feat(code-review): wire start.md to the run-prefix loop (PLN-1229 Phase 3)

[wongk]

Summary

PLN-1229 Phase 3 — orchestrator wiring. The payoff phase: rewires the /code-review orchestrator (start.md) to run the deterministic prefix via the run-prefix loop (built + parity-proven in #173/#174) instead of walking ~19 helper stages one at a time. This collapses the prefix's serial model turns into a handful of orchestrator turns — the single biggest turn-count reduction in a review — while keeping the per-stage walk as a live, documented fallback (no behavior lost).

This is a prose-only change (orchestrator command + one skill + a SCHEMA doc fix). The runner it drives is already byte-parity-proven; this phase's safety net is careful authoring + an expert review of the rewire.

What changed

commands/start.md

• New "Deterministic Prefix — run-prefix loop" section: invoke run-prefix, then dispatch on next_action —
  • needs_singleton → invoke the singleton-dispatch skill, then re-invoke --resume-from <resume_stage>;
  • hygiene_exit → Gate A present + EXIT;
  • ready_for_reviewers → print the cache/fast-path notices, read CACHE_DIR from cache_config.json, hand off to the Walker Contract at stage_20_spawn_reviewers;
  • error → fall back to the per-stage walk from failed_stage.
• Reframed the Walker Contract as (a) the always-walked reviewer/verify/present tail and (b) the labeled per-stage prefix fallback. Gate A/B now execute inside run-prefix (their prose is the fallback recipe); Gate C/D still fire in the tail. Updated the "Batch deterministic helper stages" note to point at run-prefix, and kept the operator todo list accurate across the now-silent prefix.

skills/singleton-dispatch/SKILL.md — step 4 now branches on caller: return to start.md for --resume-from in the run-prefix flow; advance the walker only in the fallback.

SCHEMA.md — corrected the route-failure failed_stage to stage_19_cache_check (matching the merged runner) and noted the shallow-tier derive-static-spec stage.

Review

An expert review traced all four next_action outcomes against the runner code and confirmed the tail resolves everything it needs from on-disk artifacts (no HIGH findings). All flagged items fixed: the stale skill step 4, fast-path notice model-name parity, the shallow-tier stage note, and the SCHEMA failed_stage value.

Testing

• Full code-review suite: 1306 passed, 3 skipped (prose-only change — confirms no code impact)
• ruff / pyright unaffected (no code changed)

Version: code-review 3.3.0 → 3.4.0 (MINOR — additive with a live fallback; touches the orchestration contract).

Completes the landable sequence Phase 0→3. Phase 4 (measure the turn/cost reduction with the cost tool) is the remaining follow-up.

🤖 Generated with Claude Code