Azure/azure-sdk-for-python azure-ai-agentserver-responses

2.2.0b2 (2026-09-24)

Added optional response metadata to ResponseEventStream.emit_failed. Invalid failure metadata is logged and
omitted so metadata validation cannot suppress the original failure response.

Closed a graceful-shutdown admission race for stored streams and bounded
DELETE waits when deferred terminal persistence is stalled.
Changed the default history fetch limit from 100 to -1 (unlimited), avoiding
automatic truncation of conversation history. Positive limits remain supported.
Restored compatibility with usage payloads that omit
ResponseUsageInputTokensDetails.cache_write_tokens.
The per-request span flush in the Responses endpoint no longer blocks the
asyncio event loop. The synchronous flush_spans() call in the request
finally block ran TracerProvider.force_flush inline, which blocks the
event loop until the exporter drains and serialises concurrent requests
behind a single export (head-of-line blocking). It now awaits the
non-blocking flush_spans_async() by default. A new AGENTSERVER_FLUSH_MODE
environment variable selects the strategy: async (default, off the event
loop), background (return the response first, flush in the background --
requires a platform drain window), or sync (legacy blocking behaviour).
Streaming requests use the same strategy for a single flush after stream
cleanup, without an additional pre-stream flush.
Scoped durable multi-turn task IDs with FOUNDRY_AGENT_SESSION_GUID when
available, preventing recreated same-name sessions from colliding with task
tombstones. Existing pre-rollout active chains remain resumable through a
legacy-ID lookup.

Optimized warm-path streaming last-byte latency: for in-process (non-resilient)
store=true streaming responses, the terminal response.completed/response.failed
event is now emitted to the client and the wire stream closed before the terminal
provider write, moving the terminal storage round-trip off the client's last-byte
path. A rare terminal-write failure now surfaces on a later GET (record stamped
storage_error) rather than on the stream. The resilient path is unchanged.
DELETE waits for pending execution and persistence before removing the response,
preventing deferred writes from recreating deleted data. Executions awaiting
their first event remain tracked for shutdown but are not publicly visible.
Reuse request-scoped history lookups and concurrent input-reference resolution without caching failed or cancelled reads.
Flush streaming telemetry after request-owned handler and iterator cleanup,
including on disconnects, and before HTTP completion instead of delaying the first stream event.
Construct generated model types on demand while preserving real TypedDict contracts and public exports.
Avoid redundant event and recovery-seed copies while retaining validation and caller-owned mutation isolation.
Raised the minimum azure-ai-agentserver-core dependency to >=2.2.0b2,
which provides the session GUID configuration and legacy task lookup used by
resilient Responses.