@empiricalrun/test-gen 0.75.0 → 0.77.0

package/dist/agent/chat/agent-loop.js

@@ -3,7 +3,7 @@ Object.defineProperty(exports, "__esModule", { value: true });
 exports.chatAgentLoop = chatAgentLoop;
 const state_1 = require("./state");
 const utils_1 = require("./utils");
-async function chatAgentLoop({ chatModel, selectedModel, reporter, streamingMessageReporter, trace, signal, featureFlags, tools, onPendingToolCall, onLLMResponse, systemPrompt, }) {
+async function chatAgentLoop({ chatModel, reporter, streamingMessageReporter, trace, signal, featureFlags, tools, onPendingToolCall, onLLMResponse, systemPrompt, }) {
     trace?.update({ input: { systemPrompt } });
     while (!chatModel.askUserForInput) {
         if (signal?.aborted) {
@@ -16,42 +16,31 @@ async function chatAgentLoop({ chatModel, selectedModel, reporter, streamingMess
                 await onPendingToolCall(toolCalls);
                 break;
             }
-            (0, utils_1.log)(`${(0, utils_1.getModelName)(selectedModel)} is working...`);
+            (0, utils_1.log)(`${(0, utils_1.getModelName)(chatModel.selectedModel)} is working...`);
             const response = await chatModel.getLLMResponse({
                 systemPrompt,
-                tools: tools.map((t) => t.schema),
-                selectedModel,
+                tools,
                 trace,
                 streamingMessageReporter,
                 hasThinkingEnabled: true,
-                hasInterleavedThinkingEnabled: featureFlags.includes("interleavedThinkingClaude"),
+                featureFlags,
             });
             if (!response) {
                 throw new Error("Error getting response from LLM");
             }
             chatModel.pushMessage(response);
             if (onLLMResponse) {
-                await onLLMResponse(response, selectedModel);
+                await onLLMResponse(response, chatModel.selectedModel);
             }
             const latest = chatModel.getHumanReadableLatestMessage();
             if (signal?.aborted) {
                 (0, utils_1.log)("[chatAgentLoop] Aborted by signal");
                 break;
             }
-            await reporter((0, state_1.chatStateFromModel)({
-                chatModel,
-                selectedModel,
-                error: null,
-            }), latest);
+            await reporter((0, state_1.chatStateFromModel)({ chatModel, error: null }), latest);
         }
         catch (error) {
-            await (0, utils_1.handleAgentError)({
-                error,
-                chatModel,
-                selectedModel,
-                reporter,
-                trace,
-            });
+            await (0, utils_1.handleAgentError)({ error, chatModel, reporter, trace });
             throw error;
         }
     }

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

@@ -1,10 +1,13 @@
+export type { IChatModel } from "@empiricalrun/llm/chat";
+export { createChatModel, SUPPORTED_CHAT_MODELS } from "@empiricalrun/llm/chat";
 export { getFileInfoFromGitHub, viewFileUsingGitHub, } from "../../file-info/adapters/github";
-export { allToolsDefinitions } from "../../tools";
-export { ChatAgent, getAgentForMode, TriageAgent } from "../base";
-export { getDefaultChatModelId, resolveChatModelBasedOnInput } from "./models";
+export type { AgentParams } from "../base";
+export { BaseAgent } from "../base";
+export type { CodeReviewResultV0, CodeReviewResultV1, CodeReviewResultV2, CodeReviewVersionedResult, } from "../code-review";
+export { MODE_TO_AGENT_MAP } from "../index";
+export { TriageAgent } from "../triage";
+export { ChatAgent } from "./index";
+export { resolveChatModelBasedOnInput } from "./models";
 export { CHAT_STATE_VERSIONS_MIGRATIONS_MAP, chatStateFromModel, createChatState, createChatStateForMessages, fetchToolCallFromId, getLatestDownloadBuildUrl, LATEST_CHAT_STATE_VERSION, migrateChatState, } from "./state";
 export { extractAttachments } from "./utils";
-export { processToolCallsAndUpdateChatState, rejectPendingToolCalls, } from "./utils/tool-calls";
-export type { IChatModel } from "@empiricalrun/llm/chat";
-export { createChatModel, SUPPORTED_CHAT_MODELS } from "@empiricalrun/llm/chat";
 //# sourceMappingURL=exports.d.ts.map

