@botpress/evals 1.19.0-beta.9 → 2.0.0-alpha.1

package/dist/client.d.ts

@@ -5,31 +5,36 @@
  * This is a send-only client. Observation (bot responses, tool calls, state)
  * is handled by the SSECollector via trace spans — not via WebSocket listeners.
  */
+import { Client as BpClient } from '@botpress/client';
 import type { ChatClient, EvalLogger } from './types';
-export declare function setDefaultChatClient(chatClient: ChatClient | null): void;
 /**
  * A send-only chat session that maintains a single client connection across turns.
  * Messages and events are fire-and-forget; the SSECollector observes results via traces.
  */
 export declare class ChatSession {
-    private webhookId;
-    private chatClient?;
+    private bpClient;
+    private botId;
+    private _chatWebhookId?;
+    private _chatBaseUrl?;
+    private _chatClient?;
     private client;
     private conversationId;
-    constructor(webhookId: string, chatClient?: ChatClient);
+    constructor(bpClient: BpClient, botId: string, _chatWebhookId?: string, _chatBaseUrl?: string, _chatClient?: ChatClient);
     connect(): Promise<void>;
     /** Invariant guard — calling any session method before connect() is a runner bug. */
     private assertConnected;
     get userId(): string;
     ensureConversation(): Promise<string>;
     /**
-     * Send a user message. Fire-and-forget — observation is via SSECollector.
+     * Open a fresh conversation under the same user (the user is fixed at
+     * connect()) and make it active. Returns the new conversation id.
      */
-    sendMessage(message: string): Promise<void>;
+    newConversation(): Promise<string>;
     /**
-     * Send an event. Fire-and-forget — observation is via SSECollector.
+     * Send a user message. Fire-and-forget — observation is via SSECollector.
      */
-    sendEvent(type: string, payload: Record<string, unknown>): Promise<void>;
+    sendMessage(message: string): Promise<void>;
+    sendEvent(payload: Record<string, unknown>): Promise<void>;
 }
 /**
  * Discover the chat integration's webhookId from a bot.
@@ -37,7 +42,14 @@ export declare class ChatSession {
  * Evals talk to the bot through the `chat` integration (via @botpress/chat).
  * If the bot doesn't have that integration installed, evals can't run.
  */
-export declare function discoverWebhookId(botId: string, token: string, apiUrl: string): Promise<string>;
+export declare function discoverWebhookId(client: BpClient, botId: string): Promise<string>;
+/**
+ * Verify the dev server's trace stream can be read before eval turns start.
+ *
+ * This only checks that `/api/traces/stream` is reachable and emits its initial
+ * snapshot. It does not prove the private runtime span-ingest path is healthy.
+ */
+export declare function assertTraceStreamReadable(devServerUrl: string, headers?: Record<string, string>, timeoutMs?: number): Promise<void>;
 /**
  * Verify the bot has at least one Conversation bound to `chat.channel`.
  *

package/dist/client.d.ts.map

@@ -1 +1 @@
-{"version":3,"file":"client.d.ts","sourceRoot":"","sources":["../src/client.ts"],"names":[],"mappings":"AAAA;;;;;;GAMG;AAIH,OAAO,KAAK,EAAE,UAAU,EAAwB,UAAU,EAAE,MAAM,SAAS,CAAA;AAM3E,wBAAgB,oBAAoB,CAAC,UAAU,EAAE,UAAU,GAAG,IAAI,GAAG,IAAI,CAExE;AAiBD;;;GAGG;AACH,qBAAa,WAAW;IAKpB,OAAO,CAAC,SAAS;IACjB,OAAO,CAAC,UAAU,CAAC;IALrB,OAAO,CAAC,MAAM,CAAoC;IAClD,OAAO,CAAC,cAAc,CAAsB;gBAGlC,SAAS,EAAE,MAAM,EACjB,UAAU,CAAC,EAAE,UAAU;IAG3B,OAAO;IAKb,qFAAqF;IACrF,OAAO,CAAC,eAAe;IAUvB,IAAI,MAAM,IAAI,MAAM,CAEnB;IAEK,kBAAkB,IAAI,OAAO,CAAC,MAAM,CAAC;IAY3C;;OAEG;IACG,WAAW,CAAC,OAAO,EAAE,MAAM,GAAG,OAAO,CAAC,IAAI,CAAC;IAMjD;;OAEG;IACG,SAAS,CAAC,IAAI,EAAE,MAAM,EAAE,OAAO,EAAE,MAAM,CAAC,MAAM,EAAE,OAAO,CAAC,GAAG,OAAO,CAAC,IAAI,CAAC;CAK/E;AAED;;;;;GAKG;AACH,wBAAsB,iBAAiB,CAAC,KAAK,EAAE,MAAM,EAAE,KAAK,EAAE,MAAM,EAAE,MAAM,EAAE,MAAM,GAAG,OAAO,CAAC,MAAM,CAAC,CAuBrG;AAED;;;;;;;;;GASG;AACH,wBAAsB,sBAAsB,CAC1C,YAAY,EAAE,MAAM,EACpB,OAAO,GAAE,MAAM,CAAC,MAAM,EAAE,MAAM,CAAM,EACpC,MAAM,GAAE,UAA0B,GACjC,OAAO,CAAC,IAAI,CAAC,CA8Cf"}
+{"version":3,"file":"client.d.ts","sourceRoot":"","sources":["../src/client.ts"],"names":[],"mappings":"AAAA;;;;;;GAMG;AAEH,OAAO,EAAE,MAAM,IAAI,QAAQ,EAAE,MAAM,kBAAkB,CAAA;AAErD,OAAO,KAAK,EAAE,UAAU,EAAE,UAAU,EAAE,MAAM,SAAS,CAAA;AAMrD;;;GAGG;AACH,qBAAa,WAAW;IAKpB,OAAO,CAAC,QAAQ;IAChB,OAAO,CAAC,KAAK;IACb,OAAO,CAAC,cAAc,CAAC;IACvB,OAAO,CAAC,YAAY,CAAC;IACrB,OAAO,CAAC,WAAW,CAAC;IARtB,OAAO,CAAC,MAAM,CAAgC;IAC9C,OAAO,CAAC,cAAc,CAAsB;gBAGlC,QAAQ,EAAE,QAAQ,EAClB,KAAK,EAAE,MAAM,EACb,cAAc,CAAC,EAAE,MAAM,EACvB,YAAY,CAAC,EAAE,MAAM,EACrB,WAAW,CAAC,EAAE,UAAU;IAG5B,OAAO;IAmBb,qFAAqF;IACrF,OAAO,CAAC,eAAe;IAUvB,IAAI,MAAM,IAAI,MAAM,CAEnB;IAEK,kBAAkB,IAAI,OAAO,CAAC,MAAM,CAAC;IAY3C;;;OAGG;IACG,eAAe,IAAI,OAAO,CAAC,MAAM,CAAC;IAOxC;;OAEG;IACG,WAAW,CAAC,OAAO,EAAE,MAAM,GAAG,OAAO,CAAC,IAAI,CAAC;IAM3C,SAAS,CAAC,OAAO,EAAE,MAAM,CAAC,MAAM,EAAE,OAAO,CAAC,GAAG,OAAO,CAAC,IAAI,CAAC;CAKjE;AAED;;;;;GAKG;AACH,wBAAsB,iBAAiB,CAAC,MAAM,EAAE,QAAQ,EAAE,KAAK,EAAE,MAAM,GAAG,OAAO,CAAC,MAAM,CAAC,CAsBxF;AAED;;;;;GAKG;AACH,wBAAsB,yBAAyB,CAC7C,YAAY,EAAE,MAAM,EACpB,OAAO,GAAE,MAAM,CAAC,MAAM,EAAE,MAAM,CAAM,EACpC,SAAS,SAAoC,GAC5C,OAAO,CAAC,IAAI,CAAC,CA+Df;AAoBD;;;;;;;;;GASG;AACH,wBAAsB,sBAAsB,CAC1C,YAAY,EAAE,MAAM,EACpB,OAAO,GAAE,MAAM,CAAC,MAAM,EAAE,MAAM,CAAM,EACpC,MAAM,GAAE,UAA0B,GACjC,OAAO,CAAC,IAAI,CAAC,CA8Cf"}

package/dist/client.js

@@ -1,7 +1,3 @@
-// src/client.ts
-import { Client as BpClient } from "@botpress/client";
-import { createRequire } from "module";
-
 // src/types.ts
 var defaultLogger = console;
 
@@ -49,33 +45,27 @@ var EvalRunnerError = class extends AdkError {
 };
 
 // src/client.ts
-var defaultChatClient = null;
-function setDefaultChatClient(chatClient) {
-  defaultChatClient = chatClient;
-}
-function getDefaultChatClient() {
-  if (defaultChatClient) return defaultChatClient;
-  const require2 = createRequire(import.meta.url);
-  const mod = require2("@botpress/chat");
-  const ChatClient = mod.default?.Client ?? mod.Client;
-  if (!ChatClient) {
-    throw new EvalRunnerError({
-      code: "CHAT_CLIENT_MISSING",
-      message: "@botpress/chat did not expose a Client export. Pass EvalRunnerConfig.chatClient explicitly."
-    });
-  }
-  return ChatClient;
-}
+var TRACE_STREAM_PREFLIGHT_TIMEOUT_MS = 1e4;
 var ChatSession = class {
-  constructor(webhookId, chatClient) {
-    this.webhookId = webhookId;
-    this.chatClient = chatClient;
+  constructor(bpClient, botId, _chatWebhookId, _chatBaseUrl, _chatClient) {
+    this.bpClient = bpClient;
+    this.botId = botId;
+    this._chatWebhookId = _chatWebhookId;
+    this._chatBaseUrl = _chatBaseUrl;
+    this._chatClient = _chatClient;
   }
   client = null;
   conversationId = null;
   async connect() {
-    const ChatClient = this.chatClient ?? getDefaultChatClient();
-    this.client = await ChatClient.connect({ webhookId: this.webhookId });
+    const webhookId = this._chatWebhookId ?? await discoverWebhookId(this.bpClient, this.botId);
+    const ChatClientCtor = this._chatClient ?? (await import(
+      /* webpackIgnore: true */
+      "@botpress/chat"
+    )).Client;
+    this.client = await ChatClientCtor.connect({
+      webhookId,
+      ...this._chatBaseUrl ? { baseApiUrl: this._chatBaseUrl } : {}
+    });
   }
   /** Invariant guard — calling any session method before connect() is a runner bug. */
   assertConnected() {
@@ -99,6 +89,16 @@ var ChatSession = class {
     }
     return this.conversationId;
   }
+  /**
+   * Open a fresh conversation under the same user (the user is fixed at
+   * connect()) and make it active. Returns the new conversation id.
+   */
+  async newConversation() {
+    const client = this.assertConnected();
+    const conv = await client.createConversation({});
+    this.conversationId = conv.conversation.id;
+    return conv.conversation.id;
+  }
   /**
    * Send a user message. Fire-and-forget — observation is via SSECollector.
    */
@@ -107,17 +107,13 @@ var ChatSession = class {
     const conversationId = await this.ensureConversation();
     await client.createMessage({ conversationId, payload: { type: "text", text: message } });
   }
-  /**
-   * Send an event. Fire-and-forget — observation is via SSECollector.
-   */
-  async sendEvent(type, payload) {
+  async sendEvent(payload) {
     const client = this.assertConnected();
     const conversationId = await this.ensureConversation();
-    await client.createEvent({ type, payload, conversationId });
+    await client.createEvent({ payload, conversationId });
   }
 };
-async function discoverWebhookId(botId, token, apiUrl) {
-  const client = new BpClient({ token, botId, apiUrl });
+async function discoverWebhookId(client, botId) {
   const { bot } = await client.getBot({ id: botId });
   const integrations = bot.integrations || {};
   const chat = Object.values(integrations).find((int) => int.name === "chat");
@@ -129,7 +125,7 @@ async function discoverWebhookId(botId, token, apiUrl) {
         "The `chat` integration is not installed on this bot \u2014 evals require it.",
         "",
         "To fix:",
-        "  1. Install it:    adk add chat",
+        "  1. Install it:    adk integrations add chat",
         "  2. Redeploy:      adk dev   (or: adk deploy)"
       ].join("\n"),
       expected: true
@@ -137,6 +133,77 @@ async function discoverWebhookId(botId, token, apiUrl) {
   }
   return webhookId;
 }
+async function assertTraceStreamReadable(devServerUrl, headers = {}, timeoutMs = TRACE_STREAM_PREFLIGHT_TIMEOUT_MS) {
+  const url = `${devServerUrl.replace(/\/$/, "")}/api/traces/stream?count=1`;
+  const abortController = new AbortController();
+  let timedOut = false;
+  let reader = null;
+  const timer = setTimeout(() => {
+    timedOut = true;
+    abortController.abort();
+  }, timeoutMs);
+  try {
+    const res = await fetch(url, {
+      headers: { ...headers, Accept: "text/event-stream" },
+      signal: abortController.signal
+    });
+    if (!res.ok) {
+      throw traceStreamPreflightError(`Trace stream pre-flight failed: ${res.status} ${res.statusText}`);
+    }
+    if (!res.body) {
+      throw traceStreamPreflightError("Trace stream pre-flight failed: response had no body.");
+    }
+    reader = res.body.getReader();
+    const decoder = new TextDecoder();
+    let buffer = "";
+    while (true) {
+      const { done, value } = await reader.read();
+      if (done) {
+        throw traceStreamPreflightError("Trace stream pre-flight failed: stream closed before snapshot.");
+      }
+      buffer += decoder.decode(value, { stream: true });
+      const parts = buffer.split("\n\n");
+      buffer = parts.pop() ?? "";
+      for (const part of parts) {
+        if (parseSSEEventName(part) === "snapshot") {
+          return;
+        }
+      }
+    }
+  } catch (err) {
+    if (err instanceof EvalRunnerError) {
+      throw err;
+    }
+    if (timedOut) {
+      throw traceStreamPreflightError(
+        `Trace stream pre-flight timed out after ${timeoutMs}ms before receiving a snapshot.`
+      );
+    }
+    throw traceStreamPreflightError(
+      `Trace stream pre-flight failed: ${err instanceof Error ? err.message : String(err)}`
+    );
+  } finally {
+    clearTimeout(timer);
+    reader?.cancel().catch(() => {
+    });
+  }
+}
+function parseSSEEventName(block) {
+  for (const line of block.split("\n")) {
+    if (line.startsWith("event:")) {
+      return line.slice("event:".length).trim();
+    }
+  }
+  return null;
+}
+function traceStreamPreflightError(message) {
+  return new EvalRunnerError({
+    code: "SSE_CONNECT_FAILED",
+    message,
+    expected: true,
+    suggestion: "Make sure the dev server is running (`adk dev`) and trace streaming is available."
+  });
+}
 async function assertChatChannelBound(devServerUrl, headers = {}, logger = defaultLogger) {
   let conversations = [];
   try {
@@ -176,7 +243,7 @@ async function assertChatChannelBound(devServerUrl, headers = {}, logger = defau
 export {
   ChatSession,
   assertChatChannelBound,
-  discoverWebhookId,
-  setDefaultChatClient
+  assertTraceStreamReadable,
+  discoverWebhookId
 };
 //# sourceMappingURL=client.js.map

