review-fix-loop

# Review-Fix Loop Spawn a reviewer subagent, fix findings, re-review until clean. Max 5 rounds. ## Process ```dot digraph review_fix_loop { "Spawn reviewer" [shape=box]; "Findings?" [shape=diamond]; "Fix all findings" [shape=box]; "Max rounds?" [shape=diamond]; "Done — clean" [shape=doublecircle]; "Done — cap hit" [shape=doublecircle]; "Spawn reviewer" -> "Findings?"; "Findings?" -> "Done — clean" [label="none"]; "Findings?" -> "Fix all findings" [label="yes"]; "Fix all findings" -> "Max rounds?" ; "Max rounds?" -> "Spawn reviewer" [label="no"]; "Max rounds?" -> "Done — cap hit" [label="5 rounds"]; } ``` ### 1. First review Spawn a reviewer agent with: - Target file(s) and what to check - Instruction to verify against actual codebase (not just the file in isolation) - `run_in_background: false` (need results to proceed) ### 2. Fix Read findings. Apply all fixes to the file(s). No partial fixes — address every finding before re-reviewing. ### 3. Re-review Spawn reviewer again with: - Same target file(s) - List of what was fixed from previous round (so reviewer verifies fixes landed AND checks for new issues) - Same instruction to verify against codebase ### 4. Repeat or exit - **Clean (no findings):** Done. Report to user. - **Findings remain:** Fix and re-review (go to step 2). - **5 rounds reached:** Stop. Report remaining issues to user — something may need human judgment. ## Reviewer prompt template ``` [Review/Re-review] this [file type]: [path] [First round: describe what to check] [Re-review: list previous findings and which were fixed. Ask reviewer to verify fixes AND find new issues.] Check against actual codebase, not just the file text. [If clean, say "Clean — no issues."] ``` ## Key behaviors - **Always run reviewer synchronously** (`run_in_background: false`) — need findings before fixing. - **Fix ALL findings before re-review** — partial fixes waste review rounds. - **Pass previous findings to re-reviewer** — reviewer needs context to verify fixes and avoid re-reporting. - **Report round count** — tell user "Clean after N rounds" or "Stopped at 5 rounds, N issues remain."