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

@@ -1 +1 @@
-{"version":3,"file":"exports.d.ts","sourceRoot":"","sources":["../../../src/agent/chat/exports.ts"],"names":[],"mappings":"AAAA,OAAO,EACL,qBAAqB,EACrB,mBAAmB,GACpB,MAAM,iCAAiC,CAAC;AACzC,OAAO,EAAE,mBAAmB,EAAE,MAAM,aAAa,CAAC;AAClD,OAAO,EAAE,SAAS,EAAE,eAAe,EAAE,WAAW,EAAE,MAAM,SAAS,CAAC;AAClE,OAAO,EAAE,qBAAqB,EAAE,4BAA4B,EAAE,MAAM,UAAU,CAAC;AAC/E,OAAO,EACL,kCAAkC,EAClC,kBAAkB,EAClB,eAAe,EACf,0BAA0B,EAC1B,mBAAmB,EACnB,yBAAyB,EACzB,yBAAyB,EACzB,gBAAgB,GACjB,MAAM,SAAS,CAAC;AACjB,OAAO,EAAE,kBAAkB,EAAE,MAAM,SAAS,CAAC;AAC7C,OAAO,EACL,kCAAkC,EAClC,sBAAsB,GACvB,MAAM,oBAAoB,CAAC;AAC5B,YAAY,EAAE,UAAU,EAAE,MAAM,wBAAwB,CAAC;AACzD,OAAO,EAAE,eAAe,EAAE,qBAAqB,EAAE,MAAM,wBAAwB,CAAC"}
+{"version":3,"file":"exports.d.ts","sourceRoot":"","sources":["../../../src/agent/chat/exports.ts"],"names":[],"mappings":"AAAA,YAAY,EAAE,UAAU,EAAE,MAAM,wBAAwB,CAAC;AACzD,OAAO,EAAE,eAAe,EAAE,qBAAqB,EAAE,MAAM,wBAAwB,CAAC;AAChF,OAAO,EACL,qBAAqB,EACrB,mBAAmB,GACpB,MAAM,iCAAiC,CAAC;AACzC,YAAY,EAAE,WAAW,EAAE,MAAM,SAAS,CAAC;AAC3C,OAAO,EAAE,SAAS,EAAE,MAAM,SAAS,CAAC;AACpC,YAAY,EACV,kBAAkB,EAClB,kBAAkB,EAClB,kBAAkB,EAClB,yBAAyB,GAC1B,MAAM,gBAAgB,CAAC;AACxB,OAAO,EAAE,iBAAiB,EAAE,MAAM,UAAU,CAAC;AAC7C,OAAO,EAAE,WAAW,EAAE,MAAM,WAAW,CAAC;AACxC,OAAO,EAAE,SAAS,EAAE,MAAM,SAAS,CAAC;AACpC,OAAO,EAAE,4BAA4B,EAAE,MAAM,UAAU,CAAC;AACxD,OAAO,EACL,kCAAkC,EAClC,kBAAkB,EAClB,eAAe,EACf,0BAA0B,EAC1B,mBAAmB,EACnB,yBAAyB,EACzB,yBAAyB,EACzB,gBAAgB,GACjB,MAAM,SAAS,CAAC;AACjB,OAAO,EAAE,kBAAkB,EAAE,MAAM,SAAS,CAAC"}

