@empiricalrun/test-gen 0.76.0 → 0.77.0

package/dist/tools/index.js

@@ -1,31 +1,30 @@
 "use strict";
 Object.defineProperty(exports, "__esModule", { value: true });
-exports.allToolsDefinitions = exports.viewFailedTestRunReportTool = exports.upgradePackagesTool = exports.updateIssueTool = exports.traceDotZipTool = exports.sendTriageSummaryTool = exports.runTestTool = exports.reviewPullRequestTool = exports.renameFileTool = exports.mergeConflictsTool = exports.listTestsForProjectTool = exports.listProjectsTool = exports.listIssuesTool = exports.listEnvironmentsTool = exports.grepTool = exports.generateTestWithBrowserAgent = exports.fetchVideoAnalysis = exports.fetchTestRunReportTool = exports.fetchSessionDiffTool = exports.fetchLastSuccessfulTestRunTool = exports.fetchFileTool = exports.fetchDiagnosisReportTool = exports.extractFramesFromVideo = exports.downloadBuildTool = exports.deleteFileTool = exports.createPullRequestTool = exports.createIssueTool = exports.triageOnlyTools = exports.testGenerationTools = exports.commonTools = void 0;
+exports.allToolsDefinitions = exports.viewFailedTestRunReportTool = exports.upgradePackagesTool = exports.updateIssueTool = exports.traceDotZipTool = exports.sendTriageSummaryTool = exports.runTestTool = exports.reviewPullRequestTool = exports.renameFileTool = exports.mergeConflictsTool = exports.listIssuesTool = exports.listEnvironmentsTool = exports.grepTool = exports.generateTestWithBrowserAgent = exports.analyseVideo = exports.fetchTestRunReportTool = exports.fetchSessionDiffTool = exports.fetchLastSuccessfulTestRunTool = exports.fetchFileTool = exports.fetchDiagnosisReportTool = exports.downloadBuildTool = exports.deleteFileTool = exports.createPullRequestTool = exports.createIssueTool = exports.triageOnlyTools = exports.testGenerationTools = exports.commonTools = void 0;
+exports.hasBuiltInTextEditor = hasBuiltInTextEditor;
 exports.textEditorToolsForModel = textEditorToolsForModel;
+exports.textViewToolsForModel = textViewToolsForModel;
 exports.toolsNeedBrowser = toolsNeedBrowser;
 exports.nameForTelemetry = nameForTelemetry;
 exports.sendToolRequestToRemoteQueue = sendToolRequestToRemoteQueue;
-const chat_1 = require("@empiricalrun/llm/chat");
 const SQSClient_1 = require("../utils/SQSClient");
 const create_pull_request_1 = require("./create-pull-request");
 Object.defineProperty(exports, "createPullRequestTool", { enumerable: true, get: function () { return create_pull_request_1.createPullRequestTool; } });
+const analyse_video_1 = require("./definitions/analyse-video");
+Object.defineProperty(exports, "analyseVideo", { enumerable: true, get: function () { return analyse_video_1.analyseVideo; } });
 const delete_file_1 = require("./definitions/delete-file");
 Object.defineProperty(exports, "deleteFileTool", { enumerable: true, get: function () { return delete_file_1.deleteFileTool; } });
 const download_build_1 = require("./definitions/download-build");
 Object.defineProperty(exports, "downloadBuildTool", { enumerable: true, get: function () { return download_build_1.downloadBuildTool; } });
-const extract_frames_from_video_1 = require("./definitions/extract-frames-from-video");
-Object.defineProperty(exports, "extractFramesFromVideo", { enumerable: true, get: function () { return extract_frames_from_video_1.extractFramesFromVideo; } });
-const fetch_video_analysis_1 = require("./definitions/fetch-video-analysis");
-Object.defineProperty(exports, "fetchVideoAnalysis", { enumerable: true, get: function () { return fetch_video_analysis_1.fetchVideoAnalysis; } });
 const grep_1 = require("./definitions/grep");
 Object.defineProperty(exports, "grepTool", { enumerable: true, get: function () { return grep_1.grepTool; } });
 const list_tests_and_projects_1 = require("./definitions/list-tests-and-projects");
-Object.defineProperty(exports, "listProjectsTool", { enumerable: true, get: function () { return list_tests_and_projects_1.listProjectsTool; } });
-Object.defineProperty(exports, "listTestsForProjectTool", { enumerable: true, get: function () { return list_tests_and_projects_1.listTestsForProjectTool; } });
 const merge_conflicts_1 = require("./definitions/merge-conflicts");
 Object.defineProperty(exports, "mergeConflictsTool", { enumerable: true, get: function () { return merge_conflicts_1.mergeConflictsTool; } });
 const rename_file_1 = require("./definitions/rename-file");
 Object.defineProperty(exports, "renameFileTool", { enumerable: true, get: function () { return rename_file_1.renameFileTool; } });
+const review_pull_request_1 = require("./definitions/review-pull-request");
+Object.defineProperty(exports, "reviewPullRequestTool", { enumerable: true, get: function () { return review_pull_request_1.reviewPullRequestTool; } });
 const run_test_1 = require("./definitions/run-test");
 Object.defineProperty(exports, "runTestTool", { enumerable: true, get: function () { return run_test_1.runTestTool; } });
 const str_replace_editor_1 = require("./definitions/str_replace_editor");
@@ -47,8 +46,6 @@ Object.defineProperty(exports, "listIssuesTool", { enumerable: true, get: functi
 Object.defineProperty(exports, "updateIssueTool", { enumerable: true, get: function () { return issues_1.updateIssueTool; } });
 const list_environments_1 = require("./list-environments");
 Object.defineProperty(exports, "listEnvironmentsTool", { enumerable: true, get: function () { return list_environments_1.listEnvironmentsTool; } });
-const review_pull_request_1 = require("./review-pull-request");
-Object.defineProperty(exports, "reviewPullRequestTool", { enumerable: true, get: function () { return review_pull_request_1.reviewPullRequestTool; } });
 const test_run_fetcher_1 = require("./test-run-fetcher");
 Object.defineProperty(exports, "fetchTestRunReportTool", { enumerable: true, get: function () { return test_run_fetcher_1.fetchTestRunReportTool; } });
 const trace_dot_zip_1 = require("./trace-dot-zip");
@@ -68,6 +65,8 @@ exports.commonTools = [
     download_build_1.downloadBuildTool,
     fetch_file_1.fetchFileTool,
     trace_dot_zip_1.traceDotZipTool,
+    list_tests_and_projects_1.listProjectsTool,
+    list_tests_and_projects_1.listTestsForProjectTool,
 ];
 exports.testGenerationTools = [
     test_gen_browser_1.generateTestWithBrowserAgent,
@@ -83,7 +82,7 @@ exports.triageOnlyTools = [
     issues_1.createIssueTool,
     issues_1.updateIssueTool,
     view_failed_test_run_report_1.viewFailedTestRunReportTool,
-    fetch_video_analysis_1.fetchVideoAnalysis,
+    analyse_video_1.analyseVideo,
     fetch_last_successful_test_run_1.fetchLastSuccessfulTestRunTool,
     triage_summary_1.sendTriageSummaryTool,
     review_pull_request_1.reviewPullRequestTool,
@@ -93,16 +92,18 @@ exports.allToolsDefinitions = [
     ...exports.testGenerationTools,
     ...exports.triageOnlyTools,
     // Individual tools that were in removed arrays
-    extract_frames_from_video_1.extractFramesFromVideo,
     fetch_session_diff_1.fetchSessionDiffTool,
     list_tests_and_projects_1.listProjectsTool,
     list_tests_and_projects_1.listTestsForProjectTool,
 ];
+function hasBuiltInTextEditor(model) {
+    return model.includes("claude") && !model.includes("haiku");
+}
 function textEditorToolsForModel(model) {
-    if (model && (0, chat_1.getProviderForModel)(model) !== "claude") {
-        return Object.values(str_replace_editor_1.textEditorTools);
-    }
-    return [];
+    return !hasBuiltInTextEditor(model) ? Object.values(str_replace_editor_1.textEditorTools) : [];
+}
+function textViewToolsForModel(model) {
+    return !hasBuiltInTextEditor(model) ? Object.values(str_replace_editor_1.textViewTools) : [];
 }
 function toolsNeedBrowser(toolCalls) {
     return toolCalls.some((toolCall) => exports.allToolsDefinitions.find((t) => t.schema.name === toolCall.name)

