---

name: commit description: Stage changes, run pre-commit hooks, and create a well-formatted git commit allowed-tools: Bash, Read, Grep, Glob

Git Commit Workflow

Create a clean, well-documented git commit following this project's conventions.

Step 1: Review changes

git status
git diff --stat

Show the user a summary of what will be committed.

Step 2: Stage files

Stage files selectively (never git add . or git add -A blindly):

# Stage specific files
git add <file1> <file2> ...

Files to NEVER commit

| Pattern | Reason | |---------|--------| | .env | Contains HUGGINGFACE_ACCESS_TOKEN | | *.pyc, __pycache__/ | Compiled Python | | .DS_Store | macOS junk | | model_output/ | Downloaded models | | data/eval_results/ | Generated evaluation data | | output/ | Pipeline outputs | | .claude/settings.local.json | Personal Claude settings | | .claude/CLAUDE.local.md | Personal Claude memory |

If $ARGUMENTS says "all" or "todo", stage all modified/untracked files except those above.

Step 3: Pre-commit hooks

This project has pre-commit hooks configured in .pre-commit-config.yaml:

• ruff (linter) - auto-fixes code style
• ruff-format (formatter) - auto-formats code
• gitleaks - scans for leaked secrets

If a hook fails:

1. The commit is aborted (it did NOT happen)
2. Fix the issues (ruff usually auto-fixes, just re-stage)
3. Create a NEW commit (never --amend after a hook failure - that would modify the wrong commit)

Step 4: Commit message format

Use Conventional Commits:

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

[optional body]

Co-Authored-By: Claude Opus 4.6 <noreply@anthropic.com>

Types

| Type | When to use | |------|-------------| | feat | New feature or functionality | | fix | Bug fix | | docs | Documentation changes (README, CLAUDE.md, comments) | | refactor | Code restructuring without behavior change | | chore | Maintenance (deps, configs, cleanup) | | style | Formatting, linting (no logic change) | | test | Adding or modifying tests | | perf | Performance improvements | | ci | CI/CD changes |

Scopes (optional, use the most relevant)

| Scope | Files | |-------|-------| | inference | model/inference/, infrastructure/inference_pipeline_api.py | | rag | application/rag/ | | eval | model/evaluation/ | | training | model/finetuning/ | | etl | application/crawlers/, configs/digital_data_etl_* | | dataset | application/dataset/ | | settings | settings.py, .env, configs/ | | deps | pyproject.toml, poetry.lock | | infra | docker-compose.yml, ZenML configs |

Examples

feat(rag): Add author filtering to Qdrant search
fix(inference): Add task parameter to HuggingFaceEndpoint
docs: Rewrite README for local-only setup
chore(deps): Upgrade langchain-huggingface to 1.2.0
refactor(eval): Replace OpenAI judge with HuggingFace

Step 5: Create the commit

Use HEREDOC for proper message formatting:

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

<body if needed>

Co-Authored-By: Claude Opus 4.6 <noreply@anthropic.com>
EOF
)"

Step 6: Verify

git log --oneline -3

Show the user the new commit hash and message.

After the commit

Do NOT push automatically. Tell the user they can push with /push or git push.

If pre-commit hook fails

1. Check what failed: usually ruff auto-fixed files
2. Re-stage the auto-fixed files: git add <fixed-files>
3. Create a NEW commit (do NOT use --amend)
4. If gitleaks fails: a secret was detected. Remove it from the staged files before retrying.