jb-clawpatch-review

name	jb-clawpatch-review
description	Use when the user mentions Clawpatch/clawpatch.ai, semantic feature review, repo-wide AI audit, persistent findings, or clawpatch init/map/review/report/fix/revalidate.

JB Clawpatch Review

Use Clawpatch for semantic repo review with persistent findings and optional one-finding-at-a-time repairs.

Clawpatch is not a replacement for jb-autoreview: use jb-autoreview for final closeout review of a local diff, branch, commit, or PR; use this skill for feature-map-driven repo audits and tracked findings.

Prerequisites

Required:

• clawpatch
• git
• a supported Clawpatch provider, usually local codex

Check setup:

clawpatch --help
clawpatch doctor

If provider/model/effort matters, pass them explicitly instead of changing silently:

clawpatch review --provider codex --model <model> --reasoning-effort high

Core Workflow

Run from the repo root unless --root <path> is explicitly needed.

clawpatch init
clawpatch map
clawpatch review --limit 10
clawpatch report

Then inspect findings before acting:

clawpatch next
clawpatch show --finding <id>

Generate structured output when the agent needs to parse results:

clawpatch status --json
clawpatch report --json
clawpatch show --finding <id> --json

Review Scope

Use batches for large repos:

clawpatch review --limit 10 --jobs 10

Narrow to a feature or project when the user asks for a specific area:

clawpatch review --feature <id>
clawpatch review --project <name-or-root>

Review changes since a ref when that is the requested scope:

clawpatch review --since origin/main

Append extra reviewer guidance with a prompt file:

clawpatch review --limit 10 --prompt-file /tmp/clawpatch-guidance.md

Finding Handling

Treat every finding as advisory.

1. Read the finding:

   clawpatch show --finding <id>
   

2. Verify it by reading the real code path and adjacent files.

3. Reject speculative, low-value, unrealistic, or over-broad findings.

4. Triage verified status:

   clawpatch triage --finding <id> --status false-positive --note "Reason"
   clawpatch triage --finding <id> --status wont-fix --note "Reason"
   clawpatch triage --finding <id> --status uncertain --note "Reason"
   

5. For accepted issues, prefer small manual fixes unless the user explicitly asks Clawpatch to patch.

Explicit Fix Workflow

Only fix one explicit finding at a time.

Before fixing:

• Confirm the finding is valid.
• Confirm the worktree is clean or intentionally dirty.
• Do not run clawpatch fix for multiple findings in one command.

clawpatch fix --finding <id>

After fixing:

git diff
clawpatch revalidate --finding <id>

Then run focused project checks from .clawpatch/project.json or repo conventions, for example:

npm test
npm run typecheck
npm run lint

If the fix is wrong or too broad, revert manually and explain why. Do not commit unless explicitly asked.

Revalidation

Revalidate one finding after a manual or Clawpatch fix:

clawpatch revalidate --finding <id>

Revalidate a changed scope when the user asks for broader confirmation:

clawpatch revalidate --since origin/main

Use filters for batches:

clawpatch revalidate --status open --severity high --limit 10

Guardrails

• Never blindly apply Clawpatch findings or patches.
• Always inspect code before accepting a finding.
• Always inspect git diff after clawpatch fix.
• Prefer manual fixes for nuanced product or architecture decisions.
• Keep fixes at the right ownership boundary; avoid broad refactors unless the bug class requires it.
• Do not push, commit, open PRs, or land changes unless explicitly requested.
• Do not run destructive git commands as part of this workflow.
• Keep .clawpatch/ state as audit data; only remove it when the user requests cleanup.
• If Clawpatch reports provider/config trouble, run clawpatch doctor --json and fix setup before retrying review.

Final Report

Include:

• exact Clawpatch commands run
• scope reviewed: feature/project/since ref/limit
• findings accepted, rejected, or triaged, with brief reasons
• fixes applied, if any
• revalidation result
• tests/checks run
• remaining open findings or next recommended command