[open-swe] feat: Add find_instances_of tool for comprehensive string search (#350)

* Apply patch

* Apply patch

* Apply patch

* Apply patch

* Apply patch

* Apply patch

* Apply patch

* Apply patch

* Apply patch

* Apply patch

* Apply patch

* Apply patch

* Apply patch

* Apply patch

* Apply patch

* cr

* cr

* fix

* cr

---------

Co-authored-by: open-swe-dev[bot] <open-swe-dev@users.noreply.github.com>
Co-authored-by: bracesproul <braceasproul@gmail.com>

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

@@ -22,6 +22,7 @@ import { getTaskPlanFromIssue } from "../../../../utils/github/issue-task.js";
 import { createRgTool } from "../../../../tools/rg.js";
 import { formatCustomRulesPrompt } from "../../../../utils/custom-rules.js";
 import { createPlannerNotesTool } from "../../../../tools/planner-notes.js";
+import { createFindInstancesOfTool } from "../../../../tools/find-instances-of.js";
 import { getMcpTools } from "../../../../utils/mcp-client.js";
 
 const logger = createLogger(LogLevel.INFO, "GeneratePlanningMessageNode");
@@ -56,6 +57,7 @@ export async function generateAction(
   const tools = [
     createRgTool(state),
     createShellTool(state),
+    createFindInstancesOfTool(state),
     createPlannerNotesTool(),
     createGetURLContentTool(),
     ...mcpTools,

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

@@ -19,6 +19,7 @@ import {
   getChangedFilesStatus,
   stashAndClearChanges,
 } from "../../../utils/github/git.js";
+import { createFindInstancesOfTool } from "../../../tools/find-instances-of.js";
 import { getRepoAbsolutePath } from "@open-swe/shared/git";
 import { daytonaClient } from "../../../utils/sandbox.js";
 import { createPlannerNotesTool } from "../../../tools/planner-notes.js";
@@ -40,8 +41,8 @@ export async function takeActions(
   const shellTool = createShellTool(state);
   const rgTool = createRgTool(state);
   const plannerNotesTool = createPlannerNotesTool();
+  const findInstancesOfTool = createFindInstancesOfTool(state);
   const getURLContentTool = createGetURLContentTool();
-
   const mcpTools = await getMcpTools(config);
 
   const allTools = [
@@ -49,6 +50,7 @@ export async function takeActions(
     rgTool,
     plannerNotesTool,
     getURLContentTool,
+    findInstancesOfTool,
     ...mcpTools,
   ];
   const toolsMap = Object.fromEntries(

