/**
* model-selection — Shared model selection utilities.
*
* Centralizes model preference lists and selection logic so tools
* don't duplicate the same heuristics. Two strategies:
*
* 1. Large-context models (Gemini 1M) — for tasks with big inputs
* (session recap, file analysis)
* 2. Cheap/fast models — for compression/distillation
* (already in compress.ts as pickCompressorModel)
*/
import { Provider } from "../providers/provider.js";
// ---------------------------------------------------------------------------
// Large-context model selection
// ---------------------------------------------------------------------------
/**
* Preferred models for large-context tasks, in priority order.
* These models can handle 500K+ tokens of input.
* Gemini 2.5 Flash is ideal: 1M context, fast, cheap on subscription quota.
*/
export const LARGE_CONTEXT_MODELS = [
"gemini-2.5-flash", // 1M context, fast, best for large inputs
"gemini-2.5-flash-lite", // 1M context, fastest
"gemini-3-flash", // likely 1M context
"gemini-3-pro", // likely 1M+ context
"gemini-2.5-pro", // 1M context, slower but smarter
];
/**
* Pick a model suitable for large-context tasks — prefers Gemini models
* with 1M+ context windows. Falls back to any available model.
*/
export async function pickLargeContextModel(
provider: Provider,
preferredModel?: string
): Promise<string | null> {
if (preferredModel) return preferredModel;
try {
const available = await provider.listModels();
if (available.length === 0) return null;
// Try preferred models in priority order
for (const preferred of LARGE_CONTEXT_MODELS) {
const match = available.find((m) =>
m.id.toLowerCase().includes(preferred)
);
if (match) return match.id;
}
// Fallback: any gemini model
const anyGemini = available.find((m) =>
m.id.toLowerCase().includes("gemini")
);
if (anyGemini) return anyGemini.id;
// Last resort: any model
return available[0].id;
} catch {
return null;
}
}
