@empiricalrun/test-gen 0.58.0 → 0.59.0

package/dist/tools/download-build.js

@@ -1,17 +1,18 @@
 "use strict";
 Object.defineProperty(exports, "__esModule", { value: true });
-exports.downloadBuildTool = void 0;
+exports.downloadBuildTool = exports.downloadBuildToolSchema = void 0;
 const zod_1 = require("zod");
 const test_build_1 = require("../test-build");
+exports.downloadBuildToolSchema = zod_1.z.object({
+    buildUrl: zod_1.z.string().describe("The URL of the build to download"),
+});
 exports.downloadBuildTool = {
     schema: {
         name: "downloadBuild",
         description: `Download a build from a build URL. If you do not have
 have a build URL, you can try getting the environment details with the getEnvironment tool.
 Environment details will include the build URL.`,
-        parameters: zod_1.z.object({
-            buildUrl: zod_1.z.string().describe("The URL of the build to download"),
-        }),
+        parameters: exports.downloadBuildToolSchema,
     },
     execute: async (input) => {
         if (!(await (0, test_build_1.hasDownloadScript)())) {

package/dist/tools/test-gen-browser.d.ts.map

@@ -1 +1 @@
-{"version":3,"file":"test-gen-browser.d.ts","sourceRoot":"","sources":["../../src/tools/test-gen-browser.ts"],"names":[],"mappings":"AACA,OAAO,KAAK,EAAE,IAAI,EAAE,MAAM,wBAAwB,CAAC;AAyFnD,eAAO,MAAM,4BAA4B,EAAE,IA8E1C,CAAC"}
+{"version":3,"file":"test-gen-browser.d.ts","sourceRoot":"","sources":["../../src/tools/test-gen-browser.ts"],"names":[],"mappings":"AACA,OAAO,KAAK,EAAE,IAAI,EAAE,MAAM,wBAAwB,CAAC;AAyFnD,eAAO,MAAM,4BAA4B,EAAE,IAyE1C,CAAC"}

package/dist/tools/test-gen-browser.js

@@ -104,34 +104,29 @@ exports.generateTestWithBrowserAgent = {
         }
         const fileBackup = await promises_1.default.readFile(fileName, "utf-8");
         try {
-            await (0, utils_1.replaceTodoWithCreateTest)({
-                testCaseName: testName,
-                testCaseSuites: testSuites,
-                testFilePath: fileName,
-            });
+            await (0, utils_1.replaceTodoWithCreateTest)(fileName);
         }
         catch (error) {
+            // Undo the TODO -> createTest and test.only changes
+            await promises_1.default.writeFile(fileName, fileBackup, "utf-8");
             return {
                 isError: true,
                 result: `Error running tool: ${error}`,
             };
         }
+        const testGenToken = (0, scenarios_1.buildTokenFromOptions)({ name: testName, file: fileName, prompt: changeToMake }, { useComputerUseAgent: true });
         const toolResult = await (0, run_1.generateTestsUsingMasterAgent)({
+            testCaseName: testName,
+            testCaseSuites: testSuites,
             testFilePath: fileName,
             filePathToUpdate: fileName,
             pwProjectsFilter: [project],
             traceId: trace?.id,
-            testGenToken: (0, scenarios_1.buildTokenFromOptions)({
-                name: testName,
-                file: fileName,
-                prompt: changeToMake,
-            }, {
-                useComputerUseAgent: true,
-            }),
+            testGenToken,
             repoDir: process.cwd(),
             editFileWithGeneratedCode: false,
         });
-        // Undo the TODO -> createTest change
+        // Undo the TODO -> createTest changes
         await promises_1.default.writeFile(fileName, fileBackup, "utf-8");
         const { isError, error, actionsSummary } = toolResult;
         if (!isError) {

package/dist/tools/test-run.d.ts.map

@@ -1 +1 @@
-{"version":3,"file":"test-run.d.ts","sourceRoot":"","sources":["../../src/tools/test-run.ts"],"names":[],"mappings":"AAAA,OAAO,KAAK,EAAE,IAAI,EAAE,MAAM,wBAAwB,CAAC;AAgDnD,eAAO,MAAM,WAAW,EAAE,IAuDzB,CAAC"}
+{"version":3,"file":"test-run.d.ts","sourceRoot":"","sources":["../../src/tools/test-run.ts"],"names":[],"mappings":"AAAA,OAAO,KAAK,EAAE,IAAI,EAAE,MAAM,wBAAwB,CAAC;AAgDnD,eAAO,MAAM,WAAW,EAAE,IAiDzB,CAAC"}

