@hiper2d/ai-agents 0.1.1 → 0.1.3

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);
@@ -905,10 +913,11 @@ var SupportedAiModels = {
     tags: ["expensive"]
   },
   [LLM_CONSTANTS.GEMINI_FLASH]: {
-    // Repointed from gemini-3.6-flash 2026-08-13 (stable picker id, same pattern as gpt).
-    // 3.7 rejects thinkingLevel 'minimal' (low|medium|high only), unlike 3.5/3.6.
-    displayName: "Gemini 3.7 Flash",
-    modelApiName: "gemini-3.7-flash",
+    // Repointed 3.6 → 3.7 (2026-08-13) → 3.8 (2026-09-02); stable picker id, same pattern as gpt.
+    // 3.7 rejected thinkingLevel 'minimal' (low|medium|high only), unlike 3.5/3.6 — 3.8 untested
+    // on 'minimal', so keep the pin at 'medium' or above.
+    displayName: "Gemini 3.8 Flash",
+    modelApiName: "gemini-3.8-flash",
     apiKeyName: API_KEY_CONSTANTS.GOOGLE,
     hasThinking: true,
     reasoningEffort: "medium",
@@ -1236,9 +1245,9 @@ var MODEL_PRICING = {
     extendedContextThresholdTokens: 2e5
   },
   [SupportedAiModels[LLM_CONSTANTS.GEMINI_FLASH].modelApiName]: {
-    // Launch pricing through 2026-12-31; doubles to $1.50/$7.50/$0.15 on 2027-01-01
-    // (ai.google.dev pricing page, fetched 2026-08-13) — ACTION NEEDED then: update these
-    // rates.
+    // 3.8 Flash (2026-09-02) launched at the same rates as 3.7. 3.7's launch pricing was
+    // scheduled to double to $1.50/$7.50/$0.15 on 2027-01-01 (ai.google.dev pricing page,
+    // fetched 2026-08-13) — ACTION NEEDED then: re-check whether 3.8 follows and update.
     // Cache storage cost ($0.50 / 1M tokens per hour) is not tracked here — the
     // schema only models per-token call costs, not time-based storage.
     inputPrice: 0.75,
@@ -1771,18 +1780,37 @@ 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: ${response.output_text}`);
-        throw new Error(this.errorMessages.invalidFormat);
+        throw new ModelInvalidResponseError(this.model, this.errorMessages.invalidFormat);
       }
       let reasoningContent = "";
       if (this.enableThinking && response.output_parsed.thinking) {
@@ -1821,6 +1849,9 @@ var Gpt5Agent = class extends AbstractAgent {
       return [response.output_parsed, reasoningContent, tokenUsage];
     } catch (error) {
       this.logger(this.logTemplates.error(this.name, error));
+      if (error instanceof ModelError) {
+        throw error;
+      }
       throw new Error(this.errorMessages.apiError(error));
     }
   }
@@ -1846,6 +1877,14 @@ var Gpt5Agent = class extends AbstractAgent {
       });
       const content = response.output_text;
       if (!content) {
+        if (response.status === "incomplete") {
+          const reason = response.incomplete_details?.reason ?? "unknown";
+          throw new ModelInvalidResponseError(
+            this.model,
+            `empty response, incomplete (${reason}) at max_output_tokens=${this.maxOutputTokens}`,
+            reason === "max_output_tokens"
+          );
+        }
         throw new Error(this.errorMessages.emptyResponse);
       }
       let tokenUsage;
@@ -1878,6 +1917,9 @@ 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));
     }
   }
@@ -4293,6 +4335,7 @@ export {
   MistralAgent,
   ModelAuthenticationError,
   ModelError,
+  ModelInvalidResponseError,
   ModelOverloadError,
   ModelQuotaExceededError,
   ModelRateLimitError,