@empiricalrun/test-gen 0.78.0 → 0.78.2

package/dist/agent/chat/state.js

@@ -4,11 +4,9 @@ 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.createChatStateForMessages = createChatStateForMessages;
 exports.chatStateFromModel = chatStateFromModel;
 exports.getUsageSummary = getUsageSummary;
 exports.loadChatState = loadChatState;
-exports.migrateChatState = migrateChatState;
 exports.saveToDisk = saveToDisk;
 exports.getLatestDownloadBuildUrl = getLatestDownloadBuildUrl;
 exports.fetchToolCallFromId = fetchToolCallFromId;
@@ -19,27 +17,20 @@ exports.CHAT_STATE_VERSIONS_MIGRATIONS_MAP = {
 };
 exports.LATEST_CHAT_STATE_VERSION = "0.1";
 exports.CHAT_STATE_PATH = path_1.default.join(process.cwd(), ".empiricalrun", "last-chat.json");
-function createChatStateForMessages({ messages, selectedModel, askUserForInput, error, }) {
-    // TODO: Add better types for messages
+function chatStateFromModel({ chatModel, error, }) {
     return {
         version: exports.LATEST_CHAT_STATE_VERSION,
-        model: selectedModel,
-        messages: messages,
-        askUserForInput: askUserForInput,
-        error: error,
-    };
-}
-function chatStateFromModel({ chatModel, error, }) {
-    return createChatStateForMessages({
+        model: chatModel.selectedModel,
         messages: chatModel.messages,
-        selectedModel: chatModel.selectedModel,
         askUserForInput: chatModel.askUserForInput,
-        error,
-    });
+        error: error,
+    };
 }
 function getUsageSummary(chatState) {
     const messages = chatState.messages;
-    const messageUsages = messages.map((msg) => msg.usage).filter(Boolean);
+    const messageUsages = messages
+        .map((msg) => "usage" in 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);
     const costs = messageUsages
@@ -71,37 +62,11 @@ function loadChatState({ resetChat, }) {
     }
     const raw = fs_1.default.readFileSync(exports.CHAT_STATE_PATH, "utf8");
     let state = JSON.parse(raw);
-    // Always migrate to the latest version after loading
-    const migratedState = migrateChatState(state);
     // Only save if migration actually changed the state
-    if (JSON.stringify(state) !== JSON.stringify(migratedState)) {
-        fs_1.default.writeFileSync(exports.CHAT_STATE_PATH, JSON.stringify(migratedState, null, 2));
+    if (JSON.stringify(state) !== JSON.stringify(state)) {
+        fs_1.default.writeFileSync(exports.CHAT_STATE_PATH, JSON.stringify(state, null, 2));
     }
-    return migratedState;
-}
-/**
- * Migrates a chat state object from an old version to the latest version.
- * Add migration logic for each version as needed.
- */
-function migrateChatState(oldState) {
-    if (!oldState || Object.keys(oldState).length === 0) {
-        return oldState;
-    }
-    if (!oldState.version) {
-        throw new Error("No version found in chat state");
-    }
-    if (!exports.CHAT_STATE_VERSIONS_MIGRATIONS_MAP[oldState.version]) {
-        throw new Error(`No migration function found for version: ${oldState.version}`);
-    }
-    if (oldState.version === exports.LATEST_CHAT_STATE_VERSION) {
-        return oldState;
-    }
-    const migrateFn = exports.CHAT_STATE_VERSIONS_MIGRATIONS_MAP[oldState.version];
-    const migrated = migrateFn(oldState);
-    return {
-        version: exports.LATEST_CHAT_STATE_VERSION,
-        ...migrated,
-    };
+    return state;
 }
 function saveToDisk(chatState) {
     const statePath = exports.CHAT_STATE_PATH;

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

@@ -1,12 +1,12 @@
 import { TraceClient } from "@empiricalrun/llm";
-import { IChatCanonicalModel, IChatModel } from "@empiricalrun/llm/chat";
-import { Attachment, ReporterFunction } from "@empiricalrun/shared-types";
+import { IChatModel } from "@empiricalrun/llm/chat";
+import { Attachment, ReporterFunction } from "@empiricalrun/shared-types/chat-agent";
 export declare const log: (...args: any[]) => void;
 export declare function getModelName(model: string): string;
-export declare function handleAgentError({ error, chatModel, reporter, trace, }: {
+export declare function handleAgentError<T>({ error, chatModel, reporter, trace, }: {
     error: unknown;
-    chatModel: IChatModel<any> | IChatCanonicalModel;
-    reporter: ReporterFunction;
+    chatModel: IChatModel<T>;
+    reporter: ReporterFunction<T>;
     trace?: TraceClient;
 }): Promise<void>;
 export declare function extractAttachments(userPrompt: string): {

