fix(anthropic): execute bash + file tools via tool node #33960
Conversation
github-actions
bot
added
integration
CodSpeed Performance ReportMerging #33960 will not alter performanceComparing Summary
Footnotes
|
mdrxy
changed the title
github-actions
bot
added
langchain
* properly register shell tool w/ tool node * inherit this improvement in claude shell tool middleware as well * allow for configuration of tool name (which allows for claude impl to use `bash` as required by the anthropic api)
| execution_policy: BaseExecutionPolicy | None = None, | ||
| redaction_rules: tuple[RedactionRule, ...] | list[RedactionRule] | None = None, | ||
| tool_description: str | None = None, | ||
| tool_name: str = SHELL_TOOL_NAME, |
There was a problem hiding this comment.
nit: inline the tool name perhaps? it's only used in one place
There was a problem hiding this comment.
I have it as an arg here so that we can override it in the anthropic implementation
There was a problem hiding this comment.
i meant to not use SHELL_TOOL_NAME and instead tool_name = "shell" , a global variable for one use case feels a bit odd
| AgentState, | ||
| ModelRequest, | ||
| ModelResponse, | ||
| _ModelRequestOverrides, |
There was a problem hiding this comment.
Do we need to make this public?
There was a problem hiding this comment.
good question, I think no for now
There was a problem hiding this comment.
I feel comfortable importing this in a partner lib
* use `override` instead of directly patching things on `ModelRequest` * rely on `ToolNode` for execution of tools related to said middleware, using `wrap_model_call` to inject the relevant claude tool specs + allowing tool node to forward them along to corresponding langchain tool implementations * making the same change for the native shell tool middleware * allowing shell tool middleware to specify a name for the shell tool (negative diff then for claude bash middleware) long term I think the solution might be to attach metadata to a tool to map the provider spec to a langchain implementation, which we could also take some lessons from on the MCP front.
overrideinstead of directly patching things onModelRequestToolNodefor execution of tools related to said middleware, usingwrap_model_callto inject the relevant claude tool specs + allowing tool node to forward them along to corresponding langchain tool implementationslong term I think the solution might be to attach metadata to a tool to map the provider spec to a langchain implementation, which we could also take some lessons from on the MCP front.