feat: Add install dependencies tool (#291)

* feat: Add install dependencies tool

* cr

* hookup prompt

* frontend

* add prompt saying deps installed

apps/open-swe/src/constants.ts

@@ -0,0 +1,12 @@
+import { SNAPSHOT_NAME } from "@open-swe/shared/constants";
+import { CreateSandboxParams } from "@daytonaio/sdk";
+
+export const DEFAULT_SANDBOX_CREATE_PARAMS: CreateSandboxParams = {
+  resources: {
+    cpu: 2,
+    memory: 4,
+    disk: 5,
+  },
+  user: "daytona",
+  image: SNAPSHOT_NAME,
+};

apps/open-swe/src/graphs/programmer/nodes/generate-message/index.ts

@@ -16,11 +16,16 @@ import { createLogger, LogLevel } from "../../../../utils/logger.js";
 import { getCurrentPlanItem } from "../../../../utils/current-task.js";
 import { getMessageContentString } from "@open-swe/shared/messages";
 import { getActivePlanItems } from "@open-swe/shared/open-swe/tasks";
-import { SYSTEM_PROMPT } from "./prompt.js";
+import {
+  DEPENDENCIES_INSTALLED_PROMPT,
+  INSTALL_DEPENDENCIES_TOOL_PROMPT,
+  SYSTEM_PROMPT,
+} from "./prompt.js";
 import { getRepoAbsolutePath } from "@open-swe/shared/git";
 import { getMissingMessages } from "../../../../utils/github/issue-messages.js";
 import { getTaskPlanFromIssue } from "../../../../utils/github/issue-task.js";
 import { createRgTool } from "../../../../tools/rg.js";
+import { createInstallDependenciesTool } from "../../../../tools/install-dependencies.js";
 
 const logger = createLogger(LogLevel.INFO, "GenerateMessageNode");
 
@@ -50,7 +55,13 @@ const formatPrompt = (state: GraphState): string => {
       state.codebaseTree || "No codebase tree generated yet.",
     )
     .replaceAll("{CURRENT_WORKING_DIRECTORY}", repoDirectory)
-    .replaceAll("{CURRENT_TASK_NUMBER}", currentPlanItem.index.toString());
+    .replaceAll("{CURRENT_TASK_NUMBER}", currentPlanItem.index.toString())
+    .replaceAll(
+      "{INSTALL_DEPENDENCIES_TOOL_PROMPT}",
+      !state.dependenciesInstalled
+        ? INSTALL_DEPENDENCIES_TOOL_PROMPT
+        : DEPENDENCIES_INSTALLED_PROMPT,
+    );
 };
 
 export async function generateAction(
@@ -64,6 +75,10 @@ export async function generateAction(
     createApplyPatchTool(state),
     createRequestHumanHelpToolFields(),
     createUpdatePlanToolFields(),
+    // Only provide the dependencies installed tool if they're not already installed.
+    ...(state.dependenciesInstalled
+      ? []
+      : [createInstallDependenciesTool(state)]),
   ];
   const modelWithTools = model.bindTools(tools, {
     tool_choice: "auto",

apps/open-swe/src/graphs/programmer/nodes/generate-message/prompt.ts

@@ -1,3 +1,6 @@
+export const INSTALL_DEPENDENCIES_TOOL_PROMPT = `* Use \`install_dependencies\` to install dependencies (skip if installation fails). IMPORTANT: You should only call this tool if you're executing a task which REQUIRES installing dependencies. Keep in mind that not all tasks will require installing dependencies.`;
+export const DEPENDENCIES_INSTALLED_PROMPT = `* Dependencies have already been installed. *`;
+
 export const SYSTEM_PROMPT = `# Identity
 
 You are a terminal-based agentic coding assistant built by LangChain. You wrap LLM models to enable natural language interaction with local codebases. You are precise, safe, and helpful.
@@ -33,6 +36,7 @@ You are currently executing a specific task from a pre-generated plan. You have
 * Use \`apply_patch\` for file edits (accepts diffs and file paths)
 * Use \`shell\` with \`touch\` to create new files (not \`apply_patch\`)
 * Always use \`workdir\` parameter instead of \`cd\` when running commands via the \`shell\` tool
+{INSTALL_DEPENDENCIES_TOOL_PROMPT}
 