package/dist/client.js.map

@@ -1,7 +1,7 @@
 {
   "version": 3,
-  "sources": ["../src/client.ts", "../src/types.ts", "../../analytics/src/errors.ts", "../src/errors.ts"],
-  "sourcesContent": ["/**\n * Chat client for eval conversations.\n * Drives conversations against a running ADK bot via @botpress/chat.\n *\n * This is a send-only client. Observation (bot responses, tool calls, state)\n * is handled by the SSECollector via trace spans \u2014 not via WebSocket listeners.\n */\n\nimport { Client as BpClient } from '@botpress/client'\nimport { createRequire } from 'module'\nimport type { ChatClient, ChatClientConnection, EvalLogger } from './types'\nimport { defaultLogger } from './types'\nimport { EvalRunnerError } from './errors'\n\nlet defaultChatClient: ChatClient | null = null\n\nexport function setDefaultChatClient(chatClient: ChatClient | null): void {\n  defaultChatClient = chatClient\n}\n\nfunction getDefaultChatClient(): ChatClient {\n  if (defaultChatClient) return defaultChatClient\n\n  const require = createRequire(import.meta.url)\n  const mod = require('@botpress/chat') as { Client?: ChatClient; default?: { Client?: ChatClient } }\n  const ChatClient = mod.default?.Client ?? mod.Client\n  if (!ChatClient) {\n    throw new EvalRunnerError({\n      code: 'CHAT_CLIENT_MISSING',\n      message: '@botpress/chat did not expose a Client export. Pass EvalRunnerConfig.chatClient explicitly.',\n    })\n  }\n  return ChatClient\n}\n\n/**\n * A send-only chat session that maintains a single client connection across turns.\n * Messages and events are fire-and-forget; the SSECollector observes results via traces.\n */\nexport class ChatSession {\n  private client: ChatClientConnection | null = null\n  private conversationId: string | null = null\n\n  constructor(\n    private webhookId: string,\n    private chatClient?: ChatClient\n  ) {}\n\n  async connect() {\n    const ChatClient = this.chatClient ?? getDefaultChatClient()\n    this.client = await ChatClient.connect({ webhookId: this.webhookId })\n  }\n\n  /** Invariant guard \u2014 calling any session method before connect() is a runner bug. */\n  private assertConnected(): ChatClientConnection {\n    if (!this.client) {\n      throw new EvalRunnerError({\n        code: 'CHAT_NOT_CONNECTED',\n        message: 'ChatSession not connected. Call connect() first.',\n      })\n    }\n    return this.client\n  }\n\n  get userId(): string {\n    return this.assertConnected().user.id\n  }\n\n  async ensureConversation(): Promise<string> {\n    const client = this.assertConnected()\n\n    if (!this.conversationId) {\n      const conv = await client.createConversation({})\n      this.conversationId = conv.conversation.id\n      return conv.conversation.id\n    }\n\n    return this.conversationId\n  }\n\n  /**\n   * Send a user message. Fire-and-forget \u2014 observation is via SSECollector.\n   */\n  async sendMessage(message: string): Promise<void> {\n    const client = this.assertConnected()\n    const conversationId = await this.ensureConversation()\n    await client.createMessage({ conversationId, payload: { type: 'text', text: message } })\n  }\n\n  /**\n   * Send an event. Fire-and-forget \u2014 observation is via SSECollector.\n   */\n  async sendEvent(type: string, payload: Record<string, unknown>): Promise<void> {\n    const client = this.assertConnected()\n    const conversationId = await this.ensureConversation()\n    await client.createEvent({ type, payload, conversationId })\n  }\n}\n\n/**\n * Discover the chat integration's webhookId from a bot.\n *\n * Evals talk to the bot through the `chat` integration (via @botpress/chat).\n * If the bot doesn't have that integration installed, evals can't run.\n */\nexport async function discoverWebhookId(botId: string, token: string, apiUrl: string): Promise<string> {\n  const client = new BpClient({ token, botId, apiUrl })\n  const { bot } = await client.getBot({ id: botId })\n  const integrations = bot.integrations || {}\n\n  const chat = Object.values(integrations).find((int) => (int as Record<string, unknown>).name === 'chat')\n  const webhookId = (chat as Record<string, unknown> | undefined)?.webhookId\n\n  if (!webhookId || typeof webhookId !== 'string') {\n    throw new EvalRunnerError({\n      code: 'CHAT_INTEGRATION_MISSING',\n      message: [\n        'The `chat` integration is not installed on this bot \u2014 evals require it.',\n        '',\n        'To fix:',\n        '  1. Install it:    adk add chat',\n        '  2. Redeploy:      adk dev   (or: adk deploy)',\n      ].join('\\n'),\n      expected: true,\n    })\n  }\n\n  return webhookId\n}\n\n/**\n * Verify the bot has at least one Conversation bound to `chat.channel`.\n *\n * The `chat` integration can be installed without any handler listening on its\n * channel \u2014 in that case messages sent via @botpress/chat produce no response\n * and the caller times out silently. Fail fast with actionable guidance instead.\n *\n * Uses the dev server's `/api/agent` endpoint, so this is a no-op against\n * production bots (no devServerUrl).\n */\nexport async function assertChatChannelBound(\n  devServerUrl: string,\n  headers: Record<string, string> = {},\n  logger: EvalLogger = defaultLogger\n): Promise<void> {\n  type ConversationSummary = { channel?: string | string[] }\n  let conversations: ConversationSummary[] = []\n  try {\n    const res = await fetch(`${devServerUrl}/api/agent`, { headers })\n    if (!res.ok) return // Dev server unreachable \u2014 skip, let downstream errors surface\n    const agent = (await res.json()) as { conversations?: ConversationSummary[] }\n    conversations = agent.conversations ?? []\n  } catch (err) {\n    // Soft pre-flight only \u2014 a missing dev server is normal (e.g. prod bots),\n    // but log the skip so a *broken* dev server (malformed /api/agent JSON)\n    // isn't invisible while every eval afterwards times out mysteriously.\n    logger?.warn(\n      `Skipping chat.channel binding pre-flight \u2014 could not query ${devServerUrl}/api/agent: ${err instanceof Error ? err.message : String(err)}`\n    )\n    return\n  }\n\n  if (conversations.length === 0) return // Bot has no conversations at all \u2014 different problem\n\n  // Mirrors the runtime's matchesChannel semantics: a handler matches `chat.channel`\n  // if it is the wildcard '*', the literal string 'chat.channel', or an array\n  // containing either.\n  const matchesChatChannel = (ch: string | string[] | undefined): boolean => {\n    if (!ch) return false\n    if (ch === '*' || ch === 'chat.channel') return true\n    return Array.isArray(ch) && (ch.includes('chat.channel') || ch.includes('*'))\n  }\n\n  const bound = conversations.some((c) => matchesChatChannel(c.channel))\n  if (bound) return\n\n  throw new EvalRunnerError({\n    code: 'CHAT_CHANNEL_UNBOUND',\n    message: [\n      'No Conversation is bound to `chat.channel` \u2014 messages via `@botpress/chat` will time out silently.',\n      '',\n      'To fix, bind a Conversation in `src/conversations/` to `chat.channel` (or `*`):',\n      \"  channel: 'chat.channel',           // explicit\",\n      \"  channel: '*',                      // wildcard \u2014 matches every channel\",\n      \"  channel: ['chat.channel', ...],    // array form\",\n      '',\n      'The `chat` integration is how evals and `adk chat --single` send messages programmatically.',\n    ].join('\\n'),\n    expected: true,\n  })\n}\n", "/**\n * Eval type definitions.\n *\n * User-facing types live in @botpress/evals so eval authors can update eval\n * tooling without tying the authoring API to the agent runtime package.\n */\n\nexport {\n  Eval,\n  type EvalDefinition,\n  type EvalSetup,\n  type ConversationTurn,\n  type TurnAssertions,\n  type OutcomeAssertions,\n  type ResponseAssertion,\n  type ToolAssertion,\n  type StateAssertion,\n  type WorkflowAssertion,\n  type TimingAssertion,\n  type MatchOperator,\n  type NumericOperator,\n} from './definition'\n\nexport type { Span, SpanStatus } from './trace'\n\nexport interface ChatClientConnection {\n  user: { id: string }\n  createConversation: (props: Record<string, unknown>) => Promise<{ conversation: { id: string } }>\n  // Broad parameter types keep this compatible with @botpress/chat's narrower\n  // generated payload unions while avoiding a dependency on those definitions.\n  createMessage: (props: any) => Promise<unknown>\n  createEvent: (props: any) => Promise<unknown>\n}\n\nexport interface ChatClient {\n  connect: (props: { webhookId: string }) => Promise<ChatClientConnection>\n}\n\n// --- Grader result (universal currency for all graders) ---\n\nexport interface GraderResult {\n  assertion: string\n  pass: boolean\n  expected: string\n  actual: string\n}\n\n// --- Tool call extracted from traces ---\n\nexport interface ToolCall {\n  name: string\n  input: Record<string, unknown>\n  output: string\n  status: string\n}\n\n// --- Turn report ---\n\nexport interface TurnReport {\n  turnNumber: number\n  userMessage: string\n  botResponse: string\n  assertions: GraderResult[]\n  pass: boolean\n  /** Time spent waiting for the bot to respond (ms) */\n  botDuration: number\n  /** Time spent fetching traces + running graders (ms) */\n  evalDuration: number\n}\n\n// --- Single eval report ---\n\nexport interface EvalReport {\n  name: string\n  description?: string\n  type?: 'capability' | 'regression'\n  tags?: string[]\n  turns: TurnReport[]\n  outcomeAssertions: GraderResult[]\n  pass: boolean\n  duration: number\n  error?: string\n  /** Stable EvalErrorCode when `error` came from a typed EvalRunnerError \u2014 lets consumers distinguish setup/config failures from bot failures. */\n  errorCode?: string\n}\n\n// --- Progress event for real-time UI updates ---\n\nexport type EvalProgressEvent =\n  | { type: 'suite_start'; totalEvals: number }\n  | { type: 'eval_start'; evalName: string; index: number; totalTurns: number }\n  | {\n      type: 'turn_start'\n      evalName: string\n      evalIndex: number\n      turnNumber: number\n      totalTurns: number\n      userMessage: string\n    }\n  | {\n      type: 'turn_complete'\n      evalName: string\n      evalIndex: number\n      turnNumber: number\n      totalTurns: number\n      turnReport: TurnReport\n    }\n  | { type: 'eval_complete'; evalName: string; index: number; report: EvalReport }\n  | { type: 'suite_complete'; report: EvalRunReport; error?: string }\n\n// --- Eval run report (full suite) ---\n\nexport interface EvalRunReport {\n  id: string\n  timestamp: string\n  evals: EvalReport[]\n  passed: number\n  failed: number\n  total: number\n  duration: number\n  filter?: {\n    names?: string[]\n    tags?: string[]\n    type?: 'capability' | 'regression'\n  }\n  /**\n   * True when the suite was aborted mid-run (e.g. user clicked Stop). The\n   * persisted report is partial \u2014 `evals` reflects only the completed evals,\n   * and `total` matches that count, not the original suite size.\n   */\n  aborted?: boolean\n}\n\n// --- Bot connection info ---\n\nexport interface BotConnection {\n  webhookId: string\n  botId: string\n  token: string\n  apiUrl: string\n}\n\n// --- Logger interface ---\n\nexport interface EvalLogger {\n  info: (message: string) => void\n  warn: (message: string) => void\n  error: (message: string) => void\n}\n\n/**\n * Fallback when no logger is injected (programmatic use outside the CLI).\n * console rather than a no-op so warnings stay visible \u2014 a silent default\n * would hide judge failures and skipped pre-flights, the exact swallow\n * anti-pattern docs/ERROR-HANDLING.md bans. The CLI always injects its own\n * structured logger via EvalRunnerConfig.logger.\n */\nexport const defaultLogger: EvalLogger = console\n\n// --- Eval runner configuration ---\n\nexport interface EvalRunnerConfig {\n  credentials: {\n    token: string\n    apiUrl: string\n    botId: string\n    workspaceId?: string\n    webhookId?: string\n  }\n  agentPath: string\n  /** URL of the dev server providing traces and config. Defaults to the discovered DevConsole URL. */\n  devServerUrl?: string\n  /** Headers for dev-server requests, e.g. X-Agent-Path when routed through the singleton proxy. */\n  devServerHeaders?: Record<string, string>\n  /**\n   * Optional host-provided chat client. The CLI passes its bundled chat client\n   * here so the eval engine does not depend on @botpress/adk.