@hiper2d/ai-agents 0.19.2 → 0.21.0

package/dist/index.d.mts

@@ -375,7 +375,6 @@ declare const LLM_CONSTANTS: {
     GEMINI_PRO: string;
     GEMINI_FLASH: string;
     GEMINI_LITE: string;
-    MISTRAL_MEDIUM: string;
     MISTRAL_SMALL: string;
     GROK: string;
     KIMI: string;
@@ -1298,7 +1297,7 @@ 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

@@ -375,7 +375,6 @@ declare const LLM_CONSTANTS: {
     GEMINI_PRO: string;
     GEMINI_FLASH: string;
     GEMINI_LITE: string;
-    MISTRAL_MEDIUM: string;
     MISTRAL_SMALL: string;
     GROK: string;
     KIMI: string;
@@ -1298,7 +1297,7 @@ 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

@@ -984,7 +984,6 @@ var LLM_CONSTANTS = {
   GEMINI_PRO: "gemini-pro",
   GEMINI_FLASH: "gemini-flash",
   GEMINI_LITE: "gemini-lite",
-  MISTRAL_MEDIUM: "mistral-medium",
   MISTRAL_SMALL: "mistral-small",
   GROK: "grok",
   KIMI: "kimi",
@@ -1038,13 +1037,19 @@ 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
@@ -1184,14 +1189,18 @@ var SupportedAiModels = {
     reasoningEffort: "medium",
     temperature: 0.7
   },
-  // Mistral models. Two hybrid entries since 2026-09-18: reasoning is off by default on the
-  // API and switched on per request with `reasoning_effort` (see mistral-agent.ts); the trace
-  // comes back as a `thinking` content chunk and works with json_schema structured output
-  // (verified live 2026-09-18 on both). Pinned to 'high' — the only level Mistral's docs
-  // describe ("full thinking chunk before the final answer"); Small is cheap enough that the
-  // extra tokens don't matter, Medium's are priced into its free-tier band via the hybrid
-  // multiplier. Live 2026-09-18 (one short schema ask each): Small 1.7s with a ~600-char
-  // trace, Medium 3.8s with ~2000 chars — Medium's trace ran ~15x its answer length.
+  // Mistral models. One hybrid entry: reasoning is off by default on the API and switched on
+  // per request with `reasoning_effort` (see mistral-agent.ts); the trace comes back as a
+  // `thinking` content chunk and works with json_schema structured output (verified live
+  // 2026-09-18). Pinned to 'high' — the only level Mistral's docs describe ("full thinking
+  // chunk before the final answer"); Small is cheap enough that the extra tokens don't
+  // matter. Live 2026-09-18 (one short schema ask): 1.7s with a ~600-char trace.
+  //
+  // Mistral Medium 3.5 (`mistral-medium-3-5`, $1.5/$7.5, $0.15 cached) was removed in 0.21.0
+  // (2026-10-10): too slow and too expensive at the only reasoning level it takes (high|none).
+  // Production 2026-10-03..10: p50 15.9s, p90 35.1s, $0.036 a request (2x Gemini 3.1 Pro,
+  // 19x Small; reasoning spiked to 15-24k tokens), ~6% of calls failed (60s timeouts, empty
+  // replies, empty JSON), and a 15-player story pipeline took 277s. Wait for Medium 4.
   //
   // Mistral Large 3 and Magistral Medium 1.2 were dropped 2026-09-18. Large 3 was retired by
   // Mistral on 2026-08-31 (the `-latest` alias still answered, but its capabilities say
@@ -1199,7 +1208,7 @@ var SupportedAiModels = {
   // and `magistral-medium-latest` is now literally an alias of Medium 3.5 — the model list
   // 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.
+  // seats already. Consumers map both retired ids onto Small (Medium itself is gone since 0.21.0).
   //
   // Mistral Large 4 (`mistral-large-4`, $0.68/$2.09, $0.07 cached) was added in 0.19.0 and
   // removed in 0.19.1 (2026-10-06/07): too slow to play. Like Small/Medium it accepts only
@@ -1207,16 +1216,6 @@ var SupportedAiModels = {
   // a realistic day-2 vote took 25-154s (half past the 60s timeout) and its parallel story
   // batches finished ~100s apart (Mistral served them one at a time: ~6 min per story). Off,
   // it answers in ~1s with no trace. Re-test before re-adding.
-  [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
-    // now serves as an alias of the same model (Medium 3 itself retired 2026-08-31).
-    modelApiName: "mistral-medium-3-5",
-    apiKeyName: API_KEY_CONSTANTS.MISTRAL,
-    hasThinking: true,
-    reasoningEffort: "high",
-    tags: ["fast", "expensive"]
-  },
   [LLM_CONSTANTS.MISTRAL_SMALL]: {
     displayName: "Mistral 4 Small",
     // Resolves to mistral-small-2603 (Small 4), which also carries the magistral-small alias.
@@ -1562,9 +1561,16 @@ 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]: {
@@ -1597,11 +1603,6 @@ var MODEL_PRICING = {
   // 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.
-  [SupportedAiModels[LLM_CONSTANTS.MISTRAL_MEDIUM].modelApiName]: {
-    inputPrice: 1.5,
-    outputPrice: 7.5,
-    cacheHitPrice: 0.15
-  },
   [SupportedAiModels[LLM_CONSTANTS.MISTRAL_SMALL].modelApiName]: {
     inputPrice: 0.15,
     outputPrice: 0.6,
@@ -1682,8 +1683,7 @@ var HYBRID_THINKING_API_NAMES = /* @__PURE__ */ new Set([
   SupportedAiModels[LLM_CONSTANTS.QWEN_FLASH].modelApiName,
   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_MEDIUM].modelApiName,
+  // same definition as Qwen. Added 2026-09-18 when the Mistral entries went thinking-on.
   SupportedAiModels[LLM_CONSTANTS.MISTRAL_SMALL].modelApiName
 ]);
 function isHybridThinkingModel(modelApiName) {
@@ -2414,7 +2414,7 @@ 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 +2814,7 @@ ${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 +2836,8 @@ 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 +3022,16 @@ ${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 +3120,16 @@ ${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)) {
@@ -5513,7 +5515,6 @@ 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_MEDIUM:
       case LLM_CONSTANTS.MISTRAL_SMALL:
         return new MistralAgent(name, instruction, model.modelApiName, key, shouldEnableThinking);
       case LLM_CONSTANTS.KIMI: