@librechat/agents 3.1.73 → 3.1.75-dev.0

package/src/tools/BashExecutor.ts

@@ -3,17 +3,23 @@ import fetch, { RequestInit } from 'node-fetch';
 import { HttpsProxyAgent } from 'https-proxy-agent';
 import { tool, DynamicStructuredTool } from '@langchain/core/tools';
 import type * as t from '@/types';
-import { imageExtRegex, getCodeBaseURL } from './CodeExecutor';
+import { getCodeBaseURL, renderFileSection } from './CodeExecutor';
 import { Constants } from '@/common';
 
 config();
 
-const imageMessage = 'Image is already displayed to the user';
 const otherMessage = 'File is already downloaded by the user';
+const inheritedFileMessage =
+  'Available as an input — already known to the user';
 const accessMessage =
   'Note: Files from previous executions are automatically available and can be modified.';
 const emptyOutputMessage =
   'stdout: Empty. Ensure you\'re writing output explicitly.\n';
+const inheritedFilesHeader =
+  'Available files (inputs, not generated by this execution):';
+const generatedFilesHeader = 'Generated files:';
+const inheritedNote =
+  'Note: Files in "Available files" are inputs the user (or a skill) already provided to the sandbox. They were not produced by this execution and you should not present them as new outputs in your response.';
 
 const baseEndpoint = getCodeBaseURL();
 const EXEC_ENDPOINT = `${baseEndpoint}/exec`;
@@ -198,20 +204,38 @@ function createBashExecutionTool(
         }
         if (result.stderr) formattedOutput += `stderr:\n${result.stderr}\n`;
         if (result.files && result.files.length > 0) {
-          formattedOutput += 'Generated files:\n';
+          /* Split inherited (read-only / unchanged-input passthroughs from
+           * codeapi) from genuine generated outputs. The LLM was previously
+           * shown skill files under "Generated files:" with the message
+           * "File is already downloaded by the user", which led it to
+           * (a) believe it had just produced files it merely referenced
+           * and (b) sometimes invent paths like /mnt/user-data/uploads/
+           * trying to find the "originals". Labeling them as inputs makes
+           * the mental model accurate. */
+          const inheritedFiles = result.files.filter(
+            (f) => f.inherited === true
+          );
+          const generatedFiles = result.files.filter(
+            (f) => f.inherited !== true
+          );
 
-          const fileCount = result.files.length;
-          for (let i = 0; i < fileCount; i++) {
-            const file = result.files[i];
-            const isImage = imageExtRegex.test(file.name);
-            formattedOutput += `- /mnt/data/${file.name} | ${isImage ? imageMessage : otherMessage}`;
+          formattedOutput += renderFileSection(
+            generatedFilesHeader,
+            generatedFiles,
+            otherMessage
+          );
+          formattedOutput += renderFileSection(
+            inheritedFilesHeader,
+            inheritedFiles,
+            inheritedFileMessage
+          );
 
-            if (i < fileCount - 1) {
-              formattedOutput += fileCount <= 3 ? ', ' : ',\n';
-            }
+          if (generatedFiles.length > 0) {
+            formattedOutput += `\n\n${accessMessage}`;
+          }
+          if (inheritedFiles.length > 0) {
+            formattedOutput += `\n\n${inheritedNote}`;
           }
-
-          formattedOutput += `\n\n${accessMessage}`;
           return [
             formattedOutput.trim(),
             {

package/src/tools/CodeExecutor.ts

@@ -15,10 +15,41 @@ export const getCodeBaseURL = (): string =>
 
 const imageMessage = 'Image is already displayed to the user';
 const otherMessage = 'File is already downloaded by the user';
+const inheritedFileMessage =
+  'Available as an input — already known to the user';
 const accessMessage =
   'Note: Files from previous executions are automatically available and can be modified.';
 const emptyOutputMessage =
   'stdout: Empty. Ensure you\'re writing output explicitly.\n';
+const inheritedFilesHeader =
+  'Available files (inputs, not generated by this execution):';
+const generatedFilesHeader = 'Generated files:';
+const inheritedNote =
+  'Note: Files in "Available files" are inputs the user (or a skill) already provided to the sandbox. They were not produced by this execution and you should not present them as new outputs in your response.';
+
+/**
+ * Renders one section of the post-execution file listing. Used by the
+ * code/bash tool formatters to keep generated outputs and inherited
+ * inputs visually separated. See BashExecutor for full docs.
+ */
+export function renderFileSection(
+  header: string,
+  files: t.FileRefs,
+  defaultMessage: string
+): string {
+  if (files.length === 0) return '';
+  let out = `${header}\n`;
+  for (let i = 0; i < files.length; i++) {
+    const file = files[i];
+    const isImage = imageExtRegex.test(file.name);
+    out += `- /mnt/data/${file.name} | ${isImage ? imageMessage : defaultMessage}`;
+    if (i < files.length - 1) {
+      out += files.length <= 3 ? ', ' : ',\n';
+    }
+  }
+  out += '\n';
+  return out;
+}
 
