Commit Message

Generate a commit message for the currently staged changes.

Scope: staged changes ONLY

Consider only what is staged. Never look at unstaged or untracked changes.

1. git diff --cached --stat — see which files are staged and the change size.
2. git diff --cached — read the actual staged hunks; base the message solely on these.
3. If git diff --cached is empty, stop and tell the user nothing is staged (do not fall back to unstaged changes).

Do not run git add. Do not consider git status working-tree (unstaged) or untracked entries.

Match the recent history

Before writing, check the history for coherence:

4. git log -1 --format=%B — read the latest commit's full message (subject AND body). This is the most important reference.
5. git log -n 15 --pretty=format:%s — scan recent subjects for the type/scope vocabulary and phrasing.

Use the latest commit to:

• Avoid duplication/contradiction: if the staged changes continue or fix work the latest commit just landed, phrase the new message as that continuation — do not re-announce what is already committed (e.g. don't say "add X" if the latest commit already added X; say "refine X" / "fix X").
• Stay consistent: reuse its scope naming, casing, and imperative style so the new message reads as the next entry in the log, not an outlier.

Then obey the format and rules below.

Format

type(scope): description

• type — one of: feat, fix, docs, refactor, test, chore (per .claude/rules/git-flow.mdc).
• scope — the primary package/area touched (e.g. group, node, db). Pick the dominant one from the staged paths.
• description — imperative mood, lowercase start, no trailing period.

Optional body (only when the change spans multiple distinct things):

• A blank line after the subject, then a bullet list.
• One bullet per distinct change, each starting with - .
• A bullet may lead with a short label: framing when it aids clarity (e.g. backward compatibility: ...).

Rules

-Why: Define Why more than What in description

• Precise & crisp: state exactly what changed, no filler ("various", "some", "improvements", "update code").
• Describe observable behavior, not plumbing: name user-facing things (endpoints, response fields/keys, capability values, config keys). Omit internal mechanics — helper/function names, parameter or signature changes, and call-chain threading — unless that mechanic is the change. e.g. "derive matter_node_id from node_id", not "derive it via MatterNodeIDFromThingName threaded through AddNode".
• Do not break lines: never hard-wrap text. The subject is one line; each bullet is one line (however long). No mid-sentence newlines.
• Subject ≤ ~72 chars when achievable without losing precision; bullets have no length limit (one line each).
• List allowed: use bullets only when there are multiple independent changes; for a single change, a one-line subject (no body) is preferred.
• Derive everything from the diff — do not invent rationale that isn't evident in the staged changes.

Self-review before output

Re-read the draft against the Rules and the latest commit, and revise until flawless:

• Cut every plumbing detail (helper/function names, signature/param changes, call-chain threading) and all filler.
• Cut any bullet that restates work the latest commit already landed.
• Ensure each bullet is exactly one distinct, observable change; merge near-duplicates, split "and"-joined ones.
• Confirm the type and scope match the history, the subject is imperative and ≤ ~72 chars, and no line is hard-wrapped.

Output

Print the final commit message in a fenced code block, ready to paste. Do not create the commit unless the user explicitly asks; if they do, branch off main first per .claude/rules/git-flow.mdc (no direct commits to main).