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,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"}
+{"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,UAAU,EACX,MAAM,wBAAwB,CAAC;AAChC,OAAO,EACL,UAAU,EAEV,gBAAgB,EACjB,MAAM,uCAAuC,CAAC;AAK/C,eAAO,MAAM,GAAG,GAAI,GAAG,MAAM,GAAG,EAAE,SAEjC,CAAC;AAcF,wBAAgB,YAAY,CAAC,KAAK,EAAE,MAAM,GAAG,MAAM,CAMlD;AAED,wBAAsB,gBAAgB,CAAC,CAAC,EAAE,EACxC,KAAK,EACL,SAAS,EACT,QAAQ,EACR,KAAK,GACN,EAAE;IACD,KAAK,EAAE,OAAO,CAAC;IACf,SAAS,EAAE,UAAU,CAAC,CAAC,CAAC,CAAC;IACzB,QAAQ,EAAE,gBAAgB,CAAC,CAAC,CAAC,CAAC;IAC9B,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"}

package/dist/agent/cli.d.ts

@@ -1,6 +1,6 @@
-import { AgentModeEnum, SupportedChatModels } from "@empiricalrun/shared-types";
+import { AgentModeEnum, SupportedChatModels } from "@empiricalrun/shared-types/chat-agent";
 export declare function fetchEnvironmentVariables(): Promise<Record<string, string>>;
-export declare function runChatAgentForCLI({ useDiskForChatState, selectedModel, initialPromptContent, agentMode, resetChat, useFSCache, }: {
+export declare function runChatAgentForCLI<T>({ useDiskForChatState, selectedModel, initialPromptContent, agentMode, resetChat, useFSCache, }: {
     selectedModel: SupportedChatModels;
     useDiskForChatState: boolean;
     initialPromptContent: string | undefined;

package/dist/agent/cli.d.ts.map

@@ -1 +1 @@
-{"version":3,"file":"cli.d.ts","sourceRoot":"","sources":["../../src/agent/cli.ts"],"names":[],"mappings":"AACA,OAAO,EACL,aAAa,EAKb,mBAAmB,EACpB,MAAM,4BAA4B,CAAC;AA0BpC,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,aAAa,CAAC;IACzB,UAAU,EAAE,OAAO,CAAC;IACpB,SAAS,EAAE,OAAO,CAAC;CACpB,iBA+JA"}
+{"version":3,"file":"cli.d.ts","sourceRoot":"","sources":["../../src/agent/cli.ts"],"names":[],"mappings":"AAEA,OAAO,EACL,aAAa,EAIb,mBAAmB,EACpB,MAAM,uCAAuC,CAAC;AA6B/C,wBAAsB,yBAAyB,IAAI,OAAO,CACxD,MAAM,CAAC,MAAM,EAAE,MAAM,CAAC,CACvB,CAkCA;AAED,wBAAsB,kBAAkB,CAAC,CAAC,EAAE,EAC1C,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,aAAa,CAAC;IACzB,UAAU,EAAE,OAAO,CAAC;IACpB,SAAS,EAAE,OAAO,CAAC;CACpB,iBA+JA"}

package/dist/agent/cli.js

@@ -53,8 +53,6 @@ async function fetchEnvironmentVariables() {
 }
 async function runChatAgentForCLI({ useDiskForChatState, selectedModel, initialPromptContent, agentMode, resetChat, useFSCache, }) {
     const enableStreaming = !useFSCache;
-    // TODO: Implement cache support in BaseAgent
-    // const cache = useFSCache ? new FilesystemLLMCache() : undefined;
     // 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}`;
@@ -94,7 +92,7 @@ async function runChatAgentForCLI({ useDiskForChatState, selectedModel, initialP
     });
     if (trace) {
         const traceUrl = trace.getTraceUrl();
-        console.log(`Starting ${selectedModel}: ${traceUrl}`);
+        console.log(`Starting Model: ${selectedModel} \nTrace URL: ${traceUrl}`);
     }
     let authType = "user-access-token";
     if (process.env.EMPIRICALRUN_API_KEY) {
@@ -117,8 +115,11 @@ async function runChatAgentForCLI({ useDiskForChatState, selectedModel, initialP
         featureFlags: [],
         chatState,
         toolExecutor,
+        repoInfoBuilder: fileInfoBuilder,
     };
-    const agent = index_1.MODE_TO_AGENT_MAP[agentMode]({ ...agentParams });
+    const agent = agentMode === "triage"
+        ? new index_1.TriageAgent(agentParams)
+        : new index_1.ChatAgent(agentParams);
     // If we have an initial prompt and no existing chat state, push it to the agent
     if (initialPromptContent && !chatState) {
         const { text, attachments } = (0, utils_1.extractAttachments)(initialPromptContent);
@@ -185,8 +186,6 @@ async function runChatAgentForCLI({ useDiskForChatState, selectedModel, initialP
                         }
                     };
                 })(),
-                trace,
-                repoInfoBuilder: fileInfoBuilder,
                 onPendingToolCall: async (toolCalls) => {
                     const { toolResults, checkpoint } = await toolExecutor.execute(toolCalls);
                     agent.processToolResults(toolCalls, toolResults, checkpoint);

package/dist/agent/code-review/index.d.ts

@@ -1,12 +1,13 @@
-import type { ToolsForLLM } from "@empiricalrun/shared-types";
+import type { ToolsForLLM } from "@empiricalrun/shared-types/chat-agent";
+import type { FileInfo } from "@empiricalrun/shared-types/test-gen";
 import { BaseAgent } from "../base";
 import { type CodeReviewResultV0, type CodeReviewResultV1, type CodeReviewResultV2, CodeReviewSeverity, CodeReviewVerdict } from "./types";
 export type { CodeReviewResultV1, CodeReviewResultV0, CodeReviewResultV2 };
 export { CodeReviewVerdict, CodeReviewSeverity };
 export type CodeReviewVersionedResult = CodeReviewResultV0 | CodeReviewResultV1 | CodeReviewResultV2;
-export declare class CodeReviewAgent extends BaseAgent {
+export declare class CodeReviewAgent<T> extends BaseAgent<T> {
     protected getTools(): ToolsForLLM;
     getResult(): CodeReviewVersionedResult | undefined;
-    protected buildSystemPrompt(repoContext: string): Promise<string>;
+    protected buildSystemPrompt(repoInfoBuilder: () => Promise<FileInfo>): Promise<string>;
 }
 //# sourceMappingURL=index.d.ts.map

