garden-desk: Command for Claude Code

Review the given pull request. REVIEW.md at the repository root is the primary review instruction set; CLAUDE.md and the files it imports are secondary general project context.

Agent assumptions (applies to all agents and subagents):

• All tools are functional. Do not test tools or make exploratory calls.
• Only call a tool when required to complete the task.
• Read the pull request only with gh pr view, gh pr diff, and the gh api .../compare and gh api .../pulls read commands named below; read repository files only with Read, Grep, and Glob. Never run git.
• Models and thinking levels are pinned: the orchestrator runs on the model and effort set by the workflow; subagents are the review-gate (Haiku, low effort), review-sonnet (Sonnet 5, high effort), and review-opus (Opus 5, medium effort) agents defined in .claude/agents/. Launch every agent with subagent_type set to one of those names and never override its model or effort. Those agents have no Agent tool, so they cannot launch agents of their own.

Waiting for agents (this runs unattended, so it must never stall or loop):

• Agents run in the background. Tell every agent its time budget in its prompt ("You have N seconds; reply with what you have before then"). After launching a group of agents, call TaskOutput once per agent with block: true and the timeout given for that step. Do not poll, sleep, schedule wakeups, or use Monitor. Only a lens agent that is still running after its first wait gets one more TaskOutput with timeout 60000; no agent gets a third.
• An agent whose final TaskOutput status is not completed, or whose output is empty or unusable, has failed: call TaskStop on it and apply the fallback given for that step.
• Never end your turn while an agent you launched is still running; every step below ends with all its agents collected or stopped.

Review modes:

• Full review: no earlier claude summary comment exists on the pull request. Review the whole gh pr diff.
• Incremental review: the newest claude summary comment ends with a Reviewed <sha> line. Review only the commits after that sha, so work is not repeated. Per REVIEW.md, an incremental review posts CRITICAL and WARNING findings only.

Review tiers:

• Small to medium: at most 400 changed lines and at most 15 changed files, after the exclusions in REVIEW.md. One review-opus agent reviews all four lenses and validates its findings.
• Large: more than 400 changed lines or more than 15 changed files. Use the full agent set.
• Measure the diff for this run and select the tier one time before a lens agent starts.

Follow these steps precisely:

1. Launch a review-gate agent that returns: whether the pull request is closed or a draft; the current head sha from gh pr view <PR> --json headRefOid; and the sha from the Reviewed <sha> line of the newest claude comment in gh pr view <PR> --comments, or none. Wait with timeout 45000. Stop if the pull request is closed or a draft. Fallback: continue with a full review.

2. Read REVIEW.md in full. Read CLAUDE.md and any file it references with @ syntax. Identify the review lenses that REVIEW.md defines under "WHAT TO REVIEW"; each inline [tag] line that introduces a bullet list is one lens.

3. Choose the mode. With a reviewed sha, run gh api repos/<owner>/<repo>/compare/<reviewed-sha>...<head-sha> --jq '{status, files: [.files[] | {filename, additions, deletions, patch}]}'. If status is ahead and files is not empty, the mode is incremental and that output is the diff to review; also collect the earlier claude inline findings with gh api repos/<owner>/<repo>/pulls/<number>/comments --jq '.[] | select(.user.login == "claude[bot]") | {path, line, body}'. If status is identical, post nothing and stop. Otherwise (no reviewed sha, behind, diverged, or a failed command) the mode is full.

   Select the tier after the mode. In full mode, use gh pr view <PR> --json files --jq '[.files[] | select((.path | test("(^|/)(pnpm-lock.yaml|package-lock.json|yarn.lock)$|(^|/)__snapshots__/")) | not)] | {files: length, lines: (map(.additions + .deletions) | add // 0)}'. In incremental mode, calculate the same counts from the compare output files array and its filename, additions, and deletions values, with the same path exclusions. Log the counts and the selected tier.