@empiricalrun/test-gen 0.55.0 → 0.56.1

package/CHANGELOG.md

@@ -1,5 +1,32 @@
 # @empiricalrun/test-gen
 
+## 0.56.1
+
+### Patch Changes
+
+- d05bb69: chore: remove fs-extra
+- e1e2564: fix: avoid global promisify calls
+- 803f61f: fix: run type checks for file inserts in text editor tools
+- Updated dependencies [d05bb69]
+  - @empiricalrun/test-run@0.8.2
+
+## 0.56.0
+
+### Minor Changes
+
+- b073084: feat: new APIs and refactor to introduce tool execute service
+
+### Patch Changes
+
+- 93d7a0b: feat: add tool call for downloading builds
+- a58ac3f: feat: add dashboard agent integration and optimize tsx handling
+- Updated dependencies [87af227]
+- Updated dependencies [3831109]
+- Updated dependencies [a58ac3f]
+- Updated dependencies [f77e33d]
+  - @empiricalrun/llm@0.15.1
+  - @empiricalrun/test-run@0.8.1
+
 ## 0.55.0
 
 ### Minor Changes

package/dist/agent/browsing/run.js

@@ -5,13 +5,13 @@ var __importDefault = (this && this.__importDefault) || function (mod) {
 Object.defineProperty(exports, "__esModule", { value: true });
 exports.generateTestsUsingMasterAgent = void 0;
 const detect_port_1 = __importDefault(require("detect-port"));
-const fs_extra_1 = __importDefault(require("fs-extra"));
+const fs_1 = __importDefault(require("fs"));
 const web_1 = require("../../bin/utils/platform/web");
 const server_1 = require("../../file/server");
 const exec_1 = require("../../utils/exec");
 const utils_1 = require("./utils");
 async function generateTestsUsingMasterAgent({ testFilePath, filePathToUpdate, pwProjectsFilter, testGenToken, repoDir, editFileWithGeneratedCode, }) {
-    if (!fs_extra_1.default.existsSync(testFilePath)) {
+    if (!fs_1.default.existsSync(testFilePath)) {
         throw new Error(`File for master agent to run not found: ${testFilePath}`);
     }
     const pm = new exec_1.ProcessManager();
@@ -69,9 +69,9 @@ async function generateTestsUsingMasterAgent({ testFilePath, filePathToUpdate, p
     // clean up the file if there is any error
     if (isError) {
         try {
-            const fileContent = await fs_extra_1.default.readFile(filePathToUpdate, "utf-8");
+            const fileContent = fs_1.default.readFileSync(filePathToUpdate, "utf-8");
             const updatedContent = (0, web_1.replaceCreateTestWithNewCode)(filePathToUpdate, fileContent, "");
-            await fs_extra_1.default.writeFile(filePathToUpdate, updatedContent, "utf-8");
+            fs_1.default.writeFileSync(filePathToUpdate, updatedContent, "utf-8");
             await (0, web_1.lintErrors)(filePathToUpdate);
         }
         catch (e) {

