- Add scripts/extract-models.ts to parse command-code npm dist file - Generate models.json (21 models, 15 pricing entries) with contextWindow and maxOutputTokens pre-filled; no nulls or hardcoded fallbacks in index.ts - Rewrite index.ts to load model list and costs from models.json - Cap gateway model maxOutputTokens at 65536 (API limit for Baseten/Vercel) - Add 'Update models' section to README documenting the generation flow - Add npm run extract-models script
253 lines
8.5 KiB
TypeScript
253 lines
8.5 KiB
TypeScript
#!/usr/bin/env -S npx tsx
|
|
|
|
/**
|
|
* Extract model & provider definitions from the command-code npm package dist file.
|
|
*
|
|
* Usage:
|
|
* npx tsx scripts/extract-models.ts [path-to-dist/index.mjs]
|
|
* npx tsx scripts/extract-models.ts (downloads latest from npm)
|
|
*
|
|
* Output: models.json
|
|
* {
|
|
* providers: { ... }, // provider key -> value map
|
|
* providerGroups: { ... }, // provider-group key -> { id, label, providers[] }
|
|
* models: [ ... ], // flattened model definitions
|
|
* pricing: [ ... ] // pricing entries (per 1M tokens, USD)
|
|
* }
|
|
*/
|
|
|
|
import { execSync } from "node:child_process";
|
|
import { existsSync, mkdirSync, readFileSync, writeFileSync } from "node:fs";
|
|
import { join } from "node:path";
|
|
|
|
// ---------------------------------------------------------------------------
|
|
// Helpers
|
|
// ---------------------------------------------------------------------------
|
|
|
|
/** Evaluate a JS object literal string safely via Function constructor. */
|
|
function parseObjectLiteral(code: string): Record<string, unknown> {
|
|
const fn = new Function(`return (${code})`);
|
|
return fn() as Record<string, unknown>;
|
|
}
|
|
|
|
// ---------------------------------------------------------------------------
|
|
// Step 1: get the dist file
|
|
// ---------------------------------------------------------------------------
|
|
|
|
function ensureDist(srcPath?: string): string {
|
|
if (srcPath) {
|
|
if (!existsSync(srcPath)) throw new Error(`File not found: ${srcPath}`);
|
|
return srcPath;
|
|
}
|
|
|
|
// Download latest from npm
|
|
const tmpDir = join(process.cwd(), ".extract-tmp");
|
|
mkdirSync(tmpDir, { recursive: true });
|
|
const tgz = execSync(`npm pack command-code --pack-destination "${tmpDir}"`, {
|
|
encoding: "utf8",
|
|
}).trim();
|
|
const tgzPath = join(tmpDir, tgz);
|
|
execSync(`tar xzf "${tgzPath}" -C "${tmpDir}"`, { encoding: "utf8" });
|
|
return join(tmpDir, "package", "dist", "index.mjs");
|
|
}
|
|
|
|
// ---------------------------------------------------------------------------
|
|
// Step 2: extract
|
|
// ---------------------------------------------------------------------------
|
|
|
|
interface ModelDef {
|
|
key: string;
|
|
id: string;
|
|
provider: string;
|
|
spec: string;
|
|
label: string;
|
|
name: string;
|
|
description: string;
|
|
reasoning: boolean;
|
|
reasoningEfforts: string[] | null;
|
|
contextWindow: number;
|
|
maxOutputTokens: number;
|
|
vendorLabel: string | null;
|
|
}
|
|
|
|
interface PricingEntry {
|
|
provider: string;
