Skip to main content

Context Passport and Burst Guard

Context Passport persists bounded logical inference-session identity and best-effort affinity hints:
Send the returned ID as X-InferCrane-Context-Passport. The in-memory route snapshot prefers the hinted binding or target only while it remains eligible; otherwise ordinary healthy routing wins and the response reports X-InferCrane-Affinity: fallback.
Context Passport does not store conversation bodies and does not guarantee durable KV state. Logical identity survives a worker; engine cache survival does not unless a separately qualified backend provides it.

Request survival

Request survival is a delegated capability contract. A runtime/backend may claim support only with a named mechanism and local or real qualification evidence. Plain vLLM does not gain request migration from InferCrane; unsupported and unknown states fail closed. InferCrane never implements distributed KV or token migration.

Burst Guard

Burst Guard combines fresh queue evidence, sustained breach/recovery intervals, qualified external health, and a hard incremental hourly cost ceiling:
It requires the existing governed external fallback policy, including privacy acknowledgement and its hard request/cost budget. Stale signals, unavailable capacity, or cost above policy produce a fail-closed decision. The evaluator persists evidence; route mutation remains owned by the existing policy controller.