Claude Code

Topic file for running pr-reviewer from inside Claude Code. Named claude-code.md on purpose: claude.md resolves as CLAUDE.md on a case-insensitive filesystem and may be auto-loaded as project instructions.

Fetch tooling

Claude Code has the Bash tool. The skill drives metadata and clone itself with gh (GitHub) or glab (GitLab).

• From inside the matching repo: /pr-reviewer (current PR branch) or /pr-reviewer <num>
• From a barebones directory: /pr-reviewer https://github.com/<org>/<repo>/pull/<n> or /pr-reviewer <org>/<repo>#<n> (GitHub), <group>/<project>!<n> (GitLab)

Models

Same aliases as the-mister's Claude Code file, so one override vocabulary works for both skills.

Effort: the Agent tool takes an effort parameter (low, medium, high, xhigh, max); the skill's "high effort" is effort: "high". If your build of the tool lacks it, paste Think hard and read every call site before concluding. into the prompt instead and record which you did. Record the alias and effort you passed, not what the agent says it is. Never pass fable.

Fan-out primitive

One Agent call per reviewer, all three in one message, foreground so you cannot proceed without their results:

Agent(subagent_type: "general-purpose", model: "sonnet", effort: "medium",
      description: "pr-reviewer A correctness #<num>",
      prompt: <reviewer prompt from references/passes.md, fully instantiated>,
      run_in_background: false)

general-purpose has Bash, which the reviewer needs for git diff and grep in the review tree. Its read-only behaviour is an instruction in the prompt, not an enforced fence; the prompt's RULES block is what holds it. Explore is read-only by construction but may refuse git commands on a detached checkout; prefer general-purpose.

The assessor is one Agent call with model: "opus"; the summary one with model: "haiku". Never use isolation: "worktree": the review tree is already pinned and a second worktree would not be.

If Agent is not in your tool list, or the first call is denied, fall back to single mode and record mode: single (fallback: <reason>) in findings.md so the report says so.

Looping primitive

Claude Code ships /loop:

/loop /pr-reviewer [<num>]

Each iteration re-pins the revision (Step 2), reuses the clone, and appends a round to findings.md. /loop does not read files: the skill ends every unfinished round with a status line, and when a round finishes the revision (converged, --rounds reached, or Latest-state review incomplete) the iteration says so plainly and the loop is stopped (by you in dynamic mode, or by the user). Publication is offered only on that final iteration, so the loop is never blocked on a question. If the author is actively pushing, /loop 10m /pr-reviewer [<num>] paces the re-pins; the --max-refresh budget caps how many restarts the pushes can cause mid-review.

Posting

Step 11 of the skill, via AskUserQuestion then gh api … /reviews --input. Always after the report, always with explicit confirmation.