Git Commit and Push

Overview

This skill provides a structured workflow for committing changes to git and pushing to GitHub. It ensures changes are reviewed, commit messages are meaningful and follow conventions, secrets are not committed, and commits are properly pushed to remote repositories. The skill adapts to project conventions, supporting both Conventional Commits and custom formats.

Execution Philosophy: This skill uses contextual autonomy - it executes immediately for safe, straightforward commits, but asks for confirmation when changes are large, complex, or potentially risky. This balances speed with safety.

When to Use This Skill

Use this skill when the user:

• Explicitly requests to "commit my work" or "push to GitHub"
• Says they want to save/commit/push their changes
• Asks to create a commit or push code
• Wants to share their work on GitHub

Workflow

Step 1: Review Changes

Before committing, review what has changed:

1. Check git status to see modified, new, and deleted files:

   git status
   

2. Analyze diff systematically:

   git diff --stat  # Overview of changes
   git diff         # Detailed line-by-line changes
   

3. Categorize changes to inform commit message:

   • New features: New files, new functions, new capabilities
   • Bug fixes: Modified logic, error handling improvements
   • Refactoring: Structure changes with no behavior change
   • Documentation: *.md files, code comments, docstrings
   • Tests: Test files, test additions/modifications
   • Configuration: Build files, dependencies, settings
   • Styling: Formatting, whitespace, code style only

4. Check recent commits to understand commit message style:

   git log --oneline -20
   

5. Present summary to the user:

   • List all modified, new, and deleted files
   • Highlight key changes by category
   • Note total lines added/removed
   • Flag any unusual patterns (large files, many deletions, etc.)

Step 2: Detect Project Convention

Determine if project uses Conventional Commits or custom format:

1. Check for Conventional Commits pattern:

   # Look for type(scope): format in recent commits
   git log --oneline -20 | grep -E "^[a-f0-9]+ (feat|fix|docs|style|refactor|test|chore|perf|ci|build|revert)(\(.+\))?:"
   

2. Analyze result:

   • If ≥50% of commits match pattern → Project uses Conventional Commits
   • Otherwise → Project uses custom format

3. Note the convention for commit message crafting

Step 3: Craft Commit Message

Create a meaningful commit message following detected convention:

If Conventional Commits Detected

Use format: type(scope): description

Common types:

• feat: New feature or capability
• fix: Bug fix
• docs: Documentation changes only
• style: Code style/formatting (no logic change)
• refactor: Code restructuring (no behavior change)
• test: Adding or modifying tests
• chore: Maintenance, dependencies, build
• perf: Performance improvements
• ci: CI/CD pipeline changes
• build: Build system or external dependencies

Optional scope: Component or module affected (e.g., auth, api, ui)

Format:

type(scope): imperative summary (max 50 chars)

Optional body explaining the change in detail.
Explain WHY, not WHAT (the diff shows what).

🤖 Generated with [Claude Code](https://claude.com/claude-code)

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

Example:

feat(evaluation): add Inspect AI integration to experiment framework

Implements Phase 4 evaluation support with:
- User-written evaluation tasks workflow
- Adapter-only checkpointing (200x storage reduction)
- Complete working examples and documentation

🤖 Generated with [Claude Code](https://claude.com/claude-code)

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

If Custom Format Detected

Match the project's existing style:

1. Summarize changes in 1-2 sentences (imperative, present tense)
2. Focus on "why" rather than "what" (diff shows what changed)
3. Follow observed conventions (capitalization, punctuation, length)
4. Keep concise but informative

Format:

Summary line (imperative mood, present tense)

Optional detailed explanation of the changes.
Focus on motivation and context.

🤖 Generated with [Claude Code](https://claude.com/claude-code)

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

Example:

Add evaluation configuration to experiment definition framework

Implements Inspect AI integration, adapter-only checkpointing support,
and complete working examples for Phase 4 implementation.

