@librechat/agents 3.1.74 → 3.1.75

package/src/messages/cache.test.ts

@@ -1,8 +1,9 @@
 import {
   AIMessage,
   BaseMessage,
-  ToolMessage,
   HumanMessage,
+  SystemMessage,
+  ToolMessage,
   MessageContentComplex,
 } from '@langchain/core/messages';
 import type Anthropic from '@anthropic-ai/sdk';
@@ -404,7 +405,107 @@ describe('addBedrockCacheControl (Bedrock cache checkpoints)', () => {
     expect(first[1]).toEqual({ cachePoint: { type: 'default' } });
   });
 
-  it('works with the example from the langchain pr (with multi-turn behavior)', () => {
+  it('preserves LangChain system message content unchanged', () => {
+    const systemContent = [
+      { type: ContentTypes.TEXT, text: 'Stable system text' },
+      { cachePoint: { type: 'default' } },
+      { type: ContentTypes.TEXT, text: 'Dynamic system text' },
+    ] as MessageContentComplex[];
+    const messages: BaseMessage[] = [
+      new SystemMessage({ content: systemContent }),
+      new HumanMessage('Hello'),
+      new AIMessage('Hi'),
+    ];
+
+    const result = addBedrockCacheControl(messages);
+
+    expect(result[0]).toBe(messages[0]);
+    expect(result[0].content).toEqual(systemContent);
+  });
+
+  it('preserves serialized system message content unchanged', () => {
+    const systemContent = [
+      { type: ContentTypes.TEXT, text: 'Stable system text' },
+      { cachePoint: { type: 'default' } },
+      { type: ContentTypes.TEXT, text: 'Dynamic system text' },
+    ] as MessageContentComplex[];
+    const messages: TestMsg[] = [
+      { role: 'system', content: systemContent },
+      { role: 'user', content: 'Hello' },
+      { role: 'assistant', content: 'Hi' },
+    ];
+
+    const result = addBedrockCacheControl(messages);
+
+    expect(result[0]).toBe(messages[0]);
+    expect(result[0].content).toEqual(systemContent);
+  });
+
+  it('strips Anthropic cache_control from LangChain system messages without moving cache points', () => {
+    const systemContent = [
+      {
+        type: ContentTypes.TEXT,
+        text: 'Stable system text',
+        cache_control: { type: 'ephemeral' },
+      } as MessageContentComplex,
+      { cachePoint: { type: 'default' } },
+      {
+        type: ContentTypes.TEXT,
+        text: 'Dynamic system text',
+        cache_control: { type: 'ephemeral' },
+      } as MessageContentComplex,
+    ] as MessageContentComplex[];
+    const messages: BaseMessage[] = [
+      new SystemMessage({ content: systemContent }),
+      new HumanMessage('Hello'),
+      new AIMessage('Hi'),
+    ];
+
+    const result = addBedrockCacheControl(messages);
+
+    expect(result[0]).not.toBe(messages[0]);
+    expect(result[0].content).toEqual([
+      { type: ContentTypes.TEXT, text: 'Stable system text' },
+      { cachePoint: { type: 'default' } },
+      { type: ContentTypes.TEXT, text: 'Dynamic system text' },
+    ]);
+    expect(systemContent[0]).toHaveProperty('cache_control');
+    expect(systemContent[2]).toHaveProperty('cache_control');
+  });
+
+  it('strips Anthropic cache_control from serialized system messages without moving cache points', () => {
+    const systemContent = [
+      {
+        type: ContentTypes.TEXT,
+        text: 'Stable system text',
+        cache_control: { type: 'ephemeral' },
+      } as MessageContentComplex,
+      { cachePoint: { type: 'default' } },
+      {
+        type: ContentTypes.TEXT,
+        text: 'Dynamic system text',
+        cache_control: { type: 'ephemeral' },
+      } as MessageContentComplex,
+    ] as MessageContentComplex[];
+    const messages: TestMsg[] = [
+      { role: 'system', content: systemContent },
+      { role: 'user', content: 'Hello' },
+      { role: 'assistant', content: 'Hi' },
+    ];
+
+    const result = addBedrockCacheControl(messages);
+
+    expect(result[0]).not.toBe(messages[0]);
+    expect(result[0].content).toEqual([
+      { type: ContentTypes.TEXT, text: 'Stable system text' },
+      { cachePoint: { type: 'default' } },
+      { type: ContentTypes.TEXT, text: 'Dynamic system text' },
+    ]);
+    expect(systemContent[0]).toHaveProperty('cache_control');
+    expect(systemContent[2]).toHaveProperty('cache_control');
+  });
+
+  it('skips serialized system messages while adding cache points to non-system turns', () => {
     const messages: TestMsg[] = [
       {
         role: 'system',
@@ -429,7 +530,7 @@ describe('addBedrockCacheControl (Bedrock cache checkpoints)', () => {
       type: ContentTypes.TEXT,
       text: 'You\'re an advanced AI assistant.',
     });
-    expect(system[1]).toEqual({ cachePoint: { type: 'default' } });
+    expect(system).toHaveLength(1);
     expect(user[0]).toEqual({
       type: ContentTypes.TEXT,
       text: 'What is the capital of France?',

