commit-staged

name	commit-staged
description	Generates and commits conventional-commits messages by analyzing staged git diffs — fast and fork-context safe. Use when the user asks to commit staged changes, needs a type-scope-description message generated from the current diff, or wants scope selection guidance and pre-commit hook handling.
argument-hint	[notes or comments to account for in the commit message]
allowed-tools	Bash(git:*), Read, Glob, Grep, Bash(grep:*), Bash(find:*), Bash(fdfind:*), Bash(prek:*), Bash(uv run prek:*), Bash(uv run pre-commit:*), Bash(pre-commit:*)
model	haiku
context	fork
user-invocable	true

Analyze these staged changes and generate commit message, then commit the changes: !uv run prek run >/dev/null 2>&1 || git add -u !git --no-pager status || true !git --no-pager diff --cached || true !git --no-pager diff --cached --stat || true

<user_notes> $ARGUMENTS </user_notes>

Commit message format

Follow conventional commits format (no footer, except for breaking change footer):

<type>(<scope>): <description>

[optional body]

Types

• feat: New feature
• fix: Bug fix
• docs: Documentation changes
• style: Code style changes (formatting, missing semicolons)
• refactor: Code refactoring
• test: Adding or updating tests
• chore: Maintenance tasks

Examples

Feature commit:

feat(auth): add JWT authentication

Implement JWT-based authentication system with:
- Login endpoint with token generation
- Token validation middleware
- Refresh token support

Bug fix:

fix(api): handle null values in user profile

Prevent crashes when user profile fields are null.
Add null checks before accessing nested properties.

Refactor:

refactor(database): simplify query builder

Extract common query patterns into reusable functions.
Reduce code duplication in database layer.

Commit message guidelines

DO:

• Use imperative mood ("add feature" not "added feature")
• Keep first line under 50 characters
• Capitalize first letter
• No period at end of summary
• Explain WHY not just WHAT in body

DON'T:

• Use vague messages like "update" or "fix stuff"
• Include technical implementation details in summary
• Write paragraphs in summary line
• Use past tense

Multi-file commits

When committing multiple related changes:

refactor(core): restructure authentication module

- Move auth logic from controllers to service layer
- Extract validation into separate validators
- Update tests to use new structure
- Add integration tests for auth flow

Breaking change: Auth service now requires config object

Scope Rules

Scope MUST identify WHERE in the codebase, NOT what type of change.

The scope is a module, component, or directory name - never a description of the change itself.

Determining Scope

1. Single module/directory: Use that module name

   • Changes to src/auth/*.py → auth
   • Changes to plugins/gitlab-skill/ → gitlab-skill

2. Multiple files in same area: Use the common parent

   • Changes to skills/python3-dev/assets/*.py → assets or python3-dev

3. Cross-cutting changes: Use the primary affected area OR omit scope

   • Config + code changes → use primary area: feat(auth): add OAuth support
   • Truly scattered changes → omit: chore: update dependencies across modules

4. Root config files: Use the config type

   • pyproject.toml linting rules → lint or ruff
   • pyproject.toml dependencies → deps
   • .github/workflows/ → ci

Scope Anti-Patterns

NEVER use these as scopes - they describe WHAT, not WHERE:

❌ Wrong	✅ Correct	Why
docs	readme, api-docs, skills	"docs" is a change type, not a location
tests	auth-tests, api	Be specific about what's being tested
types	models, api	Types belong to a module
refactor	(use as type, not scope)	"refactor" is a type, not a location
bugfix	(use fix as type)	"bugfix" is a type, not a location

Scope Examples

By domain:

• feat(auth): add JWT authentication
• fix(payments): handle currency conversion
• refactor(users): extract validation logic

By layer:

• feat(api): add user profile endpoint
• fix(db): resolve connection pool leak
• chore(ci): update Node version to 20

By plugin/skill:

• feat(gitlab-skill): add MR approval support
• fix(python3-dev): correct shebang detection
• docs(commit-staged): clarify scope selection

Breaking changes

Indicate breaking changes clearly:

feat(api)!: restructure API response format

BREAKING CHANGE: All API responses now follow JSON:API spec

Previous format:
{ "data": {...}, "status": "ok" }

New format:
{ "data": {...}, "meta": {...} }

Migration guide: Update client code to handle new response structure

Template workflow

1. Review changes: git diff --staged
2. Identify type: Is it feat, fix, refactor, etc.?
3. Determine scope: What part of the codebase?
4. Write summary: Brief, imperative description
5. Add body: Explain why and what impact
6. Note breaking changes: If applicable

Interactive commit helper

Use git add -p for selective staging:

# Stage changes interactively
git add -p

# Review what's staged
git diff --staged

# Commit with message
git commit -m "type(scope): description"

Amending commits

Fix the last commit message:

# Amend commit message only
git commit --amend

# Amend and add more changes
git add forgotten-file.js
git commit --amend --no-edit

Best practices

1. Atomic commits - One logical change per commit
2. Test before commit - Ensure code works
3. Reference issues - Include issue numbers if applicable
4. Keep it focused - Don't mix unrelated changes
5. Write for humans - Future you will read this

Commit message checklist

• Type is appropriate (feat/fix/docs/etc.)
• Scope identifies WHERE (module/directory), NOT what type of change
• Scope is NOT a banned word: docs, tests, types, refactor, bugfix
• Summary is under 50 characters
• Summary uses imperative mood
• Body explains WHY not just WHAT
• Breaking changes are clearly marked
• Related issue numbers are included

Finally, you will commit the changes with the generated commit message, without using --no-verify:

git commit -m "<commit-message>"

If you are blocked by a pre-commit or prek hook, you should stop, and announce your intended commit message, and that it was blocked, and requires a linting-agent to be run against the issues found.