manage-precommit

A Claude Code skill that builds a repository's .pre-commit-config.yaml from a small curated catalog — and merges into an existing config without clobbering what is already there.

Two invariants:

• Latest, pinned. Every repo it adds gets the newest release tag, fetched live at run time. Nothing is hardcoded.
• Never clobber. An existing config keeps its comments, formatting, top-level keys, repo revs, and any hooks outside the catalog. Only missing pieces are added, and nothing is duplicated.

Catalog

Mermaid ships no offline linter — its real parser only runs in a browser — so the bundled hook renders each diagram with mermaid-cli and fails the commit on a parse error.

Requirements

• pre-commit 4.0+
• Python 3.10+ and git. No third-party Python packages — the skill is installed by symlink and its scripts run under your system python3, so anything it needed would have to be installed by hand on every machine.
• For the mermaid hook only: Node.js, plus a Chromium/Chrome the hook's mermaid-cli can drive (it downloads one on first use if none is reusable). In CI, a container, or on a distro that restricts unprivileged user namespaces, Chromium's sandbox cannot start — set MERMAID_LINT_NO_SANDBOX=1 to run it without one. Opt-in, because that removes a real boundary around a browser rendering text out of the repository. The hook says so itself when it hits that failure, and reports it as an environment problem rather than an invalid diagram.

Install

Both routes install a symlink into ~/.claude/skills/, never a copy — so there is only ever one set of files, and nothing can drift out of sync. Restart Claude Code to pick it up.

As a package

# not on PyPI yet -- install straight from the repository:
pipx install git+https://github.com/grammy-jiang/manage-precommit.git
manage-precommit install

The package is an installer and nothing else; all the work lives in the skill files it links. --dry-run prints what would happen, refusals included; --dest DIR overrules the default location; --force acts on something that is not ours. manage-precommit uninstall removes the link and leaves the package alone.

It refuses to touch anything it did not create: a real directory there may be a hand-written skill, and a link pointing elsewhere is not its to remove.

From a checkout

git clone git@github.com:grammy-jiang/manage-precommit.git
cd manage-precommit
make install     # ~/.claude/skills/manage-precommit -> ./src/manage_precommit/skill
make uninstall

This links the working tree, so an edit is live on the next Claude Code restart with no rebuild.

Usage

Invoke the skill and answer the prompts:

/manage-precommit

It scans the repo, proposes hooks (always-on plus ones matching what the repo actually contains), merges the selection, installs the git hook, runs the suite, shows the diff, and — only with confirmation — commits and pushes just the pre-commit setup files.

The engine also runs standalone (S=src/manage_precommit/skill/scripts):

python3 $S/precommit.py --catalog                      # list catalog keys
python3 $S/precommit.py --dir /path/to/repo --detect   # inspect existing config
python3 $S/precommit.py --dir /path/to/repo --recommend  # what the repo calls for, and why
python3 $S/precommit.py --dir /path/to/repo --force \
    --templates-file keys.txt --facts-out /tmp/facts.json

How a run flows

flowchart TD
  A[Scan repo] --> B[Propose hooks]
  B --> C{User selects}
  C --> D[Merge templates<br/>pin latest versions]
  D --> E[Install + run hooks]
  E --> F{All pass?}
  F -->|no| G[Report, stop]
  F -->|yes| H[Review diff]
  H --> I{Commit?}
  I -->|yes| J[Commit setup files only]
  J --> K[Push, with force gated<br/>behind a compare]
  I -->|no| L[Leave in working tree]

Layout

src/manage_precommit/skill/SKILL.md       the procedure Claude Code reads
                    skill/scripts/precommit.py  catalog, detect, recommend, merge, verify
                    skill/scripts/gitwork.py    status, commit, push-plan, push, facts
                    skill/scripts/config.py     the config scanner and additive writer
                    skill/scripts/summary.py    the end-of-run summary
                    skill/scripts/shared.py     sanitiser, no-follow reader, JSON contract
                    skill/templates/            one YAML fragment per catalog entry
                    skill/assets/               files copied into the target repo
                    skill/references/           on-demand detail (force-push, worked example)
tests/                                     pytest suite

The checkout and the installed tree are the same paths: nothing is remapped, so a path in a traceback traces back here by relative position.

How the merge keeps its promise

Each catalog entry is a YAML fragment with a __REV__ or __NPM__ placeholder. The engine substitutes the latest upstream version and inserts the fragment as text — it never re-emits the file. Every byte outside an inserted block is carried across untouched, and the write is rejected unless the original can be reconstructed from the result by deleting exactly the blocks that were added.

Reading is a strict line scanner rather than a YAML library. It refuses anything it cannot prove it understands — anchors, aliases, merge keys, flow sequences where a block is expected, more than one document, tabs — and says which line. A refusal is an exit code; a guess would be a wrong answer that looks right.

Dogfooding

This repo uses its own hooks. scripts/lint-mermaid.mjs is a symlink to src/manage_precommit/skill/assets/lint-mermaid.mjs, so the copy this repo runs and the payload it ships to other repos cannot drift apart.

Development

pip install -e '.[dev]'
python3 -m pytest        # 119 tests; no test touches the network
python3 -m ruff check . && python3 -m ruff format --check .
python3 -m mypy