refine

Iron Law

A fix without a review is not a fix.

Red Flags

• "I fixed the wording but the meaning is basically the same" → STOP, because refinement must preserve the author's intent — changing meaning during polish is an unauthorized rewrite.
• "The score went up so the fix must be good" → STOP, because you must compare verdicts between iterations, not just scores — a higher score with new Critical findings is a regression.
• "Let me do one more pass, I can make it even better" → STOP, because the plateau rule (same verdict for 2 consecutive iterations) exists to prevent infinite loops — stop when improvement stalls.
• "I'll skip the completion gate, the last review was already APPROVED" → STOP, because the final /second-claude-code:review --preset quick check is mandatory — never declare done without it.
• "The reviewers disagreed so I'll just pick the feedback I like" → STOP, because reviewer consensus drives refinement priority — Critical and Major findings from any reviewer must be addressed before Minor ones.

Refine

Run review-fix cycles until a draft meets a target score or verdict, with resumable state on disk.

When to Use

• A draft exists but is not good enough yet
• The user wants a specific quality bar
• Another skill produced output that needs iterative refinement

Workflow

1. Read the current file and record a baseline hash. If --dod is provided, parse the semicolon-separated string into a numbered criteria list and store in state as dod_criteria.
2. Run /second-claude-code:review — this MUST dispatch actual subagents per the review skill spec. Do NOT simulate review inline or merge reviewer perspectives into one pass. When --dod is active, inject the following block into each reviewer's context alongside any --promise text:

   ## DoD Checklist — evaluate each criterion independently
   1. [criterion text]
   2. [criterion text]
   ...
   For EACH criterion, append a line: `DoD-N: PASS` or `DoD-N: FAIL — <reason>`
   

   After reviewers return, extract DoD-N: PASS/FAIL lines and compute per-criterion consensus (majority across reviewers). Store results in state dod_results.
3. Apply fixes. When --dod is active, prioritize FAIL DoD criteria first (up to 3 total including general feedback). If all DoD criteria already pass, fall back to the standard top-3 general feedback.
4. Re-run /second-claude-code:review (with DoD checklist re-injected) and keep the new baseline only if the verdict improves; otherwise revert. If mmbridge was used in the original review, run mmbridge resume before dispatching the full re-review to get the external reviewer's preliminary assessment of fixes (see MMBridge Refinement Enhancement below). Revert strategy:
   • Before reverting a git-tracked file: run git diff --name-only <file>. If the file shows uncommitted user changes that are NOT from this refine iteration (i.e., changes that predate baseline_hash), warn the user and abort the revert unless they explicitly confirm. Never silently overwrite uncommitted work.
   • Path validation: confirm the file path resolves within the project root — reject any path containing ../ traversal or resolving outside the working directory.
   • For git-tracked files with no external uncommitted changes: use git checkout -- <file>.
   • For non-git files (e.g., in ${CLAUDE_PLUGIN_DATA}): restore from baseline_content in refine-active.json.
5. Stop when the target is met, --max is reached, or the verdict plateaus (same verdict for 2 consecutive iterations with no severity reduction). When --dod is active, the target is only considered met when all DoD criteria pass AND the score/verdict target is satisfied.
6. Completion gate: Before declaring done, run /second-claude-code:review with --preset quick (a parameter passed to /second-claude-code:review, not a refine option) one final time (with DoD checklist if active). Only exit on APPROVED or MINOR FIXES and all DoD criteria PASS. If it returns MUST FIX or NEEDS IMPROVEMENT, or any DoD criterion is FAIL, continue refining.

Options

Flag	Values	Default
--max	1-10	3
--target	score (e.g., 4.5) or verdict (e.g., APPROVED)	APPROVED
--promise	text injected into each reviewer's context as a constraint	none
--dod	semicolon-separated success criteria checklist (e.g. "no factual errors; every section has examples")	none
--file	path to the draft to iterate on	required
--review	path to initial review report (skips first review cycle)	none

When called from a pipeline with input_from, the first file is the draft (--file) and the second (if present) is the initial review report (--review).