apps/open-swe/src/tools/find-instances-of.ts

@@ -0,0 +1,154 @@
+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 { createFindInstancesOfToolFields } from "@open-swe/shared/open-swe/tools";
+import { getRepoAbsolutePath } from "@open-swe/shared/git";
+import { z } from "zod";
+import { wrapScript } from "../utils/wrap-script.js";
+
+const logger = createLogger(LogLevel.INFO, "FindInstancesOfTool");
+
+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 createFindInstancesOfTool(
+  state: Pick<GraphState, "sandboxSessionId" | "targetRepository">,
+) {
+  const findInstancesOfFields = createFindInstancesOfToolFields(
+    state.targetRepository,
+  );
+  const formatFindInstancesOfCommand = (
+    input: z.infer<typeof findInstancesOfFields.schema>,
+  ): string[] => {
+    const args = ["rg"];
+
+    // Always include these flags for consistent output
+    args.push("--color", "never", "--line-number", "--heading");
+
+    // Add context lines (3 above and 3 below)
+    args.push("-A", "3", "-B", "3");
+
+    // Handle case sensitivity
+    if (!input.case_sensitive) {
+      args.push("-i");
+    }
+
+    // Handle word matching
+    if (input.match_word) {
+      args.push("--word-regexp");
+    }
+
+    // Handle file inclusion/exclusion patterns
+    if (input.exclude_files) {
+      args.push("-g", `!${input.exclude_files}`);
+    }
+
+    if (input.include_files) {
+      args.push("-g", input.include_files);
+    }
+
+    // For literal string matching (not regex)
+    args.push("--fixed-strings");
+
+    // Add the search query as the last argument (ensure it's properly quoted)
+    const formattedQuery = `'${input.query.replace(/^'|'$/g, "")}'`;
+    args.push(formattedQuery);
+
+    return args;
+  };
+
+  const findInstancesOfTool = tool(
+    async (
+      input: z.infer<typeof findInstancesOfFields.schema>,
+    ): Promise<{ result: string; status: "success" | "error" }> => {
+      let sandbox: Sandbox | undefined;
+      try {
+        const state = getCurrentTaskInput<GraphState>();
+        const { sandboxSessionId } = state;
+        if (!sandboxSessionId) {
+          logger.error(
+            "FAILED TO RUN COMMAND: No sandbox session ID provided",
+            {
+              input,
+            },
+          );
+          throw new Error(
+            "FAILED TO RUN COMMAND: No sandbox session ID provided",
+          );
+        }
+
+        const repoRoot = getRepoAbsolutePath(state.targetRepository);
+
+        sandbox = await daytonaClient().get(sandboxSessionId);
+        const command = formatFindInstancesOfCommand(input);
+        logger.info("Running find_instances_of command", {
+          command: command.join(" "),
+          repoRoot,
+        });
+
+        const response = await sandbox.process.executeCommand(
+          wrapScript(command.join(" ")),
+          repoRoot,
+          DEFAULT_ENV,
+          TIMEOUT_SEC,
+        );
+
+        let successResult = response.result;
+
+        if (
+          response.exitCode === 1 ||
+          (response.exitCode === 127 && response.result.startsWith("sh: 1: "))
+        ) {
+          logger.info("Exit code 1. no results found", {
+            ...response,
+          });
+          successResult = `Exit code 1. No results found.\n\n${response.result}`;
+        } else if (response.exitCode > 1) {
+          logger.error("Failed to run find_instances_of command", {
+            error: response.result,
+            error_result: response,
+            input,
+          });
+          throw new Error("Command failed. Exit code: " + response.exitCode);
+        }
+
+        return {
+          result: successResult,
+          status: "success",
+        };
+      } catch (e) {
+        const errorFields = getSandboxErrorFields(e);
+        if (errorFields) {
+          logger.error("Failed to run find_instances_of command", {
+            input,
+            error: errorFields,
+          });
+          throw new Error("Command failed. Exit code: " + errorFields.exitCode);
+        }
+
+        logger.error(
+          "Failed to run find_instances_of command: " +
+            (e instanceof Error ? e.message : "Unknown error"),
+          {
+            error: e,
+            input,
+          },
+        );
+        throw new Error(
+          "FAILED TO RUN FIND_INSTANCES_OF COMMAND: " +
+            (e instanceof Error ? e.message : "Unknown error"),
+        );
+      }
+    },
+    findInstancesOfFields,
+  );
+
+  return findInstancesOfTool;
+}

apps/open-swe/src/tools/rg.ts

@@ -11,13 +11,7 @@ import {
   formatRgCommand,
 } from "@open-swe/shared/open-swe/tools";
 import { getRepoAbsolutePath } from "@open-swe/shared/git";
-
+import { wrapScript } from "../utils/wrap-script.js";
-const wrapScript = (command: string): string => {
-  return `script --return --quiet -c "$(cat <<'OPEN_SWE_X'
-${command}
-OPEN_SWE_X
-)" /dev/null`;
-};
 
 const logger = createLogger(LogLevel.INFO, "RgTool");
 

apps/open-swe/src/utils/wrap-script.ts

@@ -0,0 +1,6 @@
+export function wrapScript(command: string): string {
+  return `script --return --quiet -c "$(cat <<'OPEN_SWE_X'
+${command}
+OPEN_SWE_X
+)" /dev/null`;
+}

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

@@ -1,6 +1,6 @@
 "use client";
 
-import { useState } from "react";
+import { JSX, useState } from "react";
 import {
   Terminal,
   FileCode,
@@ -14,6 +14,7 @@ import {
   CloudDownload,
   Search,
   Globe,
+  Hash,
 } from "lucide-react";
 import {
   createApplyPatchToolFields,
@@ -21,10 +22,18 @@ import {
   createInstallDependenciesToolFields,
   createTakePlannerNotesFields,
   createGetURLContentToolFields,
+  createFindInstancesOfToolFields,
   formatRgCommand,
   RipgrepCommand,
 } from "@open-swe/shared/open-swe/tools";
 import { z } from "zod";
+import {
+  Tooltip,
+  TooltipContent,
+  TooltipProvider,
+  TooltipTrigger,
+} from "../ui/tooltip";
+import { cn } from "@/lib/utils";
 
 // Used only for Zod type inference.
 const dummyRepo = { owner: "dummy", repo: "dummy" };
@@ -40,6 +49,8 @@ const plannerNotesTool = createTakePlannerNotesFields();
 type PlannerNotesToolArgs = z.infer<typeof plannerNotesTool.schema>;
 const getURLContentTool = createGetURLContentToolFields();
 type GetURLContentToolArgs = z.infer<typeof getURLContentTool.schema>;
+const findInstancesOfTool = createFindInstancesOfToolFields(dummyRepo);
+type FindInstancesOfToolArgs = z.infer<typeof findInstancesOfTool.schema>;
 
