@hiper2d/ai-agents 0.17.0 → 0.19.0

package/dist/index.mjs

@@ -713,6 +713,14 @@ var ModelInvalidResponseError = class extends ModelError {
     this.truncated = truncated;
   }
 };
+var ModelTimeoutError = class extends ModelError {
+  timeoutMs;
+  constructor(modelType, timeoutMs) {
+    super(`${modelType} did not respond within ${Math.round(timeoutMs / 1e3)}s`, modelType);
+    this.name = "ModelTimeoutError";
+    this.timeoutMs = timeoutMs;
+  }
+};
 var ModelRefusalError = class extends ModelError {
   reason;
   constructor(modelType, message = `${modelType} refused to answer (stop_reason: refusal)`, reason) {
@@ -798,6 +806,9 @@ var LLM_CONSTANTS = {
   GEMINI_PRO: "gemini-pro",
   GEMINI_FLASH: "gemini-flash",
   GEMINI_LITE: "gemini-lite",
+  // Reused 2026-10-06 for Mistral Large 4: the old id (Large 3) was retired 2026-09-18 and
+  // consumers had mapped it away, so the version-free id was free again.
+  MISTRAL_LARGE: "mistral-large",
   MISTRAL_MEDIUM: "mistral-medium",
   MISTRAL_SMALL: "mistral-small",
   GROK: "grok",
@@ -816,6 +827,7 @@ var LLM_CONSTANTS = {
   MUSE_SPARK: "muse-spark"
 };
 var DEFAULT_MAX_OUTPUT_TOKENS = 8192;
+var DEFAULT_REQUEST_TIMEOUT_MS = 6e4;
 var SupportedAiModels = {
   // Claude Fable 5.1 (2026-09-02, was Fable 5) - frontier reasoning model. Thinking is always on (no non-thinking variant).
   [LLM_CONSTANTS.CLAUDE_FABLE]: {
@@ -911,9 +923,11 @@ var SupportedAiModels = {
   // ($4/$20 → $2/$10) and Luna halved ($0.20/$1.20 → $0.10/$0.50). Note GPT-6 Sol now
   // undercuts GPT-5.6 Terra ($2/$12) on output at the same input rate, which makes the Terra
   // slot largely redundant — retiring it is a product call, not a catalog one.
+  // Sol → GPT-6.1 Sol on 2026-09-29 (id verified against GET /v1/models; gpt-6-sol is still
+  // served). Same price as GPT-6 Sol; its no-param default effort still measures `medium`.
   [LLM_CONSTANTS.GPT_SOL]: {
-    displayName: "GPT-6 Sol",
-    modelApiName: "gpt-6-sol",
+    displayName: "GPT-6.1 Sol",
+    modelApiName: "gpt-6.1-sol",
     apiKeyName: API_KEY_CONSTANTS.OPENAI,
     reasoningEffort: "medium",
     hasThinking: true,
@@ -1002,6 +1016,21 @@ var SupportedAiModels = {
   // returns `mistral-medium-3-5`, `mistral-medium-3`, `mistral-medium-latest` and
   // `magistral-medium-latest` as one entry — so Magistral seats had silently become Medium
   // seats already. Consumers map both retired ids onto the surviving pair.
+  //
+  // Mistral Large 4 added 2026-10-06 under the freed `mistral-large` picker id. The model list
+  // reports reasoning: true (Large 3 said false), 512K context; same reasoning_effort switch,
+  // but only 'high' and 'none' are accepted (medium/low are a 400). Live 2026-10-06 at high:
+  // a chat-turn schema ask took 14-21s with 1.3-2k output tokens (~$0.003), one constrained
+  // pick took 49s — it thinks long, so it is cheap but slow.
+  [LLM_CONSTANTS.MISTRAL_LARGE]: {
+    displayName: "Mistral Large 4",
+    // `mistral-large-4-0` is the same entry; `mistral-large-latest` still points at Large 3.
+    modelApiName: "mistral-large-4",
+    apiKeyName: API_KEY_CONSTANTS.MISTRAL,
+    hasThinking: true,
+    reasoningEffort: "high",
+    tags: ["slow", "cheap"]
+  },
   [LLM_CONSTANTS.MISTRAL_MEDIUM]: {
     displayName: "Mistral Medium 3.5",
     // The docs' explicit 3.5 id. It was `mistral-medium-3` until 2026-09-18, which Mistral
@@ -1265,6 +1294,7 @@ var MODEL_PRICING = {
   // As with Astra, OpenAI's table doesn't restate the short/long boundary, so we keep the
   // 272k threshold the 5.6 siblings use. Cache WRITES (Sol $2.50/$5, Luna $0.125/$0.25) are
   // not modelled: OpenAI caching is automatic and the API only reports hits.
+  // GPT-6.1 Sol (2026-09-29) kept GPT-6 Sol's rates exactly, cache writes included.
   [SupportedAiModels[LLM_CONSTANTS.GPT_SOL].modelApiName]: {
     inputPrice: 2,
     outputPrice: 10,
@@ -1390,7 +1420,13 @@ var MODEL_PRICING = {
   // Mistral models (mistral.ai/pricing/api, verified 2026-09-18). Cached tokens bill at 10% of
   // the input price (documented on the prompt_cache_key param; no per-model cached prices
   // published). Reasoning tokens are counted inside completion_tokens — no separate
-  // reasoning_tokens field — so the output rate already covers the trace.