package/dist/agent/chat/exports.js

@@ -1,17 +1,21 @@
 "use strict";
 Object.defineProperty(exports, "__esModule", { value: true });
-exports.SUPPORTED_CHAT_MODELS = exports.createChatModel = exports.rejectPendingToolCalls = exports.processToolCallsAndUpdateChatState = 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.TriageAgent = exports.getAgentForMode = exports.ChatAgent = exports.allToolsDefinitions = exports.viewFileUsingGitHub = exports.getFileInfoFromGitHub = void 0;
+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.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, "ChatAgent", { enumerable: true, get: function () { return base_1.ChatAgent; } });
-Object.defineProperty(exports, "getAgentForMode", { enumerable: true, get: function () { return base_1.getAgentForMode; } });
-Object.defineProperty(exports, "TriageAgent", { enumerable: true, get: function () { return base_1.TriageAgent; } });
+Object.defineProperty(exports, "BaseAgent", { enumerable: true, get: function () { return base_1.BaseAgent; } });
+var index_1 = require("../index");
+Object.defineProperty(exports, "MODE_TO_AGENT_MAP", { enumerable: true, get: function () { return index_1.MODE_TO_AGENT_MAP; } });
+var triage_1 = require("../triage");
+Object.defineProperty(exports, "TriageAgent", { enumerable: true, get: function () { return triage_1.TriageAgent; } });
+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; } });
@@ -24,9 +28,3 @@ Object.defineProperty(exports, "LATEST_CHAT_STATE_VERSION", { enumerable: true,
 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 tool_calls_1 = require("./utils/tool-calls");
-Object.defineProperty(exports, "processToolCallsAndUpdateChatState", { enumerable: true, get: function () { return tool_calls_1.processToolCallsAndUpdateChatState; } });
-Object.defineProperty(exports, "rejectPendingToolCalls", { enumerable: true, get: function () { return tool_calls_1.rejectPendingToolCalls; } });
-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

@@ -1,11 +1,7 @@
-import { SupportedChatModels } from "@empiricalrun/shared-types";
-export declare function fetchEnvironmentVariables(): Promise<Record<string, string>>;
-export declare function runChatAgentForCLI({ useDiskForChatState, selectedModel, initialPromptContent, useTriage, resetChat, useFSCache, }: {
-    selectedModel: SupportedChatModels;
-    useDiskForChatState: boolean;
-    initialPromptContent: string | undefined;
-    useTriage: boolean;
-    useFSCache: boolean;
-    resetChat: boolean;
-}): Promise<void>;
+import type { ToolsForLLM } from "@empiricalrun/shared-types";
+import { BaseAgent } from "../base";
+export declare class ChatAgent extends BaseAgent {
+    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":"AAEA,OAAO,EAML,mBAAmB,EACpB,MAAM,4BAA4B,CAAC;AAwCpC,wBAAsB,yBAAyB,IAAI,OAAO,CACxD,MAAM,CAAC,MAAM,EAAE,MAAM,CAAC,CACvB,CAkCA;AAED,wBAAsB,kBAAkB,CAAC,EACvC,mBAAmB,EACnB,aAAa,EACb,oBAAoB,EACpB,SAAS,EACT,SAAS,EACT,UAAU,GACX,EAAE;IACD,aAAa,EAAE,mBAAmB,CAAC;IACnC,mBAAmB,EAAE,OAAO,CAAC;IAC7B,oBAAoB,EAAE,MAAM,GAAG,SAAS,CAAC;IACzC,SAAS,EAAE,OAAO,CAAC;IACnB,UAAU,EAAE,OAAO,CAAC;IACpB,SAAS,EAAE,OAAO,CAAC;CACpB,iBAyKA"}
+{"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