\n   */\n  chatClient?: ChatClient\n  onProgress?: (event: EvalProgressEvent) => void\n  /** Logger for warnings and errors. Defaults to console. */\n  logger?: EvalLogger\n  /**\n   * Called for unexpected (internal-bug) eval failures that runEval folds\n   * into report.error \u2014 they never propagate to the CLI's command boundary,\n   * so without this hook they'd be invisible to error tracking. The CLI binds\n   * this to telemetry.captureException; evals itself stays telemetry-free.\n   */\n  onException?: (error: unknown, properties?: Record<string, unknown>) => void\n  /**\n   * Optional abort signal \u2014 when fired, the suite stops starting new evals.\n   * The currently running eval continues to completion (LLM calls in flight\n   * cannot always be cancelled), but no further evals are started.\n   */\n  signal?: AbortSignal\n  /** Bot-level eval options from agent.config.ts. Cascades: eval > agent config > default. */\n  evalOptions?: {\n    idleTimeout?: number\n    /** LLM judge pass threshold (integer, 1-5). Values outside this range are clamped. Cascades: eval > agent config > default (3). */\n    judgePassThreshold?: number\n    /** Model to use for llm_judge assertions (e.g. 'openai:gpt-4o'). Defaults to 'fast'. */\n    judgeModel?: string\n  }\n}\n\n// --- Eval suite filter ---\n\nexport interface EvalFilter {\n  names?: string[]\n  tags?: string[]\n  type?: 'capability' | 'regression'\n}\n\n// --- Transformer types (trace-driven eval engine) ---\n\nexport interface StateMutation {\n  type: string // 'bot', 'user', 'conversation'\n  changedKeys: string[]\n  previous: unknown\n  current: unknown\n  swappedToFile: boolean // data['swapped_to_file'] === true\n}\n\nexport interface WorkflowSpan {\n  name: string\n  status: 'running' | 'ok' | 'error'\n  statusFinal?: string // 'completed', 'continue', 'failed'\n}\n\nexport interface TurnData {\n  messages: string[]\n  toolCalls: ToolCall[]\n  stateMutations: StateMutation[]\n  workflowSpans: WorkflowSpan[]\n  handlerDuration: number\n  handlerStatus: 'ok' | 'error'\n}\n", "/**\n * Shared base error for ADK packages. Lives in `@botpress/analytics` because\n * it's the dependency-light package every surface already shares (cli, adk,\n * ui) \u2014 so all typed errors can genuinely inherit one base instead of\n * mirroring its shape.\n *\n * The policy lives in docs/ERROR-HANDLING.md. The short version: every error\n * we *intend* to throw should carry a stable `code` (so PostHog error\n * tracking can slice issues by code) and an `expected` flag separating\n * user/environment conditions (bad input, not logged in, network down \u2014 render\n * a friendly message, low alert priority) from internal bugs (invariant\n * violations \u2014 full stack, page someone).\n *\n * The runtime package is out of scope and keeps its own `Errors.AbstractError`\n * hierarchy. The adk typed families (`DependencyError`, `Agent0*Error`) extend\n * this base; new code should start here.\n */\nexport class AdkError<Code extends string = string> extends Error {\n  /**\n   * Marker for cross-module-instance checks \u2014 `instanceof` breaks when a\n   * package is duplicated in the dependency graph.\n   * @internal\n   */\n  public static readonly __IS_ADK_BASE_ERROR = true\n\n  /**\n   * Stable SCREAMING_SNAKE_CASE identifier, e.g. 'BOT_NOT_FOUND'. Subclasses\n   * pin `Code` to a literal union (`class EvalRunnerError extends\n   * AdkError<EvalErrorCode>`) so typos are compile errors and `code` narrows\n   * in switches.\n   */\n  readonly code: Code\n  /**\n   * True for user/environment conditions the developer can fix themselves\n   * (validation, auth, connectivity). False for internal bugs. Defaults to\n   * false so unclassified errors surface as bugs until someone triages them.\n   */\n  readonly expected: boolean\n  /** Structured context. Sanitized before any telemetry leaves the machine. */\n  readonly details?: Record<string, unknown>\n  /** Actionable next step shown to the user, e.g. \"Run 'adk login' first.\" */\n  readonly suggestion?: string\n\n  constructor(opts: {\n    code: Code\n    message: string\n    expected?: boolean\n    details?: Record<string, unknown>\n    suggestion?: string\n    cause?: unknown\n  }) {\n    super(opts.message, opts.cause !== undefined ? { cause: opts.cause } : undefined)\n    this.name = this.constructor.name\n    this.code = opts.code\n    this.expected = opts.expected ?? false\n    // Conditional assignment keeps the optional fields absent-not-undefined,\n    // which downstream tsconfigs with exactOptionalPropertyTypes require.\n    if (opts.details !== undefined) {\n      this.details = opts.details\n    }\n    if (opts.suggestion !== undefined) {\n      this.suggestion = opts.suggestion\n    }\n  }\n}\n\n/**\n * Duck-typed guard that survives duplicated module instances in the\n * dependency graph, where `instanceof AdkError` returns false for an error\n * constructed from another copy of this module. No manual prototype walking\n * needed: static properties are inherited through the constructor chain, so\n * a subclass's constructor resolves the marker in one lookup.\n */\nexport function isAdkError(error: unknown): error is AdkError {\n  if (error instanceof AdkError) return true\n  if (!(error instanceof Error)) return false\n  return (error.constructor as unknown as Record<string, unknown> | undefined)?.__IS_ADK_BASE_ERROR === true\n}\n", "import { AdkError } from '@botpress/analytics'\n\n/**\n * Stable error codes for the eval engine. Typed as a literal union so a typo'd\n * code is a compile error and PostHog issue-slicing stays consistent.\n * Conventions per docs/ERROR-HANDLING.md.\n */\nexport const EVAL_ERROR_CODES = [\n  'EVAL_LOAD_FAILED',\n  'EVAL_FILE_EMPTY',\n  'EVAL_DUPLICATE_NAME',\n  'EVAL_SEED_NO_CONVERSATION',\n  'EVAL_NO_CONVERSATION_ID',\n  'EVAL_TURN_CONFIG_INVALID',\n  'SSE_CONNECT_FAILED',\n  'SSE_NO_BODY',\n  'CHAT_CLIENT_MISSING',\n  'CHAT_NOT_CONNECTED',\n  'CHAT_INTEGRATION_MISSING',\n  'CHAT_CHANNEL_UNBOUND',\n] as const\n\nexport type EvalErrorCode = (typeof EVAL_ERROR_CODES)[number]\n\n/**\n * Typed error for the eval engine. Named EvalRunnerError (not EvalError) to\n * avoid shadowing the JS built-in `EvalError`.\n */\nexport class EvalRunnerError extends AdkError<EvalErrorCode> {}\n"],
-  "mappings": ";AAQA,SAAS,UAAU,gBAAgB;AACnC,SAAS,qBAAqB;;;ACoJvB,IAAM,gBAA4B;;;AC5IlC,IAAM,WAAN,cAAqD,MAAM;AAAA;AAAA;AAAA;AAAA;AAAA;AAAA,EAMhE,OAAuB,sBAAsB;AAAA;AAAA;AAAA;AAAA;AAAA;AAAA;AAAA,EAQpC;AAAA;AAAA;AAAA;AAAA;AAAA;AAAA,EAMA;AAAA;AAAA,EAEA;AAAA;AAAA,EAEA;AAAA,EAET,YAAY,MAOT;AACD,UAAM,KAAK,SAAS,KAAK,UAAU,SAAY,EAAE,OAAO,KAAK,MAAM,IAAI,MAAS;AAChF,SAAK,OAAO,KAAK,YAAY;AAC7B,SAAK,OAAO,KAAK;AACjB,SAAK,WAAW,KAAK,YAAY;AAGjC,QAAI,KAAK,YAAY,QAAW;AAC9B,WAAK,UAAU,KAAK;AAAA,IACtB;AACA,QAAI,KAAK,eAAe,QAAW;AACjC,WAAK,aAAa,KAAK;AAAA,IACzB;AAAA,EACF;AACF;;;ACpCO,IAAM,kBAAN,cAA8B,SAAwB;AAAC;;;AHd9D,IAAI,oBAAuC;AAEpC,SAAS,qBAAqB,YAAqC;AACxE,sBAAoB;AACtB;AAEA,SAAS,uBAAmC;AAC1C,MAAI,kBAAmB,QAAO;AAE9B,QAAMA,WAAU,cAAc,YAAY,GAAG;AAC7C,QAAM,MAAMA,SAAQ,gBAAgB;AACpC,QAAM,aAAa,IAAI,SAAS,UAAU,IAAI;AAC9C,MAAI,CAAC,YAAY;AACf,UAAM,IAAI,gBAAgB;AAAA,MACxB,MAAM;AAAA,MACN,SAAS;AAAA,IACX,CAAC;AAAA,EACH;AACA,SAAO;AACT;AAMO,IAAM,cAAN,MAAkB;AAAA,EAIvB,YACU,WACA,YACR;AAFQ;AACA;AAAA,EACP;AAAA,EANK,SAAsC;AAAA,EACtC,iBAAgC;AAAA,EAOxC,MAAM,UAAU;AACd,UAAM,aAAa,KAAK,cAAc,qBAAqB;AAC3D,SAAK,SAAS,MAAM,WAAW,QAAQ,EAAE,WAAW,KAAK,UAAU,CAAC;AAAA,EACtE;AAAA;AAAA,EAGQ,kBAAwC;AAC9C,QAAI,CAAC,KAAK,QAAQ;AAChB,YAAM,IAAI,gBAAgB;AAAA,QACxB,MAAM;AAAA,QACN,SAAS;AAAA,MACX,CAAC;AAAA,IACH;AACA,WAAO,KAAK;AAAA,EACd;AAAA,EAEA,IAAI,SAAiB;AACnB,WAAO,KAAK,gBAAgB,EAAE,KAAK;AAAA,EACrC;AAAA,EAEA,MAAM,qBAAsC;AAC1C,UAAM,SAAS,KAAK,gBAAgB;AAEpC,QAAI,CAAC,KAAK,gBAAgB;AACxB,YAAM,OAAO,MAAM,OAAO,mBAAmB,CAAC,CAAC;AAC/C,WAAK,iBAAiB,KAAK,aAAa;AACxC,aAAO,KAAK,aAAa;AAAA,IAC3B;AAEA,WAAO,KAAK;AAAA,EACd;AAAA;AAAA;AAAA;AAAA,EAKA,MAAM,YAAY,SAAgC;AAChD,UAAM,SAAS,KAAK,gBAAgB;AACpC,UAAM,iBAAiB,MAAM,KAAK,mBAAmB;AACrD,UAAM,OAAO,cAAc,EAAE,gBAAgB,SAAS,EAAE,MAAM,QAAQ,MAAM,QAAQ,EAAE,CAAC;AAAA,EACzF;AAAA;AAAA;AAAA;AAAA,EAKA,MAAM,UAAU,MAAc,SAAiD;AAC7E,UAAM,SAAS,KAAK,gBAAgB;AACpC,UAAM,iBAAiB,MAAM,KAAK,mBAAmB;AACrD,UAAM,OAAO,YAAY,EAAE,MAAM,SAAS,eAAe,CAAC;AAAA,EAC5D;AACF;AAQA,eAAsB,kBAAkB,OAAe,OAAe,QAAiC;AACrG,QAAM,SAAS,IAAI,SAAS,EAAE,OAAO,OAAO,OAAO,CAAC;AACpD,QAAM,EAAE,IAAI,IAAI,MAAM,OAAO,OAAO,EAAE,IAAI,MAAM,CAAC;AACjD,QAAM,eAAe,IAAI,gBAAgB,CAAC;AAE1C,QAAM,OAAO,OAAO,OAAO,YAAY,EAAE,KAAK,CAAC,QAAS,IAAgC,SAAS,MAAM;AACvG,QAAM,YAAa,MAA8C;AAEjE,MAAI,CAAC,aAAa,OAAO,cAAc,UAAU;AAC/C,UAAM,IAAI,gBAAgB;AAAA,MACxB,MAAM;AAAA,MACN,SAAS;AAAA,QACP;AAAA,QACA;AAAA,QACA;AAAA,QACA;AAAA,QACA;AAAA,MACF,EAAE,KAAK,IAAI;AAAA,MACX,UAAU;AAAA,IACZ,CAAC;AAAA,EACH;AAEA,SAAO;AACT;AAYA,eAAsB,uBACpB,cACA,UAAkC,CAAC,GACnC,SAAqB,eACN;AAEf,MAAI,gBAAuC,CAAC;AAC5C,MAAI;AACF,UAAM,MAAM,MAAM,MAAM,GAAG,YAAY,cAAc,EAAE,QAAQ,CAAC;AAChE,QAAI,CAAC,IAAI,GAAI;AACb,UAAM,QAAS,MAAM,IAAI,KAAK;AAC9B,oBAAgB,MAAM,iBAAiB,CAAC;AAAA,EAC1C,SAAS,KAAK;AAIZ,YAAQ;AAAA,MACN,mEAA8D,YAAY,eAAe,eAAe,QAAQ,IAAI,UAAU,OAAO,GAAG,CAAC;AAAA,IAC3I;AACA;AAAA,EACF;AAEA,MAAI,cAAc,WAAW,EAAG;AAKhC,QAAM,qBAAqB,CAAC,OAA+C;AACzE,QAAI,CAAC,GAAI,QAAO;AAChB,QAAI,OAAO,OAAO,OAAO,eAAgB,QAAO;AAChD,WAAO,MAAM,QAAQ,EAAE,MAAM,GAAG,SAAS,cAAc,KAAK,GAAG,SAAS,GAAG;AAAA,EAC7E;AAEA,QAAM,QAAQ,cAAc,KAAK,CAAC,MAAM,mBAAmB,EAAE,OAAO,CAAC;AACrE,MAAI,MAAO;AAEX,QAAM,IAAI,gBAAgB;AAAA,IACxB,MAAM;AAAA,IACN,SAAS;AAAA,MACP;AAAA,MACA;AAAA,MACA;AAAA,MACA;AAAA,MACA;AAAA,MACA;AAAA,MACA;AAAA,MACA;AAAA,IACF,EAAE,KAAK,IAAI;AAAA,IACX,UAAU;AAAA,EACZ,CAAC;AACH;",
-  "names": ["require"]
+  "sources": ["../src/types.ts", "../../analytics/src/errors.ts", "../src/errors.ts", "../src/client.ts"],
+  "sourcesContent": ["/**\n * Eval type definitions.\n *\n * User-facing types live in @botpress/evals so eval authors can update eval\n * tooling without tying the authoring API to the agent runtime package.\n */\n\nimport type { EvalDefinition as _EvalDefinition } from './definition'\n\nexport {\n  Eval,\n  type EvalDefinition,\n  type EvalSetup,\n  type ConversationTurn,\n  type TurnAssertions,\n  type OutcomeAssertions,\n  type ResponseAssertion,\n  type ToolAssertion,\n  type StateAssertion,\n  type WorkflowAssertion,\n  type TimingAssertion,\n  type MatchOperator,\n  type NumericOperator,\n} from './definition'\n\nexport type { Span, SpanStatus } from './spans/trace'\nexport type { WaitOptions, WorkflowWaitOptions } from './spans/span-source'\nimport type { SpanSource } from './spans/span-source'\nexport type { SpanSource }\n\n/**\n * The @botpress/chat Client *class* (constructor) \u2014 not an instance. Callers\n * inject it (the CLI passes its bundled chat client) and the runner calls the\n * static `.connect()`. Typed via `typeof import(...)` so this types-only module\n * doesn't pull in a value import.\n */\nexport type ChatClient = typeof import('@botpress/chat').Client\n\n// --- Grader result (universal currency for all graders) ---\n\nexport interface GraderResult {\n  assertion: string\n  pass: boolean\n  expected: string\n  actual: string\n  /**\n   * True when the assertion could not actually be evaluated (e.g. the llm_judge had no credentials,\n   * returned an empty response, or errored). Distinguishes a genuine pass from a \"didn't run\" pass.\n   * Callers that demand a trustworthy verdict (e.g. the benchmark) should treat `skipped` as a failure.