review-changes

name

review-changes

description

Run a multi-agent adversarial review over the current working-tree changes. Fans out one finder per dimension (correctness, security, performance, maintainability) over the git diff, adversarially verifies each finding to cut false positives, and returns ranked blockers vs warnings. Invoke after completing a substantive change, or when the Stop review gate asks for it. This reviews a finished local diff, not a remote PR (use /review for a PR).

review-changes

Launch a multi-agent review of the working-tree diff through the Workflow tool, then act on the confirmed blockers. This is the worker behind the always-on Stop review gate (review-gate.py); it is also fine to invoke directly.

When to run

• The Stop gate blocked with "run the review-changes skill", or
• you just finished a substantive change and want the adversarial pass before declaring it done.

Skip for genuinely trivial changes (a typo, a one-line doc/comment, a config value with no logic). Say so in one line and stop.

How to run

1. Find the change context: cwd = repo root via git rev-parse --show-toplevel (fall back to the current directory if not in a git tree). The review targets git diff plus git diff --staged; if both are empty because the change was already committed, the finders fall back to git show HEAD.

2. Call the Workflow tool with the committed review script (do not inline a script of your own):

   Workflow({
     scriptPath: "~/.claude/skills/review-changes/review.workflow.js",
     args: { cwd: "<repo root>" }
   })
   

   The script ships next to this skill; expand ~ to the absolute home path. The workflow runs in the background and notifies on completion. It fans out one code-reviewer finder per dimension over the diff, pipelines each finding through an independent skeptic that tries to refute it, and returns the surviving findings ranked, with Correctness and Security marked as blockers.

   Fallback — Workflow tool unavailable. Not every harness/session exposes a Workflow tool. If the call fails because the tool does not exist, do not stall or improvise a different review: run the same shape manually. Spawn one code-reviewer agent per dimension — correctness, security, performance, maintainability — over the same diff (git diff + git diff --staged, falling back to git show HEAD), all in a single message so they run concurrently, each restricted to its one dimension. Then adversarially verify each finding yourself (try to refute it against the actual code) before reporting, and keep the same ranking: Correctness and Security are blockers.

3. When it completes, report the verdict to the user, blockers first. Fix the confirmed Correctness/Security findings in the same turn, or state plainly why a finding is declined. Performance and Maintainability are the author's call. If the workflow returns zero confirmed findings, say so in one line and move on.

Notes

• The review is read-only; it never edits. You apply the fixes.
• The workflow reuses the global code-reviewer agent as its finder, so findings follow the same severity and evidence discipline as a manual reviewer pass.