yesod.work

Roadmap / Coder-Plan Aware Scheduling

Spend the plans first.
Meter second.

The design sketch behind the Coder-Plan Aware Scheduling initiative: subscription coder plans are prepaid capacity with windows and limits — the dispatcher should schedule against them like any other scarce resource.

Init-03 · Design sketch

The problem

A serious factory operator holds several subscription coder plans at once — Codex, Claude, z.ai, Kimi. Each is prepaid capacity with its own shape: usage windows, rate limits, session ceilings, reset schedules. Today the dispatcher routes by capability and metered cost tier, which quietly wastes the resource already paid for: a plan hour that expires unused is pure loss, while a metered token spent when plan capacity sat idle is an unforced error.

Treat every coder plan as a schedulable resource. Spend plan capacity first, meter second, and let the windows shape the schedule.

The spend page already shows the symptom: runs on subscription plans meter $0, so the ledger cannot even see the factory’s largest real input cost.

What it takes

Plan-usage telemetry

Per-provider adapters that know each plan’s window, limits, and remaining capacity — observed, not guessed. This is the prerequisite for everything else, and it also fixes the $0-metering blind spot: a plan-attributed run can carry a notional value even when no invoice exists.

Scheduling policies

The dispatcher gains a policy layer above the capability tiers:

  • plan-first routing — work that any plan-covered model can do goes there before a metered model is considered;
  • window packing — as a plan window nears reset with capacity left, the scheduler pulls eligible backlog forward to consume it;
  • metered overflow — metered spend becomes the explicit overflow valve, visible as a decision rather than a default.

A two-ledger cost model

Every run records both its metered cost (what an invoice would show) and its plan draw (which plan absorbed it, valued against the plan’s price). Cost-per-outcome then works across both ledgers, and the scoreboard can finally answer: which plans earn their subscription?

Routing experiments

The tier verdicts hint that models have niches — the workhorse, the heavy-lifter, the sprinter. Deliberate routing experiments across plans turn those hints into policy: which work does each plan absorb best, at what failure cost?

Where it starts

The first note is not written yet — deliberately. The initiative starts when the plan-telemetry contract is designed, because every policy above depends on trustworthy capacity signals. Until then this sheet records the intent: the factory should be as deliberate about whose capacity runs the work as it already is about which model does.

Project
Coder-Plan Scheduling
Sheet
Design sketch
Drafted
2026-08-18
Drawn by
Stephen + the factory

No factory notes yet — the initiative starts when the plan-telemetry contract is designed. Context: factory spend (where plan runs meter $0 today) and model economics.