package/dist/tools/merge-conflicts/index.d.ts.map

@@ -1 +1 @@
-{"version":3,"file":"index.d.ts","sourceRoot":"","sources":["../../../src/tools/merge-conflicts/index.ts"],"names":[],"mappings":"AAAA,OAAO,KAAK,EAAE,IAAI,EAAE,MAAM,4BAA4B,CAAC;AA2CvD,eAAO,MAAM,kBAAkB,EAAE,IAiDhC,CAAC"}
+{"version":3,"file":"index.d.ts","sourceRoot":"","sources":["../../../src/tools/merge-conflicts/index.ts"],"names":[],"mappings":"AAAA,OAAO,KAAK,EAAE,IAAI,EAAE,MAAM,4BAA4B,CAAC;AA8CvD,eAAO,MAAM,kBAAkB,EAAE,IAiDhC,CAAC"}

package/dist/tools/merge-conflicts/index.js

@@ -11,7 +11,7 @@ const merge_conflicts_1 = require("../definitions/merge-conflicts");
 const git_1 = require("../executor/utils/git");
 function runMergeMain(repoPath) {
     try {
-        (0, child_process_1.execSync)("git merge --no-ff origin/main", { stdio: "pipe", cwd: repoPath });
+        (0, child_process_1.execSync)("git merge --no-ff origin/main -m 'Merge remote-tracking branch origin/main [skip ci]'", { stdio: "pipe", cwd: repoPath });
         return "CLEAN";
     }
     catch (error) {

