Pr Review Toolkit

PR Review Toolkit

Portable PR review skill adapted for the project plugin. The upstream Anthropic pr-review-toolkit inspired the review lenses; this implementation keeps the procedure and rubrics in references/ so Codex, Claude Code, and Copilot can use the same source of truth.

Review Lenses

• code: project rules, bugs, contracts, maintainability, security, and performance.
• tests: behavioral coverage, edge cases, negative paths, and brittle tests.
• errors: silent failures, broad catches, hidden fallbacks, and unhelpful user-facing errors.
• comments: comment accuracy, stale documentation, redundant comments, and missing rationale.
• types: invariants, validation boundaries, encapsulation, and illegal states.
• simplify: clarity and refactoring opportunities while preserving behavior.
• all: all applicable lenses. This is the default.

Dispatch

Prefer a fresh-context reviewer when the current tool supports subagents. The reviewer must read references/review-procedure.md and references/review-rubrics.md before producing findings.

For Codex, use references/review-panel-protocol.md to select the bounded native-subagent panel before falling back. The default panel is three Luna/max reviewers grouped by risk; it is not a six-agent lens swarm.

• Claude Code: run Agent(subagent_type: "project:pr-review-toolkit", ...) if the bundled agent is available; otherwise follow the fallback branch.
• Codex CLI: invoke the project-preflight skill's bundled <project-plugin-root>/bin/project-preflight --format json executable by absolute path first; do not assume the plugin's bin/ directory is on PATH. When the scope is all (default) or the caller requests a pre-merge review, native_agent_status=available, and no single/inline opt-out is present, dispatch the bounded panel in references/review-panel-protocol.md through the current native subagent surface. Use gpt-5.6-luna at max effort by default, cap the panel at three reviewers, require structured accounting, synthesize and verify findings, and close completed agents. If native subagents are unavailable, or the caller selects a single/inline review, use the existing inline procedure. The copied agents/templates/pr-review-toolkit.toml remains an opt-in configured-agent path when the current Codex surface lacks direct native dispatch.
• Copilot CLI: plugin installation exposes this skill and, through .github/plugin/plugin.json, the .agent.md reviewer in agents/templates/. Invoke that configured agent when the current Copilot surface supports custom agents. If running outside a plugin install, copy agents/templates/pr-review-toolkit.agent.md to .github/agents/ or ~/.copilot/agents/. If no custom agent is available, use the fallback branch.
• Fallback: run the review inline in the current context after reading both reference files. This is the portable default path when preflight reports no native subagents or the caller opts out.

Inputs To Collect

• Scope: PR number, branch range, staged changes, unstaged changes, or explicit files.
• Lens selection: one or more of code, tests, errors, comments, types, simplify, or all.
• Mode: advisory review by default; pre-merge selects the bounded panel when available; delta is the round-2 review of fix commits only (one reviewer); single and inline opt out; edit mode is allowed only when the user explicitly asks to simplify/refine or implement review fixes.
• Round: a review that gates a merge is round 1 (full) or round 2 (delta). There is no round 3 unless the user asks.
• Overrides: user-supplied model, effort, panel size, lens grouping, or fallback choice always wins over defaults.

Required Workflow

1. Read references/review-procedure.md.
2. Read references/review-rubrics.md.
3. For Codex automatic dispatch, read references/review-panel-protocol.md and run the preflight gate before choosing panel versus fallback.
4. When the review gates a merge, read references/review-convergence.md: it decides whether round 2 runs, where fixes go, and when to merge.
5. Inspect the local project rules before generic preferences: nearest AGENTS.md, CLAUDE.md, .rules/, and relevant framework guidance.
6. Determine the diff scope with git status, git diff, git diff --cached, gh pr view, or gh pr diff as appropriate. For delta, the scope is the fix diff only.
7. Select applicable lenses. Use code for every review unless the user explicitly excludes it.
8. If a panel ran, account for every reviewer and synthesize/deduplicate findings; independently verify load-bearing claims and close completed one-off agents.
9. Report findings first, ordered by severity, with concrete file and line references.
10. If no actionable findings exist, say so clearly and list any residual test gaps or verification limits.

Output Shape

Use this structure unless the user asks for a different format:

markdownCopy code
**Findings**
- Critical: ...
- Important: ...
- Moderate: ...

**Questions**
- ...

**Residual Risk**
- ...

For simplify in edit mode, make the smallest behavior-preserving patch, run the relevant tests or checks, and summarize what changed.

Attribution

Inspired by Anthropic's pr-review-toolkit plugin in anthropics/claude-code/plugins/pr-review-toolkit. The upstream plugin README identifies the plugin license as MIT; this project implementation is written as an original project skill with attribution rather than a verbatim copy.