@librechat/agents 3.4.6 → 3.5.0

package/src/scripts/bedrock-cache-debug.ts

@@ -1,250 +0,0 @@
-/**
- * Debug script to investigate cache token omission in Bedrock responses.
- *
- * This script:
- * 1. Makes a streaming call to Bedrock and logs the raw metadata event
- * 2. Shows exactly what fields the AWS SDK returns in usage (including cache tokens)
- * 3. Shows what our handleConverseStreamMetadata produces vs what it should produce
- * 4. Makes a multi-turn call to trigger caching and verify cache tokens appear
- */
-import { config } from 'dotenv';
-config();
-import { concat } from '@langchain/core/utils/stream';
-import { HumanMessage } from '@langchain/core/messages';
-import {
-  BedrockRuntimeClient,
-  ConverseStreamCommand,
-} from '@aws-sdk/client-bedrock-runtime';
-import type { AIMessageChunk } from '@langchain/core/messages';
-import { CustomChatBedrockConverse } from '@/llm/bedrock';
-
-const region = process.env.BEDROCK_AWS_REGION ?? 'us-east-1';
-const credentials = {
-  accessKeyId: process.env.BEDROCK_AWS_ACCESS_KEY_ID!,
-  secretAccessKey: process.env.BEDROCK_AWS_SECRET_ACCESS_KEY!,
-};
-
-const MODEL_ID = 'us.anthropic.claude-sonnet-4-5-20250929-v1:0';
-
-// A long system prompt to increase likelihood of cache usage
-// Bedrock requires minimum 1024 tokens for prompt caching to activate
-const SYSTEM_PROMPT = `You are an expert assistant. Here is a large context block to help trigger cache behavior:
-
-${Array(200).fill('This is padding content to make the prompt large enough to trigger Bedrock prompt caching. The minimum requirement for Anthropic models on Bedrock is 1024 tokens in the cached prefix. We need to ensure this prompt is well above that threshold. ').join('')}
-
-When answering, be brief and direct.`;
-
-async function rawSdkCall(): Promise<void> {
-  console.log('='.repeat(60));
-  console.log('TEST 1: Raw AWS SDK call - inspect metadata.usage directly');
-  console.log('='.repeat(60));
-
-  const client = new BedrockRuntimeClient({ region, credentials });
-
-  // First call - should create cache
-  // Use cachePoint block to explicitly enable prompt caching
-  console.log('\n--- Call 1 (cache write expected) ---');
-  const command1 = new ConverseStreamCommand({
-    modelId: MODEL_ID,
-    system: [{ text: SYSTEM_PROMPT }, { cachePoint: { type: 'default' } }],
-    messages: [{ role: 'user', content: [{ text: 'What is 2+2?' }] }],
-    inferenceConfig: { maxTokens: 100 },
-  });
-
-  const response1 = await client.send(command1);
-  if (response1.stream) {
-    for await (const event of response1.stream) {
-      if (event.metadata != null) {
-        console.log('\nRAW metadata event (Call 1):');
-        console.dir(event.metadata, { depth: null });
-        console.log('\nRAW metadata.usage:');
-        console.dir(event.metadata.usage, { depth: null });
-        console.log('\nSpecific cache fields:');
-        console.log(
-          '  cacheReadInputTokens:',
-          (event.metadata.usage as unknown as Record<string, unknown>)
-            ?.cacheReadInputTokens
-        );
-        console.log(
-          '  cacheWriteInputTokens:',
-          (event.metadata.usage as unknown as Record<string, unknown>)
-            ?.cacheWriteInputTokens
-        );
-      }
-    }
-  }
-
-  // Second call - should read from cache
-  console.log('\n--- Call 2 (cache read expected) ---');
-  const command2 = new ConverseStreamCommand({
-    modelId: MODEL_ID,
-    system: [{ text: SYSTEM_PROMPT }, { cachePoint: { type: 'default' } }],
-    messages: [
-      { role: 'user', content: [{ text: 'What is 2+2?' }] },
-      { role: 'assistant', content: [{ text: '4' }] },
-      { role: 'user', content: [{ text: 'And what is 3+3?' }] },
