fix: Handle apply patch errors better (#27)

apps/open-swe/scripts/run-e2e.ts

@@ -41,7 +41,7 @@ Once you're done, ensure you've documented the development process in the readme
     },
     config: {
       configurable,
-      recursion_limit: 200,
+      recursion_limit: 400,
     },
     ifNotExists: "create",
     streamSubgraphs: true,
@@ -85,7 +85,7 @@ async function resumeGraph(threadId: string) {
     },
     config: {
       configurable,
-      recursion_limit: 200,
+      recursion_limit: 400,
     },
     streamSubgraphs: true,
     streamMode: "updates",

apps/open-swe/scripts/run-from-plan.ts

@@ -19,8 +19,15 @@ async function runFromPlan() {
     messages: [
       {
         role: "user",
-        content:
+        content: `The project is structured as a monorepo, with some apps located inside the /apps directory. In this directory, there is an /auth directory. This directory only contains the scaffolding for a new app in the monorepo, but is not yet implemented. Please take the following plan/task description and implement it in the /auth directory:
-          "This repo contains the react/next.js code for my persona/portfolio site. It currently has static values set for the number of stars on the repositories I highlight. I want this to be accurate, but I do NOT want it to make requests to GitHub every time a user visits. Instead, please implement a solution which will run once a day, fetch the number of stars from a list of repos, then write them to vercel's KV store. Finally, update the UI to make a request to the KV store when the user visits my page and render the accurate star counts.",
+This monorepo is for an AI coding agent. The app runs and edits the code in the cloud in a sandboxed environment. Right now, we require users to generate a GitHub PAT, which we store in a .env file and can use to authenticate with GitHub. This is not idea, and instead we want to have a github oauth app which users can authenticate with.
+Please implement a new auth server inside the /auth directory which can do this.
+You will not have any access to secrets, so you will not be able to run the server to test it.
+I want the server to be able to authenticate users with GitHub, such that we will be able to take the following actions:
+1. clone repositories they give us access to
+2. checkout existing and create new branches on the repositories they give us access to
+3. make pull requests and push changes to the repositories they give us access to
+Once you're done, ensure you've documented the development process in the readme of this new app.`,
       },
       {
         id: "toolu_01RNEjViZFBXgCSNMwqKvkVL",
@@ -116,8 +123,8 @@ async function runFromPlan() {
     "thread_id" | "assistant_id"
   > = {
     target_repository: {
-      owner: "bracesproul",
+      owner: "langchain-ai",
-      repo: "personal-site",
+      repo: "open-swe",
     },
   };
 
@@ -144,7 +151,7 @@ async function runFromPlan() {
     },
     config: {
       configurable,
-      recursion_limit: 200,
+      recursion_limit: 400,
     },
     ifNotExists: "create",
     streamSubgraphs: true,

apps/open-swe/src/nodes/generate-message.ts

@@ -44,6 +44,7 @@ You MUST adhere to the following criteria when executing the task:
   - Each summary message will also include a short description of the task it completed, how it did so, and every change it made to the codebase during this task. This section will be titled 'Repository modifications summary'.
 - All changes are automatically committed, so you should not worry about creating backups, or committing changes.
 - Use \`apply_patch\` to edit files. This tool accepts diffs and file paths. It will then apply the given diff to the file.
+- You should NOT try to create empty files with \`apply_patch\`. If you need to create a file, use the \`shell\` tool, and pass \`touch <file path>\` to create the file.
 - When using the \`shell\` tool, always take advantage of the \`workdir\` parameter to run commands inside the repo directory. You should not try to generate a command with \`cd <some path>\` as passing that path to \`workdir\` is much more efficient.
 - Do not try to install dependencies, or run a server, compile the code, etc., unless you are explicitly asked to.
 - If completing the user's task requires writing or modifying files:

apps/open-swe/src/utils/read-write.ts

@@ -6,6 +6,33 @@ import { traceable } from "langsmith/traceable";
 
 const logger = createLogger(LogLevel.INFO, "ReadWriteUtil");
 
+async function handleCreateFile(
+  sandbox: Sandbox,
+  filePath: string,
+  args?: {
+    workDir?: string;
+  },
+) {
+  try {
+    const touchCommand = `touch "${filePath}"`;
+    const touchOutput = await sandbox.commands.run(touchCommand, {
+      cwd: args?.workDir,
+    });
+    return touchOutput;
+  } catch (e) {
+    const errorFields = getSandboxErrorFields(e);
+    if (errorFields) {
+      return errorFields;
+    }
+    return {
+      exitCode: 1,
+      error: e instanceof Error ? e.message : String(e),
+      stdout: "",
+      stderr: "",
+    };
+  }
+}
+
 async function readFileFunc(
   sandbox: Sandbox,
   filePath: string,
@@ -42,6 +69,19 @@ async function readFileFunc(
       output: readOutput.stdout,
     };
   } catch (e: any) {
+    if (e instanceof Error && e.message.includes("No such file or directory")) {
+      const createOutput = await handleCreateFile(sandbox, filePath, args);
+      if (createOutput.exitCode !== 0) {
+        return {
+          success: false,
+          output: `FAILED TO EXECUTE READ COMMAND for sandbox '${filePath}'. Error: ${(e as Error).message || String(e)}`,
+        };
+      } else {
+        // If the file was created successfully, try reading it again.
+        return readFile(sandbox, filePath, args);
+      }
+    }
+
     logger.error(
       `Exception while trying to read file '${filePath}' from sandbox via cat:`,
       {