@hiper2d/ai-agents 0.22.1 → 0.23.0

package/dist/index.d.mts

@@ -372,6 +372,7 @@ declare const LLM_CONSTANTS: {
     GPT_SOL: string;
     GPT: string;
     GPT_MINI: string;
+    GEMINI_ARGON: string;
     GEMINI_PRO: string;
     GEMINI_FLASH: string;
     GEMINI_LITE: string;

package/dist/index.d.ts

@@ -372,6 +372,7 @@ declare const LLM_CONSTANTS: {
     GPT_SOL: string;
     GPT: string;
     GPT_MINI: string;
+    GEMINI_ARGON: string;
     GEMINI_PRO: string;
     GEMINI_FLASH: string;
     GEMINI_LITE: string;

package/dist/index.js

@@ -981,6 +981,8 @@ 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",
@@ -1138,6 +1140,21 @@ 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",
@@ -1573,6 +1590,15 @@ 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,
@@ -5504,6 +5530,7 @@ 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: