@broberg/ai-sdk 0.55.3 → 0.56.1

package/dist/index.js

@@ -3,25 +3,28 @@ import {
   ModelUnavailableError,
   allEntries,
   assertOverrideProvider,
+  describeTier,
   listModels,
   providerIds,
   resetRegistry,
   resolveModel,
   resolveTier,
   setAvailability
-} from "./chunk-PDGFAD66.js";
+} from "./chunk-VMQSIO3J.js";
 import {
   DEFAULT_CLIP_SEC,
   MEDIA_PRICING,
   MEDIA_PRICING_CHECKED_AT,
-  PRICING,
   PRICING_STALE_AFTER_DAYS,
   getMediaPrice,
-  getPrice,
   listMediaPrices,
   pricingFreshness,
   pricingGeneratedAt
-} from "./chunk-ILZ33HZO.js";
+} from "./chunk-BOZYLPIM.js";
+import {
+  PRICING,
+  getPrice
+} from "./chunk-QVMGFHCF.js";
 
 // src/cost/default-hosts.ts
 var DEFAULT_BASE_URLS = {
@@ -3303,6 +3306,8 @@ var callOptions = {
   override: tierSpecSchema.partial().optional(),
   fallback: z.array(z.union([tierSchema, tierSpecSchema])).optional(),
   purpose: z.string().optional(),
+  /** F077.3 — you MEAN to call a floating "-latest" id (no warning, no strictPinning error). */
+  allowFloating: z.boolean().optional(),
   /** Consumer-defined attribution dimensions (e.g. {tenantId}) ridden into the
    *  cost sink for per-tenant/per-customer cost breakdown (F011). */
   labels: z.record(z.string(), z.string()).optional()
@@ -3525,6 +3530,8 @@ var aiConfigSchema = z.object({
   /** F039.2 — client-wide opt-out from prompt caching (default: on where the
    *  provider supports it). A per-call `promptCache` wins over this. */
   promptCache: z.boolean().optional(),
+  /** F077.3 — refuse any "-latest" model id unless the call passes allowFloating: true. */
+  strictPinning: z.boolean().optional(),
   defaults: z.record(tierSchema, tierSpecSchema).optional(),
   // Functions can't be deeply validated — z.custom asserts the TS type and
   // passes the value through untouched.
@@ -3541,8 +3548,8 @@ var aiConfigSchema = z.object({
 import { randomUUID } from "crypto";
 
 // src/version.ts
-var VERSION = "0.55.3";
-var SDK_TAG = "@broberg/ai-sdk@0.55.3";
+var VERSION = "0.56.1";
+var SDK_TAG = "@broberg/ai-sdk@0.56.1";
 