🤖 Generated with [Claude Code](https://claude.com/claude-code)

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

Step 4: Assess Commit Risk

Determine whether this commit is "safe" (autonomous) or "risky" (requires confirmation):

Safe commits (execute autonomously):

• ≤ 5 files changed
• ≤ 500 total lines changed
• Not committing to main/master branch
• All changes are related (same category from Step 1)
• No unusual patterns (massive deletions, binary files, etc.)
• User specified which files to commit

Risky commits (require confirmation):

• 5 files changed

• 500 total lines changed

• Committing to main/master branch
• Mixed change types (features + fixes + refactoring)
• Many untracked files (unclear what to commit)
• Large files (> 100KB)
• Unusual patterns detected

If risky, show summary and ask user:

⚠️ Large/complex commit detected

Files: 15 files changed (+2,500 lines, -300 lines)
Types: Features (8 files), Documentation (5 files), Tests (2 files)
Branch: main
Message: "Add evaluation configuration and examples"

This appears to contain multiple logical changes.
Options:
  [Y] Proceed with single commit
  [n] Cancel (I'll help you split it)
  [s] Show detailed file list

If safe, proceed autonomously to Step 5.

Step 5: Check for Secrets

Before staging, scan for sensitive information:

1. Check unstaged changes for secret patterns:

   git diff | grep -E "(api[_-]?key|api[_-]?secret|password|token|secret[_-]?key|private[_-]?key|aws[_-]?access)" -i
   

2. Check for specific patterns:

   • AWS Access Keys: AKIA[0-9A-Z]{16}
   • Generic API keys: [a-zA-Z0-9_-]{32,}
   • Private keys: -----BEGIN.*PRIVATE KEY-----
   • GitHub tokens: ghp_, gho_, ghs_, ghr_
   • Bearer tokens: Bearer [a-zA-Z0-9._-]+
   • Passwords in configs: password\s*[:=]\s*["']?[^"'\s]+

3. Check for sensitive files:

   git status --short | grep -E "\.env|credentials|secrets|\.pem|\.key|\.p12"
   

4. If secrets detected:

   • ❌ STOP immediately
   • Show user the matched patterns (without revealing full values)
   • Explain the risk
   • Ask user to:
     • Remove secrets from code
     • Use environment variables or secret management
     • Add file to .gitignore if appropriate
   • Do not proceed until secrets are removed

5. If no secrets detected:

   • ✅ Proceed to staging

Step 6: Stage and Commit Changes

1. Stage relevant files:

   git add <file1> <file2> ...
   

   Staging guidelines:

   • Stage files related to this logical change
   • Do NOT stage unrelated changes (save for separate commit)
   • Determine files to stage based on change analysis from Step 1
   • If autonomous: Announce which files are being staged
   • If confirmed: User already saw file list in confirmation

2. Verify staged changes:

   git diff --staged --stat
   

3. Announce or show commit message:

   • If autonomous: Announce message briefly ("Committing: 'Add feature X'")
   • If confirmed: User already saw message in confirmation

4. Create commit with formatted message:

   git commit -m "$(cat <<'EOF'
   <commit message here>
   
   🤖 Generated with [Claude Code](https://claude.com/claude-code)
   
   Co-Authored-By: Claude <noreply@anthropic.com>
   EOF
   )"
   

5. Verify commit was created:

   git log -1 --stat
   

Step 7: Push to Remote

1. Check remote status:

   git status  # Check if branch tracks remote
   

2. Push to GitHub:

   # If branch already tracks remote
   git push
   
   # If new branch (first push)
   git push -u origin <branch-name>
   

3. Confirm success:

   • Report push result to user
   • Show number of commits pushed
   • Provide remote URL if available
   • Note any warnings or issues

Important Guidelines

Safety Rules

• Never push force to main/master without explicit user confirmation
• Never skip hooks (--no-verify) unless explicitly requested by user
• Never commit secrets - stop and warn user if detected
• Check authorship before amending commits (don't amend others' work)
• Check for large files - warn if files > 100KB (may need Git LFS)
• Always verify what's being committed before executing

Best Practices

1. Review before committing - show user what will be committed
2. Meaningful messages - explain why changes were made, not just what
3. Atomic commits - each commit should be a single logical change
4. Follow conventions - detect and match project's commit style
5. Add co-author attribution - include Claude Code footer
6. Keep commits focused - don't mix refactoring with features

Commit Scope Guidelines

Good commit scope:

