@empiricalrun/test-gen 0.76.0 → 0.78.0

package/dist/agent/chat/exports.js

@@ -1,11 +1,12 @@
 "use strict";
 Object.defineProperty(exports, "__esModule", { value: true });
-exports.SUPPORTED_CHAT_MODELS = exports.createChatModel = exports.extractAttachments = exports.migrateChatState = exports.LATEST_CHAT_STATE_VERSION = exports.getLatestDownloadBuildUrl = exports.fetchToolCallFromId = exports.createChatStateForMessages = exports.createChatState = exports.chatStateFromModel = exports.CHAT_STATE_VERSIONS_MIGRATIONS_MAP = exports.resolveChatModelBasedOnInput = exports.getDefaultChatModelId = exports.ChatAgent = exports.TriageAgent = exports.MODE_TO_AGENT_MAP = exports.BaseAgent = exports.allToolsDefinitions = exports.viewFileUsingGitHub = exports.getFileInfoFromGitHub = void 0;
+exports.extractAttachments = exports.migrateChatState = exports.LATEST_CHAT_STATE_VERSION = exports.getLatestDownloadBuildUrl = exports.fetchToolCallFromId = exports.chatStateFromModel = exports.CHAT_STATE_VERSIONS_MIGRATIONS_MAP = exports.resolveChatModelBasedOnInput = exports.ChatAgent = exports.TriageAgent = exports.MODE_TO_AGENT_MAP = exports.BaseAgent = exports.viewFileUsingGitHub = exports.getFileInfoFromGitHub = exports.SUPPORTED_CHAT_MODELS = exports.createChatModel = void 0;
+var chat_1 = require("@empiricalrun/llm/chat");
+Object.defineProperty(exports, "createChatModel", { enumerable: true, get: function () { return chat_1.createChatModel; } });
+Object.defineProperty(exports, "SUPPORTED_CHAT_MODELS", { enumerable: true, get: function () { return chat_1.SUPPORTED_CHAT_MODELS; } });
 var github_1 = require("../../file-info/adapters/github");
 Object.defineProperty(exports, "getFileInfoFromGitHub", { enumerable: true, get: function () { return github_1.getFileInfoFromGitHub; } });
 Object.defineProperty(exports, "viewFileUsingGitHub", { enumerable: true, get: function () { return github_1.viewFileUsingGitHub; } });
-var tools_1 = require("../../tools");
-Object.defineProperty(exports, "allToolsDefinitions", { enumerable: true, get: function () { return tools_1.allToolsDefinitions; } });
 var base_1 = require("../base");
 Object.defineProperty(exports, "BaseAgent", { enumerable: true, get: function () { return base_1.BaseAgent; } });
 var index_1 = require("../index");
@@ -15,19 +16,13 @@ Object.defineProperty(exports, "TriageAgent", { enumerable: true, get: function
 var index_2 = require("./index");
 Object.defineProperty(exports, "ChatAgent", { enumerable: true, get: function () { return index_2.ChatAgent; } });
 var models_1 = require("./models");
-Object.defineProperty(exports, "getDefaultChatModelId", { enumerable: true, get: function () { return models_1.getDefaultChatModelId; } });
 Object.defineProperty(exports, "resolveChatModelBasedOnInput", { enumerable: true, get: function () { return models_1.resolveChatModelBasedOnInput; } });
 var state_1 = require("./state");
 Object.defineProperty(exports, "CHAT_STATE_VERSIONS_MIGRATIONS_MAP", { enumerable: true, get: function () { return state_1.CHAT_STATE_VERSIONS_MIGRATIONS_MAP; } });
 Object.defineProperty(exports, "chatStateFromModel", { enumerable: true, get: function () { return state_1.chatStateFromModel; } });
-Object.defineProperty(exports, "createChatState", { enumerable: true, get: function () { return state_1.createChatState; } });
-Object.defineProperty(exports, "createChatStateForMessages", { enumerable: true, get: function () { return state_1.createChatStateForMessages; } });
 Object.defineProperty(exports, "fetchToolCallFromId", { enumerable: true, get: function () { return state_1.fetchToolCallFromId; } });
 Object.defineProperty(exports, "getLatestDownloadBuildUrl", { enumerable: true, get: function () { return state_1.getLatestDownloadBuildUrl; } });
 Object.defineProperty(exports, "LATEST_CHAT_STATE_VERSION", { enumerable: true, get: function () { return state_1.LATEST_CHAT_STATE_VERSION; } });
 Object.defineProperty(exports, "migrateChatState", { enumerable: true, get: function () { return state_1.migrateChatState; } });
 var utils_1 = require("./utils");
 Object.defineProperty(exports, "extractAttachments", { enumerable: true, get: function () { return utils_1.extractAttachments; } });
-var chat_1 = require("@empiricalrun/llm/chat");
-Object.defineProperty(exports, "createChatModel", { enumerable: true, get: function () { return chat_1.createChatModel; } });
-Object.defineProperty(exports, "SUPPORTED_CHAT_MODELS", { enumerable: true, get: function () { return chat_1.SUPPORTED_CHAT_MODELS; } });

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

@@ -1,7 +1,7 @@
-import type { ToolDefinition } from "@empiricalrun/shared-types";
+import type { ToolsForLLM } from "@empiricalrun/shared-types";
 import { BaseAgent } from "../base";
 export declare class ChatAgent extends BaseAgent {
-    protected getTools(): ToolDefinition[];
+    protected getTools(): ToolsForLLM;
     protected buildSystemPrompt(repoContext?: string): Promise<string>;
 }
 //# sourceMappingURL=index.d.ts.map

