Commit Tidy

Topics

Analyze staged/unstaged changes and recommend whether to split into multiple commits.

When to use

• Before committing large changesets
• User asks "should I split this commit?"
• Reviewing changes that touch many files
• Ensuring atomic, reviewable commits

Mandatory Invocation Gate (HARD STOP)

• 5+ Files / Large-Scale Modification Gate: Whenever 5+ files are modified/added/deleted, or a broad cross-directory refactoring, skill renaming, or multi-component edit occurs, invoking commit-tidy before committing is MANDATORY (HARD STOP). Monolithic single-commit attempts without a commit-tidy split & staging review are strictly prohibited.

• Repeated Red→Green Cycle Gate: Whenever a branch accumulates two or more Red→Green pairs, invoking commit-tidy before the first push of that branch is MANDATORY (HARD STOP), regardless of how few files the cycles touch. The 5+ files gate above does not cover this case — TDD cycles typically rewrite one file repeatedly, so a three-cycle branch can reach seven commits while never tripping a file-count threshold.

  Two consequences make the range-level judgement necessary even when each cycle felt self-contained:

  Consequence	Why it needs the range, not the cycle
  Repeated-single-file squash candidates	"Squash-scan scope" below scans the range grouped by file. A branch where every commit touches the same file is precisely that scan's target, and no single cycle can see the pattern
  Push-limit collision	A repo may cap commits per push (.githooks/pre-push, PUSH_MAX_COMMITS). Seven commits against a limit of five forces either a staged push or an override — both decisions belong before the push, not after

  The verdict may well be "keep them all" — different Conventional Commit types (test / feat / docs) are disqualified from squashing by "Don't squash" below, and when each Red encodes a different discovery (especially one found by executing the code after an earlier Green) squashing would fabricate a history in which every test predates every implementation. The requirement is that the judgement was made, not that it collapses anything.

  #	Don't	Do
  1	Treat each Red→Green pair as its own unit task, satisfy the per-task commit obligation, and push the accumulated branch without a range-level pass	From the second Green onward, run commit-tidy over origin/<base>..HEAD before the first push
  2	Invoke commit-tidy only after the branch is already pushed	A squash then requires a force-push, which narrows the options to ones needing CI checks and user approval. Order is the whole point
  3	Skip the gate because the cycles touch a single file	Single-file is the trigger, not an exemption — see the table above

Squash-scan scope (HARD STOP)

A user-named commit range is the minimum scope, never the maximum. The moment any squash candidate is found — whether self-discovered or pointed at by the user — scan the entire unpushed range (git log --name-only @{u}..HEAD) grouped by file for the same repeated-single-file pattern before proposing a squash plan. See staging-discipline.md "Full-range squash-candidate scan" for the procedure. Presenting a squash plan for only the range the user mentioned, while an identical streak sits elsewhere in the same unpushed history, is a violation — the assistant surfaces the full picture, not just the part the user already knew about.

Conflict-residue commits (HARD STOP)

Before proposing any squash, reword, reorder, or push plan, grep the range's commit messages for conflict residue (git log --oneline <base>..<branch> | grep -iE 'conflict'). Any hit is disqualified from every fast path in this skill until its diff has been reviewed hunk by hunk, with an ask per trade-off group. A clean git status and a marker scan that finds no <<<<<<< prove nothing here — they detect abandoned resolutions, while the dangerous case is one that was completed with the wrong side kept. See conflict-commit-review.md for the scoping, evidence, grouping, and verification procedure.

Split Decision Criteria

Split when

1. Unrelated functionality changes

   • Feature A + Bug fix B → 2 commits
   • UI change + API change (if independent) → 2 commits

2. Wide file spread

   • Changes span 5+ directories with no common purpose
   • Frontend + Backend + Config all modified

3. Mixed change types

   • Refactoring + New feature → 2 commits
   • Formatting + Logic change → 2 commits
   • Dependency update + Code change → 2 commits

4. Large diff size

   • 500+ lines changed across unrelated areas
   • Multiple components modified independently

Reset & Rollback Commit Tidy Rule (HARD STOP)

• Mandatory Amend on Rollback: When recovering initial configurations or adding files after a repository reset/rollback to the Initial Commit state, do not accumulate new commits; always merge changes into the existing initialization commit via git commit --amend.
• Clean Single Initial Commit: Organize initial repository setups (CLAUDE.md, index.md, log.md, pages/) into a single clean, self-contained initial commit to maximize history readability.

