@empiricalrun/test-gen 0.79.4 → 0.79.6

package/dist/tools/list-tests-and-projects/index.js

@@ -37,7 +37,7 @@ function parseObjectString(objectString) {
 exports.listProjectsTool = {
     ...list_tests_and_projects_1.listProjectsTool,
     execute: async ({ repoPath }) => {
-        const projects = await (0, test_run_1.getProjectsFromPlaywrightConfig)(test_run_1.Platform.WEB, repoPath);
+        const projects = await (0, test_run_1.getProjectsFromPlaywrightConfig)(repoPath);
         return {
             isError: false,
             result: JSON.stringify(projects),

package/dist/tools/rename-file/index.d.ts.map

@@ -1 +1 @@
-{"version":3,"file":"index.d.ts","sourceRoot":"","sources":["../../../src/tools/rename-file/index.ts"],"names":[],"mappings":"AAAA,OAAO,KAAK,EAAE,IAAI,EAAc,MAAM,uCAAuC,CAAC;AAS9E,eAAO,MAAM,cAAc,EAAE,IA+E5B,CAAC"}
+{"version":3,"file":"index.d.ts","sourceRoot":"","sources":["../../../src/tools/rename-file/index.ts"],"names":[],"mappings":"AAAA,OAAO,KAAK,EAAE,IAAI,EAAc,MAAM,uCAAuC,CAAC;AAQ9E,eAAO,MAAM,cAAc,EAAE,IAkF5B,CAAC"}

package/dist/tools/rename-file/index.js

@@ -7,9 +7,9 @@ exports.renameFileTool = void 0;
 const child_process_1 = require("child_process");
 const promises_1 = require("fs/promises");
 const path_1 = __importDefault(require("path"));
-const web_1 = require("../../bin/utils/platform/web");
 const rename_file_1 = require("../definitions/rename-file");
 const helpers_1 = require("../file-operations/shared/helpers");
+const validators_1 = require("../utils/validators");
 exports.renameFileTool = {
     ...rename_file_1.renameFileTool,
     execute: async ({ input, repoPath }) => {
@@ -68,10 +68,13 @@ exports.renameFileTool = {
             }
             stageFile(oldFilePath);
             stageFile(newFilePath);
-            const tscResult = await (0, web_1.runTypescriptCompiler)(repoPath);
-            if (!tscResult.success) {
+            const validationResult = await (0, validators_1.runValidators)(repoPath, [
+                normalizedOldPath,
+                normalizedNewPath,
+            ]);
+            if (!validationResult.success) {
                 return {
-                    result: `File renamed from '${normalizedOldPath}' to '${normalizedNewPath}'. However, type checks are failing with errors:\n\n${tscResult.errors.join("\n")}`,
+                    result: `File renamed from '${normalizedOldPath}' to '${normalizedNewPath}'. However, validation is failing with errors:\n\n${validationResult.errors.join("\n")}`,
                     isError: false,
                 };
             }

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

@@ -1 +1 @@
-{"version":3,"file":"run-test.d.ts","sourceRoot":"","sources":["../../src/tools/run-test.ts"],"names":[],"mappings":"AAAA,OAAO,KAAK,EAEV,IAAI,EAEL,MAAM,uCAAuC,CAAC;AA2C/C,eAAO,MAAM,WAAW,EAAE,IAsIzB,CAAC"}
+{"version":3,"file":"run-test.d.ts","sourceRoot":"","sources":["../../src/tools/run-test.ts"],"names":[],"mappings":"AAAA,OAAO,KAAK,EAEV,IAAI,EAEL,MAAM,uCAAuC,CAAC;AA2C/C,eAAO,MAAM,WAAW,EAAE,IAmKzB,CAAC"}

package/dist/tools/run-test.js

@@ -47,7 +47,8 @@ exports.runTestTool = {
             const warnMsg = "R2 credentials not found: report artifacts will not be uploaded";
             logger?.warn(warnMsg) || console.warn(warnMsg);
         }
-        const { filePath, project } = input;
+        const { filePath } = input;
+        let { project } = input;
         // Normalize the path to support both /repo/... and relative paths
         const normalizedFilePath = (0, helpers_1.normalizePath)(filePath);
         const absoluteFilePath = path_1.default.join(repoPath, normalizedFilePath);
@@ -63,8 +64,33 @@ exports.runTestTool = {
             };
         }
         try {
-            const { projects } = await (0, test_run_1.listProjectsAndTests)(repoPath);
-            if (!projects.includes(project)) {
+            const { projects, projectsToTestCases } = await (0, test_run_1.listProjectsAndTests)(repoPath);
+            if (!project) {
+                // Auto-select project if the test file belongs to only one project
+                const matchingProjects = projects.filter((p) => projectsToTestCases[p]?.some((tc) => normalizedFilePath.endsWith(tc.filePath)));
+                if (matchingProjects.length === 1) {
+                    project = matchingProjects[0];
+                    logger?.info("Auto-selected project", { project }) ||
+                        console.log("Auto-selected project:", project);
+                }
+                else if (matchingProjects.length === 0) {
+                    const errorMsg = `Could not find any project containing the test file: ${normalizedFilePath}. Valid project names are: ${projects.join(", ")}`;
+                    logger?.error("No matching project found", {
+                        filePath: normalizedFilePath,
+                        validProjects: projects,
+                    }) || console.error("No matching project found:", normalizedFilePath);
+                    return { isError: true, result: errorMsg };
+                }
+                else {
+                    const errorMsg = `Multiple projects contain this test file: ${matchingProjects.join(", ")}. Please specify one using the 'project' parameter.`;
+                    logger?.error("Multiple matching projects", {
+                        filePath: normalizedFilePath,
+                        matchingProjects,
+                    }) || console.error("Multiple matching projects:", matchingProjects);
+                    return { isError: true, result: errorMsg };
+                }
+            }
+            else if (!projects.includes(project)) {
                 const errorMsg = `Invalid project name: ${project}. Valid project names are: ${projects.join(", ")}`;
                 logger?.error("Invalid project name", {
                     project,
@@ -84,17 +110,17 @@ exports.runTestTool = {
                 result: `Error reading playwright config: ${error}`,
             };
         }
-        if (input.buildUrl) {
+        if (input.envSlug) {
             envOverrides = {
                 ...envOverrides,
-                BUILD_URL: input.buildUrl,
+                ENV_SLUG: input.envSlug,
             };
         }
         try {
             logger?.info("Starting test execution", {
                 testName: input.testName,
                 filePath: normalizedFilePath,
-                project: input.project,
+                project,
                 suites: input.suites,
             }) || console.log("Starting test execution:", input.testName);
             const result = await (0, test_run_1.runSingleTest)({
@@ -102,7 +128,7 @@ exports.runTestTool = {
                 suites: input.suites,
                 filePath: normalizedFilePath,
                 repoDir: repoPath,
-                projects: [input.project],
+                projects: [project],
                 envOverrides,
                 stdout: logger?.createPrefixedStream?.("stdout"),
                 stderr: logger?.createPrefixedStream?.("stderr"),

