P06 is a public guided and runnable Academy lesson in this preview. Its shared action cards are the public course route.
Complete P05 in the same Academy fork and clone. Before Prepare, return to main and confirm that
your worktree is clean: git status --short should print nothing. Keep a native terminal open at
the clone root for Academy Prepare, Check, and Reset. Keep one codeArbiter harness open at that
same clone for inspection, the context audit, and the two governed commits.
This is a recovery lesson, not a cleanup lesson. The prepared scenario contains a stale
.codearbiter/CONTEXT.md, a matching stale provenance record, and an unrelated
docs/preserved-note.md. You will prove that the note survived byte-for-byte. Never delete,
recreate, normalize, or “tidy” it.
Native-terminal commands are entered directly and never start with !. codeArbiter commands and
agent messages are entered in your selected harness and never start with !. Every action card
names the actor and the surface so you do not have to infer where a command belongs.
The prepared context says Workshop Queue report output is JSON-only. The prepared source proves
the opposite: stable text is the default and structured JSON is optional. The old lifecycle link
also points to ADR-0002 even though P05 accepted ADR-0005. The provenance record names the older
CLI object 042746e43698e5d2a6de4c536f1024f893aef805; the prepared source object is
5b41fb168a8b258cfae7eebc46e8b9ea7696ba56.
You will make exactly two ordered single-parent commits after Prepare. The first changes only the
context and provenance record. The second introduces only a canonical recovery handoff. Academy
Check independently compares those commits with the prepared baseline and verifies that the
unrelated note has identical raw Git-object bytes before and after the repair.
At the root of your Academy clone on a clean main branch, run Prepare in a native terminal. Academy creates the numbered attempt and freezes the stale context, provenance record, and unrelated note before you inspect anything.
Why this matters: A prepared commit gives Check a fixed before-state and keeps a failed recovery attempt inspectable.
Academy prints academy/P06-context-drift-recovery/ATTEMPT_NUMBER and leaves you on that numbered branch.
Evidence
The prepared commit is the trusted baseline for every later byte and history comparison.
If that does not happen
If Prepare names a dirty worktree or branch condition, preserve the message and resolve only that condition. Do not reset, rebase, or delete an earlier attempt.
ATTEMPT_NUMBER means the number Academy prints, such as 1; do not type that word literally.
Stay on the numbered branch until Check passes or Reset creates a preserved retry.
2
Your host
Have the agent expose the contradictory evidence
In your selected codeArbiter harness, ask the agent to read the prepared context, provenance, preserved note, current CLI source, and ADR-0005 without editing. The agent must show why the context is stale before you select a recovery route.
Why this matters: The correction is justified by a tracked source contradiction, not by guessing that documentation is old.
Any OS · Claude Code
Read, do not edit, .codearbiter/CONTEXT.md, .codearbiter/.provenance/CONTEXT.json, docs/preserved-note.md, workshop_queue/cli.py, and .codearbiter/decisions/0005-terminal-blocked-ticket-lifecycle.md. Show the stale report claim, the source behavior that contradicts it, the recorded source object ID, the prepared source object ID, and the unrelated note boundary. Do not select a route, edit, stage, or commit.
Any OS · Codex
Read, do not edit, .codearbiter/CONTEXT.md, .codearbiter/.provenance/CONTEXT.json, docs/preserved-note.md, workshop_queue/cli.py, and .codearbiter/decisions/0005-terminal-blocked-ticket-lifecycle.md. Show the stale report claim, the source behavior that contradicts it, the recorded source object ID, the prepared source object ID, and the unrelated note boundary. Do not select a route, edit, stage, or commit.
Any OS · Pi
Read, do not edit, .codearbiter/CONTEXT.md, .codearbiter/.provenance/CONTEXT.json, docs/preserved-note.md, workshop_queue/cli.py, and .codearbiter/decisions/0005-terminal-blocked-ticket-lifecycle.md. Show the stale report claim, the source behavior that contradicts it, the recorded source object ID, the prepared source object ID, and the unrelated note boundary. Do not select a route, edit, stage, or commit.
Expected result
You can see that stable text is the default, structured JSON is optional, ADR-0005 replaced the lifecycle reference, and docs/preserved-note.md is outside the repair.
Evidence
This inspection supports your decision, but Check later verifies committed objects rather than an agent transcript.
If that does not happen
If any named evidence file is missing or differs from the prepared scenario, stop and preserve the attempt. Use Reset for a fresh prepared baseline; do not substitute evidence.
