review-loop

Drive a review-fix-rerun cycle on the given PR using one or more review engines. The orchestrator is engine-agnostic — all engine-specific work routes through adapters described in references/engine-adapter-contract.md.

Inputs:

• PR_NUMBER (optional; defaults to PR for current branch).
• ENGINES (one or more of: bugbot, copilot, claude; defaults to bugbot,copilot,claude — all engines whose adapter is present + retrigger tool is reachable. Reachability caveat (A2): claude is in the default set only on repos where its action workflow (claude-code-review.yml) is installed; where absent, find_claude_comments.sh emits CLAUDE_NOT_INSTALLED=true and claude self-excludes from the default so a bare /review-loop doesn't block to HUNG on an un-provisioned engine. An explicit /review-loop <PR> claude still attempts it and degrades loudly — see references/engine-claude.md § Tool invocations.)

This skill is invoked via commands/review-loop.md — the single review entry point. Pass ENGINES as bugbot, copilot, claude, or any subset (e.g. bugbot,copilot,claude); single-engine runs are just a one-element list.

Before you trigger engines — wrap docs first if this is a doc-heavy / IDEA PR. If the PR carries substantial docs (IDEA file, plan, index, devlog, guides), run a bare /wrap (the --scope=docs default) FIRST, then trigger engines — so the reviewer sees docs at their merged shape and doc-consistency findings land in this single review pass instead of as post-review drift. /wrap never merges, so it is safe to run before review; this is the wrap-before-review single-review cadence — one pass over the wrapped PR covers code + docs (no separate deliverables/docs passes), then /land merges. Mechanics: skills/wrap/references/WRAP_BEFORE_REVIEW.md. (Code-only PRs: skip — go straight to the loop.)

Pre-flight — un-draft the PR. /work opens PRs as draft so the push-triggered Claude engine doesn't bill a review on every WIP commit (see references/engine-claude.md § Push-triggered model). The review stage is where that flips: if claude is in ENGINES and gh pr view <PR> --json isDraft -q .isDraft is true, mark it ready (gh pr ready <PR>) before Phase 1 — a draft PR makes claude no-op (the run fires but posts nothing → reads SILENT), so its verdict is unavailable until ready. Bugbot/Copilot are unaffected by draft state. Skip the un-draft only if the user explicitly wants the PR to stay draft.

Hard bounds (enforced by the loop)

• max_commits_per_session = 20
• max_active_work_minutes = 240 (excludes ScheduleWakeup sleep time; 240 covers large multi-engine surface-migration PRs whose fix-cycle count legitimately accumulates past 180)
• max_idle_polls = 10 (consecutive wakes with no new finding AND no new push, across all engines). Lowered from 20: with the 1200s Phase 4 backstop (step 1), 10 idle polls ≈ 3.3h wall-clock — within max_active_work_minutes, and the Monitor accelerator (references/MONITOR_ACCELERATION.md) ends most waits on a real event rather than a poll, so this count is a backstop, not the primary cadence. New-push detection: Phase 4 compares the scratch file's last_push_sha against git rev-parse HEAD on each wake; if they differ (e.g. an out-of-band push by another process or the user), reset idle_polls=0, update scratch last_push_sha, and re-enter Phase 1 to fetch fresh state for the new SHA. Without this check the counter accumulates forever past a push the loop didn't initiate.
• Targeted tests only inside the loop; broader regression deferred to hand-back
• Feature branch only — never main (per RULE_git-safety)
• Batch fixes per cycle into one commit, not one-per-finding
• Commits: standard RULE_git-safety applies — feature branches are the agent's sandbox, so the loop commits and pushes Tier 1 fixes autonomously. Tier 2 still needs explicit per-finding fix-direction approval. Protected-branch guardrails remain in force: never main, never merge into a protected branch, never force-push to protected, never --no-verify.