 const SUPPORTED_LANGUAGES = [
   'py',
@@ -196,20 +227,33 @@ function createCodeExecutionTool(
         }
         if (result.stderr) formattedOutput += `stderr:\n${result.stderr}\n`;
         if (result.files && result.files.length > 0) {
-          formattedOutput += 'Generated files:\n';
+          /* See BashExecutor for the rationale: split inherited (read-only
+           * passthrough) inputs from real generated outputs so the LLM
+           * doesn't conflate skill files with newly-produced artifacts. */
+          const inheritedFiles = result.files.filter(
+            (f) => f.inherited === true
+          );
+          const generatedFiles = result.files.filter(
+            (f) => f.inherited !== true
+          );
 
-          const fileCount = result.files.length;
-          for (let i = 0; i < fileCount; i++) {
-            const file = result.files[i];
-            const isImage = imageExtRegex.test(file.name);
-            formattedOutput += `- /mnt/data/${file.name} | ${isImage ? imageMessage : otherMessage}`;
+          formattedOutput += renderFileSection(
+            generatedFilesHeader,
+            generatedFiles,
+            otherMessage
+          );
+          formattedOutput += renderFileSection(
+            inheritedFilesHeader,
+            inheritedFiles,
+            inheritedFileMessage
+          );
 
-            if (i < fileCount - 1) {
-              formattedOutput += fileCount <= 3 ? ', ' : ',\n';
-            }
+          if (generatedFiles.length > 0) {
+            formattedOutput += `\n\n${accessMessage}`;
+          }
+          if (inheritedFiles.length > 0) {
+            formattedOutput += `\n\n${inheritedNote}`;
           }
-
-          formattedOutput += `\n\n${accessMessage}`;
           return [
             formattedOutput.trim(),
             {

package/src/tools/ProgrammaticToolCalling.ts

@@ -5,7 +5,7 @@ import { HttpsProxyAgent } from 'https-proxy-agent';
 import { tool, DynamicStructuredTool } from '@langchain/core/tools';
 import type { ToolCall } from '@langchain/core/messages/tool';
 import type * as t from '@/types';
-import { imageExtRegex, getCodeBaseURL } from './CodeExecutor';
+import { getCodeBaseURL, renderFileSection } from './CodeExecutor';
 import { Constants } from '@/common';
 
 config();
@@ -14,8 +14,14 @@ config();
 // Constants
 // ============================================================================
 
-const imageMessage = 'Image is already displayed to the user';
 const otherMessage = 'File is already downloaded by the user';
+const inheritedFileMessage =
+  'Available as an input — already known to the user';
+const inheritedFilesHeader =
+  'Available files (inputs, not generated by this execution):';
+const generatedFilesHeader = 'Generated files:';
+const inheritedNote =
+  'Note: Files in "Available files" are inputs the user (or a skill) already provided to the sandbox. They were not produced by this execution and you should not present them as new outputs in your response.';
 const accessMessage =
   'Note: Files from previous executions are automatically available and can be modified.';
 const emptyOutputMessage =
@@ -552,20 +558,29 @@ export function formatCompletedResponse(
   }
 
   if (response.files && response.files.length > 0) {
-    formatted += 'Generated files:\n';
-
-    const fileCount = response.files.length;
-    for (let i = 0; i < fileCount; i++) {
-      const file = response.files[i];
-      const isImage = imageExtRegex.test(file.name);
-      formatted += `- /mnt/data/${file.name} | ${isImage ? imageMessage : otherMessage}`;
+    /* See BashExecutor for the rationale: split inherited (read-only
+     * passthrough) inputs from real generated outputs so the LLM doesn't
+     * conflate skill files with newly-produced artifacts. */
+    const inheritedFiles = response.files.filter((f) => f.inherited === true);
+    const generatedFiles = response.files.filter((f) => f.inherited !== true);
+
+    formatted += renderFileSection(
+      generatedFilesHeader,
+      generatedFiles,
+      otherMessage
+    );
+    formatted += renderFileSection(
+      inheritedFilesHeader,
+      inheritedFiles,
+      inheritedFileMessage
+    );
 