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

@@ -1 +1 @@
-{"version":3,"file":"index.d.ts","sourceRoot":"","sources":["../../../src/agent/chat/index.ts"],"names":[],"mappings":"AAAA,OAAO,KAAK,EAAE,cAAc,EAAE,MAAM,4BAA4B,CAAC;AAQjE,OAAO,EAAE,SAAS,EAAE,MAAM,SAAS,CAAC;AAGpC,qBAAa,SAAU,SAAQ,SAAS;IACtC,SAAS,CAAC,QAAQ,IAAI,cAAc,EAAE;cAmBtB,iBAAiB,CAAC,WAAW,CAAC,EAAE,MAAM,GAAG,OAAO,CAAC,MAAM,CAAC;CAkHzE"}
+{"version":3,"file":"index.d.ts","sourceRoot":"","sources":["../../../src/agent/chat/index.ts"],"names":[],"mappings":"AAAA,OAAO,KAAK,EAAE,WAAW,EAAE,MAAM,4BAA4B,CAAC;AAS9D,OAAO,EAAE,SAAS,EAAE,MAAM,SAAS,CAAC;AAIpC,qBAAa,SAAU,SAAQ,SAAS;IACtC,SAAS,CAAC,QAAQ,IAAI,WAAW;cAajB,iBAAiB,CAAC,WAAW,CAAC,EAAE,MAAM,GAAG,OAAO,CAAC,MAAM,CAAC;CAuGzE"}

package/dist/agent/chat/index.js

@@ -4,20 +4,19 @@ exports.ChatAgent = void 0;
 const tools_1 = require("../../tools");
 const base_1 = require("../base");
 const pw_utils_docs_1 = require("./prompt/pw-utils-docs");
+const test_case_def_1 = require("./prompt/test-case-def");
 class ChatAgent extends base_1.BaseAgent {
     getTools() {
-        const tools = [
+        const custom = [
             ...tools_1.commonTools,
             ...tools_1.testGenerationTools,
-            // Model-specific tools
+            tools_1.analyseVideo,
             ...(0, tools_1.textEditorToolsForModel)(this.selectedModel),
         ];
-        // Feature flag: video analysis (if not already included)
-        if (this.featureFlags?.includes("useVideoAnalysis") &&
-            !tools.some((tool) => tool.schema.name === "fetchVideoAnalysis")) {
-            tools.push(tools_1.fetchVideoAnalysis);
-        }
-        return tools;
+        return {
+            custom,
+            builtInTextEditor: (0, tools_1.hasBuiltInTextEditor)(this.selectedModel),
+        };
     }
     async buildSystemPrompt(repoContext) {
         if (!repoContext) {
@@ -46,39 +45,26 @@ fixed with modifications to the test code, and it is your job to do that.
 
 # Tools
 
-You are given a set of tools to help you fulfill the user's request. Read their descriptions to
-understand what each tool does.
-
-For example, if the user asks you to run a test, you could use the runTest tool.
-Once the test is run, you will receive the results in the form of a JSON object.
-Summarize the results in a few sentences.
-
-If the user provides a diagnosis URL, you can use the fetchDiagnosisDetails tool
-to get more information about the test case and its results.
-
-If the user provides a test run URL, you can use the fetchTestRunDetails tool
-to get detailed information about a specific test run.
+You are given a set of tools (= functions) to help you fulfill the user's request. Read their descriptions
+to understand what each tool does. You are HIGHLY encouraged to use tools and generate a high quality output.
 
-Or if the user asks you to modify a test, you could use the generateTestWithBrowserAgent tool. If you suspect
-that a UI selector needs to be updated, using the browser agent is a good idea.
+For example,
 
-Before using generateTestWithBrowserAgent, you need to prepare the test code for the browser agent.
-You can do this by using the strReplaceEditor or the text editor tool to add a TODO comment to the test
-code. This comment explains to the browser agent what it needs to do.
+1. If you are adding or modifying a test, run the test to ensure it works well.
+2. Every test run generates report artifacts, and you have tools to read markdown files, see screenshots and videos. Use them to understand the test results
+   and fix test issues.
+3. Use the grep and file view tools to read the codebase. While the codebase only has Playwright tests, there are often patterns that can be reused.
 
-For example, if the expected modification is to click on a login button, you could add the following comment. 
-
-// TODO(agent): Click on the login button
+# Proactiveness
 
-The position of the comment is important: the browser agent will look for this comment and replace it with 
-the actual code to click on the login button. If you are fixing a failing test, your comment should be 
-around the failing line of code, so that it can be replaced/modified.
+You are encouraged to be proactive. Use the tools, explore the repo, and accomplish the task that the user has given to you. As long as your actions
+are aligned with the user's objective, you should lean on being more proactive than not.
 
-# Proactiveness
+If you are supposed to add or modify a test, understand the context, make changes, run the test, and go over this loop until you have achieved
+the user's goal - OR, ask the user for clarifications if there's some confusion. 
 
-1. You are allowed to be proactive, but ONLY for read-only tool calls: like searching for content, reading files, fetching data from tools, and
-running Playwright tests.
-2. For any read-write tool calls (e.g. modifying any file), you should share your plan and get the user's approval before proceeding.
+If you are unsure about next steps, or have some evidence to modify the user's objective, ask for clarification with a brief summary of 
+your findings and 1-2 questions for the user to help you understand the user's objective better.
 
 # Rules to follow
 
@@ -123,6 +109,8 @@ if (await saveButton.isVisible()) {
 `;
         return `${preamble}
 
+${test_case_def_1.testCasesDefinitionPrompt}
+
 # Recipes
 ${pw_utils_docs_1.playwrightUtilsDocs}
 

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

@@ -1,5 +1,3 @@
 import { SupportedChatModels } from "@empiricalrun/shared-types";
-export declare function getDefaultChatModelId(): SupportedChatModels;
 export declare const resolveChatModelBasedOnInput: (requestedModel?: string) => SupportedChatModels;
-export declare const modelLabels: Record<SupportedChatModels, string>;
 //# sourceMappingURL=models.d.ts.map