package/dist/agent/browsing/utils.d.ts.map

@@ -1 +1 @@
-{"version":3,"file":"utils.d.ts","sourceRoot":"","sources":["../../../src/agent/browsing/utils.ts"],"names":[],"mappings":"AAAA,OAAO,EAAE,WAAW,EAAE,MAAM,mBAAmB,CAAC;AAChD,OAAO,KAAK,EAAe,QAAQ,EAAE,MAAM,4BAA4B,CAAC;AAIxE,OAAO,EAAE,IAAI,EAAE,MAAM,YAAY,CAAC;AAClC,OAAO,EAAE,oBAAoB,EAAE,MAAM,iBAAiB,CAAC;AAsBvD,wBAAgB,QAAQ,CAAC,GAAG,EAAE,GAAG,GAAG,GAAG,IAAI,MAAM,CAKhD;AAED,wBAAgB,wBAAwB,CAAC,KAAK,EAAE,MAAM,EAAE,UAIvD;AAiFD,wBAAsB,yBAAyB,CAAC,EAC9C,YAAY,EACZ,YAAY,EACZ,cAAc,GACf,EAAE;IACD,YAAY,EAAE,MAAM,CAAC;IACrB,YAAY,EAAE,MAAM,CAAC;IACrB,cAAc,EAAE,MAAM,EAAE,CAAC;CAC1B,iBA2BA;AAED,wBAAsB,cAAc,CAAC,EACnC,YAAY,EACZ,cAAc,EACd,QAAQ,GACT,EAAE;IACD,YAAY,EAAE,MAAM,CAAC;IACrB,cAAc,EAAE,MAAM,EAAE,CAAC;IACzB,QAAQ,EAAE,MAAM,CAAC;CAClB,iBAoBA;AAED,wBAAsB,yBAAyB,CAAC,EAC9C,QAAQ,EACR,QAAQ,EACR,KAAK,GACN,EAAE;IACD,QAAQ,EAAE,QAAQ,CAAC;IACnB,QAAQ,EAAE,MAAM,CAAC;IACjB,KAAK,CAAC,EAAE,WAAW,CAAC;CACrB,GAAG,OAAO,CAAC,MAAM,CAAC,CAyDlB;AAyBD,wBAAsB,wBAAwB,CAAC,IAAI,EAAE,IAAI,iBA2HxD;AAED;;;GAGG;AACH,wBAAsB,oBAAoB,CACxC,OAAO,EAAE,MAAM,GACd,OAAO,CAAC,oBAAoB,CAAC,CAM/B;AAWD,wBAAsB,oBAAoB,CACxC,gBAAgB,EAAE,oBAAoB,GACrC,OAAO,CAAC,MAAM,EAAE,CAAC,CAQnB;AAED;;;;;GAKG;AACH,wBAAsB,iBAAiB,CACrC,YAAY,EAAE,MAAM,EACpB,gBAAgB,EAAE,oBAAoB,EACtC,gBAAgB,GAAE,MAAM,EAAU,GACjC,OAAO,CAAC,MAAM,CAAC,CA+CjB;AAED,qBAAa,eAAe;IACd,OAAO,CAAC,SAAS;gBAAT,SAAS,EAAE,MAAM;IACrC,OAAO,CAAC,aAAa,CAAqB;YAE5B,mBAAmB;YAUnB,gBAAgB;IAsBjB,OAAO;IAoBb,SAAS;CAKjB"}
+{"version":3,"file":"utils.d.ts","sourceRoot":"","sources":["../../../src/agent/browsing/utils.ts"],"names":[],"mappings":"AAAA,OAAO,EAAE,WAAW,EAAE,MAAM,mBAAmB,CAAC;AAChD,OAAO,KAAK,EAAe,QAAQ,EAAE,MAAM,4BAA4B,CAAC;AAIxE,OAAO,EAAE,IAAI,EAAE,MAAM,YAAY,CAAC;AAClC,OAAO,EAAE,oBAAoB,EAAE,MAAM,iBAAiB,CAAC;AAwBvD,wBAAgB,QAAQ,CAAC,GAAG,EAAE,GAAG,GAAG,GAAG,IAAI,MAAM,CAKhD;AAED,wBAAgB,wBAAwB,CAAC,KAAK,EAAE,MAAM,EAAE,UAIvD;AAiFD,wBAAsB,yBAAyB,CAAC,EAC9C,YAAY,EACZ,YAAY,EACZ,cAAc,GACf,EAAE;IACD,YAAY,EAAE,MAAM,CAAC;IACrB,YAAY,EAAE,MAAM,CAAC;IACrB,cAAc,EAAE,MAAM,EAAE,CAAC;CAC1B,iBA2BA;AAED,wBAAsB,cAAc,CAAC,EACnC,YAAY,EACZ,cAAc,EACd,QAAQ,GACT,EAAE;IACD,YAAY,EAAE,MAAM,CAAC;IACrB,cAAc,EAAE,MAAM,EAAE,CAAC;IACzB,QAAQ,EAAE,MAAM,CAAC;CAClB,iBAoBA;AAED,wBAAsB,yBAAyB,CAAC,EAC9C,QAAQ,EACR,QAAQ,EACR,KAAK,GACN,EAAE;IACD,QAAQ,EAAE,QAAQ,CAAC;IACnB,QAAQ,EAAE,MAAM,CAAC;IACjB,KAAK,CAAC,EAAE,WAAW,CAAC;CACrB,GAAG,OAAO,CAAC,MAAM,CAAC,CAyDlB;AAyBD,wBAAsB,wBAAwB,CAAC,IAAI,EAAE,IAAI,iBA2HxD;AAED;;;GAGG;AACH,wBAAsB,oBAAoB,CACxC,OAAO,EAAE,MAAM,GACd,OAAO,CAAC,oBAAoB,CAAC,CA0B/B;AAWD,wBAAsB,oBAAoB,CACxC,gBAAgB,EAAE,oBAAoB,GACrC,OAAO,CAAC,MAAM,EAAE,CAAC,CAQnB;AAED;;;;;GAKG;AACH,wBAAsB,iBAAiB,CACrC,YAAY,EAAE,MAAM,EACpB,gBAAgB,EAAE,oBAAoB,EACtC,gBAAgB,GAAE,MAAM,EAAU,GACjC,OAAO,CAAC,MAAM,CAAC,CA+CjB;AAED,qBAAa,eAAe;IACd,OAAO,CAAC,SAAS;gBAAT,SAAS,EAAE,MAAM;IACrC,OAAO,CAAC,aAAa,CAAqB;YAE5B,mBAAmB;YAUnB,gBAAgB;IAsBjB,OAAO;IAoBb,SAAS;CAKjB"}

