azure-ai-agentserver-invocations_1.3.0b1

1.3.0b1 (2026-10-26)

Features Added

• Capture invocation-local original-subject assertions for HTTP, WebSocket and
  typed Voice handlers. Model-header propagation is an explicit per-call opt-in.
• Added immutable InputDataPart and ResponseOutputDataDone for structured application data in Bridge protocol 1.0. Existing callbacks and explicit ResponseNone silent completion remain application-owned.
• Added a standalone /voice_conversation_engine WebSocket route to the experimental VoiceAgentServerHost, reusing the existing typed Bridge event relay alongside /invocations_ws. Existing invocations_ws behavior and generic InvocationAgentServerHost raw WebSocket handlers are unchanged.
• Aligned the experimental Voice caller-interruption contract with user.barge_in, UserBargeIn, UserBargeInCallback, and VoiceAgentServerHost.on_user_barge_in, preserving playback snapshots and application-owned cancellation/history. Voice Live senders must use the same event contract.

Bugs Fixed

• Reject malformed platform attribution before binding HTTP context, and
  finalize WebSocket connections with a classified error on invalid metadata.
• Upgrade both typed Voice WebSocket endpoints before closing with code 1011
  on malformed platform attribution, instead of rejecting the HTTP handshake.
• Apply the Invocations task-runtime switch to both typed Voice WebSocket
  endpoints, including composed Voice/Responses hosts where Responses
  independently enables the task manager.

Other Changes

• Raised the minimum azure-ai-agentserver-core dependency to 2.3.0b2 for
  protocol-scoped task-runtime support and original-subject attribution.
  Publish that Core version before this package.
• Invocations continues to opt into resilient tasks with
  set_resilient_tasks_enabled(True) before host startup. Responses
  protocol-scoped task-runtime activation does not change that global switch,
  so enabling Responses does not implicitly activate Invocations.