@broberg/ai-sdk 0.28.0 → 0.30.0

package/dist/index.js

@@ -1,34 +1,16 @@
 import {
+  DEFAULT_TIER_MAP,
   ModelUnavailableError,
   listModels,
   providerIds,
   resetRegistry,
   resolveModel,
+  resolveTier,
   setAvailability
-} from "./chunk-IT7HNKLY.js";
+} from "./chunk-V2PD522L.js";
 import {
   getPrice
-} from "./chunk-LUPUAAHZ.js";
-
-// src/routing/tier-map.ts
-var DEFAULT_TIER_MAP = {
-  fast: { provider: "mistral", model: "mistral-small-latest", transport: "http" },
-  smart: { provider: "mistral", model: "mistral-large-latest", transport: "http" },
-  powerful: { provider: "mistral", model: "mistral-large-latest", transport: "http" },
-  cheap: { provider: "mistral", model: "mistral-small-latest", transport: "http" },
-  // Vision: small-latest (vision-capable, cheap EU) is the default; override to
-  // mistral-large-latest for demanding image/spatial/composition work.
-  vision: { provider: "mistral", model: "mistral-small-latest", transport: "http" },
-  // Native video understanding — Gemini leads; flash-lite is the cheap default (F019).
-  // NOT Anthropic → out of the F030 phase-out (its own EU epic if/when needed).
-  video: { provider: "gemini", model: "gemini-2.5-flash-lite", transport: "http" },
-  // NOT Anthropic → out of F030 (EU-embedding migration is its own future epic).
-  embedding: { provider: "openai", model: "text-embedding-3-small", transport: "http" }
-};
-function resolveTier(tier, override, configMap) {
-  const base = configMap?.[tier] ?? DEFAULT_TIER_MAP[tier];
-  return { ...base, ...override };
-}
+} from "./chunk-LKVCPMVI.js";
 
 // src/transport/http.ts
 async function httpTransport(req) {
@@ -541,6 +523,7 @@ function makeOpenAICompatibleAdapter(config) {
     if (req.maxTokens !== void 0) body.max_tokens = req.maxTokens;
     if (req.temperature !== void 0) body.temperature = req.temperature;
     if (req.responseFormat === "json") body.response_format = { type: "json_object" };
+    if (req.promptCacheKey !== void 0) body.prompt_cache_key = req.promptCacheKey;
     if (config.costFromResponseField) body.usage = { include: true };
     const res = await httpTransport({
       spec: req.spec,
@@ -563,13 +546,19 @@ function makeOpenAICompatibleAdapter(config) {
     const toolCalls = msg?.tool_calls?.map(
       (tc) => fromProviderToolCall(tc, "openai")
     );
+    const cachedIn = data.usage?.prompt_tokens_details?.cached_tokens;
     const usage = freshUsage({
       provider: config.name,
       model: req.spec.model,
       transport: "http",
       capability: "chat",
-      inputTokens: data.usage?.prompt_tokens ?? 0,
-      outputTokens: data.usage?.completion_tokens ?? 0
+      // prompt_tokens INCLUDES the cached ones; computeCost adds cacheReadTokens on
+      // top of inputTokens, so billing the raw figure would charge the cached prefix
+      // twice — at full rate AND at the cache rate.
+      inputTokens: (data.usage?.prompt_tokens ?? 0) - (cachedIn ?? 0),
+      outputTokens: data.usage?.completion_tokens ?? 0,
+      // Absent field stays undefined: "not reported" is not the same as "zero cached".
+      ...cachedIn === void 0 ? {} : { cacheReadTokens: cachedIn }
     });
     if (config.costFromResponseField && typeof data.usage?.cost === "number") {
       usage.costUsd = data.usage.cost;
@@ -2637,6 +2626,9 @@ var callOptions = {
   labels: z.record(z.string(), z.string()).optional()
 };
 var chatInputSchema = z.object({
+  /** F039 — reuse a cached prompt prefix on Mistral at 10% of the input rate.
+   *  A stable id for "the same conversation" (conversation id, session id). */
+  promptCacheKey: z.string().optional(),
   prompt: z.string().optional(),
   messages: z.array(messageSchema).optional(),
   system: z.string().optional(),
@@ -2761,7 +2753,15 @@ var ttsInputSchema = z.object({
   text: z.string(),
   voice: z.string(),
   /** F037: voice to use if `voice` is one we know the provider has retired. Without
-   *  it a retired voice throws VoiceUnavailableError rather than reaching the API. */
+   *  it a retired voice throws VoiceUnavailableError rather than reaching the API.
+   *
+   *  THINK BEFORE SETTING THIS. It is right for batch work where any acceptable voice
+   *  will do (a generated podcast, a bulk render). It is WRONG where the voice is an
+   *  identity a human recognises — a brand voice, a house narrator: falling back means
+   *  your product suddenly speaks with a stranger's voice and says nothing about why.
+   *  There, leave it unset and let the throw alarm you. Silence WITH a message beats
+   *  the wrong audio WITHOUT one. (torrent-search-api's call on their own house voice,
+   *  2026-08-11 — they took the alarm over the fallback, deliberately.) */
   voiceFallback: z.string().optional(),
   lang: z.string().optional(),
   format: z.string().optional(),
@@ -2790,8 +2790,8 @@ var aiConfigSchema = z.object({
 });
 
 // src/version.ts
-var VERSION = "0.28.0";
-var SDK_TAG = "@broberg/ai-sdk@0.28.0";
+var VERSION = "0.30.0";
+var SDK_TAG = "@broberg/ai-sdk@0.30.0";
 
 // src/cost/sinks/upmetrics.ts
 function upmetricsSink(config) {
@@ -3075,7 +3075,7 @@ function createAI(config = {}) {
         invoke: async (spec) => {
           const adapter = pickProvider(spec.provider);
           if (!adapter.chat) throw new Error(`createAI: provider "${spec.provider}" does not support chat`);
-          return adapter.chat({ messages, spec, tools: input.tools, maxTokens: input.maxTokens, temperature: input.temperature, responseFormat: input.responseFormat });
+          return adapter.chat({ messages, spec, tools: input.tools, maxTokens: input.maxTokens, temperature: input.temperature, responseFormat: input.responseFormat, promptCacheKey: input.promptCacheKey });
         }
       });
     },