fix: strip <think> tokens from reasoning model output

[benjamin7007]

Problem

Models with thinking/reasoning capabilities (DeepSeek-R1, MiniMax-M2.7, QwQ, Qwen3, etc.) include <think>...</think> blocks in their response content when used via OpenAI-compatible API endpoints. These internal reasoning tokens leak into:

1. Agent output — downstream nodes receive thinking tokens as part of the input
2. Timeline content — execution logs show raw thinking blocks
3. Final workflow result — end users see <think> tags in the output

Root Cause

OpenAIProvider._deserialize_chat_response() and _append_chat_response_output() pass raw content from model responses without filtering reasoning tokens.

Fix

Add _strip_thinking_tokens() classmethod to OpenAIProvider:

• Uses regex <think>.*?</think>\s* with re.DOTALL to strip thinking blocks
• Fast path: skips regex if <think> substring not found (zero-cost for non-thinking models)
• Applied in both deserialization paths (_deserialize_chat_response and _append_chat_response_output)

Testing

Verified with MiniMax-M2.7 (thinking model) in a Writer→Reviewer workflow:

• Before fix: <think> blocks leaked into Reviewer input and final output
• After fix: Clean output, no thinking tokens visible

Notes

• This is a minimal, targeted fix in the OpenAI provider only
• The Gemini provider uses a different content structure (MessageBlock) and would need separate handling if Gemini models add thinking tokens
• No existing tests were broken