0098 — Broker picks a cheaper model for the request
Opted-in people can have a relay call sent to a cheaper model in the same family.
Decided. When a person is opted in, relay asks a cheap model on the same provider whether the request needs a light, standard, or frontier model, then substitutes a cheaper catalog model in that family when it still fits. A static routing rule wins first. A failed classification keeps the model the caller asked for.
Broker does not know job titles. The flag defaults off. A platform admin turns it on per email.
The classifier's tokens are their own usage row. Routing savings on the main call stay the requested price minus the actual cost.
Canonical source: docs/decisions/0098-broker-picks-a-cheaper-model-for-the-request.md.
0070 — HTTP tracing and durable trace context
OpenTelemetry records HTTP execution, while outbox context and workflow identity remain durable metadata.
0100 — Broker routing is on by default, and a credential is personal or service
Every caller is routed unless an admin turns it off, requests are graded by a Go check, Jev, then a chat model, and a service credential has its own budget.