[diegosouzapw]

Context

Proactive doc review of Xiaomi MiMo (api.xiaomimimo.com/v1, the native xiaomi-mimo provider) — same "new model, check the call flow for param mismatches" pass we did for DeepSeek V4 / MiniMax. Focus: don't send a param extra / missing / under a different name and take a 400. MiMo's validator is strict ("400 Param Incorrect" across several upstream issues), unlike DeepSeek which ignored unknowns silently — so extra-param risk is real here.

The one documented hard 400 (replay reasoning_content on tool-call turns) was already handled (xiaomi-mimo ∈ REASONING_REPLAY_PROVIDERS). This PR closes four remaining mismatches found in the audit.

Changes

1. Thinking control → MiMo's native thinking:{type} (fix)

MiMo controls chain-of-thought only via top-level thinking:{type:"enabled"|"disabled"} and does not understand OpenAI's reasoning_effort/reasoning. OmniRoute's OpenAI path carried reasoning intent as reasoning_effort, and the claude→openai translator can leave a Claude-shaped thinking:{type, budget_tokens} — so the client's on/off choice was dropped and budget_tokens/reasoning_effort rode along as extra params a strict validator can reject.

New open-sse/services/mimoThinking.ts::normalizeMimoThinking (wired in chatCore for provider==="xiaomi-mimo"):

• any thinking object → reduced to just {type} (disabled stays; enabled/adaptive/other → enabled),
• reasoning_effort / reasoning → removed.

It deliberately does not synthesize thinking:{type:"enabled"} from a bare reasoning_effort — mimo-v2-omni is documented non-thinking, so forcing it on could turn a silently-ignored param into a hard error. (Inverse situation to the DeepSeek reasoning_effort tier remap in base.ts, which becomes a no-op here once reasoning_effort is removed.)

2. Vision false-positive on the *-pro chat models (fix)

Per Xiaomi's docs, only mimo-v2.5 and mimo-v2-omni accept images; the *-pro chat models are text-only. modelSpecs marked mimo-v2.5-pro/mimo-v2-pro as vision-capable, and models.dev mislabels them (hermes-agent#18884) — and resolveVisionCapability lets a synced attachment:true win first, which would route an image request to a blind model (the #4071 failure mode).

• Corrected modelSpecs (supportsVision: false for both *-pro).
• Added a hard override in resolveVisionCapability (checked before the synced branch) for documented text-only ids, so a wrong synced attachment:true can't win. Anchored so mimo-v2.5-pro never matches the multimodal mimo-v2.5.

3. Catalog completeness

• Registered the missing mimo-v2-pro chat model (served via the gitlawb gateway but absent from the native provider).
• Registered the missing mimo-v2-tts speech model in the audio registry.

Validation (TDD — Hard Rule #18)

• New suites RED→GREEN: mimo-thinking-normalize.test.ts (9), model-capabilities-mimo-vision-override.test.ts (4 — seeds a wrong attachment:true and proves the override wins).
• Aligned two existing tests that enshrined the wrong vision verdict (t31-…-model-specs, xiaomi-mimo-provider) — correction, not masking.
• Adjacent regression cluster: 159/159 (provider-models-config, combo-vision-aware-routing, vision-detection-consistency, mistral-vision-sync, audio-speech-handler, reasoning-cache/replay, no-thinking-alias, …).
• typecheck:core clean · lint 0 · check:file-size OK (chatCore baseline 5086→5095) · check:any-budget:t11 PASS · docs-sync/docs-counts PASS.

Audited but intentionally not changed (reported for the record)

• ASR mimo-v2.5-asr is unwired (it's a custom /v1/chat/completions + input_audio + asr_options call, not OpenAI /v1/audio/transcriptions). Wiring it is a standalone feature, out of scope here.
• max_tokens vs max_completion_tokens: MiMo's docs canonicalize max_completion_tokens but no evidence it rejects max_tokens; left as a watch-item (renaming blindly is riskier than the status quo).
• temperature range [0,1.5] (vs OpenAI [0,2]): a client sending an extreme value could 400, but per-provider clamping is invasive — left as a watch-item.
• Tool-call message shape (XiaomiMiMo/MiMo#44): an upstream-acknowledged, unresolved strict-validation bug — not safely patchable speculatively.

[gemini-code-assist]

Warning

You have reached your daily quota limit. Please wait up to 24 hours and I will start processing your requests again!

[chatgpt-codex-connector]

You have reached your Codex usage limits for code reviews. You can see your limits in the Codex usage dashboard.