review-loop

name	review-loop
description	Automated iterative review loop. Runs full review on explicitly specified files, fixes all findings, commits, and repeats until the first clean pass (no CRITICAL/WARNING). Each iteration runs in an isolated sub-agent to avoid context overflow. Usage: /review-loop src/path/file1.ts src/path/file2.ts

Review Loop Skill

Automated iterative code review loop that runs until the first clean pass (no CRITICAL/WARNING findings). Each iteration is isolated in a sub-agent to prevent context overflow.

Usage

/review-loop src/hooks/useFileOperations.ts src/hooks/useMarqueeSelection.ts

Files must be explicitly provided as arguments. Never auto-select files from CODE_REVIEW.md.

Workflow

SETUP

1. Parse file list from arguments (space-separated paths after /review-loop)
2. If no files provided, ask the user once: "Which files should I review? Please provide the file paths."
3. Note the working directory (CWD)
4. Set iteration = 1

LOOP

Spawn a general-purpose sub-agent (via the Agent tool) for each iteration with the following prompt template:

---

Sub-agent prompt template:

You are performing iteration {N} of an automated code review loop for OwnChart.
Working directory: {CWD}

STEP 1 — REVIEW:
Use the review skill. As in CODE_REVIEW.md mentioned, review the following files,
regardless of whether they are already marked as reviewed or not:
{FILES_LIST}

Invoke the /review skill for the files listed above. Wait for the full review report.

STEP 2 — FIX:
Fix ALL findings from STEP 1 (CRITICAL + WARNING + NOTE).
Additionally fix everything a senior developer would fix for future-proofing and
robustness — even if not flagged by the checklist.
Do NOT ask for user confirmation. Apply all fixes directly.

STEP 3 — VERIFY:
Run: npm run ci:local
If it fails, fix all issues before proceeding. Keep fixing until ci:local passes.

STEP 4 — COMMIT:
Commit all changes using Conventional Commits format.
Use "fix(review):" prefix if fixing bugs/issues, or "refactor(review):" if restructuring.
Example: git commit -m "fix(review): address review findings in useFileOperations"

STEP 5 — COUNT AND RETURN:
Count the findings from the STEP 1 review report:
- critical = number of CRITICAL findings
- warning = number of WARNING findings
- notes = number of NOTE findings

End your response with this EXACT line (no other text after it):
REVIEW_RESULT: critical=N warning=N notes=N

Replace N with the actual counts from the review report.

---

AFTER EACH SUB-AGENT RETURNS

1. Find and parse the REVIEW_RESULT: line from the sub-agent's response
2. Extract critical, warning, notes counts

If critical > 0 OR warning > 0:

• Print: "Iteration {N} complete: {critical} critical, {warning} warning, {notes} notes fixed. Starting iteration {N+1}..."
• Increment iteration
• Spawn a new sub-agent (GOTO LOOP)

If critical == 0 AND warning == 0:

• Update CODE_REVIEW.md: find each file in the list and check off its review checkbox (change [ ] to [x])
• Print final summary:

  Loop complete after {N} iteration(s). Files are clean (0 critical, 0 warning, {notes} notes).
  Files reviewed:
  - {file1}
  - {file2}
  ...
  CODE_REVIEW.md updated.
  

• STOP

Key Constraints

• Files from args only — never read CODE_REVIEW.md to select files
• Sub-agents are general-purpose with full tool access (so they can invoke the /review skill)
• npm run ci:local must pass before every commit
• Conventional Commits required — fix: → PATCH bump, refactor: → no release
• CODE_REVIEW.md updated only on final iteration by the main session (not sub-agents)
• No user prompts during the loop — fully autonomous after initial file confirmation

Error Handling

• If sub-agent doesn't return a REVIEW_RESULT: line, print a warning and ask Martin whether to retry or stop
• If npm run ci:local keeps failing inside a sub-agent, the sub-agent should report the error and stop — the main session will surface this to Martin
• Maximum recommended iterations: 10 (if still failing after 10, stop and report)