@hiper2d/ai-agents 0.9.0 → 0.11.0

package/dist/index.mjs

@@ -958,10 +958,15 @@ var SupportedAiModels = {
     reasoningEffort: "minimal",
     tags: ["fast", "cheap"]
   },
-  // Always-on reasoning (xAI default effort "high", cannot be disabled) — no non-thinking sibling
+  // Always-on reasoning — no non-thinking sibling. Bumped 4.6 → 4.7 on 2026-09-22 (id verified
+  // against GET /v1/models). Same price, same 200K extended-context boundary; the context
+  // window is 500K. NOTE 4.7 newly accepts `reasoning_effort` (low | medium | high | xhigh,
+  // default high) where 4.6 had no knob at all. GrokAgent still sends none, so behaviour is
+  // unchanged — but Grok is the app's most expensive model per turn (94% of its output is
+  // reasoning), so a lower pin is the obvious lever to measure before anything else.
   [LLM_CONSTANTS.GROK]: {
-    displayName: "Grok 4.6",
-    modelApiName: "grok-4.6",
+    displayName: "Grok 4.7",
+    modelApiName: "grok-4.7",
     apiKeyName: API_KEY_CONSTANTS.GROK,
     hasThinking: true,
     temperature: 0.7
@@ -1373,8 +1378,9 @@ var MODEL_PRICING = {
     cacheHitPrice: 0.015
   },
   // Grok models. Cached price is per-model on xAI (not a uniform ratio):
-  // grok-4.6 is $0.50/M cached vs $2.00/M input, and all rates double for prompts
-  // >= 200K tokens, per docs.x.ai/developers/models (verified 2026-08-12).
+  // grok-4.7 is $0.50/M cached vs $2.00/M input, and all rates double for prompts
+  // >= 200K tokens, per docs.x.ai/docs/models (re-verified 2026-09-22 for 4.7).
+  // 4.7 carries over 4.6's rates unchanged — the bump is capability-only, not a reprice.
   [SupportedAiModels[LLM_CONSTANTS.GROK].modelApiName]: {
     inputPrice: 2,
     outputPrice: 6,
@@ -1524,6 +1530,7 @@ var DEEPSEEK_REASONING_EFFORTS = ["low", "high", "max"];
 var FUGU_REASONING_EFFORTS = ["high", "xhigh", "max"];
 var META_REASONING_EFFORTS = ["minimal", "low", "medium", "high", "xhigh", "max"];
 var MISTRAL_REASONING_EFFORTS = ["minimal", "low", "medium", "high", "xhigh"];
+var XAI_REASONING_EFFORTS = ["low", "medium", "high", "xhigh"];
 function clampReasoningEffort(effort, allowed) {
   const rank = REASONING_EFFORT_SCALE.indexOf(effort);
   let best = allowed[0];
@@ -1545,6 +1552,7 @@ var toDeepSeekEffort = (effort) => clampReasoningEffort(effort, DEEPSEEK_REASONI
 var toMetaEffort = (effort) => clampReasoningEffort(effort, META_REASONING_EFFORTS);
 var toFuguEffort = (effort) => clampReasoningEffort(effort, FUGU_REASONING_EFFORTS);
 var toMistralEffort = (effort) => clampReasoningEffort(effort, MISTRAL_REASONING_EFFORTS);
+var toXaiEffort = (effort) => clampReasoningEffort(effort, XAI_REASONING_EFFORTS);
 
 // src/pricing/token-usage-utils.ts
 function extractTokenUsage(response) {
@@ -3782,6 +3790,16 @@ ${schemaDescription}`;
       throw new Error(this.errorMessages.apiError(error));
     }
   }
+  /**
+   * A stable key per (game, bot) so xAI routes this conversation's requests to the same
+   * server. Without it a request often lands on a cache-cold machine and the whole prompt
+   * bills at full input price — xAI calls this out as the most common caching mistake.
+   * Falls back to the bot name alone outside a game (story generation, tests), which is
+   * still stabler than sending nothing.
+   */
+  get promptCacheKey() {
+    return this.gameId ? `${this.gameId}:${this.name}` : this.name;
+  }
   createResponse(input, jsonMode) {
     return this.client.responses.create({
       model: this.model,
@@ -3792,9 +3810,16 @@ ${schemaDescription}`;
       // ever starts truncating — measured turns peak far below the shared default.
       max_output_tokens: this.maxOutputTokens,
       // We manage conversation state ourselves; encrypted reasoning is only
-      // returned for unstored responses.
+      // returned for unstored responses. Grok 4.7 returns reasoning.encrypted_content
+      // on the Responses API whether or not `include` asks for it; we keep asking so
+      // the behaviour is identical if the model is ever pointed back at 4.6.
       store: false,
       include: ["reasoning.encrypted_content"],
+      prompt_cache_key: this.promptCacheKey,
+      // Grok 4.7 is the first Grok with an effort knob (4.6 had none) and defaults to
+      // "high". Only send it when the catalog asks for a level, so an unset entry keeps
+      // the provider default rather than us silently picking one.
+      ...this.reasoningEffort ? { reasoning: { effort: toXaiEffort(this.reasoningEffort) } } : {},
       ...jsonMode ? { text: { format: { type: "json_object" } } } : {}
     });
   }
@@ -5207,6 +5232,7 @@ export {
   VOICE_MODEL_PRICING,
   VOICE_PROVIDER_API_KEY,
   VoiceAgentFactory,
+  XAI_REASONING_EFFORTS,
   ZodSchemaConverter,
   applySpend,
   buildGoogleTtsPrompt,
@@ -5287,6 +5313,7 @@ export {
   toMetaEffort,
   toMistralEffort,
   toOpenAIEffort,
+  toXaiEffort,
   transcribeWithGemini,
   transcribeWithOpenAi,
   validateResponse