review-and-simplify

name	review-and-simplify
description	MANDATORY pre-commit orchestrator — fans out review lenses (code-review, security-review) in parallel, then runs a behavior-preserving simplify pass, before any commit is proposed.

review-and-simplify

When: after implement-and-evaluate reports all tasks done, all Critérios de Aceitação met, and gates green — before any commit is proposed. This is a MANDATORY pre-commit step. A commit MUST NOT be proposed without it having run and come back clean. Skipping it is not allowed, regardless of how small or "obviously fine" the diff looks.

Output: an aggregated findings report (by severity) from all lenses, a simplified diff with gates confirmed green, and — only once both are clean — a proposed commit message awaiting human approval.

Steps

1. Resolve scope. Resolve the active feature from .specify/state (fallback: newest dir under specs/). Determine the diff to review: the feature's tree-snapshot diff (git diff BASE_TREE HEAD_TREE, per the no-commit-rule snapshotting in subagent-driven-development) or, if no snapshots were taken, the working tree diff.

2. Resolve the lens set from the registry. Read .specify/review-lenses.txt — one skill name per line, ignoring blank lines and # comments. That file is the source of truth for which lenses run. Core seeds it with the stack-agnostic lenses:

   • code-review — quality against the constitution and checklist-template.md.
   • security-review — the security checklist (secrets, allow-list authz, injection, SSRF, crypto, supply-chain).

   Packs append stack-specific lenses to the same file without touching this orchestration — e.g. the ui-review pack appends perf-review + a11y-review, and the api-review pack appends api-contract + migration-safety. If the registry is missing (older install), fall back to the two agnostic lenses above.

   Dispatch every listed lens IN PARALLEL. Follow dispatching-parallel-agents — issue all lens dispatches in the same response so they run concurrently, not sequentially. Each lens is read-only and returns findings by severity (Critical/Important/Minor) with file:line.

3. Aggregate findings. Any Critical or Important finding from either lens MUST be fixed: hand it back to fix-runner (small, localized fix) or implement-and-evaluate (larger scope), then re-dispatch the lens that raised it to confirm the fix. Do not proceed to step 4 with open Critical/Important findings. Minor findings are logged for the human, not blocking.

4. Run simplify on the changed code. It edits for readability only (behavior preserved exactly), then re-runs .specify/gates/run-gates.sh and the project's tests itself to confirm everything is still green. If simplify reports it had to revert, treat the diff as unchanged and proceed.

5. If simplify changed non-trivial logic shape (restructured control flow, extracted/merged functions — not just renames or formatting), re-run the lens whose territory that touches (typically code-review; security-review too if the touched code was security-relevant) before proceeding.

6. Pre-commit GATE. Only once: all lenses are clean (no open Critical/Important), simplify has run, and gates/tests are green — propose the commit message(s) to the human and WAIT for explicit approval. Per the repo's commit-at-milestone convention, do not commit yourself; subagents never commit either.

Hard rules

• Skipping this step, or proposing a commit without having run it to a clean state, is not allowed.
• The lenses run in parallel, not sequentially — that's the point of dispatching both in one response.
• simplify runs only after the lenses are clean, never before (no point cleaning up code that's about to be sent back for a Critical fix).
• Do not commit. This orchestrator's job ends at proposing the commit message and waiting for approval.

Next: human approval, then commit (executed only after explicit approval — never by this skill or any subagent it dispatches).