Azure/azure-sdk-for-python azure-ai-agentserver-core_2.3.0b1

on GitHub

2.3.0b1 (2026-10-15)

Features Added

• Added AgentServerHost.readiness_handler for replacing the default
  GET /readiness response with application-specific readiness logic. When no
  handler is registered, the existing healthy response remains unchanged.
• Added the private POST /_agent/before-snapshot and
  POST /_agent/after-restore lifecycle endpoints, with equivalent
  POST /_before_snapshot and POST /_after_restore aliases, for safely
  capturing and restoring hosted agent processes. Applications can register
  before_snapshot_handler and after_restore_handler callbacks; lifecycle
  calls are serialized and idempotent, and restored session environment
  overrides are applied before the after-restore callback runs. Protocol
  subclasses can extend the lifecycle through protected hooks without
  bypassing the host's required hydration and rollback behavior. Later
  materializations reset omitted overrides to their captured values and expose
  the restored session through the ambient request context during callbacks.
  Environment-backed AgentConfig values are refreshed in place before
  callbacks run, and AgentSessionContext is marked experimental.

Bugs Fixed

• Preserve an observed final terminal marker before recovery-time TTL
  compaction, keeping closed replay logs closed across subsequent restarts.
• Retain replay backing ownership through deletion: keep POSIX inode locks
  until unlink and Windows sidecar locks until log removal. Acquire Windows
  ownership before opening the log, reject stale or replaced locked descriptors,
  and retry cleanup without unlinking a successor's backing file.
• Release owned replay-file and writer-lock resources on every rehydration
  failure, including callback errors and cancellation. Surface cleanup errors
  with their original failure context and retain exact registry cleanup owners
  when resource release must be retried.
• Fail closed after compaction replaces the log but cannot establish a usable
  writer. Retire stale handles, reject further writes, wake subscribers with an
  explicit error, and reacquire ownership before deleting a failed replacement.
  Keep the original writer usable when preparation fails before replacement.
• Open existing replay logs in non-creating mode during cold lookup and
  compaction reopen, preventing concurrent removal from recreating an empty
  log. Preserve explicit creation and propagate backing-lock acquisition errors.
• Delete cold file-backed replay logs without deserializing their events, so
  malformed logs remain removable after restart. Preserve locking and retryable
  cleanup when file access, lock acquisition, or removal fails.
• Apply explicit task-list limits to the total returned records for both hosted
  and local providers, clamped to 100. Keep omitted or None limits as complete
  enumeration so recovery and response-ownership scans include every page.
• Reclaim per-ID stream lifecycle locks after their last holder or waiter exits,
  including cancellation, without splitting concurrent lookup, creation, or deletion.
• Enumerate every matching local durable task when no listing limit is
  requested, matching hosted task enumeration. Reject missing or repeated
  hosted continuation cursors instead of returning an incomplete task list.
• Restore existing file-backed replay logs during stream lookup and deletion
  after restart without creating absent logs. Serialize lookup, creation, and
  deletion for the same stream ID within the registry.
• Propagate replay-file deletion failures without installing a successful
  deletion tombstone, allowing cleanup to be retried.
• Make Windows replay lock-file removal failures retryable without closing
  an already-released file descriptor or reporting successful deletion.

Other Changes

• Removed the experimental marker (docstring note and one-time runtime warning)
  from the resilient task public APIs (TaskContext, TaskRun, @task,
  @multi_turn_task, set_resilient_tasks_enabled, task exceptions, and
  retry policy).
• Changed the default SSE keep-alive interval to 15 seconds. Set
  SSE_KEEPALIVE_INTERVAL=0 to disable keep-alive comments.