\n   */\n  skipped?: boolean\n}\n\n// --- Tool call extracted from traces ---\n\nexport interface ToolCall {\n  name: string\n  input: Record<string, unknown>\n  output: string\n  status: string\n}\n\n// --- Turn report ---\n\nexport interface TurnReport {\n  turnIndex: number\n  userMessage: string\n  botResponse: string\n  assertions: GraderResult[]\n  pass: boolean\n  /** Time spent waiting for the bot to respond (ms) */\n  botDuration: number\n  /** Time spent fetching traces + running graders (ms) */\n  evalDuration: number\n}\n\n// --- Single eval report ---\n\nexport interface EvalReport {\n  name: string\n  description?: string\n  type?: 'capability' | 'regression'\n  tags?: string[]\n  turns: TurnReport[]\n  outcomeAssertions: GraderResult[]\n  pass: boolean\n  duration: number\n  error?: string\n  /** Stable EvalErrorCode when `error` came from a typed EvalRunnerError \u2014 lets consumers distinguish setup/config failures from bot failures. */\n  errorCode?: string\n}\n\n// --- Progress event for real-time UI updates ---\n\nexport type EvalProgressEvent =\n  | { type: 'suite_start'; totalEvals: number }\n  | { type: 'eval_start'; evalName: string; index: number; totalTurns: number }\n  | {\n      type: 'turn_start'\n      evalName: string\n      evalIndex: number\n      turnIndex: number\n      totalTurns: number\n      userMessage: string\n    }\n  | {\n      type: 'turn_complete'\n      evalName: string\n      evalIndex: number\n      turnIndex: number\n      totalTurns: number\n      turnReport: TurnReport\n    }\n  | { type: 'eval_complete'; evalName: string; index: number; report: EvalReport }\n  | { type: 'suite_complete'; report: EvalRunReport; error?: string }\n\n// --- Eval run report (full suite) ---\n\nexport interface EvalRunReport {\n  id: string\n  timestamp: string\n  evals: EvalReport[]\n  passed: number\n  failed: number\n  total: number\n  duration: number\n  filter?: {\n    names?: string[]\n    tags?: string[]\n    type?: 'capability' | 'regression'\n  }\n  /**\n   * True when the suite was aborted mid-run (e.g. user clicked Stop). The\n   * persisted report is partial \u2014 `evals` reflects only the completed evals,\n   * and `total` matches that count, not the original suite size.\n   */\n  aborted?: boolean\n}\n\n// --- Bot connection info ---\n\nexport interface BotConnection {\n  client: import('@botpress/client').Client\n  botId: string\n}\n\n// --- Logger interface ---\n\nexport interface EvalLogger {\n  info: (message: string) => void\n  warn: (message: string) => void\n  error: (message: string) => void\n}\n\n/**\n * Fallback when no logger is injected (programmatic use outside the CLI).\n * console rather than a no-op so warnings stay visible \u2014 a silent default\n * would hide judge failures and skipped pre-flights, the exact swallow\n * anti-pattern docs/ERROR-HANDLING.md bans. The CLI always injects its own\n * structured logger via EvalRunnerConfig.logger.\n */\nexport const defaultLogger: EvalLogger = console\n\n// --- Eval runner configuration ---\n\nexport interface EvalRunnerConfig {\n  /** Authenticated Botpress client. Used for state seeding, workflow triggers, webhook discovery, and LLM judge. */\n  client: import('@botpress/client').Client\n  botId: string\n  /** Path to the agent directory. Used to discover eval files from `<agentPath>/evals/`. Not needed when `definitions` is provided. */\n  agentPath?: string\n  /** Pre-loaded eval definitions. When provided, `agentPath` is not used for eval discovery. */\n  definitions?: _EvalDefinition[]\n  /** URL of the dev server providing traces and config. Defaults to the discovered DevConsole URL. */\n  devServerUrl?: string\n  /** Headers for dev-server requests, e.g. X-Agent-Path when routed through the singleton proxy. */\n  devServerHeaders?: Record<string, string>\n  /**\n   * Optional host-provided chat client. The CLI passes its bundled chat client\n   * here so the eval engine does not depend on @botpress/adk.\n   */\n  chatClient?: ChatClient\n  onProgress?: (event: EvalProgressEvent) => void | Promise<void>\n  /** Logger for warnings and errors. Defaults to console. */\n  logger?: EvalLogger\n  /**\n   * Called for unexpected (internal-bug) eval failures that runEval folds\n   * into report.error \u2014 they never propagate to the CLI's command boundary,\n   * so without this hook they'd be invisible to error tracking. The CLI binds\n   * this to telemetry.captureException; evals itself stays telemetry-free.\n   */\n  onException?: (error: unknown, properties?: Record<string, unknown>) => void\n  /**\n   * Optional abort signal \u2014 when fired, the suite stops starting new evals.\n   * The currently running eval continues to completion (LLM calls in flight\n   * cannot always be cancelled), but no further evals are started.\n   */\n  signal?: AbortSignal\n  /** Bot-level eval options from agent.config.ts. Cascades: eval > agent config > default. */\n  evalOptions?: {\n    idleTimeout?: number\n    /** LLM judge pass threshold (integer, 1-5). Values outside this range are clamped. Cascades: eval > agent config > default (3). */\n    judgePassThreshold?: number\n    /** Model to use for llm_judge assertions (e.g. 'openai:gpt-4o'). Defaults to 'fast'. */\n    judgeModel?: string\n  }\n  /** Factory for creating span sources. Each eval gets its own instance. Defaults to LocalSpanSource (SSE from dev server). */\n  createSpanSource?: () => SpanSource\n  /** Externally-provided run ID. When set, the runner uses this instead of generating its own. */\n  runId?: string\n  /** Pre-resolved chat integration webhook ID. Skips the getBot() discovery call when provided. */\n  chatWebhookId?: string\n  /** Base URL for the chat service (e.g. https://chat.botpress.dev). Defaults to https://chat.botpress.cloud. */\n  chatBaseUrl?: string\n}\n\n// --- Eval suite filter ---\n\nexport interface EvalFilter {\n  names?: string[]\n  tags?: string[]\n  type?: 'capability' | 'regression'\n}\n\n// --- Transformer types (trace-driven eval engine) ---\n\nexport interface StateMutation {\n  type: string // 'bot', 'user', 'conversation'\n  changedKeys: string[]\n  previous: unknown\n  current: unknown\n  swappedToFile: boolean // data['swapped_to_file'] === true\n}\n\nexport interface WorkflowSpan {\n  name: string\n  status: 'running' | 'ok' | 'error'\n  statusFinal?: string // 'completed', 'continue', 'failed'\n}\n\nexport interface TurnData {\n  messages: string[]\n  toolCalls: ToolCall[]\n  stateMutations: StateMutation[]\n  workflowSpans: WorkflowSpan[]\n  handlerDuration: number\n  handlerStatus: 'ok' | 'error'\n}\n", "/**\n * Shared base error for ADK packages. Lives in `@botpress/analytics` because\n * it's the dependency-light package every surface already shares (cli, adk,\n * ui) \u2014 so all typed errors can genuinely inherit one base instead of\n * mirroring its shape.\n *\n * The policy lives in docs/ERROR-HANDLING.md. The short version: every error\n * we *intend* to throw should carry a stable `code` (so PostHog error\n * tracking can slice issues by code) and an `expected` flag separating\n * user/environment conditions (bad input, not logged in, network down \u2014 render\n * a friendly message, low alert priority) from internal bugs (invariant\n * violations \u2014 full stack, page someone).\n *\n * The runtime package is out of scope and keeps its own `Errors.AbstractError`\n * hierarchy. The adk typed families (`DependencyError`, `Agent0*Error`) extend\n * this base; new code should start here.\n */\nexport class AdkError<Code extends string = string> extends Error {\n  /**\n   * Marker for cross-module-instance checks \u2014 `instanceof` breaks when a\n   * package is duplicated in the dependency graph.\n   * @internal\n   */\n  public static readonly __IS_ADK_BASE_ERROR = true\n\n  /**\n   * Stable SCREAMING_SNAKE_CASE identifier, e.g. 'BOT_NOT_FOUND'. Subclasses\n   * pin `Code` to a literal union (`class EvalRunnerError extends\n   * AdkError<EvalErrorCode>`) so typos are compile errors and `code` narrows\n   * in switches.\n   */\n  readonly code: Code\n  /**\n   * True for user/environment conditions the developer can fix themselves\n   * (validation, auth, connectivity). False for internal bugs. Defaults to\n   * false so unclassified errors surface as bugs until someone triages them.\n   */\n  readonly expected: boolean\n  /** Structured context. Sanitized before any telemetry leaves the machine. */\n  readonly details?: Record<string, unknown>\n  /** Actionable next step shown to the user, e.g. \"Run 'adk login' first.\" */\n  readonly suggestion?: string\n\n  constructor(opts: {\n    code: Code\n    message: string\n    expected?: boolean\n    details?: Record<string, unknown>\n    suggestion?: string\n    cause?: unknown\n  }) {\n    super(opts.message, opts.cause !== undefined ? { cause: opts.cause } : undefined)\n    this.name = this.constructor.name\n    this.code = opts.code\n    this.expected = opts.expected ?? false\n    // Conditional assignment keeps the optional fields absent-not-undefined,\n    // which downstream tsconfigs with exactOptionalPropertyTypes require.\n    if (opts.details !== undefined) {\n      this.details = opts.details\n    }\n    if (opts.suggestion !== undefined) {\n      this.suggestion = opts.suggestion\n    }\n  }\n}\n\n/**\n * Duck-typed guard that survives duplicated module instances in the\n * dependency graph, where `instanceof AdkError` returns false for an error\n * constructed from another copy of this module. No manual prototype walking\n * needed: static properties are inherited through the constructor chain, so\n * a subclass's constructor resolves the marker in one lookup.\n */\nexport function isAdkError(error: unknown): error is AdkError {\n  if (error instanceof AdkError) return true\n  if (!(error instanceof Error)) return false\n  return (error.constructor as unknown as Record<string, unknown> | undefined)?.__IS_ADK_BASE_ERROR === true\n}\n", "import { AdkError } from '@botpress/analytics'\n\n/**\n * Stable error codes for the eval engine. Typed as a literal union so a typo'd\n * code is a compile error and PostHog issue-slicing stays consistent.\n * Conventions per docs/ERROR-HANDLING.md.\n */\nexport const EVAL_ERROR_CODES = [\n  'EVAL_LOAD_FAILED',\n  'EVAL_FILE_EMPTY',\n  'EVAL_DUPLICATE_NAME',\n  'EVAL_SEED_NO_CONVERSATION',\n  'EVAL_NO_CONVERSATION_ID',\n  'EVAL_TURN_CONFIG_INVALID',\n  'SSE_CONNECT_FAILED',\n  'SSE_NO_BODY',\n  'CHAT_CLIENT_MISSING',\n  'CHAT_NOT_CONNECTED',\n  'CHAT_INTEGRATION_MISSING',\n  'CHAT_CHANNEL_UNBOUND',\n] as const\n\nexport type EvalErrorCode = (typeof EVAL_ERROR_CODES)[number]\n\n/**\n * Typed error for the eval engine. Named EvalRunnerError (not EvalError) to\n * avoid shadowing the JS built-in `EvalError`.\n */\nexport class EvalRunnerError extends AdkError<EvalErrorCode> {}\n", "/**\n * Chat client for eval conversations.\n * Drives conversations against a running ADK bot via @botpress/chat.\n *\n * This is a send-only client. Observation (bot responses, tool calls, state)\n * is handled by the SSECollector via trace spans \u2014 not via WebSocket listeners.\n */\n\nimport { Client as BpClient } from '@botpress/client'\nimport type { AuthenticatedClient as AuthedChatClient } from '@botpress/chat'\nimport type { ChatClient, EvalLogger } from './types'\nimport { defaultLogger } from './types'\nimport { EvalRunnerError } from './errors'\n\nconst TRACE_STREAM_PREFLIGHT_TIMEOUT_MS = 10_000\n\n/**\n * A send-only chat session that maintains a single client connection across turns.\n * Messages and events are fire-and-forget; the SSECollector observes results via traces.\n */\nexport class ChatSession {\n  private client: AuthedChatClient | null = null\n  private conversationId: string | null = null\n\n  constructor(\n    private bpClient: BpClient,\n    private botId: string,\n    private _chatWebhookId?: string,\n    private _chatBaseUrl?: string,\n    private _chatClient?: ChatClient\n  ) {}\n\n  async connect() {\n    const webhookId = this._chatWebhookId ?? (await discoverWebhookId(this.bpClient, this.botId))\n\n    // Prefer the host-injected client. The CLI passes its bundled CJS chat client\n    // (working node http adapter); only fall back to a dynamic import when nothing\n    // was injected \u2014 e.g. the cloud runtime, where @botpress/chat resolves to a\n    // real node_modules dep. In a bun-compiled CLI binary the dynamic import\n    // resolves to the browser ESM bundle whose axios http adapter is null\n    // (\"Adapter 'http' is not available in the build\"), so the injected client is\n    // required there.\n    const ChatClientCtor =\n      this._chatClient ?? (await import(/* webpackIgnore: true */ '@botpress/chat' as string)).Client\n\n    this.client = await ChatClientCtor.connect({\n      webhookId,\n      ...(this._chatBaseUrl ? { baseApiUrl: this._chatBaseUrl } : {}),\n    })\n  }\n\n  /** Invariant guard \u2014 calling any session method before connect() is a runner bug. */\n  private assertConnected(): AuthedChatClient {\n    if (!this.client) {\n      throw new EvalRunnerError({\n        code: 'CHAT_NOT_CONNECTED',\n        message: 'ChatSession not connected. Call connect() first.',\n      })\n    }\n    return this.client\n  }\n\n  get userId(): string {\n    return this.assertConnected().user.id\n  }\n\n  async ensureConversation(): Promise<string> {\n    const client = this.assertConnected()\n\n    if (!this.conversationId) {\n      const conv = await client.createConversation({})\n      this.conversationId = conv.conversation.id\n      return conv.conversation.id\n    }\n\n    return this.conversationId\n  }\n\n  /**\n   * Open a fresh conversation under the same user (the user is fixed at\n   * connect()) and make it active. Returns the new conversation id.