Wrap all streaming SDK errors in `ModelHTTPError`/`ModelAPIError` (#4889)

Support `server_message_id` in `VercelAIEventStream` (#4579)

feat: smart instruction caching — Anthropic/Bedrock cache boundary at static/dynamic split (#4865)

docs: add new open source libs to capabilities section (#4957)

Add `ThreadExecutor` capability and `Agent.using_thread_executor()` (#4942)

fix: run output validators for image outputs (#4938)

Add `defer_loading` to tools and toolsets to enable Tool Search (#4090)

Remove input/output types from `operation.cost` metric (#4944)

fix: replace pytest collection with AST parsing in check_cassettes (#4949)

feat: add local `WebFetchTool` and have `WebFetch` capability use it when provider lacks builtin support (#4906)

Add `agent` to `RunContext` (#4922)

Have `ImageGeneration` capability auto-fallback to subagent with imagegen model if main model lacks imagegen (#4910)

include `gen_ai.operation.name` in agent runs (#4936)

refactor: remove OpenAI deprecated models (#4879)

Update Mistral integration to mistralai SDK v2 (#4736)