CodeFerret

Many lenses. One review.

CodeFerret reviews a pull request through several independent code review skills at once, then merges their findings into a single review with inline comments. Each finding records which lenses found it, so agreement between them is visible rather than collapsed.

permissions:
    contents: write # resolveReviewThread needs it; contents: read disables thread resolution
    pull-requests: write

steps:
    - uses: pocketarc/codeferret@v1
      with:
          claude-code-oauth-token: ${{ secrets.CLAUDE_CODE_OAUTH_TOKEN }}

That is the whole setup. The action checks the repository out and installs what it needs.

@v1 moves with each 1.x release. Pin a full version, @v1.0.0, to hold a revision.

See review/README.md for the inputs, how to add a lens, and why each part is built the way it is.

action.yml   the composite action
lenses/      bundled review skills, packaged as a Claude Code plugin
review/      prompts, schemas, and the scripts a run uses

Testing it against itself

main holds only the tool. The test fixture lives on two branches that are never merged:

Branch	Contents
test/fixture	A clean Laravel-style PHP backend and Next-style TypeScript frontend, plus the coding standards and issue spec a review is scored against.
test/fixture-defects	Branched from test/fixture, adding a change that deliberately mixes real defects with plausible-looking non-defects.

Two pull requests exercise different things:

• test/fixture-defects against test/fixture is the realistic case. Its diff touches 7 files, and some findings root-cause into files the diff does not touch, which is what exercises the out-of-diff anchoring path.
• test/fixture-defects against main puts the whole fixture in one diff, so every lens reads every file.

The fixture mirrors PHP and TypeScript so that language-specific review guidance gets exercised on both sides.

The scoring key is deliberately not in this repository. A reviewer that can read the answers is not being measured.