@empiricalrun/test-gen 0.78.0 → 0.78.2

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

@@ -9,9 +9,9 @@ const fs_1 = __importDefault(require("fs"));
 const path_1 = __importDefault(require("path"));
 const merge_conflicts_1 = require("../definitions/merge-conflicts");
 const git_1 = require("../executor/utils/git");
-function runMergeMain(repoPath) {
+function runMerge(repoPath, targetBranch) {
     try {
-        (0, child_process_1.execSync)("git merge --no-ff origin/main -m 'Merge remote-tracking branch origin/main [skip ci]'", { stdio: "pipe", cwd: repoPath });
+        (0, child_process_1.execSync)(`git merge --no-ff origin/${targetBranch} -m 'Merge remote-tracking branch origin/${targetBranch} [skip ci]'`, { stdio: "pipe", cwd: repoPath });
         return "CLEAN";
     }
     catch (error) {
@@ -42,26 +42,39 @@ function getConflictedFilesWithContents(repoPath) {
 }
 exports.mergeConflictsTool = {
     ...merge_conflicts_1.mergeConflictsTool,
-    execute: async ({ repoPath }) => {
+    execute: async ({ repoPath, chatSession }) => {
+        const targetBranch = chatSession?.baseBranchName || "main";
         try {
             // Fetch main branch (git clone runs --single-branch), and then run merge
-            (0, child_process_1.execSync)("git remote set-branches --add origin main", { cwd: repoPath });
-            (0, child_process_1.execSync)("git fetch origin main", { cwd: repoPath });
+            (0, child_process_1.execSync)(`git remote set-branches --add origin ${targetBranch}`, {
+                cwd: repoPath,
+            });
+            (0, child_process_1.execSync)(`git fetch origin ${targetBranch}`, { cwd: repoPath });
+            // Configure git user for merge
+            (0, child_process_1.execSync)(`git config user.name "${git_1.GIT_USER_NAME}"`, { cwd: repoPath });
+            (0, child_process_1.execSync)(`git config user.email "${git_1.GIT_USER_EMAIL}"`, { cwd: repoPath });
             // Run merge command and commit (with or without conflict)
-            const mergeResult = runMergeMain(repoPath);
+            const mergeResult = runMerge(repoPath, targetBranch);
             const branchName = await (0, git_1.getCurrentBranchName)(repoPath);
             if (mergeResult === "CLEAN") {
                 (0, child_process_1.execSync)(`git push origin ${branchName}`, { cwd: repoPath });
                 return {
                     isError: false,
-                    result: `Merge from main was successful, without any conflicts.`,
+                    result: `Merge from ${targetBranch} was successful, without any conflicts.`,
                 };
             }
             else if (mergeResult === "CONFLICTS") {
                 const files = getConflictedFilesWithContents(repoPath);
+                const filePaths = Object.keys(files);
+                if (filePaths.length === 0) {
+                    return {
+                        isError: true,
+                        result: `Merge failed with exit code 1, but no conflicted files were found.`,
+                    };
+                }
                 (0, git_1.stageAndCommitFilesAsBotUser)({
-                    commitMessage: "Merge from main with conflicts",
-                    files: Object.keys(files),
+                    commitMessage: `Merge from ${targetBranch} with conflicts`,
+                    files: filePaths,
                     repoPath,
                 });
                 (0, child_process_1.execSync)(`git push origin ${branchName}`, { cwd: repoPath });
@@ -71,7 +84,7 @@ exports.mergeConflictsTool = {
                     .join("\n\n");
                 return {
                     isError: false,
-                    result: `Merge from main is committed, with conflicts. Use text edit tools to resolve them.\n\n${filesResult}`,
+                    result: `Merge from ${targetBranch} is committed, with conflicts. Use text edit tools to resolve them.\n\n${filesResult}`,
                 };
             }
             else {

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

@@ -1,3 +1,3 @@
-import type { Tool } from "@empiricalrun/shared-types";
+import type { Tool } from "@empiricalrun/shared-types/chat-agent";
 export declare const renameFileTool: Tool;
 //# sourceMappingURL=index.d.ts.map

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,4BAA4B,CAAC;AAQnE,eAAO,MAAM,cAAc,EAAE,IA6E5B,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,IA4E5B,CAAC"}

