fix(subagent): preserve reasoning_content in assistant messages

[lailoo]

Summary

• Subagent's _run_subagent() was silently dropping reasoning_content and thinking_blocks when building assistant messages for the conversation history
• Providers like Deepseek Reasoner mandate reasoning_content on every assistant message when thinking mode is active, causing a 400 BadRequestError on the second LLM round-trip
• The main AgentLoop already preserves these fields correctly via ContextBuilder.add_assistant_message() — this fix aligns the subagent loop with that behavior

Root Cause

In subagent.py, the assistant message was built as a plain dict without reasoning_content:

# Before (broken)
messages.append({
    "role": "assistant",
    "content": response.content or "",
    "tool_calls": tool_call_dicts,
})

Deepseek's API returns 400 with:

Missing `reasoning_content` field in the assistant message at message index 2

Fix

Conditionally include reasoning_content and thinking_blocks when the provider returns them, matching how ContextBuilder.add_assistant_message() handles this in the main agent loop:

# After (fixed)
assistant_msg: dict[str, Any] = {
    "role": "assistant",
    "content": response.content or "",
    "tool_calls": tool_call_dicts,
}
if response.reasoning_content is not None:
    assistant_msg["reasoning_content"] = response.reasoning_content
if response.thinking_blocks:
    assistant_msg["thinking_blocks"] = response.thinking_blocks
messages.append(assistant_msg)

Verification

Bug reproduced and fix verified by calling Deepseek Reasoner API (deepseek-reasoner) directly:

STEP 1: First LLM call → tool_calls + reasoning_content returned ✅
STEP 2: ❌ BUG — send assistant message WITHOUT reasoning_content
  → DeepseekException: Missing `reasoning_content` field in the assistant message at message index 2
STEP 3: ✅ FIX — send assistant message WITH reasoning_content
  → API success, normal response returned

Full test output:

======================================================================
STEP 2: BUG -- assistant message WITHOUT reasoning_content
======================================================================
  Message keys: ['role', 'content', 'tool_calls']
  Calling Deepseek API with buggy messages...

  BUG REPRODUCED!
  Error: Error calling LLM: litellm.BadRequestError: DeepseekException -
  {"error":{"message":"Missing `reasoning_content` field in the assistant
  message at message index 2."}}

======================================================================
STEP 3: FIX -- assistant message WITH reasoning_content
======================================================================
  Message keys: ['role', 'content', 'tool_calls', 'reasoning_content']
  Calling Deepseek API with fixed messages...

  SUCCESS!

======================================================================
SUMMARY
======================================================================
  BUG:  Without reasoning_content -> API error
  FIX:  With reasoning_content -> API success

  Issue #1834 confirmed and fix is correct!

Test plan

• Reproduced bug with real Deepseek Reasoner API (missing reasoning_content → 400 error)
• Verified fix with real Deepseek Reasoner API (with reasoning_content → success)
• Non-reasoning providers (Claude, GPT) unaffected — fields only added when present

Closes #1834

🤖 Generated with Claude Code