The Kanban Crew — Orchestrator, Specialist, Judge

The Crew turns your Sprint Board into a self-checking assembly line. Pick one model for each of three roles — an Orchestrator to plan, a Specialist to do the work, and a Judge to grade it — and the board runs each card until the Judge is satisfied. It’s the Vibe Code workflow’s judged loop, wired straight into the board, with a live token + cost estimate on every card so you always see the bill before you spend it.

The controls live in the crew bar across the top of the Kanban pane.

The three roles

• Orchestrator — a strong reasoning model, used sparingly. It breaks a goal into tasks with acceptance criteria and does a cheap self-check on each result before spending a Judge pass.
• Specialist — the worker. It’s one of your open agent windows (Claude, Codex, Grok…), dispatched the task exactly like a normal card assignment. Pick the cheapest model that can clear the bar.
• Judge — grades each result against its acceptance criteria on an absolute rubric (correctness, consistency, security, bugs) and returns a score plus a list of defects.

Setting up the crew

1. Open the board (open kanban or Ctrl+Shift+K) and name the project (pencil at the top).
2. In the crew bar, fill Orchestrator, Specialist, and Judge — each field is typeable: enter an open agent window’s name (Alex), a CLI (Claude Code), or any model from the roster, with suggestions as you type. Or click Auto to let the Model Council assign roles for you (see below). Typing an agent as Specialist makes that window the board’s preferred worker. A CLI role (Claude Code, Codex, Gemini…) gets a second dropdown to pin the exact model it runs with — pick Claude Code and then, say, Fable 5 or Opus 4.8; leave it on default model to use the CLI’s own setting.
3. Set the Pass ≥ bar (default 95%) and Attempts cap (default 3) to taste.

Manual vs. Auto

• Manual — you pin each role yourself, choosing from your detected CLIs and the OpenRouter catalog.
• Auto — a short representative probe runs through the Council arena, and the blind judge’s scores assign the roles: the strongest reasoner orchestrates, a cheap-but-capable model works, and a strong (or local) model judges.

Running work

Break a goal into cards — with an Orchestrator picked and the project named, click Decompose. It splits your project goal into cards, each with its own acceptance criteria.

Run the whole board — click Run crew to send every unfinished card through the loop: the Specialist attempts it, the Judge grades it, and a card that misses the bar is refined and retried up to the Attempts cap. You need a Judge model and at least one open agent window.

Run a single card — open any card to set its Size (S/M/L, which drives the estimate) and Acceptance criteria — what does “done to 95%” look like? — then use Run with Crew right there in the card.

Each card ships only when the Judge clears it; at the attempt cap the crew surfaces its best-so-far rather than silently accepting it.

Reading the cost

• Per card — the card modal shows a live ~$0.42 · ~38k tok estimate that updates with the task size and your chosen models.
• Whole board — Board est. in the crew bar totals every unfinished card, plus one-time planning and the final Judge pass.
• Costs from CLI workers are flagged est — estimated at API-equivalent rates, since your CLI subscription is what actually bills them. OpenRouter roles are metered exactly. Real runs feed a calibration table, so the numbers get sharper the more you use it.

Why it beats a plain dispatch

• Quality is gated, not hoped-for: nothing is marked done until the Judge scores it past your bar, and a security or correctness defect blocks it outright.
• Spend is right-sized: each task runs on the cheapest capable model, a cheap self-check runs before the expensive Judge pass, and only failing tasks are re-run — never the whole board.
• You see the bill first: every card and the board as a whole carry a token + cost estimate before a single run starts.

Crew activity — runs, pass-rate, average attempts, and average score — also rolls up into the Command Center, so you can see how the board is performing over time.