Safely orchestrate Conventional Commits for staged Git changes, or all working-tree changes when the user explicitly asks to include everything. Use when asked to write a commit message, split staged or working-tree changes, organize a messy index, or draft commit text without pushing. Auto-detects commit language; explicit phrases like 用中文提交 or 请使用中文拆分提交所有的改动 override.
Safely orchestrate Conventional Commits for staged Git changes, or all working-tree changes when the user explicitly asks to include everything. Use when asked to write a commit message, split staged or working-tree changes, organize a messy index, or draft commit text without pushing. Auto-detects commit language; explicit phrases like 用中文提交 or 请使用中文拆分提交所有的改动 override.
Use this workflow in order: preflight -> split plan -> classify -> compose -> commit/draft -> verify.
Decide the active change authority and output language before doing anything else:
staged-only is the default. Respect the current index and treat unstaged or untracked files as context only.
all-changes is allowed only when the user explicitly asks to include everything, such as "all changes", "所有改动", "全部改动", "不管有没有 stage", or "包括未跟踪文件". In this mode, the skill may rebuild the index from the full working tree and should treat any existing partial staging as intentionally overridden by the user.
commit-language is detected, not fixed-default, because forcing one language fights repos (and users) whose history is already in the other. Resolve it in this order, then use it for scope, subject, body, and explanatory output:
An explicit instruction in the user's message — "use English", "英文提交", "用中文", "中文提交", or the legacy phrase 请使用中文拆分提交所有的改动. Honor whichever language the user names.
The dominant language of the user's current request message.
The repository's own habit, sampled in Preflight from git log -n 20 --format=%s. Match what the repo already does so your commit doesn't read as an outlier.
Fall back to English only when none of the above gives a clear signal.
Language is orthogonal to emoji, [AI], and trailers: detecting Chinese never changes whether [AI] or agent trailers attach.
agent-mode is on by default whenever this skill runs (the caller is an agent). It injects [AI] in the header, attaches Agent-Task / Agent-Model / Generated-By trailers, and applies the Why-line rule for feat / fix / refactor / perf. Turn it off only when the user explicitly says "no AI tag", "不要 AI 标记", "不加 agent trailer", or equivalent.
1. Preflight
Inspect git status --short first. Then inspect the active change set:
staged-only: git diff --staged --stat and git diff --staged
all-changes: git diff --stat, git diff, git diff --staged --stat, and git diff --staged
If rtk is available, prefer rtk git status, rtk git diff --staged, and rtk git diff for model-visible inspection.
Explicitly note:
the active change authority (staged-only or all-changes)
staged changes ready to commit
unstaged changes that might make the index misleading
untracked files that are intentionally excluded vs accidentally forgotten
Branch immediately on preflight results:
staged-only + no staged changes: stop and tell the user to stage files first.
all-changes + no staged, unstaged, or untracked changes: stop and say there is nothing to commit.
The active change set exists but is obviously mixed and cannot be safely separated from inspection alone: do not improvise a commit. Output a split plan and stop.
The user explicitly asked only for commit text, a draft, or suggestions: continue through classification and composition, but do not run git commit.
Detect agent context (skip when user disabled agent-mode):
Resolve agent-model from the model currently running this skill (e.g. claude-opus-4-8, claude-sonnet-4-6, gpt-5-codex). This value is required.
Resolve agent-task by trying, in order: (a) explicit task ID or issue URL in the user message, (b) closes #N / refs #N mentioned by the user, (c) ticket ID extracted from the current branch name, (d) Agent-Task value from the previous commit on this branch, (e) fallback to unspecified.
Resolve agent-prompt-ref only when a stable prompt reference exists; otherwise leave empty.
Detect checkpoint-mode: triggered by user words such as checkpoint, 打个 checkpoint, 先存一下, WIP, [WIP], work in progress, 先提交一下，待会再整理.
