create-agent

Create Agent

ℹ️ First, set expectations: before anything else, print one short line with this skill's version and its most recent change — the top entry of metadata.changelog above — e.g. create-agent vX.Y — recent: <summary>. Then proceed.

Scaffold a complete Claude Code agent from scratch. The agent will be Trinity-compatible and ready for development with playbook-based skill creation.

HEADLESS MODE (read first)

You are headless when the AskUserQuestion tool is not available — typically when this wizard runs inside a deployed Trinity agent (a builder agent) rather than a person's Claude Code session. Interactive sessions ignore this section entirely.

In headless mode every AskUserQuestion below is replaced by these rules:

1. The brief is the answer. Take purpose, name, destination, skills, plugins, credentials and schedules from the argument. Destination defaults to ~/builds/[agent-name]; plugins default to agent-dev + trinity.
2. Decorative choices take the default silently — the avatar prompt, widget layout, wording. Never ask about them.
3. Real gaps go out as ONE ask, never as guesses. A real gap is something the generated agent cannot work without and the brief does not say: the data sources it watches, the accounts it posts to, a currency or locale, an external system's name. Collect every gap first, then raise a single mcp__trinity__ask_operator — request_id: "create-agent:[agent-name]:gaps", type: "question", title: "[Agent Display Name] — N details needed before I build it", question: the numbered gaps, each with the default you would use. Write ~/builds/[agent-name].pending.md (the brief + the request_id + the gap list), report that you are waiting on the ask, and end the turn. On the next invocation for the same agent, read the pending file and mcp__trinity__get_my_ask(request_id): answered → merge the answers and continue from Step 2; cancelled or dismissed → stop and say so (do not raise the same ask again); expired → continue with the stated defaults and list them as assumptions in the Completion summary. Keep the title ≤120 characters and the question ≤4,000 (longer is refused); the gaps are one decision — "confirm or correct these defaults" — never a list of unrelated approvals. Never write REPLACE-ME placeholders for a gap you could have asked about.
4. One approval before anything leaves the workspace. The Step 6 "present each skill outline" gate and the Step 13/14 questions fold into a single approval ask at the end — see Step 14 → Headless. Nothing is pushed or created before it is approved.
5. Schedules stay declared enabled: false (the default below). The created agent is non-production until a person reviews it.