@hiper2d/ai-agents 0.19.2 → 0.20.0

package/dist/index.d.mts

@@ -1298,7 +1298,7 @@ declare abstract class AbstractAgent {
      * 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?: ReasoningEffort;

package/dist/index.d.ts

@@ -1298,7 +1298,7 @@ declare abstract class AbstractAgent {
      * 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?: ReasoningEffort;

package/dist/index.js

@@ -1038,13 +1038,19 @@ 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
@@ -1562,9 +1568,16 @@ 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]: {
@@ -2414,7 +2427,7 @@ 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;
@@ -2814,6 +2827,7 @@ ${instruction}`);
 var import_sdk = require("@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;
   /**
@@ -2835,7 +2849,8 @@ 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() {
@@ -3020,16 +3035,16 @@ ${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)) {
@@ -3118,16 +3133,16 @@ ${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)) {