@hiper2d/ai-agents 0.1.3 → 0.1.5

package/dist/index.mjs

@@ -787,6 +787,7 @@ var LLM_CONSTANTS = {
   DEEPSEEK_PRO: "deepseek-pro",
   // GPT-5.6 family. 'gpt' and 'gpt-mini' are stable picker ids carried over from the
   // GPT-5.5 / GPT-5.4-mini era so existing consumers keep working across the repoint.
+  GPT_ASTRA: "gpt-astra",
   GPT_SOL: "gpt-sol",
   GPT: "gpt",
   GPT_MINI: "gpt-mini",
@@ -871,6 +872,17 @@ var SupportedAiModels = {
     tags: ["cheap"]
   },
   // Models with always-on reasoning
+  // 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.
+  [LLM_CONSTANTS.GPT_ASTRA]: {
+    displayName: "GPT-6 Astra",
+    modelApiName: "gpt-6-astra",
+    apiKeyName: API_KEY_CONSTANTS.OPENAI,
+    hasThinking: true,
+    temperature: 1,
+    tags: ["expensive"]
+  },
   // GPT-5.6 family (promoted July 2026 when the limited preview opened up):
   // sol is the flagship, terra the mainline, luna the cheap tier.
   [LLM_CONSTANTS.GPT_SOL]: {
@@ -1142,6 +1154,19 @@ var DEEPSEEK_PEAK_SCHEDULE = {
   weekendOffPeak: { utcOffsetHours: 8 }
 };
 var MODEL_PRICING = {
+  // OpenAI GPT-6 Astra (developers.openai.com/api/docs/pricing, 2026-09-03): $10/$50 cache-hit $1
+  // short context, $20/$75 cache-hit $2 long context. OpenAI's pricing table doesn't restate
+  // the boundary; we assume the same 272k threshold as the GPT-5.6 siblings. Cache writes
+  // ($12.50/$25) are not modelled — caching is automatic and we only see hits.
+  [SupportedAiModels[LLM_CONSTANTS.GPT_ASTRA].modelApiName]: {
+    inputPrice: 10,
+    outputPrice: 50,
+    cacheHitPrice: 1,
+    extendedContextInputPrice: 20,
+    extendedContextOutputPrice: 75,
+    extendedContextCacheHitPrice: 2,
+    extendedContextThresholdTokens: 272e3
+  },
   // OpenAI GPT-5.6 models
   // Sol repriced 2026-08-30 (developers.openai.com/api/docs/pricing): $4/$20 short context,
   // $8/$30 past the long-context threshold — the same 272k boundary its siblings use.
@@ -1740,7 +1765,6 @@ ${msg.content}` };
 
 // src/agents/gpt-5-agent.ts
 import OpenAI from "openai";
-import { z as z2 } from "zod";
 import { zodTextFormat } from "openai/helpers/zod";
 var Gpt5Agent = class extends AbstractAgent {
   client;
@@ -1774,12 +1798,7 @@ var Gpt5Agent = class extends AbstractAgent {
         `System: ${this.instruction}`,
         ...this.prepareMessages(messages).map((msg) => `${msg.role === "user" ? "User" : "Assistant"}: ${msg.content}`)
       ].join("\n\n");
-      let schemaToSend = zodSchema;
-      if (this.enableThinking && zodSchema instanceof z2.ZodObject) {
-        schemaToSend = zodSchema.extend({
-          thinking: z2.string().describe("Your internal chain-of-thought reasoning process used to arrive at the final answer.")
-        });
-      }
+      const schemaToSend = zodSchema;
       let response;
       try {
         response = await this.client.responses.parse({
@@ -4258,6 +4277,7 @@ var AgentFactory = class {
       case LLM_CONSTANTS.CLAUDE_HAIKU:
         return new ClaudeAgent(name, instruction, model.modelApiName, key, shouldEnableThinking);
       // Always-on reasoning models
+      case LLM_CONSTANTS.GPT_ASTRA:
       case LLM_CONSTANTS.GPT_SOL:
       case LLM_CONSTANTS.GPT:
       case LLM_CONSTANTS.GPT_MINI: