smart-commit（智能提交）

Triggers: When the user says "commit", "smart commit", or /commit.

When to Use

Triggers: When the user says "commit", "smart commit", or /commit.

• Committing staged changes in any git repository
• When you want a well-structured conventional commit message
• When you want automatic detection of common commit issues before pushing

Process

Step 1 — Read working-tree state

Run these commands and collect their output:

git status --porcelain          # full working-tree state: staged, unstaged, untracked
git diff --cached               # full diff for content analysis (staged changes)

Also check:

git log --oneline -5            # recent commits to match message style

Parse the git status --porcelain output and assign a staging-status tag to each file:

• Lines where XY = "??" → tag untracked
• Lines where X is A, M, R, D, or C (index column non-space) → tag staged

- If X = R (rename), split the entry on " -> " and include both the old path and the new path; both receive tag staged

• Lines where X = " " and Y is M, D, or T (worktree column non-space) → tag unstaged
• When both X and Y are non-space (staged change also has worktree modifications), the index change takes precedence: tag staged

If git status --porcelain returns empty output, report:

Nothing to commit. Working tree is clean.

And stop — do not proceed.

Step 1b — Group detected files

Using the file paths collected in Step 1, apply the following grouping heuristic in priority order. Each detected file is assigned to exactly one group — the first rule that matches wins, and the file is not reconsidered by later rules. Each file's staging-status tag (assigned in Step 1) travels unchanged through the grouping step and is preserved in the group record.

Rule 1 — Test files (highest priority, applied before any directory check):

• Paths matching *.test.*, *.spec.*, _test.*, or *_test.* → group test
• This rule applies regardless of which directory the file lives in.

Rule 2 — Config/infra files:

• Root-level files (no subdirectory) matching *.json, *.yaml, *.yml, *.toml, *.sh, or *.env* → group chore

Rule 3 — Docs files:

• Paths under docs/ → group docs
• Root-level files matching *.md or README* → group docs

Rule 4 — Directory prefix:

• Remaining files are grouped by their first path segment (e.g., skills/smart-commit/SKILL.md → group skills; hooks/smart-commit-context.js → group hooks)

Fallback:

• Files that match none of the rules above (e.g., a root-level foo.rb with no subdirectory and no recognized extension) → group misc

No detected file may appear in more than one group. Every detected file must appear in exactly one group.

Single-group fast-path: If grouping produces exactly one group → skip the multi-commit plan and fall through to Step 2 unchanged. Behavior is identical to the pre-grouping version of this skill. If the single group contains any files tagged unstaged or untracked, print Auto-staging N file(s): <list of those files> then issue git add <unstaged/untracked files> before proceeding to Step 2.

Multi-group branch: If grouping produces two or more groups → proceed to Step 1c (multi-commit plan) before executing any commit.

Step 1c — Present multi-commit plan

For each group identified in Step 1b, run Steps 2 and 3 independently using a scoped diff limited to that group's files:

git diff --cached -- <file-1> <file-2> ...   # scoped to this group's file list only

Use this scoped diff (not the full git diff --cached) as the input for both the commit message generation (Step 2) and the issue detection (Step 3) for that group.

ERROR collection before display: Collect all ERROR-severity findings across every group before printing anything to the user. If any group yields one or more ERRORs:

• Print all ERRORs found across all groups.
• Print: Commit plan blocked. Fix the errors above before committing.
• Stop — do not display the plan, do not proceed to any git commit.

Display the full plan (only when no ERRORs are present):

This annotation MUST appear before any git add or git commit is issued, so the user can review the exact staging-status of every file before confirming.

## Smart Commit Plan (N commits)

────────────────────────────────────
Commit 1 of N — [label]
Files: file-a.md [staged], file-b.md [unstaged]
Proposed message:
  type(scope): short description
────────────────────────────────────
Commit 2 of N — [label]
Files: skills/smart-commit/SKILL.md [untracked]
Proposed message:
  type(scope): short description
────────────────────────────────────

Issues detected across all groups:
  ⚠️  WARNING in commit 1: <message>

Each file in the Files: line is annotated with its staging-status marker: [staged], [unstaged], or [untracked].

Display any WARNING-severity findings in the plan summary (non-blocking). If there are no issues, omit the "Issues detected" section.

Plan-level confirmation prompt:

Proceed with all commits? [commit all / step-by-step / abort]

• commit all → proceed to Step 5, multi-group "commit all" path
• step-by-step → proceed to Step 5, multi-group "step-by-step" path
• abort → print Commit plan aborted. No files were staged. Working tree is unchanged. and stop; no git add or git commit is executed

Step 2 — Analyze changes

From the diff output, determine:

Step 3 — Detect issues

Scan git diff --cached output for these patterns. Report as ERROR (blocking) or WARNING (non-blocking):

Step 4 — Present summary

Output a structured report before committing:

## Smart Commit Summary

**Staged:** N files  +X −Y lines

**Proposed commit message:**
──────────────────────────────
type(scope): short description

- bullet change 1
- bullet change 2
──────────────────────────────

**Issues detected:**
⛔ ERROR: .env.local is staged — unstage with: git restore --staged .env.local
⚠️  WARNING: console.log in pages/api/auth.js (2 occurrences)
⚠️  WARNING: TODO: left in domain/payment/PaymentService.js

Then, based on severity:

• ERRORs present → Print: Commit blocked. Fix the errors above before committing. Do NOT proceed.
• Warnings only → Ask: Proceed with commit? [y / edit message / abort]
• No issues → Ask: Proceed with commit? [y / edit message / abort]

If the user chooses edit message: ask them to provide the new message, then use it verbatim. If the user chooses abort: stop, print Commit aborted.