package/dist/tools/test-run.js

@@ -39,32 +39,27 @@ exports.runTestTool = {
     },
     execute: async (input) => {
         let reportUrl = undefined;
-        let projectName = undefined;
-        let testRunId = undefined;
+        let envOverrides = undefined;
         if (hasCloudflareCredentials()) {
-            projectName = "test-gen-chat-agent";
-            testRunId = Date.now().toString();
+            const projectName = "test-gen-chat-agent";
+            const testRunId = Date.now().toString();
             reportUrl = buildReportUrl(projectName, testRunId);
+            envOverrides = {
+                PROJECT_NAME: projectName,
+                TEST_RUN_GITHUB_ACTION_ID: testRunId,
+            };
         }
         else {
             console.warn("R2 credentials not found: report artifacts will not be uploaded");
         }
         const { testName, suites, fileName, project } = input;
         try {
-            // {"project":"chromium","suites":[],"fileName":"tests/quizizz-for-work/group.spec.ts","testName":"Create a group"}
-            // This runs all tests - TODO: Debug this, should only run the testName
             const result = await (0, test_run_1.runSingleTest)({
                 testName,
                 suites,
                 fileName,
                 projects: [project],
-                // Adding these to enforce report artifacts are uploaded
-                envOverrides: projectName && testRunId
-                    ? {
-                        PROJECT_NAME: projectName,
-                        TEST_RUN_GITHUB_ACTION_ID: testRunId,
-                    }
-                    : undefined,
+                envOverrides,
             });
             return {
                 result: buildResult({

package/dist/utils/checkpoint.d.ts.map

@@ -1 +1 @@
-{"version":3,"file":"checkpoint.d.ts","sourceRoot":"","sources":["../../src/utils/checkpoint.ts"],"names":[],"mappings":"AAAA,OAAO,EAAE,eAAe,EAAE,MAAM,wBAAwB,CAAC;AAIzD,wBAAsB,gBAAgB,CACpC,SAAS,EAAE,eAAe,EAAE,EAC5B,UAAU,EAAE,MAAM,iBAcnB"}
+{"version":3,"file":"checkpoint.d.ts","sourceRoot":"","sources":["../../src/utils/checkpoint.ts"],"names":[],"mappings":"AAAA,OAAO,EAAE,eAAe,EAAE,MAAM,wBAAwB,CAAC;AAIzD,wBAAsB,gBAAgB,CACpC,SAAS,EAAE,eAAe,EAAE,EAC5B,UAAU,EAAE,MAAM,iBAqBnB"}

package/dist/utils/checkpoint.js

@@ -5,7 +5,9 @@ const git_1 = require("./git");
 async function createCheckpoint(toolCalls, branchName) {
     const filesChanged = await (0, git_1.getFilesChanged)();
     const toolsWithUpdatedFiles = toolCalls
-        .filter((tc) => tc.input.path && filesChanged.includes(tc.input.path))
+        .filter((tc) => "path" in tc.input &&
+        tc.input.path &&
+        filesChanged.includes(tc.input.path))
         .map((toolCall) => ({
         name: toolCall.name,
         path: toolCall.input.path,

package/dist/utils/exec.d.ts

@@ -1,12 +1,12 @@
 export declare class ProcessManager {
     private childProcess;
-    execute(command: string[], options: {
+    execute(command: string, args: string[], options: {
         env?: Record<string, string>;
     }): Promise<number>;
     terminate(): void;
     isRunning(): boolean;
 }
-export declare function cmd(command: string[], options: {
+export declare function cmd(command: string, args: string[], options: {
     env?: Record<string, string>;
 }): Promise<number>;
 //# sourceMappingURL=exec.d.ts.map

