Skip to content

feat: a review skill proves how a change hurts a person - #19

Merged
Lutherwaves merged 1 commit into
mainfrom
feat/review-skill
Oct 22, 2026
Merged

Lutherwaves merged 1 commit into
mainfrom
feat/review-skill

Conversation

@Lutherwaves

Copy link
Copy Markdown
Owner

Summary

A new review skill: the review method itself, for a pull request, a branch or a diff of any size. The reviewer proves how a change hurts a real person and hands over the fix and the test that pins it.

  • Run before you read: build, lint, tests, the race detector. Probes run only in a throwaway worktree or temp dir, never in the checkout.
  • Hunt only what no existing test catches: real environments, ranked harm, consent and credentials, success only on success, the change's own claims against the code, seams, concurrency, hot paths, timeouts, silent skips.
  • Each finding: a bold harm headline with file:line, the concrete situation, evidence ("confirmed by running" with input → output, or "from reading"), the smallest fix, and the test that pins it. Summary first, "Declined to judge" and a verdict last.
  • Nothing is posted to the forge without the person's yes.

The routing table now sends "Review a diff, a branch or a PR" to builder review, used inside superpowers requesting-code-review or gstack /review. The README, AGENTS.md and CONTRIBUTING.md count nine skills; plugin.json gains the code-review keyword.

Type of change

  • feat — new feature

Checklist

  • PR title follows Conventional Commits
  • bash scripts/validate.sh passes locally (with the scrub). On macOS's bash 3.2 the script itself fails to parse (a heredoc inside $(...) at line 103, also on main); it passed with that heredoc moved out in a temporary copy. CI runs a newer bash.
  • skills/review/SKILL.md has valid frontmatter (name + description)
  • README updated
  • No machine-specific paths, tokens, or personal identifiers committed
  • Tested locally by installing the plugin in a Claude Code session — not done; the eval runs the skill with --plugin-dir

Notes for reviewer

Eval: tests/evals/review/ is a tiny notes app whose change breaks all three of its own description's claims while its tests pass: a token logged into a file the README says to attach to bug reports, a "your notes are safe" message counted over failed uploads, and an ignore rule compared raw against cleaned paths. Without the skill (4 runs) the agent finds the bugs but shows none of the six expect.md behaviours. With the skill (3 iterations), the final run shows all six. In 3 of 8 runs the headless permission check blocked the probe commands in both arms; the skill then marks findings "from reading" and says what it could not run. In one baseline run the agent applied the patch to the checkout itself, which the skill's rule prevents.

🤖 Generated with Claude Code

The review skill is the method a reviewer follows inside a review flow:
run the build and the tests first, hunt only what no existing test
catches, prove each harm by running a probe in a throwaway worktree,
and write each finding with its evidence, smallest fix and pinning test.
The review ends with the claims checked, what was declined and a verdict,
and nothing is posted without the person's yes.

The routing table sends "review a diff" to it, inside the superpowers
and gstack review flows. The eval fixture seeds a leaked token, a false
success message and a raw-versus-cleaned path seam.

Co-Authored-By: Claude Opus 5.5 <noreply@anthropic.com>
@Lutherwaves
Lutherwaves merged commit e8c987f into main Oct 22, 2026
2 checks passed
@github-actions github-actions Bot mentioned this pull request Oct 22, 2026
Sign up for free to join this conversation on GitHub. Already have an account? Sign in to comment

Labels

None yet

Projects

None yet

Development

Successfully merging this pull request may close these issues.

1 participant