Run this codeArbiter command in the selected harness. It is a codeArbiter command, not a shell command, so it never begins with !. Read the finding before choosing the route.
Why this matters: The host audit makes the available context-recovery routes visible before the scope is changed.
Any OS · Claude Code
/ca:context-check
Any OS · Codex
$ca-context-check
Any OS · Pi
/ca-context-check
Any OS · Pi
/skill:ca-context-check
Expected result
The agent identifies the contradicted context claim and offers a scoped recovery decision instead of silently rewriting broad repository context.
Evidence
The final handoff records your declared route; Check does not prove that the host command ran.
If that does not happen
If the host reports an unrelated blocker, preserve its finding and resolve that blocker first. Do not claim the audit ran by writing a replacement receipt.
4
YouActive harness
Select the bounded re-scout decision
Choose scoped re-scout because the prepared source contradicts the recorded claim and ADR-0005 superseded the older lifecycle reference. Do not choose an alternate route or a broader documentation rewrite.
Why this matters: Re-scout refreshes one source-backed claim while preserving unrelated evidence and the prepared exercise boundary.
Expected result
The next instruction is limited to the stale report statement, its lifecycle link, and the provenance entry that names the source.
Evidence
The handoff later declares re-scout, but that declaration is not proof of a host invocation or of your reasoning process.
If that does not happen
If you cannot explain the contradiction from the prepared files, return to the evidence inspection instead of making a route choice.
re-scout is the sole permitted recovery route because the prepared source directly contradicts
the recorded claim. It refreshes this bounded evidence; it does not authorize an alternate route,
a broad context rewrite, or a new baseline.
5
Your host
Ask the agent for the exact two-file correction
Send the bounded correction instruction in the selected harness. The agent edits and stages only the two correction files, then stops for your review; it does not commit the handoff yet.
Why this matters: Separating the correction from the handoff makes the two-commit topology reconstructable.
Any OS · Claude Code
Apply only the P06 scoped re-scout correction. In .codearbiter/CONTEXT.md replace the stale report sentence with exactly: Workshop Queue report output defaults to stable text and supports structured JSON with --format json. Update the lifecycle reference from ADR-0002 to ADR-0005. In .codearbiter/.provenance/CONTEXT.json update only the sole source hash to 5b41fb168a8b258cfae7eebc46e8b9ea7696ba56. Do not edit docs/preserved-note.md, create the handoff, commit, or touch any other path. Stage exactly those two correction files and stop for my review.
Any OS · Codex
Apply only the P06 scoped re-scout correction. In .codearbiter/CONTEXT.md replace the stale report sentence with exactly: Workshop Queue report output defaults to stable text and supports structured JSON with --format json. Update the lifecycle reference from ADR-0002 to ADR-0005. In .codearbiter/.provenance/CONTEXT.json update only the sole source hash to 5b41fb168a8b258cfae7eebc46e8b9ea7696ba56. Do not edit docs/preserved-note.md, create the handoff, commit, or touch any other path. Stage exactly those two correction files and stop for my review.
Any OS · Pi
Apply only the P06 scoped re-scout correction. In .codearbiter/CONTEXT.md replace the stale report sentence with exactly: Workshop Queue report output defaults to stable text and supports structured JSON with --format json. Update the lifecycle reference from ADR-0002 to ADR-0005. In .codearbiter/.provenance/CONTEXT.json update only the sole source hash to 5b41fb168a8b258cfae7eebc46e8b9ea7696ba56. Do not edit docs/preserved-note.md, create the handoff, commit, or touch any other path. Stage exactly those two correction files and stop for my review.
Expected result
The staged path list contains exactly .codearbiter/CONTEXT.md and .codearbiter/.provenance/CONTEXT.json.
Evidence
The first learner commit is allowed to contain only the corrected context and provenance objects.
If that does not happen
If any other path is staged or the note changed, stop. Preserve the attempt and use Reset rather than unstaging around an unclear correction.
6
You
Review the first commit boundary yourself
Use the selected copyable request to inspect the staged correction before committing. Confirm that only the exact stale statement, lifecycle reference, and one provenance hash changed.
Why this matters: A small staged review prevents a correct concept from becoming an oversized recovery commit.
Any OS · Claude Code
Run git diff --cached --name-only and git diff --cached. Show both outputs and assess only whether the staged correction contains exactly .codearbiter/CONTEXT.md and .codearbiter/.provenance/CONTEXT.json, changes the stated stale context/provenance facts, and leaves docs/preserved-note.md untouched. Do not change files, stage, commit, or clear the boundary; stop for my decision.
