The usage gate

Before implement or address-review starts a run, its gate (Implement.gate) checks the account, then the usage gate reads current usage and decides whether one more task fits. review and re-review have no gate, and neither does a human-initiated flow, since a person starts each of its runs. Every refusal is recorded as the tick’s skip reason.

Agent	Usage comes from
claude	Claude’s usage endpoint, with the account’s login token
codex	the rate limits its own sessions log under $CODEX_HOME/sessions
opencode	nothing; its runs aren’t gated on usage

The decision

Usage is a fraction of each window’s quota, 0–1.

1. Five-hour ceiling. Refuse when five-hour usage ≥ fiveHourCeiling.

2. Weekly headroom. Otherwise compute

   headroom = (targetEndUsage − weekly used) − humanRatePerHour × hours until weekly reset

   and launch only when headroom ≥ taskCostEstimate.

So the gate spends toward targetEndUsage by the end of the week while leaving room for the use you’ll add yourself.

Key	Default	Meaning
targetEndUsage	0.95	weekly usage to aim for at reset
taskCostEstimate	0.04	weekly quota one task is expected to cost
fiveHourCeiling	0.4	five-hour usage at which no task starts
humanRatePerHour	0.005	weekly quota you use per hour, reserved for you

~/.config/chernabog/config.yaml

usageGate:
  targetEndUsage: 0.95
  taskCostEstimate: 0.04
  fiveHourCeiling: 0.4
  humanRatePerHour: 0.005

usageGate is all four keys or none. It applies to every account on the machine.

An account can opt out with usageGate: none, such as one that logs in with a token from claude setup-token, which cannot read usage. Its runs launch with the account still checked, and only the subscription’s own limits stop a session:

accounts:
  ci: { oauthTokenEnv: CLAUDE_CODE_OAUTH_TOKEN, usageGate: none }

Example. 60 % used, 50 h to reset: 0.95 − 0.60 − 0.005 × 50 = 0.10, which covers a 0.04 task.