package/dist/agent/code-review/index.d.ts.map

@@ -1 +1 @@
-{"version":3,"file":"index.d.ts","sourceRoot":"","sources":["../../../src/agent/code-review/index.ts"],"names":[],"mappings":"AAAA,OAAO,KAAK,EAAmB,WAAW,EAAE,MAAM,4BAA4B,CAAC;AAG/E,OAAO,EAAE,SAAS,EAAE,MAAM,SAAS,CAAC;AAEpC,OAAO,EACL,KAAK,kBAAkB,EACvB,KAAK,kBAAkB,EACvB,KAAK,kBAAkB,EACvB,kBAAkB,EAClB,iBAAiB,EAClB,MAAM,SAAS,CAAC;AAEjB,YAAY,EAAE,kBAAkB,EAAE,kBAAkB,EAAE,kBAAkB,EAAE,CAAC;AAC3E,OAAO,EAAE,iBAAiB,EAAE,kBAAkB,EAAE,CAAC;AAEjD,MAAM,MAAM,yBAAyB,GACjC,kBAAkB,GAClB,kBAAkB,GAClB,kBAAkB,CAAC;AAEvB,qBAAa,eAAgB,SAAQ,SAAS;IAC5C,SAAS,CAAC,QAAQ,IAAI,WAAW;IAWjC,SAAS,IAAI,yBAAyB,GAAG,SAAS;cAmBlC,iBAAiB,CAAC,WAAW,EAAE,MAAM,GAAG,OAAO,CAAC,MAAM,CAAC;CA0HxE"}
+{"version":3,"file":"index.d.ts","sourceRoot":"","sources":["../../../src/agent/code-review/index.ts"],"names":[],"mappings":"AAAA,OAAO,KAAK,EAEV,WAAW,EACZ,MAAM,uCAAuC,CAAC;AAC/C,OAAO,KAAK,EAAE,QAAQ,EAAE,MAAM,qCAAqC,CAAC;AAGpE,OAAO,EAAE,SAAS,EAAE,MAAM,SAAS,CAAC;AAGpC,OAAO,EACL,KAAK,kBAAkB,EACvB,KAAK,kBAAkB,EACvB,KAAK,kBAAkB,EACvB,kBAAkB,EAClB,iBAAiB,EAClB,MAAM,SAAS,CAAC;AAEjB,YAAY,EAAE,kBAAkB,EAAE,kBAAkB,EAAE,kBAAkB,EAAE,CAAC;AAC3E,OAAO,EAAE,iBAAiB,EAAE,kBAAkB,EAAE,CAAC;AAEjD,MAAM,MAAM,yBAAyB,GACjC,kBAAkB,GAClB,kBAAkB,GAClB,kBAAkB,CAAC;AAEvB,qBAAa,eAAe,CAAC,CAAC,CAAE,SAAQ,SAAS,CAAC,CAAC,CAAC;IAClD,SAAS,CAAC,QAAQ,IAAI,WAAW;IAWjC,SAAS,IAAI,yBAAyB,GAAG,SAAS;cAmBlC,iBAAiB,CAC/B,eAAe,EAAE,MAAM,OAAO,CAAC,QAAQ,CAAC,GACvC,OAAO,CAAC,MAAM,CAAC;CAiInB"}

package/dist/agent/code-review/index.js

@@ -4,6 +4,7 @@ exports.CodeReviewAgent = exports.CodeReviewSeverity = exports.CodeReviewVerdict
 const tools_1 = require("../../tools");
 const fetch_session_diff_1 = require("../../tools/fetch-session-diff");
 const base_1 = require("../base");
+const repo_1 = require("../chat/prompt/repo");
 const parser_1 = require("./parser");
 const types_1 = require("./types");
 Object.defineProperty(exports, "CodeReviewSeverity", { enumerable: true, get: function () { return types_1.CodeReviewSeverity; } });
@@ -25,9 +26,7 @@ class CodeReviewAgent extends base_1.BaseAgent {
             ? messages[messages.length - 1]
             : undefined;
         const lastMessageTextPart = lastMessage
-            ? lastMessage.parts
-                .filter((p) => "text" in p)
-                .find((p) => "text" in p && !!p.text)
+            ? lastMessage.parts.find((p) => "text" in p && !!p.text && !("thinking" in p))
             : undefined;
         const textPart = lastMessageTextPart;
         const text = textPart?.text.trim();
@@ -36,7 +35,9 @@ class CodeReviewAgent extends base_1.BaseAgent {
         }
         return (0, parser_1.convertXmlToV2Format)(text);
     }
-    async buildSystemPrompt(repoContext) {
+    async buildSystemPrompt(repoInfoBuilder) {
+        const repoInfo = await repoInfoBuilder();
+        const repoContext = await (0, repo_1.getRepoContextPrompt)(repoInfo);
         return `
 You are an expert code reviewer that specializes in reviewing Playwright test code. You are
 provided with tools to fetch diff and pull-request metadata for a code review, where a test has been added, test modified,
@@ -122,6 +123,8 @@ Return these as XML tags with markdown inside them
 - Use locators instead of selectors: waitForSelector, $, $$ are bad - use locators instead (e.g. locator.waitFor)
 - If the test relies on some Playwright APIs that do not auto-wait (e.g. isVisible(), count()), we need to ensure 
   they are used AFTER some action that ensures the page has loaded. If nothing, at least it should have a waitForTimeout
+- For checkbox elements, use \`setChecked(true or false)\` instead of clicks. Click action on checkbox toggles the state and requires
+  a conditional, whereas \`setChecked\` allows specifying the end expected state, which is easier to maintain.
 
