| // Resolving a model id against the catalog, for the launchers. | |
| // | |
| // The subtlety this module exists for: `catalog()` fetches `/v1/models` with | |
| // `?available=true`, so a model missing from that response has TWO possible | |
| // meanings — it does not exist, or it exists and is merely out of quota right | |
| // now. Treating absence as "unknown model" reports a healthy pinned model as a | |
| // typo the moment its provider hits a rate limit. Only the UNFILTERED roster | |
| // can tell those apart, so `--model` is validated against that. | |
| import type { CatalogModel } from './types.js'; | |
| export interface ResolvedModel { | |
| id: string; | |
| /** Context window in tokens, or undefined when the catalog does not publish | |
| * one. Undefined is a fact — an unpublished window is not the same as a | |
| * known 128k, and callers must not invent one. */ | |
| contextWindow?: number; | |
| /** The catalog lists this model but it is not currently servable. */ | |
| unavailable: boolean; | |
| } | |
| export class UnknownModelError extends Error {} | |
| export function contextWindowOf(model: CatalogModel): number | undefined { | |
| return model.context_window ?? model.context_length ?? undefined; | |
| } | |
| /** | |
| * The launcher's default when the caller did not pin one: the first servable | |
| * concrete model, falling back to `auto`. | |
| * | |
| * Deliberately NOT the same default as the setup generators, which prefer | |
| * `auto`. A generated config file is long-lived and should let the router | |
| * choose per request; a launch pins one id into ANTHROPIC_MODEL for the whole | |
| * session, and naming a concrete model there is what makes the session's model | |
| * visible and reproducible. Unifying the two silently changes what | |
| * `freellmapi launch` runs. | |
| */ | |
| export function defaultModelId(models: CatalogModel[]): string { | |
| return models.find(model => model.id !== 'auto' && model.available !== false)?.id ?? 'auto'; | |
| } | |
| /** | |
| * Resolve the model a launcher should pin. | |
| * | |
| * `available` is the filtered roster the CLI already fetches; `full` is the | |
| * unfiltered one. Pass the same array twice only when the unfiltered roster | |
| * could not be fetched — the caller then loses the ability to distinguish an | |
| * unknown id from an unavailable one, which is exactly the ambiguity this | |
| * function exists to remove. | |
| */ | |
| export function resolveLaunchModel( | |
| requested: string | undefined, | |
| available: CatalogModel[], | |
| full: CatalogModel[] = available, | |
| ): ResolvedModel { | |
| const id = requested ?? defaultModelId(available); | |
| const known = full.find(model => model.id === id) ?? available.find(model => model.id === id); | |
| if (!known) { | |
| // `auto` is served by the router itself and need not appear in any roster. | |
| if (id === 'auto') return { id, unavailable: false }; | |
| const suggestions = full | |
| .map(model => model.id) | |
| .filter(candidate => candidate.includes(id) || id.includes(candidate)) | |
| .slice(0, 5); | |
| throw new UnknownModelError( | |
| `No model '${id}' in this gateway's catalog.` | |
| + (suggestions.length ? ` Did you mean: ${suggestions.join(', ')}?` : '') | |
| + ' Run `freellmapi list-models` to see what is registered.', | |
| ); | |
| } | |
| return { | |
| id, | |
| contextWindow: contextWindowOf(known), | |
| // Absent from the filtered roster but present in the full one = registered | |
| // but not servable right now (quota, cooldown, a disabled key). | |
| unavailable: known.available === false || !available.some(model => model.id === id), | |
| }; | |
| } | |