 // src/cost/sinks/upmetrics.ts
 function isRetryableStatus(status) {
@@ -3883,12 +3890,12 @@ var DEFAULT_BFL_REFERENCE_SPEC = {
   model: "flux-2-max",
   transport: "http"
 };
-var DEFAULT_OCR_SPEC = { provider: "mistral", model: "mistral-ocr-latest", transport: "http" };
-var DEFAULT_MODERATION_SPEC = { provider: "mistral", model: "mistral-moderation-latest", transport: "http" };
-var DEFAULT_JUDGE_SPEC = { provider: "typesafe", model: "jev-latest", transport: "http" };
+var DEFAULT_OCR_SPEC = { provider: "mistral", model: "mistral-ocr-4-1", transport: "http" };
+var DEFAULT_MODERATION_SPEC = { provider: "mistral", model: "mistral-moderation-2603", transport: "http" };
+var DEFAULT_JUDGE_SPEC = { provider: "typesafe", model: "jev-1.13.0", transport: "http" };
 var DEFAULT_PODCAST_SPEC = { provider: "elevenlabs", model: "eleven_v3", transport: "http" };
 var DEFAULT_TTS_SPEC = { provider: "elevenlabs", model: "eleven_multilingual_v2", transport: "http" };
-var DEFAULT_BATCH_SPEC = { provider: "mistral", model: "mistral-small-latest", transport: "http" };
+var DEFAULT_BATCH_SPEC = { provider: "mistral", model: "mistral-small-2603", transport: "http" };
 var ROUTED_SPECS = [
   DEFAULT_IMAGE_SPEC,
   DEFAULT_LORA_IMAGE_SPEC,
@@ -3987,6 +3994,23 @@ function createAI(config = {}) {
     const r = resolveModel(spec.model, { fallback: cfg.availability.fallback, provider: spec.provider });
     return r.fellBack ? { ...spec, model: r.model } : spec;
   }
+  const warnedFloating = /* @__PURE__ */ new Set();
+  function checkPinning(spec, allowFloating) {
+    if (!spec.model?.endsWith("-latest") || allowFloating) return false;
+    const key = `${spec.provider}:${spec.model}`;
+    if (cfg.strictPinning) {
+      throw new Error(
+        `createAI: "${key}" is a floating -latest alias and strictPinning is on. Use a dated id (describeTier() names the pinned ones) or pass allowFloating: true for this call.`
+      );
+    }
+    if (!warnedFloating.has(key)) {
+      warnedFloating.add(key);
+      console.warn(
+        `[@broberg/ai-sdk] "${key}" floats: the provider can move it to another model and price without a release. Pin a dated id, or pass allowFloating: true to say you mean it. (F077)`
+      );
+    }
+    return true;
+  }
   async function runCapability(opts) {
     const routes = [
       applyAvailability(opts.primary),
@@ -3997,11 +4021,13 @@ function createAI(config = {}) {
     let lastErr;
     for (let i = 0; i < routes.length; i++) {
       const spec = routes[i];
+      const floating = checkPinning(spec, opts.allowFloating);
       await preflight(spec, opts.estIn, opts.estOut);
       const t0 = performance.now();
       try {
         const res = await opts.invoke(spec);
         enrich(res.usage, opts.capability, i === 0 ? opts.tier : void 0, opts.purpose, performance.now() - t0, opts.labels);
+        if (floating) res.usage.floating = true;
         await settle(res.usage);
         await report(res.usage);
         return res;
@@ -4057,6 +4083,7 @@ function createAI(config = {}) {
     let lastErr;
     for (let i = 0; i < routes.length; i++) {
       const spec = routes[i];
+      const floating = checkPinning(spec, input.allowFloating);
       await preflight(spec, estIn, estOut);
       const adapter = pickProvider(spec.provider);
       if (!adapter.chatStream) {
@@ -4083,6 +4110,7 @@ function createAI(config = {}) {
           if (ev.type === "text" || ev.type === "tool_call") emitted = true;
           if (ev.type === "usage") {
             enrich(ev.usage, "chat", i === 0 ? tier : void 0, input.purpose, performance.now() - t0, input.labels);
+            if (floating) ev.usage.floating = true;
             await settle(ev.usage);
             await report(ev.usage);
           }
@@ -4128,6 +4156,7 @@ function createAI(config = {}) {
         tier,
         purpose: input.purpose,
         labels: input.labels,
+        allowFloating: input.allowFloating,
         estIn,
         estOut: input.maxTokens ?? 512,
         invoke: async (spec) => {
@@ -4149,6 +4178,7 @@ function createAI(config = {}) {
         tier,
         purpose: input.purpose,
         labels: input.labels,
+        allowFloating: input.allowFloating,
         estIn: estTokens(input.prompt) + 1e3,
         // prompt + ~1k image payload
         estOut: 512,
@@ -4170,6 +4200,7 @@ function createAI(config = {}) {
         tier,
         purpose: input.purpose,
         labels: input.labels,
+        allowFloating: input.allowFloating,
         estIn: estTokens(input.prompt) + 4e3,
         // prompt + video tokens (native video ≈ frames)
         estOut: 512,
@@ -4192,6 +4223,7 @@ function createAI(config = {}) {
         tier,
         purpose: input.purpose,
         labels: input.labels,
+        allowFloating: input.allowFloating,
         estIn,
         estOut: estIn,
         invoke: async (spec) => {
@@ -4216,6 +4248,7 @@ function createAI(config = {}) {
         capability: "image",
         purpose: input.purpose,
         labels: input.labels,
+        allowFloating: input.allowFloating,
         estIn: 0,
         // image cost is not token-based
         estOut: 0,
@@ -4248,6 +4281,7 @@ function createAI(config = {}) {
         capability: "animate",
         purpose: input.purpose,
         labels: input.labels,
+        allowFloating: input.allowFloating,
         estIn: 0,
         // video cost is per-second, not token-based
         estOut: 0,
@@ -4273,6 +4307,7 @@ function createAI(config = {}) {
         capability: "trainStyle",
         purpose: input.purpose,
         labels: input.labels,
+        allowFloating: input.allowFloating,
         estIn: 0,
         // training is priced flat by fal, not token-based
         estOut: 0,
@@ -4299,6 +4334,7 @@ function createAI(config = {}) {
         capability: "ocr",
         purpose: input.purpose,
         labels: input.labels,
+        allowFloating: input.allowFloating,
         estIn: 0,
         // OCR cost is per-page, not token-based
         estOut: 0,
@@ -4318,6 +4354,7 @@ function createAI(config = {}) {
         capability: "moderation",
         purpose: input.purpose,
         labels: input.labels,
+        allowFloating: input.allowFloating,
         estIn: items.reduce((n, s) => n + estTokens(s), 0),
         estOut: 0,
         invoke: async (spec) => {
@@ -4336,6 +4373,7 @@ function createAI(config = {}) {
         capability: "judge",
         purpose: input.purpose,
         labels: input.labels,
+        allowFloating: input.allowFloating,
         estIn: estTokens(stateText) + estTokens(JSON.stringify(input.questions)),
         estOut: 0,
         // output tokens are free on Jev
@@ -4362,6 +4400,7 @@ function createAI(config = {}) {
         capability: "podcast",
         purpose: input.purpose,
         labels: input.labels,
+        allowFloating: input.allowFloating,
         estIn: chars,
         // per-character cost (not token-based)
         estOut: 0,
@@ -4381,6 +4420,7 @@ function createAI(config = {}) {
         capability: "tts",
         purpose: input.purpose,
         labels: input.labels,
+        allowFloating: input.allowFloating,
         estIn: input.text.length,
         // per-character cost
         estOut: 0,
@@ -4411,6 +4451,7 @@ function createAI(config = {}) {
         tier,
         purpose: input.purpose,
         labels: input.labels,
+        allowFloating: input.allowFloating,
         estIn: text.reduce((n, t) => n + estTokens(t), 0),
         estOut: 0,
         invoke: async (spec) => {
@@ -4429,6 +4470,7 @@ function createAI(config = {}) {
         capability: "transcribe",
         purpose: input.purpose,
         labels: input.labels,
+        allowFloating: input.allowFloating,
         estIn: 0,
         estOut: 0,
         invoke: async (spec) => {
@@ -5091,6 +5133,7 @@ export {
   deepseekAdapter,
   defaultBaseUrl,
   defaultProviders,
+  describeTier,
   discordSink,
   elevenlabsAdapter,
   embeddingInputSchema,