package/dist/agent/browsing/utils.js

@@ -4,14 +4,15 @@ var __importDefault = (this && this.__importDefault) || function (mod) {
 };
 Object.defineProperty(exports, "__esModule", { value: true });
 exports.TeardownManager = exports.detectProjectName = exports.getValidProjectNames = exports.readPlaywrightConfig = exports.injectPwLocatorGenerator = exports.prepareFileForMasterAgent = exports.markTestAsOnly = exports.replaceTodoWithCreateTest = exports.prepareBrowsingAgentTask = exports.isRegExp = void 0;
-const fs_extra_1 = __importDefault(require("fs-extra"));
+const fs_1 = __importDefault(require("fs"));
 const minimatch_1 = require("minimatch");
 const path_1 = __importDefault(require("path"));
 const ts_morph_1 = require("ts-morph");
-const api_1 = __importDefault(require("tsx/cjs/api"));
+// For TypeScript type safety
+let tsxImport = null;
 const logger_1 = require("../../bin/logger");
 const context_1 = require("../../bin/utils/context");
-const fs_1 = require("../../bin/utils/fs");
+const fs_2 = require("../../bin/utils/fs");
 const web_1 = require("../../bin/utils/platform/web");
 const create_test_block_1 = require("../codegen/create-test-block");
 const fix_ts_errors_1 = require("../codegen/fix-ts-errors");
@@ -32,10 +33,10 @@ async function addImportForCreateTest(testFilePath) {
     // Instead of using "@empiricalrun/test-gen", we use the local dist file
     // This is to avoid assuming that the test-gen package is installed in the project
     const importSource = path_1.default.join(__dirname, "../../../dist/index.js");
-    if (!fs_extra_1.default.existsSync(importSource)) {
+    if (!fs_1.default.existsSync(importSource)) {
         throw new Error(`createTest import source not found at ${importSource}`);
     }
-    await fs_extra_1.default.writeFile(testFilePath, (0, web_1.addNewImport)(await fs_extra_1.default.readFile(testFilePath, "utf-8"), ["createTest"], importSource));
+    fs_1.default.writeFileSync(testFilePath, (0, web_1.addNewImport)(fs_1.default.readFileSync(testFilePath, "utf-8"), ["createTest"], importSource));
 }
 async function prepareFileForUpdateScenario({ testCase, specPath, trace, }) {
     const { name, suites } = testCase;
@@ -61,7 +62,7 @@ async function prepareFileForUpdateScenario({ testCase, specPath, trace, }) {
         },
     });
     const scopeVariables = await (0, lexical_scoped_vars_1.getLexicalScopedVars)({
-        file: await fs_extra_1.default.readFile(createTestFilePath, "utf-8"),
+        file: fs_1.default.readFileSync(createTestFilePath, "utf-8"),
         referencePoint: "await createTest",
         trace: fetchScopeVariablesSpan,
     });
@@ -91,7 +92,7 @@ async function prepareFileForUpdateScenario({ testCase, specPath, trace, }) {
 async function replaceTodoWithCreateTest({ testFilePath, testCaseName, testCaseSuites, }) {
     // This method is an alternative to prepareFileForUpdateScenario
     // TODO: Does not support scoped variables and updates in POM files
-    const fileContent = await fs_extra_1.default.readFile(testFilePath, "utf-8");
+    const fileContent = fs_1.default.readFileSync(testFilePath, "utf-8");
     const todoRegex = /\/\/ TODO\(agent(?:\s+on\s+(\w+))?\):\s*(.*)/;
     const todoMatch = fileContent.match(todoRegex);
     if (!todoMatch) {
@@ -99,7 +100,7 @@ async function replaceTodoWithCreateTest({ testFilePath, testCaseName, testCaseS
     }
     const [, pageVarName] = todoMatch;
     const pageVariable = pageVarName || "page"; // Default to "page" if not specified
-    await fs_extra_1.default.writeFile(testFilePath, fileContent.replace(todoRegex, (_, __, todoText) => `await createTest("${todoText.replace(/"/g, '\\"')}", ${pageVariable});`));
+    fs_1.default.writeFileSync(testFilePath, fileContent.replace(todoRegex, (_, __, todoText) => `await createTest("${todoText.replace(/"/g, '\\"')}", ${pageVariable});`));
     await addImportForCreateTest(testFilePath);
     await markTestAsOnly({
         testCaseName,
@@ -109,7 +110,7 @@ async function replaceTodoWithCreateTest({ testFilePath, testCaseName, testCaseS
 }
 exports.replaceTodoWithCreateTest = replaceTodoWithCreateTest;
 async function markTestAsOnly({ testCaseName, testCaseSuites, specPath, }) {
-    const testFileContent = await fs_extra_1.default.readFile(specPath, "utf-8");
+    const testFileContent = fs_1.default.readFileSync(specPath, "utf-8");
     const { testBlock, testNode } = (0, web_1.getTypescriptTestBlock)({
         scenarioName: testCaseName,
         content: testFileContent,
@@ -120,7 +121,7 @@ async function markTestAsOnly({ testCaseName, testCaseSuites, specPath, }) {
     // add test.only / describe.only to the spec file so that only that block is executed
     if (!isFileMarkedSerial) {
         const updatedTestFileContent = newContentsWithTestOnly(testFileContent, testBlock, testBlock, parentDescribe?.getText() || "");
-        await fs_extra_1.default.writeFile(specPath, updatedTestFileContent);
+        fs_1.default.writeFileSync(specPath, updatedTestFileContent);
     }
 }
 exports.markTestAsOnly = markTestAsOnly;
@@ -132,17 +133,17 @@ async function prepareFileForMasterAgent({ testCase, specPath, trace, }) {
     const { name, suites } = testCase;
     // check if the spec file exists
     // if no then create a new file with test and expect imports
-    if (!fs_extra_1.default.existsSync(specPath)) {
+    if (!fs_1.default.existsSync(specPath)) {
         const fileCreateSpan = prepareFileSpan?.span({
             name: "create-file",
             input: { specPath },
         });
-        await fs_extra_1.default.createFile(specPath);
+        fs_1.default.writeFileSync(specPath, "");
         const fileContentWithImports = (0, web_1.addNewImport)("", ["test", "expect"], (0, web_1.getFixtureImportPath)(specPath));
-        await fs_extra_1.default.writeFile(specPath, fileContentWithImports, "utf-8");
+        fs_1.default.writeFileSync(specPath, fileContentWithImports, "utf-8");
         fileCreateSpan?.end({ output: { specPath, fileContentWithImports } });
     }
-    const existingContents = await fs_extra_1.default.readFile(specPath, "utf-8");
+    const existingContents = fs_1.default.readFileSync(specPath, "utf-8");
     const { testBlock } = (0, web_1.getTypescriptTestBlock)({
         scenarioName: name,
         suites,
@@ -154,13 +155,13 @@ async function prepareFileForMasterAgent({ testCase, specPath, trace, }) {
             testCase,
             file: specPath,
         });
-        await fs_extra_1.default.writeFile(specPath, `${existingContents} \n\n ${newTestBlock}`, "utf-8");
+        fs_1.default.writeFileSync(specPath, `${existingContents} \n\n ${newTestBlock}`, "utf-8");
         const updatedContent = (0, web_1.injectCodeSnippetBySuiteChain)({
             testFileContent: existingContents,
             suites: testCase.suites,
             codeSnippet: `\n\n${newTestBlock}`,
         });
-        await fs_extra_1.default.writeFile(specPath, updatedContent, "utf-8");
+        fs_1.default.writeFileSync(specPath, updatedContent, "utf-8");
     }
     const updatePath = await prepareFileForUpdateScenario({
         testCase,
@@ -189,7 +190,7 @@ async function injectPwLocatorGenerator(page) {
         pathToInstalledTestGen = process.cwd();
     }
     const annotateElementPath = path_1.default.join(pathToInstalledTestGen, "dist", "browser-injected-scripts", "annotate-elements.js");
-    if (!fs_extra_1.default.existsSync(annotateElementPath)) {
+    if (!fs_1.default.existsSync(annotateElementPath)) {
         throw new Error(`annotate-elements.js not found at path: ${annotateElementPath}`);
     }
     const remoteScriptResponses = await Promise.all([
@@ -198,7 +199,7 @@ async function injectPwLocatorGenerator(page) {
     ].map((url) => fetch(url)));
     const scripts = await Promise.all([
         ...remoteScriptResponses.map((r) => r.text()),
-        fs_extra_1.default.readFile(annotateElementPath, "utf-8"),
+        fs_1.default.readFileSync(annotateElementPath, "utf-8"),
     ]);
     page.on("load", async () => {
         try {
@@ -295,8 +296,25 @@ exports.injectPwLocatorGenerator = injectPwLocatorGenerator;
  * @return {*}  {Promise<PlaywrightTestConfig>}
  */
 async function readPlaywrightConfig(repoDir) {
+    if (typeof window !== "undefined") {
+        throw new Error("readPlaywrightConfig cannot be used in browser environments");
+    }
+    else {
+        // Only initialize on server side
+        // This will only execute on the server
+        await import("tsx/cjs/api")
+            .then((module) => {
+            tsxImport = module;
+        })
+            .catch((err) => {
+            console.error("Failed to import tsx:", err);
+        });
+    }
+    if (!tsxImport) {
+        throw new Error("tsx module not available");
+    }
     const [lastDir] = repoDir.split("/").reverse();
-    const playwrightConfig = (await api_1.default.require("./playwright.config.ts", `${repoDir}/${lastDir}`)).default;
+    const playwrightConfig = (await tsxImport.require("./playwright.config.ts", `${repoDir}/${lastDir}`)).default;
     return playwrightConfig;
 }
 exports.readPlaywrightConfig = readPlaywrightConfig;
@@ -377,7 +395,7 @@ class TeardownManager {
     teardownFiles = [];
     async getAllTeardownFiles() {
         const teardownFileRegex = /.*\.teardown\.ts/;
-        const teardownFiles = await (0, fs_1.readFilesInDirectory)(this.directory, (fileName) => teardownFileRegex.test(fileName));
+        const teardownFiles = await (0, fs_2.readFilesInDirectory)(this.directory, (fileName) => teardownFileRegex.test(fileName));
         return teardownFiles;
     }
     async skipTeardownFile(filePath) {
@@ -413,7 +431,7 @@ class TeardownManager {
     }
     unskipAll() {
         this.teardownFiles.forEach(({ filePath, content }) => {
-            fs_extra_1.default.writeFileSync(filePath, content, "utf-8");
+            fs_1.default.writeFileSync(filePath, content, "utf-8");
         });
     }
 }