@@ -1,204 +1,125 @@
 "use strict";
 Object.defineProperty(exports, "__esModule", { value: true });
-exports.fetchEnvironmentVariables = fetchEnvironmentVariables;
-exports.runChatAgentForCLI = runChatAgentForCLI;
-const llm_1 = require("@empiricalrun/llm");
-const chat_1 = require("@empiricalrun/llm/chat");
-const picocolors_1 = require("picocolors");
-const client_1 = require("../../dashboard/client");
-const reader_1 = require("../../file-info/adapters/file-system/reader");
-const human_in_the_loop_1 = require("../../human-in-the-loop");
-const validation_1 = require("../../recorder/validation");
-const executor_1 = require("../../tools/executor");
-const git_1 = require("../../tools/executor/utils/git");
+exports.ChatAgent = void 0;
+const tools_1 = require("../../tools");
 const base_1 = require("../base");
-const filesystem_cache_1 = require("./filesystem-cache");
-const state_1 = require("./state");
-const utils_1 = require("./utils");
-function stopCriteria(userPrompt) {
-    return userPrompt?.toLowerCase() === "stop";
-}
-function concludeAgent(chatModel, useDiskForChatState, selectedModel, error) {
-    console.log(`\n${(0, picocolors_1.gray)("Usage summary -> " + (0, state_1.getUsageSummary)(chatModel))}`);
-    if (useDiskForChatState) {
-        (0, state_1.saveToDisk)(chatModel.messages, selectedModel, chatModel.askUserForInput, error);
+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 custom = [
+            ...tools_1.commonTools,
+            ...tools_1.testGenerationTools,
+            tools_1.analyseVideo,
+            ...(0, tools_1.textEditorToolsForModel)(this.selectedModel),
+        ];
+        return {
+            custom,
+            builtInTextEditor: (0, tools_1.hasBuiltInTextEditor)(this.selectedModel),
+        };
     }
+    async buildSystemPrompt(repoContext) {
+        if (!repoContext) {
+            throw new Error(`Chat agent needs repo context`);
+        }
+        const preamble = `
+You are a helpful assistant that can answer questions and help with tasks related to writing and maintaining Playwright tests.
+
+You are working on a test code repository that contains Playwright tests and other related files. Your working directory
+has been checked out on a git branch, and your actions that edit files will be automatically committed. If you make any
+file edits, you must create a pull request after your work is done, to get your work reviewed and merged by a human.
+
+# Your capabilities
+
+- Adding new Playwright tests or helper methods
+- Going through test reports and identifying app issues versus test issues
+- Modifying existing tests to adapt to changes in the application
+- Modifying repo configuration (e.g. in playwright.config.ts) and dependencies (e.g. in package.json)
+
+# Going through test reports
+
+- App issues: app issues caught by test failures, like UI issues, API endpoint issues, etc. These are issues that
+will be reported to an app developer to investigate and fix.
+- Test issues: Playwright tests can become outdated when app code changes. These are issues that need to be
+fixed with modifications to the test code, and it is your job to do that.
+
+# Tools
+
+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.
+
+For example,
+
+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.
+
+# Proactiveness
+
+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.
+
+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. 
+
+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
+
+You must follow these rules while adding new tests or modifying existing tests. There can be exceptions to these rules, but
+ONLY when explicitly asked for by the user.
+
+1. You can't delete some steps from the test to make it pass. The test needs to accomplish its objective (which is to validate a particular user scenario)
+2. Do not add any conditional logic or try catch blocks in a test. A good test deterministically tests a user scenario
+3. Trust Playwright's ability to auto-wait while taking actions on elements.
+   - Example 1: Do not add checks on locator.isVisible() before clicking on it: Playwright already waits for visibility on locator.click()
+   - Example 2: Do not add page.waitForLoadState after a page.goto: Playwright already waits for page "load" event in page.goto()
+4. Do not add waitForTimeout or waitForLoadState in a test. Playwright will automatically wait for the page to load.
+5. Try/catch blocks are a code smell for tests: you should not use them.
+6. Do not use then() or catch() syntax in a test. Use async/await only
+