|
|
id: string;
|
|
category: string;
|
|
promptCost: number;
|
|
completionCost: number;
|
|
cacheWrite5mCost: number;
|
|
cacheWrite1hCost: number;
|
|
cacheHitCost: number;
|
|
}
|
|
|
|
interface ProviderGroup {
|
|
id: string;
|
|
label: string;
|
|
shortLabel: string;
|
|
description: string;
|
|
providers: string[];
|
|
}
|
|
|
|
function extract(code: string) {
|
|
// --- Wt: provider constants ---
|
|
const wtMatch = code.match(/Wt=\{([^}]+)\}/);
|
|
if (!wtMatch) throw new Error("Cannot find Wt");
|
|
const wtRaw = "{" + wtMatch[1] + "}";
|
|
const wt: Record<string, string> = {};
|
|
for (const m of wtRaw.matchAll(/(\w+):"(\w[-\w]*)"/g)) {
|
|
wt[m[1]] = m[2];
|
|
}
|
|
console.log("Providers:", wt);
|
|
|
|
// --- an: model definitions ---
|
|
// an={...}).SONNET
|
|
const anIdx = code.indexOf("an={");
|
|
if (anIdx < 0) throw new Error("Cannot find an");
|
|
const anEndIdx = code.indexOf("}).SONNET", anIdx);
|
|
if (anEndIdx < 0) throw new Error("Cannot find end of an");
|
|
let anCode = code.substring(anIdx + 1, anEndIdx + 2); // "an={...})"
|
|
anCode = anCode.replace(/^\(an=/, "").replace(/\)$/, "");
|
|
|
|
// Replace minified JS idioms
|
|
anCode = anCode.replace(/\bQt\b/g, JSON.stringify("vercel-ai-gateway"));
|
|
anCode = anCode.replace(/\bon\b/g, JSON.stringify("chatComplete"));
|
|
anCode = anCode.replace(/\bsn\b/g, JSON.stringify("responses"));
|
|
anCode = anCode.replace(/Wt\.([A-Z_]+)/g, (_, key: string) =>
|
|
JSON.stringify(wt[key]),
|
|
);
|
|
anCode = anCode.replace(/!0/g, "true");
|
|
anCode = anCode.replace(/!1/g, "false");
|
|
|
|
const an = parseObjectLiteral(anCode);
|
|
|
|
// --- Yt: pricing (provider -> model array) ---
|
|
const ytIdx = code.indexOf("Yt={[");
|
|
if (ytIdx < 0) throw new Error("Cannot find Yt");
|
|
// Find the matching closing brace for Yt
|
|
let depth = 1;
|
|
let ytEndIdx = ytIdx + 4;
|
|
while (depth > 0 && ytEndIdx < code.length) {
|
|
if (code[ytEndIdx] === "{") depth++;
|
|
else if (code[ytEndIdx] === "}") depth--;
|
|
ytEndIdx++;
|
|
}
|
|
let ytCode = code.substring(ytIdx + 3, ytEndIdx); // "{ ... }"
|
|
ytCode = ytCode.replace(/Wt\.([A-Z_]+)/g, (_, key: string) =>
|
|
JSON.stringify(wt[key]),
|
|
);
|
|
ytCode = ytCode.replace(/!0/g, "true");
|
|
ytCode = ytCode.replace(/!1/g, "false");
|
|
const yt = parseObjectLiteral(ytCode);
|
|
|
|
// --- pn: provider groups ---
|
|
const pnIdx = code.indexOf('pn={"command-code"');
|
|
if (pnIdx < 0) throw new Error("Cannot find pn");
|
|
const pnEndIdx = code.indexOf(",__name(buildModelGroups", pnIdx);
|
|
if (pnEndIdx < 0) throw new Error("Cannot find end of pn");
|
|
let pnCode = code.substring(pnIdx + 3, pnEndIdx);
|
|
pnCode = pnCode.replace(/Wt\.([A-Z_]+)/g, (_, key: string) =>
|
|
JSON.stringify(wt[key]),
|
|
);
|
|
pnCode = pnCode.replace(/!0/g, "true");
|
|
pnCode = pnCode.replace(/!1/g, "false");
|
|
pnCode = pnCode.replace(/,\s*$/, "");
|
|
const pn = parseObjectLiteral(pnCode);
|
|
|
|
// --- Defaults for fields the CLI doesn't provide per-model ---
|
|
// maxOutputTokens: use the minimum across all providers for each model.
