Git Commit

Purpose

Produce reviewable commits: a clear subject line and a body that explains what changed and why. Staging is whole-tree by default so one commit reflects everything currently pending on the branch, not a cherry-pick of what this session edited.

Authority

• git commit allowed when the user explicitly asks you to commit, or a checked-in workflow skill (for example execute-plan) authorizes commits as part of its procedure.
• Message text only when the user wants a commit message but did not ask you to run Git: use the subject and body rules below, optionally after git status / git diff for context. Do not git add or git commit.
• If committing would be a surprise—no explicit commit request and no workflow granting commit authority—stop before git commit and confirm intent.

Use When

• A commit with a strong message is requested or workflow-authorized.
• The user wants help composing commit message text (with or without you running git commit—see Authority).

Do Not Use When

• Git writes are forbidden by context and nothing overrides that.

Staging scope

• Default (this skill alone): Stage all changes in the working tree that Git will track—modified, deleted, and untracked files—typically git add -A at the repo root. That includes edits made by other agents, the user, or tooling. .gitignore still applies (do not fight it to force secrets into the commit).
• Parent workflow override: If the active procedure explicitly limits what may be staged (for example execute-plan phase commits: only files for the completed phase), follow that staging scope and still use the message rules below.
• If git status shows something surprising (unexpected paths, credentials, huge artifacts), surface it before committing; do not silently drop co-workers’ changes to keep the commit “tidy.”

Procedure

1. Inspect full git status and git diff (and git diff --staged if anything is already staged). Understand everything that will be included.
2. Stage per Staging scope above.
3. Subject line (about 50 characters target, 72 hard cap):
   • Imperative mood: "Add", "Fix", "Refactor" — not past tense ("Added", "Fixed") or third-person singular ("Adds", "Fixes") as the subject.
   • The body may include issue closers such as Fixes #123 when that is the project convention.
   • Be specific about area or behavior, not "Update code" or "WIP".
4. Body (one blank line after the subject; wrap near 72 columns):
   • What: bullets or short paragraphs covering all substantive areas in this commit (not only the files you touched this turn).
   • Why: motivation, tradeoffs, what was broken or awkward, or why this approach over alternatives.
   • Context (optional): issue links, plan paths, or follow-ups when they help the next reader.
5. Commit (commit path only), for example in bash or zsh:
   • git commit -m "subject" -m $'paragraph...\n\n- bullet'
   • or git commit with an editor when the body is long.
6. Verify (commit path only): git show --stat HEAD matches intent (entire staged set landed).
7. Post-commit audit (commit path only): after the commit succeeds, follow Post-commit audit. Run ruff check . first; then scan only the HEAD diff for bug risks, god modules, needless complexity, and duplication (lightweight—see What this is not). Tell the user about findings in the same turn (paths, severity, next step if obvious)—as heads-up, not scolding. The commit is already done; do not delay the commit, imply it was a mistake, or push amend/revert unless the user asks. Do not silently “fix” findings unless the user asked; surfacing is the goal. If everything passes and nothing worrisome stands out in the diff, you may omit commentary.

On the message-only path, perform step 1 as needed for context, skip staging and steps 5–7, and output the subject and body (use the template below).

Message template

Add/improve/fix <behavior> in <area>

What:
- ...

Why:
- ...

Post-commit audit

Run this only after git commit completes successfully (same repo, same branch).

Intent (non-blocking)

• Commit always wins. Nothing here vetoes, reorders, or shames a commit that already landed. People should keep committing; this step is extra awareness for the author, not policy for the team.
• Alerts, not gates. Ruff or review findings are heads-up so you can fix forward (ruff check . failed after commit—here is the output) or choose to ignore. Do not treat red Ruff as “the commit was wrong.”
• Optional follow-up. Suggest a fix commit or local cleanup only when helpful; never pressure amend/revert unless the user explicitly wants that.
• Deeper review: For format check + tests, A–F grades, optional multi-agent passes, and broader scope, the user should invoke the audit skill (message contains audit)—do not inflate post-commit into a full audit.