 // Common props for all action types
 type BaseActionProps = {
@@ -91,6 +102,13 @@ type GetURLContentActionProps = BaseActionProps &
     output?: string;
   };
 
+type FindInstancesOfActionProps = BaseActionProps &
+  Partial<FindInstancesOfToolArgs> & {
+    actionType: "find_instances_of";
+    output?: string;
+    errorCode?: number;
+  };
+
 export type ActionItemProps =
   | (BaseActionProps & { status: "loading" })
   | ShellActionProps
@@ -98,7 +116,8 @@ export type ActionItemProps =
   | RgActionProps
   | InstallDependenciesActionProps
   | PlannerNotesActionProps
-  | GetURLContentActionProps;
+  | GetURLContentActionProps
+  | FindInstancesOfActionProps;
 
 export type ActionStepProps = {
   actions: ActionItemProps[];
@@ -113,8 +132,88 @@ const ACTION_GENERATING_TEXT_MAP = {
   [installDependenciesTool.name]: "Installing dependencies...",
   [plannerNotesTool.name]: "Saving notes...",
   [getURLContentTool.name]: "Fetching URL content...",
+  [findInstancesOfTool.name]: "Finding instances...",
 };
 
+function ToolIconWithTooltip({
+  toolNamePretty,
+  icon,
+}: {
+  toolNamePretty: string;
+  icon: JSX.Element;
+}) {
+  return (
+    <TooltipProvider>
+      <Tooltip>
+        <TooltipTrigger asChild>{icon}</TooltipTrigger>
+        <TooltipContent>{toolNamePretty}</TooltipContent>
+      </Tooltip>
+    </TooltipProvider>
+  );
+}
+
+function MatchCaseIcon({ matchCase }: { matchCase: boolean }) {
+  return (
+    <TooltipProvider>
+      <Tooltip>
+        <TooltipTrigger
+          className={cn(
+            "rounded-sm border border-gray-300 px-1 py-[2px] text-xs dark:border-gray-600",
+            matchCase
+              ? "border-blue-500 bg-blue-500/80 text-white"
+              : "bg-gray-100 text-gray-800 dark:bg-gray-800 dark:text-gray-200",
+          )}
+        >
+          <p className="font-mono">Aa</p>
+        </TooltipTrigger>
+        <TooltipContent>Match case {matchCase ? "on" : "off"}</TooltipContent>
+      </Tooltip>
+    </TooltipProvider>
+  );
+}
+
+function MatchWholeWordIcon({ matchWholeWord }: { matchWholeWord: boolean }) {
+  return (
+    <TooltipProvider>
+      <Tooltip>
+        <TooltipTrigger
+          className={cn(
+            "rounded-sm border border-gray-300 px-1 py-[2px] text-xs dark:border-gray-600",
+            matchWholeWord
+              ? "border-blue-500 bg-blue-500/80 text-white"
+              : "bg-gray-100 text-gray-800 dark:bg-gray-800 dark:text-gray-200",
+          )}
+        >
+          <span className="relative inline-block px-0.5">
+            <p className="relative z-10 m-0 text-center font-mono">ab</p>
+            <div
+              className={cn(
+                "absolute bottom-0 left-0 h-1/5 w-[1px]",
+                matchWholeWord ? "bg-white" : "bg-gray-600 dark:bg-gray-300",
+              )}
+            ></div>
+            <div
+              className={cn(
+                "absolute right-0 bottom-0 left-0 h-[1px] w-full",
+                matchWholeWord ? "bg-white" : "bg-gray-600 dark:bg-gray-300",
+              )}
+            ></div>
+            <div
+              className={cn(
+                "absolute right-0 bottom-0 h-1/5 w-[1px]",
+                matchWholeWord ? "bg-white" : "bg-gray-600 dark:bg-gray-300",
+              )}
+            ></div>
+          </span>
+        </TooltipTrigger>
+        <TooltipContent>
+          Match whole word {matchWholeWord ? "on" : "off"}
+        </TooltipContent>
+      </Tooltip>
+    </TooltipProvider>
+  );
+}
+
 function ActionItem(props: ActionItemProps) {
   const [expanded, setExpanded] = useState(false);
 
@@ -159,6 +258,8 @@ function ActionItem(props: ActionItemProps) {
         return props.success
           ? "URL content fetched"
           : "Failed to fetch URL content";
+      } else if (props.actionType === "find_instances_of") {
+        return props.success ? "Search completed" : "Search failed";
       }
     }
 