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

@@ -1 +1 @@
-{"version":3,"file":"models.d.ts","sourceRoot":"","sources":["../../../src/agent/chat/models.ts"],"names":[],"mappings":"AACA,OAAO,EAAE,mBAAmB,EAAE,MAAM,4BAA4B,CAAC;AAIjE,wBAAgB,qBAAqB,IAAI,mBAAmB,CAM3D;AAED,eAAO,MAAM,4BAA4B,GACvC,iBAAiB,MAAM,KACtB,mBA6BF,CAAC;AAEF,eAAO,MAAM,WAAW,EAAE,MAAM,CAAC,mBAAmB,EAAE,MAAM,CAOzD,CAAC"}
+{"version":3,"file":"models.d.ts","sourceRoot":"","sources":["../../../src/agent/chat/models.ts"],"names":[],"mappings":"AACA,OAAO,EAAE,mBAAmB,EAAE,MAAM,4BAA4B,CAAC;AAOjE,eAAO,MAAM,4BAA4B,GACvC,iBAAiB,MAAM,KACtB,mBAsBF,CAAC"}

package/dist/agent/chat/models.js

@@ -1,38 +1,24 @@
 "use strict";
 Object.defineProperty(exports, "__esModule", { value: true });
-exports.modelLabels = exports.resolveChatModelBasedOnInput = void 0;
-exports.getDefaultChatModelId = getDefaultChatModelId;
+exports.resolveChatModelBasedOnInput = void 0;
 const chat_1 = require("@empiricalrun/llm/chat");
 const model_1 = require("../../utils/model");
-function getDefaultChatModelId() {
-    const DEFAULT_CHAT_MODEL_ID = "claude-sonnet-4-20250514";
-    if (!chat_1.SUPPORTED_CHAT_MODELS.some((m) => m.id === DEFAULT_CHAT_MODEL_ID)) {
-        throw new Error("Default chat model is not in SUPPORTED_CHAT_MODELS");
-    }
-    return DEFAULT_CHAT_MODEL_ID;
-}
+const DEFAULT_AUTO_FALLBACK_MODEL = "claude-sonnet-4-5-20250929";
 const resolveChatModelBasedOnInput = (requestedModel) => {
     const sanitizedModelInput = requestedModel?.trim().toLowerCase();
     if (!sanitizedModelInput || sanitizedModelInput === "auto") {
-        const autoSelectedModelId = getDefaultChatModelId();
-        return autoSelectedModelId;
+        return DEFAULT_AUTO_FALLBACK_MODEL;
     }
-    // requestedModel is already a valid SupportedChatModels string
-    if (sanitizedModelInput &&
-        Object.values(chat_1.SUPPORTED_CHAT_MODELS).some((model) => model.id.toLowerCase() === sanitizedModelInput)) {
-        const matchedModel = chat_1.SUPPORTED_CHAT_MODELS.find((model) => model.id.toLowerCase() === sanitizedModelInput);
-        if (matchedModel) {
-            return matchedModel.id;
-        }
+    // Check for exact match first
+    const matchedModel = chat_1.SUPPORTED_CHAT_MODELS.find((model) => model.id.toLowerCase() === sanitizedModelInput);
+    if (matchedModel) {
+        return matchedModel.id;
     }
-    const resolvedModelId = model_1.ARGS_TO_MODEL_MAP[sanitizedModelInput];
-    if (!resolvedModelId) {
-        throw new Error(`Invalid chat model: ${requestedModel}`);
+    // Check for CLI args match (smaller aliases for full model names)
+    const resolvedModelId = model_1.CLI_ARGS_TO_MODEL_MAP[sanitizedModelInput];
+    if (resolvedModelId) {
+        return resolvedModelId;
     }
-    return resolvedModelId;
+    throw new Error(`Invalid chat model: ${requestedModel}`);
 };
 exports.resolveChatModelBasedOnInput = resolveChatModelBasedOnInput;
-exports.modelLabels = chat_1.SUPPORTED_CHAT_MODELS.reduce((acc, model) => ({
-    ...acc,
-    [model.id]: model.label,
-}), {});

package/dist/agent/chat/prompt/pw-utils-docs.d.ts

@@ -1,2 +1,2 @@
-export declare const playwrightUtilsDocs = "\nYou can refer to the following recipes to learn how to write tests for different scenarios.\n\n<email-automation>\n\n# Email automation\n\n## Example usage\n\n### Dynamic email\n\nThis dynamically generates a random email address that can \nbe used for the test (e.g. invite a new user).\n\n```ts\nimport { EmailClient } from \"@empiricalrun/playwright-utils\";\nimport { expect } from \"@playwright/test\";\n\nconst client = new EmailClient();\nconst address = client.getAddress();\n\n// Input the `address` in the application\n// that sends the email.\n\n// Get email received on the `address`\nconst email = await client.waitForEmail();\nexpect(\n  email.links.find((l) => l.text === \"Join your team\")\n).toBeTruthy();\n```\n\n### Static email\n\nThis uses a known (static) email that can be used to login\ninto an application.\n\nThis needs an email id (e.g. `test-login-user`). The email id\nis appended with the domain (managed internally) to get the full\nemail address.\n\n```ts\nimport { EmailClient } from \"@empiricalrun/playwright-utils\";\n\nconst emailId = `test-login-user`;\n\nconst client = new EmailClient({ emailId });\nconst address = client.getAddress(); // Returns full address with domain\n\n// Get email received on the `address`\nconst email = await client.waitForEmail();\n\n// Get login OTP\nconst loginCode = email.codes[0];\n```\n\n</email-automation>\n\n";
+export declare const playwrightUtilsDocs = "\nYou can refer to the following recipes to learn how to write tests for different scenarios.\n\n<email-automation>\n\n# Email automation\n\n## Example usage\n\n### Dynamic email\n\nThis dynamically generates a random email address that can \nbe used for the test (e.g. invite a new user).\n\n```ts\nimport { EmailClient } from \"@empiricalrun/playwright-utils\";\nimport { expect } from \"@playwright/test\";\n\nconst client = new EmailClient();\nconst address = client.getAddress();\n\n// Input the `address` in the application\n// that sends the email.\n\n// Get email received on the `address`\nconst email = await client.waitForEmail();\nexpect(\n  email.links.find((l) => l.text === \"Join your team\")\n).toBeTruthy();\n```\n\n### Static email\n\nThis uses a known (static) email that can be used to login\ninto an application.\n\nThis needs an email id (e.g. `test-login-user`). The email id\nis appended with the domain (managed internally) to get the full\nemail address.\n\n```ts\nimport { EmailClient } from \"@empiricalrun/playwright-utils\";\n\nconst emailId = `test-login-user`;\n\nconst client = new EmailClient({ emailId });\nconst address = client.getAddress(); // Returns full address with domain\n\n// Get email received on the `address`\nconst email = await client.waitForEmail();\n\n// Get login OTP\nconst loginCode = email.codes[0];\n```\n\n</email-automation>\n\n<fixtures>\n\nThe playwright-utils package provides fixtures that wrap around Playwright's built-in\n`page`, `context` fixtures to provide a mouse highlighter (which makes it easier to\nsee actions taken in a video).\n\nTo use this, you can use the `baseTestFixture` and `extendExpect` imports\nin your fixtures file.\n\n```ts\nimport { test as base, expect as baseExpect } from \"@playwright/test\";\nimport { baseTestFixture, extendExpect } from \"@empiricalrun/playwright-utils/test\";\n\nexport const test = baseTestFixture(base);\nexport const expect = extendExpect(baseExpect);\n```\n\n### Get a new browser context\n\nThis package provides a fixture `customContextPageProvider` which is a good way to create\na fresh, new browser context, and a page inside it.\n\nThere are two benefits of using this to create contexts or pages:\n1. Videos get recorded and attached to the test report\n2. Mouse highlights are available\n\n```ts\nimport { test, expect } from \"./fixtures\";\n\ntest(\"Example test\", async ({ page: builtInPage, customContextPageProvider }) => {\n  // builtInPage is from default browser context\n  const { page: newPage, context } = await customContextPageProvider();\n  // newPage is from this other browser context\n});\n```\n\ncustomContextPageProvider can accept options to customize the browser context.\n\n```\n  customContextPageProvider: (\n    options?: BrowserContextOptions,\n  ) => Promise<{ context: BrowserContext; page: Page }>;\n```\n\nFor example, pass { storageState: undefined } to create a new browser context without\nthe auth state of the current browser context. This is useful for multi-user scenarios.\n\n</fixtures>\n\n";
 //# sourceMappingURL=pw-utils-docs.d.ts.map

package/dist/agent/chat/prompt/pw-utils-docs.d.ts.map

@@ -1 +1 @@
-{"version":3,"file":"pw-utils-docs.d.ts","sourceRoot":"","sources":["../../../../src/agent/chat/prompt/pw-utils-docs.ts"],"names":[],"mappings":"AAoDA,eAAO,MAAM,mBAAmB,i3CAO/B,CAAC"}
+{"version":3,"file":"pw-utils-docs.d.ts","sourceRoot":"","sources":["../../../../src/agent/chat/prompt/pw-utils-docs.ts"],"names":[],"mappings":"AAqGA,eAAO,MAAM,mBAAmB,w/FAW/B,CAAC"}

package/dist/agent/chat/prompt/pw-utils-docs.js

@@ -1,6 +1,7 @@
 "use strict";
 Object.defineProperty(exports, "__esModule", { value: true });
 exports.playwrightUtilsDocs = void 0;
+// Ref: packages/playwright-utils/docs/email.md
 const emailRecipe = `
 # Email automation
 
@@ -52,6 +53,53 @@ const email = await client.waitForEmail();
 const loginCode = email.codes[0];
 \`\`\`
 `;
+// Ref: packages/playwright-utils/docs/fixtures.md
+const fixturesRecipe = `
+The playwright-utils package provides fixtures that wrap around Playwright's built-in
+\`page\`, \`context\` fixtures to provide a mouse highlighter (which makes it easier to
+see actions taken in a video).
+
+To use this, you can use the \`baseTestFixture\` and \`extendExpect\` imports
+in your fixtures file.
+
+\`\`\`ts
+import { test as base, expect as baseExpect } from "@playwright/test";
+import { baseTestFixture, extendExpect } from "@empiricalrun/playwright-utils/test";
+
+export const test = baseTestFixture(base);
+export const expect = extendExpect(baseExpect);
+\`\`\`
+
+### Get a new browser context
+
+This package provides a fixture \`customContextPageProvider\` which is a good way to create
+a fresh, new browser context, and a page inside it.
+
+There are two benefits of using this to create contexts or pages:
+1. Videos get recorded and attached to the test report
+2. Mouse highlights are available
+
+\`\`\`ts
+import { test, expect } from "./fixtures";
+
+test("Example test", async ({ page: builtInPage, customContextPageProvider }) => {
+  // builtInPage is from default browser context
+  const { page: newPage, context } = await customContextPageProvider();
+  // newPage is from this other browser context
+});
+\`\`\`
+
+customContextPageProvider can accept options to customize the browser context.
+
+\`\`\`
+  customContextPageProvider: (
+    options?: BrowserContextOptions,
+  ) => Promise<{ context: BrowserContext; page: Page }>;
+\`\`\`
+
+For example, pass { storageState: undefined } to create a new browser context without
+the auth state of the current browser context. This is useful for multi-user scenarios.
+`;
 exports.playwrightUtilsDocs = `
 You can refer to the following recipes to learn how to write tests for different scenarios.
 
@@ -59,4 +107,8 @@ You can refer to the following recipes to learn how to write tests for different
 ${emailRecipe}
 </email-automation>
 
+<fixtures>
+${fixturesRecipe}
+</fixtures>
+
 `;

package/dist/agent/chat/prompt/repo.d.ts.map

@@ -1 +1 @@
-{"version":3,"file":"repo.d.ts","sourceRoot":"","sources":["../../../../src/agent/chat/prompt/repo.ts"],"names":[],"mappings":"AAAA,OAAO,EAAE,QAAQ,EAAE,MAAM,4BAA4B,CAAC;AAsDtD,wBAAsB,iBAAiB,CAAC,QAAQ,EAAE,QAAQ,mBA0CzD"}
+{"version":3,"file":"repo.d.ts","sourceRoot":"","sources":["../../../../src/agent/chat/prompt/repo.ts"],"names":[],"mappings":"AAAA,OAAO,EAAE,QAAQ,EAAE,MAAM,4BAA4B,CAAC;AAuDtD,wBAAsB,iBAAiB,CAAC,QAAQ,EAAE,QAAQ,mBA2BzD"}

package/dist/agent/chat/prompt/repo.js

@@ -18,8 +18,8 @@ async function getAllMarkdownFiles(directory) {
         .filter((file) => file.type === "file" && file.path.endsWith(".md"))
         .map(async (file) => {
         return {
-            // TODO: Give full path
             name: path_1.default.basename(file.path),
+            path: file.path,
             content: file.type === "file" ? await file.getContent() : "",
         };
     }));
@@ -35,13 +35,14 @@ async function knowledgeContext(repoInfo) {
     const mdFiles = await getAllMarkdownFiles(knowledgeDir);
     const knowledge = mdFiles.map((file) => {
         return `
-    <knowledge_file>
-    <file_name>${file.name}</file_name>
-    <file_content>
-    ${file.content}
-    </file_content>
-    </knowledge_file>
-    `;
+<knowledge_file>
+<file_name>${file.name}</file_name>
+<file_path>${file.path}</file_path>
+<file_content>
+${file.content}
+</file_content>
+</knowledge_file>
+`;
     });
     return knowledge.join("\n");
 }
@@ -59,25 +60,13 @@ While specifying paths to files, use relative paths from the current working dir
 - Correct path: "tests/lesson.spec.ts"
 - Incorrect path: "/repo/tests/lesson.spec.ts" or "${repoInfo.name}/tests/lesson.spec.ts"
 `;
-    // try {
-    //   const playwrightConfig = await readPlaywrightConfig(process.cwd());
-    //   const validProjectNames = await getValidProjectNames(playwrightConfig);
-    //   if (validProjectNames.length > 0) {
-    //     REPO_CONTEXT_PROMPT += `
-    //   ## Playwright configuration
-    //   This repo is configured with these Playwright projects (in the playwright.config.ts file):
-    //   ${validProjectNames.map((name) => `- ${name}`).join("\n")}
-    //   `;
-    //   }
-    // } catch (error) {
-    //   console.warn("Failed to read playwright config", error);
-    // }
     const knowledge = await knowledgeContext(repoInfo);
     if (knowledge.length > 0) {
         REPO_CONTEXT_PROMPT += `
 ## Repo-specific knowledge
 
-You can use the following knowledge with your tasks.
+This repo knowledge files that will help you understand the repo and application better.
+
 <knowledge>
 ${knowledge}
 </knowledge>

package/dist/agent/chat/prompt/test-case-def.d.ts

@@ -0,0 +1,2 @@
+export declare const testCasesDefinitionPrompt = "\n# Specifying test cases for tools/functions\n\nMany tools and functions that you use require specifying a test case with the following properties: \n- Test case name: Name of the test (specified to the \"test(...)\" method)\n- Suites: Refers to describe blocks where this test is located, specified as an array of strings\n- File path: Path to the file - must be relative to the repo directory\n\n<example_1>\n// Contents of tests/example.spec.ts\nimport { test, expect } from \"./fixtures\";\n\ntest(\"has title\", async ({ page }) => {\n  await page.goto(\"https://playwright.dev/\");\n  await expect(page).toHaveTitle(/Playwright/);\n});\n\n// Specifying this test case\n- Test case name: \"has title\"\n- Suites: [] (since this test has no describe blocks)\n- File path: \"tests/example.spec.ts\"\n</example_1>\n\n<example_2>\n// Contents of tests/sidebar/nav.spec.ts\nimport { test, expect } from \"./fixtures\";\n\ntest.describe(\"Sidebar Navigation\", () => {\n  test(\"collapse sidebar and expand it by clicking settings gear icon\", async ({ page }) => {\n    await page.goto(\"/\");\n    await page.getByRole('button', { name: 'Collapse sidebar' }).click();\n    await page.getByRole('navigation').getByRole('button').click();\n  });\n});\n\n// Specifying this test case\n- Test case name: \"collapse sidebar and expand it by clicking settings gear icon\"\n- Suites: [\"Sidebar Navigation\"]\n- File path: \"tests/sidebar/nav.spec.ts\"\n</example_2>\n";
+//# sourceMappingURL=test-case-def.d.ts.map

package/dist/agent/chat/prompt/test-case-def.d.ts.map

@@ -0,0 +1 @@
+{"version":3,"file":"test-case-def.d.ts","sourceRoot":"","sources":["../../../../src/agent/chat/prompt/test-case-def.ts"],"names":[],"mappings":"AAAA,eAAO,MAAM,yBAAyB,47CAwCrC,CAAC"}

package/dist/agent/chat/prompt/test-case-def.js

@@ -0,0 +1,44 @@
+"use strict";
+Object.defineProperty(exports, "__esModule", { value: true });
+exports.testCasesDefinitionPrompt = void 0;
+exports.testCasesDefinitionPrompt = `
+# Specifying test cases for tools/functions
+
+Many tools and functions that you use require specifying a test case with the following properties: 
+- Test case name: Name of the test (specified to the "test(...)" method)
+- Suites: Refers to describe blocks where this test is located, specified as an array of strings
+- File path: Path to the file - must be relative to the repo directory
+
+<example_1>
+// Contents of tests/example.spec.ts
+import { test, expect } from "./fixtures";
+
+test("has title", async ({ page }) => {
+  await page.goto("https://playwright.dev/");
+  await expect(page).toHaveTitle(/Playwright/);
+});
+
+// Specifying this test case
+- Test case name: "has title"
+- Suites: [] (since this test has no describe blocks)
+- File path: "tests/example.spec.ts"
+</example_1>
+
+<example_2>
+// Contents of tests/sidebar/nav.spec.ts
+import { test, expect } from "./fixtures";
+
+test.describe("Sidebar Navigation", () => {
+  test("collapse sidebar and expand it by clicking settings gear icon", async ({ page }) => {
+    await page.goto("/");
+    await page.getByRole('button', { name: 'Collapse sidebar' }).click();
+    await page.getByRole('navigation').getByRole('button').click();
+  });
+});
+
+// Specifying this test case
+- Test case name: "collapse sidebar and expand it by clicking settings gear icon"
+- Suites: ["Sidebar Navigation"]
+- File path: "tests/sidebar/nav.spec.ts"
+</example_2>
+`;