-    ],
-    inferenceConfig: { maxTokens: 100 },
-  });
-
-  const response2 = await client.send(command2);
-  if (response2.stream) {
-    for await (const event of response2.stream) {
-      if (event.metadata != null) {
-        console.log('\nRAW metadata event (Call 2):');
-        console.dir(event.metadata, { depth: null });
-        console.log('\nRAW metadata.usage:');
-        console.dir(event.metadata.usage, { depth: null });
-        console.log('\nSpecific cache fields:');
-        console.log(
-          '  cacheReadInputTokens:',
-          (event.metadata.usage as unknown as Record<string, unknown>)
-            ?.cacheReadInputTokens
-        );
-        console.log(
-          '  cacheWriteInputTokens:',
-          (event.metadata.usage as unknown as Record<string, unknown>)
-            ?.cacheWriteInputTokens
-        );
-      }
-    }
-  }
-}
-
-async function wrapperStreamCallNoCachePoint(): Promise<void> {
-  console.log('\n' + '='.repeat(60));
-  console.log(
-    'TEST 2: CustomChatBedrockConverse stream (NO cachePoint) - check usage_metadata'
-  );
-  console.log('='.repeat(60));
-  console.log('(Without cachePoint, Bedrock does NOT return cache tokens)');
-
-  const model = new CustomChatBedrockConverse({
-    model: MODEL_ID,
-    region,
-    credentials,
-    maxTokens: 100,
-    streaming: true,
-    streamUsage: true,
-  });
-
-  console.log('\n--- Wrapper Call (no cachePoint) ---');
-  const messages1 = [new HumanMessage(SYSTEM_PROMPT + '\n\nWhat is 2+2?')];
-  let finalChunk1: AIMessageChunk | undefined;
-
-  for await (const chunk of await model.stream(messages1)) {
-    finalChunk1 = finalChunk1 ? concat(finalChunk1, chunk) : chunk;
-  }
-
-  console.log(
-    '\nFinal usage_metadata:',
-    JSON.stringify(finalChunk1!.usage_metadata)
-  );
-  console.log('(No cache tokens expected since no cachePoint block was sent)');
-}
-
-async function wrapperStreamCallWithCachePoint(): Promise<void> {
-  console.log('\n' + '='.repeat(60));
-  console.log(
-    'TEST 3: Raw SDK with cachePoint -> verify handleConverseStreamMetadata extracts cache tokens'
-  );
-  console.log('='.repeat(60));
-
-  // We use the raw SDK with cachePoint to trigger caching, then verify
-  // that our handleConverseStreamMetadata function properly extracts cache fields
-  const { handleConverseStreamMetadata } = await import(
-    '@/llm/bedrock/utils/message_outputs'
-  );
-
-  const client = new BedrockRuntimeClient({ region, credentials });
-
-  // Call 1 - establish cache
-  console.log('\n--- Call 1 (cache write) ---');
-  const command1 = new ConverseStreamCommand({
-    modelId: MODEL_ID,
-    system: [{ text: SYSTEM_PROMPT }, { cachePoint: { type: 'default' } }],
-    messages: [{ role: 'user', content: [{ text: 'What is 2+2?' }] }],
-    inferenceConfig: { maxTokens: 100 },
-  });
-
-  const response1 = await client.send(command1);
-  if (response1.stream) {
-    for await (const event of response1.stream) {
-      if (event.metadata != null) {
-        console.log('Raw usage:', JSON.stringify(event.metadata.usage));
-
-        // Test our handler
-        const chunk = handleConverseStreamMetadata(event.metadata, {
-          streamUsage: true,
-        });
-        const msg = chunk.message as AIMessageChunk;
-        console.log(
-          'handleConverseStreamMetadata output usage_metadata:',
-          JSON.stringify(msg.usage_metadata)
-        );
-
-        const hasDetails = msg.usage_metadata?.input_token_details != null;
-        console.log(
-          `Has input_token_details: ${hasDetails}`,
-          hasDetails
-            ? JSON.stringify(msg.usage_metadata!.input_token_details)
-            : '(MISSING - BUG!)'