package/src/messages/cache.ts

@@ -14,6 +14,10 @@ type MessageWithContent = {
   content?: string | MessageContentComplex[];
 };
 
+type MessageContentWithCacheControl = MessageContentComplex & {
+  cache_control?: unknown;
+};
+
 /**
  * Deep clones a message's content to prevent mutation of the original.
  */
@@ -101,6 +105,40 @@ function cloneMessage<T extends MessageWithContent>(
   return cloned;
 }
 
+function stripAnthropicCacheControlFromBlocks(
+  content: MessageContentComplex[]
+): { content: MessageContentComplex[]; modified: boolean } {
+  let modified = false;
+  const strippedContent = content.map((block) => {
+    if (!('cache_control' in block)) {
+      return block;
+    }
+
+    const cloned: MessageContentWithCacheControl = { ...block };
+    delete cloned.cache_control;
+    modified = true;
+    return cloned;
+  });
+
+  return { content: strippedContent, modified };
+}
+
+function sanitizeBedrockSystemMessage<T extends MessageWithContent>(
+  message: T
+): T {
+  const content = message.content;
+  if (!Array.isArray(content)) {
+    return message;
+  }
+
+  const stripped = stripAnthropicCacheControlFromBlocks(content);
+  if (!stripped.modified) {
+    return message;
+  }
+
+  return cloneMessage(message, stripped.content);
+}
+
 /**
  * Anthropic API: Adds cache control to the appropriate user messages in the payload.
  * Strips ALL existing cache control (both Anthropic and Bedrock formats) from all messages,
@@ -310,11 +348,24 @@ export function addBedrockCacheControl<
 
   for (let i = updatedMessages.length - 1; i >= 0; i--) {
     const originalMessage = updatedMessages[i];
-    const isToolMessage =
+    const messageType =
       'getType' in originalMessage &&
-      typeof originalMessage.getType === 'function' &&
-      originalMessage.getType() === 'tool';
+      typeof originalMessage.getType === 'function'
+        ? originalMessage.getType()
+        : undefined;
+    const messageRole =
+      'role' in originalMessage && typeof originalMessage.role === 'string'
+        ? originalMessage.role
+        : undefined;
+
+    const isSystemMessage =
+      messageType === 'system' || messageRole === 'system';
+    if (isSystemMessage) {
+      updatedMessages[i] = sanitizeBedrockSystemMessage(originalMessage);
+      continue;
+    }
 
+    const isToolMessage = messageType === 'tool' || messageRole === 'tool';
     const content = originalMessage.content;
     const hasArrayContent = Array.isArray(content);
     const isEmptyString = typeof content === 'string' && content === '';

package/src/specs/anthropic.simple.test.ts

@@ -376,6 +376,67 @@ describe(`${capitalizeFirstLetter(provider)} Streaming Tests`, () => {
     );
   });
 