@@ -173,7 +274,8 @@ function ActionItem(props: ActionItemProps) {
       props.actionType === "shell" ||
       props.actionType === "rg" ||
       props.actionType === "install_dependencies" ||
-      props.actionType === "get_url_content"
+      props.actionType === "get_url_content" ||
+      props.actionType === "find_instances_of"
     ) {
       return !!props.output;
     } else if (props.actionType === "apply-patch") {
@@ -187,23 +289,66 @@ function ActionItem(props: ActionItemProps) {
 
   // Render the header icon based on action type
   const renderHeaderIcon = () => {
+    const defaultIconStyling = "text-muted-foreground mr-2 size-3.5";
     if (props.status === "loading" || !("actionType" in props)) {
       // In loading state, we don't know the type yet, use a generic icon
-      return <Loader2 className="text-muted-foreground mr-2 size-3.5" />;
+      return (
+        <ToolIconWithTooltip
+          toolNamePretty="Loading"
+          icon={<Loader2 className={cn(defaultIconStyling)} />}
+        />
+      );
     }
 
     if (props.actionType === "planner_notes") {
-      return <FileText className="text-muted-foreground mr-2 size-3.5" />;
+      return (
+        <ToolIconWithTooltip
+          toolNamePretty="Planner Notes"
+          icon={<FileText className={cn(defaultIconStyling)} />}
+        />
+      );
     } else if (props.actionType === "install_dependencies") {
-      return <CloudDownload className="text-muted-foreground mr-2 size-3.5" />;
+      return (
+        <ToolIconWithTooltip
+          toolNamePretty="Install Dependencies"
+          icon={<CloudDownload className={cn(defaultIconStyling)} />}
+        />
+      );
     } else if (props.actionType === "apply-patch") {
-      return <FileCode className="text-muted-foreground mr-2 size-3.5" />;
+      return (
+        <ToolIconWithTooltip
+          toolNamePretty="Apply Patch"
+          icon={<FileCode className={cn(defaultIconStyling)} />}
+        />
+      );
     } else if (props.actionType === "rg") {
-      return <Search className="text-muted-foreground mr-2 size-3.5" />;
+      return (
+        <ToolIconWithTooltip
+          toolNamePretty="Ripgrep (rg)"
+          icon={<Search className={cn(defaultIconStyling)} />}
+        />
+      );
     } else if (props.actionType === "get_url_content") {
-      return <Globe className="text-muted-foreground mr-2 size-3.5" />;
+      return (
+        <ToolIconWithTooltip
+          toolNamePretty="Get URL Contents"
+          icon={<Globe className={cn(defaultIconStyling)} />}
+        />
+      );
+    } else if (props.actionType === "find_instances_of") {
+      return (
+        <ToolIconWithTooltip
+          toolNamePretty="Find Instances of"
+          icon={<Hash className={cn(defaultIconStyling)} />}
+        />
+      );
     } else {
-      return <Terminal className="text-muted-foreground mr-2 size-3.5" />;
+      return (
+        <ToolIconWithTooltip
+          toolNamePretty="Tool Call"
+          icon={<Terminal className={cn(defaultIconStyling)} />}
+        />
+      );
     }
   };
 
@@ -237,6 +382,32 @@ function ActionItem(props: ActionItemProps) {
       );
     }
 
+    if (props.actionType === "find_instances_of") {
+      return (
+        <div className="flex-1">
+          <div className="flex items-center gap-2">
+            <code className="text-foreground/80 text-xs font-normal">
+              {props.query}
+            </code>
+            <div className="bg-border h-4 w-[1px] dark:bg-white"></div>
+            <MatchCaseIcon matchCase={!!props.case_sensitive} />
+            <MatchWholeWordIcon matchWholeWord={!!props.match_word} />
+          </div>
+          {(props.include_files || props.exclude_files) && (
+            <div className="text-muted-foreground mt-0.5 text-xs font-normal">
+              {props.include_files && (
+                <span>Include: {props.include_files}</span>
+              )}
+              {props.include_files && props.exclude_files && <span> | </span>}
+              {props.exclude_files && (
+                <span>Exclude: {props.exclude_files}</span>
+              )}
+            </div>
+          )}
+        </div>
+      );
+    }
+
     if (
       props.actionType === "shell" ||
       props.actionType === "install_dependencies"
@@ -305,6 +476,7 @@ function ActionItem(props: ActionItemProps) {
     if (
       (props.actionType === "shell" ||
         props.actionType === "rg" ||
+        props.actionType === "find_instances_of" ||
         props.actionType === "install_dependencies") &&
       props.output
     ) {

