[alex-clickhouse]

Stacked on #542. After setup in the browser, the agent is set up by talking to it, in both
modes. A worker set up without the browser with a task description still sets itself up.

What changes:

• One onboarding conversation. Finish opens one web chat session, "Onboarding", for the
  person who finished setup. The agent speaks first: a hidden first message (not in the
  transcript) tells it to say that it is set up by talking to it, offer what it can help
  with (personal installs lead with the personal assistant), and ask which items apply.
  For engineering work it asks which repositories to onboard and studies them. It proposes
  cron jobs with gates, the plan-and-approve loop, statuses (the built-in backlog and
  plan_review), tags, skills and, for a worker, the TASK.md sections. It changes nothing
  before the person confirms, then applies, summarizes and calls the new onboarding_done
  tool.
• Without the browser. nerve init --terminal or --non-interactive with a task
  description records an autonomous onboarding. At the next start with a model, one session
  without a person reads the description and the worker guide, fills TASK.md, creates
  skills, cron jobs and tasks, notifies the operator, and calls onboarding_done. It runs
  once. With an empty description, the worker waits for the conversation in the web UI.
  The terminal prompt is optional and says which path follows.
• Onboarding state. Migration v060 adds one row: requested, running, done or skipped,
  and the kind (conversation or autonomous). Finish is idempotent (also two clicks at once)
  and always opens the conversation. With a pending restart, the next daemon start opens
  it. A worker without the old TASK.md marker stays done. The "ended without a plan"
  state and "Try again" are removed.
• Mode step. Personal "works for you", worker "works for a team"; an optional note
  instead of the required task description and the starter buttons.
• Do not miss it. The Ready screen leads into the conversation. Until onboarding is done
  or skipped, a banner on every page except the conversation says the agent waits, with
  Continue (it can be hidden until the next page). When setup is complete and nobody pressed
  Finish (also an install from before), the banner offers Start, which runs Finish and opens
  the conversation. While a worker sets itself up, the banner only links to that session.
  The setup page offers Start and Skip onboarding (with a confirmation). Every signed-in
  person sees the banner, because the conversation sets up the shared agent.
• Effective job state. The Automation step shows each built-in job as the scheduler
  runs it: system.yaml with jobs.yaml jobs of the same id replacing it (the same
  merge). Such a job is read-only there, and saving the step leaves it alone.

Verification: backend tests for the changed code, npm run build, Vitest for the changed
specs, migration number check; CI runs the full suites. Screenshots of the Mode step, Ready
screen, banner, setup page and the first agent message (a real run), in light and dark.

Known gaps: the hidden instructions can be shortened by a context compaction. A failed
first turn keeps its session with the error, and onboarding goes back to pending so the
next Finish or start opens a new one. An autonomous session cut by a restart stays open;
the banner links to it.

🤖 Generated with Claude Code