onboard-agentlisted

<!-- User-invoked: registering an agent is the human's decision and its timing is theirs, so this skill pays no permanent context for a reach nobody uses. The description above is written for the human who types it, not for a model deciding whether to fire it. --> # Onboarding a new agent into the crew - **Can:** probe a CLI's capabilities, run the containment battery, build the wrapper and register the agent. - **Must:** run the whole battery — including the commit test and the fabrication test — and record the scorecard as evidence. - **Cannot:** register an agent that has not passed, and cannot grant one git access or the complex tier. Adding a CLI agent as an MCP server is mostly mechanical. **The part that is not mechanical is proving it stays in its lane** — and that is the part everyone skips, because a delegated agent that misbehaves still returns a confident, well-formatted success message. This skill is the procedure, and its centre is a **test battery you run before trusting anything**. > Every claim in this file was measured while onboarding `agy` and `copilot`. The failures described > are real failures observed on real CLIs, not hypotheticals. ## Rule zero — the battery is not optional **An agent that has not passed §3 does not get registered.** Not "registered with a warning", not "registered for read-only". A CLI you have not tested is a CLI whose containment you are assuming, and the whole point of this kit is that assumed guarantees are the