Build a Feature End to End
/ca:feature is the one entry point for new implementation work. Run it with a one-line description of what you want built:
/ca:feature "add rate limiting to the public search endpoint"Host syntax
Claude Code uses/ca:feature ”…”; Codex uses $ca-feature ”…”;
Pi uses /ca-feature ”…”. Shared artifact formats do not imply identical host authority.
The HTML writer needs the installed engine’s capabilities. An explicit attestation claim separately
uses the typed-artifact host boundary.The request authorizes the agreed feature work. The lane maintains useful artifacts, follows real dependencies and checks the resulting behavior. Ask about consequential missing choices and respect explicit user pauses; routine artifact updates do not create another permission layer.
The work, from your side
Intent → specification → plan → reviewed delivery
Full-lane overview. New typed HTML work requires the applicable installed-host capabilities. Small-lane work and retained Markdown pairs follow the distinct paths below.
You set the outcome
Review the specification
- What you get
- A specification with observable acceptance criteria, exclusions and verification oracles.
- Before moving on
- Correct missing behavior and resolve consequential choices. Preserve decisions already made and any requested review pause.
You review the approach
Inspect the linked plan
- What you get
- Tasks with exact paths, dependencies and checks linked to the specification criteria.
- Before moving on
- Check criterion coverage, real dependencies and current specification binding. In-scope maintenance uses the existing authorization.
codeArbiter works and reports progress
Test, implement, review
- What you get
- Implementation progress, actual check results, relevant reviewer findings and pending work.
- Before moving on
- Resolve applicable blocking findings. REVIEW and DONE describe implementation progress, not independent verification.
You retain delivery authority
Inspect the commit and PR
- What you get
- A selectively staged commit and a pull request with fresh verification for the current source.
- Before moving on
- Check implemented scope, criterion coverage and unresolved findings. A PR is not a merge or a deployment.
Reading map, not a captured run or live status. Only the actual artifacts, approvals and fresh checks establish what has completed.
See one complete example
Section titled “See one complete example”Your first feature follows the same saved-search CSV export from a request through a corrected criterion, linked plan, actual fixture tests, review and PR checklist. It distinguishes illustrative human dialogue from captured fixture evidence. This page is the lane reference for applying that pattern to your own work.
You can state the outcome naturally; explicit commands select the same existing lane. A clear request does not replace the user decisions or permissions required by that lane.
First: full lane, small lane, or resume
Section titled “First: full lane, small lane, or resume”The command classifies the request before it writes anything.
| Path | When it applies | First durable evidence |
|---|---|---|
Full lane | Any scope touch, new public surface, new dependency, more than two implementation files, or uncertainty | A maintained spec under .codearbiter/specs/ |
Small lane | At most two implementation files, no sensitive or architectural scope touch, and one to three testable criteria | Inline criteria reflecting the request plus one append-only triage.log line |
Resume | A matching spec or plan already exists | The existing pipeline re-enters at its furthest completed checkpoint |
The small lane trims ceremony, not gates: it still starts with test obligations, clears review and commit-gate, and finishes through a PR. If the classifier is uncertain, the full lane wins.
An interrupted run is not silently restarted. Name the same slug or ask for resumable work.
The owning lane inspects the furthest valid checkpoint, not merely the latest filename. An
usable spec without a plan returns to planning; an incomplete spec resumes its missing requirements.
Inspect IN_PROGRESS work and outstanding workers before redispatch. REVIEW and DONE record
implementation progress; actual checks determine verification. Resume and recover
explains the continuation boundaries.
Refine the Idea Into a Spec
Section titled “Refine the Idea Into a Spec”The brainstorming skill frames the outcome, selects a viable approach, makes acceptance observable, writes the specification and reviews it. It uses existing answers and asks only about consequential unresolved choices. Ordinary method choices stay within the agent’s scope. An explicitly requested pause remains effective.
Verification obligations come from agreed criteria and applicable contracts or security controls. One criterion may require several checks. The spec remains the working agreement for the feature.
Break the Spec Into a Plan
Section titled “Break the Spec Into a Plan”The writing-plans skill creates coherent tasks with useful outcomes, relevant paths, actual dependencies and concrete verification. Task size follows useful progress and review, without a compulsory duration or a separate agent lifecycle for each small edit.
For new full-lane work, the spec and plan are written as
.codearbiter/specs/<slug>.html and .codearbiter/plans/<slug>.html through the
installed typed-artifact engine. Existing Markdown pairs stay on their exact
legacy path. Readiness diagnostics identify missing task structure or criterion coverage.