-        );
-      }
-    }
-  }
-
-  // Call 2 - read from cache
-  console.log('\n--- Call 2 (cache read) ---');
-  const command2 = new ConverseStreamCommand({
-    modelId: MODEL_ID,
-    system: [{ text: SYSTEM_PROMPT }, { cachePoint: { type: 'default' } }],
-    messages: [
-      { role: 'user', content: [{ text: 'What is 2+2?' }] },
-      { role: 'assistant', content: [{ text: '4' }] },
-      { role: 'user', content: [{ text: 'What is 3+3?' }] },
-    ],
-    inferenceConfig: { maxTokens: 100 },
-  });
-
-  const response2 = await client.send(command2);
-  if (response2.stream) {
-    for await (const event of response2.stream) {
-      if (event.metadata != null) {
-        console.log('Raw usage:', JSON.stringify(event.metadata.usage));
-
-        const chunk = handleConverseStreamMetadata(event.metadata, {
-          streamUsage: true,
-        });
-        const msg = chunk.message as AIMessageChunk;
-        console.log(
-          'handleConverseStreamMetadata output usage_metadata:',
-          JSON.stringify(msg.usage_metadata)
-        );
-
-        const hasDetails = msg.usage_metadata?.input_token_details != null;
-        console.log(
-          `Has input_token_details: ${hasDetails}`,
-          hasDetails
-            ? JSON.stringify(msg.usage_metadata!.input_token_details)
-            : '(MISSING - BUG!)'
-        );
-      }
-    }
-  }
-}
-
-async function main(): Promise<void> {
-  console.log('Bedrock Cache Token Debug Script');
-  console.log(`Model: ${MODEL_ID}`);
-  console.log(`Region: ${region}\n`);
-
-  await rawSdkCall();
-  await wrapperStreamCallNoCachePoint();
-  await wrapperStreamCallWithCachePoint();
-}
-
-main().catch((err) => {
-  console.error('Fatal error:', err);
-  process.exit(1);
-});

package/src/scripts/bedrock-content-aggregation-test.ts

@@ -1,266 +0,0 @@
-import { config } from 'dotenv';
-config();
-import { HumanMessage, BaseMessage } from '@langchain/core/messages';
-import type { UsageMetadata } from '@langchain/core/messages';
-import * as t from '@/types';
-import { ChatModelStreamHandler, createContentAggregator } from '@/stream';
-import { createCodeExecutionTool } from '@/tools/CodeExecutor';
-import { ToolEndHandler, ModelEndHandler } from '@/events';
-import { GraphEvents, ContentTypes, Providers } from '@/common';
-import { getLLMConfig } from '@/utils/llmConfig';
-import { Run } from '@/run';
-
-const conversationHistory: BaseMessage[] = [];
-let _contentParts: t.MessageContentComplex[] = [];
-const collectedUsage: UsageMetadata[] = [];
-
-async function testBedrockContentAggregation(): Promise<void> {
-  const instructions =
-    'You are a helpful AI assistant with coding capabilities. When answering questions, be thorough in your reasoning.';
-  const { contentParts, aggregateContent } = createContentAggregator();
-  _contentParts = contentParts as t.MessageContentComplex[];
-
-  const customHandlers = {
-    [GraphEvents.TOOL_END]: new ToolEndHandler(),
-    [GraphEvents.CHAT_MODEL_END]: new ModelEndHandler(collectedUsage),
-    [GraphEvents.CHAT_MODEL_STREAM]: new ChatModelStreamHandler(),
-    [GraphEvents.ON_RUN_STEP_COMPLETED]: {
-      handle: (
-        event: GraphEvents.ON_RUN_STEP_COMPLETED,
-        data: t.StreamEventData
-      ): void => {
-        const result = (data as unknown as { result: t.ToolEndEvent }).result;
-        console.log(
-          `[ON_RUN_STEP_COMPLETED] stepId=${result.id} index=${result.index} type=${result.type} tool=${result.tool_call?.name ?? 'n/a'}`
-        );
-        aggregateContent({
-          event,
-          data: data as unknown as { result: t.ToolEndEvent },
-        });
-      },
-    },
-    [GraphEvents.ON_RUN_STEP]: {