package/dist/tools/safe-bash/index.d.ts.map

@@ -1 +1 @@
-{"version":3,"file":"index.d.ts","sourceRoot":"","sources":["../../../src/tools/safe-bash/index.ts"],"names":[],"mappings":"AAAA,OAAO,KAAK,EAAE,IAAI,EAAc,MAAM,uCAAuC,CAAC;AA2H9E,eAAO,MAAM,YAAY,EAAE,IA6E1B,CAAC"}
+{"version":3,"file":"index.d.ts","sourceRoot":"","sources":["../../../src/tools/safe-bash/index.ts"],"names":[],"mappings":"AAAA,OAAO,KAAK,EAAE,IAAI,EAAc,MAAM,uCAAuC,CAAC;AA0I9E,eAAO,MAAM,YAAY,EAAE,IAoF1B,CAAC"}

package/dist/tools/safe-bash/index.js

@@ -3,8 +3,8 @@ Object.defineProperty(exports, "__esModule", { value: true });
 exports.safeBashTool = void 0;
 const child_process_1 = require("child_process");
 const util_1 = require("util");
-const web_1 = require("../../bin/utils/platform/web");
 const safe_bash_1 = require("../definitions/safe-bash");
+const validators_1 = require("../utils/validators");
 const execAsync = (0, util_1.promisify)(child_process_1.exec);
 const DANGEROUS_PATTERNS = [
     /git\s+checkout/i,
@@ -47,6 +47,19 @@ async function getGitStatus(cwd) {
         return { staged: "Unable to get status", unstaged: "Unable to get status" };
     }
 }
+function getChangedFilesFromGitStatus(gitStatus) {
+    const files = [];
+    for (const section of [gitStatus.staged, gitStatus.unstaged]) {
+        for (const line of section.split("\n")) {
+            // Format: "M\tfile.ts", "A\tfile.ts", "??\tfile.ts"
+            const match = line.match(/^[A-Z?]+\t(.+)$/);
+            if (match?.[1]) {
+                files.push(match[1]);
+            }
+        }
+    }
+    return files;
+}
 function containsDangerousPattern(script) {
     for (const pattern of DANGEROUS_PATTERNS) {
         if (pattern.test(script)) {
@@ -55,7 +68,7 @@ function containsDangerousPattern(script) {
     }
     return null;
 }
-function formatResult({ stdout, stderr, gitStatus, tscResult, }) {
+function formatResult({ stdout, stderr, gitStatus, validationResult, }) {
     const parts = [];
     if (stdout.trim()) {
         parts.push(`## stdout\n\`\`\`\n${stdout.trim()}\n\`\`\``);
@@ -63,13 +76,13 @@ function formatResult({ stdout, stderr, gitStatus, tscResult, }) {
     if (stderr.trim()) {
         parts.push(`## stderr\n\`\`\`\n${stderr.trim()}\n\`\`\``);
     }
-    if (tscResult) {
-        if (!tscResult.success) {
-            const errors = tscResult.errors;
-            parts.push(`## tsc result: Errors\n\`\`\`\n${errors.join("\n")}\n\`\`\``);
+    if (validationResult) {
+        if (!validationResult.success) {
+            const errors = validationResult.errors;
+            parts.push(`## Validation result: Errors\n\`\`\`\n${errors.join("\n")}\n\`\`\``);
         }
         else {
-            parts.push(`## tsc result: No errors`);
+            parts.push(`## Validation result: No errors`);
         }
     }
     parts.push(`## Staged (will be committed)\n\`\`\`\n${gitStatus.staged}\n\`\`\``);
@@ -78,7 +91,7 @@ function formatResult({ stdout, stderr, gitStatus, tscResult, }) {
 }
 exports.safeBashTool = {
     ...safe_bash_1.safeBashTool,
-    execute: async ({ input, repoPath, logger }) => {
+    execute: async ({ input, repoPath, logger, environmentOverrides, }) => {
         const { script, timeout = 60000 } = input;
         const initialBranch = await getCurrentBranch(repoPath);
         if (!initialBranch) {
@@ -101,6 +114,7 @@ exports.safeBashTool = {
                 cwd: repoPath,
                 timeout,
                 shell: "/bin/bash",
+                env: { ...process.env, ...environmentOverrides },
             });
             stdout = result.stdout;
             stderr = result.stderr;
@@ -130,14 +144,15 @@ exports.safeBashTool = {
             };
         }
         const gitStatus = await getGitStatus(repoPath);
-        const tscResult = await (0, web_1.runTypescriptCompiler)(repoPath);
+        const changedFiles = getChangedFilesFromGitStatus(gitStatus);
+        const validationResult = await (0, validators_1.runValidators)(repoPath, changedFiles);
         return {
             isError: false,
             result: formatResult({
                 stdout,
                 stderr,
                 gitStatus,
-                tscResult,
+                validationResult,
             }),
         };
     },

