interactive-diff-review（交互式差异审查）

Git Diff Review Skill

Interactive hunk-by-hunk code review. Resume → Parse → Review → Verify → Report → Apply → Commit.

Commands

Parse $ARGUMENTS to determine the command and diff target:

Target is optional: <commit>, <commit_a> <commit_b>, or omit for auto-detect (staged → workspace).

Commands can be combined: /interactive-diff-review --apply --commit will apply fixes then commit if verify passes.

Command Routing

1. Parse $ARGUMENTS — extract command flags and remaining positional args (diff target).
2. If no command flag is present → default to review.
3. Route to the corresponding section below.

Workflow Checklist

Copy this checklist and check off items as you complete them:

Diff Review Progress:
- [ ] Step 0: Check for Existing Progress
  - [ ] 0.1 Check if diff-review-progress.md exists
  - [ ] 0.2 If exists, parse header and count pending vs reviewed hunks
  - [ ] 0.3 Ask user: Resume or Restart?
  - [ ] 0.4 If resume → skip to Step 2 (pending hunks only)
  - [ ] 0.5 If restart → delete progress file, continue to Step 1
- [ ] Step 1: Resolve & Parse
  - [ ] 1.1 Run resolve_diff.py to get diff + language
  - [ ] 1.2 Verify no error, inform user of diff source + language
  - [ ] 1.3 Run parse_hunks.py to split into hunks
  - [ ] 1.4 Handle edge cases ⚠️ Load references/edge-cases.md
  - [ ] 1.5 Write diff-review-progress.md ⚠️ Load references/progress-template.md
- [ ] Step 2: Interactive Review ⚠️ CORE LOOP
  - [ ] Load references/review-format.md
  - [ ] 2.1 Display hunk (index/total, file, type, diff body)
  - [ ] 2.2 Provide analysis (summary, impact, risks, suggestions)
  - [ ] 2.3 Collect decision via ask_user_input (Accept / Reject / Skip)
  - [ ] 2.4 If rejected, ask reason (open-ended text follow-up)
  - [ ] 2.5 Update diff-review-progress.md (hunk status, analysis, decision, reviewed counter)
  - [ ] 2.6 Show running tally, advance to next hunk
  - [ ] 2.7 Repeat 2.1–2.6 until all hunks done
  - [ ] 2.8 Loop back for any skipped hunks
- [ ] Step 3: Coverage Verification ⚠️ REQUIRED
  - [ ] 3.1 Re-run resolve_diff.py + parse_hunks.py
  - [ ] 3.2 Compare new hunks against progress file entries (file + header + body)
  - [ ] 3.3 If new/changed hunks → append to progress file, review them (back to Step 2)
  - [ ] 3.4 Confirm full coverage, set Phase to "verified"
- [ ] Step 4: Generate Report
  - [ ] Load references/report-template.md
  - [ ] 4.1 Read diff-review-progress.md as input source
  - [ ] 4.2 Build Markdown report (detected language, diff order)
  - [ ] 4.3 Save to diff-review-report.md in the project root
  - [ ] 4.4 Set Phase to "completed" in progress file
  - [ ] 4.5 Present to user

Step 0: Check for Existing Progress

Before starting any review work, check if a progress file already exists.

0.1 Check for progress file

Look for diff-review-progress.md in the project root.

• If the file does not exist → proceed to Step 1.
• If the file exists → continue to 0.2.

0.2 Parse progress state

Read diff-review-progress.md and extract:

• Phase: current workflow phase
• Reviewed counter: X / N
• Pending hunks: count of hunks with Status: pending
• Diff Source and Target Args: for context

0.3 Ask user

Present the state to the user and ask:

• If there are pending hunks (Phase = reviewing): "Found an existing review in progress (X/N hunks reviewed, diff source:...). Resume or restart?"

- Resume: Skip to Step 2, only iterate through pending hunks. - Restart: Delete progress file, proceed to Step 1.

• If all hunks are reviewed (Phase = verified or completed): "Found a completed review (N/N hunks reviewed). Regenerate the report, or restart a fresh review?"

- Regenerate: Skip to Step 4. - Restart: Delete progress file, proceed to Step 1.

• If Phase = parsing (interrupted during Step 1): "Found an incomplete progress file (parsing was interrupted). Restarting fresh."

- Delete progress file, proceed to Step 1.

0.4 Resume flow

When resuming, re-read the progress file to build the hunk list. Filter to only Status: pending hunks. Enter Step 2 with these hunks. The progress file's existing reviewed hunks are preserved.

0.5 Restart flow

Delete diff-review-progress.md and proceed to Step 1 from scratch.

Step 1: Resolve & Parse

1.1 Run resolver

python scripts/resolve_diff.py                       # auto: staged → workspace
python scripts/resolve_diff.py <commit>              # single commit
python scripts/resolve_diff.py <commit_a> <commit_b> # commit range

Returns JSON: {source, ref, language, diff, error}

1.2 Verify & inform

If error is non-null → report to user and stop. Otherwise tell the user which diff source was resolved and what language was detected.

