@hiper2d/ai-agents 0.19.1 → 0.20.0

package/dist/index.mjs

@@ -860,13 +860,19 @@ var SupportedAiModels = {
     reasoningEffort: "high",
     tags: ["expensive"]
   },
+  // Haiku moved 4.5 → 5.5 on 2026-10-08 (`claude-haiku-5-5`, verified live with the platform
+  // key). 10x cheaper ($0.10/$0.50 up to 100K-token prompts), 1M context / 128K output.
+  // Adaptive thinking now (the `enabled` + budget_tokens mode 4.5 used is rejected), API
+  // default effort `medium`; we pin `high` like the other Claude entries. It reads Haiku 4.5
+  // thinking blocks, so games moving onto it keep the bots' earlier reasoning. 'slow' tag
+  // dropped: Anthropic lists it as the fastest model in the lineup.
   [LLM_CONSTANTS.CLAUDE_HAIKU]: {
-    displayName: "Claude 4.5 Haiku",
-    modelApiName: "claude-haiku-4-5",
+    displayName: "Claude 5.5 Haiku",
+    modelApiName: "claude-haiku-5-5",
     apiKeyName: API_KEY_CONSTANTS.ANTHROPIC,
     hasThinking: true,
-    thinkingBudgetTokens: 1024,
-    tags: ["slow", "cheap"]
+    reasoningEffort: "high",
+    tags: ["cheap"]
   },
   // DeepSeek V4 models — thinking-only entries (non-thinking variants retired 2026-08-05).
   // reasoningEffort pinned to 'low' 2026-08-30: at the provider default ('high', no budget
@@ -961,6 +967,10 @@ var SupportedAiModels = {
     apiKeyName: API_KEY_CONSTANTS.GOOGLE,
     hasThinking: true,
     reasoningEffort: "high",
+    // Pre-3.6: still honours sampling params, so it keeps the 0.2 GoogleAgent always sent
+    // before 2026-10-07. 3.6+ ignore them and future Gemini models reject them — leave
+    // `temperature` unset on every newer Gemini entry (GoogleAgent then sends none).
+    temperature: 0.2,
     tags: ["expensive"]
   },
   [LLM_CONSTANTS.GEMINI_FLASH]: {
@@ -980,6 +990,8 @@ var SupportedAiModels = {
     apiKeyName: API_KEY_CONSTANTS.GOOGLE,
     hasThinking: true,
     reasoningEffort: "minimal",
+    // Pre-3.6, same as Pro: keeps the 0.2 it always had.
+    temperature: 0.2,
     tags: ["fast", "cheap"]
   },
   // Always-on reasoning — no non-thinking sibling. Bumped 4.6 → 4.7 on 2026-09-22 (id verified
@@ -1378,9 +1390,16 @@ var MODEL_PRICING = {
     cacheHitPrice: 0.2
   },
   [SupportedAiModels[LLM_CONSTANTS.CLAUDE_HAIKU].modelApiName]: {
-    inputPrice: 1,
-    outputPrice: 5,
-    cacheHitPrice: 0.1
+    // Haiku 5.5 (2026-10-07): the first Claude with a long-context tier — every rate is 5x
+    // for prompts over 100K tokens (cache reads + writes count toward the prompt size,
+    // which is what AnthropicAgent passes as inputTokens).
+    inputPrice: 0.1,
+    outputPrice: 0.5,
+    cacheHitPrice: 0.01,
+    extendedContextInputPrice: 0.5,
+    extendedContextOutputPrice: 2.5,
+    extendedContextCacheHitPrice: 0.05,
+    extendedContextThresholdTokens: 1e5
   },
   // Google models
   [SupportedAiModels[LLM_CONSTANTS.GEMINI_PRO].modelApiName]: {
@@ -2230,7 +2249,7 @@ var AbstractAgent = class {
    * Reasoning-depth knobs, resolved once from the catalog like maxOutputTokens and, like it,
    * overridable per instance for calls whose profile differs from a turn (story generation
    * runs deeper). Each provider speaks one dialect — effort (DeepSeek, GLM, Gemini, Claude
-   * adaptive) or a token budget (Qwen, Claude Haiku) — and reads only the field it
+   * adaptive) or a token budget (Qwen, Claude Haiku 4.5) — and reads only the field it
    * understands; the other is ignored. Subclasses read these when building a request.
    */
   reasoningEffort;
@@ -2630,6 +2649,7 @@ ${instruction}`);
 import { Anthropic } from "@anthropic-ai/sdk";
 var rejectsDisabledThinking = (model) => model.includes("fable") || model.includes("opus-5-5");
 var usesBetweenToolsForOff = (model) => model.includes("sonnet-5-5");
+var usesAdaptiveThinking = (model) => !model.includes("haiku-4");
 var ClaudeAgent = class extends AbstractAgent {
   client;
   /**
@@ -2651,7 +2671,8 @@ var ClaudeAgent = class extends AbstractAgent {
   //             state writes (a lynch, the night resolution, the summary rewrite, the new
   //             day), so every call inside one of those windows reads it.
