@hiper2d/ai-agents 0.1.2 → 0.1.4

package/dist/index.mjs

@@ -705,6 +705,14 @@ var ModelQuotaExceededError = class extends ModelError {
     this.name = "ModelQuotaExceededError";
   }
 };
+var ModelInvalidResponseError = class extends ModelError {
+  truncated;
+  constructor(modelType, detail, truncated = false) {
+    super(`${modelType} failed to produce a valid response: ${detail}`, modelType);
+    this.name = "ModelInvalidResponseError";
+    this.truncated = truncated;
+  }
+};
 var ModelRefusalError = class extends ModelError {
   constructor(modelType, message = `${modelType} refused to answer (stop_reason: refusal)`) {
     super(message, modelType);
@@ -779,6 +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",
@@ -863,6 +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]: {
@@ -1134,6 +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.
@@ -1772,18 +1805,37 @@ var Gpt5Agent = class extends AbstractAgent {
           thinking: z2.string().describe("Your internal chain-of-thought reasoning process used to arrive at the final answer.")
         });
       }
-      const response = await this.client.responses.parse({
-        model: this.model,
-        instructions: this.instruction,
-        input,
-        max_output_tokens: this.maxOutputTokens,
-        text: {
-          format: zodTextFormat(schemaToSend, "response_schema")
+      let response;
+      try {
+        response = await this.client.responses.parse({
+          model: this.model,
+          instructions: this.instruction,
+          input,
+          max_output_tokens: this.maxOutputTokens,
+          text: {
+            format: zodTextFormat(schemaToSend, "response_schema")
+          }
+        });
+      } catch (error) {
+        if (error instanceof SyntaxError) {
+          throw new ModelInvalidResponseError(
+            this.model,
+            `malformed JSON output \u2014 the generation was cut off at the ${this.maxOutputTokens}-token output cap or went off the rails (${error.message})`
+          );
         }
-      });
+        throw error;
+      }
+      if (response.status === "incomplete") {
+        const reason = response.incomplete_details?.reason ?? "unknown";
+        throw new ModelInvalidResponseError(
+          this.model,
+          `response incomplete (${reason}) at max_output_tokens=${this.maxOutputTokens}`,
+          reason === "max_output_tokens"
+        );
+      }
       if (!response.output_parsed) {
         this.logger(`Parsing failed. Raw content: ${response.output_text}`);
-        throw new Error(this.errorMessages.invalidFormat);
+        throw new ModelInvalidResponseError(this.model, this.errorMessages.invalidFormat);
       }
       let reasoningContent = "";
       if (this.enableThinking && response.output_parsed.thinking) {
@@ -1822,6 +1874,9 @@ var Gpt5Agent = class extends AbstractAgent {
       return [response.output_parsed, reasoningContent, tokenUsage];
     } catch (error) {
       this.logger(this.logTemplates.error(this.name, error));
+      if (error instanceof ModelError) {
+        throw error;
+      }
       throw new Error(this.errorMessages.apiError(error));
     }
   }
@@ -1847,6 +1902,14 @@ var Gpt5Agent = class extends AbstractAgent {
       });
       const content = response.output_text;
       if (!content) {
+        if (response.status === "incomplete") {
+          const reason = response.incomplete_details?.reason ?? "unknown";
+          throw new ModelInvalidResponseError(
+            this.model,
+            `empty response, incomplete (${reason}) at max_output_tokens=${this.maxOutputTokens}`,
+            reason === "max_output_tokens"
+          );
+        }
         throw new Error(this.errorMessages.emptyResponse);
       }
       let tokenUsage;
@@ -1879,6 +1942,9 @@ var Gpt5Agent = class extends AbstractAgent {
       return [content, "", tokenUsage];
     } catch (error) {
       this.logger(this.logTemplates.error(this.name, error));
+      if (error instanceof ModelError) {
+        throw error;
+      }
       throw new Error(this.errorMessages.apiError(error));
     }
   }
@@ -4217,6 +4283,7 @@ 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:
@@ -4294,6 +4361,7 @@ export {
   MistralAgent,
   ModelAuthenticationError,
   ModelError,
+  ModelInvalidResponseError,
   ModelOverloadError,
   ModelQuotaExceededError,
   ModelRateLimitError,