Git Commit

Git Commit

Create one or more atomic commits from the working tree. Autonomous by default — execute without asking unless the user says "show plan first" or "confirm before commit".

Metadata

Run from repo root:

node "${CLAUDE_PLUGIN_ROOT}/skills/git-commit/scripts/git-changes.mjs"
echo "---recent-subjects---"
git log --pretty=%s -n 20 2>/dev/null || true

If in_repo: no, stop: "This directory is not a git repository."

If status shows clean working tree, stop: "Nothing to commit."

Workflow

Step 1: Check remote state (does NOT block committing)

Measure whether the remote moved while work was in progress. This is informational — it never stops the commit. Committing while behind is fine; the merge happens later in git-sync.

git fetch origin 2>/dev/null || true
git rev-list --left-right --count @{u}...HEAD 2>/dev/null   # output: "<behind>	<ahead>"

Remember the behind number (the first one) — surface it in Step 6. Do not run git pull here, and do not skip committing for any remote state.

Step 2: Analyze changes

Use the Metadata block (---status---, ---diffstat---, ---recent-subjects---).

• For small changes (≲5 lines per file in diffstat), infer intent from path + counts — skip full git diff
• Run git diff <path> only when intent is unclear from filename and diffstat
• For untracked directories (?? path/), treat contents as the change unless the directory is huge

Step 3: Plan commit groups

Scan for distinct concerns. Split when obvious; one commit when ambiguous.

Apply the grouping rules in references/conventions.md — file-level groups, file-type ordering across multiple commits, small diffs to a single commit.

Draft messages matching ---recent-subjects--- style. If history is empty or mixed, use the default types from references/conventions.md.

Step 4: Safety check

Before staging, scan paths for secrets: .env, credentials, *.pem, *.key, token files. Exclude them and warn if present.

Never stage with git add -A or git add .. Stage explicit file paths only.

Step 5: Commit

Commit on the current branch. Do not create feature branches; commits go straight to wherever HEAD is (typically main).

For each group:

git add path/to/file1 path/to/file2 && git commit -m "$(cat <<'EOF'
type: subject line

Optional body explaining why.
EOF
)"

Step 6: Confirm

git log --oneline -n <number-of-commits-created>
git status --short --branch

Report: commit hash(es), subject line(s), files per commit.

If the behind count from Step 1 was greater than 0, the local branch is now diverged from the remote. Add a line: "⚠ The remote advanced by N commits while you worked — run git-sync to merge them in before git-push." Do not push a diverged branch.

Authorship

• Never add co-author trailers or AI attribution
• Write messages as if the user wrote them

Optional confirmation mode

Only when user explicitly requests review: present the plan (files per commit + messages) and wait for approval before Step 5.

Guardrails

• Stage explicit file paths only — never git add -A or git add ..
• Never stage .env, credentials, *.pem, *.key, or token files; warn if present.
• Never create feature branches — commit to the current branch.
• Never push a diverged branch — if the remote advanced, git-sync integrates first.

Order

Ordering across operations is owned by blake-os:git-ops.