Azure/azure-sdk-for-python azure-ai-agentserver-responses_2.4.0b1

on GitHub

2.4.0b1 (2026-10-15)

Bugs Fixed

• Propagate source cancellation through SSE keep-alive streams instead of
  treating interrupted durable admission as successful stream completion.
• Signal accepted durable task shutdown before awaiting response execution, so
  explicit and checkpoint-based recovery exits preserve recoverable input.
  Rebind cold-recovery cancellation to the handler's exact event before
  publication, allowing cancellation to finish the tracked execution.
• Atomically close fresh response reservations and unpublished execution
  admission when shutdown takes its drain snapshot. Preserve exact accepted
  executions for terminal finalization, track handler work before its first
  event, and return the existing HTTP 503 shutdown response for late creates.
• Validate durable user partitions as strings or None before constructing
  platform, lifecycle, or runtime identities. Invalid partitions never fall
  back to anonymous storage. Fail the response only with independently proven
  original task references and current-input/deletion guards; otherwise settle
  only the malformed input and leave response, replay, and references untouched.
• Allow only the current execution's exact record or request context to finalize
  while DELETE checks eligibility, so rejected in-flight DELETE requests cannot
  suppress terminal SSE events or response persistence.
• Reject runtime publication once DELETE retains cleanup ownership, and bind
  fresh admission to its exact request context so late callbacks cannot replace
  a successor or clear a completed deletion marker.
• Partition local in-memory and file-backed responses, items, history, and
  conversation indexes by the platform user key, keeping anonymous local state
  separate.
• Release exact caller-scoped incarnation references when durable inputs settle
  before handler dispatch, preserving successor entries and admission deferrals.
• Read retained DELETE cleanup ownership through the authorized response
  provider and noncreating replay lookup, without allocating replacement streams.
• Retire stopped nonterminal runtime executions under scoped recovery admission,
  preventing early settlement or cancellation from retaining stale live state.
• Settle malformed current durable response incarnations as failed without
  re-invoking execution, while retaining deletion fences and rejecting stale inputs.
• Check all durable task owners before reclaiming empty replay or fencing DELETE,
  including older conversation tasks beyond the first listing page.
• Partition live execution, cancellation state, SSE streams, and resilient
  task identifiers by the platform user key. Different users can use the same
  public response ID independently. Duplicate creation within one user's
  partition returns HTTP 409 while execution or replay is retained.
• Require a successful user-scoped provider lookup before replaying stored SSE
  events; storage errors no longer fall through to cached replay.
• Require Core 2.3.0b1 so retained file replay is discovered after restart for
  duplicate admission, authorized replay, and deletion.
• Return an error when replay cleanup fails during DELETE and retain ownership
  state so the owner can retry cleanup.
• Reserve the caller-scoped response ID throughout DELETE cleanup so a new
  POST cannot reuse the ID before stream, runtime, and provider cleanup finish.
• Finish retained replay with the persisted terminal snapshot when crash
  recovery cannot resume execution, preventing indefinite SSE waits.
• Remove newly allocated replay streams when durable task admission fails or
  shutdown rejects the request before execution starts, allowing response ID
  reuse after cleanup. Emit startup errors only on the requesting HTTP stream
  without recreating rejected replay or closing another execution's stream.
• Release streaming create reservations and stop disconnect monitors even
  when sending HTTP headers fails before the first body iteration.
• Shield non-streaming create reservation cleanup from request cancellation
  so an interrupted request does not prevent later reuse of the response ID.
• Reclaim empty crash-abandoned file replay only when no caller-scoped runtime,
  response provider, or durable task retains ownership.
• Retain DELETE ownership through provider cleanup failures and cancellation,
  and fence deleted durable response inputs without deleting unrelated turns.
  Recovery participates in scoped admission and cannot resurrect deleted responses.
• Scope durable deletion fences and process-local recovery references by a
  private response incarnation, allowing a fresh same-ID turn on a retained
  conversation task without admitting an old deleted recovery.

Breaking Changes

• FileResponseStore now persists data under user-scoped
  partitions-v1/{anonymous|user-hash} directories. Existing files in the
  previous shared layout remain unchanged but are not read or migrated
  automatically, because their user ownership cannot be established safely.
  Back up and explicitly migrate any required local data into the correct
  partition only after independently verifying its ownership.
• Identified users now use private user-scoped stream and task keys. Existing
  shared replay logs and task chains are not adopted for those users because
  their ownership cannot be verified. Drain active work before upgrading.
  Anonymous lifecycle keys retain their previous format.