fix(opencode): add effort reasoning_options for mimo-v2.6 (flash, pro, flash-free)

[akraad]

Summary

Replaces reasoning_options = [] (which clients render as "no caller control" and hide the effort picker) with a verified effort list on the three MiMo-V2.6 entries:

• providers/opencode-go/models/mimo-v2.6-flash.toml → effort [none, low, medium, high]
• providers/opencode-go/models/mimo-v2.6-pro.toml → effort [none, low, medium, high]
• providers/opencode/models/mimo-v2.6-flash-free.toml → effort [none, low, medium, high]

Host classification

Both opencode (Zen gateway) and opencode-go are multi-lab relays → options come from this-host live test + lab docs + same-lab peers; no contradiction.

Wire

reasoning_effort (OpenAI-compatible body), serialized by the opencode client from reasoning_options effort values.

Evidence (live probe 2026-09-22, POST https://opencode.ai/zen/go/v1/chat/completions)

• reasoning_effort = none | low | medium | high → 200; none yields completion_tokens_details.reasoning_tokens = 0.
• xhigh, max, minimal, adaptive, default, "off" → 400 Invalid request parameters (upstream) — so those levels are not authored.
• thinking.type = disabled → 200 / 0 reasoning tokens as well; per the effort-with-none rule no separate toggle is added.
• Response headers: x-opencode-endpoint-id: mimo, x-opencode-upstream-model-id: mimo-v2.6-flash.

Free tier disclosure: opencode/mimo-v2.6-flash-free cannot be probed from outside OpenCode (403 FreeTierError); its values come from the live zen/go test on the same base_model (xiaomi/mimo-v2.6-flash) plus the kilo relay's effort option for xiaomi/mimo-v2.6-*. Happy to adjust if a maintainer can verify from within OpenCode.

Peers / baseline

• lab docs: https://mimo.mi.com/docs/api/chat/openai-api (thinking.type enabled|disabled)
• openrouter xiaomi/mimo-v2.6-* → toggle; kilo xiaomi/mimo-v2.6-* → effort [none, high]
• deepseek-v4-flash on the same providers → effort (style template)

Validation

bun validate passes (exit 0); generated JSON shows reasoning_options: [{"type":"effort","values":["none","low","medium","high"]}] for all three entries.

[github-actions]

Action items

• [high] [violation] providers/opencode-go/models/mimo-v2.6-flash.toml:6 - Check: Reasoning options baseline = lab + same-surface peers for this model; do not invent low/medium/high when that set is narrower; extra levels need this-host docs or live meaningful effect (not HTTP accept alone). Why: Xiaomi first-party MiMo is toggle-only (thinking.type enabled|disabled); OpenRouter xiaomi/mimo-v2.6-* is toggle-only; kilo peers use effort ["none", "high"]. Authoring ["none", "low", "medium", "high"] on opencode-go/opencode from 200 vs 400 probes overstates graded control and will show a false L/M/H picker. Same change on mimo-v2.6-pro.toml and opencode/models/mimo-v2.6-flash-free.toml. Action: Align all three files to baseline—prefer toggle (lab/OpenRouter wire) or effort ["none", "high"] (kilo-style off/on)—unless you show meaningful effect differences for low vs medium vs high (e.g. distinct reasoning token behavior), not just status codes.