package/dist/agent/chat/agent-loop.d.ts

@@ -1,9 +1,11 @@
 import { TraceClient } from "@empiricalrun/llm";
 import { IChatModel } from "@empiricalrun/llm/chat";
+import { ToolCallService } from "../../tool-call-service";
 import { ReporterFunction, SupportedChatModels } from "./types";
-export declare function chatAgentLoop({ chatModel, selectedModel, reporter, trace, }: {
+export declare function chatAgentLoop({ chatModel, selectedModel, reporter, trace, toolCallService, }: {
     chatModel: IChatModel<any>;
     selectedModel: SupportedChatModels;
+    toolCallService: ToolCallService;
     reporter: ReporterFunction;
     trace?: TraceClient;
 }): Promise<void>;

package/dist/agent/chat/agent-loop.d.ts.map

@@ -1 +1 @@
-{"version":3,"file":"agent-loop.d.ts","sourceRoot":"","sources":["../../../src/agent/chat/agent-loop.ts"],"names":[],"mappings":"AAAA,OAAO,EAAE,WAAW,EAAE,MAAM,mBAAmB,CAAC;AAChD,OAAO,EACL,UAAU,EAIX,MAAM,wBAAwB,CAAC;AAOhC,OAAO,EAAE,gBAAgB,EAAE,mBAAmB,EAAE,MAAM,SAAS,CAAC;AAgBhE,wBAAsB,aAAa,CAAC,EAClC,SAAS,EACT,aAAa,EACb,QAAQ,EACR,KAAK,GACN,EAAE;IACD,SAAS,EAAE,UAAU,CAAC,GAAG,CAAC,CAAC;IAC3B,aAAa,EAAE,mBAAmB,CAAC;IACnC,QAAQ,EAAE,gBAAgB,CAAC;IAC3B,KAAK,CAAC,EAAE,WAAW,CAAC;CACrB,iBAiDA"}
+{"version":3,"file":"agent-loop.d.ts","sourceRoot":"","sources":["../../../src/agent/chat/agent-loop.ts"],"names":[],"mappings":"AAAA,OAAO,EAAE,WAAW,EAAE,MAAM,mBAAmB,CAAC;AAChD,OAAO,EACL,UAAU,EAIX,MAAM,wBAAwB,CAAC;AAGhC,OAAO,EAAE,eAAe,EAAE,MAAM,yBAAyB,CAAC;AAG1D,OAAO,EAAE,gBAAgB,EAAE,mBAAmB,EAAE,MAAM,SAAS,CAAC;AAgBhE,wBAAsB,aAAa,CAAC,EAClC,SAAS,EACT,aAAa,EACb,QAAQ,EACR,KAAK,EACL,eAAe,GAChB,EAAE;IACD,SAAS,EAAE,UAAU,CAAC,GAAG,CAAC,CAAC;IAC3B,aAAa,EAAE,mBAAmB,CAAC;IACnC,eAAe,EAAE,eAAe,CAAC;IACjC,QAAQ,EAAE,gBAAgB,CAAC;IAC3B,KAAK,CAAC,EAAE,WAAW,CAAC;CACrB,iBAoCA"}

package/dist/agent/chat/agent-loop.js

@@ -3,8 +3,6 @@ Object.defineProperty(exports, "__esModule", { value: true });
 exports.chatAgentLoop = void 0;
 const chat_1 = require("@empiricalrun/llm/chat");
 const picocolors_1 = require("picocolors");
-const tool_call_service_1 = require("../../tool-call-service");
-const str_replace_editor_1 = require("../../tools/str_replace_editor");
 const prompt_1 = require("./prompt");
 const state_1 = require("./state");
 function getModelName(model) {
@@ -20,31 +18,18 @@ const log = (...args) => {
     console.log((0, picocolors_1.gray)(args.join(" ")));
 };
 const isRemote = process.env.TOOL_EXECUTION_IS_REMOTE === "true" || false;
-async function chatAgentLoop({ chatModel, selectedModel, reporter, trace, }) {
+async function chatAgentLoop({ chatModel, selectedModel, reporter, trace, toolCallService, }) {
     const systemPrompt = await (0, prompt_1.buildSystemPrompt)();
     trace?.update({ input: { systemPrompt } });
-    const toolCallService = new tool_call_service_1.ToolCallService();
-    const { tools } = await toolCallService.getTools(selectedModel);
+    const { tools } = await toolCallService.getTools();
     while (!chatModel.askUserForInput) {
         const toolCalls = chatModel.getPendingToolCalls();
+        console.log("toolCalls", toolCalls);
         if (toolCalls.length) {
-            const toolResults = [];
-            for (const call of toolCalls) {
-                const args = JSON.stringify(call.input);
-                log(`Executing tool ${call.name} with args: ${args}`);
-                let callResponse = await toolCallService.execute({
-                    tool: {
-                        name: call.name,
-                        input: call.input,
-                    },
-                }, isRemote);
-                if (callResponse.isError) {
-                    log(`Tool ${call.name} failed: ${callResponse.result}`);
-                }
-                else {
-                    log(`Tool ${call.name} completed`);
-                }
-                toolResults.push(callResponse);
+            const toolResults = await toolCallService.execute(toolCalls, isRemote);
+            if (isRemote) {
+                log(`Tool call remote execution in progress`);
+                break;
             }
             chatModel.pushToolResultsMessage(toolCalls, toolResults);
         }
@@ -62,6 +47,5 @@ async function chatAgentLoop({ chatModel, selectedModel, reporter, trace, }) {
         const latest = chatModel.getHumanReadableLatestMessage();
         await reporter((0, state_1.chatStateFromModel)(chatModel, selectedModel), latest);
     }
-    (0, str_replace_editor_1.cleanupBackupFiles)(process.cwd());
 }
 exports.chatAgentLoop = chatAgentLoop;