 ### Tool Usage Best Practices
 

apps/open-swe/src/graphs/programmer/nodes/open-pr.ts

@@ -159,7 +159,10 @@ export async function openPullRequest(
   return {
     messages: newMessages,
     internalMessages: newMessages,
-    // If the sandbox was successfully deleted, we can remove it from the state.
+    // If the sandbox was successfully deleted, we can remove it from the state & reset the dependencies installed flag.
-    ...(sandboxDeleted && { sandboxSessionId: undefined }),
+    ...(sandboxDeleted && {
+      sandboxSessionId: undefined,
+      dependenciesInstalled: false,
+    }),
   };
 }

apps/open-swe/src/graphs/programmer/nodes/take-action.ts

@@ -20,6 +20,8 @@ import { daytonaClient } from "../../../utils/sandbox.js";
 import { getCodebaseTree } from "../../../utils/tree.js";
 import { getRepoAbsolutePath } from "@open-swe/shared/git";
 import { shouldDiagnoseError } from "../utils/tool-message-error.js";
+import { createInstallDependenciesTool } from "../../../tools/install-dependencies.js";
+import { createRgTool } from "../../../tools/rg.js";
 
 const logger = createLogger(LogLevel.INFO, "TakeAction");
 
@@ -41,9 +43,13 @@ export async function takeAction(
 
   const applyPatchTool = createApplyPatchTool(state);
   const shellTool = createShellTool(state);
+  const rgTool = createRgTool(state);
+  const installDependenciesTool = createInstallDependenciesTool(state);
   const toolsMap = {
     [applyPatchTool.name]: applyPatchTool,
     [shellTool.name]: shellTool,
+    [rgTool.name]: rgTool,
+    [installDependenciesTool.name]: installDependenciesTool,
   };
 
   const toolCalls = lastMessage.tool_calls;
@@ -107,6 +113,13 @@ export async function takeAction(
 
   const toolCallResults = await Promise.all(toolCallResultsPromise);
 
+  let wereDependenciesInstalled: boolean | null = null;
+  toolCallResults.forEach((toolCallResult) => {
+    if (toolCallResult.name === installDependenciesTool.name) {
+      wereDependenciesInstalled = toolCallResult.status === "success";
+    }
+  });
+
   // Always check if there are changed files after running a tool.
   // If there are, commit them.
   const sandbox = await daytonaClient().get(state.sandboxSessionId);
@@ -142,6 +155,9 @@ export async function takeAction(
     internalMessages: toolCallResults,
     ...(branchName && { branchName }),
     codebaseTree,
+    ...(wereDependenciesInstalled !== null && {
+      dependenciesInstalled: wereDependenciesInstalled,
+    }),
   };
   return new Command({
     goto: shouldRouteDiagnoseNode ? "diagnose-error" : "progress-plan-step",

apps/open-swe/src/graphs/shared/initialize-sandbox.ts

@@ -11,16 +11,14 @@ import {
   pullLatestChanges,
 } from "../../utils/github/git.js";
 import { getCodebaseTree } from "../../utils/tree.js";
-import {
+import { DO_NOT_RENDER_ID_PREFIX } from "@open-swe/shared/constants";
-  DO_NOT_RENDER_ID_PREFIX,
-  SNAPSHOT_NAME,
-} from "@open-swe/shared/constants";
 import {
   CustomNodeEvent,
   INITIALIZE_NODE_ID,
 } from "@open-swe/shared/open-swe/custom-node-events";
 import { Sandbox } from "@daytonaio/sdk";
 import { AIMessage, BaseMessage } from "@langchain/core/messages";
+import { DEFAULT_SANDBOX_CREATE_PARAMS } from "../../constants.js";
 
 const logger = createLogger(LogLevel.INFO, "InitializeSandbox");
 