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

@@ -64,7 +64,6 @@ exports.renameFileTool = {
             }
             stageFile(oldFilePath);
             stageFile(newFilePath);
-            // Run TypeScript compilation check
             const tscResult = await (0, web_1.runTypescriptCompiler)(repoPath);
             if (!tscResult.success) {
                 return {

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

@@ -1,3 +1,3 @@
-import { Tool } from "@empiricalrun/shared-types";
+import { Tool } from "@empiricalrun/shared-types/chat-agent";
 export declare const reviewPullRequestTool: Tool;
 //# sourceMappingURL=index.d.ts.map

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,EAAE,IAAI,EAAc,MAAM,4BAA4B,CAAC;AAS9D,eAAO,MAAM,qBAAqB,EAAE,IA2FnC,CAAC"}
+{"version":3,"file":"index.d.ts","sourceRoot":"","sources":["../../../src/tools/review-pull-request/index.ts"],"names":[],"mappings":"AAAA,OAAO,EAEL,IAAI,EAEL,MAAM,uCAAuC,CAAC;AAU/C,eAAO,MAAM,qBAAqB,EAAE,IA0FnC,CAAC"}

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

@@ -3,11 +3,12 @@ Object.defineProperty(exports, "__esModule", { value: true });
 exports.reviewPullRequestTool = void 0;
 const code_review_1 = require("../../agent/code-review");
 const executor_1 = require("../../agent/code-review/executor");
+const file_info_1 = require("../../file-info");
 const review_pull_request_1 = require("../definitions/review-pull-request");
-const CODE_REVIEW_MODEL = "claude-sonnet-4-20250514";
+const CODE_REVIEW_MODEL = "claude-sonnet-4-5-20250929";
 exports.reviewPullRequestTool = {
     ...review_pull_request_1.reviewPullRequestTool,
-    execute: async ({ input, apiClient, chatSession, environmentOverrides, trace, featureFlags, repoPath, }) => {
+    execute: async ({ input, apiClient, environmentOverrides, trace, featureFlags, repoPath, }) => {
         try {
             const { sessionUrl } = input;
             if (!sessionUrl) {
@@ -29,7 +30,7 @@ exports.reviewPullRequestTool = {
                 },
             });
             const codeReviewToolExecutor = new executor_1.CodeReviewToolExecutor({
-                chatSession: chatSession || null,
+                chatSession: null,
                 repoPath,
                 apiClient,
                 trace: span,
@@ -42,6 +43,7 @@ exports.reviewPullRequestTool = {
                 chatState: undefined,
                 toolExecutor: codeReviewToolExecutor,
                 trace: span,
+                repoInfoBuilder: () => (0, file_info_1.getFileInfoFromFS)(repoPath),
             };
             const agent = new code_review_1.CodeReviewAgent({ ...agentParams });
             // Push the review prompt to the agent
@@ -50,7 +52,6 @@ exports.reviewPullRequestTool = {
             while (!agent.askUserForInput) {
                 await agent.runLoop({
                     reporter: async () => { },
-                    trace: span,
                 });
             }
             const agentResult = agent.getResult();

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

@@ -1,3 +1,3 @@
-import type { Tool } from "@empiricalrun/shared-types";
+import type { Tool } from "@empiricalrun/shared-types/chat-agent";
 export declare const runTestTool: Tool;
 //# sourceMappingURL=run-test.d.ts.map

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,EAAU,IAAI,EAAc,MAAM,4BAA4B,CAAC;AA0C3E,eAAO,MAAM,WAAW,EAAE,IA4GzB,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;AA0C/C,eAAO,MAAM,WAAW,EAAE,IA4GzB,CAAC"}

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

@@ -0,0 +1,3 @@
+import type { Tool } from "@empiricalrun/shared-types/chat-agent";
+export declare const safeBashTool: Tool;
+//# sourceMappingURL=index.d.ts.map

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

@@ -0,0 +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;AA4G9E,eAAO,MAAM,YAAY,EAAE,IAuE1B,CAAC"}

