Skip to content
  • AI Toolkit plugin
  • v0.12.0
  • 6 skills
  • any language

review: code and PR review skills

Run a PR through two independent correctness passes, a type-safety check and a comment audit, and read one report where every finding has a file and a line.

  1. /plugin marketplace add noppu-labs/ai-toolkit
  2. /plugin install review@ai-toolkit

01 — What it does

What a review run produces.

  • passes

    Two correctness passes

    pr-review runs correctness twice: through the harness's code-review skill when there is one, otherwise correctness-review, and again with a subagent reading the diff by hand. A finding both passes made is labelled both passes.

  • outcomes

    Correctness findings end in an outcome

    A correctness finding names a path:line and ends in crash, wrong data shown, wrong data persisted, harmless or question, which separates a bug from an edge case you can live with.

  • refutation

    Suspects stay until refuted

    correctness-review drops a candidate only when the code disproves it, and quotes the line that does. A candidate it can neither confirm nor refute stays in, marked uncertain, with the fact that would decide it.

  • types

    Type findings propose a shape

    type-safety-review rules a diff or a path against numbered rules (PHP-1 to PHP-5, TS-1 to TS-4, PY-1 to PY-5), quotes the line and proposes the type to use instead, without editing any file.

  • comments

    Comment verdicts without edits

    comment-audit gives each added comment one of DELETE, TRIM, MOVE, KEEP, UNSURE or WRONG. It only writes a report unless you pass --apply, which makes one commit and does not push.

  • authors

    Findings written for the author

    pr-comments rewrites the report for whoever opened the PR: one comment per finding, with a code such as COR-01, a severity taken from a fixed label and, when the fix fits in six lines, a GitHub suggestion block.

02 — Skills

The review skills.

03 — How it works

What one run looks like.

  1. pr-review takes one PR number, or several in merge order for a stack, and resolves each one's base and head with gh. Each PR in a stack is reviewed against the head of the PR below it, so its diff holds only its own changes.
  2. It counts added plus deleted lines, minus lockfiles and binaries. With code-review listed, the count sets its level: high up to 300 lines, xhigh up to 3000, max above. Otherwise it sets how many correctness-review subagents run: one, three or four.
  3. It builds a structural brief for each PR, with investigate:brief when the investigate plugin is installed and plain git commands when it is not, and pastes the brief into the stage subagents' prompts.
  4. The correctness passes, the type-safety stage and the comment audit are dispatched in the same turn, so they run in parallel where the harness allows. The orchestrator reads their reports and never opens the diff itself.
  5. It merges what came back into one report per PR, with Correctness, Type safety and Comments sections. Every finding keeps its path:line, so it can be pasted as a review comment.
  6. pr-comments turns that report into comments for the author and gives a verdict of approve, comment or request_changes. A stage missing from the report is never treated as clean.

04 — Requirements

What you need.

  • The gh CLI, able to read the pull request, for pr-review. Its origin remote has to be the repository gh resolves pull requests against.
  • Claude Code 2.1.248 or later (verified on 2.1.289) with dynamic workflows enabled, to run the pr-review-stages workflow. Without them pr-review follows the same steps by hand.
  • Optional: the investigate plugin for a richer brief, and access to context7 so the correctness stage can check library calls against their docs, though the review goes on without either.
  • node on PATH, for pr-comments and for the task-output hook, which matches reliably on Claude Code 2.1.176 or later.

05 — Questions

Before you install.

How does the review plugin relate to investigate?

They are separate plugins and each works alone. When investigate is installed, pr-review builds each PR's structural brief with investigate:brief. Without it, pr-review builds a lighter brief from git diff and git grep.

Does the review plugin work in OpenCode?

Yes. The same skills ship as the npm package @noppu-labs/opencode-review, which targets OpenCode V2, and you call a skill with @pr-review. There pr-review uses correctness-review in place of code-review, has no workflow, and the task-output hook is left out.

Can I install it with npx skills add?

Yes. That route installs the skills only, so the workflow is missing and pr-review runs by hand. It also leaves ${CLAUDE_PLUGIN_ROOT} unresolved, so for OpenCode the npm package is the supported route.

Which languages does it cover?

The type-safety rules exist for PHP, TypeScript and Python. The correctness review checks each diff against the pitfalls of whichever language it is written in.

Does it change my code or post to the pull request?

correctness-review and type-safety-review are read-only. pr-comments leaves posting to whoever calls it. comment-audit edits comments and docs, never code, in one local commit, and only with --apply.

What is the task-output hook for?

It is a PreToolUse hook on Bash and Read that denies a call naming a .output file when that file is a symlink, as a background subagent's transcript is. Without it, such a read stops an unattended run on a permission prompt, and only the Claude Code plugin has it.

More plugins

All plugins