Context

run_agent.py:_supports_reasoning_extra_body() short-circuits to False for any base_url not on its whitelist (nousresearch.com, ai-gateway.vercel.sh, github models, openrouter with specific model prefixes, lmstudio with detected capability). The MiMo direct endpoint isn't on that list, and there's no is_xiaomi branch in agent/transports/chat_completions.py's top-level reasoning_effort block either. So neither top-level reasoning_effort nor extra_body.reasoning is forwarded to MiMo.

The conservative default is sensible in principle (avoids forwarding fields some upstreams reject), but I wasn't sure whether MiMo actually honors reasoning_effort server-side, so I tested it.

TL;DR — the conservative default is correct. MiMo accepts the field at the schema layer (rejects xhigh with Input should be 'low', 'medium' or 'high'), but the four levels (none / low / medium / high) produce statistically indistinguishable reasoning depth, accuracy, and reasoning content on a real AIME problem.

Setup

• Model: mimo-v2.5-pro

• Endpoint: MiMo OpenAI-compatible /v1/chat/completions, called directly with httpx (bypassing hermes — the goal was to characterize server-side behavior in isolation)

• Prompt: AIME 2025 II Problem 2, sourced verbatim from HF yentinglin/aime_2025 row 16

  "Find the sum of all positive integers $n$ such that $n+2$ divides $3(n+3)(n^2+9)$."

  Ground truth: 49

• 4 effort levels (none, low, medium, high) × N=5 trials each

• temperature=0, max_tokens=16000

• prompt_tokens_details.cached_tokens = 256/304 across all 20 trials, so prompt cache hit rate is constant — not a confound.

Results

Mann-Whitney U (two-tailed, normal approximation*) on reasoning_tokens distributions, all 6 pairs:

none vs low      p=0.117    indistinguishable
none vs medium   p=0.117    indistinguishable
none vs high     p=0.917    indistinguishable
low vs medium    p=0.602    indistinguishable
low vs high      p=0.917    indistinguishable
medium vs high   p=0.754    indistinguishable

*Normal approximation; exact p with N=5 vs N=5 would be slightly higher (~0.15 for the lowest U). No pair reaches p<0.05 either way. Within-effort std (especially high at 623) dwarfs between-effort mean spread (1619–1948 = 329).

Reasoning content inspection

I read the first ~1500 chars of message.reasoning_content for high trial 0 (rt=2871, longest), high trial 1 (rt=1473, shortest of high), and low trial 2 (rt=1392, shortest of low). All three follow the same solution path within that window:

1. Substitute d = n+2, so n = d-2
2. Compute mod d: (d+1) ≡ 1, (d²-4d+13) ≡ 13
3. Conclude d | 39
4. Enumerate divisors d ∈ {3, 13, 39}, sum n = 1+11+37 = 49

The extra tokens in high trial 0 look (from skimming) like verbose self-hedges ("But careful: Is that correct? ... However, there is a subtlety:") around the same argument structure rather than additional reasoning steps. The variance appears more consistent with sampling noise within a single reasoning policy than with effort-controlled depth.

(I also ran 5 trials per effort on a trivial retrievable prompt — 24-game with 8,8,3,3 — and saw the same indistinguishability there. So the flat result isn't an artifact of "AIME P2 was already easy enough for low effort"; it holds at both difficulty extremes I tested.)

Suggestion

The current behavior is correct — please don't add is_xiaomi to the top-level reasoning_effort block in agent/transports/chat_completions.py or extend _supports_reasoning_extra_body() to include xiaomi. Doing so would just ship a no-op field to MiMo.

Two minor follow-ups, no strong opinion:

1. Inline comment in _supports_reasoning_extra_body() documenting that xiaomi/MiMo was empirically tested and excluded — would prevent a future PR from "completing" the list without checking server behavior.
2. Optional UX nicety: if agent.reasoning_effort is set in config.yaml while model.provider is one for which the field won't be forwarded, surface a one-time warning at startup. Right now the config silently has no effect for these providers.

Caveats

• N=5 per effort. Power to detect a small effect is limited; the no-effect conclusion rests on the combination of (a) flat means, (b) overlapping ranges, (c) identical solution paths in the inspected reasoning content, and (d) 100% accuracy across all efforts.
• Single problem (AIME 2025 II P2) plus a trivial sanity-check prompt. Not tested on code, longer reasoning chains, or multi-turn agentic scenarios.
• Tested only mimo-v2.5-pro; other MiMo variants not covered.
• Server-side behavior may change. Tested 2026-04-29.
• This experiment characterizes the MiMo endpoint in isolation. The hermes → MiMo end-to-end path was not exercised, since hermes already (correctly) doesn't forward the field.