Helix Start

Orchestrator cloud context (check before setup)

If HELIX_ORCHESTRATOR_RUN_ID or HELIX_SKILL_CREDENTIAL_FILE is set, this skill runs in Helix Console. This section takes precedence over local setup and browser instructions below.

• First call the plugin's bin/helix orchestrator-protocol (Windows: bin/helix.cmd orchestrator-protocol). It must print 1. Otherwise stop with an actionable request to update the Helix plugin/CLI; never run an older binary's setup commands.
• Use the normal helix <surface> up, push, state, and wait commands. The CLI selects cloud transport and the assigned project automatically. Keep the same slug within this task.
• Do not start local servers or Fleet, open a browser, curl localhost, use --local, log in, or change profiles/projects. Do not read or print the credential file. Do not remove the injected environment variables.
• up returns the Console session URL. Use that URL wherever the local instructions request Board/Fleet links. Events go to Console; human responses and gate approvals still come from the person through the normal wait protocol.
• On cloud or compatibility errors, report the error. Never fall back to local mode. Outside this context, follow the local behavior below unchanged.

The front door. The user describes a job; you pick the surface it deserves, say why in one sentence, and launch it. The suite is only as good as the routing into it, and the honest answer is sometimes "no surface at all".

The seven answers

Fleet is not an answer here. It is the overview of many sessions, not a way to run one. Archive is not one either; it records a session after the fact rather than running it. If the job is "write up / archive / snapshot what happened", skip routing and invoke helix-archive directly.

Infer first; ask only what the description doesn't say

Read the job description and commit. When the pick is clear, state it with its one-sentence why and launch. No interview. Ask only when the description genuinely underdetermines the pick, at most three questions, in a single round, drawn from the three axes that decide it:

1. Who drives. Will this need your calls as we go, or mostly my work after an initial brief?
2. Is there a bar. Can we write a crisp definition of done before starting?
3. Will you stay. At the keyboard throughout, checking in occasionally, or back at the end?

Riffing with no decisions yet is spark; deciding together is canvas; a crisp bar to iterate against is loop; a one-way door with several possible shapes is arena; work too wide for one context is swarm; walking away is stream; minutes of work is none.

Arena and swarm both cost real money before they return anything, so both earn a fourth question when the pick is close: is this worth several agents at once? An arena over a settled shape, or a swarm over work one agent could read in a single pass, is an expensive way to reach the answer you already had.

Jobs that span phases

Shaping and building are often the same job. Recommend the sequence: spark to diverge if nothing is scoped, canvas to ratify the brief or the quality bar, then loop or stream to build against it. Launch only the first surface. The hand-off to the next one is that session's closing decision, not something scheduled here.

Launch

Invoke the chosen skill with the job as its topic: helix-spark, helix-canvas, helix-loop, helix-arena, helix-swarm, or helix-stream. Follow that skill's own rules from there, slug confirmation included; never guess a slug here. For none, just do the work; mention the routing only if the user asked about it.

When the answer was none: offer the archive at the end

A session that ran in the terminal leaves nothing behind but the transcript and the diff. When the work is done, look at what it produced. If it holds at least one thing worth finding again — a decision and its reason, a finding that changed the approach, a constraint discovered the hard way, an artifact someone else will need — ask, once, in one line:

Worth keeping? I can archive this session to Helix Archive — decisions, findings, artifacts — so it can be found later.

On yes, invoke helix-archive and follow its rules (it asks for a level, then curates). On no, or no answer, drop it; do not ask again in the same session.

Do not offer it when the session was only a lookup, a one-line fix, a question answered from the code, or work with no decision in it. An archive of nothing teaches the user to say no to the next one. The test is not how long the session ran but whether a future reader would open it: if you cannot name the one card that reader would want, there is nothing to offer.

If the user overrides your pick, the correction is signal: note in memory what kind of job it was and where they wanted it, so the next routing is better.