Resolve requirements needed for the affected task while continuing independent authorized work.
Inspect the working agreement
Section titled “Inspect the working agreement”Open the actual specification and plan. In the specification, inspect caller, trigger, observable behavior, exclusions and each verification oracle. In the plan, follow those stable criterion IDs to tasks, dependencies, paths and named checks. Ask for a scoped revision where the connection is missing. Use the native typed writer for revisions, preserving human edits and stable references. Opening HTML does not create an approval event. Use Review specifications and plans for the inspection checklist.
A typed plan can exist as an incomplete draft preview. Required coverage is many-to-many; the specification owns criterion identities. The native engine preserves source identity and concurrent-write checks. A structural failure prevents claiming a successful save; a draft label alone does not prevent already-authorized work.
Implement Test-First, Batch by Batch
Section titled “Implement Test-First, Batch by Batch”The executing-plans skill selects meaningful batches and hands their scope to the implementation owner. Choose direct implementation or useful delegation. Review the combined change at the scope its risk and applicable policy warrant.
The tdd skill runs six phases for each task, in order:
- Obligations and approach. Map claims to meaningful checks before changing code.
- Baseline. Prefer test-first feature checks. Existing implementation may use current criterion-focused checks and meaningful negative controls. Confirmed fixes require genuine regression RED; refactors use passing characterization baselines.
- Implementation. Make the scoped change and run proving checks and affected existing tests.
- Obligation verification. Match claims to observed results, leaving failures and untested requirements outstanding.
- Coverage. Meet the declared profile’s required metrics and critical-path requirements, with unavailable measurements explicit.
- Lint and handoff. Run applicable quality tools and return actual results and pending checks.
After a batch, the coordinator reports changes, checks and next work, then continues within the authorization. An explicitly requested pause, actual consequential choice or applicable host restriction stops affected work. Checkpoint numbering is not itself a dependency.
Feature coordinates coherent batches; sprint owns full-plan autonomous execution and ends at an open PR. Both preserve real user decisions and current checks.
Commit selected work
Section titled “Commit selected work”When a batch is ready, run:
/ca:commitThe commit-gate skill is the only path to a git commit. It runs nine phases: permission, branch, classification, verification (tests, lint, secrets scan), behavioral proof, diff review, selective stage, message, and commit. A finding at any phase stops the commit.
The gate does not auto-decide hard gates. A secret in the diff, a crypto change, or a branch violation is a true stop that requires your explicit resolution. “It looks right” is not authorization.
Open the Pull Request
Section titled “Open the Pull Request”When the branch is ready to propose for integration, run:
/ca:prThe pr command verifies the current branch’s base-to-HEAD diff and applies its reviewer matrix. Existing applicable results can be reused; missing or affected checks run before readiness is claimed. Unresolved BLOCK findings must be corrected. Opening a PR does not authorize a merge, and exact-head hosted CI remains required before an authorized merge.
Know that the lane is complete
Section titled “Know that the lane is complete”Before calling the feature done, verify all of these:
- criteria still represent the agreed outcome and any material changes have their owner’s decision;
- implementation progress is accurate;
DONEdoes not imply verification; - each required claim has adequate current evidence, with missing or failed checks explicit;
- applicable local checks and review cover the actual source;
- any requested commit or PR was created through its owner under the existing authorization;
- merge, publication and live-host qualification are reported only when actually established.
Report implementation, verification, review and delivery separately. A plan label or one focused test cannot stand in for all four, and an unfinished unrelated task does not erase useful progress.
Common stops
Section titled “Common stops”| Stop | What it means | Resume path |
|---|---|---|
A consequential unresolved decision | The actual owner must choose | Ask once, preserve settled answers and continue independent tasks |
A task has no exact path or verification | The plan is underspecified | Return that task to writing-plans |
A confirmed fix lacks target RED | The regression baseline is missing | Observe the actual defect before repair; setup failure is insufficient |
A reviewer reports BLOCK | A required correction remains | Resolve the finding and rerun affected verification |
The session ends | Durable state remains on disk | Reinvoke feature; do not start over |
Related
Section titled “Related”| Skill or command | Where |
|---|---|
/ca:feature | |
/ca:commit | |
/ca:pr | |
brainstorming | |
writing-plans | |
tdd | |
executing-plans | |
commit-gate |
For the concepts behind gated lanes and gate strengths, see Concepts.