\n   */\n  async newConversation(): Promise<string> {\n    const client = this.assertConnected()\n    const conv = await client.createConversation({})\n    this.conversationId = conv.conversation.id\n    return conv.conversation.id\n  }\n\n  /**\n   * Send a user message. Fire-and-forget \u2014 observation is via SSECollector.\n   */\n  async sendMessage(message: string): Promise<void> {\n    const client = this.assertConnected()\n    const conversationId = await this.ensureConversation()\n    await client.createMessage({ conversationId, payload: { type: 'text', text: message } })\n  }\n\n  async sendEvent(payload: Record<string, unknown>): Promise<void> {\n    const client = this.assertConnected()\n    const conversationId = await this.ensureConversation()\n    await client.createEvent({ payload, conversationId })\n  }\n}\n\n/**\n * Discover the chat integration's webhookId from a bot.\n *\n * Evals talk to the bot through the `chat` integration (via @botpress/chat).\n * If the bot doesn't have that integration installed, evals can't run.\n */\nexport async function discoverWebhookId(client: BpClient, botId: string): Promise<string> {\n  const { bot } = await client.getBot({ id: botId })\n  const integrations = bot.integrations || {}\n\n  const chat = Object.values(integrations).find((int) => (int as Record<string, unknown>).name === 'chat')\n  const webhookId = (chat as Record<string, unknown> | undefined)?.webhookId\n\n  if (!webhookId || typeof webhookId !== 'string') {\n    throw new EvalRunnerError({\n      code: 'CHAT_INTEGRATION_MISSING',\n      message: [\n        'The `chat` integration is not installed on this bot \u2014 evals require it.',\n        '',\n        'To fix:',\n        '  1. Install it:    adk integrations add chat',\n        '  2. Redeploy:      adk dev   (or: adk deploy)',\n      ].join('\\n'),\n      expected: true,\n    })\n  }\n\n  return webhookId\n}\n\n/**\n * Verify the dev server's trace stream can be read before eval turns start.\n *\n * This only checks that `/api/traces/stream` is reachable and emits its initial\n * snapshot. It does not prove the private runtime span-ingest path is healthy.\n */\nexport async function assertTraceStreamReadable(\n  devServerUrl: string,\n  headers: Record<string, string> = {},\n  timeoutMs = TRACE_STREAM_PREFLIGHT_TIMEOUT_MS\n): Promise<void> {\n  const url = `${devServerUrl.replace(/\\/$/, '')}/api/traces/stream?count=1`\n  const abortController = new AbortController()\n  let timedOut = false\n  let reader: ReadableStreamDefaultReader<Uint8Array> | null = null\n\n  const timer = setTimeout(() => {\n    timedOut = true\n    abortController.abort()\n  }, timeoutMs)\n\n  try {\n    const res = await fetch(url, {\n      headers: { ...headers, Accept: 'text/event-stream' },\n      signal: abortController.signal,\n    })\n\n    if (!res.ok) {\n      throw traceStreamPreflightError(`Trace stream pre-flight failed: ${res.status} ${res.statusText}`)\n    }\n\n    if (!res.body) {\n      throw traceStreamPreflightError('Trace stream pre-flight failed: response had no body.')\n    }\n\n    reader = res.body.getReader()\n    const decoder = new TextDecoder()\n    let buffer = ''\n\n    while (true) {\n      const { done, value } = await reader.read()\n      if (done) {\n        throw traceStreamPreflightError('Trace stream pre-flight failed: stream closed before snapshot.')\n      }\n\n      buffer += decoder.decode(value, { stream: true })\n      const parts = buffer.split('\\n\\n')\n      buffer = parts.pop() ?? ''\n\n      for (const part of parts) {\n        if (parseSSEEventName(part) === 'snapshot') {\n          return\n        }\n      }\n    }\n  } catch (err) {\n    if (err instanceof EvalRunnerError) {\n      throw err\n    }\n\n    if (timedOut) {\n      throw traceStreamPreflightError(\n        `Trace stream pre-flight timed out after ${timeoutMs}ms before receiving a snapshot.`\n      )\n    }\n\n    throw traceStreamPreflightError(\n      `Trace stream pre-flight failed: ${err instanceof Error ? err.message : String(err)}`\n    )\n  } finally {\n    clearTimeout(timer)\n    reader?.cancel().catch(() => {})\n  }\n}\n\nfunction parseSSEEventName(block: string): string | null {\n  for (const line of block.split('\\n')) {\n    if (line.startsWith('event:')) {\n      return line.slice('event:'.length).trim()\n    }\n  }\n  return null\n}\n\nfunction traceStreamPreflightError(message: string): EvalRunnerError {\n  return new EvalRunnerError({\n    code: 'SSE_CONNECT_FAILED',\n    message,\n    expected: true,\n    suggestion: 'Make sure the dev server is running (`adk dev`) and trace streaming is available.',\n  })\n}\n\n/**\n * Verify the bot has at least one Conversation bound to `chat.channel`.\n *\n * The `chat` integration can be installed without any handler listening on its\n * channel \u2014 in that case messages sent via @botpress/chat produce no response\n * and the caller times out silently. Fail fast with actionable guidance instead.\n *\n * Uses the dev server's `/api/agent` endpoint, so this is a no-op against\n * production bots (no devServerUrl).\n */\nexport async function assertChatChannelBound(\n  devServerUrl: string,\n  headers: Record<string, string> = {},\n  logger: EvalLogger = defaultLogger\n): Promise<void> {\n  type ConversationSummary = { channel?: string | string[] }\n  let conversations: ConversationSummary[] = []\n  try {\n    const res = await fetch(`${devServerUrl}/api/agent`, { headers })\n    if (!res.ok) return // Dev server unreachable \u2014 skip, let downstream errors surface\n    const agent = (await res.json()) as { conversations?: ConversationSummary[] }\n    conversations = agent.conversations ?? []\n  } catch (err) {\n    // Soft pre-flight only \u2014 a missing dev server is normal (e.g. prod bots),\n    // but log the skip so a *broken* dev server (malformed /api/agent JSON)\n    // isn't invisible while every eval afterwards times out mysteriously.\n    logger?.warn(\n      `Skipping chat.channel binding pre-flight \u2014 could not query ${devServerUrl}/api/agent: ${err instanceof Error ? err.message : String(err)}`\n    )\n    return\n  }\n\n  if (conversations.length === 0) return // Bot has no conversations at all \u2014 different problem\n\n  // Mirrors the runtime's matchesChannel semantics: a handler matches `chat.channel`\n  // if it is the wildcard '*', the literal string 'chat.channel', or an array\n  // containing either.\n  const matchesChatChannel = (ch: string | string[] | undefined): boolean => {\n    if (!ch) return false\n    if (ch === '*' || ch === 'chat.channel') return true\n    return Array.isArray(ch) && (ch.includes('chat.channel') || ch.includes('*'))\n  }\n\n  const bound = conversations.some((c) => matchesChatChannel(c.channel))\n  if (bound) return\n\n  throw new EvalRunnerError({\n    code: 'CHAT_CHANNEL_UNBOUND',\n    message: [\n      'No Conversation is bound to `chat.channel` \u2014 messages via `@botpress/chat` will time out silently.',\n      '',\n      'To fix, bind a Conversation in `src/conversations/` to `chat.channel` (or `*`):',\n      \"  channel: 'chat.channel',           // explicit\",\n      \"  channel: '*',                      // wildcard \u2014 matches every channel\",\n      \"  channel: ['chat.channel', ...],    // array form\",\n      '',\n      'The `chat` integration is how evals and `adk chat --single` send messages programmatically.',\n    ].join('\\n'),\n    expected: true,\n  })\n}\n"],
+  "mappings": ";AAiKO,IAAM,gBAA4B;;;AChJlC,IAAM,WAAN,cAAqD,MAAM;AAAA;AAAA;AAAA;AAAA;AAAA;AAAA,EAMhE,OAAuB,sBAAsB;AAAA;AAAA;AAAA;AAAA;AAAA;AAAA;AAAA,EAQpC;AAAA;AAAA;AAAA;AAAA;AAAA;AAAA,EAMA;AAAA;AAAA,EAEA;AAAA;AAAA,EAEA;AAAA,EAET,YAAY,MAOT;AACD,UAAM,KAAK,SAAS,KAAK,UAAU,SAAY,EAAE,OAAO,KAAK,MAAM,IAAI,MAAS;AAChF,SAAK,OAAO,KAAK,YAAY;AAC7B,SAAK,OAAO,KAAK;AACjB,SAAK,WAAW,KAAK,YAAY;AAGjC,QAAI,KAAK,YAAY,QAAW;AAC9B,WAAK,UAAU,KAAK;AAAA,IACtB;AACA,QAAI,KAAK,eAAe,QAAW;AACjC,WAAK,aAAa,KAAK;AAAA,IACzB;AAAA,EACF;AACF;;;ACpCO,IAAM,kBAAN,cAA8B,SAAwB;AAAC;;;ACd9D,IAAM,oCAAoC;AAMnC,IAAM,cAAN,MAAkB;AAAA,EAIvB,YACU,UACA,OACA,gBACA,cACA,aACR;AALQ;AACA;AACA;AACA;AACA;AAAA,EACP;AAAA,EATK,SAAkC;AAAA,EAClC,iBAAgC;AAAA,EAUxC,MAAM,UAAU;AACd,UAAM,YAAY,KAAK,kBAAmB,MAAM,kBAAkB,KAAK,UAAU,KAAK,KAAK;AAS3F,UAAM,iBACJ,KAAK,gBAAgB,MAAM;AAAA;AAAA,MAAiC;AAAA,IAA0B,GAAG;AAE3F,SAAK,SAAS,MAAM,eAAe,QAAQ;AAAA,MACzC;AAAA,MACA,GAAI,KAAK,eAAe,EAAE,YAAY,KAAK,aAAa,IAAI,CAAC;AAAA,IAC/D,CAAC;AAAA,EACH;AAAA;AAAA,EAGQ,kBAAoC;AAC1C,QAAI,CAAC,KAAK,QAAQ;AAChB,YAAM,IAAI,gBAAgB;AAAA,QACxB,MAAM;AAAA,QACN,SAAS;AAAA,MACX,CAAC;AAAA,IACH;AACA,WAAO,KAAK;AAAA,EACd;AAAA,EAEA,IAAI,SAAiB;AACnB,WAAO,KAAK,gBAAgB,EAAE,KAAK;AAAA,EACrC;AAAA,EAEA,MAAM,qBAAsC;AAC1C,UAAM,SAAS,KAAK,gBAAgB;AAEpC,QAAI,CAAC,KAAK,gBAAgB;AACxB,YAAM,OAAO,MAAM,OAAO,mBAAmB,CAAC,CAAC;AAC/C,WAAK,iBAAiB,KAAK,aAAa;AACxC,aAAO,KAAK,aAAa;AAAA,IAC3B;AAEA,WAAO,KAAK;AAAA,EACd;AAAA;AAAA;AAAA;AAAA;AAAA,EAMA,MAAM,kBAAmC;AACvC,UAAM,SAAS,KAAK,gBAAgB;AACpC,UAAM,OAAO,MAAM,OAAO,mBAAmB,CAAC,CAAC;AAC/C,SAAK,iBAAiB,KAAK,aAAa;AACxC,WAAO,KAAK,aAAa;AAAA,EAC3B;AAAA;AAAA;AAAA;AAAA,EAKA,MAAM,YAAY,SAAgC;AAChD,UAAM,SAAS,KAAK,gBAAgB;AACpC,UAAM,iBAAiB,MAAM,KAAK,mBAAmB;AACrD,UAAM,OAAO,cAAc,EAAE,gBAAgB,SAAS,EAAE,MAAM,QAAQ,MAAM,QAAQ,EAAE,CAAC;AAAA,EACzF;AAAA,EAEA,MAAM,UAAU,SAAiD;AAC/D,UAAM,SAAS,KAAK,gBAAgB;AACpC,UAAM,iBAAiB,MAAM,KAAK,mBAAmB;AACrD,UAAM,OAAO,YAAY,EAAE,SAAS,eAAe,CAAC;AAAA,EACtD;AACF;AAQA,eAAsB,kBAAkB,QAAkB,OAAgC;AACxF,QAAM,EAAE,IAAI,IAAI,MAAM,OAAO,OAAO,EAAE,IAAI,MAAM,CAAC;AACjD,QAAM,eAAe,IAAI,gBAAgB,CAAC;AAE1C,QAAM,OAAO,OAAO,OAAO,YAAY,EAAE,KAAK,CAAC,QAAS,IAAgC,SAAS,MAAM;AACvG,QAAM,YAAa,MAA8C;AAEjE,MAAI,CAAC,aAAa,OAAO,cAAc,UAAU;AAC/C,UAAM,IAAI,gBAAgB;AAAA,MACxB,MAAM;AAAA,MACN,SAAS;AAAA,QACP;AAAA,QACA;AAAA,QACA;AAAA,QACA;AAAA,QACA;AAAA,MACF,EAAE,KAAK,IAAI;AAAA,MACX,UAAU;AAAA,IACZ,CAAC;AAAA,EACH;AAEA,SAAO;AACT;AAQA,eAAsB,0BACpB,cACA,UAAkC,CAAC,GACnC,YAAY,mCACG;AACf,QAAM,MAAM,GAAG,aAAa,QAAQ,OAAO,EAAE,CAAC;AAC9C,QAAM,kBAAkB,IAAI,gBAAgB;AAC5C,MAAI,WAAW;AACf,MAAI,SAAyD;AAE7D,QAAM,QAAQ,WAAW,MAAM;AAC7B,eAAW;AACX,oBAAgB,MAAM;AAAA,EACxB,GAAG,SAAS;AAEZ,MAAI;AACF,UAAM,MAAM,MAAM,MAAM,KAAK;AAAA,MAC3B,SAAS,EAAE,GAAG,SAAS,QAAQ,oBAAoB;AAAA,MACnD,QAAQ,gBAAgB;AAAA,IAC1B,CAAC;AAED,QAAI,CAAC,IAAI,IAAI;AACX,YAAM,0BAA0B,mCAAmC,IAAI,MAAM,IAAI,IAAI,UAAU,EAAE;AAAA,IACnG;AAEA,QAAI,CAAC,IAAI,MAAM;AACb,YAAM,0BAA0B,uDAAuD;AAAA,IACzF;AAEA,aAAS,IAAI,KAAK,UAAU;AAC5B,UAAM,UAAU,IAAI,YAAY;AAChC,QAAI,SAAS;AAEb,WAAO,MAAM;AACX,YAAM,EAAE,MAAM,MAAM,IAAI,MAAM,OAAO,KAAK;AAC1C,UAAI,MAAM;AACR,cAAM,0BAA0B,gEAAgE;AAAA,MAClG;AAEA,gBAAU,QAAQ,OAAO,OAAO,EAAE,QAAQ,KAAK,CAAC;AAChD,YAAM,QAAQ,OAAO,MAAM,MAAM;AACjC,eAAS,MAAM,IAAI,KAAK;AAExB,iBAAW,QAAQ,OAAO;AACxB,YAAI,kBAAkB,IAAI,MAAM,YAAY;AAC1C;AAAA,QACF;AAAA,MACF;AAAA,IACF;AAAA,EACF,SAAS,KAAK;AACZ,QAAI,eAAe,iBAAiB;AAClC,YAAM;AAAA,IACR;AAEA,QAAI,UAAU;AACZ,YAAM;AAAA,QACJ,2CAA2C,SAAS;AAAA,MACtD;AAAA,IACF;AAEA,UAAM;AAAA,MACJ,mCAAmC,eAAe,QAAQ,IAAI,UAAU,OAAO,GAAG,CAAC;AAAA,IACrF;AAAA,EACF,UAAE;AACA,iBAAa,KAAK;AAClB,YAAQ,OAAO,EAAE,MAAM,MAAM;AAAA,IAAC,CAAC;AAAA,EACjC;AACF;AAEA,SAAS,kBAAkB,OAA8B;AACvD,aAAW,QAAQ,MAAM,MAAM,IAAI,GAAG;AACpC,QAAI,KAAK,WAAW,QAAQ,GAAG;AAC7B,aAAO,KAAK,MAAM,SAAS,MAAM,EAAE,KAAK;AAAA,IAC1C;AAAA,EACF;AACA,SAAO;AACT;AAEA,SAAS,0BAA0B,SAAkC;AACnE,SAAO,IAAI,gBAAgB;AAAA,IACzB,MAAM;AAAA,IACN;AAAA,IACA,UAAU;AAAA,IACV,YAAY;AAAA,EACd,CAAC;AACH;AAYA,eAAsB,uBACpB,cACA,UAAkC,CAAC,GACnC,SAAqB,eACN;AAEf,MAAI,gBAAuC,CAAC;AAC5C,MAAI;AACF,UAAM,MAAM,MAAM,MAAM,GAAG,YAAY,cAAc,EAAE,QAAQ,CAAC;AAChE,QAAI,CAAC,IAAI,GAAI;AACb,UAAM,QAAS,MAAM,IAAI,KAAK;AAC9B,oBAAgB,MAAM,iBAAiB,CAAC;AAAA,EAC1C,SAAS,KAAK;AAIZ,YAAQ;AAAA,MACN,mEAA8D,YAAY,eAAe,eAAe,QAAQ,IAAI,UAAU,OAAO,GAAG,CAAC;AAAA,IAC3I;AACA;AAAA,EACF;AAEA,MAAI,cAAc,WAAW,EAAG;AAKhC,QAAM,qBAAqB,CAAC,OAA+C;AACzE,QAAI,CAAC,GAAI,QAAO;AAChB,QAAI,OAAO,OAAO,OAAO,eAAgB,QAAO;AAChD,WAAO,MAAM,QAAQ,EAAE,MAAM,GAAG,SAAS,cAAc,KAAK,GAAG,SAAS,GAAG;AAAA,EAC7E;AAEA,QAAM,QAAQ,cAAc,KAAK,CAAC,MAAM,mBAAmB,EAAE,OAAO,CAAC;AACrE,MAAI,MAAO;AAEX,QAAM,IAAI,gBAAgB;AAAA,IACxB,MAAM;AAAA,IACN,SAAS;AAAA,MACP;AAAA,MACA;AAAA,MACA;AAAA,MACA;AAAA,MACA;AAAA,MACA;AAAA,MACA;AAAA,MACA;AAAA,IACF,EAAE,KAAK,IAAI;AAAA,IACX,UAAU;AAAA,EACZ,CAAC;AACH;",
+  "names": []
 }

