@modelprofile.com/flexharness 8.2.0 → 8.4.0

package/ts_compaction/smartagent.compaction.ts

@@ -24,6 +24,13 @@ When constructing the summary, try to stick to this template:
 
 export interface ICompactMessagesOptions {
   abortSignal?: AbortSignal;
+  /**
+   * Receives the usage of every model call the compaction makes, once per call: `reported` with the
+   * provider's usage, or `unreported` for a call that failed or was aborted, retried attempts
+   * included. A context compactor passes its own options through, so the calls reach the session's
+   * `onUsage`. Must not throw.
+   */
+  reportUsage?: plugins.TAgentModelCallUsageReporter;
 }
 
 /**
@@ -32,7 +39,7 @@ export interface ICompactMessagesOptions {
  *
  * @param model    The same model used by runAgent, or a cheaper small model
  * @param messages The full message history that overflowed
- * @param options  Invocation options, including cancellation
+ * @param options  Invocation options, including cancellation and usage reporting
  * @returns A minimal ModelMessage[] containing the summary as context
  */
 export async function compactMessages(
@@ -40,20 +47,60 @@ export async function compactMessages(
   messages: plugins.ModelMessage[],
   options: ICompactMessagesOptions = {},
 ): Promise<plugins.ModelMessage[]> {
-  const result = await plugins.generateText({
+  const prompt: plugins.ModelMessage[] = [
+    ...messages,
+    { role: 'user', content: COMPACTION_PROMPT },
+  ];
+  const reportUsage = options.reportUsage;
+  if (!reportUsage) {
+    const result = await plugins.generateText({
+      model,
+      abortSignal: options.abortSignal,
+      allowSystemInMessages: true,
+      messages: prompt,
+    });
+    return summaryMessages(result.text);
+  }
+  const modelCalls = new plugins.AgentModelCallUsageRecorder(model, reportUsage);
+  // Each attempt, retries included, is one model call: it starts in the middleware, which also
+  // settles an attempt that throws; a call whose response arrives ends in onLanguageModelCallEnd.
+  const observedModel = plugins.wrapLanguageModel({
     model,
-    abortSignal: options.abortSignal,
-    allowSystemInMessages: true,
-    messages: [
-      ...messages,
-      { role: 'user', content: COMPACTION_PROMPT },
-    ],
+    middleware: {
+      wrapGenerate: async ({ doGenerate, params }) => {
+        modelCalls.start();
+        try {
+          return await doGenerate();
+        } catch (error) {
+          modelCalls.settle(params.abortSignal?.aborted ? 'aborted' : 'failed');
+          throw error;
+        }
+      },
+    },
   });
+  try {
+    const result = await plugins.generateText({
+      model: observedModel,
+      abortSignal: options.abortSignal,
+      allowSystemInMessages: true,
+      messages: prompt,
+      onLanguageModelCallEnd: ({ modelId, usage }) => {
+        modelCalls.end(modelId, usage);
+      },
+    });
+    modelCalls.settle('missing');
+    return summaryMessages(result.text);
+  } catch (error) {
+    modelCalls.settle(options.abortSignal?.aborted ? 'aborted' : 'failed');
+    throw error;
+  }
+}
 
+function summaryMessages(summary: string): plugins.ModelMessage[] {
   return [
     {
       role: 'user',
-      content: `[Previous conversation summary]\n\n${result.text}\n\n[End of summary. Continue from here.]`,
+      content: `[Previous conversation summary]\n\n${summary}\n\n[End of summary. Continue from here.]`,
     },
   ];
 }