Workflow

1 — Scope

• Prefer the active pull request, current git diff, or an explicit file list from the user.
• If there is nothing to review, ask the user to point at a branch, PR, diff, or file set.

2 — Build a Change Summary

Before fanning out, build a concise change summary. Do not paste raw diffs into the subagent prompts — each subagent has tools to read files and inspect changes itself.

The summary should include:

1. Intent — what the change is trying to accomplish (from PR description, commit messages, or conversation context).
2. Changed files — list each file path with a one-line description of what changed (e.g., "added input validation", "refactored into helper").
3. Risk areas — anything risky: new dependencies, auth changes, hot-path modifications, API surface changes.
4. How to inspect — branch/PR/commit info so subagents can use their tools.

Keep the summary under ~50 lines. Subagents get better results reading code in context than scanning a wall of diff.

3 — Fan Out

Launch 2–4 parallel subagents using the area prompts below. Each subagent works in isolation — do not share one area's findings with another before synthesis.

Use unnamed subagents (no custom agent needed). Each gets a self-contained prompt with its area, the change summary, and the return format.

4 — Synthesize

When all subagents return:

1. Deduplicate findings that overlap across areas (e.g., a correctness bug that also shows up as a missing test).
2. Order by severity: breaking > wrong behavior > security > missing coverage > performance > product.
3. Apply the signal filter — drop anything that wouldn't block a PR.
4. If no blocking issues survive, say so and mention any meaningful testing gaps.

5 — Save Findings

Always save the synthesized findings to session memory at /memories/session/review.md. This makes them available for follow-up turns, fix planning, and cross-referencing with future reviews.

6 — Fix or Report

• Review-only: if the user said "only review", "just review", or "read-only", stop after reporting.
• Default (review-and-fix):
  1. For each fixable finding, launch one or more parallel Explore subagents to investigate the fix — this is faster and deeper than reading files inline. Give each Explore agent the finding, the relevant file paths, and ask it to return the specific change needed (what to replace, where).
  2. Apply fixes in severity order based on the Explore agents' recommendations.
  3. After applying, re-check the changed lines and diagnostics to validate.
  4. Report what was fixed and any remaining items that need the user's input.

Signal Filter

Keep only findings a senior engineer would block a PR for:

• Will fail to compile, type-check, or produce wrong results
• Clear, citable violation of workspace coding standards

…

council-review

调用多模型交叉审查代码变更、PR与高风险修改，辅助发现缺陷与争议点

autoreview

对本地或PR分支执行结构化代码审查，帮助提交或发布前发现问题。