pydantic-ai 2.43.0 fixes a Temporal tool opt-out bug

pydantic-ai 2.43.0 changes how Temporal tool opt-out decisions get made. Before this release, whether a tool was skipped depended in part on whether MCP had been imported. MCP had no business deciding runtime behavior for an unrelated tool check. The new version keys the check on operation kind instead, per pydantic-ai's 2.43.0 release notes.

That's hard to catch in testing. If a test environment imports MCP for some other reason and production doesn't, or the reverse, a tool opt-out that worked in one environment silently stops working in the other. Nothing throws an error. The behavior just changes underneath you.

The same release fixes text part boundaries in OpenAIChatModel after tool calls. Boundaries were getting mangled when parsing interleaved text and tool output from that model. That's a correctness fix, not a new feature, so anything downstream that assumed clean boundaries was working around a bug it didn't know it had.

Separately, Logfire 5.1.0 fixes an OpenTelemetry logging deadlock inside force_flush(), according to Logfire's 5.1.0 release notes. Flushing traces at process exit in an agent run is exactly the code path this touches. A deadlock there looks like a hung process with no useful stack trace, which makes it expensive to diagnose blind.

Neither release note says how long the MCP-import coupling existed before someone noticed, or how many Temporal workflows shipped with it.