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,EAGV,IAAI,EACL,MAAM,uCAAuC,CAAC;AA0B/C,eAAO,MAAM,4BAA4B,EAAE,IAsP1C,CAAC"}
+{"version":3,"file":"test-gen-browser.d.ts","sourceRoot":"","sources":["../../src/tools/test-gen-browser.ts"],"names":[],"mappings":"AAQA,OAAO,KAAK,EAGV,IAAI,EACL,MAAM,uCAAuC,CAAC;AAkB/C,eAAO,MAAM,4BAA4B,EAAE,IA2Q1C,CAAC"}

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

@@ -4,14 +4,14 @@ var __importDefault = (this && this.__importDefault) || function (mod) {
 };
 Object.defineProperty(exports, "__esModule", { value: true });
 exports.generateTestWithBrowserAgent = void 0;
+const model_1 = require("@empiricalrun/cua/model");
+const pw_codegen_1 = require("@empiricalrun/cua/pw-codegen");
+const pw_pause_1 = require("@empiricalrun/cua/pw-codegen/pw-pause");
 const test_run_1 = require("@empiricalrun/test-run");
 const fs_1 = __importDefault(require("fs"));
 const path_1 = __importDefault(require("path"));
 const run_1 = require("../agent/browsing/run");
 const utils_1 = require("../agent/browsing/utils");
-const model_1 = require("../agent/cua/model");
-const pw_codegen_1 = require("../agent/cua/pw-codegen");
-const pw_pause_1 = require("../agent/cua/pw-codegen/pw-pause");
 const utils_2 = require("../artifacts/utils");
 const web_1 = require("../bin/utils/platform/web");
 const scenarios_1 = require("../bin/utils/scenarios");
@@ -21,7 +21,7 @@ const helpers_1 = require("./file-operations/shared/helpers");
 exports.generateTestWithBrowserAgent = {
     ...test_gen_browser_1.generateTestWithBrowserAgent,
     execute: async ({ input, repoPath, trace, collectArtifacts, environmentOverrides = {}, chatSession, logger, }) => {
-        const { project } = input;
+        let { project } = input;
         // Normalize the path to support both /repo/... and relative paths
         const normalizedFilePath = (0, helpers_1.normalizePath)(input.filePath);
         const absoluteFilePath = path_1.default.join(repoPath, normalizedFilePath);
@@ -32,8 +32,27 @@ exports.generateTestWithBrowserAgent = {
             };
         }
         try {
-            const { projects } = await (0, test_run_1.listProjectsAndTests)(repoPath);
-            if (!projects.includes(project)) {
+            const { projects, projectsToTestCases } = await (0, test_run_1.listProjectsAndTests)(repoPath);
+            if (!project) {
+                const matchingProjects = projects.filter((p) => projectsToTestCases[p]?.some((tc) => normalizedFilePath.endsWith(tc.filePath)));
+                if (matchingProjects.length === 1) {
+                    project = matchingProjects[0];
+                    logger.info("Auto-selected project", { project });
+                }
+                else if (matchingProjects.length === 0) {
+                    return {
+                        isError: true,
+                        result: `Could not find any project containing the test file: ${normalizedFilePath}. Valid project names are: ${projects.join(", ")}`,
+                    };
+                }
+                else {
+                    return {
+                        isError: true,
+                        result: `Multiple projects contain this test file: ${matchingProjects.join(", ")}. Please specify one using the 'project' parameter.`,
+                    };
+                }
+            }
+            else if (!projects.includes(project)) {
                 return {
                     isError: true,
                     result: `Invalid project name: ${project}. Valid project names are: ${projects.join(", ")}`,
@@ -96,10 +115,10 @@ exports.generateTestWithBrowserAgent = {
             // Running browser agent in headed implies using page.pause method for codegen
             RUN_PLAYWRIGHT_HEADED: "true",
         };
-        if (input.buildUrl) {
+        if (input.envSlug) {
             envOverrides = {
                 ...envOverrides,
-                BUILD_URL: input.buildUrl,
+                ENV_SLUG: input.envSlug,
             };
         }
         const testGenToken = (0, scenarios_1.buildTokenFromOptions)({ name: testName, file: normalizedFilePath }, { useComputerUseAgent: true });

package/dist/tools/utils/validators.d.ts

@@ -0,0 +1,7 @@
+interface ValidationResult {
+    success: boolean;
+    errors: string[];
+}
+export declare function runValidators(repoDir: string, changedFiles: string[]): Promise<ValidationResult>;
+export {};
+//# sourceMappingURL=validators.d.ts.map