State

Data directory: ${CLAUDE_PLUGIN_DATA} is set by the plugin system. If unavailable, fall back to .data/ relative to the plugin root. Before writing state files, verify the directory exists with mkdir -p.

Save active state to ${CLAUDE_PLUGIN_DATA}/state/refine-active.json with:

{"goal":"...","file":"...","current_iteration":2,"max":3,"verdicts":["NEEDS IMPROVEMENT","APPROVED"],"scores":[],"baseline_hash":"...","baseline_content":null,"is_git_tracked":true,"feedback_log":[],"dod_criteria":["no factual errors","every section has examples"],"dod_results":[[true,false],[true,true]]}

• dod_criteria: array of criterion strings parsed from --dod. Empty array when --dod is not used.

• dod_results: array of arrays — one boolean array per iteration, positionally matching dod_criteria. true = PASS (majority of reviewers agreed), false = FAIL.

• baseline_content: stores the full file content when is_git_tracked is false (non-git files cannot be reverted via git checkout)

• is_git_tracked: set at step 1 by running git ls-files --error-unmatch <file>. Determines the revert strategy.

MMBridge Refinement Enhancement

When mmbridge is detected (see references/mmbridge-integration.md) and the previous review cycle used --external, the refine skill can leverage mmbridge's session continuity for smarter iterations.

Followup — Clarifying Review Findings

When a review finding is ambiguous or the editor needs more context before fixing:

mmbridge followup --tool kimi --prompt "<specific question about a finding>" --latest --json > /tmp/mmbridge-followup-${RUN_ID}.json

• --latest: reuses the most recent review session for this project
• Use this BEFORE applying fixes when a Critical or Major finding's intent is unclear
• Parse the followup JSON response and provide it to the editor alongside the original finding

Resume — Re-review After Fixes

After the editor applies fixes (Step 3), before dispatching a full internal re-review (Step 4):

mmbridge resume --action followup -y --json > /tmp/mmbridge-resume-${RUN_ID}.json

• This asks the original external reviewer to evaluate the fixes against its earlier findings
• The resume result is merged as supplemental context into the next /second-claude-code:review cycle
• If resume indicates all external findings are addressed, it counts as a positive signal for the consensus gate

When to Use

• Followup: when a review finding (especially from [external: mmbridge]) lacks actionable detail
• Resume: at every re-review cycle if mmbridge was used in the original review
• Skip both: if the original review did not use --external, or if mmbridge is not installed

Cost Note

Followup and resume reuse existing mmbridge sessions — they are cheaper than dispatching a full new review.

Output

Return the final draft plus an iteration log showing verdict progression and major changes. When --dod is active, append DoD status per iteration:

Round 0 (baseline):  2.0/5  ||||..............  MUST FIX
  DoD: [x] no factual errors  [ ] every section has examples

Round 1 (post-edit): 3.8/5  |||||||||||||||...  MINOR FIXES  (+1.8)
  DoD: [x] no factual errors  [x] every section has examples  ✓ ALL PASS

Gotchas

• Do not claim improvement without comparing verdicts between iterations.
• Revert through git checkout -- <file> for git-tracked files, or restore from baseline_content for non-git files. Never rely on memory alone. Always check git diff --name-only <file> before reverting — abort with a warning if uncommitted user changes are detected outside this refine pass's own edits.
• Never revert a file whose path contains ../ or resolves outside the project root.
• Do not simulate review inline — always dispatch through /second-claude-code:review.
• Stop early if the verdict stops improving across iterations.
• The completion gate is mandatory — never skip the final /second-claude-code:review --preset quick check.

Subagents

reviewer: { skill: /second-claude-code:review, constraint: "return score plus ranked feedback" }
editor: { model: opus, tools: [Read, Edit], constraint: "apply only the top 3 feedback items" }