1.3 Save diff & run parser

Extract the diff field from Step 1.1's JSON output and write it to a temporary file, then parse:

# Write the diff field to a temp file (agent should use Write tool or shell redirect)
# e.g. write resolve output's .diff value to /tmp/diff.txt

python scripts/parse_hunks.py --file /tmp/diff.txt

Returns JSON: {total, hunks: [{index, file, header, body, type, status, reason}]}

1.4 Edge cases

Load references/edge-cases.md and handle accordingly.

1.5 Write progress file

Load references/progress-template.md for the format specification.

Create diff-review-progress.md in the project root with:

• Header metadata (date, diff source, language, target args, total hunks, reviewed = 0/N, phase = reviewing)
• One section per hunk with: type, status = pending, header, empty analysis/decision, diff body

Set Phase to reviewing (parsing is complete, ready for review loop).

Step 2: Interactive Review

Before starting this step, load references/review-format.md for display format, analysis fields, and decision collection details.

All UI text uses the detected language from Step 1. Technical terms remain in English.

Loop through each hunk (or pending hunks only if resuming): display → analyze → collect decision → update progress file → show tally → next.

Progress file updates

After each hunk decision (2.3/2.4), use the Edit tool to update diff-review-progress.md:

1. Update the hunk's Status — change pending to ✅ Accepted or ❌ Rejected.
2. Fill in Analysis — replace the empty Analysis section with the summary, impact, risks, and suggestions.
3. Fill in Decision — replace the empty Decision section with the accept/reject decision and rejection reason if applicable.
4. Update the Reviewed counter — increment the Reviewed: X / N line in the header.

The progress file is the source of truth — if context is lost, the review can resume from it.

After the first pass, loop back for any skipped hunks.

Step 3: Coverage Verification

1. Re-run resolve_diff.py + parse_hunks.py with same arguments as Step 1 (read Target Args from progress file if needed).
2. Read diff-review-progress.md to get the reviewed hunks.
3. Compare new hunks against progress file entries by file + header + body.
4. New/changed hunks found → append them to the progress file as pending, update Total Hunks, then go back to Step 2 for those hunks only.
5. Confirm full coverage. Set Phase to verified in the progress file. Only proceed to Step 4 when all hunks are decided.

Step 4: Generate Report

Before starting this step, load references/report-template.md for the full template.

1. Read diff-review-progress.md as the input source — all hunk data, analysis, and decisions come from the progress file.
2. Build the report in the detected language (technical terms in English), ordered by original diff sequence.
3. Save to diff-review-report.md in the project root and present to user.
4. Set Phase to completed in the progress file.

Command: --verify

Validates that the review is complete and current.

1. Check diff-review-report.md exists in project root. If not → error: "No report found. Run review first."
2. Re-run resolve_diff.py + parse_hunks.py with same target args to get current hunks.
3. Parse the report's Overview table to extract reviewed hunks (file + type + status).
4. Compare:

- Every current hunk has a matching entry in the report (by file path + hunk header + body hash). - No extra entries in report that don't exist in current diff (stale entries).

1. Check that all hunks have status ✅ Accepted (no ❌ Rejected, no missing).
2. Output result:

- PASS: "Verification passed. All N hunks reviewed and accepted. Ready to commit." - FAIL — unreviewed hunks: List the unmatched hunks. - FAIL — rejected hunks: List the rejected hunks with their reasons. - FAIL — stale report: Report doesn't match current diff. Suggest re-running review.

Command: --apply

Fixes code based on review suggestions for rejected hunks.

1. Check diff-review-report.md exists. If not → error.
2. Parse the report to find all ❌ Rejected hunks and their rejection reasons + analysis suggestions.
3. For each rejected hunk (in diff order): a. Read the relevant file and locate the code region from the hunk. b. Based on the rejection reason and the review analysis, generate a fix. c. Apply the fix using Edit tool. d. Show the user what was changed and why.
4. After all fixes applied, inform the user:

- "Applied fixes for N rejected hunks. Run --verify or re-run review to confirm."

Note: --apply does NOT re-run the review. The user should run review again or --verify after applying.

Command: --commit

Generates and executes a git commit. Requires verification to pass first.

1. Run the --verify flow internally (same as above).
2. If verify FAILS → stop and show the failure reason. Do not commit.
3. If verify PASSES: a. Parse the report to build a commit message:

- Subject line: summarize the changes (max 72 chars), in detected language. - Body: list the key changes from the report overview (accepted hunks grouped by file). b. Show the proposed commit message to the user for confirmation. c. On user approval: - git add -A (or stage the specific files from the review). - git commit with the message. d. Show the commit result.

Parameter Reference

Language

• Detected from git log --oneline -10 by resolve_diff.py.
• All UI, analysis, and report use detected language; technical terms stay English.
• Mixed-language commits → most frequent wins.
• No history → fall back to user's conversation language.

File Structure

git-diff-review/
├── SKILL.md                          # This file — workflow orchestration
├── scripts/
│   ├── resolve_diff.py               # Diff source + language detection
│   └── parse_hunks.py                # Unified diff → structured hunks
└── references/
    ├── review-format.md              # Step 2: display, analysis, decisions
    ├── report-template.md            # Step 4: Markdown report template
    ├── progress-template.md          # Steps 1-2: progress document format
    └── edge-cases.md                 # Step 1.4: edge case handling

Generated Files (project root)