+There are few exceptions to these rules. BEFORE applying any of the following exceptions, you MUST share your plan with the user and get their approval.
+
+## Exceptions for conditional logic
+
+There are few exceptions where you can add conditional logic to a test. If the application UI reveals some UI elements on certain conditions, we can add conditional logic.
+
+For example, a form view shows a "Save" button only when the form is dirty. In this case, we will have to check if the "Save" button is visible before clicking on it. To do this,
+follow this pattern:
+
+\`\`\`
+const saveButton = page.getByRole('button', { name: 'Save' });
+if (await saveButton.isVisible()) {
+  await saveButton.click();
 }
-async function fetchEnvironmentVariables() {
-    const projectApiKey = process.env.EMPIRICALRUN_API_KEY;
-    let apiClient;
-    let queryParams = {};
-    if (projectApiKey) {
-        apiClient = new client_1.DashboardAPIClient({
-            authType: "project-api-key",
-            projectApiKey: process.env.EMPIRICALRUN_API_KEY,
-        });
-    }
-    else {
-        apiClient = new client_1.DashboardAPIClient({
-            authType: "user-access-token",
-        });
-        const repoName = await (0, validation_1.validatePackageJson)(process.cwd());
-        queryParams = { project_repo_name: repoName };
-    }
-    const data = await apiClient.request("/api/environment-variables", { method: "GET", params: queryParams });
-    if (!data.data || !data.data.environment_variables) {
-        console.error("Failed to fetch environment variables:", data);
-        throw new Error("Failed to fetch environment variables");
-    }
-    const envVars = data.data.environment_variables.reduce((acc, envVar) => {
-        acc[envVar.name] = envVar.value;
-        return acc;
-    }, {});
-    return envVars;
+\`\`\`
+
+Note that locator.isVisible() DOES NOT wait for the element to be visible. If the element in question shows up after a delay, we have no option but to add a waitForTimeout.
+
+\`\`\`
+const saveButton = page.getByRole('button', { name: 'Save' });
+await page.waitForTimeout(100); // Wait for the element to be visible -- only if necessary.
+if (await saveButton.isVisible()) {
+  await saveButton.click();
 }
-async function runChatAgentForCLI({ useDiskForChatState, selectedModel, initialPromptContent, useTriage, resetChat, useFSCache, }) {
-    let chatState;
-    const enableStreaming = !useFSCache;
-    const cache = useFSCache ? new filesystem_cache_1.FilesystemLLMCache() : undefined;
-    if (resetChat) {
-        (0, state_1.clearChatState)();
-    }
-    if (useDiskForChatState) {
-        chatState = (0, state_1.loadChatState)();
-    }
-    // TODO: Store branch name in chat state so that we don't recreate it every time
-    const randomId = crypto.randomUUID().substring(0, 8);
-    const branchName = `branch-${randomId}`;
-    await (0, git_1.checkoutBranch)(branchName, process.cwd());
-    let messagesLoadedFromDisk = chatState?.messages || [];
-    let chatModel = (0, chat_1.createChatModel)(messagesLoadedFromDisk, selectedModel, undefined, cache);
-    chatModel.validateEnvVarsForAuth();
-    if (initialPromptContent && chatModel.messages.length === 0) {
-        chatModel.pushUserMessage(initialPromptContent, []);
-    }
-    else if (initialPromptContent && chatModel.messages.length > 0) {
-        console.warn(`Ignoring initial prompt because we have existing messages.`);
-    }
-    if (chatModel.askUserForInput) {
-        // Show last message to the user for context when we loaded from disk
-        const latest = chatModel.getHumanReadableLatestMessage();
-        if (latest) {
-            console.log(`${(0, picocolors_1.blue)(latest.role)}: ${latest.textMessage}`);
-        }
-    }
-    if (chatState && chatState.error) {
-        // Reset error state as we are attempting a retry
-        chatState.error = null;
-    }
-    const handleSigInt = () => {
-        concludeAgent(chatModel, useDiskForChatState, selectedModel, null);
-        process.exit(0);
-    };
-    process.once("SIGINT", handleSigInt);
-    process.once("SIGTERM", handleSigInt);
-    let userPrompt;
-    let reporterFunc = async (chatState, latest) => {
-        if (useDiskForChatState) {
-            (0, state_1.saveToDisk)(chatState.messages, selectedModel, chatState.askUserForInput, chatState.error);
-        }
-        if (latest) {
-            if (!enableStreaming) {
-                console.log(`${(0, picocolors_1.blue)(latest.role)}: ${latest.textMessage}`);
-            }
-            else {
-                process.stdout.write(`\n`);
-            }
-        }
-    };
-    const trace = (0, llm_1.createLangfuseTrace)({
