CC-Review Agent Skill

--- name: cc-review version: 2.7.0 triggers: - 深度 review 这个方案 - review 这个复杂需求 - review 这个 bug 修复 - 做一次 cc-review - deep review this plan - review this implementation deeply - check code smells - review complexity hotspots - 复杂度专项 review - run cc-review reads: - PLAYBOOK.md - references/review-methods.md - references/plan-review-branch.md - references/implementation-review-branch.md - references/e2e-and-plugin-verification.md - references/complexity-optimization-playbook.md - references/complexity-report-template.md - scripts/collect-review-context.sh - scripts/analyze_complexity.py - references/checklist-contract.md writes: - path: current response durability: ephemeral required: true - path: devflow/changes/<change-key>/task.md durability: durable required: false when: plan or investigation contract review changes the task contract effects: - optional deep review - read-only reviewer agent dispatch - finding aggregation - complexity hotspot review - reroute recommendation entry_gate: - Classify the review target as plan, implementation, PR, or mixed. - Read only the task, PR, diff, code, tests, logs, screenshots, and docs needed to review the requested scope. - Use Git history and current diff as the only durable review memory; do not load or create process files. - For repeat reviews, use `git diff <old>...HEAD` or `scripts/collect-review-context.sh` to narrow the delta before re-reviewing. - Freeze the requested scope before finding smells; report only issues inside the change blast radius or clearly amplified by it. - When the scope includes loops, rendering, repeated scans, database/API iteration, large inputs, or performance-sensitive paths, select the built-in complexity facet and use the local scanner/reference copy only as leads. - Subagents are optional read-only reviewers; their raw output stays in the conversation and is not saved to files. exit_criteria: - Findings are listed first, ordered by severity and backed by concrete file/line, command, diff, UI, log, or missing-evidence proof. - Every finding has impact, recommendation, and route: cc-plan, cc-do, cc-check, cc-act, or stop. - Findings include an ASCII Branch Chain when they involve plan scope, root cause, implementation behavior, PR diff risk, or code smell propagation; each non-trivial chain records evidence, diagnosis, causal path, Phenomenal/Essential/Philosophical layers, at least three upstream layers, and at least three downstream layers. Tree connector characters stay ASCII while node text follows the configured output language. - Plan or investigation review findings are written into the relevant `task.md` section before exit. - Implementation review findings are returned with concrete repair options and a blocking user choice; no repair happens until the user selects an option. - In-scope code smells are either findings, explicit defers, or clean with reason. - Selected review facets or changed surfaces are checked, skipped with reason, or blocked; no artificial finding cap was applied. - If the complexity facet is selected, findings include current pattern, estimated current complexity, recommended change, estimated complexity after, risk level, and tests or measurements needed. - Subagent findings, when used, are accepted, merged, downgraded, or rejected by the main thread before final output. - If no issues are found, the answer says so and names residual test or evidence risk. - No process file was created. reroutes: - when: Plan assumptions, scope, architecture, design, or DX contracts are wrong or incomplete. target: cc-plan - when: Implementation findings require code, test, docs, UI behavior, logs, or PR text changes. target: cc-do - when: Deep review is clean and only fresh evidence verification remains. target: cc-check recovery_modes: - name: scope-reset when: The review starts drifting into unrelated historical debt. action: Return to the current diff, task.md, or PR scope and discard out-of-scope notes. tool_budget: read_files: 30 search_steps: 18 shell_commands: 18 ---