Any OS · Codex
Run git diff --cached --name-only and git diff --cached. Show both outputs and assess only whether the staged correction contains exactly .codearbiter/CONTEXT.md and .codearbiter/.provenance/CONTEXT.json, changes the stated stale context/provenance facts, and leaves docs/preserved-note.md untouched. Do not change files, stage, commit, or clear the boundary; stop for my decision.
Any OS · Pi
Run git diff --cached --name-only and git diff --cached. Show both outputs and assess only whether the staged correction contains exactly .codearbiter/CONTEXT.md and .codearbiter/.provenance/CONTEXT.json, changes the stated stale context/provenance facts, and leaves docs/preserved-note.md untouched. Do not change files, stage, commit, or clear the boundary; stop for my decision.
Expected result
You can account for both staged paths and see no change to docs/preserved-note.md.
Evidence
Check reconstructs the committed path boundary; it does not authenticate that you personally viewed this review.
If that does not happen
If the review shows an unrelated change, keep the attempt as evidence and use Reset for a new numbered baseline. Do not amend or rebase committed work.
7
Your host
Commit the correction through codeArbiter
After your staged-boundary review, run the codeArbiter commit gate in the selected harness. This command never begins with !.
Why this matters: The correction must become the first of exactly two linear commits after Prepare.
Any OS · Claude Code
/ca:commit
Any OS · Codex
$ca-commit
Any OS · Pi
/ca-commit
Any OS · Pi
/skill:ca-commit
Expected result
The first descendant of Prepare contains exactly the two correction paths and leaves the worktree clean.
Evidence
Check reads commit ancestry and changed paths, not a commit message chosen by a learner.
If that does not happen
If the commit gate blocks, preserve its finding and resolve the named issue without expanding the two-file correction scope.
8
You
Generate the digest-bound recovery handoff
In a native terminal at the Academy clone root, run the local Academy helper. It derives a candidate only from committed Git objects and leaves it untracked and unstaged for your review.
Why this matters: The lesson teaches evidence-led recovery, not hand-authoring canonical JSON or blob hashes.
Only .codearbiter/reports/academy/P06-recovery.json is untracked. It names re-scout and has identical preserved-note digests; it is not staged or committed.
Evidence
The helper derives a candidate from committed objects. Check later recomputes it and does not prove that a host command ran.
If that does not happen
If the helper refuses, preserve its message and use Reset for a fresh attempt. Do not create, overwrite, stage, or hand-edit a substitute handoff.
The helper writes one untracked candidate. It does not stage, commit, reset, run Check, contact a
network service, or replace a report that already exists.
9
You
Stage only the generated handoff
In the native terminal at the Academy clone root, stage only the generated handoff. Do not stage the preserved note, correction files, or any substitute report.
Why this matters: The learner-visible staging boundary keeps the generated evidence separate from the earlier correction and makes the second commit reviewable.
The staged path list contains only .codearbiter/reports/academy/P06-recovery.json, ready for the second-commit review.
Evidence
Only this staged handoff path may become the second commit; the next action reviews that exact boundary.
If that does not happen
If the staged path list contains any other entry, stop and use Reset for a fresh attempt. Do not unstage around an unclear boundary or stage a replacement report.
10
You
Review the handoff as the second commit
Use the selected copyable request to inspect the staged handoff before committing. Confirm it is the only staged file and binds the re-scout route, prepared/correction commits, and preserved-note digests.
Why this matters: The handoff is evidence about the repair, so its own commit must stay separate from the correction it describes.
Any OS · Claude Code
Run git diff --cached --name-only and git diff --cached. Show both outputs and assess only whether .codearbiter/reports/academy/P06-recovery.json is the sole staged path, names re-scout, identifies the prepared and correction commits, and preserves identical note digests. Do not change files, stage, commit, or clear the boundary; stop for my decision.
Any OS · Codex
Run git diff --cached --name-only and git diff --cached. Show both outputs and assess only whether .codearbiter/reports/academy/P06-recovery.json is the sole staged path, names re-scout, identifies the prepared and correction commits, and preserves identical note digests. Do not change files, stage, commit, or clear the boundary; stop for my decision.
Any OS · Pi
Run git diff --cached --name-only and git diff --cached. Show both outputs and assess only whether .codearbiter/reports/academy/P06-recovery.json is the sole staged path, names re-scout, identifies the prepared and correction commits, and preserves identical note digests. Do not change files, stage, commit, or clear the boundary; stop for my decision.
