skills/jb-clawpatch-review/SKILL.md
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.
npx skillsauth add bjesuiter/skills jb-clawpatch-reviewInstall this skill globally with one command. Works with Claude Code, Cursor, and Windsurf.
3 of 9 scanners reported clean
Some scanners were skipped, did not run, or reported a non-clean status. Review each row below.
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.
Required:
clawpatchgitcodexCheck 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
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
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
Treat every finding as advisory.
Read the finding:
clawpatch show --finding <id>
Verify it by reading the real code path and adjacent files.
Reject speculative, low-value, unrealistic, or over-broad findings.
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"
For accepted issues, prefer small manual fixes unless the user explicitly asks Clawpatch to patch.
Only fix one explicit finding at a time.
Before fixing:
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.
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
git diff after clawpatch fix..clawpatch/ state as audit data; only remove it when the user requests cleanup.clawpatch doctor --json and fix setup before retrying review.Include:
development
Use when the user asks for autoreview, Codex/Claude second-model review, or final review of dirty changes, a branch, commit, or PR before ship.
testing
Use when the user asks to cut, prepare, publish, tag, or verify a release, especially npm/package releases.
tools
Use when adding, writing, fixing, or exposing a script for the Tuna macOS launcher.
development
Use when smoke-testing a local web app, reproducing frontend bugs, verifying browser flows, or gathering evidence through PinchTab.