implement

Implement the work in the spec or the ticket the user names. Fetch it in full first — the fetch-spec or fetch-ticket verb — and for a ticket, fetch its parent spec too; state its title before starting, and where the reference is ambiguous, ask. The loop config translates the storage verbs: it is docs/agents/cantrips-loop.md, and when that doc is absent the plugin defaults (defaults.md) govern.

Where the spec agreed a test seam, drive /tdd at that seam — the seams are already approved, so test at them without re-asking. Where no seam was agreed, implement directly: state explicit verification criteria up front (what observable result proves it works), then verify against them before calling the work done.

Run typechecking regularly, single test files regularly, and the full test suite once at the end.

When working from a ticket, tick its acceptance criteria as each one is verified. When every criterion is verified, resolve the ticket — the resolve-ticket verb from the same loop config, the ticked criteria standing as evidence. Resolve on the criteria without second-guessing: a wrong resolve is one human reopen away. The spec itself stays as published — no skill ever closes a spec; closing the feature is the human's act through the backend's native machinery.

Done when every criterion is verified, the ticket (if any) is resolved, and the full suite passes → close with a flow pointer (the message's final paragraph, a blockquote in full italics opening Next: — or Next steps: over one bullet per pointer — each pointer naming its skill the way this skill was itself invoked, same prefix and namespace, and ending in a one-clause rationale after an em dash) naming the review tail, a chain rendered whole and in order, all in this session with the working diff as context:

• /simplify (user-invoked) — an optional quality pass, rendered as the command to type with one proof-test:<path> token per proving test of the whole run and bare where the run has none: a proving test is a /tdd test at an agreed seam or a test the agent wrote or changed and ran to tick a criterion, so a test never run for proof, a pre-existing test that merely passed, and a criterion verified by hand name none.
• /review-gate [--fix | --loop] (user-invoked) — suggest low for a trivial or mechanical diff, high for a large, cross-cutting, or risky one, medium otherwise; pair any level with --fix to apply the findings once, or with --loop to converge the gate to green.
• /commit (user-invoked) — once the gate stands.

Where this session engaged a deputy through /deputy and has not dismissed it, a choice is put to the deputy as ROUNDS.md says and rendered as the one step it chose, any other closing is rendered as written, and each rendered step that names a skill is then taken: one marked (user-invoked) by notifying the user where the harness has a tool for it and waiting for them to type it, any other by invoking it, notifying nobody.