 ## Deprecated patterns
 - test.describe.serial(...) is not prefered: use test.describe.configure({ mode: "serial" }) if the tests need to be serial
@@ -156,6 +159,9 @@ Return these as XML tags with markdown inside them
 
 # Repo context
 ${repoContext}
+
+# Reference
+Today's date is ${new Date().toDateString()}
 `;
     }
 }

package/dist/agent/code-review/types.d.ts

@@ -1,36 +1,3 @@
-export declare enum CodeReviewVerdict {
-    Approved = "Approved",
-    Rejected = "Rejected"
-}
-export declare enum CodeReviewSeverity {
-    MergeBlocking = "merge-blocking",
-    Warning = "warning"
-}
-export type CodeReviewLineComment = {
-    file: string;
-    line_start: number;
-    line_end: number;
-    severity: CodeReviewSeverity;
-    message: string;
-};
-export type CodeReviewResultV2 = {
-    version: "2.0";
-    last_commit: string;
-    describe_code_change: string;
-    line_comments: CodeReviewLineComment[];
-    verdict: CodeReviewVerdict;
-};
-export type CodeReviewResultV1 = {
-    version: "1.0";
-    last_commit: string | null;
-    code_review_comments: string | null;
-    describe_code_change: string | null;
-    merge_blocking_issues: string | null;
-    best_practices_and_warnings: string | null;
-    verdict: "Approved" | "Rejected" | null;
-};
-export type CodeReviewResultV0 = {
-    version: "0.1";
-    result: string;
-};
+export type { CodeReviewLineComment, CodeReviewResultV0, CodeReviewResultV1, CodeReviewResultV2, } from "@empiricalrun/shared-types/code-review";
+export { CodeReviewSeverity, CodeReviewVerdict, } from "@empiricalrun/shared-types/code-review";
 //# sourceMappingURL=types.d.ts.map

package/dist/agent/code-review/types.d.ts.map

@@ -1 +1 @@
-{"version":3,"file":"types.d.ts","sourceRoot":"","sources":["../../../src/agent/code-review/types.ts"],"names":[],"mappings":"AAAA,oBAAY,iBAAiB;IAC3B,QAAQ,aAAa;IACrB,QAAQ,aAAa;CACtB;AAED,oBAAY,kBAAkB;IAC5B,aAAa,mBAAmB;IAChC,OAAO,YAAY;CACpB;AAED,MAAM,MAAM,qBAAqB,GAAG;IAClC,IAAI,EAAE,MAAM,CAAC;IACb,UAAU,EAAE,MAAM,CAAC;IACnB,QAAQ,EAAE,MAAM,CAAC;IACjB,QAAQ,EAAE,kBAAkB,CAAC;IAC7B,OAAO,EAAE,MAAM,CAAC;CACjB,CAAC;AAEF,MAAM,MAAM,kBAAkB,GAAG;IAC/B,OAAO,EAAE,KAAK,CAAC;IACf,WAAW,EAAE,MAAM,CAAC;IACpB,oBAAoB,EAAE,MAAM,CAAC;IAC7B,aAAa,EAAE,qBAAqB,EAAE,CAAC;IACvC,OAAO,EAAE,iBAAiB,CAAC;CAC5B,CAAC;AAEF,MAAM,MAAM,kBAAkB,GAAG;IAC/B,OAAO,EAAE,KAAK,CAAC;IACf,WAAW,EAAE,MAAM,GAAG,IAAI,CAAC;IAC3B,oBAAoB,EAAE,MAAM,GAAG,IAAI,CAAC;IACpC,oBAAoB,EAAE,MAAM,GAAG,IAAI,CAAC;IACpC,qBAAqB,EAAE,MAAM,GAAG,IAAI,CAAC;IACrC,2BAA2B,EAAE,MAAM,GAAG,IAAI,CAAC;IAC3C,OAAO,EAAE,UAAU,GAAG,UAAU,GAAG,IAAI,CAAC;CACzC,CAAC;AAEF,MAAM,MAAM,kBAAkB,GAAG;IAC/B,OAAO,EAAE,KAAK,CAAC;IAEf,MAAM,EAAE,MAAM,CAAC;CAChB,CAAC"}
+{"version":3,"file":"types.d.ts","sourceRoot":"","sources":["../../../src/agent/code-review/types.ts"],"names":[],"mappings":"AAAA,YAAY,EACV,qBAAqB,EACrB,kBAAkB,EAClB,kBAAkB,EAClB,kBAAkB,GACnB,MAAM,wCAAwC,CAAC;AAEhD,OAAO,EACL,kBAAkB,EAClB,iBAAiB,GAClB,MAAM,wCAAwC,CAAC"}