package/dist/agent/chat/exports.d.ts

@@ -0,0 +1,8 @@
+import { IChatModel } from "@empiricalrun/llm/chat";
+import { chatAgentLoop } from "./agent-loop";
+import { createChatModel } from "./model";
+import { ChatStateOnDisk } from "./state";
+import { ReporterFunction, SupportedChatModels } from "./types";
+export { chatAgentLoop, createChatModel };
+export type { ChatStateOnDisk, IChatModel, ReporterFunction, SupportedChatModels, };
+//# sourceMappingURL=exports.d.ts.map

package/dist/agent/chat/exports.d.ts.map

@@ -0,0 +1 @@
+{"version":3,"file":"exports.d.ts","sourceRoot":"","sources":["../../../src/agent/chat/exports.ts"],"names":[],"mappings":"AAAA,OAAO,EAAE,UAAU,EAAE,MAAM,wBAAwB,CAAC;AAEpD,OAAO,EAAE,aAAa,EAAE,MAAM,cAAc,CAAC;AAC7C,OAAO,EAAE,eAAe,EAAE,MAAM,SAAS,CAAC;AAC1C,OAAO,EAAE,eAAe,EAAE,MAAM,SAAS,CAAC;AAC1C,OAAO,EAAE,gBAAgB,EAAE,mBAAmB,EAAE,MAAM,SAAS,CAAC;AAEhE,OAAO,EAAE,aAAa,EAAE,eAAe,EAAE,CAAC;AAE1C,YAAY,EACV,eAAe,EACf,UAAU,EACV,gBAAgB,EAChB,mBAAmB,GACpB,CAAC"}

package/dist/agent/chat/exports.js

@@ -0,0 +1,7 @@
+"use strict";
+Object.defineProperty(exports, "__esModule", { value: true });
+exports.createChatModel = exports.chatAgentLoop = void 0;
+const agent_loop_1 = require("./agent-loop");
+Object.defineProperty(exports, "chatAgentLoop", { enumerable: true, get: function () { return agent_loop_1.chatAgentLoop; } });
+const model_1 = require("./model");
+Object.defineProperty(exports, "createChatModel", { enumerable: true, get: function () { return model_1.createChatModel; } });

package/dist/agent/chat/index.d.ts.map

@@ -1 +1 @@
-{"version":3,"file":"index.d.ts","sourceRoot":"","sources":["../../../src/agent/chat/index.ts"],"names":[],"mappings":"AAYA,OAAO,EAAoB,mBAAmB,EAAE,MAAM,SAAS,CAAC;AAiBhE,wBAAsB,kBAAkB,CAAC,EACvC,mBAAmB,EACnB,aAAa,EACb,oBAAoB,GACrB,EAAE;IACD,aAAa,EAAE,mBAAmB,CAAC;IACnC,mBAAmB,EAAE,OAAO,CAAC;IAC7B,oBAAoB,EAAE,MAAM,GAAG,SAAS,CAAC;CAC1C,iBAoFA;AAuBD,wBAAsB,wBAAwB,CAAC,EAC7C,aAAa,EACb,aAAa,GACd,EAAE;IACD,aAAa,EAAE,mBAAmB,CAAC;IACnC,aAAa,EAAE,MAAM,CAAC;CACvB,iBAyCA"}
+{"version":3,"file":"index.d.ts","sourceRoot":"","sources":["../../../src/agent/chat/index.ts"],"names":[],"mappings":"AAaA,OAAO,EAAoB,mBAAmB,EAAE,MAAM,SAAS,CAAC;AAiBhE,wBAAsB,kBAAkB,CAAC,EACvC,mBAAmB,EACnB,aAAa,EACb,oBAAoB,GACrB,EAAE;IACD,aAAa,EAAE,mBAAmB,CAAC;IACnC,mBAAmB,EAAE,OAAO,CAAC;IAC7B,oBAAoB,EAAE,MAAM,GAAG,SAAS,CAAC;CAC1C,iBAsFA;AAuBD,wBAAsB,wBAAwB,CAAC,EAC7C,aAAa,EACb,aAAa,GACd,EAAE;IACD,aAAa,EAAE,mBAAmB,CAAC;IACnC,aAAa,EAAE,MAAM,CAAC;CACvB,iBA2CA"}

package/dist/agent/chat/index.js

@@ -4,6 +4,7 @@ exports.runChatAgentForDashboard = exports.runChatAgentForCLI = void 0;
 const llm_1 = require("@empiricalrun/llm");
 const picocolors_1 = require("picocolors");
 const human_in_the_loop_1 = require("../../human-in-the-loop");
+const tool_call_service_1 = require("../../tool-call-service");
 const git_1 = require("../../utils/git");
 const agent_loop_1 = require("./agent-loop");
 const model_1 = require("./model");
@@ -83,11 +84,13 @@ async function runChatAgentForCLI({ useDiskForChatState, selectedModel, initialP
         }
         else {
             // TODO: Should we pass a loader function? That would allow us to show a spinner
+            const toolCallService = new tool_call_service_1.ToolCallService(null, selectedModel);
             await (0, agent_loop_1.chatAgentLoop)({
                 chatModel,
                 selectedModel,
                 reporter: reporterFunc,
                 trace,
+                toolCallService,
             });
         }
     }
@@ -125,6 +128,7 @@ async function runChatAgentForDashboard({ chatSessionId, selectedModel, }) {
             chatSessionId,
         },
     });
+    const toolCallService = new tool_call_service_1.ToolCallService(chatSessionId, selectedModel);
     await (0, git_1.checkoutBranch)(branchName);
     let chatModel = (0, model_1.createChatModel)(chatState.messages, selectedModel);
     let reporterFunc = async (chatState, latest) => {
@@ -147,6 +151,7 @@ async function runChatAgentForDashboard({ chatSessionId, selectedModel, }) {
         selectedModel,
         reporter: reporterFunc,
         trace,
+        toolCallService,
     });
     await (0, git_1.commitLocalAndPushBranchToRemote)(branchName);
 }

