fix(tools): forward explicit nulls through the LangChain adapter

Closed
GL
10d ago
+18-10

What

StackOneTool.to_langchain() no longer drops None arguments in _run. The LangChain adapter now forwards exactly what the model sent, the same as to_pydantic_ai_tool already does.

Why

The filter was added in #199, when args_schema was a pydantic model: BaseTool then filled every unsupplied optional with None, so dropping None was the only way to tell "omitted" from "sent". Since the toolset became MCP-only, args_schema is the served JSON Schema dict. With a dict schema BaseTool passes the tool-call args through untouched, so an omitted optional is already absent. The filter now only removes something real: a null the model sent on purpose (for example to clear a nullable field), which the server never sees.

Probe against the locked langchain-core (1.6.3):

Model sent_run received before_run receives now
{"path_id": "x"}{"path_id": "x"}{"path_id": "x"}
{"path_id": "x", "body_after": null}{"path_id": "x"}{"path_id": "x", "body_after": null}

Risk

A caller that relied on LangChain stripping explicit nulls will now send them. Header arguments are unaffected: a null header value is still dropped by the header sanitiser before the call.

How verified

  • Replaced the old test (which asserted the drop) with a parametrized test_langchain_forwards_the_model_arguments_as_sent that invokes the converted tool with a real tool call; the null case fails on main and passes here.
  • uv run ruff check ., uv run ruff format --check ., uv run ty check stackone_ai examples, uv run pytest -q: 872 passed.

Gates (head c67101b)

  • Forge: ran, slop meter +17 -10, no findings.
  • Security scan: ran, traced None through the nested and flat header sanitising paths; nulls are dropped there, other arguments were already forwarded verbatim.
  • Codex security review: ran, no findings.
  • /security-review: ran on the same scope, no findings.

Requested by guillaume@stackone.com in Slack (repo audit thread)

šŸ¤– Generated with Claude Code


Summary by cubic

Makes the LangChain adapter forward explicit nulls to the tool call instead of dropping them, matching what to_pydantic_ai_tool already does. Previously every None argument was filtered because a pydantic args schema materialised unsupplied optionals as None; now the schema is a JSON Schema dict, so BaseTool passes the tool-call arguments through untouched — an omitted optional stays absent, and a null the model sent on purpose reaches the server.

  • Replaces the old test (which asserted the drop) with parametrized tests invoking the converted tool with a real tool call; the null case fails on main and passes here.

Risk

  • Callers that relied on LangChain stripping explicit nulls will now send them.
  • Header arguments are unaffected: a null header value is still dropped by the header sanitiser before the call.