package/dist/tools/utils/validators.d.ts.map

@@ -0,0 +1 @@
+{"version":3,"file":"validators.d.ts","sourceRoot":"","sources":["../../../src/tools/utils/validators.ts"],"names":[],"mappings":"AAcA,UAAU,gBAAgB;IACxB,OAAO,EAAE,OAAO,CAAC;IACjB,MAAM,EAAE,MAAM,EAAE,CAAC;CAClB;AAgED,wBAAsB,aAAa,CACjC,OAAO,EAAE,MAAM,EACf,YAAY,EAAE,MAAM,EAAE,GACrB,OAAO,CAAC,gBAAgB,CAAC,CAgC3B"}

package/dist/tools/utils/validators.js

@@ -0,0 +1,93 @@
+"use strict";
+var __importDefault = (this && this.__importDefault) || function (mod) {
+    return (mod && mod.__esModule) ? mod : { "default": mod };
+};
+Object.defineProperty(exports, "__esModule", { value: true });
+exports.runValidators = runValidators;
+const environments_1 = require("@empiricalrun/shared-types/environments");
+const owners_1 = require("@empiricalrun/shared-types/owners");
+const fs_1 = __importDefault(require("fs"));
+const js_yaml_1 = __importDefault(require("js-yaml"));
+const path_1 = __importDefault(require("path"));
+const web_1 = require("../../bin/utils/platform/web");
+function validateOwnersYaml(repoDir) {
+    const filePath = path_1.default.join(repoDir, owners_1.OWNERS_FILE_PATH);
+    if (!fs_1.default.existsSync(filePath)) {
+        return {
+            success: false,
+            errors: [`OWNERS.yaml file not found at ${owners_1.OWNERS_FILE_PATH}`],
+        };
+    }
+    const content = fs_1.default.readFileSync(filePath, "utf8");
+    let parsed;
+    try {
+        parsed = js_yaml_1.default.load(content);
+    }
+    catch (e) {
+        return {
+            success: false,
+            errors: [`Failed to parse OWNERS.yaml: ${e?.message}`],
+        };
+    }
+    const result = (0, owners_1.parseOwnersConfig)(parsed);
+    if ("errors" in result) {
+        return {
+            success: false,
+            errors: result.errors,
+        };
+    }
+    return { success: true, errors: [] };
+}
+function validateEnvironmentsYaml(repoDir) {
+    const filePath = path_1.default.join(repoDir, environments_1.ENVIRONMENTS_FILE_PATH);
+    if (!fs_1.default.existsSync(filePath)) {
+        return {
+            success: false,
+            errors: [`ENVIRONMENTS.yaml file not found at ${environments_1.ENVIRONMENTS_FILE_PATH}`],
+        };
+    }
+    const content = fs_1.default.readFileSync(filePath, "utf8");
+    let parsed;
+    try {
+        parsed = js_yaml_1.default.load(content);
+    }
+    catch (e) {
+        return {
+            success: false,
+            errors: [`Failed to parse ENVIRONMENTS.yaml: ${e?.message}`],
+        };
+    }
+    const result = (0, environments_1.parseEnvironmentsConfig)(parsed);
+    if ("errors" in result) {
+        return {
+            success: false,
+            errors: result.errors,
+        };
+    }
+    return { success: true, errors: [] };
+}
+async function runValidators(repoDir, changedFiles) {
+    const allErrors = [];
+    const tscResult = await (0, web_1.runTypescriptCompiler)(repoDir);
+    if (!tscResult.success) {
+        allErrors.push(...tscResult.errors);
+    }
+    const hasOwnersChange = changedFiles.some((f) => f === owners_1.OWNERS_FILE_PATH || f.endsWith("OWNERS.yaml"));
+    if (hasOwnersChange) {
+        const ownersResult = validateOwnersYaml(repoDir);
+        if (!ownersResult.success) {
+            allErrors.push(...ownersResult.errors);
+        }
+    }
+    const hasEnvironmentsChange = changedFiles.some((f) => f === environments_1.ENVIRONMENTS_FILE_PATH || f.endsWith("ENVIRONMENTS.yaml"));
+    if (hasEnvironmentsChange) {
+        const environmentsResult = validateEnvironmentsYaml(repoDir);
+        if (!environmentsResult.success) {
+            allErrors.push(...environmentsResult.errors);
+        }
+    }
+    return {
+        success: allErrors.length === 0,
+        errors: allErrors,
+    };
+}