package/dist/agent/chat/repo.js

@@ -4,22 +4,22 @@ var __importDefault = (this && this.__importDefault) || function (mod) {
 };
 Object.defineProperty(exports, "__esModule", { value: true });
 exports.getRepoContext = void 0;
-const fs_extra_1 = __importDefault(require("fs-extra"));
+const fs_1 = __importDefault(require("fs"));
 const path_1 = __importDefault(require("path"));
 const repo_tree_1 = require("../../utils/repo-tree");
 const utils_1 = require("../browsing/utils");
 async function getAllMarkdownFiles() {
     const dir = path_1.default.join(process.cwd(), ".empiricalrun");
-    if (!fs_extra_1.default.existsSync(dir)) {
+    if (!fs_1.default.existsSync(dir)) {
         return [];
     }
-    const files = await fs_extra_1.default.readdir(dir);
+    const files = fs_1.default.readdirSync(dir);
     return files
         .filter((file) => file.endsWith(".md"))
         .map((file) => {
         return {
             name: file,
-            content: fs_extra_1.default.readFileSync(path_1.default.join(dir, file), "utf8"),
+            content: fs_1.default.readFileSync(path_1.default.join(dir, file), "utf8"),
         };
     });
 }

package/dist/agent/codegen/fix-ts-errors.js

@@ -5,7 +5,7 @@ var __importDefault = (this && this.__importDefault) || function (mod) {
 Object.defineProperty(exports, "__esModule", { value: true });
 exports.validateAndFixTypescriptErrors = void 0;
 const llm_1 = require("@empiricalrun/llm");
-const fs_extra_1 = __importDefault(require("fs-extra"));
+const fs_1 = __importDefault(require("fs"));
 const logger_1 = require("../../bin/logger");
 const web_1 = require("../../bin/utils/platform/web");
 const constants_1 = require("../../constants");
@@ -21,7 +21,7 @@ async function validateAndFixTypescriptErrors({ trace, logger = new logger_1.Cus
     const maxIteration = 2;
     let counter = 0;
     while (errors.length > 0) {
-        const fileContent = await fs_extra_1.default.readFile(file, "utf-8");
+        const fileContent = fs_1.default.readFileSync(file, "utf-8");
         counter += 1;
         if (counter > maxIteration) {
             trace?.event({ name: "code-fix-iteration-max-out" });
@@ -58,7 +58,7 @@ async function validateAndFixTypescriptErrors({ trace, logger = new logger_1.Cus
         });
         const response = message?.content || "";
         const readWriteFileSpan = trace?.span({ name: "write-to-file" });
-        await fs_extra_1.default.writeFile(file, response, "utf-8");
+        fs_1.default.writeFileSync(file, response, "utf-8");
         readWriteFileSpan?.end({ output: { response } });
         trace?.event({ name: "lint-file" });
         await (0, web_1.lintErrors)(file);

package/dist/agent/codegen/generate-code-apply-changes.js

@@ -5,7 +5,7 @@ var __importDefault = (this && this.__importDefault) || function (mod) {
 Object.defineProperty(exports, "__esModule", { value: true });
 exports.generateCodeAndApplyChanges = exports.systemPromptBuilderForRepoEdit = void 0;
 const llm_1 = require("@empiricalrun/llm");
-const fs_extra_1 = __importDefault(require("fs-extra"));
+const fs_1 = __importDefault(require("fs"));
 const path_1 = require("path");
 const constants_1 = require("../../constants");
 const types_1 = require("./types");
@@ -348,8 +348,8 @@ async function generateCodeAndApplyChanges({ task, trace, logger, getRelevantFil
                     newCode: args.code,
                     reason: args.reason,
                 });
-                await fs_extra_1.default.mkdir((0, path_1.dirname)(args.filePath), { recursive: true });
-                await fs_extra_1.default.writeFile(args.filePath, args.code, "utf-8");
+                fs_1.default.mkdirSync((0, path_1.dirname)(args.filePath), { recursive: true });
+                fs_1.default.writeFileSync(args.filePath, args.code, "utf-8");
                 console.log(`Created file: ${args.filePath}`);
             })();
         }));
@@ -360,7 +360,7 @@ async function generateCodeAndApplyChanges({ task, trace, logger, getRelevantFil
         // Filter out the tool calls which are for replacing code in existing files
         const fileChanges = strReplaceToolCalls
             .map((toolCall) => JSON.parse(toolCall.function.arguments))
-            .filter((f) => f.filePath && fs_extra_1.default.existsSync(f.filePath));
+            .filter((f) => f.filePath && fs_1.default.existsSync(f.filePath));
         // We add all the suggested changes to the updatedFiles array
         // This is used to validate and format files later
         updatedFiles.push(...fileChanges);

package/dist/agent/codegen/repo-edit.js

@@ -4,7 +4,7 @@ var __importDefault = (this && this.__importDefault) || function (mod) {
 };
 Object.defineProperty(exports, "__esModule", { value: true });
 exports.repoEditAgent = exports.generateCodeUsingRepoAgent = void 0;
-const fs_extra_1 = __importDefault(require("fs-extra"));
+const fs_1 = __importDefault(require("fs"));
 const context_1 = require("../../bin/utils/context");
 const web_1 = require("../../bin/utils/platform/web");
 const reporter_1 = require("../../reporter");
@@ -75,7 +75,7 @@ async function repoEditAgent({ trace, task, logger, }) {
     });
     await Promise.all(updates.map((f) => {
         return (async () => {
-            if (fs_extra_1.default.existsSync(f.filePath)) {
+            if (fs_1.default.existsSync(f.filePath)) {
                 await (0, web_1.lintErrors)(f.filePath);
             }
         })();

