Consolidate reasoning_content echo detection, add Xiaomi MiMo support

[stanleylai]

Bug Description

Xiaomi MiMo (mimo-v2.5-pro) requires reasoning_content to be echoed back on assistant messages in multi-turn conversations. Without it, the API returns HTTP 400:

"The reasoning_content in the thinking mode must be passed back to the API."

This is the same requirement as DeepSeek V4 and Kimi/Moonshot — Hermes already has fixes for those, but MiMo was not detected by any of the three existing methods (_needs_deepseek_tool_reasoning, _needs_kimi_tool_reasoning, _needs_thinking_reasoning_pad).

Impact: Anyone using MiMo via any provider (opencode-go, OpenRouter, custom endpoints) hits a hard failure on every message after the first turn.

Root Cause

Three separate detection methods existed for the same underlying requirement (echo reasoning_content), each with its own provider-specific logic. MiMo wasn't in any of them, and adding yet another copy would compound the sprawl.

Fix

Consolidate the three detection methods into a single _needs_reasoning_content_echo() backed by a module-level _REASONING_ECHO_PROVIDERS table. Add Xiaomi MiMo entries:

• _REASONING_ECHO_PROVIDERS table at module level — tuple of (provider, model_substr, host) rows, checked as OR
• _needs_reasoning_content_echo() — one method replacing three, iterates the table
• Xiaomi entries: "mimo-" model substring (hyphen avoids false positives on the common ML acronym) + api.xiaomimimo.com host fallback
• Deleted: _needs_deepseek_tool_reasoning(), _needs_kimi_tool_reasoning(), _needs_thinking_reasoning_pad()
• Updated all call sites and comments to mention MiMo
• Net: +107 / −65 lines across 2 files

How to Verify

1. hermes chat -q "What model are you?" -m mimo-v2.5-pro --provider opencode-go — turn 1 works
2. hermes chat --resume <session_id> -q "What did I just ask?" — turn 2 works (previously HTTP 400)
3. grep reasoning_content ~/.hermes/logs/agent.log — no 400 errors after restart
4. python -m pytest tests/run_agent/test_deepseek_reasoning_content_echo.py -v — 44/44 pass

Test Plan

• Unit tests pass (44/44)
• Live verification: multi-turn MiMo conversation works, zero 400s in logs
• Regression: DeepSeek and Kimi detection still works (covered by existing parametrized tests)
• Negative test: non-MiMo models (e.g. qwen3.5-plus) correctly return False

Risk Assessment

Low — The consolidated method preserves exact semantics for DeepSeek and Kimi (same match logic, same rows). MiMo detection uses "mimo-" with hyphen to avoid false positives. Kimi detection is now case-insensitive (.lower()s first), which is harmless since providers are already normalized at init.

[teknium1]

Thanks for the MiMo echo-back work. The reported runtime failure is already covered on current main: run_agent.py:5478-5483 includes MiMo in _needs_thinking_reasoning_pad(), and run_agent.py:5521-5535 implements Xiaomi/model/host detection. The replay path applies that predicate at agent/agent_runtime_helpers.py:2768.

Problems

• The proposed consolidation removes the current hot-path cache. Main caches the result by provider, model, and base URL at run_agent.py:5474-5483; the submitted table lookup performs row and host checks on each call.

Suggested changes

• If this refactor is salvaged, retain _thinking_pad_cache and its current key/invalidation semantics while centralizing the underlying matching data.
• Update against the current helper layout; agent/agent_runtime_helpers.py:2763-2896 now owns the replay forwarding and provider-switch reapplication logic.

Automated hermes-sweeper review.

[teknium1]

Please retain the current _thinking_pad_cache behavior when consolidating this predicate. On main, run_agent.py:5474-5483 caches the provider/model/base-URL result because this check is called repeatedly in the agent loop.