providers: MiniMax MiniMax-M3 has no variants and /variants command missing (V2 beta) #36888

Summary

On the V2 beta, the preconfigured MiniMax (MiniMax.io) and MiniMax Token Plan (MiniMax.io) providers ship with no variants entries for MiniMax-M3, even though the model supports a thinking mode upstream. Both providers expose MiniMax-M3 in /models, but the TUI has no /variants command and Ctrl+T does not cycle anything on that model, so the user has no way to enable thinking from the UI.

Environment

• opencode version: opencode2 v0.0.0-next-15527 (beta/next channel)
• OS: Fedora Linux 7.1.3-200.fc44.x86_64
• Terminal: $TERM=xterm-256color, $COLORTERM=truecolor
• Shell: /bin/bash
• Install/channel: next (beta)
• Active plugins: @prevalentware/opencode-goal-plugin

Reproduction

1. Install the V2 beta: npm install -g @opencode-ai/cli@next (binary is opencode2).
2. Start the TUI. No opencode.jsonc overrides are required; both providers come from the built-in catalog.
3. Run /connect and select either MiniMax (MiniMax.io) or MiniMax Token Plan (MiniMax.io). Complete authentication.
4. Run /models and select MiniMax/MiniMax-M3.
5. Try to change the reasoning level: type /variants and/or press Ctrl+T.

Both providers (MiniMax and MiniMax-coding-plan) exhibit the same behavior.

Expected Behavior

/variants opens a picker that lists at least one variant (e.g. thinking or high) so the user can enable the model's thinking mode. The picker should be invokable via /variants and via the documented Ctrl+T cycle keybind for the currently selected model.

Actual Behavior

• /variants is not registered as a command in the TUI — typing it does nothing and it is not in command autocomplete.
• Ctrl+T has no effect on MiniMax-M3; no variant cycle UI appears.
• Inspecting the providers via the V2 API shows the catalog entries contain no variants array for MiniMax-M3:

  opencode2 api get /api/provider/minimax
  opencode2 api get /api/provider/minimax-coding-plan

  Both return a provider entry with package: aisdk:@ai-sdk/anthropic, base https://api.MiniMax.io/anthropic/v1, and the anthropic-beta: interleaved-thinking-2025-05-14,fine-grained-tool-streaming-2025-05-14 header — but no variants for MiniMax-M3.
• Selecting MiniMax-M3 and sending a prompt uses the base model with no thinking enabled; there is no UI affordance to opt in.

Additional Context

• Provider package: aisdk:@ai-sdk/anthropic (Anthropic-compatible endpoint at https://api.MiniMax.io/anthropic/v1).
• The preconfigured anthropic-beta header (interleaved-thinking-2025-05-14,...) confirms the upstream supports thinking; the V2 catalog simply does not surface it.
• This affects both MiniMax and MiniMax-coding-plan provider entries — same package, same endpoint, same missing variants.
• Related open issue covering the same gap more broadly for V2: feat(v2): unify reasoning options and OpenAI pro mode #36769 ("feat(v2): unify reasoning options and OpenAI pro mode"). Similar variant exposure bugs in other providers: ollama-cloud: reasoning variants not passing parameter to Ollama API #36638 (ollama-cloud), GPT-5.6 models missing max reasoning effort variant #36141 / [FEATURE]: expose reasoning effort "max" for GPT-5.6 (Codex Ultra parity) #36391 (GPT-5.6 max effort), bug(provider): GPT-5.6 *-pro models do not send reasoning.mode "pro" on Responses API #36392 (GPT-5.6 *-pro reasoning.mode "pro").

Partial workaround: define variants explicitly in user config, e.g.

{
  "$schema": "https://opencode.ai/config.json",
  "providers": {
    "MiniMax": {
      "models": {
        "MiniMax-M3": {
          "variants": [
            { "id": "thinking", "settings": { "thinking": { "type": "enabled", "budgetTokens": 8192 } } },
            { "id": "high",     "settings": { "thinking": { "type": "enabled", "budgetTokens": 8192 } } },
            { "id": "low",      "settings": { "thinking": { "type": "disabled" } } }
          ]
        }
      }
    }
  }
}

After this workaround the variants appear in the model picker and Ctrl+T cycles through them, but /variants itself is still not registered, and the same workaround has to be repeated under MiniMax-coding-plan.