• ✅ One logical change per commit
• ✅ Related files together (e.g., code + corresponding test)
• ✅ Can be described in one sentence
• ✅ Typically < 500 lines (unless it's a rename/refactor)
• ✅ All changes serve the same purpose

Signs to split commits:

• ❌ Using "and" more than once in commit message
• ❌ Changes span multiple unrelated features
• ❌ Mix of bug fixes and new features
• ❌ Experimental changes alongside production code
• ❌ "WIP", "misc fixes", or "various updates" messages

When to split:

# If you have unrelated changes, stage selectively:
git add -p  # Interactively choose hunks to stage

# Or stage specific files:
git add file1.py file1_test.py  # First commit: feature A
git commit -m "feat: add feature A"

git add file2.py file2_test.py  # Second commit: feature B
git commit -m "feat: add feature B"

Error Handling

If commit fails:

• Check for pre-commit hooks modifying files
• Review error message carefully
• May need to stage hook-modified files and amend
• Only amend if: (1) user requested it OR (2) fixing pre-commit hook changes AND it's safe (check authorship, not pushed)

If secrets detected:

• Stop immediately, do not commit
• Explain which patterns matched
• Guide user to remove secrets
• Suggest using environment variables or secret management
• Consider adding .gitignore entries

If push fails:

• Check remote branch status (might need pull first)
• Check for merge conflicts
• Verify remote URL is correct
• Check authentication/permissions
• Report error to user with suggested fixes

If large files detected:

• Warn user about file sizes
• Suggest Git LFS for files > 100KB
• Ask if files should be committed or added to .gitignore

Example Interaction

Example 1: Conventional Commits Project

User: "I'd like to commit my work and push to GitHub"

Assistant:

1. Runs git status and git diff --stat
2. Analyzes: "2 files changed (+1,269 lines)"
3. Categorizes: "New feature (eval config), documentation"
4. Risk assessment: ✅ Safe (2 files, 1,269 lines, feature branch)
5. Checks convention: Detects Conventional Commits (15/20 matches)
6. Scans for secrets: ✅ None detected
7. Announces: "Committing: feat(evaluation): add Inspect AI integration"
8. Executes: git add, git commit, git push
9. Reports: "✓ Committed 795efaf and pushed to GitHub"

Example 2: Large Commit Requiring Confirmation

User: "commit my changes"

Assistant:

1. Analyzes: "15 files changed (+2,847 lines, -156 lines)"
2. Categorizes: "Features (8 files), docs (5 files), tests (2 files)"
3. Risk assessment: ⚠️ Risky (>5 files, >500 lines, mixed types)
4. Scans for secrets: ✅ None detected
5. Shows confirmation:

   ⚠️ Large commit with mixed changes detected
   
   Files: 15 (+2,847 lines, -156 lines)
   Categories: Features (8), Documentation (5), Tests (2)
   Message: "Add evaluation framework with examples and tests"
   
   Consider splitting into:
   - Commit 1: Core evaluation framework (8 files)
   - Commit 2: Documentation (5 files)
   - Commit 3: Tests (2 files)
   
   Proceed with single commit? [Y/n/split]
   

6. Waits for user response

Example 3: Secrets Detected (Always Blocks)

User: "commit my changes"

Assistant:

1. Reviews changes: Modified config.py (+10 lines)
2. Scans for secrets: ❌ FOUND pattern matching API key
3. STOPS: "⚠️ Security Warning: Detected potential API key in config.py"
4. Shows matched line: api_key = "sk_live_..."
5. Advises: "Please use environment variables instead: api_key = os.getenv('API_KEY')"
6. Does not proceed until user confirms secrets are removed

Notes

• This skill follows the git safety protocol from the Bash tool documentation
• Commit message format includes Claude Code attribution as per project conventions
• Always run commands sequentially (staging, committing, pushing) not in parallel
• Adapts to project conventions rather than enforcing a single standard
• Uses contextual autonomy: executes immediately for safe commits, confirms for risky ones
• Security checks are non-negotiable - always blocks on secrets
• Thresholds: Safe ≤ 5 files, ≤ 500 lines, not main/master, single change type

References

• Conventional Commits Specification
• Git Commit Best Practices
• Pre-commit Framework
• Git Hooks Documentation