package/dist/video-core/index.d.ts.map

@@ -1 +1 @@
-{"version":3,"file":"index.d.ts","sourceRoot":"","sources":["../../src/video-core/index.ts"],"names":[],"mappings":"AAAA,OAAO,EAGL,cAAc,EACd,KAAK,EACN,MAAM,uCAAuC,CAAC;AAC/C,OAAO,EAEL,mBAAmB,EACnB,qBAAqB,EACtB,MAAM,qCAAqC,CAAC;AAC7C,OAAO,EAGL,KAAK,eAAe,EACrB,MAAM,WAAW,CAAC;AAgBnB,KAAK,eAAe,GAAG;IACrB,GAAG,EAAE,MAAM,CAAC;IACZ,SAAS,EAAE,MAAM,CAAC;IAClB,MAAM,EAAE,mBAAmB,CAAC;IAC5B,gBAAgB,EAAE,MAAM,CAAC;IACzB,UAAU,CAAC,EAAE,OAAO,CAAC;IACrB,MAAM,CAAC,EAAE,eAAe,CAAC;CAC1B,CAAC;AAEF,wBAAsB,gBAAgB,CAAC,EACrC,GAAG,EACH,SAAS,EACT,MAAM,EACN,gBAAgB,EAChB,UAAU,EACV,MAAM,GACP,EAAE,eAAe;;;;;GAqIjB"}
+{"version":3,"file":"index.d.ts","sourceRoot":"","sources":["../../src/video-core/index.ts"],"names":[],"mappings":"AACA,OAAO,EAGL,cAAc,EACd,KAAK,EACN,MAAM,uCAAuC,CAAC;AAC/C,OAAO,EAEL,mBAAmB,EACnB,qBAAqB,EACtB,MAAM,qCAAqC,CAAC;AAC7C,OAAO,EAGL,KAAK,eAAe,EACrB,MAAM,WAAW,CAAC;AAgBnB,KAAK,eAAe,GAAG;IACrB,GAAG,EAAE,MAAM,CAAC;IACZ,SAAS,EAAE,MAAM,CAAC;IAClB,MAAM,EAAE,mBAAmB,CAAC;IAC5B,gBAAgB,EAAE,MAAM,CAAC;IACzB,UAAU,CAAC,EAAE,OAAO,CAAC;IACrB,MAAM,CAAC,EAAE,eAAe,CAAC;CAC1B,CAAC;AAEF,wBAAsB,gBAAgB,CAAC,EACrC,GAAG,EACH,SAAS,EACT,MAAM,EACN,gBAAgB,EAChB,UAAU,EACV,MAAM,GACP,EAAE,eAAe;;;;;GA0IjB"}

package/dist/video-core/index.js

@@ -1,6 +1,7 @@
 "use strict";
 Object.defineProperty(exports, "__esModule", { value: true });
 exports.runVideoAnalysis = runVideoAnalysis;
+const node_fs_1 = require("node:fs");
 const videostil_1 = require("videostil");
 const constants_1 = require("../constants");
 const model_limits_1 = require("./model-limits");
@@ -98,6 +99,10 @@ async function runVideoAnalysis({ url, videoHash, params, workingDirectory, skip
         await Promise.all(uploadPromises);
     }
     const usage = (0, utils_1.aggregateUsageFromMessages)(allMessages);
+    // Clean up local working directory (frames are already in memory / uploaded)
+    await node_fs_1.promises
+        .rm(workingDirectory, { recursive: true, force: true })
+        .catch(() => { });
     return {
         result,
         error: finalError,

package/package.json

@@ -1,6 +1,6 @@
 {
   "name": "@empiricalrun/test-gen",
-  "version": "0.79.4",
+  "version": "0.79.6",
   "publishConfig": {
     "registry": "https://registry.npmjs.org/",
     "access": "public"
@@ -10,10 +10,6 @@
   },
   "main": "dist/index.js",
   "exports": {
-    "./agent/master/run": {
-      "types": "./dist/agent/master/run.d.ts",
-      "default": "./dist/agent/master/run.js"
-    },
     "./chat": {
       "types": "./dist/agent/chat/exports.d.ts",
       "default": "./dist/agent/chat/exports.js"
@@ -73,6 +69,7 @@
     "dotenv": "^16.4.5",
     "express": "^4.19.2",
     "ignore": "^5.3.1",
+    "js-yaml": "^4.1.0",
     "inquirer": "^12.4.2",
     "istextorbinary": "^9.5.0",
     "lodash.isequal": "^4.5.0",
@@ -80,7 +77,7 @@
     "minimatch": "^10.0.1",
     "nanoid": "^5.0.7",
     "open": "^10.1.2",
-    "openai": "^6.10.0",
+    "openai": "^6.27.0",
     "ora": "^8.1.0",
     "picocolors": "^1.0.1",
     "pixelmatch": "^6.0.0",
@@ -92,12 +89,15 @@
     "typescript": "^5.3.3",
     "zod": "^4.0.1",
     "@empiricalrun/ast-parser": "^0.0.10",
-    "@empiricalrun/llm": "^0.25.2",
+    "@empiricalrun/cua": "^0.3.0",
+    "@empiricalrun/dashboard-client": "^0.2.0",
+    "@empiricalrun/shared-types": "0.12.1",
+    "@empiricalrun/llm": "^0.26.0",
     "@empiricalrun/r2-uploader": "^0.9.1",
     "@empiricalrun/trace-utils": "^0.1.0",
     "@empiricalrun/reporter": "^0.27.0",
     "@empiricalrun/test-run": "^0.14.2",
-    "videostil": "0.3.4"
+    "videostil": "0.3.5"
   },
   "devDependencies": {
     "@playwright/test": "1.57.0",
@@ -106,6 +106,7 @@
     "@types/detect-port": "^1.3.5",
     "@types/express": "^4.17.21",
     "@types/js-levenshtein": "^1.1.3",
+    "@types/js-yaml": "^4.0.9",
     "@types/lodash.isequal": "^4.5.8",
     "@types/mime": "3.0.0",
     "@types/node": "^24.3.0",
@@ -114,8 +115,7 @@
     "js-levenshtein": "^1.1.6",
     "playwright": "1.57.0",
     "serve-handler": "^6.1.6",
-    "ts-patch": "^3.3.0",
-    "@empiricalrun/shared-types": "0.12.0"
+    "ts-patch": "^3.3.0"
   },
   "scripts": {
     "dev": "tspc --build --watch",