-      if (i < fileCount - 1) {
-        formatted += fileCount <= 3 ? ', ' : ',\n';
-      }
+    if (generatedFiles.length > 0) {
+      formatted += `\n\n${accessMessage}`;
+    }
+    if (inheritedFiles.length > 0) {
+      formatted += `\n\n${inheritedNote}`;
     }
-
-    formatted += `\n\n${accessMessage}`;
   }
 
   return [

package/src/tools/ToolNode.ts

@@ -33,6 +33,7 @@ import {
 } from '@/utils/truncation';
 import { safeDispatchCustomEvent } from '@/utils/events';
 import { executeHooks } from '@/hooks';
+import { toLangChainContent } from '@/messages/langchain';
 import { Constants, GraphEvents, CODE_EXECUTION_TOOLS } from '@/common';
 import {
   buildReferenceKey,
@@ -1282,7 +1283,10 @@ export class ToolNode<T = any> extends RunnableCallable<T, T> {
       if (msg.skillName != null) additional_kwargs.skillName = msg.skillName;
 
       converted.push(
-        new HumanMessage({ content: msg.content, additional_kwargs })
+        new HumanMessage({
+          content: toLangChainContent(msg.content),
+          additional_kwargs,
+        })
       );
     }
     return converted;

package/src/tools/__tests__/ProgrammaticToolCalling.test.ts

@@ -664,6 +664,66 @@ for member in team:
       expect(output).toContain('chart.png');
       expect(output).toContain('Image is already displayed to the user');
     });
+
+    it('splits inherited inputs from generated outputs into distinct sections', () => {
+      const response: t.ProgrammaticExecutionResponse = {
+        status: 'completed',
+        stdout: 'analysis done\n',
+        stderr: '',
+        files: [
+          { id: 'g1', name: 'report.pdf' },
+          { id: 'i1', name: 'pptx/SKILL.md', inherited: true },
+          { id: 'i2', name: 'pptx/scripts/clean.py', inherited: true },
+          { id: 'g2', name: 'chart.png' },
+        ],
+        session_id: 'sess_abc123',
+      };
+
+      const [output, artifact] = formatCompletedResponse(response);
+
+      /* Generated section lists only outputs the run produced. */
+      const generatedIdx = output.indexOf('Generated files:');
+      const inheritedIdx = output.indexOf('Available files (inputs');
+      expect(generatedIdx).toBeGreaterThan(-1);
+      expect(inheritedIdx).toBeGreaterThan(generatedIdx);
+
+      /* Slice each section so we can assert membership without
+       * cross-talk between the two listings. */
+      const generatedSection = output.slice(generatedIdx, inheritedIdx);
+      const inheritedSection = output.slice(inheritedIdx);
+
+      expect(generatedSection).toContain('report.pdf');
+      expect(generatedSection).toContain('chart.png');
+      expect(generatedSection).not.toContain('SKILL.md');
+
+      expect(inheritedSection).toContain('pptx/SKILL.md');
+      expect(inheritedSection).toContain('pptx/scripts/clean.py');
+      expect(inheritedSection).toContain('Available as an input');
+
+      /* The artifact still carries every file so the host can still
+       * thread per-file ids through to subsequent calls. */
+      expect(artifact.files).toHaveLength(4);
+    });
+
+    it('omits the Generated files header when every entry is inherited', () => {
+      const response: t.ProgrammaticExecutionResponse = {
+        status: 'completed',
+        stdout: 'cat: ok\n',
+        stderr: '',
+        files: [
+          { id: 'i1', name: 'pptx/SKILL.md', inherited: true },
+          { id: 'i2', name: 'pptx/editing.md', inherited: true },
+        ],
+        session_id: 'sess_abc123',
+      };
+
+      const [output] = formatCompletedResponse(response);
+
+      expect(output).not.toContain('Generated files:');
+      expect(output).toContain('Available files (inputs');
+      expect(output).toContain('pptx/SKILL.md');
+      expect(output).toContain('pptx/editing.md');
+    });
   });
 
   describe('createProgrammaticToolCallingTool - Manual Invocation', () => {

package/src/types/graph.ts

@@ -1,14 +1,5 @@
 // src/types/graph.ts
-import type {
-  START,
-  StateType,
-  UpdateType,
-  StateGraph,
-  StateGraphArgs,
-  StateDefinition,
-  CompiledStateGraph,
-  BinaryOperatorAggregate,
-} from '@langchain/langgraph';
+import type { START, StateGraph, StateGraphArgs } from '@langchain/langgraph';
 import type { BindToolsInput } from '@langchain/core/language_models/chat_models';
 import type {
   BaseMessage,
@@ -129,76 +120,40 @@ export type Workflow<
   N extends string = string,
 > = StateGraph<T, U, N>;
 