Step 5 — Execute commit

After confirmation, execute:

git commit -m "$(cat <<'EOF'
type(scope): short description

- bullet change 1
- bullet change 2
EOF
)"

On success, report:

✅ Committed: <hash> — type(scope): short description

On failure, report the full git error and suggest remediation.

Multi-group "commit all" path (when user chose commit all at Step 1c):

Iterate through every group in the order shown in the plan. For each group: 0. For each file in this group tagged unstaged or untracked, issue git add <file>; skip files already tagged staged.

1. Execute git commit with the group's proposed message.
2. Print the resulting commit hash immediately after each commit fires.
3. Do not present any intermediate prompt between groups.

After all groups have been committed, print the full-execution summary (see below).

Multi-group "step-by-step" path (when user chose step-by-step at Step 1c):

Before each group's commit, print the per-commit confirmation block:

## Commit N of M — [label]

**Files:** file-a.md, file-b.md
**Proposed message:**
──────────────────────────────
type(scope): short description
- bullet 1
──────────────────────────────

Proceed? [y / edit message / skip / abort remaining]

Apply the user's choice before moving to the next group:

• y → for each file in this group tagged unstaged or untracked, issue git add <file>; skip files already tagged staged; then execute git commit with the proposed message; print the resulting hash; continue to the next group.
• edit message → prompt: Enter replacement message: — use the provided text verbatim for this commit only; for each file in this group tagged unstaged or untracked, issue git add <file>; skip files already tagged staged; then execute and print the hash; continue to the next group.
• skip → no git add is issued for this group; leave this group's files in their original state without committing; move to the next group.
• abort remaining → no git add is issued for remaining groups; stop processing after the current group (do not commit any subsequent groups); print the partial-execution summary.

Partial-execution summary (triggered when all groups complete or when abort remaining is chosen mid-sequence):

When all groups succeeded (or were skipped without aborting):

N of N commits executed.
  ✅ <hash-1> — type(scope): description (commit 1)
  ✅ <hash-2> — type(scope): description (commit 2)
  ...

When execution was aborted mid-sequence (M commits executed, remaining groups not committed):

M of N commits executed.
  ✅ <hash-1> — type(scope): description (commit 1)
  ✅ <hash-2> — type(scope): description (commit 2)

Not committed — files remain in original state:
  Commit <next> of N — [label]
    Files: file-x.md [staged], file-y.md [unstaged]
    Proposed message: type(scope): description
  Commit <next+1> of N — [label]
    Files: file-z.md [untracked]
    Proposed message: type(scope): description

Do NOT attempt to undo or roll back any commits that have already been executed.

Anti-Patterns

Don't generate vague messages

# Bad
git commit -m "update"
git commit -m "fix stuff"
git commit -m "changes"

# Good
git commit -m "fix(auth): prevent session expiry on page reload"
git commit -m "feat(player): add shuffle mode to playlist"

Don't bypass ERRORs

If a .env file or credential pattern is detected, stop unconditionally. Do not commit even if the user asks to skip.

Don't stage files that were not confirmed

Issue git add only for groups the user explicitly confirmed; never batch-stage all detected files upfront before the plan confirmation.

Quick Reference

Rules

• Never commit without first presenting the generated message summary and waiting for explicit user confirmation
• Detect and block commits that contain secrets (API keys, tokens, passwords) found by pattern matching in the diff
• Detect and warn on console.log, debugger, TODO, and large binary files in the staged diff before committing
• Commit messages must follow conventional commits format (type(scope): description); generated messages that cannot be classified must default to chore: with a descriptive subject
• Detect files from the full working tree via git status --porcelain; assign each file a staging-status tag (staged, unstaged, untracked) during Step 1
• Auto-stage only the files of a confirmed group — issue git add <files> for files tagged unstaged or untracked immediately before that group's git commit; never re-add already-staged files
• For skipped or aborted groups, issue no git add — their files must remain in the exact state they were in when the skill was invoked
• When detected files span two or more functional groups, present the full multi-commit plan before executing any commit
• ERROR conditions in any group block the entire multi-commit plan — no partial execution allowed when ERRORs are present
• Every commit in a multi-commit sequence uses the proposed message without additional trailers
• The grouping heuristic is applied in priority order: test → config/infra → docs → directory prefix → misc fallback; no file may appear in more than one group
• The single-group fast-path preserves exact backward compatibility — no behavior change when all staged files resolve to one group