Uak Start

---
name: uak-start
description: "Use at the start of a session with no task in hand: take the next task (uak next) and get its worktree ready, claim included. Run this only when the user invokes it by name ($uak-start or /uak-start); never start it on your own."
disable-model-invocation: true
---
<!-- Generated by scripts/gen_harness.py from .claude/commands/uak-start.md; edit that file, not this one. -->
Arguments (what the user typed after the command, `<arguments>` below): [ID] [agent-name]
Goal: from zero to editing code in under 2 min. Arguments: <arguments>
1. Already in a worktree with `.uak-env` and its task still CLAIMED? Follow its `Next` (`bash .uak/bin/uak status --task ID --summary`, ID from `.uak-env`). Do not claim another.
2. Otherwise run `bash .uak/bin/uak next` and take its candidate (or the argument). None? Its `Why not` lines say what blocks each task; run `bash .uak/bin/uak start` for every missing step and pass the human-only ones (approve the plan, push to main) to the human. Read only its section: `grep -n -A25 '^## ID' .uak/TASKS.md`. If it is not there, it came via `uak plan`, and `uak status --task ID` shows Verify and Paths.
3. Identity: the argument or `UAK_AGENT`. Otherwise ask ONCE (format `person-N`).
4. `bash .uak/bin/wt new ID --next "<exact first step>"` (no `--agent`: it reuses a free identity of this machine or takes the next number). On CONFLICT (exit 3), go back to `next` and pick another; don't retry.
5. Ask the human to `cd <worktree>` and open the agent there, or continue there yourself if you can.
6. Read the task thread: `bash .uak/bin/uak inbox --task ID`.
7. Get context cheaply: send up to 3 `uak-scout` subagents in parallel with precise questions (where is X, which tests cover Y), instead of reading files yourself. If your plan touches more than 2 files or a contract, consult the advisor (or `/uak-advise plan`) before editing.
8. Run `/uak-loop` on the first criterion. When it is green, commit and push the branch.