skills/tribunal/SKILL.mdSource-bound opt-in audit, guarded hash-bound stranded-lock recovery, thirteen-card roster, review/verify, durable triage, report and separate follow-up completion. Selected gate
Check source-bound status Same source: audit or follow-up Drift / unbound / invalid: no reuse Terminal or new scope: fresh run
Related entries /tribunal
Pinned evidence S29 · core/surface/skills/tribunal/SKILL.mdSource-bound opt-in audit, guarded hash-bound stranded-lock recovery, thirteen-card roster, review/verify, durable triage, report and separate follow-up completion. Selected gate
Inventory → profiles → cost Mechanical facts before model work Bound concrete evidence inputs User acknowledges cost + settings
Related entries /tribunal
Pinned evidence S29 · core/surface/skills/tribunal/SKILL.mdSource-bound opt-in audit, guarded hash-bound stranded-lock recovery, thirteen-card roster, review/verify, durable triage, report and separate follow-up completion. Selected state
Allocate a fresh bound run start returns run_id + run_dir Persist source.json before reviews Record actual budget and capacity
Related entries /tribunal
Pinned evidence S29 · core/surface/skills/tribunal/SKILL.mdSource-bound opt-in audit, guarded hash-bound stranded-lock recovery, thirteen-card roster, review/verify, durable triage, report and separate follow-up completion. Selected warn
Retain history; do not relabel Changed evidence needs a fresh binding. A deliberate abort records run-aborted.
Related entries /tribunal
Pinned evidence S29 · core/surface/skills/tribunal/SKILL.mdSource-bound opt-in audit, guarded hash-bound stranded-lock recovery, thirteen-card roster, review/verify, durable triage, report and separate follow-up completion. Selected state
Unchanged-source resume audit → unfinished recorded waves follow-up → pending dispositions only
Related entries /tribunal
Pinned evidence S29 · core/surface/skills/tribunal/SKILL.mdSource-bound opt-in audit, guarded hash-bound stranded-lock recovery, thirteen-card roster, review/verify, durable triage, report and separate follow-up completion. Selected state
Inventory ≠ risk judgment inventory.json + readable projection Risk map adds contract judgment Optional mappers answer only gaps No repository scripts run by discovery
Related entries /tribunal
Pinned evidence S29 · core/surface/skills/tribunal/SKILL.mdSource-bound opt-in audit, guarded hash-bound stranded-lock recovery, thirteen-card roster, review/verify, durable triage, report and separate follow-up completion. Selected skill
Applicability + bounded packets Account for every lens card Declare finding/evidence scopes Only bound evidence may expand Recheck source before dispatch
Related entries /tribunal
Pinned evidence S29 · core/surface/skills/tribunal/SKILL.mdSource-bound opt-in audit, guarded hash-bound stranded-lock recovery, thirteen-card roster, review/verify, durable triage, report and separate follow-up completion. Selected agent
Generic reviewer · MODE review One assignment per active lens At most 5 concurrent; host-bounded Own findings / pending-leads files No child dispatch or shared-log edits
Related entries /tribunal
Pinned evidence S29 · core/surface/skills/tribunal/SKILL.mdSource-bound opt-in audit, guarded hash-bound stranded-lock recovery, thirteen-card roster, review/verify, durable triage, report and separate follow-up completion. Selected state
Persist and account for output Import leads through the helper Coordinator records actual events Check source before accepting output Missing output remains incomplete
Related entries /tribunal
Pinned evidence S29 · core/surface/skills/tribunal/SKILL.mdSource-bound opt-in audit, guarded hash-bound stranded-lock recovery, thirteen-card roster, review/verify, durable triage, report and separate follow-up completion. Selected skill
Root-cause triage candidates Preserve provenance and old keys Calibrate severity + confidence Retain serious counterarguments
Related entries /tribunal
Pinned evidence S29 · core/surface/skills/tribunal/SKILL.mdSource-bound opt-in audit, guarded hash-bound stranded-lock recovery, thirteen-card roster, review/verify, durable triage, report and separate follow-up completion. Selected agent
Fresh reviewer · MODE verify All serious / promoted serious claims Expensive inferential recommendations Attempt disproof with needed evidence
Related entries /tribunal
Pinned evidence S29 · core/surface/skills/tribunal/SKILL.mdSource-bound opt-in audit, guarded hash-bound stranded-lock recovery, thirteen-card roster, review/verify, durable triage, report and separate follow-up completion. Selected gate
Eligibility + confidence gate Independent confirmed / narrowed Keep only the supported surviving claim Helper refusal cannot be log-edited away
Related entries /tribunal
Pinned evidence S29 · core/surface/skills/tribunal/SKILL.mdSource-bound opt-in audit, guarded hash-bound stranded-lock recovery, thirteen-card roster, review/verify, durable triage, report and separate follow-up completion. Selected out
