✦ Features 5
- export key-aware extractor and make it model-configurablemodellanguage
- add cache_key to CacheConfig for key-routed cache providersmodel
- add internal InProcessTaskEngineasynclanguage
- add sessionId getter to Agentdevxsessions
- add model routingmodellanguage
⊘ Fixes 11
- count Gemini tool-use tokens as input and thinking tokens as outputmodel
- redact blocked content when guardrail trace is disabledmodelinterventions
- harden vercel document testmodel
- harden mantle routing integ tests
- restore always() in upload-metrics if guard
- grant Bidi integration test permissionsbidirectional-streaming
- emit semconv-compliant cache usage attributesotel
- forward cancellation signal to OpenAIasyncmodel
- support content blocks in AfterToolsEvent.endTurn and simplify delegationmultiagenthooks
- treat empty integration test reports as a no-op
- count cached tokens in context-size and compaction baselinecontext