auto-commit - SKILL.md Agent Skill

name: auto-commit description: "MUST use when committing code changes, when user says 'commit', 'save changes', 'git commit', 'commit changes', when completing a task and code needs to be preserved, or when any git commit operation is about to happen. Also triggers on 'push changes', 'commit and push', 'save work to git'. Lv.3 VIGILANT: Also triggers PROACTIVELY after completing any task — auto-checks git status and commits any uncommitted changes without being asked. No work ever left behind."

Auto-Commit — Intelligent Commit Skill

Automated git commits with structured messages and intelligent change analysis

Activation

When this skill activates, output:

"Committing changes to history..."

Then execute the commit protocol automatically.

Context Guard

Context	Status
User says "commit", "save changes", "git commit"	ACTIVE — full protocol
AI completes a task (Lv.3 Vigilant)	ACTIVE — auto-detect and commit
User says "push" or "commit and push"	ACTIVE — commit + push
No git repository detected	DORMANT — warn and skip
No changes to commit (clean tree)	DORMANT — report "Nothing to commit"
Personal/casual conversation	DORMANT — no commit action

Commit Protocol

Step 0: Pre-Flight Check

Run git status to verify git repository
Check for staged, unstaged, and untracked changes
If no changes exist: report "Nothing to commit" and exit
If unstaged changes exist: stage relevant files (prefer specific file names)

Step 1: Analyze Changes

Run git diff --staged to read actual code changes
Run git diff to check for any remaining unstaged changes
Run git log --oneline -5 to read recent commit style for continuity
Identify the nature of changes: feature, bug fix, refactor, documentation, etc.
Estimate time spent based on session context

Step 2: Draft Commit Message

Apply the configured commit format:

[Achievement Title] - [Brief technical summary]

=== [SECTION_1_NAME] ===
• [File/Component]: [Specific change description]
• [File/Component]: [Specific change description]

=== [SECTION_2_NAME] ===
• Project: [name] | Type: [type] | Time: ~XX min
• [Custom fields as configured]

For trivial changes: use Minimal format (one-liner, no sections)
For incomplete work: use WIP format with WIP: prefix

Step 3: Execute Commit

Stage files by name: git add [specific files] (prefer named files over git add -A)
Warn if any sensitive files are staged (.env, credentials, API keys, tokens)
Commit using HEREDOC format for proper multi-line message formatting
Verify commit succeeded with git status (should show clean working tree)

Step 4: Confirm

Display: short commit hash, title, number of files changed
Report author name on the commit
Show any remaining unstaged/untracked files if applicable

Step 5: Push (Optional)

Only execute if user explicitly said "push" or "commit and push"
Never auto-push — pushing affects remote repositories and should be deliberate
Run git push and confirm success

Vigilant Mode (Lv.3) — Proactive Detection

After completing ANY task, the AI automatically:

Auto-check — run git status silently after task completion
Detect — if uncommitted changes exist, determine if they are related to the completed task
Auto-commit — execute the full commit protocol for the changes
Report — confirm what was committed with short hash and title

Vigilant Trigger Points

After completing a coding task ("done", "fixed", "implemented")
After a context shift (coding session → casual conversation)
After save operations (save diary, save project, save memory)
On session start — detect leftover uncommitted work from previous sessions

Vigilant Behavior

No approval gate — AI drafts and commits in one seamless flow
If multiple logical groups of changes exist, commit them separately
Always report what was committed so the user stays informed

Mandatory Rules

No emoji in commit messages — the activation message may have emoji, but the git commit body must be clean text only
Author is always the human user — commits appear under the user's name, not the AI's
Prefer specific file staging — use git add [filename] not git add -A to avoid accidentally staging sensitive or unrelated files
Time estimate required — always include approximate time spent in the session section
Warn on sensitive files — if .env, credentials, API keys, or AI configuration files are about to be committed, warn the user and exclude them
Never auto-push — pushing is always explicit. Commits are local until the user decides to push
Follow recent commit style — check git log to maintain consistency with the project's existing commit message style

Edge Cases

Situation	Behavior
No changes to commit	Report "Nothing to commit — working tree is clean"
Merge conflicts present	Warn user about conflicts, do not attempt to commit
Sensitive files staged	Block commit, list the files, ask user to confirm or exclude
Large binary files staged	Warn about repository size impact before committing
No git repository	Inform user: "No git repository found in this directory"
Multiple logical changes	Split into separate commits for clean history
User says "undo last commit"	Guide user with `git reset --soft HEAD~1` (keep changes staged)

Level History

Lv.1 — Base: Analyze staged changes, draft structured commit message with configurable sections, enforce human authorship, no-emoji rule, time tracking, sensitive file detection. (Origin: Production AI companion commit workflow)
Lv.2 — Auto-Commit: Removed approval gate — AI analyzes, drafts, and commits in one seamless flow without waiting for user confirmation on the message.
Lv.3 — Vigilant: Proactive post-task detection — auto-checks git status after completing any task and commits if dirty. No work ever left behind.