Factory docs, home
Page navigation

See the input → responsibility → output diagram.

Status: Passive single-unit capacity forecasts. Build review records accepted snapshots and current verification gates. Broader units/accounts and E36 workflows remain separate.

An account's spare capacity can be shared by several Products and the owner's own work. Capacity Planning shows the observations and subtracts each settled charge, unresolved original ceiling and reserve from every relevant quota window. It retains deficits and missing evidence instead of treating an unknown account as empty.

A wide teal bob carrying three nesting measuring cups.

Meet the quartermaster. Measure from work already reviewed. Gives a capacity range only from comparable reviewed work, with its assumptions stated. Meet the whole Plumb family.

A completion range needs comparable independently reviewed work, complete coverage and an explicit future scenario: full work cycles release successively at acceptance, stay inside the observed cost/time envelope, and use only the fixed residual allowance. Those are assumptions supplied by the caller, not confident predictions. Old unfinished observations remain old; unused reserves cannot quietly become new capacity.

Try the installed example. The agent reference gives source and verification links. A forecast starts no work and makes no promise. Multi-account forecasts, material timing, durable proposals and delivery remain outside this packet in E36.

Honest limits:

  • Observations, account scope and future scenarios are supplied by the caller; no account or provider is queried.
  • Completion ranges depend on comparable reviewed work and the stated scenario. Unknown evidence remains unknown.
  • Forecasts grant no capacity, reserve no funds and dispatch no work. CI and broader E36 acceptance remain separate gates.

Factory use

The Factory's first use reads the discovery schedule's weekly ledger and reports how many more estimated tokens the Factory would admit for one subscription account this week. It answers from the schedule as recorded and changes nothing:

node scripts/factory-capacity.ts --journal /path/to/existing.sqlite \
  --account claude --lane alpha

The first output line is the full forecast; the second says the same in plain words. Settled calls count at their reported usage. Calls still running, lost or without a usage report count at their full bound. A closed usage window counts as nothing left until its reset, and the plain words say so. A call from an earlier week that is still unresolved does not count against this week, but it keeps its lane's completion range unknown.

Without --at the forecast is as of now. An earlier --at is refused when the schedule changed after it, or when the latest commission change has no recorded time yet: the schedule is only ever read as it is now.

Honest limits. The headroom is the Factory's own weekly admission limit, not the provider's quota, which the schedule does not observe. When the schedule no longer holds every call of the week, or the account could not be admitted, the headroom is unknown rather than estimated. A completion range needs reviewed samples and stated assumptions from a caller file. Its value rises once the schedule records live use.

Shared value implementation: common contracts. Owning errors, limits, stored identities and authority remain as specified above.

Source: docs/guide/capacity-planning.md