review-team（审查小组）

Review Team

Use this skill when review quality matters and one reviewer would miss things.

Load references in this order

• references/scope-discovery.md
• references/intent-discovery.md
• references/preflight-risk-model.md
• references/persona-selection.md
• references/orchestration.md
• references/subagent-template.md
• references/findings-schema.json
• references/merge-normalization.md
• references/arbiter-pass.md
• references/action-synthesis.md
• references/severity-rubric.md
• references/output-format.md
• references/personas/catalog.md
• Only the selected persona files in references/personas/

Workflow

1. Resolve review scope

- If there are working-copy changes, treat those as the default scope. - If working copy is clean, inspect unpushed commits and decide if they are the intended scope. - If scope is ambiguous, ask the user one direct question with options.

1. Understand intent before reviewing

- Gather intent from conversation context, commit messages, PR/bead references, and docs. - If intent is still fuzzy, ask one short clarifying question.

1. Run preflight profile and risk tagging

- Measure file count, changed lines, file types, and critical-path indicators. - Apply references/preflight-risk-model.md and record risk tags. - Use this profile as input for persona scoring and concurrency plan.

1. Choose personas and assignment matrix

- Score personas with deterministic rules from references/persona-selection.md. - Build disjoint ownership slices by directory, service, or commit chunk. - Duplicate overloaded personas for large scopes (-a, -b, -c).

1. Announce execution note before spawn

- Publish scope, intent, risk summary, reviewer designations, and wave plan. - This is progress reporting, not a blocking confirmation step.

1. Spawn sub-agents with hard constraints

- Respect platform sub-agent limits when known. - If no clear limit is available, use a conservative cap of 6 concurrent sub-agents and run in waves. - Sub-agents are reviewers only; they do not edit code. - Each sub-agent must return JSON that matches references/findings-schema.json.

1. Merge and normalize findings

- Apply references/merge-normalization.md for validation, dedupe, and reconciliation. - Reconcile severity with references/severity-rubric.md. - If conflicts remain after merge rules, run targeted arbitration from references/arbiter-pass.md. - Keep only actionable findings with concrete code evidence. - Suppress findings below 0.50 confidence.

1. Synthesize top actions

- Derive grouped remediation actions from merged findings using references/action-synthesis.md. - Prioritize actions by severity and confidence.

1. Produce one final report

- Organize primarily by severity. - Include secondary domain summary plus coverage gaps. - Do not output an overall pass/fail verdict. - Keep the report issues-first and actionable.

Non-negotiables

• No style-only nits unless they carry real operational risk.
• Every finding must include concrete file and line references.
• Every finding must include evidence and confidence.
• Suppress findings below 0.50 confidence.
• Sub-agent output must conform to references/findings-schema.json.
• If no actionable findings exist, say so directly and list residual risks/testing gaps.