-      handle: (event: GraphEvents.ON_RUN_STEP, data: t.RunStep) => {
-        const toolCalls =
-          data.stepDetails.type === 'tool_calls' && data.stepDetails.tool_calls
-            ? (
-                data.stepDetails.tool_calls as Array<{
-                  name?: string;
-                  id?: string;
-                }>
-              )
-                .map((tc) => `${tc.name ?? '?'}(${tc.id ?? '?'})`)
-                .join(', ')
-            : 'none';
-        console.log(
-          `[ON_RUN_STEP] stepId=${data.id} index=${data.index} type=${data.type} stepIndex=${data.stepIndex} toolCalls=[${toolCalls}]`
-        );
-        aggregateContent({ event, data });
-      },
-    },
-    [GraphEvents.ON_RUN_STEP_DELTA]: {
-      handle: (
-        event: GraphEvents.ON_RUN_STEP_DELTA,
-        data: t.RunStepDeltaEvent
-      ) => {
-        const tcNames =
-          data.delta.tool_calls
-            ?.map(
-              (tc) =>
-                `${tc.name ?? '?'}(args=${(tc.args ?? '').substring(0, 30)}...)`
-            )
-            .join(', ') ?? 'none';
-        console.log(
-          `[ON_RUN_STEP_DELTA] stepId=${data.id} type=${data.delta.type} toolCalls=[${tcNames}]`
-        );
-        aggregateContent({ event, data });
-      },
-    },
-    [GraphEvents.ON_MESSAGE_DELTA]: {
-      handle: (
-        event: GraphEvents.ON_MESSAGE_DELTA,
-        data: t.MessageDeltaEvent
-      ) => {
-        const preview = Array.isArray(data.delta.content)
-          ? data.delta.content
-              .map(
-                (c) =>
-                  `${c.type}:"${String((c as Record<string, unknown>).text ?? (c as Record<string, unknown>).think ?? '').substring(0, 40)}"`
-              )
-              .join(', ')
-          : String(data.delta.content).substring(0, 40);
-        console.log(
-          `[ON_MESSAGE_DELTA] stepId=${data.id} content=[${preview}]`
-        );
-        aggregateContent({ event, data });
-      },
-    },
-    [GraphEvents.ON_REASONING_DELTA]: {
-      handle: (
-        event: GraphEvents.ON_REASONING_DELTA,
-        data: t.ReasoningDeltaEvent
-      ) => {
-        const preview = Array.isArray(data.delta.content)
-          ? data.delta.content
-              .map(
-                (c) =>
-                  `${c.type}:"${String((c as Record<string, unknown>).think ?? '').substring(0, 40)}"`
-              )
-              .join(', ')
-          : '?';
-        console.log(
-          `[ON_REASONING_DELTA] stepId=${data.id} content=[${preview}]`
-        );
-        aggregateContent({ event, data });
-      },
-    },
-  };
-
-  const baseLlmConfig = getLLMConfig(Providers.BEDROCK);
-
-  const llmConfig = {
-    ...baseLlmConfig,
-    model: 'global.anthropic.claude-opus-4-6-v1',
-    maxTokens: 16000,
-    additionalModelRequestFields: {
-      thinking: { type: 'enabled', budget_tokens: 10000 },
-    },
-  };
-
-  const run = await Run.create<t.IState>({
-    runId: 'bedrock-content-aggregation-test',
-    graphConfig: {
-      instructions,
-      type: 'standard',
-      tools: [createCodeExecutionTool()],
-      llmConfig,
-    },
-    returnContent: true,
-    skipCleanup: true,
-    customHandlers: customHandlers as t.RunConfig['customHandlers'],
-  });
-
-  const streamConfig = {
-    configurable: {
-      thread_id: 'bedrock-content-aggregation-thread',
-    },
-    streamMode: 'values',
-    version: 'v2' as const,
-  };
-
-  const userMessage = `im testing edge cases with our code interpreter. i know we can persist files, but what happens when we put them in directories?`;
-  conversationHistory.push(new HumanMessage(userMessage));
-
-  console.log('Running Bedrock content aggregation test...\n');
-  console.log(`Prompt: "${userMessage}"\n`);