@@ -30,6 +28,7 @@ type InitializeSandboxState = {
   sandboxSessionId?: string;
   codebaseTree?: string;
   messages?: BaseMessage[];
+  dependenciesInstalled?: boolean;
 };
 
 export async function initializeSandbox(
@@ -207,7 +206,7 @@ export async function initializeSandbox(
   emitStepEvent(baseCreateSandboxAction, "pending");
   let sandbox: Sandbox;
   try {
-    sandbox = await daytonaClient().create({ image: SNAPSHOT_NAME });
+    sandbox = await daytonaClient().create(DEFAULT_SANDBOX_CREATE_PARAMS);
     emitStepEvent(baseCreateSandboxAction, "success");
   } catch {
     emitStepEvent(
@@ -332,5 +331,6 @@ export async function initializeSandbox(
     targetRepository,
     codebaseTree,
     messages: createEventsMessage(),
+    dependenciesInstalled: false,
   };
 }

apps/open-swe/src/tests/sandbox.int.test.ts

@@ -1,9 +1,10 @@
 /* eslint-disable no-console */
 import { test, expect } from "@jest/globals";
 import { daytonaClient } from "../utils/sandbox.js";
-import { SNAPSHOT_NAME } from "@open-swe/shared/constants";
+import { SANDBOX_ROOT_DIR } from "@open-swe/shared/constants";
+import { DEFAULT_SANDBOX_CREATE_PARAMS } from "../constants.js";
 
-test("Can execute rg commands", async () => {
+test.skip("Can execute rg commands", async () => {
   const githubToken = process.env.GITHUB_PAT;
   if (!githubToken) {
     throw new Error("GITHUB_PAT environment variable is not set");
@@ -12,10 +13,7 @@ test("Can execute rg commands", async () => {
   const client = daytonaClient();
 
   console.log("Setting up sandbox...");
-  const sandbox = await client.create({
+  const sandbox = await client.create(DEFAULT_SANDBOX_CREATE_PARAMS);
-    image: SNAPSHOT_NAME,
-    user: "daytona",
-  });
   console.log("Setup sandbox:", sandbox.id);
 
   const repoUrlWithToken = `https://x-access-token:${githubToken}@github.com/langchain-ai/open-swe.git`;
@@ -41,3 +39,51 @@ OPEN_SWE_X
 
   expect(testRes.exitCode).toBe(0);
 });
+
+const DEFAULT_ENV = {
+  // Prevents corepack from showing a y/n download prompt which causes the command to hang
+  COREPACK_ENABLE_DOWNLOAD_PROMPT: "0",
+};
+
+test("Installing dependencies", async () => {
+  const githubToken = process.env.GITHUB_PAT;
+  if (!githubToken) {
+    throw new Error("GITHUB_PAT environment variable is not set");
+  }
+
+  const client = daytonaClient();
+
+  console.log("Setting up sandbox...");
+  const sandbox = await client.create(DEFAULT_SANDBOX_CREATE_PARAMS);
+  console.log("Setup sandbox:", sandbox.id);
+
+  try {
+    const repoUrlWithToken = `https://x-access-token:${githubToken}@github.com/langchain-ai/open-swe.git`;
+    const cloneCommand = `git clone ${repoUrlWithToken}`;
+
+    console.log("Cloning repo...");
+    const cloneRes = await sandbox.process.executeCommand(
+      cloneCommand,
+      SANDBOX_ROOT_DIR,
+    );
+    expect(cloneRes.exitCode).toBe(0);
+
+    const installCommand = "yarn install";
+    const installRes = await sandbox.process.executeCommand(
+      installCommand,
+      `${SANDBOX_ROOT_DIR}/open-swe`,
+      DEFAULT_ENV,
+      120, // 120 seconds timeout
+    );
+    console.log(
+      `install res status: ${installRes.exitCode}\ninstall res output: ${installRes.result}`,
+    );
+    console.log(
+      `Install res exit code: ${installRes.exitCode}\nInstall res output: ${installRes.result}`,
+    );
+    expect(installRes.exitCode).toBe(0);
+  } finally {
+    await sandbox.delete();
+    console.log("Deleted sandbox:", sandbox.id);
+  }
+}, 300_000); // 5 minutes timeout