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

@@ -0,0 +1,128 @@
+"use strict";
+Object.defineProperty(exports, "__esModule", { value: true });
+exports.safeBashTool = void 0;
+const child_process_1 = require("child_process");
+const util_1 = require("util");
+const safe_bash_1 = require("../definitions/safe-bash");
+const execAsync = (0, util_1.promisify)(child_process_1.exec);
+const DANGEROUS_PATTERNS = [
+    /git\s+checkout/i,
+    /git\s+switch/i,
+    /git\s+reset\s+--hard/i,
+    /git\s+push/i,
+    /git\s+commit/i,
+    /\.git\/HEAD/i,
+    /symbolic-ref\s+HEAD/i,
+];
+async function getCurrentBranch(cwd) {
+    try {
+        const { stdout } = await execAsync("git rev-parse --abbrev-ref HEAD", {
+            cwd,
+        });
+        return stdout.trim();
+    }
+    catch {
+        return null;
+    }
+}
+async function getGitStatus(cwd) {
+    try {
+        const { stdout: stagedOutput } = await execAsync("git diff --cached --name-status", { cwd });
+        const { stdout: unstagedOutput } = await execAsync("git diff --name-status", { cwd });
+        const { stdout: untrackedOutput } = await execAsync("git ls-files --others --exclude-standard", { cwd });
+        const staged = stagedOutput.trim() || "No staged changes";
+        const unstagedLines = unstagedOutput.trim();
+        const untrackedLines = untrackedOutput
+            .trim()
+            .split("\n")
+            .filter(Boolean)
+            .map((f) => `??\t${f}`)
+            .join("\n");
+        const unstaged = [unstagedLines, untrackedLines].filter(Boolean).join("\n") ||
+            "No unstaged changes";
+        return { staged, unstaged };
+    }
+    catch {
+        return { staged: "Unable to get status", unstaged: "Unable to get status" };
+    }
+}
+function containsDangerousPattern(script) {
+    for (const pattern of DANGEROUS_PATTERNS) {
+        if (pattern.test(script)) {
+            return pattern.source;
+        }
+    }
+    return null;
+}
+function formatResult({ stdout, stderr, gitStatus, }) {
+    const parts = [];
+    if (stdout.trim()) {
+        parts.push(`## stdout\n\`\`\`\n${stdout.trim()}\n\`\`\``);
+    }
+    if (stderr.trim()) {
+        parts.push(`## stderr\n\`\`\`\n${stderr.trim()}\n\`\`\``);
+    }
+    parts.push(`## Staged (will be committed)\n\`\`\`\n${gitStatus.staged}\n\`\`\``);
+    parts.push(`## Unstaged\n\`\`\`\n${gitStatus.unstaged}\n\`\`\``);
+    return parts.join("\n\n");
+}
+exports.safeBashTool = {
+    ...safe_bash_1.safeBashTool,
+    execute: async ({ input, repoPath, logger }) => {
+        const { script, timeout = 60000 } = input;
+        const initialBranch = await getCurrentBranch(repoPath);
+        if (!initialBranch) {
+            return {
+                isError: true,
+                result: "Not in a git repository or unable to determine branch",
+            };
+        }
+        const dangerousPattern = containsDangerousPattern(script);
+        if (dangerousPattern) {
+            return {
+                isError: true,
+                result: `Script rejected: contains forbidden pattern '${dangerousPattern}'`,
+            };
+        }
+        let stdout = "";
+        let stderr = "";
+        try {
+            const result = await execAsync(script, {
+                cwd: repoPath,
+                timeout,
+                shell: "/bin/bash",
+            });
+            stdout = result.stdout;
+            stderr = result.stderr;
+        }
+        catch (error) {
+            const execError = error;
+            logger?.error("Error executing safe bash", {
+                tool: "safeBashTool",
+                error: execError.message || String(error),
+            });
+            const gitStatus = await getGitStatus(repoPath);
+            return {
+                isError: true,
+                result: formatResult({
+                    stdout: execError.stdout || "",
+                    stderr: execError.stderr || execError.message || "Unknown error",
+                    gitStatus,
+                }),
+            };
+        }
+        const finalBranch = await getCurrentBranch(repoPath);
+        if (finalBranch !== initialBranch) {
+            const gitStatus = await getGitStatus(repoPath);
+            return {
+                isError: true,
+                result: `Branch changed from '${initialBranch}' to '${finalBranch}' - execution rejected\n\n${formatResult({ stdout, stderr, gitStatus })}`,
+            };
+        }
+        const gitStatus = await getGitStatus(repoPath);
+        return {
+            isError: false,
+            result: formatResult({ stdout, stderr, gitStatus }),
+        };
+    },
+};

