--- name: assistant-review description: "Reviews code and diffs (the current uncommitted changes, named files, a branch or a pull request), specs, plans, tests, documents, Jira tickets, and refinement material, fixes actionable findings, and runs one fresh re-review, or reports only in audit mode. Use for code review, C# pull requests, readiness checks on such material, and the workflow Review phase; QA runs only when required. Assessing a skill itself belongs to assistant-skill-creator; formal review of a delivered diff stays here." --- # Review ## Shared operating guidance Load [the common outcome guidance and shared stagnation policy](references/operating-guidance.md). Find scoped defects/regressions in diff and connected callers/contracts/tests/ configuration, including outside-diff consumers. Findings need real situations, file/line evidence, impact, calibrated confidence and smallest useful fixes. Style/speculation stays non-blocking. Preserve carried reduced-testing boundary/source through criteria, verification evidence and material snapshot refs. Additional tests require a distinct important uncovered failure and inadequate-existing-evidence rationale; scores/generic coverage advice cannot expand scope. Return material expansion to selection/approval. ## Authority and scope For "review this", resolve explicit scope, uncommitted work, carried state and current files. Missing required material blocks. Report-only selects audit; clear edit authority permits scoped review-fix. Ask on ambiguity before inspection or mutation. Ownership grants no edits. Standalone review runs without a task journal. Use local company-safe checks. External operations retain approval. Redact secrets and private data; preserve dirty work; return out-of-scope findings to workflow. ## Review process At structured boundaries load `contracts/index.yaml`. For behavior or cosmetic changes, an impact artifact, or locality/shared/unresolved claims, load `change_impact` and [change-impact.md](references/change-impact.md) before planning or dispatch; keep caller-only guidance outside Reviewer bundles. Structured batches load [review-loop.md](references/review-loop.md) before the first batch; v10 rules: - Formal review uses newly created independent actors and self-contained context (Codex `fork_turns=none`), bound to retained native creation/handoff/source authority. Missing isolation stays incomplete; no author fallback. - Each frozen batch attempts every expected pass: trivial/small light review (controller_intensity light) uses one newly created reviewer covering both concerns (another after a fix); light review records the compact result unless admission, a Pack or expanded impact requires structure (plan review: persisted/executable admission only); medium+ keeps at least two sibling-blind passes. Resolve `reviewer_context` for each pass, or `lifecycle_reviewer_context` only when assigned_scope or coverage_obligations affect lifecycle boundaries; synchronous/cosmetic scope alone uses ordinary context. Load the selected handoff before dispatch/return, and `return_validation` before aggregation. Account for every expected pass; incomplete coverage returns `HAS_REMAINING_ITEMS`. - Standalone review checks the user's specification. Audit retains mismatches as findings and finishes one read-only batch. Review-fix repairs authorized mismatches and requires Spec PASS before quality review. Workflow-composed review consumes current due-check evidence; acceptance review owns spec checks. - Review relevant spec, regression, testing, maintainability, bugfix, semantic/behavioral contract, loop-safety and security concerns. Apply matching [checklists](references/review-checklists.md) and [principles](references/review-principles.md); exclude unrelated sections. Medium+ retains its Design Coherence Pass. For a changed function with nested branches or loops, use the advisory [cognitive-complexity estimate](references/cognitive-complexity.md); it never blocks by itself. - Review carried Architecture Decision Packs, including invalid ones, using the matching checklist. Preserve freshness, types/lifecycle/compatibility/quality evidence and `review_intensive` independent challenge. - Before workflow-composed mutation, preserve or refresh the canonical verification selection under `F_REVIEW_VERIFICATION_SELECTION`. An active TDD finding goes back to the Build owner with its identities, oracle, scope, IDs, and bindings; the review fixer cannot bypass RED. Standalone review-fix retains build/tests. - After fixes, verify due selected/binding checks, freeze the source and run a fresh complete batch. No-build admission requires no required/completed fix or due selected/executable binding check. Change-impact projection cannot replace snapshot/coverage closure. Review 11.0 replaced the change-impact checker receipt with a checklist result; 10.0 packets stay current-readable ([change-impact.md](references/change-impact.md)). At entry, load `legacy_history` and [legacy-compatibility.md](references/legacy-compatibility.md) only for persisted pre-10.0 packets or v9 records; historical evidence cannot certify a current clean batch. ## QA and stopping QA stays separate from code review and runs after Build and Code Reviewer evidence only when explicit QA/acceptance, accepted Done Contract, harness acceptance, domain-scored scope, or scoped UI/visual/product/UX/docs/DX acceptance requires it. Size, delegation, source changes, and generic acceptance labels do not trigger QA. Carry approved preparation QA obligations and exact result/source bindings. Load [qa-evaluation-loop.md](references/qa-evaluation-loop.md) when required and [domain-rubrics.md](references/domain-rubrics.md) only for scoped domain quality. Required independent acceptance judgment uses a newly created fresh-context actor; unavailable isolation remains incomplete. Pure check execution stays discretionary. Audit stops after its first started batch, without fixes. Review-fix runs round 1, authorized fixes and validation, then fresh round 2. Round 3+ requires changed files, unresolved findings, validation failure, regression/drift, or changed hypothesis; score alone is insufficient. Load [score-tracking.md](references/score-tracking.md) only at a structured boundary for score bookkeeping, where Review contracts load the `current_round` EXIT gate or the `completion` `rubric_summary` artifact: workflow-composed review, a task journal or persisted result, a QA consumer, or a requested serialized result. A standalone review returned only to the user is not a structured boundary for score bookkeeping: it still loads its handoff and `return_validation` selectors, and reports findings, coverage and limits, with reviewer rubric dimensions as returned and no score progression or drift labels. Require a recorded pivot/restart decision before another pass only for evidenced shared semantic guard stagnation or an applicable material change. Scores alone never require a decision, evaluator reset or pause. ## Request and plan coverage For review_subject plan or acceptance, load request_plan (`request_plan_reconciliation`). One fresh `plan_coverage_and_feasibility` plan-review pass checks original requirements, decisions, feasibility/order, oracles, completion allocation and avoidable work. Corrections receive fresh closure review before approval; already-running approved plans retain authority. Final acceptance review solely owns request/plan/delivery reconciliation, including task-wide work and spec obligations: missing criteria, extra scope, changed-files or verification mismatch. It uses the common guide's [coverage check](references/operating-guidance.md#request-alignment). Missing work/findings remains `HAS_REMAINING_ITEMS`. ## Result Report scope, findings/fixes, verification and limits. No findings means "No material findings within the reviewed scope and available evidence"; `CLEAN` is an enum, not proof. Retain `contracts/output.yaml`, canonical reviewed scope, batch/round evidence, required checks and QA. A standalone review returned only to the user retains no `rubric_summary` score progression. `CLEAN`/`ISSUES_FIXED` require complete current coverage and no remaining must-fix/should-fix items; failed/blocked passes never qualify. Omit empty checklists and internal protocol from human summaries. Load [review-finding-permanent-rule.md](references/review-finding-permanent-rule.md) only for evidence of recurring process gaps or high-impact repeatable failures. Findings require no automatic permanent rule/framework. Retain classification envelopes for existing mapped all-ID distillation fixtures; omit them from ordinary human summaries.