package/tsconfig.tsbuildinfo

@@ -1 +1 @@
-{"root":["./src/index.ts","./src/logger.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/text-content.ts","./src/actions/constants/index.ts","./src/actions/utils/index.ts","./src/agent/cli.ts","./src/agent/index.ts","./src/agent/base/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/state.ts","./src/agent/chat/types.ts","./src/agent/chat/utils.ts","./src/agent/chat/prompt/pw-utils-docs.ts","./src/agent/chat/prompt/repo.ts","./src/agent/chat/prompt/test-case-def.ts","./src/agent/chat/prompt/trace-utils-docs.ts","./src/agent/code-review/index.ts","./src/agent/code-review/types.ts","./src/agent/code-review/xml-parser.ts","./src/agent/code-review/executor/index.ts","./src/agent/cua/computer.ts","./src/agent/cua/index.ts","./src/agent/cua/model.ts","./src/agent/cua/pw-codegen/element-from-point.ts","./src/agent/cua/pw-codegen/factory.ts","./src/agent/cua/pw-codegen/index.ts","./src/agent/cua/pw-codegen/types.ts","./src/agent/cua/pw-codegen/version.ts","./src/agent/cua/pw-codegen/pw-event-sink/index.ts","./src/agent/cua/pw-codegen/pw-pause/for-recorder.ts","./src/agent/cua/pw-codegen/pw-pause/index.ts","./src/agent/cua/pw-codegen/pw-pause/ipc.ts","./src/agent/cua/pw-codegen/pw-pause/patch.ts","./src/agent/cua/pw-codegen/pw-pause/types.ts","./src/agent/cua/pw-codegen/pw-pause/utils.ts","./src/agent/fast-triage/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/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/planner/run-time-planner.ts","./src/agent/planner/run.ts","./src/agent/triage/index.ts","./src/artifacts/index.ts","./src/artifacts/utils.ts","./src/auth/cli-auth.ts","./src/auth/index.ts","./src/auth/token-store.ts","./src/bin/environments.ts","./src/bin/index.ts","./src/bin/setup.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/dashboard/client.ts","./src/dashboard/index.ts","./src/dashboard/tool-response-from-sandbox.ts","./src/dashboard/tool-response.ts","./src/dashboard/totp.ts","./src/dashboard/types.ts","./src/errors/index.ts","./src/file/client.ts","./src/file/server.ts","./src/file-info/index.ts","./src/file-info/adapters/file-system/index.ts","./src/file-info/adapters/file-system/reader.ts","./src/file-info/adapters/github/index.ts","./src/file-info/adapters/github/reader.ts","./src/generate-summary/frame-sampling.ts","./src/generate-summary/generate-error-stack-summary.ts","./src/generate-summary/generate-failed-step-screenshot-diff-summary.ts","./src/generate-summary/generate-grouped-summary.ts","./src/generate-summary/merge-summary.ts","./src/generate-summary/pick-videos-for-comparison.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/recorder/env-variables.ts","./src/recorder/index.ts","./src/recorder/request.ts","./src/recorder/temp-files.ts","./src/recorder/upload.ts","./src/recorder/validation.ts","./src/telemetry/index.ts","./src/test-build/index.ts","./src/tools/diagnosis-fetcher.ts","./src/tools/index.ts","./src/tools/list-environments.ts","./src/tools/run-test.ts","./src/tools/test-gen-browser.ts","./src/tools/analyse-video/index.ts","./src/tools/create-pull-request/index.ts","./src/tools/create-pull-request/utils.ts","./src/tools/definitions/analyse-video.ts","./src/tools/definitions/delete-file.ts","./src/tools/definitions/download-build.ts","./src/tools/definitions/grep.ts","./src/tools/definitions/list-tests-and-projects.ts","./src/tools/definitions/merge-conflicts.ts","./src/tools/definitions/publish-markdown-report.ts","./src/tools/definitions/rename-file.ts","./src/tools/definitions/run-test.ts","./src/tools/definitions/safe-bash.ts","./src/tools/definitions/str_replace_editor.ts","./src/tools/definitions/test-gen-browser.ts","./src/tools/definitions/trace-dot-zip.ts","./src/tools/definitions/upgrade-packages.ts","./src/tools/definitions/utils.ts","./src/tools/delete-file/index.ts","./src/tools/download-build/index.ts","./src/tools/executor/base.ts","./src/tools/executor/index.ts","./src/tools/executor/types.ts","./src/tools/executor/utils/checkpoint.ts","./src/tools/executor/utils/git.ts","./src/tools/executor/utils/index.ts","./src/tools/executor/utils/pr-description.ts","./src/tools/fetch-file/index.ts","./src/tools/fetch-file/utils.ts","./src/tools/fetch-session-diff/index.ts","./src/tools/file-operations/create.ts","./src/tools/file-operations/index.ts","./src/tools/file-operations/insert.ts","./src/tools/file-operations/replace.ts","./src/tools/file-operations/shared/helpers.ts","./src/tools/file-operations/view/index.ts","./src/tools/firecrawl-scrape/index.ts","./src/tools/grep/index.ts","./src/tools/grep/types.ts","./src/tools/grep/ripgrep/index.ts","./src/tools/issues-v1/index.ts","./src/tools/issues-v1/list-issues.ts","./src/tools/issues-v1/utils.ts","./src/tools/issues-v2/create-issue.ts","./src/tools/issues-v2/set-issue-description.ts","./src/tools/issues-v2/update-issue.ts","./src/tools/list-tests-and-projects/index.ts","./src/tools/merge-conflicts/index.ts","./src/tools/publish-markdown-report/index.ts","./src/tools/rename-file/index.ts","./src/tools/safe-bash/index.ts","./src/tools/test-run-fetcher/index.ts","./src/tools/test-run-fetcher/types.ts","./src/tools/trace-dot-zip/index.ts","./src/tools/trace-dot-zip/types.ts","./src/tools/trace-dot-zip/utils/console-trace.ts","./src/tools/trace-dot-zip/utils/extract-screenshots.ts","./src/tools/trace-dot-zip/utils/extract-steps.ts","./src/tools/trace-dot-zip/utils/extract-zip.ts","./src/tools/trace-dot-zip/utils/network-trace.ts","./src/tools/triage-summary/index.ts","./src/tools/triage-summary/types.ts","./src/tools/triage-summary/utils.ts","./src/tools/upgrade-packages/index.ts","./src/tools/upgrade-packages/utils.ts","./src/tools/utils/queue.ts","./src/tools/utils/validate-schema.ts","./src/types/handlebars.d.ts","./src/types/index.ts","./src/utils/SQSClient.ts","./src/utils/env.ts","./src/utils/exec.ts","./src/utils/file.ts","./src/utils/hash.ts","./src/utils/html.ts","./src/utils/index.ts","./src/utils/json.ts","./src/utils/model.ts","./src/utils/playwright-test-id.ts","./src/utils/repo-tree.ts","./src/utils/slug.ts","./src/utils/string.ts","./src/utils/stripAnsi.ts","./src/utils/url-validation.ts","./src/utils/dedup/dedup-image.ts","./src/utils/dedup/find-threshold.ts","./src/video-core/index.ts","./src/video-core/model-limits.ts","./src/video-core/storage-manager.ts","./src/video-core/types.ts","./src/video-core/utils.ts"],"version":"5.8.3"}