Eligible phase plan keep / combine only Preserve narrowed claims Emit wave-triaged after plan
Related entries /tribunal
Pinned evidence S29 · core/surface/skills/tribunal/SKILL.mdSource-bound opt-in audit, guarded hash-bound stranded-lock recovery, thirteen-card roster, review/verify, durable triage, report and separate follow-up completion. Selected warn
Refuted Preserve result and rationale Not a confirmed issue Not kept repair work
Related entries /tribunal
Pinned evidence S29 · core/surface/skills/tribunal/SKILL.mdSource-bound opt-in audit, guarded hash-bound stranded-lock recovery, thirteen-card roster, review/verify, durable triage, report and separate follow-up completion. Selected warn
Inconclusive / shared context Record verify-required Keep investigation work separate Do not invent independent confirmation
Related entries /tribunal
Pinned evidence S29 · core/surface/skills/tribunal/SKILL.mdSource-bound opt-in audit, guarded hash-bound stranded-lock recovery, thirteen-card roster, review/verify, durable triage, report and separate follow-up completion. Selected state
Genuine design fork decision-required, not a disguised factual uncertainty or automatic ADR Leads need explicit dispositions
Related entries /tribunal
Pinned evidence S29 · core/surface/skills/tribunal/SKILL.mdSource-bound opt-in audit, guarded hash-bound stranded-lock recovery, thirteen-card roster, review/verify, durable triage, report and separate follow-up completion. Selected out
Report + manifest projections Source binding + actual coverage Supported work and unresolved claims Lead and follow-up dispositions report-written → follow-up
Related entries /tribunal
Pinned evidence S29 · core/surface/skills/tribunal/SKILL.mdSource-bound opt-in audit, guarded hash-bound stranded-lock recovery, thirteen-card roster, review/verify, durable triage, report and separate follow-up completion. Selected gate
Issue filing disposition Dedup against actual existing / partial receipts Default: show selected eligible issue commands Execute only with explicit filing authorization issues-filed OR explicit filing-skipped
Related entries /tribunal
Pinned evidence S29 · core/surface/skills/tribunal/SKILL.mdSource-bound opt-in audit, guarded hash-bound stranded-lock recovery, thirteen-card roster, review/verify, durable triage, report and separate follow-up completion. Selected gate
Separate optional public telemetry Preview the full bounded aggregate payload No code / paths / source hashes / finding text Explicit per-run consent before sending telemetry-sent OR explicit telemetry-skipped
Related entries /tribunal
Pinned evidence S29 · core/surface/skills/tribunal/SKILL.mdSource-bound opt-in audit, guarded hash-bound stranded-lock recovery, thirteen-card roster, review/verify, durable triage, report and separate follow-up completion. Selected warn
Unanswered or failed follow-up Keep pending; preserve real partial receipts On resume, check current source again Do not rerun completed lenses by default
Related entries /tribunal
Pinned evidence S29 · core/surface/skills/tribunal/SKILL.mdSource-bound opt-in audit, guarded hash-bound stranded-lock recovery, thirteen-card roster, review/verify, durable triage, report and separate follow-up completion. Selected gate
Both dispositions resolved Filing done or explicitly skipped Telemetry done or explicitly skipped Source-bound follow-up rules still apply
Related entries /tribunal
Pinned evidence S29 · core/surface/skills/tribunal/SKILL.mdSource-bound opt-in audit, guarded hash-bound stranded-lock recovery, thirteen-card roster, review/verify, durable triage, report and separate follow-up completion. Selected out
run-completed Terminal signal, not just a written report Audit history stays durable No automatic repair, commit, merge or release
Related entries /tribunal
Pinned evidence S29 · core/surface/skills/tribunal/SKILL.mdSource-bound opt-in audit, guarded hash-bound stranded-lock recovery, thirteen-card roster, review/verify, durable triage, report and separate follow-up completion. Selected state
Account for every outcome Account for every finding and lead. The report includes eligible plans and unsupported, refuted or unresolved claims in separately labeled sections. Repeat review/triage for each recorded wave; report after the complete recorded scope is accounted for. A no-kept-work wave still has an explicit plan projection.
Related entries /tribunal
Pinned evidence S29 · core/surface/skills/tribunal/SKILL.mdSource-bound opt-in audit, guarded hash-bound stranded-lock recovery, thirteen-card roster, review/verify, durable triage, report and separate follow-up completion. S30 · core/pysrc/_tribunalrunlib.pyImplemented helper API, per-run writer guard, hash-bound stranded-lock archive and audit/follow-up/terminal event vocabulary; cooperative source-bound persistence.