Ac Review

1. Use $ac-memory to validate the active task state structure before review.
2. Set CURRENT_ROLE: reviewer.
3. Compare actual changes against the Contract:

• Scope
• Forbidden
• Steps
• Checks

4. If the original user request or Contract explicitly matches an exact review trigger, lock review mode before judging findings:

• 开发态联调 review -> dev-stage integration review
• 问题修复 review -> fix / regression review
• no trigger -> default code review

5. If the Contract / Evidence / Gaps explicitly describe development-stage assumptions, backend contract premises, local cache policy, or non-goals, review against those boundaries first.
6. Verify the changed artifacts are internally consistent:

• repo paths and module names match the current project
• docs-only checks were recorded honestly
• skill descriptions still explain what they do and when to use them
• for dev-stage integration tasks, unresolved items that depend on unverified backend behavior should be framed as Gaps / assumption / open question unless code or evidence clearly contradicts the declared boundary
• low-risk non-copy-edit work does not require review by default when it stays within a single docs / skills / template surface, or when it only aligns wording across a few coupled workflow assets without changing workflow semantics
• escalate when workflow routing semantics, Execution Contract / review conclusions, AGENTS core rules, shared templates, or coupled terminology across multiple workflow assets also change

7. Update 验收证据（Evidence） with concrete proof.
8. Update 未验证清单（Gaps） with residual risks, missing checks, and untested continue/doc workflows. 8.5. If review findings or follow-up comments appear, record each item in Review Findings（闭环） and close it explicitly with exactly one outcome:

• fixed
• rejected with evidence
• accepted as gap
• requires re-plan 8.6. Update the active task state file with the closure result before pass/fail conclusion, and keep .agents/state/INDEX.md aligned. 8.7. If a finding changes scope, files, checks, assumptions, or rollback strategy, set PLAN_FROZEN=false and return to $ac-plan.

9. When review reveals repeated mistakes, cross-document drift, or reusable workflow lessons, recommend $self-improving-agent after review to write candidate patterns into its own memory/.
10. Decide outcome:

• Pass: set CURRENT_ROLE: reviewer, set WORKFLOW_STATUS: completed, keep PLAN_FROZEN=true, then return control to $ac-workflow for final summary closeout and final [STATE] echo when helpful
• Fail: set CURRENT_ROLE: reviewer, set WORKFLOW_STATUS: active, set PLAN_FROZEN=false, and route back to $ac-plan

Hard rules:

• Base conclusions on Contract and evidence only.
• Do not skip Gaps when any verification is missing.
• Do not pass work that still contains known cross-document inconsistencies.
• Do not leave Review Findings（闭环） entries without an explicit closure outcome.
• Do not leave WORKFLOW_STATUS: active after a pass conclusion.
• Do not require $self-improving-agent for pass/fail; it is an optional post-review follow-up.
• Do not treat $ac-review as the final summary owner; passed work must return to $ac-workflow for 📝 总结.
• Do not treat development-only cache cleanup, debug placeholders, or declared backend assumptions as confirmed regressions unless release expectations or evidence clearly reject those assumptions.
• Do not switch into dev-stage or fix-review mode from fuzzy wording; only the exact trigger phrase or an explicit Contract record can do that.