package/dist/utils/exec.d.ts.map

@@ -1 +1 @@
-{"version":3,"file":"exec.d.ts","sourceRoot":"","sources":["../../src/utils/exec.ts"],"names":[],"mappings":"AAmBA,qBAAa,cAAc;IACzB,OAAO,CAAC,YAAY,CAA6B;IAE3C,OAAO,CACX,OAAO,EAAE,MAAM,EAAE,EACjB,OAAO,EAAE;QAAE,GAAG,CAAC,EAAE,MAAM,CAAC,MAAM,EAAE,MAAM,CAAC,CAAA;KAAE,GACxC,OAAO,CAAC,MAAM,CAAC;IAkDlB,SAAS,IAAI,IAAI;IASjB,SAAS,IAAI,OAAO;CAGrB;AAED,wBAAsB,GAAG,CACvB,OAAO,EAAE,MAAM,EAAE,EACjB,OAAO,EAAE;IAAE,GAAG,CAAC,EAAE,MAAM,CAAC,MAAM,EAAE,MAAM,CAAC,CAAA;CAAE,GACxC,OAAO,CAAC,MAAM,CAAC,CAGjB"}
+{"version":3,"file":"exec.d.ts","sourceRoot":"","sources":["../../src/utils/exec.ts"],"names":[],"mappings":"AAmBA,qBAAa,cAAc;IACzB,OAAO,CAAC,YAAY,CAA6B;IAE3C,OAAO,CACX,OAAO,EAAE,MAAM,EACf,IAAI,EAAE,MAAM,EAAE,EACd,OAAO,EAAE;QAAE,GAAG,CAAC,EAAE,MAAM,CAAC,MAAM,EAAE,MAAM,CAAC,CAAA;KAAE,GACxC,OAAO,CAAC,MAAM,CAAC;IAmDlB,SAAS,IAAI,IAAI;IASjB,SAAS,IAAI,OAAO;CAGrB;AAED,wBAAsB,GAAG,CACvB,OAAO,EAAE,MAAM,EACf,IAAI,EAAE,MAAM,EAAE,EACd,OAAO,EAAE;IAAE,GAAG,CAAC,EAAE,MAAM,CAAC,MAAM,EAAE,MAAM,CAAC,CAAA;CAAE,GACxC,OAAO,CAAC,MAAM,CAAC,CAGjB"}

package/dist/utils/exec.js

@@ -21,7 +21,8 @@ function isAcceptableExit(exitCode, signal) {
 }
 class ProcessManager {
     childProcess = null;
-    async execute(command, options) {
+    async execute(command, args, options) {
+        console.log(`Executing command: ${command} with args: ${args}`);
         if (this.childProcess) {
             throw new Error("Process is already running");
         }
@@ -31,7 +32,7 @@ class ProcessManager {
                 rejectFunc(new Error("Command cannot be empty"));
                 return;
             }
-            const p = (0, child_process_1.spawn)(command[0], command.slice(1), {
+            const p = (0, child_process_1.spawn)(command, args, {
                 env: { ...process_1.default.env, ...options.env },
                 detached: true, // Create process group so we can terminate all child processes
             });
@@ -77,7 +78,7 @@ class ProcessManager {
     }
 }
 exports.ProcessManager = ProcessManager;
-async function cmd(command, options) {
+async function cmd(command, args, options) {
     const manager = new ProcessManager();
-    return manager.execute(command, options);
+    return manager.execute(command, args, options);
 }

package/package.json

@@ -1,6 +1,6 @@
 {
   "name": "@empiricalrun/test-gen",
-  "version": "0.58.0",
+  "version": "0.59.0",
   "publishConfig": {
     "registry": "https://registry.npmjs.org/",
     "access": "public"
@@ -60,9 +60,9 @@
     "tsx": "^4.16.2",
     "typescript": "^5.3.3",
     "zod": "^3.23.8",
-    "@empiricalrun/llm": "^0.15.4",
+    "@empiricalrun/llm": "^0.16.0",
     "@empiricalrun/r2-uploader": "^0.3.9",
-    "@empiricalrun/test-run": "^0.8.5"
+    "@empiricalrun/test-run": "^0.9.0"
   },
   "devDependencies": {
     "@playwright/test": "1.47.1",
@@ -76,7 +76,7 @@
     "playwright": "1.47.1",
     "serve-handler": "^6.1.6",
     "ts-patch": "^3.3.0",
-    "@empiricalrun/shared-types": "0.0.8"
+    "@empiricalrun/shared-types": "0.1.0"
   },
   "scripts": {
     "dev": "tspc --build --watch",