package/dist/definition.d.ts

@@ -62,11 +62,16 @@ export interface TurnAssertions {
 export interface ConversationTurn {
     user?: string;
     event?: {
-        type: string;
-        payload?: Record<string, unknown>;
+        payload: Record<string, unknown>;
     };
     expectSilence?: boolean;
     assert?: TurnAssertions;
+    /**
+     * Run this turn in a NEW conversation under the same user. The prior
+     * transcript is absent, so recall here can only come from user-scoped
+     * persistence, not context. Ignored on the first turn.
+     */
+    newConversation?: boolean;
 }
 export interface OutcomeAssertions {
     state?: StateAssertion[];
@@ -93,6 +98,7 @@ export interface EvalDefinition {
     outcome?: OutcomeAssertions;
     options?: {
         idleTimeout?: number;
+        /** @deprecated No-op: the LLM judge is boolean (zai.check); there is no score threshold. */
         judgePassThreshold?: number;
     };
 }
@@ -106,6 +112,7 @@ export declare class Eval implements EvalDefinition {
     readonly outcome?: OutcomeAssertions;
     readonly options?: {
         idleTimeout?: number;
+        /** @deprecated No-op: the LLM judge is boolean (zai.check); there is no score threshold. */
         judgePassThreshold?: number;
     };
     constructor(def: EvalDefinition);