package/dist/agent/code-review/types.js

@@ -1,13 +1,6 @@
 "use strict";
 Object.defineProperty(exports, "__esModule", { value: true });
-exports.CodeReviewSeverity = exports.CodeReviewVerdict = void 0;
-var CodeReviewVerdict;
-(function (CodeReviewVerdict) {
-    CodeReviewVerdict["Approved"] = "Approved";
-    CodeReviewVerdict["Rejected"] = "Rejected";
-})(CodeReviewVerdict || (exports.CodeReviewVerdict = CodeReviewVerdict = {}));
-var CodeReviewSeverity;
-(function (CodeReviewSeverity) {
-    CodeReviewSeverity["MergeBlocking"] = "merge-blocking";
-    CodeReviewSeverity["Warning"] = "warning";
-})(CodeReviewSeverity || (exports.CodeReviewSeverity = CodeReviewSeverity = {}));
+exports.CodeReviewVerdict = exports.CodeReviewSeverity = void 0;
+var code_review_1 = require("@empiricalrun/shared-types/code-review");
+Object.defineProperty(exports, "CodeReviewSeverity", { enumerable: true, get: function () { return code_review_1.CodeReviewSeverity; } });
+Object.defineProperty(exports, "CodeReviewVerdict", { enumerable: true, get: function () { return code_review_1.CodeReviewVerdict; } });

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

@@ -1,5 +1,5 @@
 import { TraceClient } from "@empiricalrun/llm";
-import { ArtifactInputData, Usage } from "@empiricalrun/shared-types";
+import { ArtifactInputData, Usage } from "@empiricalrun/shared-types/chat-agent";
 import { Page } from "playwright";
 export type BrowserAgentResult = Array<{
     type: "text";

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

@@ -1 +1 @@
-{"version":3,"file":"index.d.ts","sourceRoot":"","sources":["../../../src/agent/cua/index.ts"],"names":[],"mappings":"AAAA,OAAO,EAAiB,WAAW,EAAE,MAAM,mBAAmB,CAAC;AAC/D,OAAO,EAAE,iBAAiB,EAAE,KAAK,EAAE,MAAM,4BAA4B,CAAC;AAStE,OAAO,EAAE,IAAI,EAAE,MAAM,YAAY,CAAC;AAoDlC,MAAM,MAAM,kBAAkB,GAAG,KAAK,CAClC;IACE,IAAI,EAAE,MAAM,CAAC;IACb,IAAI,EAAE,MAAM,CAAC;CACd,GACD;IACE,IAAI,EAAE,YAAY,CAAC;IACnB,UAAU,EAAE,iBAAiB,CAAC;CAC/B,CACJ,CAAC;AAEF,wBAAsB,+BAA+B,CAAC,EACpD,IAAI,EACJ,IAAI,EACJ,KAAK,EACL,8BAAsC,GACvC,EAAE;IACD,IAAI,EAAE,IAAI,CAAC;IACX,IAAI,EAAE,MAAM,CAAC;IACb,KAAK,CAAC,EAAE,WAAW,CAAC;IACpB,8BAA8B,CAAC,EAAE,OAAO,CAAC;CAC1C,GAAG,OAAO,CAAC;IACV,IAAI,EAAE,MAAM,CAAC;IACb,WAAW,EAAE,MAAM,EAAE,CAAC;IACtB,MAAM,EAAE,kBAAkB,CAAC;IAC3B,KAAK,EAAE,KAAK,CAAC;CACd,CAAC,CAuND"}
+{"version":3,"file":"index.d.ts","sourceRoot":"","sources":["../../../src/agent/cua/index.ts"],"names":[],"mappings":"AAAA,OAAO,EAAiB,WAAW,EAAE,MAAM,mBAAmB,CAAC;AAC/D,OAAO,EACL,iBAAiB,EACjB,KAAK,EACN,MAAM,uCAAuC,CAAC;AAS/C,OAAO,EAAE,IAAI,EAAE,MAAM,YAAY,CAAC;AAmElC,MAAM,MAAM,kBAAkB,GAAG,KAAK,CAClC;IACE,IAAI,EAAE,MAAM,CAAC;IACb,IAAI,EAAE,MAAM,CAAC;CACd,GACD;IACE,IAAI,EAAE,YAAY,CAAC;IACnB,UAAU,EAAE,iBAAiB,CAAC;CAC/B,CACJ,CAAC;AAEF,wBAAsB,+BAA+B,CAAC,EACpD,IAAI,EACJ,IAAI,EACJ,KAAK,EACL,8BAAsC,GACvC,EAAE;IACD,IAAI,EAAE,IAAI,CAAC;IACX,IAAI,EAAE,MAAM,CAAC;IACb,KAAK,CAAC,EAAE,WAAW,CAAC;IACpB,8BAA8B,CAAC,EAAE,OAAO,CAAC;CAC1C,GAAG,OAAO,CAAC;IACV,IAAI,EAAE,MAAM,CAAC;IACb,WAAW,EAAE,MAAM,EAAE,CAAC;IACtB,MAAM,EAAE,kBAAkB,CAAC;IAC3B,KAAK,EAAE,KAAK,CAAC;CACd,CAAC,CAuND"}

