review

name	review
description	Post-implementation self-review. Ревью кода, ревью изменений, code review, проверка, самопроверка после реализации.

Post-Implementation Self-Review

Phase 4. The trigger list, the self-review checklist and the decision-pinning check live in .claude/rules/workflow.md and .claude/rules/documentation.md. Never auto-fix — present findings + a fix plan, wait for approval.

1. Workflow compliance

━━━ WORKFLOW COMPLIANCE ━━━
TDD:         [PASS / VIOLATION — tests after impl, or missing for new endpoints]
Checklist:   [PASS / SKIPPED — reason / N/A]   (.claude/rules-scoped/integration.md)
Doc markers: [PASS / MISSING — list]           (ARCH on boundaries, INVARIANT+Why, module docstrings)
Acceptance:  [PASS — live-drive evidence exists / VIOLATION — no live observation / EXEMPT — no runtime surface]
Gates:       [pre-commit-gates.sh output]

Acceptance per workflow.md Phase 4: a diff touching src/ needs a live drive of the affected flow — streaming diffs show the frames from curl -N, access diffs show BOTH the allowed and the refused key, config diffs cross an actual hot reload — with the output captured verbatim. tests/docs/config-only diffs are EXEMPT and say so.

2. Diff review

git diff --stat (only the intended files) → read the full diff for incomplete access guards, duplicated literals, formatting, config-schema drift → grep old identifiers after any rename, including monkeypatch/patch(...) string targets in tests/.

3. Architecture conformance

Align with CLAUDE.md: thin gateway (no business logic beyond routing), provider isolation, config-driven, fail fast. Access check before existence check. Errors through create_error(ErrorType, …). Request context via core.context.request_context. No direct os.getenv in providers/services.

4. Decision-pinning check

Per documentation.md: a bugfix in a repeatedly-regressing area must add or tighten an INVARIANT: / ARCH: / WHY: with a Why:. Flag any new INVARIANT: lacking a Why:, and any marker contradicting the user's latest stated rule. A new marker that met no pin trigger is flagged as noise to drop, not kept as harmless.

5. Tests & gates

New or changed logic has tests in tests/. Run the breadth the entanglement table demands, then .claude/scripts/pre-commit-gates.sh unpiped.

6. Receiving findings — verify before implementing

Applies to findings from ANY source that is not this session's own diff read: /code-review, ultra, a subagent reviewer, the user relaying someone else's report. A finding is a hypothesis about this codebase, not an instruction.

• Check it against the code before touching anything: does the claimed path exist, is the current shape deliberate (INVARIANT: / WHY: / DEBT: on the line), does the suggestion break a caller. A finding that survives none of these is answered with technical reasoning, not implemented.
• Unclear on ANY item ⇒ implement NONE of them yet. Items in one review interact; doing the understood four and asking about the fifth bakes in an ordering nobody chose.
• Push back in the same card. Rejecting a finding is a normal outcome and belongs in Residual with its reason — silently dropping it reads as agreement.

7. Summary — card form (subagent-contract.md)

What moved · Evidence (checks actually run, output verbatim) · Residual (findings + fix plan, awaiting approval) · Outside scope. Wait for approval.