| name | pr-review-fix-loop |
| description | Use when a PR or branch needs automated review and fixing, when the user says "review and fix", "clean up this PR", "fix PR issues", or wants to iterate a PR to passing state |
PR Review-Fix Loop
Orchestrate an automated review-fix loop for a PR using subagents. You are the controller only - you dispatch subagents and manage the loop. You NEVER read code, edit files, run tests, or fix issues yourself.
Core principle: Controller dispatches, subagents work. Fresh context per review. Loop until clean or max iterations.
Input Resolution
digraph input {
"Argument provided?" [shape=diamond];
"Use provided PR/branch" [shape=box];
"On a branch (not main)?" [shape=diamond];
"gh pr list --head <branch>" [shape=plaintext];
"PR found?" [shape=diamond];
"Use found PR" [shape=box];
"Ask user for PR number or branch" [shape=box];
"Ready" [shape=doublecircle];
"Argument provided?" -> "Use provided PR/branch" [label="yes"];
"Argument provided?" -> "On a branch (not main)?" [label="no"];
"On a branch (not main)?" -> "gh pr list --head <branch>" [label="yes"];
"On a branch (not main)?" -> "Ask user for PR number or branch" [label="no"];
"gh pr list --head <branch>" -> "PR found?";
"PR found?" -> "Use found PR" [label="yes"];
"PR found?" -> "Ask user for PR number or branch" [label="no"];
"Use provided PR/branch" -> "Ready";
"Use found PR" -> "Ready";
"Ask user for PR number or branch" -> "Ready";
}
The Loop
digraph loop {
rankdir=TB;
"Start: Have PR number" [shape=doublecircle];
"iteration < max (default 10)?" [shape=diamond];
"Dispatch 3 review subagents IN PARALLEL" [shape=box];
"Collect all issues from all 3" [shape=box];
"Any issues found?" [shape=diamond];
"Dispatch fixer subagent with ALL issues" [shape=box];
"Fixer commits and pushes" [shape=box];
"Wait for CI to start" [shape=box];
"Increment iteration" [shape=box];
"STOP: Report remaining issues to user" [shape=octagon, style=filled, fillcolor=red, fontcolor=white];
"Done: PR is clean" [shape=doublecircle];
"Start: Have PR number" -> "iteration < max (default 10)?";
"iteration < max (default 10)?" -> "Dispatch 3 review subagents IN PARALLEL" [label="yes"];
"iteration < max (default 10)?" -> "STOP: Report remaining issues to user" [label="no"];
"Dispatch 3 review subagents IN PARALLEL" -> "Collect all issues from all 3";
"Collect all issues from all 3" -> "Any issues found?";
"Any issues found?" -> "Done: PR is clean" [label="no"];
"Any issues found?" -> "Dispatch fixer subagent with ALL issues" [label="yes"];
"Dispatch fixer subagent with ALL issues" -> "Fixer commits and pushes";
"Fixer commits and pushes" -> "Wait for CI to start";
"Wait for CI to start" -> "Increment iteration";
"Increment iteration" -> "iteration < max (default 10)?";
}
Subagent Dispatch
Review Phase (3 subagents in parallel via single Task tool message)
| # | Subagent Type | Model | Prompt essence |
|---|
| 1 | superpowers:code-reviewer | default | Review PR #{number} diff (base..HEAD). Return [CRITICAL/IMPORTANT/MINOR] file:line - description or NO ISSUES. |
| 2 | general-purpose | haiku | Check CI status (gh pr checks), CodeQL alerts (gh api code-scanning/alerts), lint results for PR #{number}. Return [CI_FAIL/CODEQL/LINT] check - description or NO ISSUES. |
| 3 | general-purpose | haiku | Check ALL comments on PR #{number} from ALL sources and ALL authors (human and bot alike). Query all three endpoints: (1) gh api repos/{owner}/{repo}/pulls/{number}/comments — inline review comments, (2) gh api repos/{owner}/{repo}/pulls/{number}/reviews — review submissions, (3) gh api repos/{owner}/{repo}/issues/{number}/comments — general comments. Every comment is potentially actionable. Report each as [COMMENT] author: summary (source) or NO ISSUES if all three endpoints return nothing actionable. |
Fix Phase (1 subagent, sequential after all reviews complete)
| Subagent Type | Model | Prompt essence |
|---|
general-purpose | default | Fix ALL issues below on branch {branch}. Prioritize CRITICAL > IMPORTANT > MINOR. After fixing, verify with: npm test, npx tsc --noEmit, and ./lint.sh. All three must pass before committing. Commit and push. Report what was fixed and what remains. {combined_issues} |
Controller Rules
You are the controller. You MUST NOT:
- Read source code, edit files, run tests/lint/build, make commits, or fix anything yourself
- ALL work goes through subagents. No exceptions, not even for "one small thing."
You MUST:
- Dispatch all 3 review subagents in parallel (single message, 3 Task calls)
- Dispatch fixer only after collecting ALL review results
- Track iteration count and report progress between iterations
- Stop at max iterations (default 10) and report remaining issues
Between iterations, report:
Iteration {n}/{max}: {summary}
- Code review: {n} issues
- CI/CodeQL: {n} issues
- Comments: {n} items
Red Flags
| Thought | Reality |
|---|
| "Let me just quickly read/fix this" | NO. Dispatch a subagent. You are the controller. |
| "One review source is enough" | NO. Always dispatch all 3 in parallel. |
| "The fixer said it's done, skip re-review" | NO. Loop back. Fresh eyes every time. |
| "3 iterations is enough" | Default is 10. Don't reduce without user consent. |
| "I'll skip comments, no one reviewed yet" | NO. Always check. Comments may appear mid-loop. |
| "Reviewer returned junk, I'll interpret it" | NO. Dispatch the reviewer again. |
| "It's just a bot comment, not actionable" | NO. Bot comments ARE actionable. Coverage, lint, security — all of it. |