apps/open-swe/src/tools/install-dependencies.ts

@@ -0,0 +1,101 @@
+import { tool } from "@langchain/core/tools";
+import { Sandbox } from "@daytonaio/sdk";
+import { GraphState } from "@open-swe/shared/open-swe/types";
+import { getCurrentTaskInput } from "@langchain/langgraph";
+import { getSandboxErrorFields } from "../utils/sandbox-error-fields.js";
+import { createLogger, LogLevel } from "../utils/logger.js";
+import { daytonaClient } from "../utils/sandbox.js";
+import { TIMEOUT_SEC } from "@open-swe/shared/constants";
+import { createInstallDependenciesToolFields } from "@open-swe/shared/open-swe/tools";
+import { getRepoAbsolutePath } from "@open-swe/shared/git";
+
+const logger = createLogger(LogLevel.INFO, "InstallDependenciesTool");
+
+const DEFAULT_ENV = {
+  // Prevents corepack from showing a y/n download prompt which causes the command to hang
+  COREPACK_ENABLE_DOWNLOAD_PROMPT: "0",
+};
+
+export function createInstallDependenciesTool(
+  state: Pick<GraphState, "sandboxSessionId" | "targetRepository">,
+) {
+  const installDependenciesTool = tool(
+    async (input): Promise<{ result: string; status: "success" | "error" }> => {
+      let sandbox: Sandbox | undefined;
+      try {
+        const state = getCurrentTaskInput<GraphState>();
+        const { sandboxSessionId } = state;
+        if (!sandboxSessionId) {
+          logger.error(
+            "FAILED TO INSTALL DEPENDENCIES: No sandbox session ID provided",
+            {
+              input,
+            },
+          );
+          throw new Error(
+            "FAILED TO INSTALL DEPENDENCIES: No sandbox session ID provided",
+          );
+        }
+
+        const repoRoot = getRepoAbsolutePath(state.targetRepository);
+
+        sandbox = await daytonaClient().get(sandboxSessionId);
+        const command = input.command.join(" ");
+        const workdir = input.workdir || repoRoot;
+        logger.info("Running install dependencies command", {
+          command,
+          workdir,
+        });
+        const response = await sandbox.process.executeCommand(
+          command,
+          workdir,
+          DEFAULT_ENV,
+          TIMEOUT_SEC * 2.5, // add a 2.5 min timeout
+        );
+
+        if (response.exitCode !== 0) {
+          logger.error("Failed to install dependencies", {
+            error: response.result,
+            error_result: response,
+            input,
+          });
+          throw new Error(
+            `Command failed. Exit code: ${response.exitCode}\nResult: ${response.result}\nStdout:\n${response.artifacts?.stdout}`,
+          );
+        }
+
+        return {
+          result: response.result,
+          status: "success",
+        };
+      } catch (e) {
+        const errorFields = getSandboxErrorFields(e);
+        if (errorFields) {
+          logger.error("Failed to install dependencies", {
+            input,
+            error: errorFields,
+          });
+          throw new Error(
+            `Command failed. Exit code: ${errorFields.exitCode}\nError: ${errorFields.result ?? errorFields.artifacts?.stdout}`,
+          );
+        }
+
+        logger.error(
+          "Failed to install dependencies: " +
+            (e instanceof Error ? e.message : "Unknown error"),
+          {
+            error: e,
+            input,
+          },
+        );
+        throw new Error(
+          "FAILED TO INSTALL DEPENDENCIES: " +
+            (e instanceof Error ? e.message : "Unknown error"),
+        );
+      }
+    },
+    createInstallDependenciesToolFields(state.targetRepository),
+  );
+
+  return installDependenciesTool;
+}

apps/open-swe/src/utils/truncate-outputs.ts

@@ -6,18 +6,17 @@ export function truncateOutput(
   output: string,
   options?: {
     /**
-     * @default 10000
+     * @default 2500
      */
     numStartCharacters?: number;
 
     /**
-     * @default 10000
+     * @default 2500
      */
     numEndCharacters?: number;
   },
 ) {
-  const { numStartCharacters = 10000, numEndCharacters = 10000 } =
+  const { numStartCharacters = 2500, numEndCharacters = 2500 } = options ?? {};
-    options ?? {};
 
   if (numStartCharacters < 0 || numEndCharacters < 0) {
     throw new Error("numStartCharacters and numEndCharacters must be >= 0");

