/ai-pr

Use when the ask is a pull request / branch review in chat (not a CI-only ask).

Prefer to delegate: task({ subagent_type: "pr", description, prompt }) runs the ladder in the background; await_subagent before you answer. Then call pr_review with the child's pr_review_draft so the Agent Review canvas opens — the human edits and publishes from there.

Structured Agent Review (/review @pr:N or the Agent Review CTA) already attaches the PR package — do not review on the primary. First turn: spawn pr (and security / a perf pr when those skills are active) in parallel, await_subagent, then one pr_review. Pass @state/pull-requests/pr-N/ so children skip gh pr diff and test suites.

Ladder (inside the pr subagent)

1. Scope — pin PR number / repo / branch before reading the diff.
2. Map — on-disk package (summary.json, all.diff, files/) when the parent names it; otherwise gh pr view + gh pr diff. Never run test suites. 2b. Graph — call code_graph({ action: "review" }) after MAP (pass paths when known). Use impact for blast radius before walking more of all.diff. Fall back to grep/read if status is unavailable/skipped.
3. Checks — skip when a package was provided; otherwise one gh pr checks for context. CI-only asks belong to ci.
4. Critique — correctness, boundaries, blast radius, contracts, tests; prefer graph hotspots, then targeted read/grep on tip when available (not diff_only).
5. Evidence — path + lines + short snippet per concrete finding.
6. Conclude — Verdict prose + pr_review_draft JSON for the parent.

Parent handoff

After await_subagent returns a draft:

1. Call pr_review once with that payload (canvas opens).
2. Keep the chat reply to one line — do not paste the full review into chat.
3. Do not post to GitHub from the agent.

Budget

Cap at about 12 bash/gh calls. Batch independent calls in the same turn. Large diffs spill under @state/tool-output/ — aggregate; never paste the full diff.