package/dist/agent/cua/index.js

@@ -6,20 +6,23 @@ Object.defineProperty(exports, "__esModule", { value: true });
 exports.createTestUsingComputerUseAgent = createTestUsingComputerUseAgent;
 const llm_1 = require("@empiricalrun/llm");
 const openai_1 = __importDefault(require("openai"));
+const logger_1 = require("../../logger");
 const computer_1 = require("./computer");
 const model_1 = require("./model");
 const element_from_point_1 = require("./pw-codegen/element-from-point");
-const pw_pause_1 = require("./pw-codegen/pw-pause");
+const factory_1 = require("./pw-codegen/factory");
 const MAX_ITERATIONS = 10;
 async function getCodegenInstance(prefersElementFromPoint) {
-    const repoDir = process.cwd();
-    const canUsePwPause = !prefersElementFromPoint && (await (0, pw_pause_1.canUsePauseCodegen)(repoDir));
-    if (canUsePwPause) {
-        console.log("[getCodegen] using PlaywrightPauseCodegen");
-        return new pw_pause_1.PlaywrightPauseCodegen();
+    if (prefersElementFromPoint) {
+        logger_1.logger.debug("[getCodegen] using ElementFromPointCodegen (preferred)");
+        return new element_from_point_1.ElementFromPointCodegen();
     }
-    console.log("[getCodegen] using ElementFromPointCodegen");
-    return new element_from_point_1.ElementFromPointCodegen();
+    const repoDir = process.cwd();
+    const { codegen, approach, version } = await (0, factory_1.createPlaywrightCodegen)(repoDir);
+    logger_1.logger.debug(`[getCodegen] using ${approach} codegen for Playwright ${version.full}`);
+    // Prepare if needed (patch-based approach requires file modifications)
+    await (0, factory_1.prepareCodegenIfNeeded)(repoDir, approach);
+    return codegen;
 }
 function artifact(screenshot, name) {
     return {
@@ -28,21 +31,29 @@ function artifact(screenshot, name) {
         data: Buffer.from(screenshot, "base64"),
     };
 }
-function stateOfTheBrowser(page) {
+async function tabsSummary(pages) {
+    let lines = [];
+    for (const p of pages) {
+        lines.push(`  - ${p.url()} - ${await p.title()}`);
+    }
+    return lines.join("\n");
+}
+async function stateOfTheBrowser(page) {
     const browserContext = page.context();
     const pages = browserContext.pages();
+    const allPagesSummary = await tabsSummary(pages);
     return `
 ## Browser window
 
 ### Current page (what you are working on)
 Current page URL: ${page.url()}
-Current page title: ${page.title()}
+Current page title: ${await page.title()}
 
 ### All pages
 Number of open pages: ${pages.length}
 
 URLs and titles:
-${pages.map((p) => `  - ${p.url()} - ${p.title()}`).join("\n")}`;
+${allPagesSummary}`;
 }
 async function createTestUsingComputerUseAgent({ page, task, trace, prefersElementFromPointCodegen = false, }) {
     const codegen = await getCodegenInstance(prefersElementFromPointCodegen);
@@ -69,7 +80,7 @@ async function createTestUsingComputerUseAgent({ page, task, trace, prefersEleme
                 content: [
                     {
                         type: "input_text",
-                        text: `Task to execute: ${task}\n\n${stateOfTheBrowser(page)}`,
+                        text: `Task to execute: ${task}\n\n${await stateOfTheBrowser(page)}`,
                     },
                     {
                         type: "input_image",
@@ -193,7 +204,7 @@ async function createTestUsingComputerUseAgent({ page, task, trace, prefersEleme
                     content: [
                         {
                             type: "input_text",
-                            text: `Action executed: ${executedActionSummary || "None"}\n\n${stateOfTheBrowser(page)}`,
+                            text: `Action executed: ${executedActionSummary || "None"}\n\n${await stateOfTheBrowser(page)}`,
                         },
                     ],
                 },

package/dist/agent/cua/pw-codegen/factory.d.ts

@@ -0,0 +1,15 @@
+/**
+ * Factory for creating the appropriate Playwright codegen implementation
+ * based on the installed Playwright version.
+ */
+import type { BasePlaywrightCodegen } from "./types";
+import { type CodegenApproach, type PlaywrightVersion } from "./version";
+export type CodegenFactoryResult = {
+    codegen: BasePlaywrightCodegen;
+    approach: CodegenApproach;
+    version: PlaywrightVersion;
+};
+export declare function createPlaywrightCodegen(repoDir: string): Promise<CodegenFactoryResult>;
+export declare function prepareCodegenIfNeeded(repoDir: string, approach: CodegenApproach): Promise<void>;
+export { type CodegenApproach, type PlaywrightVersion };
+//# sourceMappingURL=factory.d.ts.map