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

@@ -1,16 +1,8 @@
-import { IChatModel } from "@empiricalrun/llm/chat";
-import { Attachment, CanonicalMessage, ChatState, ChatStateError, PendingToolCall, SupportedChatModels, ToolResult } from "@empiricalrun/shared-types";
+import { IChatCanonicalModel, IChatModel } from "@empiricalrun/llm/chat";
+import { CanonicalMessage, ChatState, ChatStateError, PendingToolCall, SupportedChatModels, ToolResult } from "@empiricalrun/shared-types";
 export declare const CHAT_STATE_VERSIONS_MIGRATIONS_MAP: Record<string, (state: any) => any>;
 export declare const LATEST_CHAT_STATE_VERSION = "0.1";
 export declare const CHAT_STATE_PATH: string;
-export declare function clearChatState(): void;
-export declare function createChatState({ userPrompt, attachments, existingState, selectedModel, error, }: {
-    userPrompt: string | undefined;
-    attachments: Attachment[];
-    existingState: ChatState | undefined;
-    selectedModel: SupportedChatModels;
-    error: ChatStateError | null;
-}): ChatState;
 export declare function createChatStateForMessages({ messages, selectedModel, askUserForInput, error, }: {
     messages: CanonicalMessage[];
     selectedModel: SupportedChatModels;
@@ -18,17 +10,19 @@ export declare function createChatStateForMessages({ messages, selectedModel, as
     error: ChatStateError | null;
 }): ChatState;
 export declare function chatStateFromModel<T>({ chatModel, error, }: {
-    chatModel: IChatModel<T>;
+    chatModel: IChatModel<T> | IChatCanonicalModel;
     error: ChatStateError | null;
 }): ChatState;
-export declare function getUsageSummary(chatModel: IChatModel<CanonicalMessage>): string;
-export declare function loadChatState(): ChatState | undefined;
+export declare function getUsageSummary(chatState: ChatState): string;
+export declare function loadChatState({ resetChat, }: {
+    resetChat: boolean;
+}): ChatState | undefined;
 /**
  * Migrates a chat state object from an old version to the latest version.
  * Add migration logic for each version as needed.
  */
 export declare function migrateChatState(oldState: any): ChatState;
-export declare function saveToDisk(messages: Array<CanonicalMessage>, selectedModel: SupportedChatModels, askUserForInput: boolean, error: ChatStateError | null): void;
+export declare function saveToDisk(chatState: ChatState): void;
 export declare function getLatestDownloadBuildUrl(messages: CanonicalMessage[]): string | null;
 export declare function fetchToolCallFromId(toolCallId: String, messages: CanonicalMessage[]): {
     toolCallQueuedAt: Date | undefined;

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

@@ -1 +1 @@
-{"version":3,"file":"state.d.ts","sourceRoot":"","sources":["../../../src/agent/chat/state.ts"],"names":[],"mappings":"AAAA,OAAO,EAIL,UAAU,EACX,MAAM,wBAAwB,CAAC;AAChC,OAAO,EACL,UAAU,EACV,gBAAgB,EAChB,SAAS,EACT,cAAc,EACd,eAAe,EACf,mBAAmB,EAEnB,UAAU,EACX,MAAM,4BAA4B,CAAC;AA+BpC,eAAO,MAAM,kCAAkC,EAAE,MAAM,CACrD,MAAM,EACN,CAAC,KAAK,EAAE,GAAG,KAAK,GAAG,CAIpB,CAAC;AAEF,eAAO,MAAM,yBAAyB,QAAQ,CAAC;AAE/C,eAAO,MAAM,eAAe,QAI3B,CAAC;AAEF,wBAAgB,cAAc,SAQ7B;AAED,wBAAgB,eAAe,CAAC,EAC9B,UAAU,EACV,WAAW,EACX,aAAa,EACb,aAAa,EACb,KAAK,GACN,EAAE;IACD,UAAU,EAAE,MAAM,GAAG,SAAS,CAAC;IAC/B,WAAW,EAAE,UAAU,EAAE,CAAC;IAC1B,aAAa,EAAE,SAAS,GAAG,SAAS,CAAC;IACrC,aAAa,EAAE,mBAAmB,CAAC;IACnC,KAAK,EAAE,cAAc,GAAG,IAAI,CAAC;CAC9B,aAYA;AAED,wBAAgB,0BAA0B,CAAC,EACzC,QAAQ,EACR,aAAa,EACb,eAAe,EACf,KAAK,GACN,EAAE;IACD,QAAQ,EAAE,gBAAgB,EAAE,CAAC;IAC7B,aAAa,EAAE,mBAAmB,CAAC;IACnC,eAAe,EAAE,OAAO,CAAC;IACzB,KAAK,EAAE,cAAc,GAAG,IAAI,CAAC;CAC9B,GAAG,SAAS,CASZ;AAED,wBAAgB,kBAAkB,CAAC,CAAC,EAAE,EACpC,SAAS,EACT,KAAK,GACN,EAAE;IACD,SAAS,EAAE,UAAU,CAAC,CAAC,CAAC,CAAC;IACzB,KAAK,EAAE,cAAc,GAAG,IAAI,CAAC;CAC9B,aAOA;AAED,wBAAgB,eAAe,CAC7B,SAAS,EAAE,UAAU,CAAC,gBAAgB,CAAC,GACtC,MAAM,CA0BR;AAED,wBAAgB,aAAa,IAAI,SAAS,GAAG,SAAS,CAarD;AAED;;;GAGG;AACH,wBAAgB,gBAAgB,CAAC,QAAQ,EAAE,GAAG,GAAG,SAAS,CAqBzD;AAED,wBAAgB,UAAU,CACxB,QAAQ,EAAE,KAAK,CAAC,gBAAgB,CAAC,EACjC,aAAa,EAAE,mBAAmB,EAClC,eAAe,EAAE,OAAO,EACxB,KAAK,EAAE,cAAc,GAAG,IAAI,QAgB7B;AA2BD,wBAAgB,yBAAyB,CACvC,QAAQ,EAAE,gBAAgB,EAAE,GAC3B,MAAM,GAAG,IAAI,CAef;AAED,wBAAgB,mBAAmB,CACjC,UAAU,EAAE,MAAM,EAClB,QAAQ,EAAE,gBAAgB,EAAE,GAC3B;IACD,gBAAgB,EAAE,IAAI,GAAG,SAAS,CAAC;IACnC,QAAQ,EAAE,eAAe,GAAG,SAAS,CAAC;IACtC,UAAU,EAAE,UAAU,GAAG,SAAS,CAAC;CACpC,CA0BA"}
+{"version":3,"file":"state.d.ts","sourceRoot":"","sources":["../../../src/agent/chat/state.ts"],"names":[],"mappings":"AAAA,OAAO,EAAE,mBAAmB,EAAE,UAAU,EAAE,MAAM,wBAAwB,CAAC;AACzE,OAAO,EACL,gBAAgB,EAChB,SAAS,EACT,cAAc,EACd,eAAe,EACf,mBAAmB,EAEnB,UAAU,EACX,MAAM,4BAA4B,CAAC;AAMpC,eAAO,MAAM,kCAAkC,EAAE,MAAM,CACrD,MAAM,EACN,CAAC,KAAK,EAAE,GAAG,KAAK,GAAG,CAGpB,CAAC;AAEF,eAAO,MAAM,yBAAyB,QAAQ,CAAC;AAE/C,eAAO,MAAM,eAAe,QAI3B,CAAC;AAEF,wBAAgB,0BAA0B,CAAC,EACzC,QAAQ,EACR,aAAa,EACb,eAAe,EACf,KAAK,GACN,EAAE;IACD,QAAQ,EAAE,gBAAgB,EAAE,CAAC;IAC7B,aAAa,EAAE,mBAAmB,CAAC;IACnC,eAAe,EAAE,OAAO,CAAC;IACzB,KAAK,EAAE,cAAc,GAAG,IAAI,CAAC;CAC9B,GAAG,SAAS,CASZ;AAED,wBAAgB,kBAAkB,CAAC,CAAC,EAAE,EACpC,SAAS,EACT,KAAK,GACN,EAAE;IACD,SAAS,EAAE,UAAU,CAAC,CAAC,CAAC,GAAG,mBAAmB,CAAC;IAC/C,KAAK,EAAE,cAAc,GAAG,IAAI,CAAC;CAC9B,aAOA;AAED,wBAAgB,eAAe,CAAC,SAAS,EAAE,SAAS,GAAG,MAAM,CA0B5D;AAED,wBAAgB,aAAa,CAAC,EAC5B,SAAS,GACV,EAAE;IACD,SAAS,EAAE,OAAO,CAAC;CACpB,GAAG,SAAS,GAAG,SAAS,CAsBxB;AAED;;;GAGG;AACH,wBAAgB,gBAAgB,CAAC,QAAQ,EAAE,GAAG,GAAG,SAAS,CAqBzD;AAED,wBAAgB,UAAU,CAAC,SAAS,EAAE,SAAS,QAQ9C;AA2BD,wBAAgB,yBAAyB,CACvC,QAAQ,EAAE,gBAAgB,EAAE,GAC3B,MAAM,GAAG,IAAI,CAef;AAED,wBAAgB,mBAAmB,CACjC,UAAU,EAAE,MAAM,EAClB,QAAQ,EAAE,gBAAgB,EAAE,GAC3B;IACD,gBAAgB,EAAE,IAAI,GAAG,SAAS,CAAC;IACnC,QAAQ,EAAE,eAAe,GAAG,SAAS,CAAC;IACtC,UAAU,EAAE,UAAU,GAAG,SAAS,CAAC;CACpC,CA0BA"}

