feat: strip thinking blocks when switching to non-thinking models

[chaizhenhua]

Summary

When the framework switches models mid-conversation (via InferenceModelOverride or fallback_models), thinking blocks from a thinking-capable model must be stripped before sending history to a model that doesn't support extended thinking. Currently no such stripping logic exists.

Problem

If conversation history contains ReasoningDelta / ReasoningEncryptedValue events from a thinking-capable model (e.g., Claude Opus), and inference falls back to a non-thinking model, the thinking content in the message history may cause API errors or undefined behavior.

Proposed solution

Add a pre-inference transform (or logic in build_messages() / to_chat_message()) that:

1. Detects whether the target model supports extended thinking
2. If not, strips all thinking/reasoning content from the message history before sending
3. Preserves the text and tool_use content blocks

This could be implemented as:

• A method on ModelDefinition indicating thinking support
• An InferenceRequestTransform that conditionally strips thinking blocks
• Or logic in the genai conversion layer (engine/convert.rs)

Key files

• crates/tirea-agentos/src/engine/convert.rs — message conversion to genai format
• crates/tirea-agentos/src/runtime/streaming.rs — ReasoningDelta, ReasoningEncryptedValue events
• crates/tirea-agentos/src/runtime/loop_runner/core.rs — build_messages()
• crates/tirea-agentos/src/composition/registry/model.rs — ModelDefinition

References

• Book Chapter 4 §4.1.3: thinking block processing rules for model switching