FastMCP tool argument models silently ignore unknown/misspelled arguments (extra=ignore default)

Description

mcp.server.fastmcp.utilities.func_metadata.ArgModelBase does not set extra in its
model_config, so it inherits Pydantic v2's default, extra="ignore":

class ArgModelBase(BaseModel):
    """A model representing the arguments to a function."""
    ...
    model_config = ConfigDict(
        arbitrary_types_allowed=True,
    )

Every per-tool *Arguments model that FastMCP generates via create_model(..., __base__=ArgModelBase, ...)
inherits this. As a result, when a client calls a tool with an argument name that doesn't
exist on the function (a typo, or a hallucinated parameter name from an LLM), Pydantic
silently drops it instead of raising a validation error. The tool then runs with that
parameter at its default value, and returns a "successful" but semantically wrong result —
with no signal anywhere that anything was off.

Reproduction

from mcp.server.fastmcp import FastMCP

mcp = FastMCP("repro")

@mcp.tool()
def read_doc(topic: str = "") -> str:
    return f"topic={topic!r}"

# Simulate what happens when a client calls with the wrong argument name:
tool = mcp._tool_manager.get_tool("read_doc")
result = tool.fn_metadata.arg_model.model_validate({"path": "something"})
print(result.model_dump_one_level())
# {'topic': ''}  <- no error, no signal that "path" was bogus

Expected behavior

An unknown argument in a tools/call request should fail validation immediately with a
clear error (e.g. Pydantic's own "Extra inputs are not permitted"), the same way a missing
required argument already does. This is especially important for MCP given the primary
caller is frequently an LLM: a wrong parameter name is a common, plausible failure mode, and
a loud, immediate error is far more useful (and self-correcting for the model) than a
silently-wrong "successful" response.

Suggested fix

Set extra="forbid" on ArgModelBase.model_config:

model_config = ConfigDict(
    arbitrary_types_allowed=True,
    extra="forbid",
)

This should be safe: the fields validated by ArgModelBase subclasses are exactly the
arguments dict of a tools/call request (i.e. exactly what's declared in the tool's
inputSchema). Protocol-level fields (_meta, progressToken, etc.) live on the
surrounding params object, not inside arguments, and Context is injected separately
via arguments_to_pass_directly — neither passes through ArgModelBase. The side effect is
that model_json_schema() will now emit "additionalProperties": false on the published
inputSchema, which seems like a feature, not a regression (it lets clients that validate
against the schema catch this class of error before the round-trip).

Scope checked

• Confirmed present in mcp 1.27.2 and 1.28.1 (latest on PyPI as of 2026-07).
• Confirmed still present on main (the in-progress v2, where fastmcp is being renamed to
  mcpserver): src/mcp/server/mcpserver/utilities/func_metadata.py has the same
  ArgModelBase without extra set.
• I didn't find an existing issue about this specific behavior (searched for "extra fields",
  "forbid", "unknown argument", "additionalProperties", "silently ignored").
• For context: the TypeScript SDK has an equivalent gap for a different underlying reason
  (Zod's default strip behavior on unknown object keys) — see
  https://github.com/modelcontextprotocol/typescript-sdk/issues/147

Happy to open a PR with the one-line change plus a regression test if that's useful.

从 src/mcp/server/mcpserver/utilities/func_metadata.py 开始，检查 ArgModelBase.model_config 以及生成的 arg_model 验证路径。添加一个使用 read_doc-style 未知参数案例的回归测试，然后运行相关的 FastMCP 测试；当未知工具参数引发验证错误且发布的 schema 拒绝额外属性时，即表示完成。