-
-  const inputs = { messages: [...conversationHistory] };
-  await run.processStream(inputs, streamConfig);
-
-  console.log('\n\n========== CONTENT PARTS ANALYSIS ==========\n');
-
-  let hasEmptyToolCall = false;
-  let hasReasoningOrderIssue = false;
-
-  for (let i = 0; i < _contentParts.length; i++) {
-    const part = _contentParts[i];
-    if (!part) {
-      console.log(`  [${i}] undefined`);
-      continue;
-    }
-
-    const partType = part.type;
-    if (partType === ContentTypes.TOOL_CALL) {
-      const tc = (part as t.ToolCallContent).tool_call;
-      if (!tc || !tc.name) {
-        hasEmptyToolCall = true;
-        console.log(`  [${i}] TOOL_CALL *** EMPTY (no tool_call data) ***`);
-      } else {
-        const outputPreview = tc.output
-          ? `output=${(tc.output as string).substring(0, 80)}...`
-          : 'no output';
-        console.log(`  [${i}] TOOL_CALL name=${tc.name} ${outputPreview}`);
-      }
-    } else if (partType === ContentTypes.THINK) {
-      const think = (part as t.ReasoningContentText).think ?? '';
-      console.log(
-        `  [${i}] THINK (${think.length} chars): "${think.substring(0, 80)}..."`
-      );
-    } else if (partType === ContentTypes.TEXT) {
-      const text = (part as t.MessageDeltaUpdate).text ?? '';
-      console.log(
-        `  [${i}] TEXT (${text.length} chars): "${text.substring(0, 80)}..."`
-      );
-    } else {
-      console.log(`  [${i}] ${partType}`);
-    }
-  }
-
-  /**
-   * Check reasoning ordering within a single invocation cycle.
-   * A tool_call resets the cycle — text before think across different
-   * invocations (e.g., text from invocation 2, think from invocation 3) is valid.
-   */
-  let lastTextInCycle: number | null = null;
-  for (let i = 0; i < _contentParts.length; i++) {
-    const part = _contentParts[i];
-    if (!part) continue;
-
-    if (part.type === ContentTypes.TOOL_CALL) {
-      lastTextInCycle = null;
-      continue;
-    }
-
-    if (part.type === ContentTypes.TEXT) {
-      lastTextInCycle = i;
-    } else if (part.type === ContentTypes.THINK && lastTextInCycle !== null) {
-      const prevText = _contentParts[lastTextInCycle] as t.MessageDeltaUpdate;
-      const thinkContent = (part as t.ReasoningContentText).think ?? '';
-      if (
-        prevText?.text &&
-        prevText.text.trim().length > 5 &&
-        thinkContent.length > 0
-      ) {
-        hasReasoningOrderIssue = true;
-        console.log(
-          `\n  *** ORDERING ISSUE (same invocation): TEXT at [${lastTextInCycle}] appears before THINK at [${i}]`
-        );
-        console.log(
-          `      Text ends with: "...${prevText.text.substring(prevText.text.length - 60)}"`
-        );
-        console.log(
-          `      Think starts with: "${thinkContent.substring(0, 60)}..."`
-        );
-      }
-    }
-  }
-
-  console.log('\n========== SUMMARY ==========\n');
-  console.log(`Total content parts: ${_contentParts.filter(Boolean).length}`);
-  console.log(
-    `Empty tool_call parts: ${hasEmptyToolCall ? 'YES (BUG)' : 'No'}`
-  );
-  console.log(
-    `Reasoning order issues: ${hasReasoningOrderIssue ? 'YES (BUG)' : 'No'}`
-  );
-  console.log('\nFull contentParts dump:');
-  console.dir(_contentParts, { depth: null });
-}
-
-process.on('unhandledRejection', (reason, promise) => {
-  console.error('Unhandled Rejection at:', promise, 'reason:', reason);
-  console.log('Content parts:');
-  console.dir(_contentParts, { depth: null });
-  process.exit(1);
-});
-
-process.on('uncaughtException', (err) => {
-  console.error('Uncaught Exception:', err);
-});
-
-testBedrockContentAggregation().catch((err) => {
-  console.error(err);
-  console.log('Content parts:');
