Merge pull request #1068 from buster-so/wells-bus-1881-running-into-prompt-issues-prompt-too-long

Wells bus 1881 running into prompt issues prompt too long

packages/ai/src/steps/analyst-agent-steps/extract-values-step/extract-values-with-llm.ts

@@ -2,8 +2,8 @@ import type { ModelMessage } from 'ai';
 import { generateObject } from 'ai';
 import { wrapTraced } from 'braintrust';
 import { z } from 'zod';
-import { Haiku35 } from '../../../llm/haiku-3-5';
-import { DEFAULT_ANTHROPIC_OPTIONS } from '../../../llm/providers/gateway';
+import { GPT5Mini } from '../../../llm';
+import { DEFAULT_OPENAI_OPTIONS } from '../../../llm/providers/gateway';
 
 // Schema for what the LLM returns
 const llmOutputSchema = z.object({
@@ -91,11 +91,11 @@ export async function extractValuesWithLLM(
     const tracedValuesExtraction = wrapTraced(
       async () => {
         const { object } = await generateObject({
-          model: Haiku35,
+          model: GPT5Mini,
           schema: llmOutputSchema,
           messages,
           temperature: 0,
-          providerOptions: DEFAULT_ANTHROPIC_OPTIONS,
+          providerOptions: DEFAULT_OPENAI_OPTIONS,
         });
 
         return object;

packages/ai/src/steps/analyst-agent-steps/generate-chat-title-step/generate-chat-title-step.ts

@@ -3,8 +3,8 @@ import { generateObject } from 'ai';
 import type { ModelMessage } from 'ai';
 import { wrapTraced } from 'braintrust';
 import { z } from 'zod';
-import { Haiku35 } from '../../../llm';
-import { DEFAULT_ANTHROPIC_OPTIONS } from '../../../llm/providers/gateway';
+import { GPT5Nano } from '../../../llm';
+import { DEFAULT_OPENAI_OPTIONS } from '../../../llm/providers/gateway';
 
 // Zod-first: define input/output schemas and export inferred types
 export const generateChatTitleParamsSchema = z.object({
@@ -56,14 +56,10 @@ async function generateTitleWithLLM(messages: ModelMessage[]): Promise<string> {
     const tracedChatTitle = wrapTraced(
       async () => {
         const { object } = await generateObject({
-          model: Haiku35,
-          headers: {
-            'anthropic-beta': 'fine-grained-tool-streaming-2025-05-14',
-            anthropic_beta: 'fine-grained-tool-streaming-2025-05-14',
-          },
+          model: GPT5Nano,
           schema: llmOutputSchema,
           messages: titleMessages,
-          providerOptions: DEFAULT_ANTHROPIC_OPTIONS,
+          providerOptions: DEFAULT_OPENAI_OPTIONS,
         });
 
         return object;

packages/ai/src/utils/index.ts

@@ -28,3 +28,5 @@ export {
 // Step retry utilities
 export { withStepRetry, createRetryableStep, runStepsWithRetry } from './with-step-retry';
 export type { StepRetryOptions } from './with-step-retry';
+
+export { extractUserAndDoneToolMessages } from './memory/extract-user-and-done-tool-messages';

packages/ai/src/utils/memory/extract-user-and-done-tool-messages.ts

@@ -0,0 +1,26 @@
+import type { ModelMessage } from 'ai';
+
+export function extractUserAndDoneToolMessages(messages: ModelMessage[]): ModelMessage[] {
+  return messages.filter((m) => {
+    // Include user messages and done tool calls and results
+    if (m.role === 'user' && typeof m.content === 'string') {
+      return true;
+    }
+
+    // Include doneTool calls
+    if (m.role === 'assistant' && Array.isArray(m.content)) {
+      return m.content.some(
+        (contentItem) => contentItem.type === 'tool-call' && contentItem.toolName === 'doneTool'
+      );
+    }
+
+    // Include doneTool results
+    if (m.role === 'tool' && Array.isArray(m.content)) {
+      return m.content.some(
+        (contentItem) => contentItem.type === 'tool-result' && contentItem.toolName === 'doneTool'
+      );
+    }
+
+    return false;
+  });
+}

packages/ai/src/workflows/analyst-agent-workflow/analyst-workflow.ts

@@ -25,6 +25,7 @@ import { CREATE_METRICS_TOOL_NAME } from '../../tools/visualization-tools/metric
 import { MODIFY_METRICS_TOOL_NAME } from '../../tools/visualization-tools/metrics/modify-metrics-tool/modify-metrics-tool';
 import { CREATE_REPORTS_TOOL_NAME } from '../../tools/visualization-tools/reports/create-reports-tool/create-reports-tool';
 import { MODIFY_REPORTS_TOOL_NAME } from '../../tools/visualization-tools/reports/modify-reports-tool/modify-reports-tool';
+import { extractUserAndDoneToolMessages } from '../../utils';
 import { withStepRetry } from '../../utils/with-step-retry';
 import type { StepRetryOptions } from '../../utils/with-step-retry';
 import {
@@ -268,6 +269,7 @@ async function runAnalystPrepSteps({
   values: ExtractValuesSearchResult;
   analysisMode: AnalysisTypeRouterResult['analysisMode'];
 }> {
+  const filteredToUserAndDoneToolMessages = extractUserAndDoneToolMessages(messages);
   const shouldInjectUserPersonalizationTodo = Boolean(userPersonalizationConfig);
   const [todos, values, , analysisMode] = await Promise.all([
     withStepRetry(
@@ -292,7 +294,7 @@ async function runAnalystPrepSteps({
     withStepRetry(
       () =>
         runExtractValuesAndSearchStep({
-          messages,
+          messages: filteredToUserAndDoneToolMessages,
           dataSourceId,
         }),
       {
@@ -310,7 +312,7 @@ async function runAnalystPrepSteps({
     withStepRetry(
       () =>
         runGenerateChatTitleStep({
-          messages,
+          messages: filteredToUserAndDoneToolMessages,
           chatId,
           messageId,
         }),
@@ -329,7 +331,7 @@ async function runAnalystPrepSteps({
     withStepRetry(
       () =>
         runAnalysisTypeRouterStep({
-          messages,
+          messages: filteredToUserAndDoneToolMessages,
           messageAnalysisMode,
         }),
       {