+{"root":["./src/index.ts","./src/logger.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/text-content.ts","./src/actions/constants/index.ts","./src/actions/utils/index.ts","./src/agent/cli.ts","./src/agent/index.ts","./src/agent/base/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/state.ts","./src/agent/chat/types.ts","./src/agent/chat/utils.ts","./src/agent/chat/prompt/pw-utils-docs.ts","./src/agent/chat/prompt/repo.ts","./src/agent/chat/prompt/test-case-def.ts","./src/agent/chat/prompt/trace-utils-docs.ts","./src/agent/code-review/index.ts","./src/agent/code-review/types.ts","./src/agent/code-review/xml-parser.ts","./src/agent/code-review/executor/index.ts","./src/agent/fast-triage/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/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/planner/run-time-planner.ts","./src/agent/planner/run.ts","./src/agent/triage/index.ts","./src/artifacts/index.ts","./src/artifacts/utils.ts","./src/auth/cli-auth.ts","./src/auth/index.ts","./src/auth/token-store.ts","./src/bin/index.ts","./src/bin/setup.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/dashboard/client.ts","./src/dashboard/index.ts","./src/dashboard/tool-response-from-sandbox.ts","./src/dashboard/tool-response.ts","./src/dashboard/totp.ts","./src/dashboard/types.ts","./src/errors/index.ts","./src/file/client.ts","./src/file/server.ts","./src/file-info/index.ts","./src/file-info/adapters/file-system/index.ts","./src/file-info/adapters/file-system/reader.ts","./src/file-info/adapters/github/index.ts","./src/file-info/adapters/github/reader.ts","./src/generate-summary/frame-sampling.ts","./src/generate-summary/generate-error-stack-summary.ts","./src/generate-summary/generate-failed-step-screenshot-diff-summary.ts","./src/generate-summary/generate-grouped-summary.ts","./src/generate-summary/merge-summary.ts","./src/generate-summary/pick-videos-for-comparison.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/recorder/env-variables.ts","./src/recorder/index.ts","./src/recorder/request.ts","./src/recorder/temp-files.ts","./src/recorder/upload.ts","./src/recorder/validation.ts","./src/telemetry/index.ts","./src/test-build/index.ts","./src/tools/diagnosis-fetcher.ts","./src/tools/index.ts","./src/tools/list-environments.ts","./src/tools/run-test.ts","./src/tools/test-gen-browser.ts","./src/tools/analyse-video/index.ts","./src/tools/api-client/index.ts","./src/tools/create-pull-request/index.ts","./src/tools/create-pull-request/utils.ts","./src/tools/definitions/analyse-video.ts","./src/tools/definitions/delete-file.ts","./src/tools/definitions/download-build.ts","./src/tools/definitions/grep.ts","./src/tools/definitions/list-tests-and-projects.ts","./src/tools/definitions/merge-conflicts.ts","./src/tools/definitions/publish-markdown-report.ts","./src/tools/definitions/rename-file.ts","./src/tools/definitions/run-test.ts","./src/tools/definitions/safe-bash.ts","./src/tools/definitions/str_replace_editor.ts","./src/tools/definitions/test-gen-browser.ts","./src/tools/definitions/trace-dot-zip.ts","./src/tools/definitions/utils.ts","./src/tools/delete-file/index.ts","./src/tools/download-build/index.ts","./src/tools/executor/base.ts","./src/tools/executor/index.ts","./src/tools/executor/types.ts","./src/tools/executor/utils/checkpoint.ts","./src/tools/executor/utils/git.ts","./src/tools/executor/utils/index.ts","./src/tools/executor/utils/pr-description.ts","./src/tools/fetch-file/index.ts","./src/tools/fetch-file/utils.ts","./src/tools/fetch-session-diff/index.ts","./src/tools/file-operations/create.ts","./src/tools/file-operations/index.ts","./src/tools/file-operations/insert.ts","./src/tools/file-operations/replace.ts","./src/tools/file-operations/shared/helpers.ts","./src/tools/file-operations/view/index.ts","./src/tools/firecrawl-scrape/index.ts","./src/tools/grep/index.ts","./src/tools/grep/types.ts","./src/tools/grep/ripgrep/index.ts","./src/tools/issues-v1/index.ts","./src/tools/issues-v1/list-issues.ts","./src/tools/issues-v1/utils.ts","./src/tools/issues-v2/create-issue.ts","./src/tools/issues-v2/set-issue-description.ts","./src/tools/issues-v2/update-issue.ts","./src/tools/list-tests-and-projects/index.ts","./src/tools/merge-conflicts/index.ts","./src/tools/publish-markdown-report/index.ts","./src/tools/rename-file/index.ts","./src/tools/safe-bash/index.ts","./src/tools/test-run-fetcher/index.ts","./src/tools/test-run-fetcher/types.ts","./src/tools/trace-dot-zip/index.ts","./src/tools/trace-dot-zip/types.ts","./src/tools/trace-dot-zip/utils/console-trace.ts","./src/tools/trace-dot-zip/utils/extract-screenshots.ts","./src/tools/trace-dot-zip/utils/extract-steps.ts","./src/tools/trace-dot-zip/utils/extract-zip.ts","./src/tools/trace-dot-zip/utils/network-trace.ts","./src/tools/triage-summary/index.ts","./src/tools/triage-summary/types.ts","./src/tools/triage-summary/utils.ts","./src/tools/utils/queue.ts","./src/tools/utils/validate-schema.ts","./src/tools/utils/validators.ts","./src/types/handlebars.d.ts","./src/types/index.ts","./src/utils/SQSClient.ts","./src/utils/env.ts","./src/utils/exec.ts","./src/utils/file.ts","./src/utils/hash.ts","./src/utils/html.ts","./src/utils/index.ts","./src/utils/json.ts","./src/utils/model.ts","./src/utils/playwright-test-id.ts","./src/utils/repo-tree.ts","./src/utils/slug.ts","./src/utils/string.ts","./src/utils/stripAnsi.ts","./src/utils/url-validation.ts","./src/utils/dedup/dedup-image.ts","./src/utils/dedup/find-threshold.ts","./src/video-core/index.ts","./src/video-core/model-limits.ts","./src/video-core/storage-manager.ts","./src/video-core/types.ts","./src/video-core/utils.ts"],"version":"5.8.3"}

