| name | code-review |
| description | Internal Harness instruction source for code-review. Route through visible Harness aliases or hook contracts instead of invoking directly. |
Code Review
Use this Skill for review-only checks. Do not edit subject code, configs,
tests, current docs, or canonical state while this Skill is active.
Read First
../../../.agents/references/code-style.md
../../../.agents/references/language-policy.md
../../../.agents/references/project-map-rule.md
../../../.agents/references/reviewer-independence.md
../../../.agents/references/review-tracing.md
../../../.agents/references/contract-gating-rule.md
./references/review-report.md
project_map.json, AGENTS.md, CLAUDE.md when present
Modes
light: targeted read-only inspection; concise line-referenced findings.
medium: post-change review; capture git metadata, changed ranges, Codex
review attempt, optional external model review, reconciliation, and report.
heavy: gate/release/docs-supporting review; independent Codex plus
external model attempts unless unavailable; unresolved critical findings
block readiness claims.
Use $code-debug for ordinary fixes after review and $harness-maintenance
for hook, skill contract, routing, or permission-policy fixes.
Required Work
- Resolve mode and review scope: base ref, included/excluded paths, generated
and runtime exclusions, staged/unstaged/committed surface.
- Capture git metadata and changed line ranges using
./references/review-report.md.
- Read subject files from disk; use
project_map.json for stable code
responsibilities when present.
- Prepare independent reviewer prompts with paths, constraints, and
objectives, not executor interpretations.
- Attempt Codex review through the available review surface, or record
codex_review_or_NOT_RUN.
- Attempt external model review only when configured. Provider-backed review
must run through
tooling/model_api/harness_external_review.py; hooks
require an active $code-review heavy session. Prefer targeted prompts from
tooling/model_api/build_review_prompt.py --scope changed; use agentic mode
only for high-rigor local read-only review. Do not send full-repository
bundles unless the operator accepts the cost.
- Verify model findings against checked-out files and diff. A model finding is
not a project fact until line-referenced and checked.
- Reconcile as
accepted, rejected_with_reason, needs_human, or
not_reproducible.
- For medium/heavy, write
.agents/state/review_traces/code-review/<YYYY-MM-DD>_runNN/review_report.md.
- Report Gate ledger for report generation, skipped reviewers, and affected
workflow gates.
Output
For light mode, lead with findings. For medium/heavy, include report path,
scope, git metadata, reviewer statuses, provider/model when used, reconciled
counts, unresolved findings, and Gate ledger.