review-changes（审查变更）

Review Changes

Orchestrate the full review pipeline: handoff → review → fix. This skill chains three phases with user checkpoints between each.

Phase 1: Create Handoff

Run the full handoff generation inline (do not invoke /create-handoff as a sub-skill — execute the steps directly):

Step 1.1: Determine the session name

1. If the user provided $ARGUMENTS, sanitize it using a shell command — not LLM interpretation — to produce a safe kebab-case string: echo "$ARGUMENTS" | tr '[:upper:]' '[:lower:]' | sed -E 's/[^a-z0-9-]+/-/g; s/^-+//; s/-+$//; s/-{2,}/-/g' Use the output as the session name. If the sanitized result is empty, fall back to option 2.
2. Otherwise, infer from the current branch name (e.g. feature/issue-24-authentication → issue-24-authentication)
3. If on main/master, ask the user what to name the session

Step 1.2: Create the session directory

SESSION_DIR=".chalk/reviews/sessions/${session_name}"
mkdir -p "$SESSION_DIR"

If a handoff file already exists with content, ask the user whether to overwrite or skip to Phase 2.

Step 1.3: Determine the base branch

1. git merge-base main HEAD
2. If that fails, try origin/main, then master, origin/master

Store as {base}.

Step 1.4: Gather context

git log --oneline {base}..HEAD
git diff --stat {base}..HEAD
git diff {base}..HEAD

Step 1.5: Detect and run project checks

Auto-detect the project's build/check tooling and run what's available:

• Node.js (package.json): detect pm from lockfile, run build/typecheck/lint/test
• Rust (Cargo.toml): cargo check, cargo test --no-run, cargo clippy
• Go (go.mod): go build, go vet, go test -short
• Python (pyproject.toml/requirements.txt): py_compile, pytest --co
• Make (Makefile with build/check/test targets): make build/check/test
• If nothing detected, note "No build system detected"

If any check fails, note the failure honestly — do NOT try to fix it.

Step 1.6: Write the handoff

Write to $SESSION_DIR/handoff.md:

# Handoff

## Scope
- Item: {item reference}
- Goal: {1-sentence summary}

## What Changed
{bullet list of logical changes}

## Files Changed
{bullet list from git diff --stat}

## Risk Areas
{bullet list of things that could break}

## Commands Run
{bullet list with pass/fail status}

## Known Gaps
{bullet list of things NOT done}

## Suggested Focus For Reviewers
{bullet list prioritized by risk}

Step 1.7: Report handoff

Show the handoff file path and a brief summary of what was captured.

Phase 2: Review

Ask the user how they want to proceed:

Present these options:

• Self-review — Claude reviews the changes right now in this session
• Copy for external review — Generate a paste-ready prompt to use with another AI (Codex, Gemini, GPT, etc.)

Option A: Self-review

If the user chooses self-review:

1. Read the full diff: git diff {base}..HEAD
2. Read the handoff from Phase 1
3. Review the changes using the same structured format as the universal reviewer template:

## Verdict
- Block merge: yes|no
- Blocking findings: P0=<n>, P1=<n>

## Findings

| ID | Severity | Category | File:Line | Issue | Failure mode | Suggested fix | Confidence |
|---|---|---|---|---|---|---|---|

## Testing Gaps
-

## Open Questions
-

1. Write the findings to $SESSION_DIR/self.findings.md
2. Present the findings summary to the user

Option B: External review

If the user chooses external review:

1. Bootstrap the review pipeline scripts if not already present — invoke /create-review which will create the scripts and templates with their exact, version-controlled content. Do NOT generate the scripts inline.
2. Generate the review pack: bash.chalk/reviews/scripts/pack.sh "{base}" "{session_name}"
3. Generate the prompt: bash.chalk/reviews/scripts/render-prompt.sh "reviewer" "" "" "" "{session_name}"
4. Copy to clipboard if possible
5. Tell the user: "Prompt copied — paste it into your reviewer. When you have the findings, save them to $SESSION_DIR/{reviewer-name}.findings.md and tell me to continue."
6. Wait for the user to come back. When they say they're ready or tell you to continue, proceed to Phase 3.

Phase 3: Fix Findings

Ask the user: "Want me to fix the findings?"

If yes:

Step 3.1: Load findings

Discover all *.findings.md files in the session directory using the Glob tool:

.chalk/reviews/sessions/{session_name}/*.findings.md

Do NOT use shell ls — use the Glob tool for safe discovery.

Step 3.2: Parse and prioritize

Extract the findings table from each file. Sort by severity: P0 > P1 > P2 > P3.

Deduplicate findings that target the same file and line (or lines within 5 of each other) across reviewers.

Step 3.3: Present findings

Show a prioritized summary table and ask the user:

• All — fix P0-P2
• Blocking only — fix P0 and P1
• Let me choose — pick specific IDs

Step 3.4: Apply fixes

For each finding:

1. Validate the file path — must be relative and within the repo. Reject absolute paths or .. traversal
2. Read the file with at least 30 lines of surrounding context
3. Design the fix using the suggested fix as guidance
4. Show the proposed fix and ask for confirmation before applying
5. Apply using Edit tool

Step 3.5: Update resolution log

Write .chalk/reviews/sessions/{session_name}/resolution.md:

# Finding Resolution Log

## Summary
- Session: {session_name}
- Item: {from session name}
- Reviewers: {list of sources}
- Decision owner:

## Findings

| ID | Severity | Source | File:Line | Decision | Notes |
|---|---|---|---|---|---|

## Follow-up Tasks
-

## Final Gate
- Build:
- Tests:
- Ready to merge: yes/no

Step 3.6: Final summary

Show results table, resolution log path, and suggest:

• Run build/tests to verify
• Run /commit to commit the fixes
• Create or update the PR

If no:

Tell the user the findings are saved at their session path and they can run /fix-findings later when ready.

Rules

• Do NOT modify source code during Phase 1 or Phase 2 — only during Phase 3 with user confirmation
• Be honest about failures in the handoff
• Always ask before proceeding to the next phase — never auto-advance
• Keep the self-review objective and rigorous — do not rubber-stamp your own changes
• All file paths from findings must be validated as relative and within the repo
• If multiple reviewers have conflicting suggestions, present both and let the user choose