feat(provider): preserve reasoning_details across turns and configure MiniMax reasoning_split via registry

[mengxr1234]

Summary

This PR improves reasoning metadata continuity across multi-turn conversations and configures MiniMax interleaved reasoning in a provider-configurable way.

Changes

• Added reasoning_details to provider response model.
• Preserved reasoning_details in assistant message history for:
  • main agent loop
  • subagent loop
• Allowed reasoning_details through message sanitization in LiteLLM provider.
• Parsed reasoning_details from LiteLLM response (message and provider-specific fields fallback).
• Moved MiniMax reasoning_split enablement to provider registry model_overrides for:
  • minimax-m2.5
  • minimax-m2.1
  • minimax-m2

Why

• Prevent loss of reasoning metadata between turns.
• Keep provider-specific behavior declarative in registry instead of hardcoding in request flow.
• Improve maintainability and extensibility for future provider-specific options.

Validation

• Ran Python compile checks on touched core files.
• Verified branch contains only 6 changed files relevant to this feature.

[chengyongru]

Review Summary

This PR addresses a real need (preserving reasoning metadata across turns), but has some issues that should be addressed before merging.

---

Issues

1. reasoning_split Parameter Passing Method (Critical)

According to litellm MiniMax documentation, reasoning_split should be passed via extra_body, not directly in kwargs:

# From litellm docs:
stream = client.chat.completions.create(
    model="minimax/MiniMax-M2.1",
    messages=[...],
    extra_body={"reasoning_split": True},  # <-- via extra_body
    stream=True,
)

However, this PR adds it directly to kwargs via model_overrides:

model_overrides=(
    ("minimax-m2.5", {"reasoning_split": True}),
    ...
)

And _apply_model_overrides does:

kwargs.update(overrides)  # adds reasoning_split directly to kwargs

Since litellm.drop_params = True is set in LiteLLMProvider.__init__, the reasoning_split parameter will likely be dropped silently because it's not a recognized litellm parameter.

Recommendation: Either:

1. Add support for extra_body in _apply_model_overrides
2. Or verify that litellm actually accepts reasoning_split as a direct kwarg (I couldn't find evidence of this)

2. Code Duplication in subagent.py

The subagent loop directly constructs the assistant message dict instead of using ContextBuilder.add_assistant_message():

# subagent.py (new code)
{
    "role": "assistant",
    "content": response.content or "",
    "tool_calls": tool_call_dicts,
    **({"reasoning_content": response.reasoning_content} if response.reasoning_content is not None else {}),
    **({"thinking_blocks": response.thinking_blocks} if response.thinking_blocks else {}),
    **({"reasoning_details": response.reasoning_details} if response.reasoning_details is not None else {}),
}

vs:

# loop.py (existing code)
messages = self.context.add_assistant_message(
    messages, response.content, tool_call_dicts,
    reasoning_content=response.reasoning_content,
    thinking_blocks=response.thinking_blocks,
    reasoning_details=response.reasoning_details,
)

This creates inconsistency and potential bugs if add_assistant_message is updated in the future.

Recommendation: Refactor subagent.py to use add_assistant_message for consistency, or extract the logic into a shared helper function.

3. Missing Tests

No test code is included to verify:

• reasoning_details is correctly preserved across turns
• model_overrides correctly applies reasoning_split to MiniMax models
• The feature works end-to-end with actual MiniMax API

---

Minor Suggestions

• Consider adding a comment explaining what reasoning_details contains (MiniMax interleaved CoT metadata per the field comment)
• The model_overrides pattern matching (pattern in model_lower) is case-insensitive but substring-based - this could accidentally match unintended models

---

Verification Needed

Before merging, please verify:

1. Does reasoning_split=True actually reach the MiniMax API when passed directly in kwargs?
2. Does the response actually contain reasoning_details in the expected format?

You can test by adding a debug log in _parse_response:

if reasoning_details:
    logger.debug(f"Got reasoning_details: {reasoning_details}")

[mengxr1234]

Thanks for the review — addressed all requested points.

What I changed

1. Switched MiniMax reasoning_split injection to extra_body={"reasoning_split": true} (instead of direct kwargs).
2. Updated model override application to merge extra_body safely.
3. Removed assistant-message duplication in subagent.py by reusing shared assistant message construction from ContextBuilder.
4. Added focused tests for MiniMax override behavior and extra_body merge.

Validation

• pytest -q tests/test_context_prompt_cache.py -k "reasoning_split or extra_body" passed
• pytest -q tests/test_commands.py -k "sanitize_messages_keeps_reasoning_details" passed
• python -m py_compile on touched files passed