+  test(`${capitalizeFirstLetter(provider)}: follow-up after assistant message with only whitespace text content`, async () => {
+    /**
+     * Regression for LibreChat discussion #12806.
+     *
+     * The Anthropic API has two distinct rejection rules (verified against
+     * the live API):
+     *   1. Strict empty `text: ''`  → rejected anywhere
+     *      "messages: text content blocks must be non-empty"
+     *   2. Whitespace-only `text: ' '` / '\n' / '\t' → rejected when the
+     *      assistant message has no other accepted blocks (no tool blocks,
+     *      no non-whitespace text)
+     *      "messages: text content blocks must contain non-whitespace text"
+     *
+     * Anthropic responses for some prompts include a whitespace-only text
+     * block as the sole text content. Re-sending that history on a
+     * follow-up turn triggers rule 2.
+     *
+     * The wire-send filter in `_formatContent` must drop any text block
+     * whose trimmed content is empty. The previous filter used strict
+     * `text === ''` only, which caught rule 1 but not rule 2.
+     */
+    const llmConfig = getLLMConfig(provider);
+    const customHandlers1 = setupCustomHandlers();
+
+    const followUpRun = await Run.create<t.IState>({
+      runId: 'repro-12806-followup',
+      graphConfig: {
+        type: 'standard',
+        llmConfig,
+        instructions: 'You are a friendly AI assistant.',
+      },
+      returnContent: true,
+      skipCleanup: true,
+      customHandlers: customHandlers1,
+    });
+
+    // Build history with an assistant message whose entire content array
+    // is a single whitespace-only text block. This is the precise shape
+    // the API rejects under rule 2 above.
+    conversationHistory = [
+      new HumanMessage('hi'),
+      new (require('@langchain/core/messages').AIMessage)({
+        content: [{ type: 'text', text: ' ' }],
+      }),
+      new HumanMessage('please respond with a short greeting'),
+    ];
+
+    // With the fix: `_formatContent` drops the whitespace text block,
+    // the assistant content becomes an empty array, and the API accepts.
+    // Without the fix: the whitespace block is forwarded and the API
+    // rejects with "messages: text content blocks must contain non-whitespace text".
+    const finalContentParts = await followUpRun.processStream(
+      { messages: conversationHistory },
+      config
+    );
+    expect(finalContentParts).toBeDefined();
+    const finalMessages = followUpRun.getRunMessages();
+    expect(finalMessages).toBeDefined();
+    expect(finalMessages?.length).toBeGreaterThan(0);
+  });
+
   test('should handle errors appropriately', async () => {
     // Test error scenarios
     await expect(async () => {

package/src/specs/summarization.test.ts

@@ -22,6 +22,8 @@ import { formatAgentMessages } from '@/messages/format';
 import { FakeListChatModel } from '@langchain/core/utils/testing';
 import * as providers from '@/llm/providers';
 
+const SUMMARY_WRAPPER_OVERHEAD_TOKENS = 33;
+
 /** Extract plain text from a SummaryContentBlock's content array (test helper). */
 function getSummaryText(summary: t.SummaryContentBlock | undefined): string {
   if (!summary) return '';
@@ -1443,7 +1445,8 @@ describe('Cross-run summary lifecycle (no API keys)', () => {
     expect(completePayload.summary!.tokenCount ?? 0).toBeGreaterThan(0);
 
     const expectedTokenCount =
-      tokenCounter(new SystemMessage(KNOWN_SUMMARY)) + 33;
+      tokenCounter(new SystemMessage(KNOWN_SUMMARY)) +
+      SUMMARY_WRAPPER_OVERHEAD_TOKENS;
     expect(completePayload.summary!.tokenCount).toBe(expectedTokenCount);
 
     const summaryBlock = completePayload.summary!;
@@ -2605,8 +2608,9 @@ const hasAnyApiKey =
     const summaryText = getSummaryText(completePayload.summary);
     const reportedTokenCount = completePayload.summary!.tokenCount ?? 0;
 
-    // Count tokens locally using the same tokenizer
-    const localTokenCount = tokenCounter(new SystemMessage(summaryText));
+    const localTokenCount =
+      tokenCounter(new SystemMessage(summaryText)) +
+      SUMMARY_WRAPPER_OVERHEAD_TOKENS;
 
     console.log(
       `  Token match: reported=${reportedTokenCount}, local=${localTokenCount}`

package/src/types/graph.ts

@@ -471,10 +471,12 @@ export interface AgentInputs {
   toolMap?: ToolMap;
   tools?: GraphTools;
   provider: Providers;
+  /** Stable/cacheable system instructions. */
   instructions?: string;
   streamBuffer?: number;
   maxContextTokens?: number;
   clientOptions?: ClientOptions;
+  /** Dynamic system tail appended after stable instructions without provider cache markers. */
   additional_instructions?: string;
   reasoningKey?: 'reasoning_content' | 'reasoning';
   /** Format content blocks as strings (for legacy compatibility i.e. Ollama/Azure Serverless) */
@@ -500,7 +502,7 @@ export interface AgentInputs {
   summarizationEnabled?: boolean;
   summarizationConfig?: SummarizationConfig;
   /** Cross-run summary from a previous run, forwarded from formatAgentMessages.
-   *  Injected into the system message via AgentContext.buildInstructionsString(). */
+   *  Injected into the dynamic system tail via AgentContext. */
   initialSummary?: { text: string; tokenCount: number };
   contextPruningConfig?: ContextPruningConfig;
   maxToolResultChars?: number;

package/src/types/run.ts

@@ -75,7 +75,9 @@ export interface AgentStateChannels {
   messages: BaseMessage[];
   next: string;
   [key: string]: unknown;
+  /** Stable/cacheable system instructions. */
   instructions?: string;
+  /** Dynamic system tail appended after stable instructions. */
   additional_instructions?: string;
 }