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

@@ -69,7 +69,7 @@ exports.renameFileTool = {
             if (!tscResult.success) {
                 return {
                     result: `File renamed from '${input.oldPath}' to '${input.newPath}'. However, type checks are failing with errors:\n\n${tscResult.errors.join("\n")}`,
-                    isError: true,
+                    isError: false,
                 };
             }
             return {

package/dist/tools/review-pull-request/index.d.ts.map

@@ -1 +1 @@
-{"version":3,"file":"index.d.ts","sourceRoot":"","sources":["../../../src/tools/review-pull-request/index.ts"],"names":[],"mappings":"AAAA,OAAO,EAAmB,IAAI,EAAc,MAAM,4BAA4B,CAAC;AAc/E,eAAO,MAAM,qBAAqB,EAAE,IAgGnC,CAAC"}
+{"version":3,"file":"index.d.ts","sourceRoot":"","sources":["../../../src/tools/review-pull-request/index.ts"],"names":[],"mappings":"AAAA,OAAO,EAAE,IAAI,EAAc,MAAM,4BAA4B,CAAC;AAU9D,eAAO,MAAM,qBAAqB,EAAE,IAiGnC,CAAC"}

package/dist/tools/review-pull-request/index.js

@@ -1,24 +1,14 @@
 "use strict";
 Object.defineProperty(exports, "__esModule", { value: true });
 exports.reviewPullRequestTool = void 0;
-const zod_1 = require("zod");
-const exports_1 = require("../../agent/chat/exports");
 const state_1 = require("../../agent/chat/state");
 const code_review_1 = require("../../agent/code-review");
-const fetch_session_diff_1 = require("../fetch-session-diff");
+const executor_1 = require("../../agent/code-review/executor");
+const review_pull_request_1 = require("../definitions/review-pull-request");
 const CODE_REVIEW_MODEL = "claude-sonnet-4-20250514";
-const ReviewPullRequestInputSchema = zod_1.z.object({
-    sessionUrl: zod_1.z.string().describe("The URL of the chat session to review"),
-});
 exports.reviewPullRequestTool = {
-    schema: {
-        name: "reviewPullRequest",
-        description: `Reviews a pull request by using CodeReviewAgent to fetch session diff and analyze it. Returns a boolean result with an explanation.`,
-        parameters: ReviewPullRequestInputSchema,
-    },
-    needsBrowser: false,
-    isInlineTool: true,
-    execute: async ({ input, apiClient }) => {
+    ...review_pull_request_1.reviewPullRequestTool,
+    execute: async ({ input, apiClient, chatSession, environmentOverrides, trace, featureFlags, repoPath, }) => {
         try {
             const { sessionUrl } = input;
             if (!sessionUrl) {
@@ -27,10 +17,12 @@ exports.reviewPullRequestTool = {
                     result: "No chat session URL provided",
                 };
             }
-            const agent = new code_review_1.CodeReviewAgent({
-                featureFlags: [],
-                selectedModel: CODE_REVIEW_MODEL,
-            });
+            if (!apiClient)
+                return {
+                    isError: true,
+                    result: "API Client not available",
+                };
+            // creating chat state to pass to model
             const userPrompt = `Review this session: ${sessionUrl}`;
             const chatState = (0, state_1.createChatState)({
                 userPrompt,
@@ -39,49 +31,43 @@ exports.reviewPullRequestTool = {
                 selectedModel: CODE_REVIEW_MODEL,
                 error: null,
             });
-            agent.initializeWithState(chatState);
-            let chatModel = (0, exports_1.createChatModel)(chatState.messages, CODE_REVIEW_MODEL);
-            while (!chatModel.askUserForInput) {
+            const span = trace?.span({
+                name: "code-review-agent",
+                input: {
+                    sessionUrl,
+                    model: CODE_REVIEW_MODEL,
+                },
+            });
+            const codeReviewToolExecutor = new executor_1.CodeReviewToolExecutor({
+                chatSession: chatSession || null,
+                repoPath,
+                apiClient,
+                trace: span,
+                featureFlags,
+                environmentOverrides: environmentOverrides || {},
+            });
+            const agentParams = {
+                featureFlags: [],
+                selectedModel: CODE_REVIEW_MODEL,
+                chatState,
+                toolExecutor: codeReviewToolExecutor,
+                trace: span,
+            };
+            const agent = new code_review_1.CodeReviewAgent({ ...agentParams });
+            while (!agent.askUserForInput) {
                 await agent.runLoop({
-                    messages: chatModel.messages,
-                    reporter: async (data) => {
-                        console.log("Review progress:", data);
-                    },
-                    onPendingToolCall: async (toolCalls) => {
-                        const [toolCall] = toolCalls;
-                        // TODO: This will only work for 1 tool - will break if
-                        // code review agent has more tools: need first class primitive
-                        // for inline tool execution
-                        if (toolCall?.name === "fetchSessionDiff") {
-                            const tool = fetch_session_diff_1.fetchSessionDiffTool;
-                            const result = await tool.execute({
-                                input: toolCall.input,
-                                apiClient,
-                                repoPath: ``,
-                                featureFlags: [],
-                            });
-                            chatModel.pushToolResultsMessage([toolCall], [result]);
-                        }
-                    },
-                    onLLMResponse: async (response, model) => {
-                        console.log(`Agent responded using ${model}:`, response);
-                    },
+                    reporter: async () => { },
+                    trace: span,
                 });
-                // Update the chatModel with the agent's final state for next iteration
-                if (agent.messages) {
-                    chatModel = (0, exports_1.createChatModel)(agent.messages, CODE_REVIEW_MODEL);
-                }
             }
-            const messages = agent.messages || [];
-            const lastMessage = messages.length
-                ? messages[messages.length - 1]
-                : undefined;
-            const lastMessageTextPart = lastMessage
-                ? lastMessage.parts
-                    .filter((p) => "text" in p)
-                    .find((p) => "text" in p && !!p.text)
-                : undefined;
-            if (!lastMessageTextPart) {
+            const agentResult = agent.getResult();
+            span?.end({
+                output: {
+                    result: agentResult,
+                    success: !!agentResult,
+                },
+            });
+            if (!agentResult) {
                 return {
                     isError: true,
                     result: "Could not extract text response from agent",
@@ -89,7 +75,7 @@ exports.reviewPullRequestTool = {
             }
             return {
                 isError: false,
-                result: lastMessageTextPart.text,
+                result: JSON.stringify(agentResult),
             };
         }
         catch (error) {