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

@@ -33,6 +33,7 @@ import {
   createTakePlannerNotesFields,
   createDiagnoseErrorToolFields,
   createGetURLContentToolFields,
+  createFindInstancesOfToolFields,
 } from "@open-swe/shared/open-swe/tools";
 import { z } from "zod";
 import { isAIMessageSDK, isToolMessageSDK } from "@/lib/langchain-messages";
@@ -63,6 +64,9 @@ type DiagnoseErrorToolArgs = z.infer<typeof diagnoseErrorTool.schema>;
 const getURLContentTool = createGetURLContentToolFields();
 type GetURLContentToolArgs = z.infer<typeof getURLContentTool.schema>;
 
+const findInstancesOfTool = createFindInstancesOfToolFields(dummyRepo);
+type FindInstancesOfToolArgs = z.infer<typeof findInstancesOfTool.schema>;
+
 function CustomComponent({
   message,
   thread,
@@ -199,6 +203,24 @@ export function mapToolMessageToActionStepProps(
       output: getContentString(message.content),
       reasoningText,
     };
+  } else if (toolCall?.name === findInstancesOfTool.name) {
+    const args = toolCall.args as FindInstancesOfToolArgs;
+    // case_sensitive and match_word both default to true.
+    const caseSensitive =
+      args.case_sensitive === undefined ? true : args.case_sensitive;
+    const matchWord = args.match_word === undefined ? true : args.match_word;
+    return {
+      actionType: "find_instances_of",
+      status,
+      success,
+      query: args.query || "",
+      case_sensitive: caseSensitive,
+      match_word: matchWord,
+      include_files: args.include_files,
+      exclude_files: args.exclude_files,
+      output: getContentString(message.content),
+      reasoningText,
+    };
   }
   return {
     status: "loading",
@@ -264,7 +286,8 @@ export function AssistantMessage({
           tc.name === rgTool.name ||
           tc.name === installDependenciesTool.name ||
           tc.name === plannerNotesTool.name ||
-          tc.name === getURLContentTool.name,
+          tc.name === getURLContentTool.name ||
+          tc.name === findInstancesOfTool.name,
       )
     : [];
 
@@ -420,6 +443,23 @@ export function AssistantMessage({
           url: args?.url || "",
           output: "",
         } as ActionItemProps;
+      } else if (toolCall.name === findInstancesOfTool.name) {
+        const args = toolCall.args as FindInstancesOfToolArgs;
+        // case_sensitive and match_word both default to true.
+        const caseSensitive =
+          args.case_sensitive === undefined ? true : args.case_sensitive;
+        const matchWord =
+          args.match_word === undefined ? true : args.match_word;
+        return {
+          actionType: "find_instances_of",
+          status: "generating",
+          query: args?.query || "",
+          case_sensitive: caseSensitive,
+          match_word: matchWord,
+          include_files: args?.include_files,
+          exclude_files: args?.exclude_files,
+          output: "",
+        } as ActionItemProps;
       } else {
         if (isShellTool) {
           const args = toolCall.args as ShellToolArgs;

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

@@ -180,6 +180,51 @@ export function formatRgCommand(cmd: RipgrepCommand): string[] {
   return args;
 }
 
+export function createFindInstancesOfToolFields(
+  targetRepository: TargetRepository,
+) {
+  const repoRoot = getRepoAbsolutePath(targetRepository);
+  const findInstancesOfSchema = z.object({
+    query: z
+      .string()
+      .describe(
+        "The query/keyword to search for. This should be a literal string, not a regex.",
+      ),
+
+    case_sensitive: z
+      .boolean()
+      .optional()
+      .default(true)
+      .describe(
+        "Whether or not to make the query search case sensitive. Defaults to true",
+      ),
+
+    match_word: z
+      .boolean()
+      .optional()
+      .default(true)
+      .describe(
+        "Whether or not to only show results which match the exact keyword. Defaults to true",
+      ),
+
+    exclude_files: z
+      .string()
+      .optional()
+      .describe("Glob pattern of files to exclude"),
+
+    include_files: z
+      .string()
+      .optional()
+      .describe("Glob pattern of files to include"),
+  });
+
+  return {
+    name: "find_instances_of",
+    schema: findInstancesOfSchema,
+    description: `Find all instances of a string in the repository. Returns results with 3 lines of context above and below each match, absolute file paths, and total result count. The working directory this command will be executed in is \`${repoRoot}\`.`,
+  };
+}
+
 export function createSetTaskStatusToolFields() {
   const setTaskStatusToolSchema = z.object({
     reasoning: z