Automated Release PR

0.11.3 (2026-05-20)

Full Changelog: v0.11.2...v0.11.3

Features

• added Pydantic AI sync, async, temporal integration (#359) (781dfe1)
• api: add schedule, checkpoints, and deployment endpoints (53b5c36)

Bug Fixes

• resolve lint and test failures from new endpoints (#360) (bdf129c)

This pull request is managed by Stainless's GitHub App.

The semver version number is based on included commit messages. Alternatively, you can manually set the version number in the title of this pull request.

For a better experience, it is recommended to use either rebase-merge or squash-merge when merging this pull request.

🔗 Stainless website
📚 Read the docs
🙋 Reach out for help or questions

Greptile Summary

This release (0.11.3) adds Pydantic AI integration for sync, async, and Temporal agent workflows, new API endpoints for agent schedules, deployments, and checkpoints, and a fix to the CoalescingBuffer shutdown sequence that eliminated duplicate-tail streaming events on the UI.

• Pydantic AI ADK modules: Three new modules (_pydantic_ai_sync.py, _pydantic_ai_async.py, _pydantic_ai_tracing.py) map pydantic-ai event streams to Agentex streaming events, matching the existing LangGraph convention. Tracing handler uses deterministic span IDs derived from (trace_id, tool_call_id) to support Temporal activity boundaries.
• New API resources: CheckpointsResource, DeploymentsResource, and SchedulesResource are fully generated from the OpenAPI spec and follow the established Stainless SDK patterns.
• CoalescingBuffer fix: Replaces task cancellation with graceful natural-exit shutdown (while True + if self._closed: return at loop bottom) to prevent duplicate Redis publishes caused by re-enqueueing items whose writes had already completed before the CancelledError was raised.

Confidence Score: 5/5

Safe to merge — all changes are additive new features or targeted bug fixes with no modifications to existing behaviour for current users.

The pydantic-ai modules are self-contained additions that do not touch existing code paths; the generated SDK resources follow the established Stainless patterns with no structural deviations; and the CoalescingBuffer fix is a well-motivated, well-documented improvement to the streaming shutdown sequence. The only trade-off introduced (no timeout on _on_flush in close()) is acceptable provided the underlying Redis client carries its own timeouts.

src/agentex/lib/core/services/adk/streaming.py — the CoalescingBuffer shutdown change is the most behaviour-sensitive modification in the PR and warrants a second set of eyes on the Redis client timeout configuration.

Important Files Changed

Sequence Diagram

sequenceDiagram
    participant Agent as Pydantic AI Agent
    participant Helper as stream_pydantic_ai_events / convert_pydantic_ai_to_agentex_events
    participant Redis as CoalescingBuffer / Redis
    participant Agentex as Agentex API

    Agent->>Helper: PartStartEvent(TextPart)
    Helper->>Agentex: "messages.create (streaming_status=IN_PROGRESS)"
    Helper->>Redis: StreamTaskMessageStart

    Agent->>Helper: PartDeltaEvent(TextPartDelta)
    Helper->>Redis: StreamTaskMessageDelta (buffered/coalesced)

    Agent->>Helper: PartEndEvent
    Helper->>Redis: CoalescingBuffer.close() final flush
    Helper->>Redis: StreamTaskMessageDone
    Helper->>Agentex: messages.update (final content)

    Agent->>Helper: PartStartEvent(ToolCallPart)
    Note over Helper: Accumulates args until PartEndEvent

    Agent->>Helper: PartEndEvent(ToolCallPart)
    Helper->>Agentex: messages.create (ToolRequestContent, full args)
    Helper->>Agentex: spans.create (tracing_handler.on_tool_start)

    Agent->>Helper: FunctionToolResultEvent
    Helper->>Agentex: messages.create (ToolResponseContent)
    Helper->>Agentex: spans.update (tracing_handler.on_tool_end)

Fix the following 2 code review issues. Work through them one at a time, proposing concise fixes.

---

### Issue 1 of 2
src/agentex/lib/adk/_modules/_pydantic_ai_async.py:30-34
The `stream` parameter has no type annotation, which makes IDE autocompletion and static analysis significantly less useful for callers. Since `pydantic-ai` is an optional dependency and the import is lazy, the annotation can use a string forward reference to avoid an unconditional import at module load time — matching how the `tracing_handler` parameter is already handled.

```suggestion
async def stream_pydantic_ai_events(
    stream: "AsyncIterator[Any]",
    task_id: str,
    tracing_handler: "AgentexPydanticAITracingHandler | None" = None,
) -> str:
```

### Issue 2 of 2
src/agentex/lib/core/services/adk/streaming.py:211-232
**Potential indefinite hang when `_on_flush` blocks**

The old implementation called `_task.cancel()` before `await self._task`, which would interrupt a stuck `_on_flush` (e.g. a Redis write that never resolves). The new approach waits for the task to exit naturally, so `close()` will block indefinitely if `_on_flush` ever hangs. Given the motivation (avoid duplicate-tail publishing from a write that completed but whose `await` hadn't returned yet), this is a reasonable trade-off — but it implicitly relies on the Redis client having its own connection-level timeout configured. Worth confirming that the `StreamRepository` implementation's write calls carry a timeout so this path can never deadlock in production.

Reviews (3): Last reviewed commit: "release: 0.11.3" | Re-trigger Greptile