Models and Reasoning¶

Model and reasoning directives refine the active backend. They do not select a framework or harness by themselves.

Select a Model¶

> adapter: pydantic
> model: groq:qwen/qwen3-32b

The model identifier is passed to the selected adapter. Its accepted syntax, availability, credentials, and provider prefix therefore depend on that adapter.

A model directive applies to following calls in the current lexical scope:

> model: fast-model

@draft(topic: str) -> str:
  >> Draft about <topic>: [text: str].
  = <text>

> model: quality-model

@review(text: str) -> str:
  >> Review feedback for <text>: [feedback: str].
  = <feedback>

Each procedure captures the top-level state present when it is defined.

Dynamic Model Expressions¶

```
selected_model = "groq:qwen/qwen3-32b"
```

> model: `selected_model`

The expression must return a string. Prefer literal models or profiles so the LSP can reason about adapter capability and configuration.

Reasoning Effort¶

Accepted values are:

• minimal
• low
• medium
• high
• xhigh
• max

> effort: high

A backtick expression may choose the value dynamically:

> effort: `"high" if args.deep else "low"`

Not every provider implements every effort level. Pydantic maps max to xhigh; DSPy receives the value as reasoning_effort; harness adapters map or validate according to their runtime.

Inheritance and Overrides¶

An inner scope inherits the outer model and effort until it replaces either:

> model: default-model
> effort: low

@deep_review(text: str) -> str:
  > effort: high
  >> Detailed review of <text>: [result: str].
  = <result>

The effort override does not discard the inherited model, tools, MCP servers, or instructions. Applying a profile merges its specified members; omitted members continue to inherit.

Choosing Effort¶

Use low effort for extraction, formatting, classification, and routine tool selection. Use higher effort for ambiguous planning, multi-source synthesis, or code reasoning where additional latency and cost are justified. Do not use effort as compensation for vague instructions or an oversized tool surface.

Errors¶

Kedi reports an invalid effort value before execution. Unknown or unavailable model identifiers, missing credentials, and unsupported model settings are reported by the selected adapter/provider. A profile that requires a model override fails if its adapter cannot construct one.