apps/web/src/components/gen-ui/action-step.tsx

@@ -11,10 +11,13 @@ import {
   ChevronUp,
   MessageSquare,
   FileText,
+  CloudDownload,
+  Search,
 } from "lucide-react";
 import {
   createApplyPatchToolFields,
   createShellToolFields,
+  createInstallDependenciesToolFields,
   formatRgCommand,
   RipgrepCommand,
 } from "@open-swe/shared/open-swe/tools";
@@ -26,6 +29,10 @@ const shellTool = createShellToolFields(dummyRepo);
 type ShellToolArgs = z.infer<typeof shellTool.schema>;
 const applyPatchTool = createApplyPatchToolFields(dummyRepo);
 type ApplyPatchToolArgs = z.infer<typeof applyPatchTool.schema>;
+const installDependenciesTool = createInstallDependenciesToolFields(dummyRepo);
+type InstallDependenciesToolArgs = z.infer<
+  typeof installDependenciesTool.schema
+>;
 
 // Common props for all action types
 type BaseActionProps = {
@@ -59,11 +66,19 @@ type RgActionProps = BaseActionProps &
     errorCode?: number;
   };
 
+type InstallDependenciesActionProps = BaseActionProps &
+  Partial<InstallDependenciesToolArgs> & {
+    actionType: "install_dependencies";
+    output?: string;
+    errorCode?: number;
+  };
+
 export type ActionItemProps =
   | (BaseActionProps & { status: "loading" })
   | ShellActionProps
   | PatchActionProps
-  | RgActionProps;
+  | RgActionProps
+  | InstallDependenciesActionProps;
 
 export type ActionStepProps = {
   actions: ActionItemProps[];
@@ -109,6 +124,8 @@ function ActionItem(props: ActionItemProps) {
         return props.success ? "Patch applied" : "Patch failed";
       } else if (props.actionType === "rg") {
         return props.success ? "Search completed" : "Search failed";
+      } else if (props.actionType === "install_dependencies") {
+        return props.success ? "Dependencies installed" : "Installation failed";
       }
     }
 
@@ -119,7 +136,11 @@ function ActionItem(props: ActionItemProps) {
   const shouldShowToggle = () => {
     if (props.status !== "done") return false;
 
-    if (props.actionType === "shell" || props.actionType === "rg") {
+    if (
+      props.actionType === "shell" ||
+      props.actionType === "rg" ||
+      props.actionType === "install_dependencies"
+    ) {
       return !!props.output;
     } else if (props.actionType === "apply-patch") {
       return !!props.diff;
@@ -135,11 +156,15 @@ function ActionItem(props: ActionItemProps) {
       return <Loader2 className="text-muted-foreground mr-2 size-3.5" />;
     }
 
-    return props.actionType === "shell" ? (
+    if (props.actionType === "install_dependencies") {
-      <Terminal className="text-muted-foreground mr-2 size-3.5" />
+      return <CloudDownload className="text-muted-foreground mr-2 size-3.5" />;
-    ) : (
+    } else if (props.actionType === "apply-patch") {
-      <FileCode className="text-muted-foreground mr-2 size-3.5" />
+      return <FileCode className="text-muted-foreground mr-2 size-3.5" />;
-    );
+    } else if (props.actionType === "rg") {
+      return <Search className="text-muted-foreground mr-2 size-3.5" />;
+    } else {
+      return <Terminal className="text-muted-foreground mr-2 size-3.5" />;
+    }
   };
 
   // Render the header content based on action type
@@ -152,7 +177,10 @@ function ActionItem(props: ActionItemProps) {
       );
     }
 
-    if (props.actionType === "shell") {
+    if (
+      props.actionType === "shell" ||
+      props.actionType === "install_dependencies"
+    ) {
       let commandStr = "";
       if (props.command) {
         if (Array.isArray(props.command)) {
@@ -215,7 +243,9 @@ function ActionItem(props: ActionItemProps) {
     if (!expanded) return null;
 
     if (
-      (props.actionType === "shell" || props.actionType === "rg") &&
+      (props.actionType === "shell" ||
+        props.actionType === "rg" ||
+        props.actionType === "install_dependencies") &&
       props.output
     ) {
       return (