package/dist/agent/chat/state.js

@@ -4,8 +4,6 @@ var __importDefault = (this && this.__importDefault) || function (mod) {
 };
 Object.defineProperty(exports, "__esModule", { value: true });
 exports.CHAT_STATE_PATH = exports.LATEST_CHAT_STATE_VERSION = exports.CHAT_STATE_VERSIONS_MIGRATIONS_MAP = void 0;
-exports.clearChatState = clearChatState;
-exports.createChatState = createChatState;
 exports.createChatStateForMessages = createChatStateForMessages;
 exports.chatStateFromModel = chatStateFromModel;
 exports.getUsageSummary = getUsageSummary;
@@ -14,59 +12,13 @@ exports.migrateChatState = migrateChatState;
 exports.saveToDisk = saveToDisk;
 exports.getLatestDownloadBuildUrl = getLatestDownloadBuildUrl;
 exports.fetchToolCallFromId = fetchToolCallFromId;
-const chat_1 = require("@empiricalrun/llm/chat");
 const fs_1 = __importDefault(require("fs"));
 const path_1 = __importDefault(require("path"));
-// Migration wrapper for v20250327.1 -> v0.1 chat state versions
-// v20250327.1 was model-specific, but v0.1 is canonical
-function migrateToV01(oldState) {
-    if (oldState.model &&
-        typeof oldState.model === "string" &&
-        Array.isArray(oldState.messages)) {
-        const provider = (0, chat_1.getProviderForModel)(oldState.model);
-        if (provider === "google") {
-            return {
-                ...oldState,
-                version: "0.1",
-                messages: oldState.messages.map(chat_1.geminiToCanonical),
-            };
-        }
-        else {
-            throw new Error(`Unsupported state for migration with model: ${oldState.model} and version: ${oldState.version}`);
-        }
-    }
-    // If not Gemini or not matching, return the old state
-    return oldState;
-}
 exports.CHAT_STATE_VERSIONS_MIGRATIONS_MAP = {
-    "20250327.1": migrateToV01,
     "0.1": (state) => state,
 };
 exports.LATEST_CHAT_STATE_VERSION = "0.1";
 exports.CHAT_STATE_PATH = path_1.default.join(process.cwd(), ".empiricalrun", "last-chat.json");
-function clearChatState() {
-    try {
-        if (fs_1.default.existsSync(exports.CHAT_STATE_PATH)) {
-            fs_1.default.unlinkSync(exports.CHAT_STATE_PATH);
-        }
-    }
-    catch {
-        // Swallow errors to avoid blocking the CLI; deletion is best-effort
-    }
-}
-function createChatState({ userPrompt, attachments, existingState, selectedModel, error, }) {
-    const messages = existingState?.messages || [];
-    const chatModel = (0, chat_1.createChatModel)(messages, selectedModel);
-    if (userPrompt) {
-        chatModel.pushUserMessage(userPrompt, attachments);
-    }
-    return createChatStateForMessages({
-        messages: chatModel.messages,
-        selectedModel,
-        askUserForInput: chatModel.askUserForInput,
-        error,
-    });
-}
 function createChatStateForMessages({ messages, selectedModel, askUserForInput, error, }) {
     // TODO: Add better types for messages
     return {
@@ -85,8 +37,8 @@ function chatStateFromModel({ chatModel, error, }) {
         error,
     });
 }
-function getUsageSummary(chatModel) {
-    const messages = chatModel.messages;
+function getUsageSummary(chatState) {
+    const messages = chatState.messages;
     const messageUsages = messages.map((msg) => msg.usage).filter(Boolean);
     const inputTokens = messageUsages.reduce((sum, usage) => sum + (usage?.tokens?.input || 0), 0);
     const outputTokens = messageUsages.reduce((sum, usage) => sum + (usage?.tokens?.output || 0), 0);
@@ -103,7 +55,17 @@ function getUsageSummary(chatModel) {
 Input: ${inputTokens.toLocaleString()} tokens / Output: ${outputTokens.toLocaleString()} tokens`;
     return msgsSummary;
 }
-function loadChatState() {
+function loadChatState({ resetChat, }) {
+    if (resetChat) {
+        try {
+            if (fs_1.default.existsSync(exports.CHAT_STATE_PATH)) {
+                fs_1.default.unlinkSync(exports.CHAT_STATE_PATH);
+            }
+        }
+        catch {
+            // Swallow errors to avoid blocking the CLI; deletion is best-effort
+        }
+    }
     if (!fs_1.default.existsSync(exports.CHAT_STATE_PATH)) {
         return undefined;
     }
@@ -141,21 +103,14 @@ function migrateChatState(oldState) {
         ...migrated,
     };
 }
-function saveToDisk(messages, selectedModel, askUserForInput, error) {
+function saveToDisk(chatState) {
     const statePath = exports.CHAT_STATE_PATH;
     // Ensure directory exists before trying to read/write
     const dirname = path_1.default.dirname(statePath);
     if (!fs_1.default.existsSync(dirname)) {
         fs_1.default.mkdirSync(dirname, { recursive: true });
     }
-    // Use the helper to build the new state
-    const newState = createChatStateForMessages({
-        messages,
-        selectedModel,
-        askUserForInput,
-        error,
-    });
-    fs_1.default.writeFileSync(statePath, JSON.stringify(newState, null, 2));
+    fs_1.default.writeFileSync(statePath, JSON.stringify(chatState, null, 2));
 }
 function findLatestToolCall(messages, toolName) {
     return messages

