fix: ignore extra inputs on MCP tool args schema

[gambletan]

Summary

Fixes #4796

MCP tools fail with Extra inputs are not permitted because CrewAI's tool execution framework injects a security_context parameter into tool call arguments, but the dynamically created Pydantic args model uses extra='forbid' (the default from create_model_from_schema), causing validation to reject the unknown field.

Changes

• lib/crewai-tools/src/crewai_tools/adapters/mcp_adapter.py: Pass __config__=ConfigDict(extra='ignore') to create_model_from_schema() so that internally-added fields like security_context are silently discarded instead of raising a validation error.
• Added ConfigDict to the pydantic import.

Root cause

create_model_from_schema() defaults to ConfigDict(extra="forbid"). When CrewAI injects security_context (containing agent_fingerprint and metadata) during parameter validation — before _run is called — the strict schema rejects it.

Test plan

• Verified that create_model_from_schema accepts __config__ as an override parameter
• MCP tool calls with security_context injection no longer raise extra_forbidden validation errors
• MCP tools still correctly validate their declared input parameters

🤖 Generated with Claude Code

---

Note

Medium Risk
Loosens Pydantic validation for MCP-adapted tool inputs by ignoring unknown fields, which could mask unexpected/typoed parameters but is scoped to the MCP adapter path.

Overview
Prevents MCP tool calls from failing validation when CrewAI injects internal arguments by creating the MCP tool args_schema with ConfigDict(extra="ignore").

This updates mcp_adapter.py to pass a custom Pydantic config into create_model_from_schema (and imports ConfigDict), so extra fields like security_context are silently dropped instead of raising extra_forbidden errors.

^{Written by Cursor Bugbot for commit 7c7ab51. This will update automatically on new commits. Configure here.}