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,EAAE,IAAI,EAAc,MAAM,4BAA4B,CAAC;AA0CnE,eAAO,MAAM,WAAW,EAAE,IAiFzB,CAAC"}
+{"version":3,"file":"run-test.d.ts","sourceRoot":"","sources":["../../src/tools/run-test.ts"],"names":[],"mappings":"AAAA,OAAO,KAAK,EAAU,IAAI,EAAc,MAAM,4BAA4B,CAAC;AA0C3E,eAAO,MAAM,WAAW,EAAE,IA4GzB,CAAC"}

package/dist/tools/run-test.js

@@ -29,7 +29,7 @@ ${JSON.stringify(truncatedSummaryJson)}
 }
 exports.runTestTool = {
     ...run_test_1.runTestTool,
-    execute: async ({ input, repoPath, collectArtifacts, environmentOverrides = {}, }) => {
+    execute: async ({ input, repoPath, collectArtifacts, environmentOverrides = {}, logger, }) => {
         let reportUrl = undefined;
         let envOverrides = { ...environmentOverrides };
         if ((0, artifacts_1.isArtifactCollectionEnabled)()) {
@@ -43,14 +43,18 @@ exports.runTestTool = {
             };
         }
         else {
-            console.warn("R2 credentials not found: report artifacts will not be uploaded");
+            const warnMsg = "R2 credentials not found: report artifacts will not be uploaded";
+            logger?.warn(warnMsg) || console.warn(warnMsg);
         }
         const { filePath } = input;
         const absoluteFilePath = path_1.default.join(repoPath, filePath);
         if (!fs_1.default.existsSync(absoluteFilePath)) {
+            const errorMsg = `File ${filePath} does not exist.`;
+            logger?.error("Test file not found", { filePath, absoluteFilePath }) ||
+                console.error("Test file not found:", filePath);
             return {
                 isError: true,
-                result: `File ${filePath} does not exist.`,
+                result: errorMsg,
             };
         }
         if (input.buildUrl) {
@@ -60,6 +64,12 @@ exports.runTestTool = {
             };
         }
         try {
+            logger?.info("Starting test execution", {
+                testName: input.testName,
+                filePath,
+                project: input.project,
+                suites: input.suites,
+            }) || console.log("Starting test execution:", input.testName);
             const result = await (0, test_run_1.runSingleTest)({
                 testName: input.testName,
                 suites: input.suites,
@@ -67,7 +77,14 @@ exports.runTestTool = {
                 repoDir: repoPath,
                 projects: [input.project],
                 envOverrides,
+                stdout: logger?.createPrefixedStream?.("stdout"),
+                stderr: logger?.createPrefixedStream?.("stderr"),
             });
+            logger?.info("Test execution completed", {
+                passed: result.hasTestPassed,
+                testCount: result.summaryJson?.suites?.length || 0,
+            }) ||
+                console.log("Test execution completed. Passed:", result.hasTestPassed);
             const artifacts = (0, utils_1.extractAttachmentsFromPlaywrightJSONReport)(result.summaryJson);
             void collectArtifacts?.(artifacts);
             return {
@@ -82,6 +99,11 @@ exports.runTestTool = {
         catch (error) {
             // Ensure we capture the full error message regardless of error type
             const errorMessage = error instanceof Error ? error.message : String(error);
+            logger?.error("Test execution failed", {
+                error: errorMessage,
+                testName: input.testName,
+                filePath,
+            }) || console.error("Test execution failed:", errorMessage);
             return {
                 artifacts: null,
                 result: JSON.stringify({ error: errorMessage }),