apps/web/src/components/thread/messages/ai.tsx

@@ -26,6 +26,7 @@ import {
   createShellToolFields,
   createSetTaskStatusToolFields,
   createRgToolFields,
+  createInstallDependenciesToolFields,
 } from "@open-swe/shared/open-swe/tools";
 import { z } from "zod";
 import { isAIMessageSDK, isToolMessageSDK } from "@/lib/langchain-messages";
@@ -41,6 +42,10 @@ const setTaskStatusTool = createSetTaskStatusToolFields();
 type SetTaskStatusToolArgs = z.infer<typeof setTaskStatusTool.schema>;
 const rgTool = createRgToolFields(dummyRepo);
 type RgToolArgs = z.infer<typeof rgTool.schema>;
+const installDependenciesTool = createInstallDependenciesToolFields(dummyRepo);
+type InstallDependenciesToolArgs = z.infer<
+  typeof installDependenciesTool.schema
+>;
 
 function CustomComponent({
   message,
@@ -148,6 +153,17 @@ export function mapToolMessageToActionStepProps(
       output: getContentString(message.content),
       reasoningText,
     };
+  } else if (toolCall?.name === installDependenciesTool.name) {
+    const args = toolCall.args as InstallDependenciesToolArgs;
+    return {
+      actionType: "install_dependencies",
+      status,
+      success,
+      command: args.command || "",
+      workdir: args.workdir || "",
+      output: getContentString(message.content),
+      reasoningText,
+    };
   }
   return {
     status: "loading",
@@ -210,7 +226,8 @@ export function AssistantMessage({
         (tc) =>
           tc.name === shellTool.name ||
           tc.name === applyPatchTool.name ||
-          tc.name === rgTool.name,
+          tc.name === rgTool.name ||
+          tc.name === installDependenciesTool.name,
       )
     : [];
 
@@ -248,6 +265,8 @@ export function AssistantMessage({
 
       const isShellTool = toolCall.name === shellTool.name;
       const isRgTool = toolCall.name === rgTool.name;
+      const isInstallDependenciesTool =
+        toolCall.name === installDependenciesTool.name;
 
       if (correspondingToolResult) {
         // If we have a tool result, map it to action props
@@ -261,6 +280,15 @@ export function AssistantMessage({
           paths: args?.paths || [],
           output: "",
         } as ActionItemProps;
+      } else if (isInstallDependenciesTool) {
+        const args = toolCall.args as InstallDependenciesToolArgs;
+        return {
+          actionType: "install_dependencies",
+          status: "generating",
+          command: args?.command || "",
+          workdir: args?.workdir || "",
+          output: "",
+        } as ActionItemProps;
       } else {
         if (isShellTool) {
           const args = toolCall.args as ShellToolArgs;

packages/shared/src/open-swe/tools.ts

@@ -179,3 +179,28 @@ export function createSetTaskStatusToolFields() {
 
   return setTaskStatusTool;
 }
+
+export function createInstallDependenciesToolFields(
+  targetRepository: TargetRepository,
+) {
+  const repoRoot = getRepoAbsolutePath(targetRepository);
+
+  const installDependenciesToolSchema = z.object({
+    command: z
+      .array(z.string())
+      .describe("The command to run to install dependencies."),
+    workdir: z
+      .string()
+      .default(repoRoot)
+      .describe(
+        `The working directory to run the command in. The default working directory this command will be executed in is the root of the repository: \`${repoRoot}\`. If you want to execute this install command inside a different location, pass a path to this field.`,
+      ),
+  });
+
+  return {
+    name: "install_dependencies",
+    description:
+      "Installs dependencies for the repository. You should only call this tool if you need to install dependencies for a specific task. Ensure you only call this tool after gathering context on how to install dependencies, such as the package manager, proper install command, etc.",
+    schema: installDependenciesToolSchema,
+  };
+}

packages/shared/src/open-swe/types.ts

@@ -205,6 +205,16 @@ export const GraphAnnotation = MessagesZodState.extend({
       fn: (_state, update) => update,
     },
   }),
+  /**
+   * Whether or not the dependencies have been installed already in the sandbox.
+   */
+  dependenciesInstalled: withLangGraph(z.custom<boolean>(), {
+    reducer: {
+      schema: z.custom<boolean>(),
+      fn: (_state, update) => update,
+    },
+    default: () => false,
+  }),
 
   // ---NOT USED---
   ui: z