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

@@ -1,3 +1,3 @@
-import type { Tool } from "@empiricalrun/shared-types";
+import type { Tool } from "@empiricalrun/shared-types/chat-agent";
 export declare const generateTestWithBrowserAgent: Tool;
 //# sourceMappingURL=test-gen-browser.d.ts.map

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,EAAiB,IAAI,EAAE,MAAM,4BAA4B,CAAC;AAqBtE,eAAO,MAAM,4BAA4B,EAAE,IA+L1C,CAAC"}
+{"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;AAuB/C,eAAO,MAAM,4BAA4B,EAAE,IAyO1C,CAAC"}

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

@@ -9,6 +9,7 @@ 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 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");
@@ -16,7 +17,7 @@ const scenarios_1 = require("../bin/utils/scenarios");
 const test_gen_browser_1 = require("./definitions/test-gen-browser");
 exports.generateTestWithBrowserAgent = {
     ...test_gen_browser_1.generateTestWithBrowserAgent,
-    execute: async ({ input, repoPath, trace, collectArtifacts, environmentOverrides = {}, featureFlags, }) => {
+    execute: async ({ input, repoPath, trace, collectArtifacts, environmentOverrides = {}, featureFlags, logger, }) => {
         const { filePath, project } = input;
         const absoluteFilePath = path_1.default.join(repoPath, filePath);
         if (!fs_1.default.existsSync(absoluteFilePath)) {
@@ -35,7 +36,7 @@ exports.generateTestWithBrowserAgent = {
             }
         }
         catch (error) {
-            console.error("Error reading playwright config:", error);
+            logger.error("Error reading playwright config:", { error });
             return {
                 isError: true,
                 result: `Error reading playwright config: ${error}`,
@@ -67,12 +68,23 @@ exports.generateTestWithBrowserAgent = {
             };
         }
         try {
-            // Prepare playwright for codegen
-            console.log("[generateTestWithBrowserAgent] Preparing playwright for codegen");
-            await (0, pw_pause_1.preparePlaywrightForCodegen)(repoPath);
+            // Prepare playwright for codegen (only needed for patch-based approach)
+            const version = await (0, pw_codegen_1.detectPlaywrightVersion)(repoPath);
+            if (version) {
+                const approach = (0, pw_codegen_1.getCodegenApproach)(version);
+                logger.info("Preparing playwright for codegen", {
+                    tool: "generateTestWithBrowserAgent",
+                    version: version.full,
+                    approach,
+                });
+                await (0, pw_codegen_1.prepareCodegenIfNeeded)(repoPath, approach);
+            }
         }
         catch (err) {
-            console.warn("[generateTestWithBrowserAgent] Error preparing playwright for codegen", err);
+            logger.warn("Error preparing playwright for codegen", {
+                tool: "generateTestWithBrowserAgent",
+                error: err,
+            });
         }
         let envOverrides = {
             ...environmentOverrides,
@@ -86,7 +98,11 @@ exports.generateTestWithBrowserAgent = {
             };
         }
         const testGenToken = (0, scenarios_1.buildTokenFromOptions)({ name: testName, file: filePath }, { useComputerUseAgent: true });
-        console.log("[generateTestWithBrowserAgent] Validations passed, starting agent");
+        logger.info("Validations passed, starting agent", {
+            tool: "generateTestWithBrowserAgent",
+            testName,
+            filePath,
+        });
         const agentResult = await (0, run_1.runBrowsingAgent)({
             testCaseName: testName,
             testCaseSuites: suites,
@@ -103,15 +119,7 @@ exports.generateTestWithBrowserAgent = {
         await (0, pw_pause_1.revertToOriginalPwCode)(repoPath);
         fs_1.default.writeFileSync(absoluteFilePath, fileBackup, "utf-8");
         const { isError, error, result: toolResult, usage } = agentResult;
-        if (isError || !toolResult) {
-            return {
-                isError,
-                result: `Browser agent failed to run successfully. Here is the error:
-${error}
-`,
-            };
-        }
-        const screenshotArtifacts = toolResult
+        const screenshotArtifacts = (toolResult || [])
             .filter((item) => item.type === "screenshot")
             .map((item) => item.screenshot)
             .map((artifact) => ({
@@ -124,6 +132,43 @@ ${error}
             ...((0, utils_2.findPlaywrightArtifacts)(repoPath) || []),
             ...screenshotArtifacts,
         ];
+        const artifactUrlMap = new Map();
+        const pathToUrlMap = new Map();
+        if (collectArtifacts && artifactsInput.length > 0) {
+            const uploadedArtifacts = await collectArtifacts(artifactsInput);
+            artifactsInput.forEach((input, index) => {
+                const uploadedArtifact = uploadedArtifacts[index];
+                if (uploadedArtifact?.url) {
+                    artifactUrlMap.set(input, uploadedArtifact.url);
+                    if ("path" in input) {
+                        const relativePath = path_1.default.relative(repoPath, input.path);
+                        pathToUrlMap.set(relativePath, uploadedArtifact.url);
+                    }
+                }
+            });
+        }
+        if (isError || !toolResult) {
+            let errorMessage = `Browser agent failed to run successfully. Here is the error:
+${error}
+`;
+            if (typeof error === "string" && pathToUrlMap.size > 0) {
+                for (const [artifactPath, url] of pathToUrlMap.entries()) {
+                    // Escape special regex characters in the path before using it in RegExp
+                    const regex = new RegExp(artifactPath.replace(/[.*+?^${}()|[\]\\]/g, "\\$&"), "g");
+                    const matchCount = (errorMessage.match(regex) || []).length;
+                    if (matchCount === 0) {
+                        logger.info("No matches found for artifact path", {
+                            artifactPath,
+                        });
+                    }
+                    errorMessage = errorMessage.replace(regex, url);
+                }
+            }
+            return {
+                isError: true,
+                result: errorMessage,
+            };
+        }
         const summaryResult = `Browser agent has finished running.
 TODO comment which it worked on: ${todoContent}
 Summary of actions it took, and the generated Playwright code:`;
@@ -137,8 +182,6 @@ Summary of actions it took, and the generated Playwright code:`;
             ];
             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 },
@@ -147,9 +190,9 @@ Summary of actions it took, and the generated Playwright code:`;
                     return { type: "text", text: item.text };
                 }
                 else if (item.type === "screenshot") {
-                    const screenshotName = screenshotArtifacts[screenshotIndex]?.name;
-                    const url = screenshotName
-                        ? artifactUrlMap.get(screenshotName)
+                    const screenshotArtifact = screenshotArtifacts[screenshotIndex];
+                    const url = screenshotArtifact
+                        ? artifactUrlMap.get(screenshotArtifact)
                         : undefined;
                     screenshotIndex++;
                     return url