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,EACV,IAAI,EAGL,MAAM,4BAA4B,CAAC;AAoBpC,eAAO,MAAM,4BAA4B,EAAE,IA0L1C,CAAC"}
+{"version":3,"file":"test-gen-browser.d.ts","sourceRoot":"","sources":["../../src/tools/test-gen-browser.ts"],"names":[],"mappings":"AACA,OAAO,KAAK,EAAiB,IAAI,EAAE,MAAM,4BAA4B,CAAC;AAqBtE,eAAO,MAAM,4BAA4B,EAAE,IA+L1C,CAAC"}

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

@@ -10,6 +10,7 @@ const path_1 = __importDefault(require("path"));
 const run_1 = require("../agent/browsing/run");
 const utils_1 = require("../agent/browsing/utils");
 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");
 const test_gen_browser_1 = require("./definitions/test-gen-browser");
@@ -53,8 +54,9 @@ exports.generateTestWithBrowserAgent = {
         }
         // Prepare the file for the browser agent
         const fileBackup = fs_1.default.readFileSync(absoluteFilePath, "utf-8");
+        let todoContent;
         try {
-            await (0, utils_1.replaceTodoWithCreateTest)(filePath, repoPath);
+            todoContent = (0, utils_1.replaceTodoWithCreateTest)(filePath, repoPath);
         }
         catch (error) {
             // Undo the TODO -> createTest and test.only changes
@@ -83,7 +85,7 @@ exports.generateTestWithBrowserAgent = {
                 BUILD_URL: input.buildUrl,
             };
         }
-        const testGenToken = (0, scenarios_1.buildTokenFromOptions)({ name: testName, file: filePath, prompt: input.changeToMake }, { useComputerUseAgent: true });
+        const testGenToken = (0, scenarios_1.buildTokenFromOptions)({ name: testName, file: filePath }, { useComputerUseAgent: true });
         console.log("[generateTestWithBrowserAgent] Validations passed, starting agent");
         const agentResult = await (0, run_1.runBrowsingAgent)({
             testCaseName: testName,
@@ -100,10 +102,7 @@ exports.generateTestWithBrowserAgent = {
         // Cleanup: Undo the TODO -> createTest changes
         await (0, pw_pause_1.revertToOriginalPwCode)(repoPath);
         fs_1.default.writeFileSync(absoluteFilePath, fileBackup, "utf-8");
-        const { isError, error, result: toolResult, artifacts, usage, } = agentResult;
-        if (artifacts) {
-            void collectArtifacts?.(artifacts);
-        }
+        const { isError, error, result: toolResult, usage } = agentResult;
         if (isError || !toolResult) {
             return {
                 isError,
@@ -112,53 +111,58 @@ ${error}
 `,
             };
         }
-        let result = [
-            {
-                type: "text",
-                text: "Browser agent has finished running. Here is the summary of actions it took, and the generated Playwright code:",
-            },
+        const screenshotArtifacts = toolResult
+            .filter((item) => item.type === "screenshot")
+            .map((item) => item.screenshot)
+            .map((artifact) => ({
+            ...artifact,
+            data: Buffer.isBuffer(artifact.data)
+                ? artifact.data
+                : Buffer.from(artifact.data),
+        }));
+        const artifactsInput = [
+            ...((0, utils_2.findPlaywrightArtifacts)(repoPath) || []),
+            ...screenshotArtifacts,
+        ];
+        const summaryResult = `Browser agent has finished running.
+TODO comment which it worked on: ${todoContent}
+Summary of actions it took, and the generated Playwright code:`;
+        if (!collectArtifacts || artifactsInput.length === 0) {
+            const result = [
+                { type: "text", text: summaryResult },
+                ...toolResult.map((item) => ({
+                    type: "text",
+                    text: item.type === "text" ? item.text : "[Screenshot]",
+                })),
+            ];
+            return { isError, result, usage };
+        }
+        const uploadedArtifacts = await collectArtifacts(artifactsInput);
+        const artifactUrlMap = new Map(uploadedArtifacts.map((a) => [a.name, a.url]));
+        let screenshotIndex = 0;
+        const result = [
+            { type: "text", text: summaryResult },
             ...toolResult.map((item) => {
                 if (item.type === "text") {
-                    return {
-                        type: "text",
-                        text: item.text,
-                    };
+                    return { type: "text", text: item.text };
                 }
                 else if (item.type === "screenshot") {
-                    const bufferData = Buffer.isBuffer(item.screenshot.data)
-                        ? item.screenshot.data
-                        : Buffer.from(item.screenshot.data);
-                    return {
-                        type: "image/png",
-                        base64Data: `${bufferData.toString("base64")}`,
-                    };
-                }
-                else {
-                    return {
-                        type: "text",
-                        text: `Unknown item type: ${item}`,
-                    };
+                    const screenshotName = screenshotArtifacts[screenshotIndex]?.name;
+                    const url = screenshotName
+                        ? artifactUrlMap.get(screenshotName)
+                        : undefined;
+                    screenshotIndex++;
+                    return url
+                        ? { type: "image/png", url }
+                        : { type: "text", text: "[Screenshot upload failed]" };
                 }
+                return { type: "text", text: "Unknown item type" };
             }),
         ];
-        const isMultiModalEnabled = featureFlags.includes("multi_modal_tool_result_from_browser_agent");
-        if (isMultiModalEnabled) {
-            return {
-                isError,
-                result,
-                usage,
-            };
-        }
-        else {
-            const textOnlyResult = result
-                .filter((item) => item.type === "text")
-                .map((item) => item.text)
-                .join("\n");
-            return {
-                isError,
-                result: textOnlyResult,
-                usage,
-            };
-        }
+        return {
+            isError,
+            result,
+            usage,
+        };
     },
 };