|
|
// Anthropic direct: 64k OpenAI direct: 128k
|
|
// DeepSeek (gateway): 384k (known to work)
|
|
// Other gateway models (Baseten/Vercel/Cloudflare/OpenRouter): 65536
|
|
const CONTEXT_WINDOW_FALLBACKS: Record<string, number> = {
|
|
"gpt-5.5": 256_000,
|
|
"zai-org/GLM-5.1": 200_000,
|
|
"MiniMaxAI/MiniMax-M2.7": 1_048_576,
|
|
"Qwen/Qwen3.6-Max-Preview": 1_000_000,
|
|
"Qwen/Qwen3.6-Plus": 1_000_000,
|
|
};
|
|
const DEFAULT_CONTEXT_WINDOW = 200_000;
|
|
|
|
function maxOutputTokensForModel(id: string, provider: string): number {
|
|
if (provider === "anthropic") return 64_000;
|
|
if (provider === "openai") return 128_000;
|
|
if (id.startsWith("deepseek/")) return 384_000;
|
|
// Gateway models — lowest common denominator across Baseten/Vercel/Cloudflare
|
|
return 65_536;
|
|
}
|
|
|
|
// --- Build output ---
|
|
const models: ModelDef[] = [];
|
|
for (const [key, obj] of Object.entries(an)) {
|
|
const m = obj as Record<string, unknown>;
|
|
const id = m.id as string;
|
|
const provider = m.provider as string;
|
|
models.push({
|
|
key,
|
|
id,
|
|
provider,
|
|
spec: (m.spec as string) || "chatComplete",
|
|
label: m.label as string,
|
|
name: m.name as string,
|
|
description: m.description as string,
|
|
reasoning: !!(
|
|
m.reasoning ??
|
|
((m.reasoningEfforts as string[])?.length ?? 0) > 0
|
|
),
|
|
reasoningEfforts: (m.reasoningEfforts as string[]) || null,
|
|
contextWindow:
|
|
(typeof m.contextWindow === "number" ? m.contextWindow : null) ??
|
|
CONTEXT_WINDOW_FALLBACKS[id] ??
|
|
DEFAULT_CONTEXT_WINDOW,
|
|
maxOutputTokens:
|
|
maxOutputTokensForModel(id, provider),
|
|
vendorLabel: (m.vendorLabel as string) || null,
|
|
});
|
|
}
|
|
|
|
const pricing: PricingEntry[] = [];
|
|
for (const [_provider, entries] of Object.entries(yt)) {
|
|
for (const entry of entries as Array<Record<string, unknown>>) {
|
|
pricing.push({
|
|
provider: entry.provider as string,
|
|
id: entry.id as string,
|
|
category: entry.category as string,
|
|
promptCost: entry.promptCost as number,
|
|
completionCost: entry.completionCost as number,
|
|
cacheWrite5mCost: (entry.cacheWrite5mCost as number) || 0,
|
|
cacheWrite1hCost: (entry.cacheWrite1hCost as number) || 0,
|
|
cacheHitCost: (entry.cacheHitCost as number) || 0,
|
|
});
|
|
}
|
|
}
|
|
|
|
const providerGroups: ProviderGroup[] = [];
|
|
for (const [key, obj] of Object.entries(pn)) {
|
|
const g = obj as Record<string, unknown>;
|
|
providerGroups.push({
|
|
id: g.id as string,
|
|
label: g.label as string,
|
|
shortLabel: (g.shortLabel as string) || "",
|
|
description: (g.description as string) || "",
|
|
providers: (g.supportedModelProviders as string[]) || [],
|
|
});
|
|
}
|
|
|
|
return { providers: wt, providerGroups, models, pricing };
|
|
}
|
|
|
|
// ---------------------------------------------------------------------------
|
|
// main
|
|
// ---------------------------------------------------------------------------
|
|
|
|
const distPath = ensureDist(process.argv[2]);
|
|
console.log("Reading:", distPath);
|
|
const code = readFileSync(distPath, "utf8");
|
|
const result = extract(code);
|
|
|
|
const outPath = join(process.cwd(), "models.json");
|
|
writeFileSync(outPath, JSON.stringify(result, null, 2), "utf8");
|
|
console.log(
|
|
`Wrote ${result.models.length} models + ${result.pricing.length} pricing entries to ${outPath}`,
|
|
);
|