AgentGuard Composite GitHub Action

The reusable AgentGuard action wraps agentguard ci in a composite GitHub Action. It runs CI policy checks against either the current working tree diff or a PR-style base/head git diff.

This action is intentionally small: it does not install AgentGuard. Install AgentGuard in the workflow before invoking the action.

Copyable workflows pin remote Actions to reviewed 40-character commit SHAs. The AgentGuard action example below uses source revision 5e93b179f8add85bd4e8d5fa330f97ae1212c109, the reviewed repository commit used for the v0.3.0 documentation line; that commit contains action/action.yml. When updating an Action pin, verify the upstream release or source revision, review the diff, update the adjacent comment, and merge through the normal PR review path. Do not enable automated Action updates that merge without human review.

Inputs

Input	Default	Description
config	agentguard.yaml	Path to the AgentGuard config file.
base	empty	Base git ref for PR-style diff comparison.
head	HEAD	Head git ref for PR-style diff comparison.
github-summary	true	Whether to append a GitHub Actions step summary.
baseline-report	empty	Prior CI or PR report used to classify findings.
pr-report	empty	Explicit path for the machine-readable comparison report.
github-annotations	false	Emit bounded annotations for new findings with safe locations.
allow-fail-result	false	Whether to exit 0 even if AgentGuard policy result is FAIL.

When base is provided, the action runs:

agentguard ci --config "$config" --base "$base" --head "$head"

When base is empty, the action uses working-tree CI mode:

agentguard ci --config "$config"

Example Workflow

Use fetch-depth: 0 for base/head diff mode. AgentGuard needs the base ref and enough git history to compute the diff. The workflow declares read-only GITHUB_TOKEN permissions because checkout and AgentGuard only need repository contents; token permissions do not sandbox commands run from your repository, so review repository-controlled workflow steps with the same care as any CI code.

name: AgentGuard

on:
  pull_request:
  push:

permissions:
  contents: read

jobs:
  agentguard:
    runs-on: ubuntu-latest
    steps:
      - uses: actions/checkout@93cb6efe18208431cddfb8368fd83d5badbf9bfd # v5.0.1
        with:
          fetch-depth: 0
          persist-credentials: false

      - uses: actions/setup-python@a309ff8b426b58ec0e2a45f0f869d46889d02405 # v6.2.0
        with:
          python-version: "3.11"

      - name: Install AgentGuard
        run: python -m pip install -e ".[dev]"

      - uses: richinmrudul/agentguard/action@5e93b179f8add85bd4e8d5fa330f97ae1212c109 # reviewed source revision for v0.3.0 docs
        with:
          config: agentguard.yaml
          base: origin/main
          head: HEAD
          github-summary: "true"

Current Limitations

• The action is composite and assumes agentguard is already on PATH.
• The action does not package or install AgentGuard yet.
• Future Docker sandboxing is separate and is not part of this action.
• The action does not call GitHub APIs or create PR comments. Optional annotations are escaped workflow commands, capped at ten, and limited to new findings with safe repository-contained file and line locations.