package/dist/utils/artifact-paths.d.ts

@@ -0,0 +1,20 @@
+/**
+ * root/ (= baseArtifactDir)
+ *  - video_<videoUrlHash>.webm (= videoFile)
+ *  - chunk_1.mp4
+ *  - consolidated_frames/ (= consolidatedFramesDir)
+ *     - frame_0001.png
+ *     - frame_0002.png
+ *     - ...
+ *  - unique_frames/ (= uniqueFramesDir)
+ *     - frame_0002.png
+ *     - ...
+ */
+export interface VideoArtifactPaths {
+    baseArtifactDir: string;
+    videoFile: string;
+    consolidatedFramesDir: string;
+    uniqueFramesDir: string;
+}
+export declare function resolveVideoArtifactPaths(artifactsPath: string, videoUrlHash: string): VideoArtifactPaths;
+//# sourceMappingURL=artifact-paths.d.ts.map

package/dist/utils/artifact-paths.d.ts.map

@@ -0,0 +1 @@
+{"version":3,"file":"artifact-paths.d.ts","sourceRoot":"","sources":["../../src/utils/artifact-paths.ts"],"names":[],"mappings":"AAEA;;;;;;;;;;;GAWG;AAEH,MAAM,WAAW,kBAAkB;IAEjC,eAAe,EAAE,MAAM,CAAC;IAExB,SAAS,EAAE,MAAM,CAAC;IAElB,qBAAqB,EAAE,MAAM,CAAC;IAE9B,eAAe,EAAE,MAAM,CAAC;CACzB;AAED,wBAAgB,yBAAyB,CACvC,aAAa,EAAE,MAAM,EACrB,YAAY,EAAE,MAAM,GACnB,kBAAkB,CAQpB"}