-        name: "chat_agent",
-        input: initialPromptContent || "",
-        tags: [selectedModel, "chat_agent"],
-    });
-    if (trace) {
-        const traceUrl = trace.getTraceUrl();
-        console.log(`Starting ${selectedModel}: ${traceUrl}`);
-    }
-    let authType = "user-access-token";
-    if (process.env.EMPIRICALRUN_API_KEY) {
-        authType = "project-api-key";
-    }
-    const apiClient = new client_1.DashboardAPIClient({
-        authType,
-    });
-    const fileInfo = await (0, reader_1.getFileInfoFromFS)(process.cwd());
-    const agentParams = { featureFlags: [], repoInfo: fileInfo, selectedModel };
-    const agent = useTriage
-        ? new base_1.TriageAgent(agentParams)
-        : new base_1.ChatAgent(agentParams);
-    while (!stopCriteria(userPrompt)) {
-        if (chatModel.askUserForInput) {
-            try {
-                userPrompt = await human_in_the_loop_1.humanLoop.getFeedback({
-                    message: "User:",
-                });
-            }
-            catch (e) {
-                // https://github.com/SBoudrias/Inquirer.js/issues/1502#issuecomment-2275991680
-                if (e instanceof Error && e.name === "ExitPromptError") {
-                    concludeAgent(chatModel, useDiskForChatState, selectedModel, null);
-                    process.exit(0);
-                }
-                concludeAgent(chatModel, useDiskForChatState, selectedModel, {
-                    message: e.message,
-                    stack: e.stack || "Stack trace not available",
-                    timestamp: new Date().toISOString(),
-                });
-                throw e;
-            }
-            if (!stopCriteria(userPrompt)) {
-                const { text, attachments } = (0, utils_1.extractAttachments)(userPrompt);
-                chatModel.pushUserMessage(text, attachments);
-            }
-        }
-        else {
-            const toolExecutor = new executor_1.ToolExecutor({
-                chatSession: null,
-                branchName,
-                repoPath: process.cwd(),
-                apiClient,
-                trace,
-                featureFlags: [],
-                environmentOverrides: await fetchEnvironmentVariables(),
-            });
-            await agent.runLoop({
-                chatModel,
-                reporter: reporterFunc,
-                streamingMessageReporter: (() => {
-                    if (!enableStreaming) {
-                        return;
-                    }
-                    let hasStarted = false;
-                    let startedRole = undefined;
-                    return async (delta, snapshot, thinking) => {
-                        if (delta) {
-                            const role = thinking ? "Thinking" : "Assistant";
-                            if (!hasStarted) {
-                                process.stdout.write(`${(0, picocolors_1.blue)(role)}: `);
-                                hasStarted = true;
-                                startedRole = role;
-                            }
-                            else if (hasStarted && role !== startedRole) {
-                                // Changing from thinking -> text block
-                                process.stdout.write("\n");
-                                process.stdout.write(`${(0, picocolors_1.blue)(role)}: `);
-                                startedRole = role;
-                            }
-                            process.stdout.write(delta);
-                        }
-                    };
-                })(),
-                trace,
-                onPendingToolCall: async (toolCalls) => {
-                    const toolResults = await toolExecutor.execute(toolCalls);
-                    chatModel.pushToolResultsMessage(toolCalls, toolResults);
-                },
-            });
-        }
+\`\`\`
+
+`;
+        return `${preamble}
+
+${test_case_def_1.testCasesDefinitionPrompt}
+
+# Recipes
+${pw_utils_docs_1.playwrightUtilsDocs}
+
+# Repo context
+${repoContext}
+
+# Reference
+Today's date is ${new Date().toDateString()}
+`;
     }
-    trace?.update({ output: { messages: chatModel.messages } });
-    await llm_1.langfuseInstance?.flushAsync();
-    const usageSummary = (0, state_1.getUsageSummary)(chatModel);
-    console.log(`\n${(0, picocolors_1.gray)("Usage summary -> " + usageSummary)}`);
 }
+exports.ChatAgent = ChatAgent;

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;AAMjE,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 = "gpt-5-codex";
 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"}