package/dist/tools/test-run-fetcher/index.d.ts

@@ -1,4 +1,4 @@
-import type { Tool } from "@empiricalrun/shared-types";
+import type { Tool } from "@empiricalrun/shared-types/chat-agent";
 export declare function extractPathAfterSourceRepo(fullPath: string): string;
-export declare const fetchTestRunReportTool: Tool;
+export declare const fetchTestRunDetailsTool: Tool;
 //# sourceMappingURL=index.d.ts.map

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

@@ -1 +1 @@
-{"version":3,"file":"index.d.ts","sourceRoot":"","sources":["../../../src/tools/test-run-fetcher/index.ts"],"names":[],"mappings":"AAAA,OAAO,KAAK,EAAE,IAAI,EAAE,MAAM,4BAA4B,CAAC;AAcvD,wBAAgB,0BAA0B,CAAC,QAAQ,EAAE,MAAM,GAAG,MAAM,CAOnE;AAED,eAAO,MAAM,sBAAsB,EAAE,IAkHpC,CAAC"}
+{"version":3,"file":"index.d.ts","sourceRoot":"","sources":["../../../src/tools/test-run-fetcher/index.ts"],"names":[],"mappings":"AAAA,OAAO,KAAK,EAAE,IAAI,EAAE,MAAM,uCAAuC,CAAC;AAYlE,wBAAgB,0BAA0B,CAAC,QAAQ,EAAE,MAAM,GAAG,MAAM,CAOnE;AAED,eAAO,MAAM,uBAAuB,EAAE,IAgHrC,CAAC"}

