agentsgen

AGENTS.md Generator (agentsgen)

Generate repo instructions, preserve handwritten rules, and catch stale command references in pull requests.

Start in one minute

Requires Python 3.11+. Install it as an isolated CLI with uv (recommended) or pipx. Run inside your repository:

uv tool install agentsgen

agentsgen init . --defaults --autodetect
agentsgen check . --ci

Creates .agentsgen.json, AGENTS.md, and RUNBOOK.md (plus starter prompts). Review the detected commands in .agentsgen.json before committing. No API key is required for these commands.

Prefer pipx? Run pipx install agentsgen.

Already have handwritten instructions? The original stays unchanged and proposals appear in AGENTS.generated.md / RUNBOOK.generated.md. Review and copy the sections you want into your original file, retaining their markers. Until then, check reports the unmanaged document; it does not silently adopt it.

See what it catches

package.json: remove scripts.test
AGENTS.md:    still recommends npm test
agentsgen check . --ci
→ DRIFT (exit 1)

The JSON report identifies the missing script and asks you to review .agentsgen.json. The checker reads files; it never runs your project commands.

Three reproducible demos cover handwritten preservation, stale-command detection, and repeatable setup without a README.

Safe updates

agentsgen update . --dry-run --print-diff
agentsgen update .
agentsgen check . --report

Only AGENTSGEN marker sections are regenerated. Text outside them stays yours. Missing files are created; existing files without markers get generated siblings. Update config when your toolchain changes; fix cannot infer your intended replacement command.

Add a pull-request guard

name: Agent instructions
on: [push, pull_request]
permissions:
  contents: read
jobs:
  check:
    runs-on: ubuntu-latest
    steps:
      - uses: actions/checkout@v7
      - uses: markoblogo/AGENTS.md_generator/.github/actions/agentsgen-guard@v0.5.1

For immutable dependencies, pin the action to the reviewed release commit SHA. Advanced guard options include optional pack checks and PR comments.

Optional context bundle

agentsgen pack . --autodetect
agentsgen check . --pack-check --ci

The bundle provides command manifests and on-demand documents under docs/ai/. Use --all only when you want pack and README snippet validation too. Without README.md, aggregate snippet checks are explicitly skipped.

What is verified

Compound commands, custom executables, dynamic Make targets, and patterns are reported as not verified. Existence does not prove that a command succeeds. Run your actual tests separately. check --format json includes warning details. The readiness score is a heuristic, not a measure of AI task quality.

ABVX ecosystem

One repository: install agentsgen directly; SET is optional. Workflow orchestration: SET adds presets, repo-local planning configuration, and reviewable workflow exports. The tested stack is agentsgen 0.5.1 + SET 0.4.0 + ID 0.5.2. See the SET guide and ecosystem integration guide.

abvx-agent-skills supplies reusable workflows; it is not installed or required by this CLI. ID supplies portable human context, with docs/ai/id-context.json as the explicit bridge. Git Tweet can publish a completed GitHub Release downstream; it is not part of generation or CI.

Documentation and contributions

• CLI and experimental features: fleet, repo maps, MCP, proof artifacts, reflection and exports.
• Recipes: Python, Next.js, monorepo and Node CLI starting configs.
• Compatibility and limits: reading a file is distinct from native client integration.
• Release checklist and contributing.

Report a bug with the package version, a minimal repo fixture, the command, and expected/actual output. Remove secrets before sharing a fixture.

Contributor setup: python3 -m venv .venv, activate it, then pip install -e ".[dev]". Run pytest -q, ruff check ., ruff format --check ., and mypy src.