Sample repository conventions so the skill adapts instead of imposing one repo's habits everywhere:
Language + style: read git log -n 20 --format=%s. Note the dominant subject language (this feeds commit-language step 3 in §0), whether subjects carry gitmoji, and whether they already use an [AI] tag.
Config: if commitlint.config.*, .commitlintrc*, .czrc, .cz.*, .gitmessage, or a CONTRIBUTING commit section exists, treat its allowed type / scope list and length rules as authoritative over this skill's defaults.
When no signal exists, fall back to this repository's defaults: emoji on, [AI] + agent trailers on, scope from the changed path.
Safety scan the active change set's file list (git diff --staged --name-only, plus untracked paths in all-changes). If it includes likely secrets (.env, *.pem, *.key, id_rsa, id_ed25519, *.p12, *.keystore) or large/binary blobs (> ~1 MB), do not commit silently. Surface each risky path with its staged/untracked/tracked state, size when practical, and why it is risky. For large files, ask the user to decide per path before staging or committing: intentionally commit it, leave it untracked, add or adjust a .gitignore pattern, or move it to Git LFS / external artifact storage. Do not treat all-changes as permission to sweep every large file into history. "Safely orchestrate" means catching leaked secrets and accidental artifacts before they enter history, not only splitting commits cleanly.
2. Split Plan
Treat split planning as a safety step, not a formatting nicety.
Group the active change set by one coherent unit at a time, usually one of:
a single feature or fix
one module or package
code vs docs vs tests when staged together by accident
In all-changes mode, it is acceptable to ignore the current staged subset only because the user explicitly asked to include everything.
Atomic check (skip in checkpoint-mode). For every candidate commit, answer:
Can the repo compile / tests pass at this commit?
Can git revert <sha> undo it without leaving the repo inconsistent?
Can one subject line + one Why line explain its intent?
If any answer is no, return to the split-plan layer instead of committing.
If the split is clean and obvious, explain the planned commit boundaries before composing messages. In all-changes mode, say whether execution will rebuild the index per commit.
If the split is ambiguous, stop at the plan. Do not perform hunk-level surgery, edit files, or guess hidden intent just to manufacture atomic commits.
Checkpoint mode branch: when checkpoint-mode is detected, skip the atomic check and prepare a single chore(wip): [AI] 🔧 [WIP] <subject> commit covering the active change set. Skip Why enforcement. Still attach agent trailers. In §6 Verify, remind the user to squash [WIP] commits before merging.
Default to emoji because this repository expects it. Opt out when the user requests no emoji, or when Preflight §5 found a target repo whose history carries no gitmoji.
Infer scope from the changed paths instead of inventing one: take the common parent of the changed files. In this repo's layout skills/<category>/<name>/… → scope <name>, and platforms/<platform>/… → scope <platform>. Changes spanning unrelated top-level areas are a signal to split (§2), not a reason to pick a vague umbrella scope. Prefer a scope that already appears in git log when it fits.
Keep type in English — the Conventional Commit keyword is the one machine-parsed token and stays stable across languages.
Render scope, subject, body, and explanatory output in the commit-language resolved in §0 (user instruction → request language → repo history → English fallback).
Mark Why-required when type ∈ {feat, fix, refactor, perf} and checkpoint-mode is off.
4. Compose
Resolve the helper path once instead of hardcoding an interpreter name. <skill-dir> below means this skill's base directory — the path announced when the skill loads. It is not an environment variable, so substitute the literal path:
Pass --agent-task <value> (use unspecified only as last-resort fallback).
Pass --agent-prompt-ref <ref> only when a stable reference exists.
When you know them, pass --confidence <high|medium|low>, --scope-risk <narrow|moderate|broad>, and --tested "<how verified>". Recommended in agent-mode but not enforced — omit a field rather than guessing its value.
For Why-required types, pass --why "<motivation>" and --require-why so the script fails loudly when Why is missing.
In checkpoint-mode, use --type chore --scope wip and prepend [WIP] to summary; skip --require-why; skip --closes / --refs.
