Pydantic AI 2.67.0 adds KeenableSearch and moves native-tool settings

The release starts with compatibility notes. Provider-specific native-tool fields now live under provider_settings, and RepoContext loads one instruction file per directory by default. Inlined nested instructions are capped at 32,000 characters per conversation unless RepoContext(nested_max_chars=...) changes the limit.

New features include the KeenableSearch capability, dequeue and set_enqueue_priority for changing a queued message before delivery, and a client option on ModalSandbox and E2BSandbox. Model support now covers Claude Fable 5.2 and Claude Mythos 5.2, Gemini 4 Carbon on GoogleModel, and OpenAI's gpt-image-2.6-flare and gpt-image-2.6-sunburst image models.

Durable execution coverage grows as well, with SubAgents(dynamic_agents=...) supported under durable execution and tool calls runnable as DBOS child workflows. Bug fixes cover Shell job handling, Gemini quota errors and streamed Mistral tool-call deltas.

Code that sets native-tool options directly on agents needs to move those fields into provider_settings. Teams using RepoContext should check the new one-file-per-directory loading and the 32,000-character cap on nested instructions. Agents built on durable execution can now use dynamic subagents, which may change how their runs replay.