Skip to content

llm-thinking.md

Latest commit

 

History

History
115 lines (91 loc) · 3.94 KB

llm-thinking.md

File metadata and controls

115 lines (91 loc) · 3.94 KB
Raw

Thinking Configuration

Each profile selects an API format, a parameter preset, and a thinking level. Model default sends no thinking override. It is not an instruction to disable reasoning. Automatic compatibility is a legacy heuristic, not a model capability detector. Use an explicit mapping when the model or gateway has different rules.

Models That Always Think

If the model always reasons without an override, choose the Model default preset. The selector then contains only Model default.

For an explicit on-only switch, select Custom JSON mapping and use:

{
  "ENABLED": {
    "thinking": {
      "type": "adaptive"
    }
  }
}

Choose Enabled to send that fragment. The selector contains only Model default and Enabled; there is no Off or invented effort level. This example applies only to APIs that accept thinking.type=adaptive. Other APIs may require a different fragment, such as {"enable_thinking": true}.

Mappings describe request parameters, not a universal capability policy. The plugin cannot force an always-on model to stop reasoning, discover an upstream model behind an alias, or guarantee that a provider honors a parameter. If a mapped field is explicitly rejected, compatibility fallback retries without the thinking overrides and caches that mapping's rejection. The model then uses its own defaults, which can still include reasoning.

MiniMax M3 and M3.1 Flash

Verified against the official documentation on 2026-10-25:

Model Chat Completions, no override Anthropic Messages, no override Can turn thinking off?
MiniMax-M3 On Off Yes
MiniMax-M3.1-Flash On On No

Use the Chat Completions API format with a base URL such as https://api.minimax.io/v1, or the Anthropic Messages format with https://api.minimax.io/anthropic. Third-party gateways may differ.

For MiniMax-M3, an explicit toggle mapping for either protocol is:

{
  "DISABLED": {
    "thinking": {
      "type": "disabled"
    }
  },
  "ENABLED": {
    "thinking": {
      "type": "adaptive"
    }
  }
}

Do not use thinking.type=enabled for these MiniMax endpoints. For MiniMax-M3.1-Flash, use Model default or the on-only mapping above. Do not configure Off or an effort value of none; the official API rejects them. MiniMax M2.x also always thinks, but its supported protocols and treatment of unsupported fields differ from the M3 family.

Adjustable Depth on M3.1 Flash

MiniMax-M3.1-Flash supports low, medium, high, xhigh, and max. The Chat Completions field is reasoning_effort; the Anthropic Messages field is output_config.effort. MiniMax-M3 does not expose this depth control.

For Chat Completions, this mapping exposes Low, Medium, High, and Max without Off:

{
  "LOW": {"reasoning_effort": "low"},
  "MEDIUM": {"reasoning_effort": "medium"},
  "HIGH": {"reasoning_effort": "high"},
  "MAX": {"reasoning_effort": "max"}
}

For Anthropic Messages, use:

{
  "LOW": {"output_config": {"effort": "low"}},
  "MEDIUM": {"output_config": {"effort": "medium"}},
  "HIGH": {"output_config": {"effort": "high"}},
  "MAX": {"output_config": {"effort": "max"}}
}

There is no separate Xhigh UI level. To request it, map an existing level to "xhigh" instead, for example "MAX": {"reasoning_effort": "xhigh"}. The custom mapping is sent verbatim; OpenAI's GPT-5.2-specific maximum-effort translation does not change MiniMax mappings.

The plugin uses visible answer text for commit messages, not separate reasoning_content or Anthropic thinking blocks. Thinking tokens still consume the response token limit, so increase that limit if the provider returns no visible answer after exhausting the budget.

Sources