+  // reasoning_tokens field — so the output rate already covers the trace. Large 4 added
+  // 2026-10-06 at its published rates ($0.07 cached, not an exact 10%).
+  [SupportedAiModels[LLM_CONSTANTS.MISTRAL_LARGE].modelApiName]: {
+    inputPrice: 0.68,
+    outputPrice: 2.09,
+    cacheHitPrice: 0.07
+  },
   [SupportedAiModels[LLM_CONSTANTS.MISTRAL_MEDIUM].modelApiName]: {
     inputPrice: 1.5,
     outputPrice: 7.5,
@@ -1488,6 +1524,7 @@ var HYBRID_THINKING_API_NAMES = /* @__PURE__ */ new Set([
   SupportedAiModels[LLM_CONSTANTS.MINIMAX].modelApiName,
   // Mistral reasons only when asked (`reasoning_effort`), and we always ask — hybrid by the
   // same definition as Qwen. Added 2026-09-18 when the two entries went thinking-on.
+  SupportedAiModels[LLM_CONSTANTS.MISTRAL_LARGE].modelApiName,
   SupportedAiModels[LLM_CONSTANTS.MISTRAL_MEDIUM].modelApiName,
   SupportedAiModels[LLM_CONSTANTS.MISTRAL_SMALL].modelApiName
 ]);
@@ -2224,6 +2261,12 @@ var AbstractAgent = class {
    */
   reasoningEffort;
   thinkingBudgetTokens;
+  /**
+   * Wall-clock limit on one ask (DEFAULT_REQUEST_TIMEOUT_MS). Past it the ask throws
+   * ModelTimeoutError; the provider request is abandoned, not cancelled. Raise it per
+   * instance for calls that legitimately run long, like maxOutputTokens.
+   */
+  requestTimeoutMs = DEFAULT_REQUEST_TIMEOUT_MS;
   instruction;
   /**
    * The instruction split on CACHE_TIER_MARKER: [shared static tier, per-bot tier].
@@ -2258,7 +2301,7 @@ var AbstractAgent = class {
     if (beforeAskHook) await beforeAskHook(this);
     const startedAt = Date.now();
     try {
-      const [result, thinking, usage, signature] = await this.doAskWithZodSchema(zodSchema, messages);
+      const [result, thinking, usage, signature] = await this.withTimeout(this.doAskWithZodSchema(zodSchema, messages));
       return [result, thinking, this.stampDuration(usage, startedAt), signature];
     } catch (error) {
       this.stampErrorDuration(error, startedAt);
@@ -2269,13 +2312,25 @@ var AbstractAgent = class {
     if (beforeAskHook) await beforeAskHook(this);
     const startedAt = Date.now();
     try {
-      const [content, thinking, usage, signature] = await this.doAskText(messages);
+      const [content, thinking, usage, signature] = await this.withTimeout(this.doAskText(messages));
       return [content, thinking, this.stampDuration(usage, startedAt), signature];
     } catch (error) {
       this.stampErrorDuration(error, startedAt);
       throw error;
     }
   }
+  /**
+   * Races the provider call against requestTimeoutMs. A hung provider (accepts the request,
+   * never answers) otherwise runs until the host kills the function with no error to show.
+   */
+  withTimeout(call) {
+    const timeoutMs = this.requestTimeoutMs;
+    let timer;
+    const timeout = new Promise((_, reject) => {
+      timer = setTimeout(() => reject(new ModelTimeoutError(this.model, timeoutMs)), timeoutMs);
+    });
+    return Promise.race([call, timeout]).finally(() => clearTimeout(timer));
+  }
   stampDuration(usage, startedAt) {
     return usage ? { ...usage, durationMs: Date.now() - startedAt } : usage;
   }
@@ -5291,6 +5346,7 @@ var AgentFactory = class {
       case LLM_CONSTANTS.DEEPSEEK_PRO:
         return new DeepSeekV2Agent(name, instruction, model.modelApiName, key, model.temperature ?? 0, shouldEnableThinking);
       // Mistral models
+      case LLM_CONSTANTS.MISTRAL_LARGE:
       case LLM_CONSTANTS.MISTRAL_MEDIUM:
       case LLM_CONSTANTS.MISTRAL_SMALL:
         return new MistralAgent(name, instruction, model.modelApiName, key, shouldEnableThinking);
@@ -5340,6 +5396,7 @@ export {
   DEEPSEEK_REASONING_EFFORTS,
   DEFAULT_LOGGING_CONFIG,
   DEFAULT_MAX_OUTPUT_TOKENS,
+  DEFAULT_REQUEST_TIMEOUT_MS,
   DeepSeekV2Agent,
   FUGU_REASONING_EFFORTS,
   FuguAgent,
@@ -5369,6 +5426,7 @@ export {
   ModelQuotaExceededError,
   ModelRateLimitError,
   ModelRefusalError,
+  ModelTimeoutError,
   ModelUnavailableError,
   OPENAI_REASONING_EFFORTS,
   OpenAiVoiceAgent,