package/dist/definition.d.ts.map

@@ -1 +1 @@
-{"version":3,"file":"definition.d.ts","sourceRoot":"","sources":["../src/definition.ts"],"names":[],"mappings":"AAAA,MAAM,MAAM,aAAa,GACrB,MAAM,GACN;IAAE,MAAM,EAAE,OAAO,CAAA;CAAE,GACnB;IAAE,QAAQ,EAAE,MAAM,CAAA;CAAE,GACpB;IAAE,YAAY,EAAE,MAAM,CAAA;CAAE,GACxB;IAAE,OAAO,EAAE,MAAM,CAAA;CAAE,GACnB;IAAE,EAAE,EAAE,OAAO,EAAE,CAAA;CAAE,GACjB;IAAE,MAAM,EAAE,OAAO,CAAA;CAAE,GACnB;IAAE,GAAG,EAAE,MAAM,CAAA;CAAE,GACf;IAAE,GAAG,EAAE,MAAM,CAAA;CAAE,CAAA;AAEnB,MAAM,MAAM,iBAAiB,GACzB;IAAE,QAAQ,EAAE,MAAM,CAAA;CAAE,GACpB;IAAE,YAAY,EAAE,MAAM,CAAA;CAAE,GACxB;IAAE,OAAO,EAAE,MAAM,CAAA;CAAE,GACnB;IAAE,SAAS,EAAE,MAAM,CAAA;CAAE,CAAA;AAGzB,MAAM,MAAM,aAAa,GACrB;IAAE,MAAM,EAAE,MAAM,CAAC;IAAC,MAAM,CAAC,EAAE,MAAM,CAAC,MAAM,EAAE,aAAa,CAAC,CAAA;CAAE,GAC1D;IAAE,UAAU,EAAE,MAAM,CAAA;CAAE,GACtB;IAAE,UAAU,EAAE,MAAM,EAAE,CAAA;CAAE,CAAA;AAE5B,MAAM,WAAW,cAAc;IAC7B,IAAI,EAAE,MAAM,CAAA;IACZ,MAAM,CAAC,EAAE,OAAO,CAAA;IAChB,OAAO,CAAC,EAAE,OAAO,CAAA;CAClB;AAED,MAAM,WAAW,iBAAiB;IAChC,IAAI,EAAE,MAAM,CAAA;IACZ,OAAO,CAAC,EAAE,OAAO,CAAA;IACjB,SAAS,CAAC,EAAE,OAAO,CAAA;CACpB;AAED,MAAM,MAAM,eAAe,GAAG;IAAE,MAAM,EAAE,MAAM,CAAA;CAAE,GAAG;IAAE,GAAG,EAAE,MAAM,CAAA;CAAE,GAAG;IAAE,GAAG,EAAE,MAAM,CAAA;CAAE,CAAA;AAEpF,MAAM,WAAW,eAAe;IAC9B,aAAa,EAAE,eAAe,CAAA;CAC/B;AAED,MAAM,WAAW,cAAc;IAC7B,QAAQ,CAAC,EAAE,iBAAiB,EAAE,CAAA;IAC9B,KAAK,CAAC,EAAE,aAAa,EAAE,CAAA;IACvB,KAAK,CAAC,EAAE,cAAc,EAAE,CAAA;IACxB,QAAQ,CAAC,EAAE,iBAAiB,EAAE,CAAA;IAC9B,MAAM,CAAC,EAAE,eAAe,EAAE,CAAA;CAC3B;AAED,MAAM,WAAW,gBAAgB;IAC/B,IAAI,CAAC,EAAE,MAAM,CAAA;IACb,KAAK,CAAC,EAAE;QAAE,IAAI,EAAE,MAAM,CAAC;QAAC,OAAO,CAAC,EAAE,MAAM,CAAC,MAAM,EAAE,OAAO,CAAC,CAAA;KAAE,CAAA;IAC3D,aAAa,CAAC,EAAE,OAAO,CAAA;IACvB,MAAM,CAAC,EAAE,cAAc,CAAA;CACxB;AAED,MAAM,WAAW,iBAAiB;IAChC,KAAK,CAAC,EAAE,cAAc,EAAE,CAAA;IACxB,QAAQ,CAAC,EAAE,iBAAiB,EAAE,CAAA;CAC/B;AAED,MAAM,WAAW,SAAS;IACxB,KAAK,CAAC,EAAE;QACN,GAAG,CAAC,EAAE,MAAM,CAAC,MAAM,EAAE,OAAO,CAAC,CAAA;QAC7B,IAAI,CAAC,EAAE,MAAM,CAAC,MAAM,EAAE,OAAO,CAAC,CAAA;QAC9B,YAAY,CAAC,EAAE,MAAM,CAAC,MAAM,EAAE,OAAO,CAAC,CAAA;KACvC,CAAA;IACD,QAAQ,CAAC,EAAE;QACT,OAAO,EAAE,MAAM,CAAA;QACf,KAAK,CAAC,EAAE,MAAM,CAAC,MAAM,EAAE,OAAO,CAAC,CAAA;KAChC,CAAA;CACF;AAED,MAAM,WAAW,cAAc;IAC7B,IAAI,EAAE,MAAM,CAAA;IACZ,WAAW,CAAC,EAAE,MAAM,CAAA;IACpB,IAAI,CAAC,EAAE,MAAM,EAAE,CAAA;IACf,IAAI,CAAC,EAAE,YAAY,GAAG,YAAY,CAAA;IAClC,KAAK,CAAC,EAAE,SAAS,CAAA;IACjB,YAAY,EAAE,gBAAgB,EAAE,CAAA;IAChC,OAAO,CAAC,EAAE,iBAAiB,CAAA;IAC3B,OAAO,CAAC,EAAE;QACR,WAAW,CAAC,EAAE,MAAM,CAAA;QACpB,kBAAkB,CAAC,EAAE,MAAM,CAAA;KAC5B,CAAA;CACF;AAED,qBAAa,IAAK,YAAW,cAAc;IACzC,QAAQ,CAAC,IAAI,EAAE,MAAM,CAAA;IACrB,QAAQ,CAAC,WAAW,CAAC,EAAE,MAAM,CAAA;IAC7B,QAAQ,CAAC,IAAI,CAAC,EAAE,MAAM,EAAE,CAAA;IACxB,QAAQ,CAAC,IAAI,CAAC,EAAE,YAAY,GAAG,YAAY,CAAA;IAC3C,QAAQ,CAAC,KAAK,CAAC,EAAE,SAAS,CAAA;IAC1B,QAAQ,CAAC,YAAY,EAAE,gBAAgB,EAAE,CAAA;IACzC,QAAQ,CAAC,OAAO,CAAC,EAAE,iBAAiB,CAAA;IACpC,QAAQ,CAAC,OAAO,CAAC,EAAE;QACjB,WAAW,CAAC,EAAE,MAAM,CAAA;QACpB,kBAAkB,CAAC,EAAE,MAAM,CAAA;KAC5B,CAAA;gBAEW,GAAG,EAAE,cAAc;CAUhC"}
+{"version":3,"file":"definition.d.ts","sourceRoot":"","sources":["../src/definition.ts"],"names":[],"mappings":"AAAA,MAAM,MAAM,aAAa,GACrB,MAAM,GACN;IAAE,MAAM,EAAE,OAAO,CAAA;CAAE,GACnB;IAAE,QAAQ,EAAE,MAAM,CAAA;CAAE,GACpB;IAAE,YAAY,EAAE,MAAM,CAAA;CAAE,GACxB;IAAE,OAAO,EAAE,MAAM,CAAA;CAAE,GACnB;IAAE,EAAE,EAAE,OAAO,EAAE,CAAA;CAAE,GACjB;IAAE,MAAM,EAAE,OAAO,CAAA;CAAE,GACnB;IAAE,GAAG,EAAE,MAAM,CAAA;CAAE,GACf;IAAE,GAAG,EAAE,MAAM,CAAA;CAAE,CAAA;AAEnB,MAAM,MAAM,iBAAiB,GACzB;IAAE,QAAQ,EAAE,MAAM,CAAA;CAAE,GACpB;IAAE,YAAY,EAAE,MAAM,CAAA;CAAE,GACxB;IAAE,OAAO,EAAE,MAAM,CAAA;CAAE,GACnB;IAAE,SAAS,EAAE,MAAM,CAAA;CAAE,CAAA;AAGzB,MAAM,MAAM,aAAa,GACrB;IAAE,MAAM,EAAE,MAAM,CAAC;IAAC,MAAM,CAAC,EAAE,MAAM,CAAC,MAAM,EAAE,aAAa,CAAC,CAAA;CAAE,GAC1D;IAAE,UAAU,EAAE,MAAM,CAAA;CAAE,GACtB;IAAE,UAAU,EAAE,MAAM,EAAE,CAAA;CAAE,CAAA;AAE5B,MAAM,WAAW,cAAc;IAC7B,IAAI,EAAE,MAAM,CAAA;IACZ,MAAM,CAAC,EAAE,OAAO,CAAA;IAChB,OAAO,CAAC,EAAE,OAAO,CAAA;CAClB;AAED,MAAM,WAAW,iBAAiB;IAChC,IAAI,EAAE,MAAM,CAAA;IACZ,OAAO,CAAC,EAAE,OAAO,CAAA;IACjB,SAAS,CAAC,EAAE,OAAO,CAAA;CACpB;AAED,MAAM,MAAM,eAAe,GAAG;IAAE,MAAM,EAAE,MAAM,CAAA;CAAE,GAAG;IAAE,GAAG,EAAE,MAAM,CAAA;CAAE,GAAG;IAAE,GAAG,EAAE,MAAM,CAAA;CAAE,CAAA;AAEpF,MAAM,WAAW,eAAe;IAC9B,aAAa,EAAE,eAAe,CAAA;CAC/B;AAED,MAAM,WAAW,cAAc;IAC7B,QAAQ,CAAC,EAAE,iBAAiB,EAAE,CAAA;IAC9B,KAAK,CAAC,EAAE,aAAa,EAAE,CAAA;IACvB,KAAK,CAAC,EAAE,cAAc,EAAE,CAAA;IACxB,QAAQ,CAAC,EAAE,iBAAiB,EAAE,CAAA;IAC9B,MAAM,CAAC,EAAE,eAAe,EAAE,CAAA;CAC3B;AAED,MAAM,WAAW,gBAAgB;IAC/B,IAAI,CAAC,EAAE,MAAM,CAAA;IACb,KAAK,CAAC,EAAE;QAAE,OAAO,EAAE,MAAM,CAAC,MAAM,EAAE,OAAO,CAAC,CAAA;KAAE,CAAA;IAC5C,aAAa,CAAC,EAAE,OAAO,CAAA;IACvB,MAAM,CAAC,EAAE,cAAc,CAAA;IACvB;;;;OAIG;IACH,eAAe,CAAC,EAAE,OAAO,CAAA;CAC1B;AAED,MAAM,WAAW,iBAAiB;IAChC,KAAK,CAAC,EAAE,cAAc,EAAE,CAAA;IACxB,QAAQ,CAAC,EAAE,iBAAiB,EAAE,CAAA;CAC/B;AAED,MAAM,WAAW,SAAS;IACxB,KAAK,CAAC,EAAE;QACN,GAAG,CAAC,EAAE,MAAM,CAAC,MAAM,EAAE,OAAO,CAAC,CAAA;QAC7B,IAAI,CAAC,EAAE,MAAM,CAAC,MAAM,EAAE,OAAO,CAAC,CAAA;QAC9B,YAAY,CAAC,EAAE,MAAM,CAAC,MAAM,EAAE,OAAO,CAAC,CAAA;KACvC,CAAA;IACD,QAAQ,CAAC,EAAE;QACT,OAAO,EAAE,MAAM,CAAA;QACf,KAAK,CAAC,EAAE,MAAM,CAAC,MAAM,EAAE,OAAO,CAAC,CAAA;KAChC,CAAA;CACF;AAED,MAAM,WAAW,cAAc;IAC7B,IAAI,EAAE,MAAM,CAAA;IACZ,WAAW,CAAC,EAAE,MAAM,CAAA;IACpB,IAAI,CAAC,EAAE,MAAM,EAAE,CAAA;IACf,IAAI,CAAC,EAAE,YAAY,GAAG,YAAY,CAAA;IAClC,KAAK,CAAC,EAAE,SAAS,CAAA;IACjB,YAAY,EAAE,gBAAgB,EAAE,CAAA;IAChC,OAAO,CAAC,EAAE,iBAAiB,CAAA;IAC3B,OAAO,CAAC,EAAE;QACR,WAAW,CAAC,EAAE,MAAM,CAAA;QACpB,4FAA4F;QAC5F,kBAAkB,CAAC,EAAE,MAAM,CAAA;KAC5B,CAAA;CACF;AAED,qBAAa,IAAK,YAAW,cAAc;IACzC,QAAQ,CAAC,IAAI,EAAE,MAAM,CAAA;IACrB,QAAQ,CAAC,WAAW,CAAC,EAAE,MAAM,CAAA;IAC7B,QAAQ,CAAC,IAAI,CAAC,EAAE,MAAM,EAAE,CAAA;IACxB,QAAQ,CAAC,IAAI,CAAC,EAAE,YAAY,GAAG,YAAY,CAAA;IAC3C,QAAQ,CAAC,KAAK,CAAC,EAAE,SAAS,CAAA;IAC1B,QAAQ,CAAC,YAAY,EAAE,gBAAgB,EAAE,CAAA;IACzC,QAAQ,CAAC,OAAO,CAAC,EAAE,iBAAiB,CAAA;IACpC,QAAQ,CAAC,OAAO,CAAC,EAAE;QACjB,WAAW,CAAC,EAAE,MAAM,CAAA;QACpB,4FAA4F;QAC5F,kBAAkB,CAAC,EAAE,MAAM,CAAA;KAC5B,CAAA;gBAEW,GAAG,EAAE,cAAc;CAUhC"}

