@drover/model

resolveModel

function resolveModel(
  spec: ModelSpec,
  opts: { runId: string; aliases?: AliasMap; env?: Env },
): Effect<ResolvedModel, ModelError>;

interface ResolvedModel {
  model: Model<KnownApi>;        // pi-ai's typed model record
  apiKey: string;
  reasoning?: ReasoningLevel;
  temperature?: number;
  maxTokens?: number;
}

Lookup order:

1. Alias map (custom ∪ DEFAULT_ALIASES).
2. Provider-prefixed slug provider:modelId.
3. Search every pi-ai provider for a model whose id equals name.

Failure modes (all ModelError):

• routing_miss: name not found.
• auth: provider’s env vars unset.
• provider_error: pi-ai threw.

Reasoning slug

"name:reasoning" parses cleanly. Valid reasoning levels: minimal | low | medium | high | xhigh.

For models that declare reasoning: true in pi-ai’s registry, omitting a level defaults to "medium" (verified necessary for gpt-5-mini and similar — without it the model returns no final text).

DEFAULT_ALIASES

All routed through OpenRouter. Override the map per project via withAliases({...}) or replace entirely.

withAliases

function withAliases(
  ...layers: ReadonlyArray<AliasMap>
): AliasMap;

Merge user overrides on top of DEFAULT_ALIASES. Later layers win.

Provider env keys

First match wins. To add a provider, set the env var and reference it in your spec / alias.