package/tsconfig.tsbuildinfo

@@ -1 +1 @@
-{"root":["./src/index.ts","./src/actions/assert.ts","./src/actions/click.ts","./src/actions/done.ts","./src/actions/fill.ts","./src/actions/goto.ts","./src/actions/hover.ts","./src/actions/index.ts","./src/actions/next-task.ts","./src/actions/press.ts","./src/actions/skill.ts","./src/actions/text-content.ts","./src/actions/constants/index.ts","./src/actions/utils/index.ts","./src/agent/browsing/index.ts","./src/agent/browsing/run.ts","./src/agent/browsing/utils.ts","./src/agent/chat/agent-loop.ts","./src/agent/chat/exports.ts","./src/agent/chat/index.ts","./src/agent/chat/prompt.ts","./src/agent/chat/repo.ts","./src/agent/chat/state.ts","./src/agent/chat/types.ts","./src/agent/codegen/create-test-block.ts","./src/agent/codegen/fix-ts-errors.ts","./src/agent/codegen/generate-code-apply-changes.ts","./src/agent/codegen/lexical-scoped-vars.ts","./src/agent/codegen/repo-edit.ts","./src/agent/codegen/run.ts","./src/agent/codegen/skills-retriever.ts","./src/agent/codegen/test-update-feedback.ts","./src/agent/codegen/types.ts","./src/agent/codegen/update-flow.ts","./src/agent/codegen/use-skill.ts","./src/agent/codegen/utils.ts","./src/agent/cua/computer.ts","./src/agent/cua/index.ts","./src/agent/cua/model.ts","./src/agent/diagnosis-agent/index.ts","./src/agent/diagnosis-agent/strict-mode-violation.ts","./src/agent/enrich-prompt/index.ts","./src/agent/enrich-prompt/utils.ts","./src/agent/infer-agent/index.ts","./src/agent/master/action-tool-calls.ts","./src/agent/master/element-annotation.ts","./src/agent/master/execute-browser-action.ts","./src/agent/master/execute-skill-action.ts","./src/agent/master/next-action.ts","./src/agent/master/planner.ts","./src/agent/master/run.ts","./src/agent/master/scroller.ts","./src/agent/master/with-hints.ts","./src/agent/master/browser-tests/cua.spec.ts","./src/agent/master/browser-tests/fixtures.ts","./src/agent/master/browser-tests/index.spec.ts","./src/agent/master/browser-tests/skills.spec.ts","./src/agent/master/icon-descriptor/index.ts","./src/agent/master/icon-descriptor/normalize-svg.ts","./src/agent/planner/run-time-planner.ts","./src/agent/planner/run.ts","./src/bin/index.ts","./src/bin/logger/index.ts","./src/bin/utils/context.ts","./src/bin/utils/index.ts","./src/bin/utils/fs/index.ts","./src/bin/utils/platform/web/index.ts","./src/bin/utils/platform/web/test-files/ts-path-import-validate.ts","./src/bin/utils/scenarios/index.ts","./src/browser-injected-scripts/annotate-elements.spec.ts","./src/constants/index.ts","./src/errors/index.ts","./src/evals/add-scenario-agent.evals.ts","./src/evals/append-create-test-agent.evals.ts","./src/evals/fetch-pom-skills-agent.evals.ts","./src/evals/infer-master-or-code-agent.evals.ts","./src/evals/master-agent.evals.ts","./src/evals/type.ts","./src/evals/update-scenario-agent.evals.ts","./src/file/client.ts","./src/file/server.ts","./src/human-in-the-loop/cli.ts","./src/human-in-the-loop/index.ts","./src/human-in-the-loop/ipc.ts","./src/page/index.ts","./src/prompts/lib/ts-transformer.ts","./src/reporter/index.ts","./src/reporter/lib.ts","./src/session/index.ts","./src/test-build/index.ts","./src/tool-call-service/index.ts","./src/tools/codegen-agent.ts","./src/tools/commit-and-create-pr.ts","./src/tools/diagnosis-fetcher.ts","./src/tools/download-build.ts","./src/tools/environment-crud.ts","./src/tools/str_replace_editor.ts","./src/tools/test-gen-browser.ts","./src/tools/test-run.ts","./src/tools/grep/index.ts","./src/tools/grep/ripgrep/index.ts","./src/tools/grep/ripgrep/types.ts","./src/tools/test-run-fetcher/index.ts","./src/tools/test-run-fetcher/types.ts","./src/tools/utils/index.ts","./src/types/handlebars.d.ts","./src/types/index.ts","./src/uploader/index.ts","./src/uploader/utils.ts","./src/utils/checkpoint.ts","./src/utils/env.ts","./src/utils/exec.ts","./src/utils/file-tree.ts","./src/utils/file.ts","./src/utils/git.ts","./src/utils/html.ts","./src/utils/index.ts","./src/utils/pw-test.ts","./src/utils/repo-tree.ts","./src/utils/slug.ts","./src/utils/string.ts"],"version":"5.8.3"}