If the user disabled agent-mode: omit --ai, omit all --agent-* flags, omit --generated-by-agent. Fall back to plain Conventional Commit.
If the Why-required check fails and Why cannot be inferred from user context: stop, return to the split-plan layer, and ask the user for the motivation. Do not fabricate a Why line.
Never hand-roll a multiline commit message when the script can express it safely.
PROHIBITED: Never include Co-Authored-By, attribution lines (e.g. 🤖 Generated with Claude Code), or push commands by default. Generated-By: agent is a structured trailer for audit grep, not an attribution line — it stays.
5. Commit Or Draft
If the user asked only for a draft, return the proposed commit text and stop.
Execution consent checkpoint: Before any git commit, display the final commit message (header + body + footer) and the list of files to be committed. Explicitly call out whether [AI] is in the header, whether Why is present, and which agent trailers will attach. Then decide whether to pause:
Proceed without an extra confirmation when the user's current request already authorizes execution, such as "commit it", "execute the commit", "commit all changes", "直接提交", "提交了", "按方案执行", or 请使用中文拆分提交所有的改动, and preflight found no secret/large-binary risk, no ambiguous split, no missing Why, and no draft-only wording.
Wait for explicit confirmation when execution was not clearly requested, when the user asked to review/plan/draft first, when the proposed file set or message differs materially from the requested scope, or when any safety gate in this workflow says to ask before committing.
If the user asked to commit and staged-only is active, commit only the safe staged set. Write the message to a file and commit with git commit -F <message-file> so PowerShell and POSIX shells behave consistently. Put the message file outside the working tree — $(git rev-parse --git-dir)/COMMIT_MSG_SKILL or the OS temp dir — never inside the repo, where the all-changes flow's git add -A would sweep it into the commit.
If the safety scan found large/binary files, wait for an explicit decision for each risky path before any broad staging command. If the correct outcome is to ignore generated artifacts, edit or ask for a .gitignore update first, then re-run preflight so ignored files are no longer part of the candidate set. Include .gitignore only when the user approved that ignore policy.
If the user asked to commit and all-changes is active for a single atomic commit, run git add -A only after the safety scan is clear or all risky files have explicit include/ignore decisions, so tracked, deleted, and safe untracked non-ignored files enter the commit set.
If the user asked to split-commit in all-changes mode, rebuild the index one commit at a time using file/path boundaries only. Use full-worktree staging plus path-based staging or unstaging as needed, but stop if the split would require hunk-level staging or other hidden reconstruction.
If rtk is available and the user wants compact feedback, rtk git commit -F <message-file> is acceptable for the final commit step.
Do not push by default. Only discuss or run git push if the user explicitly asked for it.
6. Verify
Read the git commit output before claiming success.
Distinguish two hook outcomes before reacting:
Hook rejected the commit (non-zero exit, message-format or lint failure): stop and report the original hook failure. Do not silently rewrite the message unless the output clearly says the format is invalid and the user asked you to fix it.
Hook rewrote files (a formatter such as prettier/black/gofmt modified tracked files and left them unstaged, aborting or staling the commit): re-inspect git status, re-stage the hook's edits, and retry the same commit message. Say that the hook reformatted files — do not treat the reformatting as your own change.
After a successful commit, summarize:
the final header
whether staged-only or all-changes mode was used
the resolved commit-language and which signal chose it (user instruction / request language / repo history)
whether emoji was included
whether [AI] tag was applied and which Agent-* trailers attached
which quality trailers attached (Confidence / Scope-risk / Tested), if any
whether the commit is a chore(wip) checkpoint
whether a Why line is present (required for feat/fix/refactor/perf)
whether untracked files were included
whether issues or breaking changes were attached
If you stopped before committing, say exactly why: no active changes, no staged changes under staged-only, ambiguous split, Why missing for Why-required type, or draft-only request.
If the branch now contains multiple chore(wip): commits, remind the user to squash them via git rebase -i <base-branch> before merging — but do not run rebase from this skill.
