Problem

MiMo reasoning/thinking models fail in multi-turn Hermes conversations with HTTP 400:

Error code: 400 - {
  "error": {
    "message": "Param Incorrect",
    "type": "upstream_error",
    "param": "The reasoning_content in the thinking mode must be passed back to the API.",
    "code": "400"
  }
}

MiMo's OpenAI-compatible API requires reasoning_content to be echoed on every assistant message when continuing a conversation in thinking mode. Hermes already implements this for DeepSeek (#15250) and Kimi (#17400) but MiMo was not enrolled in the enforcement set.

Root cause

_needs_thinking_reasoning_pad() gates all three enforcement branches:

1. _build_assistant_message — pins reasoning_content at creation time for tool-call turns
2. _copy_reasoning_content_for_api tier 1 — upgrades stale "" → " " on replay
3. _copy_reasoning_content_for_api tier 4 — unconditional " " pad for all assistant turns

The method only OR'd in _needs_deepseek_tool_reasoning() and _needs_kimi_tool_reasoning(). MiMo was missing, so all three branches were skipped for provider="xiaomi" / xiaomimimo.com / mimo-* sessions.

Fix

Add _needs_mimo_tool_reasoning() detecting three signals (mirrors the pattern for Kimi):

• provider == "xiaomi" (case-insensitive)
• base_url host matches xiaomimimo.com
• model name starts with "xiaomi/" or "mimo-" (covers catalog and bare forms)

OR it into _needs_thinking_reasoning_pad(). All existing downstream branches activate for MiMo automatically — no other provider behaviour changes.

Tests

tests/run_agent/test_mimo_reasoning_content_echo.py — 30 hermetic tests, modelled on the existing DeepSeek suite:

• TestNeedsMimoToolReasoning — all three detection signals, case-insensitive, negative/substring cases
• TestNeedsThinkingReasoningPadIncludesMimo — MiMo activates pad; DeepSeek/Kimi still active; unrelated provider off
• TestCopyReasoningContentForApiMimo — poisoned history → " ", real content preserved verbatim, stale "" upgraded, cross-provider leak guard, base URL match, non-MiMo untouched
• TestBuildAssistantMessageMimo — parametrized over all three detection signals; streamed reasoning promoted over pad; real reasoning_content preserved

All 30 new tests pass. Existing test_deepseek_reasoning_content_echo.py (36 tests) unaffected.

Visual

flowchart TD
    A["_needs_thinking_reasoning_pad()"] --> B["_needs_deepseek_tool_reasoning()"]
    A --> C["_needs_kimi_tool_reasoning()"]
    A --> D["_needs_mimo_tool_reasoning() ← NEW"]
    D --> E{"provider == 'xiaomi'?"}
    D --> F{"base_url host = xiaomimimo.com?"}
    D --> G{"model starts with 'xiaomi/' or 'mimo-'?"}
    E -- yes --> H[enforce echo-back]
    F -- yes --> H
    G -- yes --> H

Closes #24443