@hiper2d/ai-agents 0.22.1 → 0.24.0

package/dist/index.mjs

@@ -803,6 +803,8 @@ var LLM_CONSTANTS = {
   GPT_SOL: "gpt-sol",
   GPT: "gpt",
   GPT_MINI: "gpt-mini",
+  // Gemini 4 Argon sits above Pro; version-free id so a later Argon repoint keeps persisted ids.
+  GEMINI_ARGON: "gemini-argon",
   GEMINI_PRO: "gemini-pro",
   GEMINI_FLASH: "gemini-flash",
   GEMINI_LITE: "gemini-lite",
@@ -819,7 +821,9 @@ 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;
@@ -960,6 +964,21 @@ var SupportedAiModels = {
   // 2.5-era thinkingBudget: 1024 (2026-08-06), which HAD been binding — so Flash Lite now
   // thinks noticeably less under its "minimal" default (0.8s/49-token votes vs 4.5s/650
   // budgeted); bump it to 'low' if its output quality visibly drops.
+  // Gemini 4 Argon (added 2026-10-13, id verified against GET /v1beta/models): Google's
+  // flagship reasoning model, 1M in / 32K out. Same thinkingLevel dialect as 3.x — accepts
+  // low|medium|high, rejects 'minimal' (probed). Pinned 'high' like Pro: on a 1.7K-token game
+  // prompt 'medium' thought 0 tokens (3.2s) while 'high' thought 622 (8.1s); the level is a
+  // 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.
+  [LLM_CONSTANTS.GEMINI_ARGON]: {
+    displayName: "Gemini 4 Argon",
+    modelApiName: "gemini-4-argon",
+    apiKeyName: API_KEY_CONSTANTS.GOOGLE,
+    hasThinking: true,
+    reasoningEffort: "high",
+    tags: ["expensive"]
+  },
   [LLM_CONSTANTS.GEMINI_PRO]: {
     displayName: "Gemini 3.1 Pro Preview",
     modelApiName: "gemini-3.1-pro-preview",
@@ -1210,6 +1229,20 @@ 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 = {}) {
@@ -1395,6 +1428,15 @@ var MODEL_PRICING = {
     extendedContextThresholdTokens: 1e5
   },
   // Google models
+  [SupportedAiModels[LLM_CONSTANTS.GEMINI_ARGON].modelApiName]: {
+    // Launch pricing (ai.google.dev pricing page, fetched 2026-10-13), one flat tier — no
+    // >200K long-context rows, unlike 3.1 Pro. DOUBLES on 2026-12-01 to $4.00 / $20.00 /
+    // $0.20 — ACTION NEEDED then. Implicit caching only (explicit caching unsupported),
+    // cached input bills at 5% of the input price. Not on Google's free tier or AI Pro plan.
+    inputPrice: 2,
+    outputPrice: 10,
+    cacheHitPrice: 0.1
+  },
   [SupportedAiModels[LLM_CONSTANTS.GEMINI_PRO].modelApiName]: {
     inputPrice: 2,
     outputPrice: 12,
@@ -1489,6 +1531,13 @@ 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([
@@ -5326,6 +5375,7 @@ var AgentFactory = class {
       case LLM_CONSTANTS.GPT:
       case LLM_CONSTANTS.GPT_MINI:
         return new Gpt5Agent(name, instruction, model.modelApiName, key, model.temperature, shouldEnableThinking);
+      case LLM_CONSTANTS.GEMINI_ARGON:
       case LLM_CONSTANTS.GEMINI_PRO:
       case LLM_CONSTANTS.GEMINI_FLASH:
       case LLM_CONSTANTS.GEMINI_LITE:
@@ -5357,6 +5407,7 @@ 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}`);