What

Enrol Xiaomi MiMo into the reasoning_content echo-back enforcement set on both the OpenAI-compatible path (run_agent.py) and the Anthropic-Messages path (agent/anthropic_adapter.py) so multi-turn MiMo conversations stop failing with HTTP 400.

Why

MiMo's OpenAI-compatible API requires reasoning_content on every prior assistant message in thinking mode (see MiMo docs). Without it, the next replay returns:

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

Hermes already ships the echo-back pattern for DeepSeek V4 thinking and Kimi / Moonshot thinking on both protocol paths; MiMo was simply not in either enforcement gate. That double omission produced four failure modes:

1. Tool-call turns without SDK-exposed reasoning_content — when MiMo streams reasoning via deltas only, the assembled ChatCompletionMessage has no top-level attribute, so the build-time pad never fired. Persisted history then lacked the field and the next request 400'd.
2. Cross-provider history — sessions opened under another provider and then /model-switched to MiMo bypassed the tier-2 leak guard, so prior chain-of-thought either leaked or was dropped.
3. Legacy reasoning_content="" sessions (pre-fix: use non-empty reasoning_content placeholder for DeepSeek V4 Pro thinking mode #17341) replayed the empty string verbatim instead of upgrading it to " ".
4. Anthropic-Messages path — MiMo running through an Anthropic-Messages-compatible gateway (xiaomimimo.com, OpenRouter mimo- slots, private proxies) had no carve-out: unsigned thinking blocks synthesised from reasoning_content were stripped along with signed Anthropic blocks, and the next replay 400'd.

Fixes #24443.

How

run_agent.py (OpenAI-compatible path)

• New _needs_mimo_tool_reasoning(), modelled on _needs_kimi_tool_reasoning(). Detection covers four signals:
  • provider == "xiaomi" (Hermes-internal id from hermes_cli/auth.py)
  • base_url host matches xiaomimimo.com
  • model starts with xiaomi/ (catalog form)
  • model starts with mimo- or contains /mimo- (bare + third-party catalog form)
• _needs_thinking_reasoning_pad() ORs the new detector in. All four existing tiers of _copy_reasoning_content_for_api and the tool-call pad branch in _build_assistant_message activate for MiMo automatically.
• Inline comments at the three pad sites updated to mention MiMo alongside DeepSeek / Kimi (refs MiMo reasoning models may fail in Hermes because reasoning_content is not preserved in chat history #24443).

agent/anthropic_adapter.py (Anthropic-Messages path)

• New _model_name_is_xiaomi_mimo() covering three slot shapes: catalog xiaomi/, bare mimo-, and embedded /mimo- for namespaced third-party slots (e.g. openrouter/mimo-v2.5-pro). Intentionally narrow — does not accept mimo_ / xiaomi-mimo- (no published catalog uses those) so lookalikes are not misdetected.
• New _is_xiaomi_mimo_anthropic_endpoint(), ORed into _preserve_unsigned_thinking alongside _is_kimi_family_endpoint and _is_deepseek_anthropic_endpoint. MiMo now uses the same strip-signed / keep-unsigned policy as Kimi /coding and DeepSeek /anthropic.
• Comment block at the thinking-block synthesis site updated to mention all three strict-replay providers.

DeepSeek / Kimi / Anthropic-native / MiniMax behaviour is untouched — the three detectors are OR-combined and mutually exclusive in practice. No schema or state migration.

A known gap is documented in the new docstring: sessions routed through a proxy or institutional gateway with a custom hostname AND no Hermes provider id set will not auto-detect; users in that configuration should set provider: xiaomi explicitly.

How to test

pytest tests/run_agent/test_mimo_reasoning_content_echo.py \
       tests/run_agent/test_deepseek_reasoning_content_echo.py \
       tests/agent/test_mimo_anthropic_thinking.py \
       tests/agent/test_deepseek_anthropic_thinking.py \
       tests/agent/test_kimi_coding_anthropic_thinking.py -q

run_agent.py suite

38 hermetic tests in tests/run_agent/test_mimo_reasoning_content_echo.py mirroring the DeepSeek echo suite:

• Detection (_needs_mimo_tool_reasoning): all four positive signals, case-insensitive provider id, third-party catalog form, plus negative guards for lookalike model names (MiniMax, Mistral, Microsoft Phi-4, mimowave substring).
• Pad gate: _needs_thinking_reasoning_pad flips on under MiMo and remains off for unrelated providers; pre-existing DeepSeek / Kimi paths unaffected.
• _copy_reasoning_content_for_api: all four tiers exercised under MiMo (tool-call pad, "" → " " upgrade, explicit content preserved, reasoning field promoted, cross-provider leak guard injects " ").
• _build_assistant_message: reasoning → reasoning_content backfill, model_extra preservation, tool-call pad with no raw reasoning_content, parametrised matrix covering attr-none / attr-absent / base-url / catalog / negative-non-mimo cases, streamed-reasoning promotion over pad, creation-time pad boundary for text-only turns.

agent/anthropic_adapter.py suite

37 hermetic tests in tests/agent/test_mimo_anthropic_thinking.py mirroring the DeepSeek/Kimi Anthropic-path suites:

• Detection (_model_name_is_xiaomi_mimo): catalog / bare / /mimo- shapes; case-insensitive; whitespace-tolerant.
• Negative guards: MiniMax, Mistral, Microsoft Phi-4, mimowave, phi-4-mimo-style, mimo_v2, xiaomi-mimo-v2, empty/non-string inputs — all must NOT match.
• Endpoint detection (_is_xiaomi_mimo_anthropic_endpoint): host-only, model-only, neither-signal cases.
• convert_messages_to_anthropic: unsigned thinking blocks preserved on tool-call replay (parametrised over four detection routes); preservation across non-latest assistant turns; signed Anthropic blocks stripped; cache_control stripped from thinking blocks; MiniMax regression guard (a too-liberal MiMo matcher would break MiniMax — explicit negative test).

Result: 101 tests pass across the new file and the pre-existing DeepSeek / Kimi / MiMo echo suites, 0 regressions. ruff check agent/anthropic_adapter.py tests/agent/test_mimo_anthropic_thinking.py run_agent.py tests/run_agent/test_mimo_reasoning_content_echo.py clean.

Platforms tested

• macOS 15 / Python 3.12 (local)

Hermetic test suite — no platform-specific code paths.