-export type CompiledWorkflow<
-  T extends BaseGraphState = BaseGraphState,
-  U extends Partial<T> = Partial<T>,
-  N extends string = string,
-> = CompiledStateGraph<T, U, N>;
-
-export type CompiledStateWorkflow = CompiledStateGraph<
-  StateType<{
-    messages: BinaryOperatorAggregate<BaseMessage[], BaseMessage[]>;
-  }>,
-  UpdateType<{
-    messages: BinaryOperatorAggregate<BaseMessage[], BaseMessage[]>;
-  }>,
-  string,
-  {
-    messages: BinaryOperatorAggregate<BaseMessage[], BaseMessage[]>;
-  },
-  {
-    messages: BinaryOperatorAggregate<BaseMessage[], BaseMessage[]>;
-  },
-  StateDefinition
+type LangChainEventStreamCallbackHandlerInput = NonNullable<
+  Parameters<Runnable['streamEvents']>[2]
 >;
 
-export type CompiledMultiAgentWorkflow = CompiledStateGraph<
-  StateType<{
-    messages: BinaryOperatorAggregate<BaseMessage[], BaseMessage[]>;
-    agentMessages: BinaryOperatorAggregate<BaseMessage[], BaseMessage[]>;
-  }>,
-  UpdateType<{
-    messages: BinaryOperatorAggregate<BaseMessage[], BaseMessage[]>;
-    agentMessages: BinaryOperatorAggregate<BaseMessage[], BaseMessage[]>;
-  }>,
-  string,
-  {
-    messages: BinaryOperatorAggregate<BaseMessage[], BaseMessage[]>;
-    agentMessages: BinaryOperatorAggregate<BaseMessage[], BaseMessage[]>;
-  },
-  {
-    messages: BinaryOperatorAggregate<BaseMessage[], BaseMessage[]>;
-    agentMessages: BinaryOperatorAggregate<BaseMessage[], BaseMessage[]>;
-  },
-  StateDefinition
->;
+export type EventStreamCallbackHandlerInput =
+  LangChainEventStreamCallbackHandlerInput & {
+    autoClose?: boolean;
+    raiseError?: boolean;
+    ignoreCustomEvent?: boolean;
+  };
 
-export type CompiledAgentWorfklow = CompiledStateGraph<
+export type WorkflowValuesStreamConfig = RunnableConfig & {
+  streamMode: 'values';
+};
+
+/**
+ * LangGraph stream output is mode-dependent (`values`, `updates`, SSE, etc.).
+ * Keep the base Runnable stream output as unknown and narrow at callsites that
+ * choose a concrete streamMode.
+ */
+export type CompiledWorkflow<
+  TInput extends BaseGraphState = BaseGraphState,
+  TOutput extends BaseGraphState = TInput,
+> = Omit<Runnable<TInput, unknown>, 'invoke'> & {
+  invoke(input: TInput, config?: RunnableConfig): Promise<TOutput>;
+};
+
+export type CompiledStateWorkflow = CompiledWorkflow;
+
+export type CompiledMultiAgentWorkflow = CompiledWorkflow<MultiAgentGraphState>;
+
+export type CompiledAgentWorfklow = CompiledWorkflow<
   AgentSubgraphState,
-  Partial<AgentSubgraphState>,
-  '__start__' | `agent=${string}` | `tools=${string}` | `summarize=${string}`,
-  {
-    messages: BinaryOperatorAggregate<BaseMessage[], BaseMessage[]>;
-    summarizationRequest: BinaryOperatorAggregate<
-      SummarizationNodeInput | undefined,
-      SummarizationNodeInput | undefined
-    >;
-  },
-  {
-    messages: BinaryOperatorAggregate<BaseMessage[], BaseMessage[]>;
-    summarizationRequest: BinaryOperatorAggregate<
-      SummarizationNodeInput | undefined,
-      SummarizationNodeInput | undefined
-    >;
-  },
-  StateDefinition,
-  {
-    [x: `agent=${string}`]: Partial<BaseGraphState>;
-    // eslint-disable-next-line @typescript-eslint/no-explicit-any
-    [x: `tools=${string}`]: any;
-    // eslint-disable-next-line @typescript-eslint/no-explicit-any
-    [x: `summarize=${string}`]: any;
-  }
+  AgentSubgraphState
 >;
 
 export type SystemRunnable =
@@ -214,21 +169,11 @@ export type SystemRunnable =
  * These are intentionally untyped to avoid coupling to library internals.
  */
 export type CompileOptions = {
-  // A checkpointer instance (e.g., MemorySaver, SQL saver)
-  // eslint-disable-next-line @typescript-eslint/no-explicit-any
-  checkpointer?: any;
+  checkpointer?: unknown;
   interruptBefore?: string[];
   interruptAfter?: string[];
 };
 
