Changes since deepagents==0.3.12
New integration packages for pluggable sandboxes: langchain-modal, langchain-daytona, and langchain-runloop. See sandboxes guide and example data analysis tutorial.
Notable changes from #1234
- Refactor the SDK's summarization middleware.
- Summarization now happens in the model node via wrap_model_call events. Due to this we retain the full message history in the graph state.
- System messages and tool definitions are accurately counted toward summarization triggers.
- Add exception handling for context overflows:
ChatOpenAIandChatAnthropicnow raise a standard LangChain exception when input exceeds a model's token limit. We catch this error and trigger summarization.
- When model strings in
create_deep_agentare prefixed with'openai:', we default to the Responses API.
chore(sdk): add support for , in allowed-tools, fail gently for bullet list (#1235)
fix(sdk): relax skills parsing for allowed tools (#1232)
chore(sdk)!: move sandbox provider back to cli (#1226)
refactor(sdk): move function-level test imports out to top (#1168)
fix(sdk): harden skills metadata parsing, improve spec compliance (#1189)
style: update brand names (#1180)
Thanks to our contributors: @sydney-runkle, @eyurtsev, @mdrxy