litellm - 💡(How to fix) Fix [Bug]: Fireworks AI rejects MCP tool schemas with unsupported JSON Schema fields (title, default) — triggered by Coralogix & DevRev MCP integrations

What happened?

When using LiteLLM with Fireworks AI models (e.g. kimi-k2p6) alongside MCP servers (Coralogix, DevRev), all requests fail with:

litellm.BadRequestError: Fireworks_aiException - {"error":{"object":"error",
"type":"invalid_request_error","message":"JSON Schema not supported: could not 
understand the instance `{'default': None, 'title': 'Page Size'}`."}}
Received Model Group=kimi-k2p6

Root Cause

MCP servers (Coralogix, DevRev, and likely others) define tools using Pydantic models, which auto-generate JSON Schema fields like title and default. LiteLLM passes these tool schemas to Fireworks AI as-is, but Fireworks AI's JSON Schema validator rejects them.

This is not covered by the existing fix in #6797 (which only addressed response_format schemas) — tool/function call schemas from MCP servers are unaffected.

Expected Behavior

LiteLLM should sanitize tool schemas before sending to Fireworks AI, stripping unsupported JSON Schema fields — similar to how strict was handled in #9774.

The fix should live in FireworksAIConfig in litellm/llms/fireworks_ai/chat/transformation.py and apply to both tool schemas and response_format schemas.

Related Issues

• #6797 (partial fix — response_format only)
• #9774 (strict param fix for Fireworks)