package/dist/agent/codegen/run.js

@@ -5,7 +5,7 @@ var __importDefault = (this && this.__importDefault) || function (mod) {
 Object.defineProperty(exports, "__esModule", { value: true });
 exports.generateTestWithCodegen = exports.createTestWithCodeAgent = void 0;
 const llm_1 = require("@empiricalrun/llm");
-const fs_extra_1 = __importDefault(require("fs-extra"));
+const fs_1 = __importDefault(require("fs"));
 const logger_1 = require("../../bin/logger");
 const context_1 = require("../../bin/utils/context");
 const web_1 = require("../../bin/utils/platform/web");
@@ -67,9 +67,9 @@ ${testCase.steps.join("\n")}`;
 exports.createTestWithCodeAgent = createTestWithCodeAgent;
 async function generateTestWithCodegen({ testCase, file, trace, }) {
     const logger = new logger_1.CustomLogger();
-    if (!fs_extra_1.default.existsSync(file)) {
+    if (!fs_1.default.existsSync(file)) {
         logger.log(`Creating a new spec file: ${file}`);
-        fs_extra_1.default.createFileSync(file);
+        fs_1.default.writeFileSync(file, "");
     }
     const context = await (0, context_1.contextForGeneration)(file);
     const { testFileContent } = context;
@@ -112,7 +112,7 @@ async function generateTestWithCodegen({ testCase, file, trace, }) {
     });
     await Promise.all(updates.fileChanges.map((f) => {
         return (async () => {
-            if (fs_extra_1.default.existsSync(f.filePath)) {
+            if (fs_1.default.existsSync(f.filePath)) {
                 await (0, web_1.lintErrors)(f.filePath);
             }
         })();

package/dist/agent/codegen/update-flow.js

@@ -5,10 +5,10 @@ var __importDefault = (this && this.__importDefault) || function (mod) {
 Object.defineProperty(exports, "__esModule", { value: true });
 exports.appendCreateTestBlock = exports.getAppendCreateTestBlockCompletion = exports.updateTest = exports.getUpdateTestCodeCompletion = void 0;
 const llm_1 = require("@empiricalrun/llm");
-const fs_extra_1 = __importDefault(require("fs-extra"));
+const fs_1 = __importDefault(require("fs"));
 const logger_1 = require("../../bin/logger");
 const context_1 = require("../../bin/utils/context");
-const fs_1 = require("../../bin/utils/fs");
+const fs_2 = require("../../bin/utils/fs");
 const web_1 = require("../../bin/utils/platform/web");
 const constants_1 = require("../../constants");
 const promptTemplate_0 = "{{#section \"system\"}}\n\nYou are a software test engineer who is given a task to update a test case. You will be provided with steps of a test\ncase scenario and you are given a snippet with interface await createTest(task, playwright_page_instance).\n\nYou need to analyze the request and place createTest snippet at the correct position and pass on the steps to this\n`createTest` method with the correct page instance.\n\nYou will be provided with current tests, fixtures and page object models for you to use and update code as per the\ntask provided to you.\nYou need to respond with file path and updated code block inside the file.\n\nHere is the list of current tests and fixtures:\n\n{{testFiles}}\n\n\nHere is the list of current page object models:\n\n{{pageFiles}}\n{{/section}}\n\n\n{{#section \"user\"}}\nFollowing is the test scenario for which you need to update the test:\n\ntest name:\n{{scenarioName}}\n\nTask:\n{{scenarioSteps}}\n\nFollow these steps to complete the task:\n\n1. Determine the task's intent:\n- The default intent is **\"add steps\"**.\n- If the task explicitly mentions (using words like \"replace\", \"remove\", \"delete\") that existing test steps should\nbe replaced or deleted, the intent is **\"update steps\"**. Identify the specific steps to be replaced or removed.\n- Do not assume that providing new or different steps implies an intent to update or replace existing steps. Unless\nthe task explicitly instructs to replace or remove existing code, interpret the intent as adding new steps to the\nexisting test.\n\n2. Identify the test block that requires updating.\n\n3. Place the `createTest` snippet:\n- Insert the `createTest` snippet at the location determined by the task, and pass the Playwright page instance.\n- If the intent is \"add\", do not alter the existing test code; simply append the `createTest` snippet based on any\nprovided location hints.\n- If the task includes location hints that don't match steps within the test, check dependent methods called from\nthe test for the update.\n- If no location hint is provided, place the `createTest` snippet at the end of the test block.\n- Even if the task includes steps that overlap with or are similar to existing steps, do not modify the existing\ncode. Unless the task explicitly instructs to replace or remove existing code, interpret the intent as adding new\nsteps to the existing test.\n\n4. Strip location hints from the task:\n- Remove any location hints (e.g., \"replace the current assertion and\") before passing the task to the `createTest`\nmethod.\n- The task passed into `createTest` should not contain any location hints. It should only contain actions that\nare required to be performed.\n- For example, if the original task: \"replace the current assertion and add a new assertion\" should be transformed to\n\"add a new assertion\" before passing it to the `createTest` method. This is the \"task_without_location_hints\"\n\nSubmission Guidelines:\n\n- Focus only on the provided test case and any related page object model methods used in the test.\n- Do not modify or add code within the `createTest` snippet.\n- Do not update or modify any other code apart from adding `createTest` code snippet.\n- Since the response will be used for search-and-replace operations, always provide the immediate parent AST node\nfor any code updates.\n- Include the full test block if any part of it is updated, preserving all unchanged code.\n- Do not use markdown syntax or backticks.\n- Respond using the following XML format:\n<reason_for_intent></reason_for_intent>\n<intent></intent>\n<location_of_update></location_of_update>\n<task_without_location_hints></task_without_location_hints>\n<file_path></file_path>\n<old_code_block></old_code_block>\n<new_code_block></new_code_block>\n<change></change>\n\n- Each `<old_code_block>` and `<new_code_block>` should contain only one test block or page object model method\n  definition. Provide separate blocks for multiple updates.\n- The `<change></change>` tag should also mention the file path being updated.\n- There should be only one `createTest` block in the `new_code_block`. The `createTest` method should be passed\nwith entire task. Do not split the task while forwarding it to `createTest`.\n- `<new_code_block>` code snippet should be syntactically correct.\n  - The code_block should not contain any import statements.\n{{/section}}";
@@ -214,8 +214,8 @@ async function appendCreateTestBlock({ testCase, file, trace, validateTypes = tr
     const logger = new logger_1.CustomLogger({ useReporter: false });
     logger.log("Appending create test block");
     const context = await (0, context_1.contextForGeneration)(file);
-    const scenarioFileContent = await fs_extra_1.default.readFile(file, "utf-8");
-    const testCodePrompt = (0, fs_1.convertFileContentsToString)([
+    const scenarioFileContent = fs_1.default.readFileSync(file, "utf-8");
+    const testCodePrompt = (0, fs_2.convertFileContentsToString)([
         {
             filePath: file,
             content: scenarioFileContent,

