@cognipeer/agent-sdk 0.8.7 → 0.8.9

package/dist/index.d.cts

@@ -2481,6 +2481,8 @@ declare class StructuredOutputManager<T> {
     /**
      * Build a nudge message for when the model didn't produce structured output.
      * Used by the agent loop when it needs to retry.
+     * User role is intentional: OpenAI-compatible chat templates such as Qwen
+     * reject system messages anywhere except the beginning of the conversation.
      */
     buildNudgeMessage(isLastAttempt: boolean): {
         role: string;

package/dist/index.d.ts

@@ -2481,6 +2481,8 @@ declare class StructuredOutputManager<T> {
     /**
      * Build a nudge message for when the model didn't produce structured output.
      * Used by the agent loop when it needs to retry.
+     * User role is intentional: OpenAI-compatible chat templates such as Qwen
+     * reject system messages anywhere except the beginning of the conversation.
      */
     buildNudgeMessage(isLastAttempt: boolean): {
         role: string;

package/dist/index.mjs

@@ -802,6 +802,28 @@ function normalizeUsage(raw) {
   };
   return normalized;
 }
+function recordUsage(state, modelName, normalized) {
+  const usageState = state.usage || { perRequest: [], totals: {} };
+  const id = `${Date.now()}_${Math.random().toString(36).slice(2, 8)}`;
+  const cachedInput = normalized.prompt_tokens_details?.cached_tokens;
+  usageState.perRequest.push({
+    id,
+    modelName,
+    usage: normalized,
+    timestamp: (/* @__PURE__ */ new Date()).toISOString(),
+    turn: usageState.perRequest.length + 1,
+    cachedInput
+  });
+  const previous = usageState.totals[modelName] || { input: 0, output: 0, total: 0, cachedInput: 0 };
+  usageState.totals[modelName] = {
+    input: previous.input + (Number(normalized.prompt_tokens) || 0),
+    output: previous.output + (Number(normalized.completion_tokens) || 0),
+    total: previous.total + (Number(normalized.total_tokens) || 0),
+    cachedInput: previous.cachedInput + (Number(cachedInput) || 0)
+  };
+  state.usage = usageState;
+  return id;
+}
 function num(...vals) {
   for (const v of vals) {
     if (v === 0) return 0;
@@ -3171,24 +3193,7 @@ function createAgentCoreNode(opts) {
       messageList: messagesWithResponse
     });
     if (normalized) {
-      const usageState = state.usage || { perRequest: [], totals: {} };
-      const id = `${Date.now()}_${Math.random().toString(36).slice(2, 8)}`;
-      const turn = usageState.perRequest.length + 1;
-      const timestamp = (/* @__PURE__ */ new Date()).toISOString();
-      const cachedInputTok = normalized.prompt_tokens_details?.cached_tokens;
-      usageState.perRequest.push({ id, modelName, usage: normalized, timestamp, turn, cachedInput: cachedInputTok });
-      const inputTok = normalized.prompt_tokens;
-      const outputTok = normalized.completion_tokens;
-      const totalTok = normalized.total_tokens;
-      const key = modelName;
-      const agg = usageState.totals[key] || { input: 0, output: 0, total: 0, cachedInput: 0 };
-      usageState.totals[key] = {
-        input: agg.input + (Number(inputTok) || 0),
-        output: agg.output + (Number(outputTok) || 0),
-        total: agg.total + (Number(totalTok) || 0),
-        cachedInput: agg.cachedInput + (Number(cachedInputTok) || 0)
-      };
-      state.usage = usageState;
+      recordUsage(state, modelName, normalized);
     }
     return { messages: messagesWithResponse, usage: state.usage };
   };
@@ -4870,11 +4875,13 @@ var StructuredOutputManager = class {
   /**
    * Build a nudge message for when the model didn't produce structured output.
    * Used by the agent loop when it needs to retry.
+   * User role is intentional: OpenAI-compatible chat templates such as Qwen
+   * reject system messages anywhere except the beginning of the conversation.
    */
   buildNudgeMessage(isLastAttempt) {
     if (this.strategy.kind === "native") {
       return {
-        role: "system",
+        role: "user",
         content: [
           "A structured output schema is active.",
           "You MUST respond with valid JSON matching the schema.",
@@ -4884,7 +4891,7 @@ var StructuredOutputManager = class {
       };
     }
     return {
-      role: "system",
+      role: "user",
       content: [
         "A structured output schema is active.",
         "You MUST now call tool `response` with the final JSON object that matches the schema.",
@@ -4902,7 +4909,7 @@ var StructuredOutputManager = class {
       const fieldIssues = error.fieldErrors.map((f) => `- "${f.path}": ${f.message} (received: ${JSON.stringify(f.received)})`).join("\n");
       const action = this.strategy.kind === "native" ? "Please output corrected JSON matching the schema." : "Please call `response` again with corrected values matching the schema.";
       return {
-        role: "system",
+        role: "user",
         content: [
           "Your previous structured output had validation errors:",
           fieldIssues,
@@ -5881,24 +5888,7 @@ function createAgentNode(opts) {
       messageList: messagesWithResponse
     });
     if (normalized) {
-      const usageState = state.usage || { perRequest: [], totals: {} };
-      const id = `${Date.now()}_${Math.random().toString(36).slice(2, 8)}`;
-      const turn = usageState.perRequest.length + 1;
-      const timestamp = (/* @__PURE__ */ new Date()).toISOString();
-      const cachedInputTok = normalized.prompt_tokens_details.cached_tokens;
-      usageState.perRequest.push({ id, modelName, usage: normalized, timestamp, turn, cachedInput: cachedInputTok });
-      const inputTok = normalized.prompt_tokens;
-      const outputTok = normalized.completion_tokens;
-      const totalTok = normalized.total_tokens;
-      const key = modelName;
-      const agg = usageState.totals[key] || { input: 0, output: 0, total: 0, cachedInput: 0 };
-      usageState.totals[key] = {
-        input: agg.input + (Number(inputTok) || 0),
-        output: agg.output + (Number(outputTok) || 0),
-        total: agg.total + (Number(totalTok) || 0),
-        cachedInput: agg.cachedInput + (Number(cachedInputTok) || 0)
-      };
-      state.usage = usageState;
+      recordUsage(state, modelName, normalized);
     }
     return { messages: messagesWithResponse, usage: state.usage };
   };
@@ -6348,6 +6338,7 @@ ${canonicalFacts.length > 0 ? canonicalFacts.map((fact) => `- ${fact.key}: ${fac
         outputTokensActual = normalized.completion_tokens;
         totalTokensActual = normalized.total_tokens;
         cachedInputTokens = normalized.prompt_tokens_details?.cached_tokens;
+        recordUsage(state, getModelName(model) || "unknown_model", normalized);
       }
     } catch (err) {
       durationMs = Date.now() - startTime;