package/dist/utils/artifact-paths.js

@@ -0,0 +1,16 @@
+"use strict";
+var __importDefault = (this && this.__importDefault) || function (mod) {
+    return (mod && mod.__esModule) ? mod : { "default": mod };
+};
+Object.defineProperty(exports, "__esModule", { value: true });
+exports.resolveVideoArtifactPaths = resolveVideoArtifactPaths;
+const node_path_1 = __importDefault(require("node:path"));
+function resolveVideoArtifactPaths(artifactsPath, videoUrlHash) {
+    const baseArtifactDir = artifactsPath;
+    return {
+        baseArtifactDir,
+        videoFile: node_path_1.default.join(baseArtifactDir, `video_${videoUrlHash}.webm`),
+        consolidatedFramesDir: node_path_1.default.join(baseArtifactDir, "consolidated_frames"),
+        uniqueFramesDir: node_path_1.default.join(baseArtifactDir, "unique_frames"),
+    };
+}

package/dist/utils/dedup-image-fs.d.ts

@@ -1,21 +1,7 @@
-/**
- * Processes an array of image file paths and returns only unique images as base64 strings.
- * This is a file system wrapper around the deduplicateImages function that processes
- * images in batches to avoid memory overflow.
- *
- * @param options - Configuration options
- * @param options.imagePaths - An array of file paths to image files.
- * @param options.batchSize - Number of images to process at a time (default: 50)
- * @param options.threshold - The maximum fraction of pixels allowed to differ to consider two images similar.
- *                    For example, 0.01 means that if less than 1% of the pixels differ, the images are considered duplicates.
- *                    Default is 0.001 (0.1%).
- * @param options.logPrefix - The ID of the test run, used for logging purposes.
- * @returns A promise that resolves to an array of unique image objects with base64 data.
- */
 export declare function deduplicateImageFiles({ imagePaths, batchSize, threshold, logPrefix, }: {
     imagePaths: string[];
-    batchSize?: number;
-    threshold?: number;
+    batchSize: number;
+    threshold: number;
     logPrefix?: string;
 }): Promise<{
     metadata: {

package/dist/utils/dedup-image-fs.d.ts.map

@@ -1 +1 @@
-{"version":3,"file":"dedup-image-fs.d.ts","sourceRoot":"","sources":["../../src/utils/dedup-image-fs.ts"],"names":[],"mappings":"AAIA;;;;;;;;;;;;;GAaG;AACH,wBAAsB,qBAAqB,CAAC,EAC1C,UAAU,EACV,SAAc,EACd,SAAiB,EACjB,SAA4B,GAC7B,EAAE;IACD,UAAU,EAAE,MAAM,EAAE,CAAC;IACrB,SAAS,CAAC,EAAE,MAAM,CAAC;IACnB,SAAS,CAAC,EAAE,MAAM,CAAC;IACnB,SAAS,CAAC,EAAE,MAAM,CAAC;CACpB,GAAG,OAAO,CAAC;IAAE,QAAQ,EAAE;QAAE,KAAK,EAAE,MAAM,CAAC;QAAC,IAAI,EAAE,MAAM,CAAA;KAAE,CAAC;IAAC,KAAK,EAAE,MAAM,CAAA;CAAE,EAAE,CAAC,CAmG1E"}
+{"version":3,"file":"dedup-image-fs.d.ts","sourceRoot":"","sources":["../../src/utils/dedup-image-fs.ts"],"names":[],"mappings":"AAKA,wBAAsB,qBAAqB,CAAC,EAC1C,UAAU,EACV,SAAS,EACT,SAAS,EACT,SAA4B,GAC7B,EAAE;IACD,UAAU,EAAE,MAAM,EAAE,CAAC;IACrB,SAAS,EAAE,MAAM,CAAC;IAClB,SAAS,EAAE,MAAM,CAAC;IAClB,SAAS,CAAC,EAAE,MAAM,CAAC;CACpB,GAAG,OAAO,CAAC;IAAE,QAAQ,EAAE;QAAE,KAAK,EAAE,MAAM,CAAC;QAAC,IAAI,EAAE,MAAM,CAAA;KAAE,CAAC;IAAC,KAAK,EAAE,MAAM,CAAA;CAAE,EAAE,CAAC,CA2G1E"}