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

@@ -1,11 +1,11 @@
 import { TraceClient } from "@empiricalrun/llm";
-import { IChatModel } from "@empiricalrun/llm/chat";
+import { IChatCanonicalModel, IChatModel } from "@empiricalrun/llm/chat";
 import { Attachment, ReporterFunction } from "@empiricalrun/shared-types";
 export declare const log: (...args: any[]) => void;
 export declare function getModelName(model: string): string;
 export declare function handleAgentError({ error, chatModel, reporter, trace, }: {
     error: unknown;
-    chatModel: IChatModel<any>;
+    chatModel: IChatModel<any> | IChatCanonicalModel;
     reporter: ReporterFunction;
     trace?: TraceClient;
 }): Promise<void>;

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

@@ -1 +1 @@
-{"version":3,"file":"utils.d.ts","sourceRoot":"","sources":["../../../src/agent/chat/utils.ts"],"names":[],"mappings":"AAAA,OAAO,EAAE,WAAW,EAAE,MAAM,mBAAmB,CAAC;AAChD,OAAO,EAAuB,UAAU,EAAE,MAAM,wBAAwB,CAAC;AACzE,OAAO,EACL,UAAU,EAEV,gBAAgB,EACjB,MAAM,4BAA4B,CAAC;AAKpC,eAAO,MAAM,GAAG,GAAI,GAAG,MAAM,GAAG,EAAE,SAEjC,CAAC;AAcF,wBAAgB,YAAY,CAAC,KAAK,EAAE,MAAM,GAAG,MAAM,CAMlD;AAED,wBAAsB,gBAAgB,CAAC,EACrC,KAAK,EACL,SAAS,EACT,QAAQ,EACR,KAAK,GACN,EAAE;IACD,KAAK,EAAE,OAAO,CAAC;IACf,SAAS,EAAE,UAAU,CAAC,GAAG,CAAC,CAAC;IAC3B,QAAQ,EAAE,gBAAgB,CAAC;IAC3B,KAAK,CAAC,EAAE,WAAW,CAAC;CACrB,iBAeA;AAED,wBAAgB,kBAAkB,CAAC,UAAU,EAAE,MAAM,GAAG;IACtD,IAAI,EAAE,MAAM,CAAC;IACb,WAAW,EAAE,UAAU,EAAE,CAAC;CAC3B,CAsBA"}
+{"version":3,"file":"utils.d.ts","sourceRoot":"","sources":["../../../src/agent/chat/utils.ts"],"names":[],"mappings":"AAAA,OAAO,EAAE,WAAW,EAAE,MAAM,mBAAmB,CAAC;AAChD,OAAO,EAGL,mBAAmB,EACnB,UAAU,EACX,MAAM,wBAAwB,CAAC;AAChC,OAAO,EACL,UAAU,EAEV,gBAAgB,EACjB,MAAM,4BAA4B,CAAC;AAKpC,eAAO,MAAM,GAAG,GAAI,GAAG,MAAM,GAAG,EAAE,SAEjC,CAAC;AAcF,wBAAgB,YAAY,CAAC,KAAK,EAAE,MAAM,GAAG,MAAM,CAMlD;AAED,wBAAsB,gBAAgB,CAAC,EACrC,KAAK,EACL,SAAS,EACT,QAAQ,EACR,KAAK,GACN,EAAE;IACD,KAAK,EAAE,OAAO,CAAC;IACf,SAAS,EAAE,UAAU,CAAC,GAAG,CAAC,GAAG,mBAAmB,CAAC;IACjD,QAAQ,EAAE,gBAAgB,CAAC;IAC3B,KAAK,CAAC,EAAE,WAAW,CAAC;CACrB,iBAaA;AASD,wBAAgB,kBAAkB,CAAC,UAAU,EAAE,MAAM,GAAG;IACtD,IAAI,EAAE,MAAM,CAAC;IACb,WAAW,EAAE,UAAU,EAAE,CAAC;CAC3B,CA4BA"}