package/dist/tools/test-run-fetcher/index.js

@@ -1,13 +1,13 @@
 "use strict";
 Object.defineProperty(exports, "__esModule", { value: true });
-exports.fetchTestRunReportTool = void 0;
+exports.fetchTestRunDetailsTool = void 0;
 exports.extractPathAfterSourceRepo = extractPathAfterSourceRepo;
 const zod_1 = require("zod");
-const urls_1 = require("../utils/urls");
 const TestRunSchema = zod_1.z.object({
-    testRunUrl: zod_1.z
-        .string()
-        .describe("The full URL of the test run (e.g. https://dash.empirical.run/sortment-tests/test-runs/20269 or with query params like ?status=failed)"),
+    testRunId: zod_1.z
+        .number()
+        .describe("Numeric identifier (e.g. 20269) for a test run"),
+    projectSlug: zod_1.z.string().describe("Slug for the project of the test run"),
 });
 function extractPathAfterSourceRepo(fullPath) {
     const parts = fullPath.split("source-repo");
@@ -18,10 +18,24 @@ function extractPathAfterSourceRepo(fullPath) {
         return parts[1].replace(/^\/|\/$/g, "");
     }
 }
-exports.fetchTestRunReportTool = {
+exports.fetchTestRunDetailsTool = {
     schema: {
         name: "fetchTestRunDetails",
-        description: "Fetch details about a test run using its URL",
+        description: `Fetch details about a test run using its numeric identifier.
+This tool returns the status of the test run, along with information about the environment
+(e.g. "staging") or build that was used for the test run.
+
+## How to use this tool?
+
+When you are given a URL like https://dash.empirical.run/<project-slug>/test-runs/20269?query=params
+- "testRunId": Extract the numeric id from the URL - in this case #20269
+- "projectSlug": Extract the project-slug from the URL 
+
+## Related tools
+
+- To learn more about a particular test case that failed in the test, you can use the fetch diagnosis tool
+  with the diagnosis id (which is passed to the test run URL as the ?detail= query parameter)
+`,
         parameters: TestRunSchema,
     },
     needsBrowser: false,
@@ -30,9 +44,11 @@ exports.fetchTestRunReportTool = {
         if (!apiClient) {
             throw new Error("Dashboard API client is not available.");
         }
-        const { testRunUrl } = input;
+        const { testRunId: runId, projectSlug } = input;
         let data = null;
-        const { repoName, runId } = (0, urls_1.extractProjectFromUrl)(testRunUrl);
+        const repoName = projectSlug.includes("-tests")
+            ? projectSlug
+            : `${projectSlug}-tests`;
         try {
             data = await apiClient.request(`/api/test-runs/${runId}?repo_name=${repoName}`, { method: "GET" });
         }
@@ -42,79 +58,64 @@ exports.fetchTestRunReportTool = {
                 result: `Failed to fetch test run details: ${error instanceof Error ? error.message : String(error)}`,
             };
         }
-        // To efficiently use input_tokens, we
-        //   1. Truncate stack trace to last 300 characters
-        //   2. Remove request/response headers from network metadata
-        //   3. Focus on failed tests only
         const testRun = data.data?.test_run;
-        if (testRun?.flattenedSummaryDetails) {
-            testRun.flattenedSummaryDetails = testRun.flattenedSummaryDetails.map((detail) => {
-                // Truncate stack trace to last 300 characters
-                if (detail.failed_run_metadata?.stack) {
-                    const stack = detail.failed_run_metadata.stack;
-                    // TODO: change to last 10 lines
-                    detail.failed_run_metadata.stack =
-                        stack.length > 300 ? `...${stack.slice(-300)}` : stack;
-                }
-                // Remove headers from network metadata
-                if (detail.network_metadata?.failed_calls) {
-                    detail.network_metadata.failed_calls =
-                        detail.network_metadata.failed_calls.map((call) => ({
-                            ...call,
-                            request: { ...call.request, headers: [] },
-                            response: { ...call.response, headers: [] },
-                        }));
-                }
-                if (detail.network_metadata?.failed_calls_within_time_range) {
-                    detail.network_metadata.failed_calls_within_time_range =
-                        detail.network_metadata.failed_calls_within_time_range.map((call) => ({
-                            ...call,
-                            request: { ...call.request, headers: [] },
-                            response: { ...call.response, headers: [] },
-                        }));
-                }
-                return detail;
-            });
+        if (!testRun) {
+            return {
+                isError: true,
+                result: `Test run not found.`,
+            };
         }
-        const failedTests = testRun?.flattenedSummaryDetails.filter((detail) => detail.status === "failed");
-        const failedTestInMarkdown = failedTests
-            ?.map((detail) => {
+        const buildInfoIfAvailable = testRun.testRun.build_url
+            ? `
+    ## Build info
+    - Build URL: ${testRun.testRun.build_url}
+    - Git branch: ${testRun.testRun.build_branch}
+    - Git commit: ${testRun.testRun.commit} 
+    `
+            : ``;
+        const failedTestsInMarkdown = (testRun.flattenedSummaryDetails || [])
+            .filter((detail) => detail.status === "failed")
+            .map((detail) => {
             const { nesting } = detail.failed_run_metadata;
             return {
                 testName: nesting.slice(1).join(" > "),
                 fileName: extractPathAfterSourceRepo(detail.failed_run_metadata.location.file),
                 project: detail.test_project,
-                detail,
+                diagnosisId: detail.slug,
             };
         })
-            .map((res) => {
+            .map((res, idx) => {
             return `
-### Test name: ${res.testName}
-File: ${res.fileName}
-Project: ${res.project}
+    ### Failed test ${idx + 1}
 