package/dist/utils/dedup-image-fs.js

@@ -1,23 +1,13 @@
 "use strict";
+var __importDefault = (this && this.__importDefault) || function (mod) {
+    return (mod && mod.__esModule) ? mod : { "default": mod };
+};
 Object.defineProperty(exports, "__esModule", { value: true });
 exports.deduplicateImageFiles = deduplicateImageFiles;
 const fs_1 = require("fs");
+const path_1 = __importDefault(require("path"));
 const dedup_image_1 = require("./dedup-image");
-/**
- * Processes an array of image file paths and returns only unique images as base64 strings.
- * This is a file system wrapper around the deduplicateImages function that processes
- * images in batches to avoid memory overflow.
- *
- * @param options - Configuration options
- * @param options.imagePaths - An array of file paths to image files.
- * @param options.batchSize - Number of images to process at a time (default: 50)
- * @param options.threshold - The maximum fraction of pixels allowed to differ to consider two images similar.
- *                    For example, 0.01 means that if less than 1% of the pixels differ, the images are considered duplicates.
- *                    Default is 0.001 (0.1%).
- * @param options.logPrefix - The ID of the test run, used for logging purposes.
- * @returns A promise that resolves to an array of unique image objects with base64 data.
- */
-async function deduplicateImageFiles({ imagePaths, batchSize = 50, threshold = 0.001, logPrefix = "dedup-image-fs", }) {
+async function deduplicateImageFiles({ imagePaths, batchSize, threshold, logPrefix = "dedup-image-fs", }) {
     const uniqueImages = [];
     console.log(`[${logPrefix}] Processing ${imagePaths.length} images in batches of ${batchSize}`);
     for (let i = 0; i < imagePaths.length; i += batchSize) {
@@ -33,9 +23,15 @@ async function deduplicateImageFiles({ imagePaths, batchSize = 50, threshold = 0
             try {
                 const buffer = await fs_1.promises.readFile(imagePath);
                 const base64 = buffer.toString("base64");
+                // Extract actual frame number from filename instead of using serial index
+                const filename = path_1.default.basename(imagePath);
+                const frameNumberMatch = filename.match(/frame_(\d+)\.png$/);
+                const frameIndex = frameNumberMatch
+                    ? parseInt(frameNumberMatch[1], 10)
+                    : i + j; // Fallback to serial index if no match
                 batchBase64Images.push({
                     metadata: {
-                        index: i + j,
+                        index: frameIndex,
                         path: imagePath,
                     },
                     image: base64,

package/dist/utils/dedup-image.d.ts

@@ -1,22 +1,9 @@
-/**
- * Processes an array of base64 image objects and returns only unique images.
- * Two images are considered "similar" if the fraction of differing pixels (when resized)
- * is below the provided threshold.
- *
- * @param options - Configuration options
- * @param options.base64Images - An array of objects, each with a metadata field and a base64-encoded image string.
- * @param options.threshold - The maximum fraction of pixels allowed to differ to consider two images similar.
- *                    For example, 0.01 means that if less than 1% of the pixels differ, the images are considered duplicates.
- *                    Default is 0.001 (0.1%).
- * @param options.logPrefix - The ID of the test run, used for logging purposes.
- * @returns A promise that resolves to an array of unique image objects.
- */
 export declare function deduplicateImages({ base64Images, threshold, logPrefix, }: {
     base64Images: {
         metadata: any;
         image: string;
     }[];
-    threshold?: number;
+    threshold: number;
     logPrefix?: string;
 }): Promise<{
     metadata: any;

package/dist/utils/dedup-image.d.ts.map

@@ -1 +1 @@
-{"version":3,"file":"dedup-image.d.ts","sourceRoot":"","sources":["../../src/utils/dedup-image.ts"],"names":[],"mappings":"AAGA;;;;;;;;;;;;GAYG;AACH,wBAAsB,iBAAiB,CAAC,EACtC,YAAY,EACZ,SAAiB,EACjB,SAAyB,GAC1B,EAAE;IACD,YAAY,EAAE;QAAE,QAAQ,EAAE,GAAG,CAAC;QAAC,KAAK,EAAE,MAAM,CAAA;KAAE,EAAE,CAAC;IACjD,SAAS,CAAC,EAAE,MAAM,CAAC;IACnB,SAAS,CAAC,EAAE,MAAM,CAAC;CACpB,GAAG,OAAO,CAAC;IAAE,QAAQ,EAAE,GAAG,CAAC;IAAC,KAAK,EAAE,MAAM,CAAA;CAAE,EAAE,CAAC,CAsE9C"}
+{"version":3,"file":"dedup-image.d.ts","sourceRoot":"","sources":["../../src/utils/dedup-image.ts"],"names":[],"mappings":"AAEA,wBAAsB,iBAAiB,CAAC,EACtC,YAAY,EACZ,SAAS,EACT,SAAyB,GAC1B,EAAE;IACD,YAAY,EAAE;QAAE,QAAQ,EAAE,GAAG,CAAC;QAAC,KAAK,EAAE,MAAM,CAAA;KAAE,EAAE,CAAC;IACjD,SAAS,EAAE,MAAM,CAAC;IAClB,SAAS,CAAC,EAAE,MAAM,CAAC;CACpB,GAAG,OAAO,CAAC;IAAE,QAAQ,EAAE,GAAG,CAAC;IAAC,KAAK,EAAE,MAAM,CAAA;CAAE,EAAE,CAAC,CA0B9C"}