Review Team
Run a fleet of specialized reviewer subagents against the same change. Each reviewer is intentionally unaware of the others and must stay within its assigned focus area.
Reviewer prompts live in
. Read only the reviewer files you plan to launch.
Modes
- single: Default. Launch reviewers once, aggregate findings, deduplicate, and report the final review.
- iterate: Launch reviewers, fix accepted findings, rerun relevant reviewers, and repeat until the review is clean or the remaining findings are intentionally deferred.
If the user does not specify a mode, use single. Use iterate only when the user explicitly asks to keep fixing, make reviewers happy, or run until clean.
Reviewer Set
Core reviewers:
reviewers/correctness-reviewer.md
reviewers/security-abuse-reviewer.md
reviewers/architecture-reviewer.md
reviewers/code-quality-conventions-reviewer.md
reviewers/simplicity-scope-reviewer.md
Specialist reviewers:
reviewers/product-ux-accessibility-reviewer.md
reviewers/performance-reliability-reviewer.md
reviewers/telemetry-observability-reviewer.md
reviewers/testing-strategy-reviewer.md
reviewers/api-compatibility-reviewer.md
reviewers/documentation-dx-reviewer.md
Default to all reviewers when the user asks for the full team. For narrower requests, launch only the relevant reviewers.
Launch Pattern
For each reviewer subagent, provide:
- The exact reviewer prompt from its Markdown file.
- The task description, PR/issue context, or user request.
- The diff or changed files to review.
- Any test results, logs, screenshots, or relevant repository context.
Add this wrapper to every reviewer prompt:
text
You are one reviewer in a fleet of independent code review agents. You cannot see the other reviewers. Stay strictly within your assigned role. Return only concrete, actionable findings with file and line references. If you find no issues in your area, say so clearly. Do not provide general commentary, praise, summaries, or duplicate concerns outside your mandate.
Prefer running independent reviewers in parallel. Do not ask multiple reviewers to solve the same broad task in the same way; their reviewer prompt is their boundary.
Aggregation
After reviewers finish:
- Merge duplicate findings.
- Drop findings that lack a concrete failure mode, user impact, security risk, or maintenance cost.
- Resolve conflicts using the codebase, tests, and stated requirements as evidence.
- Rank findings by severity and practical importance.
- Present findings first, with file and line references.
The final review should not expose raw reviewer transcripts unless the user asks for them.
Iterate Mode
In iterate mode:
- Run the initial reviewer set.
- Decide which findings are valid and should be fixed.
- Implement fixes directly.
- Run tests or checks appropriate to the changes.
- Rerun only the reviewers relevant to the changed areas or unresolved findings.
- Continue until no actionable findings remain, or until any remaining findings are explicitly documented as out of scope, false positives, or accepted tradeoffs.
Keep the user informed between iterations. Do not loop indefinitely; stop when additional iterations are not producing material new findings.
Final Response
For single mode, return the aggregated review.
For iterate mode, return:
- What was fixed.
- What verification ran.
- Which reviewers were rerun.
- Any remaining findings or accepted tradeoffs.