-Failure data below
+    Test case name: ${res.testName}
+    File: ${res.fileName}
+    Project: ${res.project}
+    Diagnosis ID: ${res.diagnosisId}
 
-\`\`\`json
-${JSON.stringify(res.detail, null, 2)}
-\`\`\`
-`;
+    `;
         })
-            .join("\n\n");
+            .join("\n");
         const markdownResponse = `
-# Test run details
+    Test run #${testRun.testRun.id}
+
+    ## Info
+    - Run ID: ${runId}
+    - Status: ${testRun.testRun.state}
+    - Started at: ${testRun.testRun.run_started_at}
+    - Ended at: ${testRun.testRun.run_ended_at}
+    - Duration: ${testRun.testRun.duration} seconds
+
+    ## Environment
+    - Name: ${testRun?.testRun.environment_name}
+    - Slug: ${testRun?.testRun.environment_slug}
+
+    ${buildInfoIfAvailable}
 
-## Run info
-- Run ID: ${runId}
-- Repository: ${repoName}
-- Started at: ${testRun?.testRun.run_started_at}
-- Ended at: ${testRun?.testRun.run_ended_at}
-- Duration: ${testRun?.testRun.duration} seconds
-- Environment name: ${testRun?.testRun.environment_name}
+    ## Failed tests
 
-## Failed tests
-${failedTestInMarkdown}
-`;
+    ${failedTestsInMarkdown.length > 0 ? failedTestsInMarkdown : "None"}
+    `;
         return {
             result: markdownResponse,
             isError: false,

package/dist/tools/trace-dot-zip/index.d.ts

@@ -1,3 +1,3 @@
-import { Tool } from "@empiricalrun/shared-types";
+import { Tool } from "@empiricalrun/shared-types/chat-agent";
 export declare const traceDotZipTool: Tool;
 //# sourceMappingURL=index.d.ts.map

package/dist/tools/trace-dot-zip/index.d.ts.map

@@ -1 +1 @@
-{"version":3,"file":"index.d.ts","sourceRoot":"","sources":["../../../src/tools/trace-dot-zip/index.ts"],"names":[],"mappings":"AAAA,OAAO,EAAE,IAAI,EAAc,MAAM,4BAA4B,CAAC;AAa9D,eAAO,MAAM,eAAe,EAAE,IAwD7B,CAAC"}
+{"version":3,"file":"index.d.ts","sourceRoot":"","sources":["../../../src/tools/trace-dot-zip/index.ts"],"names":[],"mappings":"AAAA,OAAO,EAAE,IAAI,EAAc,MAAM,uCAAuC,CAAC;AAazE,eAAO,MAAM,eAAe,EAAE,IAwD7B,CAAC"}