   // GM prompts have no marker → single block, same behavior as before. Haiku 4.5 needs a
-  // 4096-token cacheable prefix, so tiers below that silently no-op on Haiku — expected.
+  // 4096-token cacheable prefix, so tiers below that silently no-op there; Haiku 5.5 caches
+  // from ~1024 tokens (measured 2026-10-08: a 1,083-token prefix was written, 492 was not).
   // A getter, not a field: `maxOutputTokens` can be raised after construction, and a field
   // initializer would snapshot the default and silently ignore the override.
   get defaultParams() {
@@ -2836,16 +2857,16 @@ ${schemaDescription}`;
         ...this.defaultParams,
         messages: anthropicMessages
       };
-      const usesAdaptiveThinking = this.model.includes("fable") || this.model.includes("opus") || this.model.includes("sonnet");
+      const adaptive = usesAdaptiveThinking(this.model);
       if (canUseThinking) {
-        if (usesAdaptiveThinking) {
+        if (adaptive) {
           params.thinking = { type: "adaptive", display: "summarized" };
           params.output_config = { effort: toAnthropicEffort(this.reasoningEffort ?? "high") };
         } else {
           params.thinking = { type: "enabled", budget_tokens: this.thinkingBudgetTokens ?? 1024 };
           params.temperature = 1;
         }
-      } else if (usesAdaptiveThinking) {
+      } else if (adaptive) {
         if (usesBetweenToolsForOff(this.model)) {
           params.thinking = { type: "between_tools" };
         } else if (!rejectsDisabledThinking(this.model)) {
@@ -2934,16 +2955,16 @@ ${schemaDescription}`;
         ...this.defaultParams,
         messages: anthropicMessages
       };
-      const usesAdaptiveThinking = this.model.includes("fable") || this.model.includes("opus") || this.model.includes("sonnet");
+      const adaptive = usesAdaptiveThinking(this.model);
       if (canUseThinking) {
-        if (usesAdaptiveThinking) {
+        if (adaptive) {
           params.thinking = { type: "adaptive", display: "summarized" };
           params.output_config = { effort: toAnthropicEffort(this.reasoningEffort ?? "high") };
         } else {
           params.thinking = { type: "enabled", budget_tokens: this.thinkingBudgetTokens ?? 1024 };
           params.temperature = 1;
         }
-      } else if (usesAdaptiveThinking) {
+      } else if (adaptive) {
         if (usesBetweenToolsForOff(this.model)) {
           params.thinking = { type: "between_tools" };
         } else if (!rejectsDisabledThinking(this.model)) {
@@ -3057,6 +3078,13 @@ function throwForEmptyGeminiResponse(model, response, fallback) {
 }
 var GoogleAgent = class extends AbstractAgent {
   client;
+  /**
+   * Sampling temperature to send, or undefined to send none. Google deprecated custom sampling
+   * (notice of 2026-10): since Gemini 3.6 Flash temperature/top_p/top_k are ignored, and its
+   * upcoming models reject them with a 400. So it is sent only for a catalog entry that sets
+   * `temperature` (the pre-3.6 models, to keep their behaviour) — never by default.
+   */
+  samplingTemperature;
   defaultConfig = {
     responseMimeType: "application/json"
   };
@@ -3072,7 +3100,9 @@ var GoogleAgent = class extends AbstractAgent {
     unsupportedRole: (role) => `Unsupported role type: ${role}`
   };
   constructor(name, instruction, model, apiKey, enableThinking = false, agentLoggingConfig = DEFAULT_LOGGING_CONFIG.agents) {
-    super(name, instruction, model, 0.2, enableThinking, agentLoggingConfig);
+    const samplingTemperature = getModelConfigByApiName(model)?.temperature;
+    super(name, instruction, model, samplingTemperature ?? 0, enableThinking, agentLoggingConfig);
+    this.samplingTemperature = samplingTemperature;
     this.client = new GoogleGenAI3({
       apiKey
     });
@@ -3171,7 +3201,7 @@ var GoogleAgent = class extends AbstractAgent {
     try {
       const googleSchema = ZodSchemaConverter.toGoogleSchema(zodSchema);
       const config = {
-        temperature: this.temperature,
+        ...this.samplingTemperature !== void 0 ? { temperature: this.samplingTemperature } : {},
         responseMimeType: "application/json",
         responseSchema: googleSchema,
         maxOutputTokens: this.maxOutputTokens,
@@ -3261,7 +3291,7 @@ var GoogleAgent = class extends AbstractAgent {
     const contents = this.convertToContents(messages);
     try {
       const config = {
-        temperature: this.temperature,
+        ...this.samplingTemperature !== void 0 ? { temperature: this.samplingTemperature } : {},
         maxOutputTokens: this.maxOutputTokens,
         systemInstruction: this.instruction
       };