fix(langchain): relax tool node validation to allow unregistered tools #9236
Conversation
This change allows middleware to intercept and handle tool calls for tools that aren't registered with ToolNode, enabling support for schema-less tools like Anthropic's text editor and memory tools. Key changes: - Updated ToolCallRequest.tool type to allow undefined - Deferred tool validation to allow wrapToolCall middleware to intercept - Added comprehensive tests for unregistered tool handling This is the JS equivalent of Python PR langchain-ai/langchain#33512 🤖 Generated with [Claude Code](https://claude.com/claude-code) Co-Authored-By: Claude <[email protected]>
|
| /** | ||
| * skip as test requires primitives from `@langchain/core` that aren't released yet | ||
| * and fails in dependency range tests, remove after next release | ||
| */ | ||
| if (process.env.LC_DEPENDENCY_RANGE_TESTS) { | ||
| return; | ||
| } |
There was a problem hiding this comment.
is this necessary here? claude code thought it was but I don't have enough context to have an opinion
There was a problem hiding this comment.
It shouldn't as we have published @langchain/core by now. We can remove all these occurrences. Long term we will revisit these environment tests and disable them if a PR contains both, changes to langchain and @langchain/core.
Use the simpler array format [AIMessage(...)] for ToolNode.invoke() calls,
matching the Python test structure, instead of unnecessarily wrapping in
{ messages: [...] }.
🤖 Generated with [Claude Code](https://claude.com/claude-code)
Co-Authored-By: Claude <[email protected]>
Add test for dict format { messages: [...] } to verify both input formats
work with unregistered tool handling, matching Python's
test_interceptor_with_dict_input_format.
🤖 Generated with [Claude Code](https://claude.com/claude-code)
Co-Authored-By: Claude <[email protected]>
| /** | ||
| * skip as test requires primitives from `@langchain/core` that aren't released yet | ||
| * and fails in dependency range tests, remove after next release | ||
| */ | ||
| if (process.env.LC_DEPENDENCY_RANGE_TESTS) { | ||
| return; | ||
| } |
There was a problem hiding this comment.
It shouldn't as we have published @langchain/core by now. We can remove all these occurrences. Long term we will revisit these environment tests and disable them if a PR contains both, changes to langchain and @langchain/core.
2 participants
Relax tool node validation to allow middleware to intercept and handle unregistered tools.
This is the JS equivalent of Python PR langchain-ai/langchain#33512
Background
Some tools, like Anthropic's text editor and memory tools, use "schema-less" tool definitions that are added to model requests but not registered as actual tools with ToolNode. These tools exist only as type identifiers (e.g.,
{"type": "text_editor_20250728"}) that the Anthropic API understands natively.Previously, ToolNode would throw an error immediately if a tool wasn't registered, before middleware could intercept the call. This prevented middleware from handling these unregistered tools.
Changes
Type Updates
ToolCallRequest.tooltype fromClientTool | ServerTooltoClientTool | ServerTool | undefinedundefinedindicates an unregistered toolToolNode Changes
runTool(): Removed immediate error throw after tool lookup - undefined tools can now proceed to middlewarebaseHandler: Usestooldirectly from request instead of redundant lookup. Added validation that returns error ToolMessage (instead of throwing) if tool is undefinedThis allows
wrapToolCallmiddleware to intercept unregistered tools by checkingrequest.tool === undefinedand short-circuiting without calling the handlerRelated
This is a prerequisite for implementing Anthropic text editor and memory middleware (Python PR langchain-ai/langchain#33384).