@hiper2d/ai-agents 0.23.0 → 0.24.1

package/dist/index.d.mts

@@ -386,6 +386,7 @@ declare const LLM_CONSTANTS: {
     QWEN_FLASH: string;
     MINIMAX: string;
     MUSE_SPARK: string;
+    MUSE_WONDER: string;
 };
 /**
  * Per-request output ceiling for ordinary requests. Reasoning tokens are billed inside this

package/dist/index.d.ts

@@ -386,6 +386,7 @@ declare const LLM_CONSTANTS: {
     QWEN_FLASH: string;
     MINIMAX: string;
     MUSE_SPARK: string;
+    MUSE_WONDER: string;
 };
 /**
  * Per-request output ceiling for ordinary requests. Reasoning tokens are billed inside this

package/dist/index.js

@@ -999,7 +999,9 @@ var LLM_CONSTANTS = {
   // MiniMax. Single M3 entry; stable id without the version for the same repoint reason.
   MINIMAX: "minimax",
   // Meta Model API (api.meta.ai). Muse Spark; version-free id so a 1.3 → 1.4 repoint is entry-only.
-  MUSE_SPARK: "muse-spark"
+  MUSE_SPARK: "muse-spark",
+  // Muse Wonder, Meta's larger model above Spark (added 2026-10-13); same version-free pattern.
+  MUSE_WONDER: "muse-wonder"
 };
 var DEFAULT_MAX_OUTPUT_TOKENS = 8192;
 var DEFAULT_REQUEST_TIMEOUT_MS = 6e4;
@@ -1147,12 +1149,17 @@ var SupportedAiModels = {
   // ceiling, so 'high' leaves the dynamic range open. No `temperature`: Gemini 4 accepts it
   // but Google deprecated sampling params (see the Pro note). Story-pipeline speed as Game
   // Master not measured yet — run the host's live story suite before trusting it there.
+  // Pinned medium 2026-10-14 (was high since it shipped on 2026-10-13). Two days of
+  // production turns at high measured p50 21s / p90 40s with three 60s timeouts, and the
+  // werewolf day-2 vote fixture took 16-34s (380-415 output tokens). At medium the same
+  // prompt took 4-11s (75-110 output tokens) with the same vote target and a reason just
+  // as specific — the same trade Grok 4.7 made on 2026-10-06.
   [LLM_CONSTANTS.GEMINI_ARGON]: {
     displayName: "Gemini 4 Argon",
     modelApiName: "gemini-4-argon",
     apiKeyName: API_KEY_CONSTANTS.GOOGLE,
     hasThinking: true,
-    reasoningEffort: "high",
+    reasoningEffort: "medium",
     tags: ["expensive"]
   },
   [LLM_CONSTANTS.GEMINI_PRO]: {
@@ -1405,6 +1412,20 @@ var SupportedAiModels = {
     temperature: 1,
     reasoningEffort: "medium",
     tags: ["slow"]
+  },
+  // Muse Wonder 1.0 (added 2026-10-13, id verified against GET /v1/models): Meta's model above
+  // Spark, Standard mode only (no Contributor tier). Same MetaAgent and effort vocabulary as
+  // Spark; probed low/medium/high on a short game prompt: 347/455/565 reasoning tokens in
+  // 5.9/7.3/8.5s, so the pin stays 'medium' like Spark. Game-length turns and the story
+  // pipeline are unmeasured — not tagged 'slow' yet; the host decides on Game Master use.
+  [LLM_CONSTANTS.MUSE_WONDER]: {
+    displayName: "Muse Wonder 1.0",
+    modelApiName: "muse-wonder-1.0",
+    apiKeyName: API_KEY_CONSTANTS.META,
+    hasThinking: true,
+    temperature: 1,
+    reasoningEffort: "medium",
+    tags: ["expensive"]
   }
 };
 function createCatalog(overrides = {}) {
@@ -1693,6 +1714,13 @@ var MODEL_PRICING = {
     inputPrice: 1.25,
     outputPrice: 4.25,
     cacheHitPrice: 0.15
+  },
+  // Standard mode (Meta Model API pricing page, 2026-10-13); Contributor mode is not offered
+  // for Wonder. Cached input at 5% of input.
+  [SupportedAiModels[LLM_CONSTANTS.MUSE_WONDER].modelApiName]: {
+    inputPrice: 4,
+    outputPrice: 20,
+    cacheHitPrice: 0.2
   }
 };
 var HYBRID_THINKING_API_NAMES = /* @__PURE__ */ new Set([
@@ -5562,6 +5590,7 @@ var AgentFactory = class {
         return new MiniMaxAgent(name, instruction, model.modelApiName, key, model.temperature, shouldEnableThinking);
       // Meta Muse Spark — Responses API with encrypted reasoning replay, effort from the catalog
       case LLM_CONSTANTS.MUSE_SPARK:
+      case LLM_CONSTANTS.MUSE_WONDER:
         return new MetaAgent(name, instruction, model.modelApiName, key, model.temperature, shouldEnableThinking);
       default:
         throw new Error(`Unknown Key: ${modelName}`);