-export type EventStreamCallbackHandlerInput =
-  Parameters<CompiledWorkflow['streamEvents']>[2] extends Omit<
-    infer T,
-    'autoClose'
-  >
-    ? T
-    : never;
-
 export type StreamChunk =
   | (ChatGenerationChunk & {
       message: AIMessageChunk;
@@ -471,10 +416,12 @@ 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 +447,7 @@ 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/llm.ts

@@ -12,7 +12,7 @@ import type {
 } from '@langchain/openai';
 import type { GoogleGenerativeAIChatInput } from '@langchain/google-genai';
 import type { ChatVertexAIInput } from '@langchain/google-vertexai';
-import type { ChatDeepSeekCallOptions } from '@langchain/deepseek';
+import type { ChatDeepSeekInput } from '@langchain/deepseek';
 import type { ChatOpenRouterCallOptions } from '@/llm/openrouter';
 import type { ChatBedrockConverseInput } from '@langchain/aws';
 import type { ChatMistralAIInput } from '@langchain/mistralai';
@@ -70,7 +70,7 @@ export type AnthropicReasoning = {
 export type GoogleThinkingConfig = {
   thinkingBudget?: number;
   includeThoughts?: boolean;
-  thinkingLevel?: string;
+  thinkingLevel?: 'THINKING_LEVEL_UNSPECIFIED' | 'LOW' | 'MEDIUM' | 'HIGH';
 };
 export type OpenAIClientOptions = ChatOpenAIFields;
 export type AnthropicClientOptions = Omit<AnthropicInput, 'thinking'> & {
@@ -93,7 +93,7 @@ export type GoogleClientOptions = GoogleGenerativeAIChatInput & {
   customHeaders?: RequestOptions['customHeaders'];
   thinkingConfig?: GoogleThinkingConfig;
 };
-export type DeepSeekClientOptions = ChatDeepSeekCallOptions;
+export type DeepSeekClientOptions = Partial<ChatDeepSeekInput>;
 export type XAIClientOptions = ChatXAIInput;
 
 export type ClientOptions =

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;
 }
 

package/src/types/stream.ts

@@ -191,7 +191,7 @@ export interface ExtendedMessageContent {
   type?: string;
   text?: string;
   input?: string;
-  index?: number;
+  index?: string | number;
   id?: string;
   name?: string;
 }

package/src/types/tools.ts

@@ -113,6 +113,15 @@ export type FileRef = {
   path?: string;
   /** Session ID this file belongs to (for multi-session file tracking) */
   session_id?: string;
+  /**
+   * `true` when the codeapi sandbox echoed this entry as an unchanged
+   * passthrough of an input the caller already owns (skill files,
+   * downloaded inputs whose hash matched the baseline, inherited
+   * `.dirkeep` markers). The tool-result formatter renders these as
+   * "Available files" rather than "Generated files" so the LLM doesn't
+   * conflate infrastructure inputs with newly-produced outputs.
+   */
+  inherited?: true;
 };
 
 export type FileRefs = FileRef[];

package/src/utils/llmConfig.ts

@@ -56,9 +56,7 @@ export const llmConfigs: Record<string, t.LLMConfig | undefined> = {
     provider: Providers.OPENROUTER,
     streaming: true,
     streamUsage: true,
-    // model: 'anthropic/claude-sonnet-4',
-    // model: 'moonshotai/kimi-k2-thinking',
-    model: 'google/gemini-3-pro-preview',
+    model: process.env.OPENROUTER_MODEL ?? 'openai/gpt-4o-mini',
     apiKey: process.env.OPENROUTER_API_KEY,
     configuration: {
       baseURL: process.env.OPENROUTER_BASE_URL,
@@ -67,9 +65,6 @@ export const llmConfigs: Record<string, t.LLMConfig | undefined> = {
         'X-Title': 'LibreChat',
       },
     },
-    reasoning: {
-      max_tokens: 8000,
-    },
     modelKwargs: {
       max_tokens: 10000,
     },