package/dist/agent/cua/pw-codegen/factory.d.ts.map

@@ -0,0 +1 @@
+{"version":3,"file":"factory.d.ts","sourceRoot":"","sources":["../../../../src/agent/cua/pw-codegen/factory.ts"],"names":[],"mappings":"AAAA;;;GAGG;AAaH,OAAO,KAAK,EAAE,qBAAqB,EAAE,MAAM,SAAS,CAAC;AACrD,OAAO,EACL,KAAK,eAAe,EAGpB,KAAK,iBAAiB,EACvB,MAAM,WAAW,CAAC;AAEnB,MAAM,MAAM,oBAAoB,GAAG;IACjC,OAAO,EAAE,qBAAqB,CAAC;IAC/B,QAAQ,EAAE,eAAe,CAAC;IAC1B,OAAO,EAAE,iBAAiB,CAAC;CAC5B,CAAC;AAEF,wBAAsB,uBAAuB,CAC3C,OAAO,EAAE,MAAM,GACd,OAAO,CAAC,oBAAoB,CAAC,CAoD/B;AAED,wBAAsB,sBAAsB,CAC1C,OAAO,EAAE,MAAM,EACf,QAAQ,EAAE,eAAe,GACxB,OAAO,CAAC,IAAI,CAAC,CAQf;AAED,OAAO,EAAE,KAAK,eAAe,EAAE,KAAK,iBAAiB,EAAE,CAAC"}

package/dist/agent/cua/pw-codegen/factory.js

@@ -0,0 +1,59 @@
+"use strict";
+Object.defineProperty(exports, "__esModule", { value: true });
+exports.createPlaywrightCodegen = createPlaywrightCodegen;
+exports.prepareCodegenIfNeeded = prepareCodegenIfNeeded;
+/**
+ * Factory for creating the appropriate Playwright codegen implementation
+ * based on the installed Playwright version.
+ */
+const logger_1 = require("../../../logger");
+const pw_event_sink_1 = require("./pw-event-sink");
+const pw_pause_1 = require("./pw-pause");
+const version_1 = require("./version");
+async function createPlaywrightCodegen(repoDir) {
+    const version = await (0, version_1.detectPlaywrightVersion)(repoDir);
+    if (!version) {
+        throw new Error("Could not detect Playwright version. Make sure @playwright/test is installed.");
+    }
+    const approach = (0, version_1.getCodegenApproach)(version);
+    logger_1.logger.info(`Detected Playwright ${version.full}, using "${approach}" codegen approach`);
+    switch (approach) {
+        case "event-sink": {
+            const canUse = await (0, pw_event_sink_1.canUseEventSinkCodegen)();
+            if (!canUse) {
+                throw new Error("Cannot use event-sink codegen. Check that RUN_PLAYWRIGHT_HEADED=true.");
+            }
+            return {
+                codegen: new pw_event_sink_1.PlaywrightEventSinkCodegen(),
+                approach,
+                version,
+            };
+        }
+        case "patch": {
+            const canUse = await (0, pw_pause_1.canUsePauseCodegen)(repoDir);
+            if (!canUse) {
+                throw new Error("Cannot use patch-based codegen. Check requirements: " +
+                    "RUN_PLAYWRIGHT_HEADED=true, PW_CODEGEN_NO_INSPECTOR=true, and patched recorder.");
+            }
+            return {
+                codegen: new pw_pause_1.PlaywrightPauseCodegen(),
+                approach,
+                version,
+            };
+        }
+        case "unsupported":
+        default:
+            throw new Error(`Unsupported Playwright version: ${version.full}. ` +
+                "Supported versions: 1.47.x, 1.53.x (patch), 1.57.0+ (event-sink).");
+    }
+}
+async function prepareCodegenIfNeeded(repoDir, approach) {
+    if (approach === "patch") {
+        logger_1.logger.info("Preparing Playwright for patch-based codegen...");
+        await (0, pw_pause_1.preparePatchCodegen)(repoDir);
+    }
+    else if (approach === "event-sink") {
+        logger_1.logger.info("Preparing Playwright for event-sink codegen...");
+        await (0, pw_event_sink_1.prepareEventSinkCodegen)(repoDir);
+    }
+}

package/dist/agent/cua/pw-codegen/index.d.ts

@@ -0,0 +1,7 @@
+/**
+ * Playwright Codegen - exports for external use
+ */
+export { type CodegenApproach, type CodegenFactoryResult, createPlaywrightCodegen, type PlaywrightVersion, prepareCodegenIfNeeded, } from "./factory";
+export type { BasePlaywrightCodegen } from "./types";
+export { detectPlaywrightVersion, getCodegenApproach, getPlaywrightVersionFromNpmList, parseVersion, } from "./version";
+//# sourceMappingURL=index.d.ts.map

package/dist/agent/cua/pw-codegen/index.d.ts.map

@@ -0,0 +1 @@
+{"version":3,"file":"index.d.ts","sourceRoot":"","sources":["../../../../src/agent/cua/pw-codegen/index.ts"],"names":[],"mappings":"AAAA;;GAEG;AAEH,OAAO,EACL,KAAK,eAAe,EACpB,KAAK,oBAAoB,EACzB,uBAAuB,EACvB,KAAK,iBAAiB,EACtB,sBAAsB,GACvB,MAAM,WAAW,CAAC;AACnB,YAAY,EAAE,qBAAqB,EAAE,MAAM,SAAS,CAAC;AACrD,OAAO,EACL,uBAAuB,EACvB,kBAAkB,EAClB,+BAA+B,EAC/B,YAAY,GACb,MAAM,WAAW,CAAC"}

