fix(mcp): use extra='ignore' in _json_schema_to_pydantic to allow CrewAI-injected security_context

[NIK-TIGER-BILL]

Problem

CrewAI's tool_usage.py injects a security_context field into every tool call's arguments before Pydantic validation (lines 1024-1045). MCP tools build their argument schema from the MCP server's inputSchema via create_model_from_schema(), which defaults to ConfigDict(extra='forbid').

Since the MCP server's schema never includes security_context, validation raises (closes #4796):

pydantic_core.ValidationError: Extra inputs are not permitted
  security_context
    Extra inputs are not permitted [type=extra_forbidden, ...]

Root cause

MCPToolResolver._json_schema_to_pydantic() calls create_model_from_schema() without overriding the default ConfigDict(extra='forbid').

Fix

Pass ConfigDict(extra='ignore') explicitly when building the schema model for MCP tools:

return create_model_from_schema(
    json_schema,
    model_name=model_name,
    enrich_descriptions=True,
    __config__=ConfigDict(extra="ignore"),   # ← allow security_context to pass through
)

Framework-injected fields are silently dropped before the model sees only the fields it declared. The MCP tool's own validation is unaffected.

Closes #4796

---

Note

Medium Risk
Relaxes Pydantic validation for MCP tool arguments by ignoring unknown fields, which could mask unexpected/typoed inputs but is scoped to MCP-derived schemas and prevents runtime failures from framework-injected metadata.

Overview
Fixes MCP tool argument validation failures by generating Pydantic models with extra="ignore" in MCPToolResolver._json_schema_to_pydantic, allowing CrewAI-injected fields like security_context to pass through without raising extra_forbidden errors.

This changes MCP tool schemas created via create_model_from_schema to silently drop unknown keys instead of rejecting them.

^{Written by Cursor Bugbot for commit 81037f4. This will update automatically on new commits. Configure here.}