tomsej /wf-review
Argument: optional absolute path to a contract. Without a contract, review the current diff. You own the review: reviewers only report; verify every finding and delegate fixes sequentially. Never edit production code yourself.
GATE = node ~/Workspace/pi-ext/wf/wf-gate.mjs
1. Scope
TARGET = sc worktree status --json → target_branch (fallback main). Run
ocr delegate preview --from TARGET --to HEAD, restore excluded product files
such as .md when they carry the change or the contract names them, then load
rules with ocr delegate rule <files>. Keep lockfiles and generated code out.
Write .wf/review-packet.md: contract path or “no contract”, refs, merge_base,
final files, diff stats, and rules. Every reviewer gets it. Without ocr, use
git diff --stat TARGET...HEAD and disclose the fallback.
2. Pick the panel
With a contract, use its exact GATE agents <spec> --json → review plan;
never infer models from YAML. Without a contract, inspect the packet and use:
- docs/tests-only with no runtime behavior or sensitive config: 1 reviewer;
- behavior changes: 2 reviewers for correctness and tests/edge cases;
- security, auth, permissions, secrets, migrations, data integrity, concurrency, public APIs, or a large cross-module diff: 3 reviewers for correctness, the relevant risk, and impact/tests.
Read PI_PROVIDER and PI_MODEL. Reviewers must differ by model from the
current session and each other. Prefer claude/fable,
codex/gpt-5.6-sol, pi/zai/glm-5.2, then
pi/anthropic/claude-opus-5. Skip matching or unavailable models, record the
fallback, and never add reviewers merely because more models exist.
3. Review — max 2 rounds
Run each round's reviewers IN PARALLEL with subagent_spawn, each in a fresh
context; never collapse them onto one model. Give each the packet, focus,
read-only mode, and these rules:
- with claude harness: the prompt MUST start with
/code-review; codex would treat/reviewas plain text, so other harnesses get no slash prefix; - with pi harness: use
sem_impactfor blast radius/tests andsem_context; - findings are CLAIMS: severity,
file:line, problem, and impact; with a contract also cite the violated criterion or scope; - Critical/High = bug, security, data loss, or uncovered criterion; Medium = performance or missing error handling with context; discard Low unless useful;
- project rules are the checklist; read-only means never
gh pr createorgh pr merge— for the codex harness this sentence is the only guard.
Save raw outputs to .wf/reviews/<HEAD>/round-<N>/<focus>.md; a missing planned
report means incomplete review. Verify every finding against the code and
explain rejections. Batch legitimate findings by 3–5 and delegate them
sequentially; only one agent may edit at a time. Each batch ends with a commit
and targeted test. Send repeatable mechanical findings to review-guards;
never edit skills, the workflow pipeline, or AGENTS.md yourself.
Run round 2 only after fixes and only on changed areas. With a contract use its planned round; standalone use one best-matching reviewer.
4. Result
With a contract, GATE verify <spec> quick must pass (max 3 repair attempts per
hypothesis). Without a contract, do not run GATE agents or verify; after
fixes run targeted project tests. An unresolved finding or missing report means
FAILED; otherwise PASS. wf-impl creates the review attest only after its final
full gate; never create it here.
Report fixes with commits, rejections with reasons, planned/used reviewers and model fallbacks, artifact paths, scope stats, and gate/test results.