Reasoning

Reasoning levels control how much internal reasoning a model does before responding. Cast combines /v1/models metadata when an endpoint provides it with a configured provider dialect for the request shape.

Levels

Level	Description
`off`	No reasoning, standard completion
`low`	Minimal reasoning
`medium`	Balanced reasoning (often the default)
`high`	Deep reasoning
`max`	Maximum reasoning effort

For models that report reasoning as a binary toggle (on/off without effort levels), the options are off and on.

Setting the Level

CLI

bash

cast -r high "refactor this function"
cast --reasoning medium "explain the session module"

Interactive

/reasoning

Opens a picker if the model supports reasoning controls. If the provider doesn't expose reasoning metadata, a message explains that the model uses its own default.

Provider protocol

/reasoning-format selects how Cast sends reasoning controls for the active saved provider. auto is the default and detects known endpoint hosts. Choose an explicit format when a compatible proxy uses a nonstandard host or needs a forced protocol.

Supported formats include OpenAI (reasoning_effort), OpenRouter (reasoning.effort), DeepSeek, Kimi, Qianfan, Qwen/DashScope, Together, xAI, Z.ai, Huawei ModelArts, and MiniMax. Generic OpenAI-compatible endpoints receive a conservative reasoning_effort request when reasoning is enabled and no control when it is off.

Saved

The reasoning level is saved to ~/.cast/settings.json per model. Switching models triggers a new reasoning selection if the new model supports different options.

How Reasoning Metadata is Discovered

Cast fetches the model list from the configured provider's /v1/models endpoint. OpenRouter exposes a rich reasoning field per model:

json

{
  "id": "qwen/qwen3-235b-a22b",
  "reasoning": {
    "mandatory": false,
    "default_enabled": true,
    "supported_efforts": ["high", "medium", "low"],
    "default_effort": "medium"
  }
}

When present, Cast reads this metadata to determine:

Whether the model supports reasoning at all
Whether it's a binary toggle or supports effort levels
Which efforts are available
What the default is

Many OpenAI-compatible providers omit this metadata. Their request dialect is still selected from the provider configuration, so their native reasoning controls and streaming reasoning fields can be normalized for the TUI and web UI.

Think Block Parsing

Some models (Qwen, DeepSeek) output reasoning in <think> blocks rather than through a structured API field. cast parses these blocks automatically:

<think>
Let me analyze this step by step...
</think>

Here's my analysis...

The thinking content is displayed separately in the TUI and excluded from the main response.

Provider Behavior

The exact off request depends on the selected dialect. For example, OpenRouter receives reasoning: { enabled: false }, OpenAI receives reasoning_effort: "none", and providers that use a thinking flag receive their native disabled value. MiniMax reasoning is always enabled; Cast requests its split-reasoning stream format so reasoning and answer text remain separate.

MiniMax's M3.1 family is the exception to "off is a request". MiniMax's own documentation (platform.minimax.io/docs/guides/text-generation, read 2026-10-23): "thinking cannot be turned off. Sending thinking: {"type": "disabled"} or effort: "none" returns 400", and the depth is set with reasoning_effort — low, medium, high, xhigh, max, defaulting to max. Cast sends reasoning_effort for those models and offers their five depths in the picker. The older MiniMax models (M3, M2.7, M2.5) keep the older thinking: { type: "adaptive" | "disabled" } vocabulary — they ignore reasoning_effort entirely.

Cast's own default for the M3.1 family is low, not the provider's max: measured on the behaviour bench with MiniMax-M3.1-Flash, three runs of thirteen tool-contract cases and three runs of the four cases that used to flip between runs, low scored 39/39 and 11/12 at about half the wall time, where max scored 36/39 and 7/12. A level left over from another provider resolves to that default rather than failing the request. Raise it for a task that needs the thinking: /reasoning in the terminal, --reasoning <level> for one run.

Two more things MiniMax's parameter table says, and cast leans on both: the thinking tokens count toward max_tokens (a budget that is too small ends the turn with finish_reason: "length" and an empty answer), and temperature and top_p are ignored — M3 and M3.1-Flash always sample at temperature 1 and top_p 0.95. A run of the same prompt is a fresh sample every time, which is why the behaviour bench measures a case three times before calling a result a regression.

When the provider doesn't report reasoning capabilities (unknown), cast sends no reasoning params, letting the provider use its own default.