package/dist/agent/codegen/utils.js

@@ -5,7 +5,7 @@ var __importDefault = (this && this.__importDefault) || function (mod) {
 Object.defineProperty(exports, "__esModule", { value: true });
 exports.applyFileChanges = exports.searchAndReplaceCode = exports.applyFileChangesForCreateTest = exports.applyFileChangesUsingStrReplace = exports.validateTypesAndFormatCode = exports.extractTestStepsSuggestions = exports.extractAppendTestUpdates = exports.extractCreateTestUpdates = exports.extractTestUpdates = void 0;
 const llm_1 = require("@empiricalrun/llm");
-const fs_extra_1 = __importDefault(require("fs-extra"));
+const fs_1 = __importDefault(require("fs"));
 const ts_morph_1 = require("ts-morph");
 const web_1 = require("../../bin/utils/platform/web");
 const fix_ts_errors_1 = require("./fix-ts-errors");
@@ -178,7 +178,7 @@ async function applyFileChangesUsingStrReplace({ trace, fileChanges, logger, })
                 logger?.error(`Unable to find the code to update in ${result.filePath}, full error:`, result);
             }
             else {
-                await fs_extra_1.default.writeFile(fileChange.filePath, updatedContent, "utf-8");
+                fs_1.default.writeFileSync(fileChange.filePath, updatedContent, "utf-8");
                 readWriteFileSpan?.end({ output: { updatedContent } });
             }
             results.push(result);
@@ -213,13 +213,13 @@ async function applyFileChangesForCreateTest({ trace, fileChanges, testgenUpdate
                 },
             });
             if (fileChange.code) {
-                const currentContent = fs_extra_1.default.existsSync(fileChange.filePath)
-                    ? await fs_extra_1.default.readFile(fileChange.filePath, "utf-8")
+                const currentContent = fs_1.default.existsSync(fileChange.filePath)
+                    ? fs_1.default.readFileSync(fileChange.filePath, "utf-8")
                     : "";
-                await fs_extra_1.default.writeFile(fileChange.filePath, currentContent
+                fs_1.default.writeFileSync(fileChange.filePath, currentContent
                     ? `${currentContent}\n\n${fileChange.code}`
                     : fileChange.code, "utf-8");
-                const updatedFileContent = await fs_extra_1.default.readFile(fileChange.filePath, "utf-8");
+                const updatedFileContent = fs_1.default.readFileSync(fileChange.filePath, "utf-8");
                 readWriteFileSpan?.end({
                     output: { fileChange, updatedFile: updatedFileContent },
                 });
@@ -240,7 +240,7 @@ async function applyFileChangesForCreateTest({ trace, fileChanges, testgenUpdate
 }
 exports.applyFileChangesForCreateTest = applyFileChangesForCreateTest;
 async function searchAndReplaceCode({ logger, fileChange, }) {
-    let contents = await fs_extra_1.default.readFile(fileChange.filePath, "utf-8");
+    let contents = fs_1.default.readFileSync(fileChange.filePath, "utf-8");
     // since we dont know what is getting updated,
     // we believe that the patch is correct and contains few before and after lines
     // to make the change unique for search & replace
@@ -292,7 +292,7 @@ async function applyFileChanges({ trace, testCase, fileChanges, logger, }) {
                 // assuming the test case getting updated
                 // maintaining the previous accuracy of the test case update
                 const readWriteFileSpan = trace?.span({ name: "write-to-file" });
-                let contents = await fs_extra_1.default.readFile(fileChange.filePath, "utf-8");
+                let contents = fs_1.default.readFileSync(fileChange.filePath, "utf-8");
                 const [prependContent, strippedContent] = await (0, web_1.stripAndPrependImports)(fileChange.newCode, testCase?.name);
                 let updatedContent = prependContent + contents + `\n\n${strippedContent}`;
                 const { testBlock } = (0, web_1.getTypescriptTestBlock)({
@@ -302,12 +302,12 @@ async function applyFileChanges({ trace, testCase, fileChanges, logger, }) {
                 });
                 contents = contents.replace(testBlock, `\n\n${strippedContent}`);
                 updatedContent = prependContent + contents;
-                await fs_extra_1.default.writeFile(fileChange.filePath, updatedContent, "utf-8");
+                fs_1.default.writeFileSync(fileChange.filePath, updatedContent, "utf-8");
                 readWriteFileSpan?.end({ output: { updatedContent } });
             }
             else {
                 const readWriteFileSpan = trace?.span({ name: "write-to-file" });
-                let contents = await fs_extra_1.default.readFile(fileChange.filePath, "utf-8");
+                let contents = fs_1.default.readFileSync(fileChange.filePath, "utf-8");
                 const project = new ts_morph_1.Project();
                 const sourceFile = project.createSourceFile("updated-code.ts", fileChange.newCode);
                 const functions = sourceFile.getFunctions();
@@ -353,7 +353,7 @@ async function applyFileChanges({ trace, testCase, fileChanges, logger, }) {
                         contents = updatedContent;
                     }
                 }
-                await fs_extra_1.default.writeFile(fileChange.filePath, contents, "utf-8");
+                fs_1.default.writeFileSync(fileChange.filePath, contents, "utf-8");
                 readWriteFileSpan?.end({ output: { contents } });
                 results.push({
                     filePath: fileChange.filePath,