@hiper2d/ai-agents 0.11.0 → 0.13.0

package/dist/index.mjs

@@ -886,10 +886,16 @@ var SupportedAiModels = {
   // GPT-6 Astra (2026-09-03): OpenAI's frontier tier above Sol. No `none` reasoning effort;
   // temperature/top_p are rejected — Gpt5Agent sends neither, so the same agent serves it.
   // The catalog temperature is only carried for the agent constructor signature.
+  // All four OpenAI entries carry reasoningEffort: 'medium' as of 2026-09-22. That is
+  // OpenAI's own current default, established by MEASUREMENT rather than docs: each model was
+  // called with no reasoning param and the effort it echoes back read off the response, since
+  // OpenAI documents the default for Sol, Luna and the 5.6 family but never states Astra's.
+  // Pinning it changes no behaviour today; it stops the default moving under us silently.
   [LLM_CONSTANTS.GPT_ASTRA]: {
     displayName: "GPT-6 Astra",
     modelApiName: "gpt-6-astra",
     apiKeyName: API_KEY_CONSTANTS.OPENAI,
+    reasoningEffort: "medium",
     hasThinking: true,
     temperature: 1,
     tags: ["expensive"]
@@ -904,6 +910,7 @@ var SupportedAiModels = {
     displayName: "GPT-6 Sol",
     modelApiName: "gpt-6-sol",
     apiKeyName: API_KEY_CONSTANTS.OPENAI,
+    reasoningEffort: "medium",
     hasThinking: true,
     temperature: 1,
     tags: ["expensive"]
@@ -912,6 +919,7 @@ var SupportedAiModels = {
     displayName: "GPT-5.6 Terra",
     modelApiName: "gpt-5.6-terra",
     apiKeyName: API_KEY_CONSTANTS.OPENAI,
+    reasoningEffort: "medium",
     hasThinking: true,
     temperature: 1,
     tags: ["fast", "expensive"]
@@ -920,6 +928,7 @@ var SupportedAiModels = {
     displayName: "GPT-6 Luna",
     modelApiName: "gpt-6-luna",
     apiKeyName: API_KEY_CONSTANTS.OPENAI,
+    reasoningEffort: "medium",
     hasThinking: true,
     temperature: 1,
     tags: ["fast", "cheap"]
@@ -969,6 +978,7 @@ var SupportedAiModels = {
     modelApiName: "grok-4.7",
     apiKeyName: API_KEY_CONSTANTS.GROK,
     hasThinking: true,
+    reasoningEffort: "high",
     temperature: 0.7
   },
   // Mistral models. Two hybrid entries since 2026-09-18: reasoning is off by default on the
@@ -1017,6 +1027,9 @@ var SupportedAiModels = {
     hasThinking: true,
     // Temperature is omitted from the request: kimi-k3 rejects any value other than 1.
     // Speed samples: 17s (2026-08-04) and 28.9s (2026-08-05) — graded into the >25s tier.
+    // Pinned 'high' 2026-09-22, down from Moonshot's 'max' default: K3 had no levels at
+    // launch and the agent hardcoded max, so this is the first time the knob exists.
+    reasoningEffort: "high",
     tags: ["very-slow", "expensive"]
   },
   // Z.AI models — thinking-only entry (non-thinking variant retired 2026-08-05)
@@ -1531,6 +1544,7 @@ 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"];
+var KIMI_REASONING_EFFORTS = ["low", "high", "max"];
 function clampReasoningEffort(effort, allowed) {
   const rank = REASONING_EFFORT_SCALE.indexOf(effort);
   let best = allowed[0];
@@ -1553,6 +1567,7 @@ var toMetaEffort = (effort) => clampReasoningEffort(effort, META_REASONING_EFFOR
 var toFuguEffort = (effort) => clampReasoningEffort(effort, FUGU_REASONING_EFFORTS);
 var toMistralEffort = (effort) => clampReasoningEffort(effort, MISTRAL_REASONING_EFFORTS);
 var toXaiEffort = (effort) => clampReasoningEffort(effort, XAI_REASONING_EFFORTS);
+var toKimiEffort = (effort) => clampReasoningEffort(effort, KIMI_REASONING_EFFORTS);
 
 // src/pricing/token-usage-utils.ts
 function extractTokenUsage(response) {
@@ -2327,6 +2342,17 @@ ${instruction}`);
       apiKey
     });
   }
+  /**
+   * `reasoning: {effort}`, sent only when the catalog names a level. Every current OpenAI
+   * model defaults to "medium" when the field is omitted — verified 2026-09-22 by calling
+   * astra/sol/terra/luna with no reasoning param and reading the effort the API echoes back,
+   * since OpenAI documents the default for Sol, Luna and the 5.6 family but not for Astra.
+   * The catalog pins all four to medium so the level is explicit in our request rather than
+   * inherited from a default OpenAI can change under us.
+   */
+  reasoningParams() {
+    return this.reasoningEffort ? { reasoning: { effort: toOpenAIEffort(this.reasoningEffort) } } : {};
+  }
   /**
    * Structured output method using Zod with OpenAI's Responses API
    * This provides better schema handling and runtime validation
@@ -2347,6 +2373,7 @@ ${instruction}`);
           input,
           max_output_tokens: this.maxOutputTokens,
           prompt_cache_key: this.promptCacheKey,
+          ...this.reasoningParams(),
           text: {
             format: zodTextFormat(schemaToSend, "response_schema")
           }
@@ -2431,7 +2458,8 @@ ${instruction}`);
         instructions: this.instruction,
         input,
         max_output_tokens: this.maxOutputTokens,
-        prompt_cache_key: this.promptCacheKey
+        prompt_cache_key: this.promptCacheKey,
+        ...this.reasoningParams()
       });
       const content = response.output_text;
       if (!content) {
@@ -3927,8 +3955,8 @@ var KimiAgent = class extends AbstractAgent {
       model: this.model,
       stream: false,
       max_tokens: this.maxOutputTokens,
-      // Moonshot's only accepted level; "max" is not in the OpenAI SDK's ReasoningEffort union.
-      reasoning_effort: "max"
+      // "max" is not in the OpenAI SDK's ReasoningEffort union, hence the cast.
+      reasoning_effort: toKimiEffort(this.reasoningEffort ?? "max")
     };
   }
   // Log message templates
@@ -5204,6 +5232,7 @@ export {
   Gpt5Agent,
   GrokAgent,
   InMemorySpendStore,
+  KIMI_REASONING_EFFORTS,
   KimiAgent,
   LLM_CONSTANTS,
   MESSAGE_ROLE,
@@ -5310,6 +5339,7 @@ export {
   toFuguEffort,
   toGeminiEffort,
   toGlmEffort,
+  toKimiEffort,
   toMetaEffort,
   toMistralEffort,
   toOpenAIEffort,