package/dist/agent/cua/computer.d.ts

@@ -1,14 +0,0 @@
-import { ResponseComputerToolCall } from "openai/resources/responses/responses.mjs";
-import type { Page } from "playwright";
-import { BasePlaywrightCodegen } from "./pw-codegen/types";
-type OpenAIComputerAction = ResponseComputerToolCall.Click | ResponseComputerToolCall.DoubleClick | ResponseComputerToolCall.Drag | ResponseComputerToolCall.Keypress | ResponseComputerToolCall.Move | ResponseComputerToolCall.Screenshot | ResponseComputerToolCall.Scroll | ResponseComputerToolCall.Type | ResponseComputerToolCall.Wait;
-export declare function getScreenshot(page: Page): Promise<string>;
-export declare function executeModelAction(page: Page, action: OpenAIComputerAction | {
-    type: "goto";
-    url: string;
-}, codegen: BasePlaywrightCodegen): Promise<{
-    actionSummary: string;
-    actionCode: string;
-}>;
-export {};
-//# sourceMappingURL=computer.d.ts.map

package/dist/agent/cua/computer.d.ts.map

@@ -1 +0,0 @@
-{"version":3,"file":"computer.d.ts","sourceRoot":"","sources":["../../../src/agent/cua/computer.ts"],"names":[],"mappings":"AAAA,OAAO,EAAE,wBAAwB,EAAE,MAAM,0CAA0C,CAAC;AACpF,OAAO,KAAK,EAAE,IAAI,EAAE,MAAM,YAAY,CAAC;AAEvC,OAAO,EAAE,qBAAqB,EAAE,MAAM,oBAAoB,CAAC;AAE3D,KAAK,oBAAoB,GACrB,wBAAwB,CAAC,KAAK,GAC9B,wBAAwB,CAAC,WAAW,GACpC,wBAAwB,CAAC,IAAI,GAC7B,wBAAwB,CAAC,QAAQ,GACjC,wBAAwB,CAAC,IAAI,GAC7B,wBAAwB,CAAC,UAAU,GACnC,wBAAwB,CAAC,MAAM,GAC/B,wBAAwB,CAAC,IAAI,GAC7B,wBAAwB,CAAC,IAAI,CAAC;AAElC,wBAAsB,aAAa,CAAC,IAAI,EAAE,IAAI,mBAO7C;AAgCD,wBAAsB,kBAAkB,CACtC,IAAI,EAAE,IAAI,EACV,MAAM,EAAE,oBAAoB,GAAG;IAAE,IAAI,EAAE,MAAM,CAAC;IAAC,GAAG,EAAE,MAAM,CAAA;CAAE,EAC5D,OAAO,EAAE,qBAAqB,GAC7B,OAAO,CAAC;IACT,aAAa,EAAE,MAAM,CAAC;IACtB,UAAU,EAAE,MAAM,CAAC;CACpB,CAAC,CAoKD"}