Expected result
The staged boundary contains one canonical report and no recreated or edited preserved note.
Evidence
Check verifies the report shape, digests, and commit order; it cannot prove who generated the JSON.
If that does not happen
If the handoff mixes another path or claims a different route, stop and use Reset. Do not amend the first correction commit.
11
Your host
Commit the handoff through codeArbiter
After your handoff review, run the codeArbiter commit gate in the selected harness. This is the second and final learner commit for P06.
Why this matters: A dedicated evidence commit preserves the ordered distinction between correction and recovery record.
Any OS · Claude Code
/ca:commit
Any OS · Codex
$ca-commit
Any OS · Pi
/ca-commit
Any OS · Pi
/skill:ca-commit
Expected result
There are exactly two ordered single-parent commits after Prepare, and git status --short prints nothing.
Evidence
The verifier requires the second commit to introduce only the handoff.
If that does not happen
If the commit gate blocks, preserve its finding and correct only the named report condition in a new numbered attempt; do not amend history.
The first learner commit after Prepare changes exactly .codearbiter/CONTEXT.md and
.codearbiter/.provenance/CONTEXT.json. The second changes exactly
.codearbiter/reports/academy/P06-recovery.json. The corrected context states:
Workshop Queue report output defaults to stable text and supports structured JSON with --format json.
The handoff declares re-scout, identifies the prepared and correction commits, records the
before/after SHA-256 values for context, provenance, and the preserved note, and has identical
before/after note digests. Immediately before Check, git status --short prints nothing.
Check prints checkpoint P06-context-drift-recovery: passed; progress: .academy/progress.json.
Evidence
A pass verifies the frozen baseline, exact two-commit topology, object bytes, provenance repair, handoff, and clean worktree.
If that does not happen
If Check fails, preserve the committed attempt and read the failed predicate. Use Reset for a fresh numbered attempt instead of rewriting committed evidence.
A pass contains checkpoint P06-context-drift-recovery: passed; progress: .academy/progress.json.
Check proves the committed state, path boundaries, topology, raw-object digests, and preserved-note
identity. It does not prove that the host command ran, that a particular agent made an edit, or why
you selected the route. Those are honest workflow observations, not final-state claims.
If Check names a failed predicate, preserve the committed attempt and compare the failure with the
matching action card. Do not amend, rebase, force-reset, or conceal an earlier attempt.
The stale claim is a source contradiction, not a date mismatch. Compare the report parser choices
and default with the context statement before you choose a route.
Keep the correction and handoff separate: the first commit has two paths, the second has one.
After Check passes, leave the completed branch intact and return to main with this native-terminal action.
13
You
Return to main after a passing Check
After Check passes, return to main in a native terminal. Leave the completed numbered attempt branch intact for later inspection.
Why this matters: A completed attempt remains evidence. Returning to main prepares the clone for the next published lesson without erasing the recovery history.
Windows
git switch main
macOS
git switch main
Linux
git switch main
Expected result
You are back on main and the completed academy/P06-context-drift-recovery/ATTEMPT_NUMBER branch still exists.
Evidence
The retained attempt branch remains the inspectable recovery record; returning to main is not part of Check.
If that does not happen
If Check did not pass, stay on the preserved attempt and use Reset instead. Do not delete, force-reset, or rebase the completed or failed branch.
If Check did not pass, use Reset to create a preserved numbered retry.
14
You
Create a preserved numbered retry
Use Reset only when the attempt needs a new baseline. Run it in a native terminal at the Academy clone root; it archives the earlier branch before preparing the next number.
Why this matters: A retained failed attempt is evidence you can inspect, not debris to erase.
Academy archives the earlier attempt and prints the next academy/P06-context-drift-recovery/ATTEMPT_NUMBER branch.
Evidence
The archived branch remains reachable while the new attempt starts from the canonical prepared scenario.
If that does not happen
If Reset stops, preserve its message and current branch. Resolve only the named repository condition; never force-reset, rebase, or delete the prior attempt.
P07 is public, guided, and runnable in this preview. Continue with its published Academy lesson after
you return to main; do not substitute unpublished source exercises for the accepted course.
Context becomes trustworthy when its claims can be traced to the source and decision that support
them. A re-scout repairs a contradiction without pretending all repository context was freshly
reviewed. The two commits separate the source-backed correction from the record that describes it;
the external verifier then reconstructs both against the prepared baseline. Keeping unrelated work
byte-identical makes recovery auditable instead of destructive.