What this is not

Stay narrow so commits stay fast:

• Do run ruff check . and read git show -p HEAD (this commit only).
• Do not run ruff format --check, full pytest, or integration tests here by default—those belong to audit, deploy, or CI unless the user’s task already called for them.

Ruff (primary static pass)

From the repository root, with dev deps installed (pip install -e '.[dev]'):

• ruff check .

If the command is missing, dependencies are not installed, or the run fails, report the full relevant output to the user with neutral framing (for example “Heads-up: ruff check . failed after your commit”). The commit stays; the user can fix in a follow-up when ready. Do not hide lint errors.

Diff scope

Use the committed change as the source of truth—e.g. git show HEAD or git show -p HEAD, and the paths touched: git diff-tree --no-commit-id --name-only -r HEAD.

Bug scan

Re-read the change as a quick self-review on changed lines only. Flag likely problems for the user (this is not a substitute for CI, audit, or the full test suite unless already part of the task):

• Logic errors, wrong conditions, off-by-one, incorrect defaults.
• Missing or broken error handling / early returns where failures are plausible.
• Obvious regressions: removed guards, weakened validation, race or stale-state risks introduced by the new code.
• API or type mismatches, impossible states, or changes that contradict the commit message.
• Training/device edge cases if the diff touches gemma_tuner/models/ or device.py.
• Security / privacy (glance): obvious secrets or PII logged in the diff; if it needs auth review, nudge toward audit instead of digging here.
• Anything that would make you say “wait, that can’t be right” on a PR.

God module scan

Align with repo philosophy: clear separation of concerns, no dumping unrelated responsibilities into one place.

Flag when this commit creates or substantially grows a module that looks like a grab-bag “does everything” file—for example:

• Size: one file approaches or exceeds the project’s per-file size guidance (this repo aims for well under ~1000 LOC per file—use judgment near that band).
• Scope creep: unrelated domains, layers, or features fused into a single module (many disparate exports, orchestration + persistence + UI helpers in one file, etc.).
• Fan-in smell: a change that makes one file the obvious choke point for unrelated call sites when a split would be natural.

When in doubt, flag lightly with reasoning; avoid crying wolf, but do not skip obvious smells to avoid bothering the user.

Needless complexity and duplication

• Needless complexity: extra layers, over-abstraction, clever patterns where a straight line would do, new dependencies or infrastructure for trivial wins, configuration explosions, or branching that obscures the actual behavior—especially when it violates simpler is better for this repo.
• Duplication: copy-pasted blocks, parallel implementations of the same rule or transform, near-identical helpers that should be one function, or “another copy” of logic that already exists elsewhere—when unifying would obviously reduce drift risk.

Flag when this commit introduces or worsens these; cite paths and why it matters.

What to output

• Ruff: pass/fail; if fail, enough output to act on, framed as post-commit awareness—not a blocked workflow.
• If other issues: short summary, bullet list with file paths, what you saw, and whether it looks like a definite bug vs. a risk vs. a maintainability smell.
• If clean: say nothing or one line—no boilerplate required.

Anti-patterns

• One-word subjects ("fix", "wip", "updates").
• A body that only repeats the subject.
• Cherry-staging only the files you personally edited when the user asked for a commit and no narrower workflow override applies—other agents’ work must ship too.
• Claiming a “single logical change” while omitting co-present dirty files without explicit user direction.
• Skipping the post-commit audit after a successful commit when this skill’s commit path ran—especially omitting ruff check . or staying silent when it fails or the diff clearly introduces risk.
• Turning post-commit into a full audit: running pytest, ruff format --check, multi-agent Task delegation, or A–F grades here—use the audit skill when the user says audit.
• Blocking mindset: implying the user should not have committed, or urging amend/revert, because of post-commit findings—unless they asked for that.

Related skills

• audit: user message includes audit; optional tests, optional multi-agent, A–F grades, broader scope.
• deploy: blocking Ruff + tests before release.