| name | meta-codereview-current-diff |
| description | Read the current uncommitted diff, run three independent reviewers (safety + tests-coverage + style) in parallel, then arbitrate a single BLOCK / BLOCK_WITH_OVERRIDE / PASS_WITH_NOTES verdict. Use before commit when you want a multi-perspective second-opinion instead of a single-reviewer agent loop. |
| kind | meta |
| meta_priority | 65 |
| always | false |
| final_text_mode | step:arbitrate |
| triggers | ["multi-reviewer diff","codereview my diff","三路 review","审查当前 diff"] |
| provenance | {"origin":"opensquilla-original","license":"Apache-2.0"} |
| composition | {"steps":[{"id":"read_diff","kind":"skill_exec","skill":"git-diff","with":"[Truncated]"},{"id":"review_safety","kind":"llm_chat","depends_on":"[Truncated]","with":"[Truncated]"},{"id":"review_tests","kind":"llm_chat","depends_on":"[Truncated]","with":"[Truncated]"},{"id":"review_style","kind":"llm_chat","depends_on":"[Truncated]","with":"[Truncated]"},{"id":"arbitrate","kind":"llm_chat","depends_on":"[Truncated]","with":"[Truncated]"}]} |
Codereview of Current Diff (Meta-Skill)
A combinator-style meta-skill that runs three independent
reviewers in parallel over the currently-uncommitted diff, then
arbitrates a single verdict with a strict priority rule.
This is the same combinator + arbitrate pattern as
meta-security-review-bundle, applied to a code-review domain. The
three reviewers each carry a tight, project-specific rubric — safety
focuses on injection / credentials / G1.6 contract; tests focuses on
the "new public surface ⇒ corresponding test" expectation; style
flags only a fixed list of antipatterns so it doesn't drift into
free-form opinions.
Trigger surface
Fire by saying multi-reviewer diff, codereview my diff, or one of the
localized triggers listed in the frontmatter. The diff is read from the working
tree (git diff --cached HEAD, falling back to git diff HEAD if nothing is
staged).
Fallback
If read_diff returns NO_DIFF, the downstream reviewers should
reply with their respective clean verdicts ("CLEAR" / "PASS" /
"CLEAN") and arbitrate emits PASS_WITH_NOTES: clean. If a reviewer
LLM step fails, the orchestrator's partial outputs are visible in
step_outputs; operator should manually re-run the failed reviewer.