+{"root":["./src/index.ts","./src/actions/assert.ts","./src/actions/click.ts","./src/actions/done.ts","./src/actions/fill.ts","./src/actions/goto.ts","./src/actions/hover.ts","./src/actions/index.ts","./src/actions/next-task.ts","./src/actions/press.ts","./src/actions/skill.ts","./src/actions/text-content.ts","./src/actions/constants/index.ts","./src/actions/utils/index.ts","./src/agent/browsing/index.ts","./src/agent/browsing/run.ts","./src/agent/browsing/utils.ts","./src/agent/chat/agent-loop.ts","./src/agent/chat/exports.ts","./src/agent/chat/index.ts","./src/agent/chat/models.ts","./src/agent/chat/prompt.ts","./src/agent/chat/repo.ts","./src/agent/chat/state.ts","./src/agent/chat/types.ts","./src/agent/codegen/create-test-block.ts","./src/agent/codegen/fix-ts-errors.ts","./src/agent/codegen/generate-code-apply-changes.ts","./src/agent/codegen/lexical-scoped-vars.ts","./src/agent/codegen/repo-edit.ts","./src/agent/codegen/run.ts","./src/agent/codegen/skills-retriever.ts","./src/agent/codegen/test-update-feedback.ts","./src/agent/codegen/types.ts","./src/agent/codegen/update-flow.ts","./src/agent/codegen/use-skill.ts","./src/agent/codegen/utils.ts","./src/agent/cua/computer.ts","./src/agent/cua/index.ts","./src/agent/cua/model.ts","./src/agent/diagnosis-agent/index.ts","./src/agent/diagnosis-agent/strict-mode-violation.ts","./src/agent/enrich-prompt/index.ts","./src/agent/enrich-prompt/utils.ts","./src/agent/infer-agent/index.ts","./src/agent/master/action-tool-calls.ts","./src/agent/master/element-annotation.ts","./src/agent/master/execute-browser-action.ts","./src/agent/master/execute-skill-action.ts","./src/agent/master/next-action.ts","./src/agent/master/planner.ts","./src/agent/master/run.ts","./src/agent/master/scroller.ts","./src/agent/master/with-hints.ts","./src/agent/master/browser-tests/cua.spec.ts","./src/agent/master/browser-tests/fixtures.ts","./src/agent/master/browser-tests/index.spec.ts","./src/agent/master/browser-tests/skills.spec.ts","./src/agent/master/icon-descriptor/index.ts","./src/agent/master/icon-descriptor/normalize-svg.ts","./src/agent/planner/run-time-planner.ts","./src/agent/planner/run.ts","./src/bin/index.ts","./src/bin/logger/index.ts","./src/bin/utils/context.ts","./src/bin/utils/index.ts","./src/bin/utils/fs/index.ts","./src/bin/utils/platform/web/index.ts","./src/bin/utils/platform/web/test-files/ts-path-import-validate.ts","./src/bin/utils/scenarios/index.ts","./src/browser-injected-scripts/annotate-elements.spec.ts","./src/constants/index.ts","./src/errors/index.ts","./src/evals/add-scenario-agent.evals.ts","./src/evals/append-create-test-agent.evals.ts","./src/evals/fetch-pom-skills-agent.evals.ts","./src/evals/infer-master-or-code-agent.evals.ts","./src/evals/master-agent.evals.ts","./src/evals/type.ts","./src/evals/update-scenario-agent.evals.ts","./src/file/client.ts","./src/file/server.ts","./src/human-in-the-loop/cli.ts","./src/human-in-the-loop/index.ts","./src/human-in-the-loop/ipc.ts","./src/page/index.ts","./src/prompts/lib/ts-transformer.ts","./src/reporter/index.ts","./src/reporter/lib.ts","./src/session/index.ts","./src/test-build/index.ts","./src/tool-call-service/index.ts","./src/tool-call-service/utils.ts","./src/tools/codegen-agent.ts","./src/tools/commit-and-create-pr.ts","./src/tools/diagnosis-fetcher.ts","./src/tools/download-build.ts","./src/tools/environment-crud.ts","./src/tools/str_replace_editor.ts","./src/tools/test-gen-browser.ts","./src/tools/test-run.ts","./src/tools/grep/index.ts","./src/tools/grep/ripgrep/index.ts","./src/tools/grep/ripgrep/types.ts","./src/tools/test-run-fetcher/index.ts","./src/tools/test-run-fetcher/types.ts","./src/tools/utils/index.ts","./src/types/handlebars.d.ts","./src/types/index.ts","./src/uploader/index.ts","./src/uploader/utils.ts","./src/utils/checkpoint.ts","./src/utils/env.ts","./src/utils/exec.ts","./src/utils/file-tree.ts","./src/utils/file.ts","./src/utils/git.ts","./src/utils/html.ts","./src/utils/index.ts","./src/utils/pw-test.ts","./src/utils/repo-tree.ts","./src/utils/slug.ts","./src/utils/string.ts"],"version":"5.8.3"}