5. Different reviewers needed
   • Changes require different domain expertise
   • Security-sensitive + general changes

Keep together when

1. Single logical change

   • Feature requires touching multiple files
   • Refactoring that must be atomic

2. Dependent changes

   • API change + caller updates
   • Schema change + migration + model update

3. Related cleanup

   • Feature + directly related tests
   • Bug fix + regression test

Squash Criteria

When analyzing multiple commits, recommend squashing as well as splitting.

Squash when

1. Same type + same purpose

   • test: A test + test: B test (tests for the same feature) → squash into 1
   • fix: typo A + fix: typo B (same review feedback) → squash into 1

2. Commits split per loop by automated agents

   • Autonomous agents like Ralph commit per loop → squash if same purpose
   • Example: proxy test in loop 1, OIDC test in loop 2 → test: add unit tests

3. Consecutive WIP commits

   • wip: in progress + feat: complete → squash into one feat

Don't squash

1. Commits with different types — keep test + chore + feat separate
2. Commits belonging to different PRs/issues
3. Independent changes that may need to be reverted

Output format (when recommending squash)

### Recommendation: Squash 2 commits → 1

**Before** (2 commits):
- 441b966a test(dt): OIDC auth, proxy, SSO tests
- e2b6503a test(dt): OIDC route tests (login, callback, me)

**After** (1 commit):
Subject: test(dt): add OIDC auth unit tests
Body:
  Consolidates OIDC unit tests from the prior per-loop splits — covers the
  auth flow, proxy interaction, SSO behavior, and route handlers
  (login / callback / me) in a single coherent test commit.

**Reasoning**: Same type (test), same feature (OIDC auth), agent loop split

The body in every recommended commit follows message-discipline.md "Default commit message structure" — body is recommended by default, free-form (not restricted to per-file enumeration), footer optional.

Instructions

Step 0: Determine scope

When ARGUMENTS specify a range (e.g., "since main", "last 3 commits", "PR #N"), analyze all changes in that range — both committed and uncommitted.

# Range specified (e.g., "X changes since main")
git log --oneline <base>..HEAD -- <path>     # committed changes
git diff <base>..HEAD --stat -- <path>        # committed diff
git diff HEAD --stat -- <path>                # uncommitted diff

The analysis must cover committed commits (squash/split candidates) + uncommitted changes (new commit candidates) as a single unified view. Do not analyze only uncommitted changes when a range is specified.

When no range is specified, default to staged + unstaged changes only.

Step 1: Analyze changes

# Check staged changes
git diff --cached --stat
git diff --cached --name-only

# Check unstaged changes
git diff --stat
git status

Step 2: Categorize files

Group changed files by:

• Feature/Component: Which feature does this belong to?
• Change type: feat, fix, refactor, style, test, docs, chore
• Directory: Are changes localized or spread out?

Step 3: Identify boundaries

Look for natural split points:

• Different conventional commit types
• Independent functionality
• Separate test files from implementation (if tests are for different features)

Step 4: Recommend split strategy

Provide specific recommendations. Every recommended commit includes a body by default (see message-discipline.md "Default commit message structure"). The body is free-form — it does not have to enumerate per-file changes.

Mandatory Interactive Ask Gate (HARD STOP): Autonomous commit execution or git push without user confirmation is STRICTLY FORBIDDEN. After presenting the split/squash recommendation, you MUST present the split options to the user via AskUserQuestion to obtain explicit approval before executing any git commit or git push. Even when git pull --rebase is required or executed, automatically running git push afterwards is strictly forbidden without a new AskUserQuestion confirmation.

## Analysis Results

### Changed Files (N files)
- src/api/... (3 files) - API endpoints
- src/components/... (2 files) - UI components
- tests/... (2 files) - Tests

### Recommendation: Split into N commits

**Commit 1**:
  Subject: feat: add user profile API
  Body:
    Adds POST /users/profile and GET /users/profile/:id endpoints backed by a
    shared validation schema. Both endpoints reuse the existing auth middleware
    and return consistent error shapes. Unit tests cover happy path plus
    validation-error branches.
  Footer (optional): Closes #<issue>
  Files:
    - src/api/user.ts
    - src/api/types.ts
    - tests/api/user.test.ts

