grai-gate

Deterministic style gate for markdown written by coding agents. One Python engine, no model calls, no network. It scores sentence burstiness and flags structural patterns that survive spellcheck. It is a style gate, not an authorship detector.

On the v2 corpus (58 human-edited docs, 63 model-generated docs), the gate flagged 7 of 63 model docs and 3 of 58 human docs. It catches visible patterns, not provenance.

GitHub Action

name: cadence
on:
  pull_request:
    paths: ['**.md', '**.mdx', '**.txt']
jobs:
  gate:
    runs-on: ubuntu-latest
    steps:
      - uses: actions/checkout@v4
      - uses: adrianpeticila/grai-gate-action@v0.1.7
        with:
          fail-on-slop: "false"

The action scores only files changed in the pull request. It fetches the base ref itself, so a shallow checkout is fine. The default is advisory: the card lands in the log and the job stays green. Set fail-on-slop: "true" to block.

Inputs: path (scan target when no pull request base is available), fail-on-slop, verbose.

Python package

pipx run grai-gate README.md

or

pip install grai-gate
grai-gate docs/

Exit codes: 0 clean, 1 slop, 2 input error. The CLI accepts -h and --help.

pre-commit

repos:
  - repo: https://github.com/adrianpeticila/grai-gate-action
    rev: v0.1.7
    hooks:
      - id: grai-gate

The hook scores markdown, mdx, and plain text. It exits 1 on slop and stops the commit. Skip one commit with SKIP=grai-gate git commit.

One engine

src/grai_gate/gate.py is vendored byte-for-byte from the Antigravity plugin in the grai repo. Run python scripts/sync_engine.py --check to prove the copies match; it exits 1 on drift.

UTM attribution

The report card links to the Pro Cockpit. Attribution resolves from GRAI_UTM_SOURCE, then GITHUB_ACTIONS, then PRE_COMMIT. The installed PyPI command defaults to pypi; a direct engine run defaults to cli.