agentharnesses-cli

CLI for agentharnesses.io — scaffold and manage Agent Harnesses.

Installation

pip install agentharnesses-cli

Usage

ahar --help

ahar init

Initialize a new harness in the current directory:

ahar init

Optionally specify a name (defaults to the directory name):

ahar init my-harness

Always scaffolds HARNESS.md — the one thing the standard actually requires:

my-harness/
└── HARNESS.md                       # entry point and agent identity

Everything else is opt-in, since it may not fit the target repo. With no flags, init asks interactively:

Set up this harness as:
  plain  -- just HARNESS.md, no Claude Code configuration
  claude -- also wire up .claude/ (skills, settings) for Claude Code
Choose [plain]: claude

Which skills should be installed?
  Install agent-harnesses? (progressive disclosure explorer -- ...) [Y/n]:
  Install harnessify? (surveys this repo and authors real routing for it -- ...) [Y/n]:

Choosing claude and at least one skill installs from agentharnesses/metaskill (cloned fresh at init time, so you always get the latest version):

├── .gitignore                       # excludes .claude/settings.json, metaskill session state
├── .leaf-detectors                  # skill=SKILL.md -- needed for agent-harnesses' own leaf detection
└── .claude/
    ├── settings.json                # registers the harness as a Claude Code plugin -- merged
    │                                 # into any existing settings.json, never overwritten
    └── skills/
        ├── agent-harnesses/          # progressive-disclosure exploration (if selected)
        └── harnessify/               # authors real routing for a bare repo (if selected)

For scripted/CI use, flags skip the dialogue entirely:

ahar init --plain                                    # HARNESS.md only, no prompt
ahar init --claude                                    # for Claude Code, both skills, no prompt
ahar init --claude --agent-harnesses --no-harnessify  # for Claude Code, agent-harnesses only
ahar init --metaskill                                 # deprecated alias for --claude with both skills

--agent-harnesses/--harnessify (and --metaskill) imply --claude on their own — no need to pass both. If a .claude/ folder already exists in the target directory, init treats that as a standing signal the repo wants Claude Code integration and installs both skills automatically, skipping the dialogue.

ahar validate

Validate a harness directory structure:

ahar validate ./my-harness

ahar read

Read a property from a harness's HARNESS.md frontmatter:

ahar read ./my-harness name
ahar read ./my-harness description

ahar prompt

Render a harness as prompt XML for agent injection:

ahar prompt ./my-harness

These commands are backed by harnesses-ref, the reference implementation for the Agent Harnesses standard.

Publishing

Releases are published to PyPI automatically when a version tag is pushed:

git tag v1.0.0
git push origin v1.0.0

The GitHub Actions workflow builds the package and publishes it via trusted publishing (no API token required). The version is derived from the tag via hatch-vcs.

License

Apache 2.0