package/dist/agent/cua/pw-codegen/index.js

@@ -0,0 +1,14 @@
+"use strict";
+Object.defineProperty(exports, "__esModule", { value: true });
+exports.parseVersion = exports.getPlaywrightVersionFromNpmList = exports.getCodegenApproach = exports.detectPlaywrightVersion = exports.prepareCodegenIfNeeded = exports.createPlaywrightCodegen = void 0;
+/**
+ * Playwright Codegen - exports for external use
+ */
+var factory_1 = require("./factory");
+Object.defineProperty(exports, "createPlaywrightCodegen", { enumerable: true, get: function () { return factory_1.createPlaywrightCodegen; } });
+Object.defineProperty(exports, "prepareCodegenIfNeeded", { enumerable: true, get: function () { return factory_1.prepareCodegenIfNeeded; } });
+var version_1 = require("./version");
+Object.defineProperty(exports, "detectPlaywrightVersion", { enumerable: true, get: function () { return version_1.detectPlaywrightVersion; } });
+Object.defineProperty(exports, "getCodegenApproach", { enumerable: true, get: function () { return version_1.getCodegenApproach; } });
+Object.defineProperty(exports, "getPlaywrightVersionFromNpmList", { enumerable: true, get: function () { return version_1.getPlaywrightVersionFromNpmList; } });
+Object.defineProperty(exports, "parseVersion", { enumerable: true, get: function () { return version_1.parseVersion; } });

package/dist/agent/cua/pw-codegen/pw-event-sink/index.d.ts

@@ -0,0 +1,43 @@
+/**
+ * Playwright Codegen using RecorderEventSink API (Playwright 1.57.0+)
+ *
+ * This implementation uses Playwright's internal RecorderEventSink interface
+ * to capture generated code without patching any files.
+ *
+ * How it works:
+ * 1. Call _enableRecorder() with mode: 'recording' and recorderMode: 'api'
+ * 2. Playwright's Recorder captures all user actions
+ * 3. Events flow to our eventSink callbacks with generated code
+ * 4. No inspector UI opens (recorderMode: 'api' skips UI)
+ *
+ * Requirements:
+ * - Playwright 1.57.0 or higher
+ * - RUN_PLAYWRIGHT_HEADED=true (recorder needs headed mode)
+ * - A small patch to recorderApp.js to allow recorderMode: 'api' with PW_CODEGEN_NO_INSPECTOR
+ */
+import type { BrowserContext, Page } from "playwright";
+import type { BasePlaywrightCodegen } from "../types";
+/**
+ * Prepares Playwright 1.57+ for event-sink codegen by patching recorderApp.js
+ * to allow recorderMode: 'api' even when PW_CODEGEN_NO_INSPECTOR is set.
+ */
+export declare function prepareEventSinkCodegen(repoDir: string): Promise<void>;
+export declare function canUseEventSinkCodegen(): Promise<boolean>;
+export declare function hasEventSinkSupport(context: BrowserContext): boolean;
+export declare class PlaywrightEventSinkCodegen implements BasePlaywrightCodegen {
+    private page;
+    private codeBuffer;
+    private lastActionPromise;
+    private resolveLastAction;
+    private lastReturnedEntry;
+    private isGettingLastAction;
+    private eventSink;
+    private notifyActionReceived;
+    private waitForNextAction;
+    initialize(page: Page): Promise<void>;
+    recordAction(): Promise<void>;
+    getCodeForLastAction(): Promise<string>;
+    private waitForPendingSignals;
+    private wrapCodeWithSignals;
+}
+//# sourceMappingURL=index.d.ts.map

package/dist/agent/cua/pw-codegen/pw-event-sink/index.d.ts.map

@@ -0,0 +1 @@
+{"version":3,"file":"index.d.ts","sourceRoot":"","sources":["../../../../../src/agent/cua/pw-codegen/pw-event-sink/index.ts"],"names":[],"mappings":"AAAA;;;;;;;;;;;;;;;;GAgBG;AAGH,OAAO,KAAK,EAAE,cAAc,EAAE,IAAI,EAAE,MAAM,YAAY,CAAC;AASvD,OAAO,KAAK,EAAE,qBAAqB,EAAE,MAAM,UAAU,CAAC;AAItD;;;GAGG;AACH,wBAAsB,uBAAuB,CAAC,OAAO,EAAE,MAAM,GAAG,OAAO,CAAC,IAAI,CAAC,CAoC5E;AA8CD,wBAAsB,sBAAsB,IAAI,OAAO,CAAC,OAAO,CAAC,CAa/D;AAED,wBAAgB,mBAAmB,CAAC,OAAO,EAAE,cAAc,GAAG,OAAO,CAEpE;AAQD,qBAAa,0BAA2B,YAAW,qBAAqB;IACtE,OAAO,CAAC,IAAI,CAAmB;IAC/B,OAAO,CAAC,UAAU,CAAsB;IACxC,OAAO,CAAC,iBAAiB,CAA4B;IACrD,OAAO,CAAC,iBAAiB,CAA2B;IACpD,OAAO,CAAC,iBAAiB,CAA2B;IACpD,OAAO,CAAC,mBAAmB,CAAS;IAEpC,OAAO,CAAC,SAAS,CAkDf;IAEF,OAAO,CAAC,oBAAoB;IAQ5B,OAAO,CAAC,iBAAiB;IASnB,UAAU,CAAC,IAAI,EAAE,IAAI,GAAG,OAAO,CAAC,IAAI,CAAC;IA6BrC,YAAY,IAAI,OAAO,CAAC,IAAI,CAAC;IAK7B,oBAAoB,IAAI,OAAO,CAAC,MAAM,CAAC;YAuD/B,qBAAqB;IAOnC,OAAO,CAAC,mBAAmB;CA0D5B"}