line-design

A line is standing work the station runs unwatched: a job lands in the INBOX, rides the belts through BAYS, and leaves at the OUTBOX. Each bay is one agent doing one step from one brief. A line scheduled untested fails at 3 a.m. with nobody there — so the design ends with a test, never with the schedule.

You design and verify; the Commander builds. station.layout is read-only: bays, belts, briefs and schedules are placed in REFIT (Build mode) through the Workflow panel.

Method

1. Read the floor first (station.layout). The lines that exist, whether routing is live, whether automation is stopped (E-STOP), bays on no line, who holds which workstation. Extend an existing line before proposing a new one.
2. Name the job and the finished output in one sentence each. "Every weekday: this week's research on X in, a 200-word sourced brief out." If the output is vague, the line has nothing to converge on.
3. Cut the stages. A stage earns a bay only when it needs a different skill, different gear, or an independent check. Common shapes: RESEARCHER → WRITER → REVIEWER, or ENGINEER → TESTER. The LINE BUDGET caps stages after the first (6 by default); fewer bays cost less.
4. Name each handoff before any brief. For every belt, state what crosses it — "bullet notes with links", "a 200-word draft". That phrase goes in the bay's HANDS OFF field, and the next brief is written against it. A handoff you cannot name is a seam you have not found.
5. Write each brief as a standing order for arriving work. The DOES field (up to 2000 characters): what the bay receives, what it does, the output shape, and what to do when the input is thin — say so, never invent.
6. Add a loop only with a real verdict. A reviewer bay before a LOOP gate ends its output with VERDICT: approved or VERDICT: revise plus what to fix; the gate reads only the last three non-empty lines. Set max tries and an escalation lane (a FIXER bay) so a draft that never passes still goes somewhere.
7. Check gear per bay. Each bay's agent needs an assigned workstation in that room (no workstation, no compute) plus the props its step uses — a DISH for web, an INTEL CAB for files. station.layout with line set shows each bay's tools.
8. Hand the Commander a build sheet. One row per bay: role, agent, DOES, HANDS OFF, gear. Then gates, OUTBOX, and the INBOX trigger (schedule, channel, watched folder or webhook).
9. Test cheapest first. The test controls are the Commander's to press; ask in this order. WATCH IT is free — a crate rides the belts and each machine says what it would do. Then TEST THIS STEP on bay 1 with a realistic test job (one real run, real cost), carrying each output forward bay by bay — or STEP THROUGH, which pauses at every handoff and delivers nothing. RUN ONE REAL JOB last: end to end, into the OUTBOX. Fix the brief behind any bad output and retest that bay.
10. Only then schedule it. The schedule is saved on the INBOX and fires only while scheduling is on (TURN SCHEDULING ON). routine.create schedules one agent, not a line.