Skip to content

Routing evidence · deterministic example

Provider availability at the decision time

Provider and CLI probes, independent quota windows, observed headroom, inferred exhaustion, freshness, and catalog coverage—kept honest when evidence is incomplete.

Decision time 2026-08-02 12:00 UTCFreshness limit 15 minutesRate lookback 1 hour

Anthropic

CLI claude

Warning
CLI availability
Available
Freshness
Fresh · 2 min
Cost at decision
List price available

Quota windows · kept separate

five-hour

Near cap
Observed usage
800 / 1,000 · 80%
Observed headroom
200 tokens
Reset
15:00 UTC

ObservedProvider quota telemetry at 11:58 UTC

InferredExhaustion 14:00 UTC from 100 tokens/hour; before reset

weekly

Healthy
Observed usage
2,000 / 10,000 · 20%
Observed headroom
8,000 tokens
Reset
Aug 7 · 12:00 UTC

ObservedProvider quota telemetry at 11:58 UTC

InferredExhaustion Aug 5 · 20:00 UTC; before reset

Model cost coverage

  • claude-sonnet-5List price available

OpenAI

CLI codex

Critical
CLI availability
Unavailable
Freshness
Fresh · 1 min
Cost at decision
List price available

Quota windows

five-hour

Exhausted
Observed usage
1,100 / 1,000 · 110%
Observed headroom
0 tokens
Reset
14:00 UTC

ObservedProvider quota telemetry at 11:59 UTC

InferredNo projection; observed headroom is exhausted

Model cost coverage

  • gpt-5.6-solList price available

Google

CLI antigravity

Unknown
CLI availability
Unknown
Freshness
Stale · 2 hours
Cost at decision
Unknown model cost

Quota windows

daily

Unknown
Observed usage
10 / 1,000 · stale
Observed headroom
Not trusted
Reset
18:00 UTC

ObservedLast observation 10:00 UTC exceeds freshness limit

InferredNo projection from stale quota

Model cost coverage

  • unrecognized-modelUnknown model cost

This snapshot does not select a model. The orchestrator establishes the correctness floor first, then consumes this evidence. Quota or cost must never lower that floor. See docs/concepts/provider-availability-snapshot.md.