package/dist/definition.js.map

@@ -1,7 +1,7 @@
 {
   "version": 3,
   "sources": ["../src/definition.ts"],
-  "sourcesContent": ["export type MatchOperator =\n  | string\n  | { equals: unknown }\n  | { contains: string }\n  | { not_contains: string }\n  | { matches: string }\n  | { in: unknown[] }\n  | { exists: boolean }\n  | { gte: number }\n  | { lte: number }\n\nexport type ResponseAssertion =\n  | { contains: string }\n  | { not_contains: string }\n  | { matches: string }\n  | { llm_judge: string }\n// | { similar_to: string } TODO: uncomment this when it gets implemented\n\nexport type ToolAssertion =\n  | { called: string; params?: Record<string, MatchOperator> }\n  | { not_called: string }\n  | { call_order: string[] }\n\nexport interface StateAssertion {\n  path: string\n  equals?: unknown\n  changed?: boolean\n}\n\nexport interface WorkflowAssertion {\n  name: string\n  entered?: boolean\n  completed?: boolean\n}\n\nexport type NumericOperator = { equals: number } | { gte: number } | { lte: number }\n\nexport interface TimingAssertion {\n  response_time: NumericOperator\n}\n\nexport interface TurnAssertions {\n  response?: ResponseAssertion[]\n  tools?: ToolAssertion[]\n  state?: StateAssertion[]\n  workflow?: WorkflowAssertion[]\n  timing?: TimingAssertion[]\n}\n\nexport interface ConversationTurn {\n  user?: string\n  event?: { type: string; payload?: Record<string, unknown> }\n  expectSilence?: boolean\n  assert?: TurnAssertions\n}\n\nexport interface OutcomeAssertions {\n  state?: StateAssertion[]\n  workflow?: WorkflowAssertion[]\n}\n\nexport interface EvalSetup {\n  state?: {\n    bot?: Record<string, unknown>\n    user?: Record<string, unknown>\n    conversation?: Record<string, unknown>\n  }\n  workflow?: {\n    trigger: string\n    input?: Record<string, unknown>\n  }\n}\n\nexport interface EvalDefinition {\n  name: string\n  description?: string\n  tags?: string[]\n  type?: 'capability' | 'regression'\n  setup?: EvalSetup\n  conversation: ConversationTurn[]\n  outcome?: OutcomeAssertions\n  options?: {\n    idleTimeout?: number\n    judgePassThreshold?: number\n  }\n}\n\nexport class Eval implements EvalDefinition {\n  readonly name: string\n  readonly description?: string\n  readonly tags?: string[]\n  readonly type?: 'capability' | 'regression'\n  readonly setup?: EvalSetup\n  readonly conversation: ConversationTurn[]\n  readonly outcome?: OutcomeAssertions\n  readonly options?: {\n    idleTimeout?: number\n    judgePassThreshold?: number\n  }\n\n  constructor(def: EvalDefinition) {\n    this.name = def.name\n    this.conversation = def.conversation\n    if (def.description !== undefined) this.description = def.description\n    if (def.tags !== undefined) this.tags = def.tags\n    if (def.type !== undefined) this.type = def.type\n    if (def.setup !== undefined) this.setup = def.setup\n    if (def.outcome !== undefined) this.outcome = def.outcome\n    if (def.options !== undefined) this.options = def.options\n  }\n}\n"],
-  "mappings": ";AAuFO,IAAM,OAAN,MAAqC;AAAA,EACjC;AAAA,EACA;AAAA,EACA;AAAA,EACA;AAAA,EACA;AAAA,EACA;AAAA,EACA;AAAA,EACA;AAAA,EAKT,YAAY,KAAqB;AAC/B,SAAK,OAAO,IAAI;AAChB,SAAK,eAAe,IAAI;AACxB,QAAI,IAAI,gBAAgB,OAAW,MAAK,cAAc,IAAI;AAC1D,QAAI,IAAI,SAAS,OAAW,MAAK,OAAO,IAAI;AAC5C,QAAI,IAAI,SAAS,OAAW,MAAK,OAAO,IAAI;AAC5C,QAAI,IAAI,UAAU,OAAW,MAAK,QAAQ,IAAI;AAC9C,QAAI,IAAI,YAAY,OAAW,MAAK,UAAU,IAAI;AAClD,QAAI,IAAI,YAAY,OAAW,MAAK,UAAU,IAAI;AAAA,EACpD;AACF;",
+  "sourcesContent": ["export type MatchOperator =\n  | string\n  | { equals: unknown }\n  | { contains: string }\n  | { not_contains: string }\n  | { matches: string }\n  | { in: unknown[] }\n  | { exists: boolean }\n  | { gte: number }\n  | { lte: number }\n\nexport type ResponseAssertion =\n  | { contains: string }\n  | { not_contains: string }\n  | { matches: string }\n  | { llm_judge: string }\n// | { similar_to: string } TODO: uncomment this when it gets implemented\n\nexport type ToolAssertion =\n  | { called: string; params?: Record<string, MatchOperator> }\n  | { not_called: string }\n  | { call_order: string[] }\n\nexport interface StateAssertion {\n  path: string\n  equals?: unknown\n  changed?: boolean\n}\n\nexport interface WorkflowAssertion {\n  name: string\n  entered?: boolean\n  completed?: boolean\n}\n\nexport type NumericOperator = { equals: number } | { gte: number } | { lte: number }\n\nexport interface TimingAssertion {\n  response_time: NumericOperator\n}\n\nexport interface TurnAssertions {\n  response?: ResponseAssertion[]\n  tools?: ToolAssertion[]\n  state?: StateAssertion[]\n  workflow?: WorkflowAssertion[]\n  timing?: TimingAssertion[]\n}\n\nexport interface ConversationTurn {\n  user?: string\n  event?: { payload: Record<string, unknown> }\n  expectSilence?: boolean\n  assert?: TurnAssertions\n  /**\n   * Run this turn in a NEW conversation under the same user. The prior\n   * transcript is absent, so recall here can only come from user-scoped\n   * persistence, not context. Ignored on the first turn.\n   */\n  newConversation?: boolean\n}\n\nexport interface OutcomeAssertions {\n  state?: StateAssertion[]\n  workflow?: WorkflowAssertion[]\n}\n\nexport interface EvalSetup {\n  state?: {\n    bot?: Record<string, unknown>\n    user?: Record<string, unknown>\n    conversation?: Record<string, unknown>\n  }\n  workflow?: {\n    trigger: string\n    input?: Record<string, unknown>\n  }\n}\n\nexport interface EvalDefinition {\n  name: string\n  description?: string\n  tags?: string[]\n  type?: 'capability' | 'regression'\n  setup?: EvalSetup\n  conversation: ConversationTurn[]\n  outcome?: OutcomeAssertions\n  options?: {\n    idleTimeout?: number\n    /** @deprecated No-op: the LLM judge is boolean (zai.check); there is no score threshold. */\n    judgePassThreshold?: number\n  }\n}\n\nexport class Eval implements EvalDefinition {\n  readonly name: string\n  readonly description?: string\n  readonly tags?: string[]\n  readonly type?: 'capability' | 'regression'\n  readonly setup?: EvalSetup\n  readonly conversation: ConversationTurn[]\n  readonly outcome?: OutcomeAssertions\n  readonly options?: {\n    idleTimeout?: number\n    /** @deprecated No-op: the LLM judge is boolean (zai.check); there is no score threshold. */\n    judgePassThreshold?: number\n  }\n\n  constructor(def: EvalDefinition) {\n    this.name = def.name\n    this.conversation = def.conversation\n    if (def.description !== undefined) this.description = def.description\n    if (def.tags !== undefined) this.tags = def.tags\n    if (def.type !== undefined) this.type = def.type\n    if (def.setup !== undefined) this.setup = def.setup\n    if (def.outcome !== undefined) this.outcome = def.outcome\n    if (def.options !== undefined) this.options = def.options\n  }\n}\n"],
+  "mappings": ";AA8FO,IAAM,OAAN,MAAqC;AAAA,EACjC;AAAA,EACA;AAAA,EACA;AAAA,EACA;AAAA,EACA;AAAA,EACA;AAAA,EACA;AAAA,EACA;AAAA,EAMT,YAAY,KAAqB;AAC/B,SAAK,OAAO,IAAI;AAChB,SAAK,eAAe,IAAI;AACxB,QAAI,IAAI,gBAAgB,OAAW,MAAK,cAAc,IAAI;AAC1D,QAAI,IAAI,SAAS,OAAW,MAAK,OAAO,IAAI;AAC5C,QAAI,IAAI,SAAS,OAAW,MAAK,OAAO,IAAI;AAC5C,QAAI,IAAI,UAAU,OAAW,MAAK,QAAQ,IAAI;AAC9C,QAAI,IAAI,YAAY,OAAW,MAAK,UAAU,IAAI;AAClD,QAAI,IAAI,YAAY,OAAW,MAAK,UAAU,IAAI;AAAA,EACpD;AACF;",
   "names": []
 }

package/dist/graders/index.js

@@ -43,60 +43,31 @@ function operatorToString(operator) {
 }
 
 // src/graders/llm.ts
-import { Client } from "@botpress/client";
 import { Cognitive } from "@botpress/cognitive";
+import { Zai } from "@botpress/zai";
 
 // src/types.ts
 var defaultLogger = console;
 
 // src/graders/llm.ts
-var DEFAULT_PASS_THRESHOLD = 3;
-function buildJudgeSystemPrompt(passThreshold) {
-  return `You are an evaluation judge for a chatbot. You will be given:
-- The user's message
-- The bot's response
-- Grading criteria
-
-Your job is to determine if the bot's response meets the criteria.
-
-Respond with a JSON object:
-{
-  "pass": true/false,
-  "score": 1-5,
-  "reason": "brief explanation"
-}
-
-Scoring guide:
-- 5: Fully meets criteria, excellent response
-- 4: Mostly meets criteria, minor issues
-- 3: Partially meets criteria, notable gaps
-- 2: Barely meets criteria, significant issues
-- 1: Does not meet criteria
-
-A score of ${passThreshold} or above is a pass.`;
-}
 var MODEL_ALIASES = ["fast", "best"];
-var _cognitive = null;
+var _zai = null;
 var _availableModels = [];
 var _judgeModel = "fast";
+var _failClosed = true;
 var _log = defaultLogger;
 function _isModelValid(model) {
   if (MODEL_ALIASES.includes(model)) return true;
   if (_availableModels.length === 0) return false;
   return _availableModels.includes(model);
 }
-async function initLLMJudge(credentials, options) {
+async function initLLMJudge(client, options) {
   const log = options?.logger ?? defaultLogger;
   _log = log;
-  const client = new Client({
-    token: credentials.token,
-    apiUrl: credentials.apiUrl,
-    botId: credentials.botId,
-    ...credentials.workspaceId !== void 0 ? { workspaceId: credentials.workspaceId } : {}
-  });
-  _cognitive = new Cognitive({ client, __experimental_beta: true });
+  _failClosed = options?.failClosed ?? true;
+  const cognitive = new Cognitive({ client, __experimental_beta: true });
   try {
-    _availableModels = Array.from((await _cognitive.fetchRemoteModels()).keys());
+    _availableModels = Array.from((await cognitive.fetchRemoteModels()).keys());
   } catch (err) {
     log.warn(`Failed to fetch available models: ${err.message}. LLM judge will be unavailable.`);
     _availableModels = [];
@@ -107,66 +78,42 @@ async function initLLMJudge(credentials, options) {
   } else if (!_isModelValid(_judgeModel)) {
     log.warn(`Configured LLM judge model "${_judgeModel}" is invalid. Run "adk models" to list available models.`);
   }
+  _zai = new Zai({ client: cognitive, modelId: _judgeModel });
 }
 async function gradeLLMJudge(botResponse, criteria, context) {
-  const threshold = Math.max(1, Math.min(5, context.passThreshold ?? DEFAULT_PASS_THRESHOLD));
+  const assertion = `llm_judge: "${criteria}"`;
   try {
-    if (!_cognitive) {
+    if (!_zai) {
       _log.warn(`LLM judge unavailable for "${criteria}": no credentials configured (initLLMJudge not called)`);
       return {
-        assertion: `llm_judge: "${criteria}"`,
-        pass: false,
+        assertion,
+        pass: !_failClosed,
+        skipped: true,
         expected: criteria,
         actual: "FAILED \u2014 LLM judge unavailable: no credentials configured"
       };
     }
     if (!_isModelValid(_judgeModel)) {
       return {
-        assertion: `llm_judge: "${criteria}"`,
+        assertion,
         pass: false,
         expected: criteria,
         actual: "FAILED \u2014 invalid LLM judge model"
       };
     }
-    const { output } = await _cognitive.generateContent({
-      model: _judgeModel,
-      temperature: 0,
-      responseFormat: "json_object",
-      systemPrompt: buildJudgeSystemPrompt(threshold),
-      messages: [
-        {
-          role: "user",
-          content: `User message: ${context.userMessage}
-
-Bot response: ${botResponse}
-
-Criteria: ${criteria}`
-        }
-      ]
-    });
-    const rawContent = output?.choices?.[0]?.content;
-    const content = typeof rawContent === "string" ? rawContent : null;
-    if (!content) {
-      _log.warn(`LLM judge returned an empty response for "${criteria}"`);
-      return {
-        assertion: `llm_judge: "${criteria}"`,
-        pass: false,
-        expected: criteria,
-        actual: "FAILED \u2014 LLM judge returned empty response"
-      };
-    }
-    const verdict = JSON.parse(content);
+    const { output } = await _zai.check({ userMessage: context.userMessage, botResponse }, criteria).result();
     return {
-      assertion: `llm_judge: "${criteria}"`,
-      pass: verdict.score >= threshold,
+      assertion,
+      pass: output.value,
       expected: criteria,
-      actual: `Score ${verdict.score}/5 (threshold: ${threshold}) \u2014 ${verdict.reason}`
+      actual: `${output.value ? "Pass" : "Fail"} \u2014 ${output.explanation}`
     };
   } catch (err) {
     _log.warn(`LLM judge failed for "${criteria}": ${err.message}`);
     return {
-      assertion: `llm_judge: "${criteria}"`,
-      pass: false,
+      assertion,
+      pass: !_failClosed,
+      skipped: true,
       expected: criteria,
       actual: `FAILED \u2014 LLM judge unavailable: ${err.message}`
     };