Thinking blocks compatibility for non-Claude providers

[lmsanch]

Problem

When switching from Claude (with extended thinking) to non-Claude providers (Fireworks, Together) mid-task, the message history contains thinking_blocks that these providers reject.

Error

litellm.BadRequestError: Fireworks_aiException - {
  "error": {
    "type": "invalid_request_error",
    "message": "Extra inputs are not permitted, field: 'messages[2].thinking_blocks'"
  }
}

Scenario

1. User starts task with Claude → Claude generates thinking_blocks in responses
2. User hits rate limit / runs out of credits
3. User switches to Fireworks/Together provider
4. Message history still contains thinking_blocks from Claude turns
5. Fireworks/Together API rejects the request

Root Cause

• Claude's extended thinking feature adds thinking_blocks to assistant messages
• These blocks are preserved in conversation history
• Non-Claude providers don't support this field and reject it
• LiteLLM's drop_params only drops request parameters, not message fields

Proposed Solution

Add a LiteLLM pre-request callback that:

1. Checks provider capabilities (from model-capabilities/*.yaml)
2. If provider doesn't support thinking, converts thinking_blocks to text content
3. Preserves reasoning information in compatible format

Conversion

# Before (Claude format):
{
  "role": "assistant", 
  "thinking_blocks": [{"type": "thinking", "thinking": "Let me analyze..."}]
}

# After (Compatible format):
{
  "role": "assistant",
  "content": [{"type": "text", "text": "[Reasoning] Let me analyze..."}]
}

Files to Create

• litellm_callbacks/thinking_compat.py - Pre-request callback
• Update start scripts to register callback in LiteLLM config