-  console.dir(_contentParts, { depth: null });
-  process.exit(1);
-});

package/src/scripts/bedrock-merge-test.ts

@@ -1,107 +0,0 @@
-import { config } from 'dotenv';
-config();
-import { HumanMessage } from '@langchain/core/messages';
-import type { AIMessageChunk } from '@langchain/core/messages';
-import { concat } from '@langchain/core/utils/stream';
-import { CustomChatBedrockConverse } from '@/llm/bedrock';
-import { modifyDeltaProperties } from '@/messages/core';
-import { Providers } from '@/common';
-
-async function testBedrockMerge(): Promise<void> {
-  const model = new CustomChatBedrockConverse({
-    model: 'us.anthropic.claude-sonnet-4-5-20250929-v1:0',
-    region: process.env.BEDROCK_AWS_REGION,
-    credentials: {
-      accessKeyId: process.env.BEDROCK_AWS_ACCESS_KEY_ID!,
-      secretAccessKey: process.env.BEDROCK_AWS_SECRET_ACCESS_KEY!,
-    },
-    maxTokens: 4000,
-    streaming: true,
-    streamUsage: true,
-    additionalModelRequestFields: {
-      thinking: { type: 'enabled', budget_tokens: 2000 },
-    },
-  });
-
-  const messages = [new HumanMessage('What is 25 * 37? Think step by step.')];
-
-  console.log('Streaming from Bedrock with thinking enabled...\n');
-
-  const stream = await model.stream(messages);
-  let finalChunk: AIMessageChunk | undefined;
-  let chunkCount = 0;
-  let firstTextLogged = false;
-
-  for await (const chunk of stream) {
-    chunkCount++;
-    const isArr = Array.isArray(chunk.content);
-    const isStr = typeof chunk.content === 'string';
-    const isTextStr = isStr && (chunk.content as string).length > 0;
-
-    if (!firstTextLogged && isTextStr) {
-      console.log(
-        `chunk ${chunkCount} (first text): contentType=string, value="${chunk.content}"`
-      );
-      console.log(
-        '  response_metadata:',
-        JSON.stringify(chunk.response_metadata)
-      );
-      firstTextLogged = true;
-    }
-
-    if (isArr) {
-      const blocks = chunk.content as Array<Record<string, unknown>>;
-      const info = blocks.map((b) => ({
-        type: b.type,
-        hasIndex: 'index' in b,
-        index: b.index,
-      }));
-      console.log(`chunk ${chunkCount}: array content, blocks:`, info);
-    }
-
-    finalChunk = finalChunk ? concat(finalChunk, chunk) : chunk;
-  }
-
-  console.log(`Total chunks received: ${chunkCount}\n`);
-
-  console.log('=== RAW concat result (before modifyDeltaProperties) ===');
-  console.log('content type:', typeof finalChunk!.content);
-  if (Array.isArray(finalChunk!.content)) {
-    console.log('content array length:', finalChunk!.content.length);
-    const types = finalChunk!.content.map((b) =>
-      typeof b === 'object' && 'type' in b ? b.type : typeof b
-    );
-    const typeCounts = types.reduce(
-      (acc, t) => {
-        acc[t ?? ''] = (acc[t ?? ''] || 0) + 1;
-        return acc;
-      },
-      {} as Record<string, number>
-    );
-    console.log('content block type counts:', typeCounts);
-  }
-
-  console.log('\ncontent:');
-  console.dir(finalChunk!.content, { depth: null });
-
-  console.log('\n=== lc_kwargs.content ===');
-  if (Array.isArray(finalChunk!.lc_kwargs.content)) {
-    console.log(
-      'lc_kwargs.content length:',
-      finalChunk!.lc_kwargs.content.length
-    );
-  }
-  console.dir(finalChunk!.lc_kwargs.content, { depth: null });
-
-  const modified = modifyDeltaProperties(Providers.BEDROCK, finalChunk);
-  console.log('\n=== After modifyDeltaProperties ===');
-  console.log('content:');
-  console.dir(modified!.content, { depth: null });
-  console.log('\nlc_kwargs.content:');
-  console.dir(modified!.lc_kwargs.content, { depth: null });
-}
-
-testBedrockMerge().catch((err) => {
-  console.error(err);
-  process.exit(1);
-});