Model selection

Model selection defines which LLM an agent uses. Choose a provider and model from your provider catalog; connection settings and credentials belong in that catalog.

Configure agent model

Skiller uses the LLM section from the agent profile first (see Agent profiles). This is the file selected in the workflow YAML:

- agent: reviewer
  config:
    file: "~/.skiller/profiles/agent.reviewer.json"

If the profile has no LLM section, Skiller uses the settings in ~/.skiller/settings/agent.json, the global agent configuration.

If neither file defines it, configuration validation fails: an agent requires a provider and model.

LLM section shape

Both files use the same structure:

{
  "llm": {
    "provider": "bedrock",
    "model": "us.anthropic.claude-opus-4-6-v1",
    "effort": 3
  }
}

Adding the LLM section to a profile replaces the entire global section. Both provider and model must be present; omitted effort is left unset rather than inherited.

Fields

Field	Description	Default
provider	Provider name in the catalog.	Required; no default
model	Model identifier declared by that provider.	Required; no default
effort	Reasoning effort from 1 to 5; support depends on the adapter and model.	Not set

Defaults

Skiller does not select a provider or model automatically when the LLM section is missing. Authentication saves a selection in the global agent configuration, which profiles can inherit.

Effort is optional. Leaving it unset does not guarantee that reasoning is disabled: adapter and model behavior determines how the request is handled.

Model effort

Effort requests a reasoning level from 1 to 5. Its meaning depends on the adapter, provider, and model. Skiller’s mappers translate the level into the corresponding API setting and apply model-specific adjustments where implemented to find the closest supported match. This is not automatic negotiation: not every model supports every level.

Higher effort requests more reasoning; it is not an output token limit or a guarantee of better results.

Effort	Base mapping
1	low
2	medium
3	high
4	xhigh
5	max

OpenAI-compatible adapter

Sends the mapped level as the API’s reasoning effort. Support depends on the provider and model; Skiller does not retry with a lower level if it is rejected. When effort is omitted, the parameter is not sent and the provider applies its own behavior.

Bedrock adapter

Applies effort to Anthropic models using adaptive thinking. Other Bedrock models ignore this setting.

The mapper adjusts specific model limitations: for Opus 4.6 and Sonnet 4.6, level 4 maps to high rather than xhigh. In the example above, level 3 requests high effort for Opus 4.6.

Without effort, the mapper requests disabled thinking, except for models it identifies as adaptive-only, where it requests adaptive thinking with medium effort.

Codex adapter

The configured effort is not applied. Both Codex transports currently send a fixed medium effort, whether or not the LLM section includes this field.