garrison-review

garrison-review

Policy-read preamble (soft - D5/D12)

At the start of every invocation, look for the compiled Orchestrator policy at ~/.garrison/orchestrator/policy.json (or $GARRISON_POLICY_PATH).

• Policy present (a Garrison run): it is the single authority. This skill carries NO model/effort pins - its execution parameters come from the policy matrix cell for its phase (matrix[<phase>][<tier>]), and its gate duties from the bindable phase-skill contract (the Orchestrator fitting's PHASE_SKILL_CONTRACT.md): do the phase's work in the run context handed to you (runDir, card, phase), write the phase's gate-status entry under the runDir, and print the phase's GATE <phase>: <verdict> line before choosing the next list.
• Policy absent (standalone, any repo): proceed with the caller-supplied context and sensible defaults - NEVER stop. Report to the caller rather than writing gate-status/run artifacts, and skip any board/run-engine steps.

Same-model (Claude) code review of the current change. A standalone code reviewer.

No longer a per-slice pipeline gate (Part 8.1). The per-slice review merged into the fresh-context garrison-adversarial-review — once the deterministic wall (build/typecheck/lint/tests) plus the fresh-context and Codex passes existed, this same-model review's unique catches were ~zero, so it was retired from the build loop. It is retained here as a fully-functional standalone code-review skill (Claude's built-in code-review): use it directly — /code-review before shipping, a second look at a diff or PR — whenever you want Claude's own same-context review.

What it runs

Invoke the built-in code-review skill on the current diff — it reviews for correctness bugs first, then reuse/simplification/efficiency cleanups, at the session effort. (--comment posts inline PR comments; --fix applies findings — use only when asked.)

If code-review is unavailable, run the equivalent yourself: read the diff (git --no-pager diff <base>...HEAD + uncommitted), and report correctness bugs first, then quality cleanups, each with file:line and a concrete fix. That is what the built-in does.

Scope

• In a garrison build: review the SLICE diff only — capture BASE=$(git rev-parse HEAD) before the slice, review git diff $BASE...HEAD + uncommitted.
• Standalone: review the diff/PR/range the user names (default: uncommitted + last commit).

Loop role + output

• In a garrison build: report findings; real correctness findings send the slice back to garrison-implement to fix (garrison owns the retry ceiling). Cheap quality nits are applied; larger ones are logged. Re-review after a fix.
• Standalone: report the findings and stop — do not auto-fix unless asked.

Print one line in the lead context: GATE review: <clean|findings(n)> — <summary>. In a build, fold the result into the slice gate-status. Distinct from garrison-adversarial-review (this same session's built-in review vs. a FRESH-CONTEXT Anthropic reviewer with no access to this session's rationale) and from garrison-codex-checkpoint (the genuine cross-model check, now run once in the final phase, not per slice).