**Commit 2**:
  Subject: feat: add profile UI component
  Body:
    Adds a Profile component that consumes the new API endpoints, including
    loading and error states. CSS extracted into a sibling module to keep the
    component file focused on behavior. Component tests stub the API client to
    exercise the loading / success / error branches independently.
  Files:
    - src/components/Profile.tsx
    - src/components/Profile.css
    - tests/components/Profile.test.tsx

### Reasoning
- API and UI can function independently
- Each can be reviewed by different reviewers

Step 4.5: Local commit-type rule self-check (HARD STOP — before Step 5)

Before executing the recommended split, run the working-tree-specific commit-type override self-check. Workspaces with <repo>/.claude/rules/*.md defining their own commit-type / version-bump mapping take precedence over the global tag-selection defaults applied in Step 4.

REPO_ROOT="$(git rev-parse --show-toplevel)"
find "$REPO_ROOT/.claude/rules/" -name '*.md' 2>/dev/null

If the find returns one or more files, grep them for commit-type semantics (recursive to match the find scope above):

grep -rliE 'commit type|conventional commit|version bump|feat|fix|topic' \
  "$REPO_ROOT/.claude/rules/" --include='*.md' 2>/dev/null

Read each matched file. If a local mapping exists (e.g., "new topic = feat:, in-place edit of existing topic = fix:"), re-classify each recommended subject under the local mapping before invoking Step 5. Example reclassification:

See message-discipline.md → "Working-tree-specific commit-type override (HARD STOP)" for the full self-check.

Step 4.6: Prior-merge regression check (HARD STOP — before presenting the split as ready)

Before declaring a split "confirmed / ready to execute", check whether any staged file's skill was already the subject of a merged PR — especially a squash-merged one. A squash-merge collapses individual commits (including "address review findings" fix commits) into one commit on the base branch. Staged changes that touch the same files can silently regress a bug that a findings-application commit already fixed, and a structural/pattern-scan diff will not reveal it.

# has this skill already got a merged PR? (repo-specific owner/repo)
gh pr list -R <owner>/<repo> --state merged --search "<skill-name>" --json number,title,mergedAt,baseRefName

# resolve the actual base (do not assume origin/main — staging repos use next-feat/next-fix)
gh pr view <N> -R <owner>/<repo> --json baseRefName -q '.baseRefName'

# if the PR had multiple commits before merge, list them — findings-application
# commits are recognizable by "review findings" / "internal review" / "address feedback"
gh api repos/<owner>/<repo>/pulls/<N>/commits --jq '.[] | {sha: .sha[0:7], message: .commit.message}'

Diff the staged content against the actual resolved base ref (not an assumed default) — and read the full diff, not a head -N or grep-filtered slice:

git diff origin/<actual-base> -- <file>

For any hunk that overlaps a findings-application commit's touched lines, and for any changed logic that is testable (regex assignment, conditional branch, script behavior), execute the changed line(s) to confirm behavior — a diff read is not a substitute for running the code.

Step 5: Execute split (if requested)

Use HEREDOC (git commit -F -) so the body and optional footer land in the commit message exactly as drafted. Per message-discipline.md, single -m "<subject>" invocations are reserved for the rare subject-only acceptable cases (typo / routine dep bump).

# Unstage all
git reset HEAD

# Stage first commit files
git add src/api/ tests/api/
git commit -F - <<'EOF'
feat: add user profile API

Adds POST /users/profile and GET /users/profile/:id endpoints backed by a
shared validation schema. Both endpoints reuse the existing auth middleware
and return consistent error shapes. Unit tests cover happy path plus
validation-error branches.
EOF

# Stage second commit files
git add src/components/ tests/components/
git commit -F - <<'EOF'
feat: add profile UI component

Adds a Profile component that consumes the new API endpoints, including
loading and error states. CSS extracted into a sibling module to keep the
component file focused on behavior. Component tests stub the API client to
exercise the loading / success / error branches independently.
EOF

Quick Reference

File spread heuristic

Change type combinations to split

Output Format

Analysis results should include:

1. List of changed files with categories
2. Whether split is needed and why
3. Specific commit splitting plan
4. Suggested commit messages for each — subject + body by default (free-form body, footer optional). See message-discipline.md "Default commit message structure". Subject-only is reserved for typo / routine dep bump
5. Per-commit body draft — free-form prose / bullets / sections, NOT restricted to per-file enumeration
6. Execution commands (if requested) — use git commit -F - <<'EOF'…EOF HEREDOC form so the drafted body / footer lands in the commit message verbatim