Review Fix Loop

Act as the parent orchestrator for a review-and-fix loop. Keep the user informed in simple, clear language while subagents do the review and fix work.

Workflow

1. Identify the target change set: local uncommitted changes, staged changes, a branch diff, or a PR. If the target is ambiguous, ask one concise clarifying question.
2. Spawn one reviewer subagent with reasoning_effort: "xhigh" and no model override. Use the default agent unless the current tool surface requires another general coding agent type.
3. Give the reviewer the target change set and the available $code-review skill. Tell it to review only and return the code-review JSON output, with no fixes.
4. Wait for the review result. If there are no findings, stop the loop and report that no actionable findings remain.
5. If findings exist, summarize them to the user before triggering fixes. Use plain language, grouped by finding, with file path, priority, risk, and the intended fix direction.
6. Send the findings back to the same reviewer subagent and instruct it to fix only those issues. The same agent that found the issues owns the fix pass.
7. Review and integrate the fix agent's changed files into the parent workspace when the subagent works in a forked workspace. Do not overwrite unrelated user or agent changes.
8. Run the relevant verification after fixes. Use project instructions for commands and choose the narrowest relevant checks.
9. Spawn a fresh xhigh reviewer subagent for the next review pass. Do not reuse the fixing reviewer for the next independent review.
10. Repeat until a fresh reviewer reports no actionable findings.

Subagent Instructions