@botpress/evals 1.19.0-beta.14 → 1.19.0-beta.16

package/dist/graders/index.js.map

@@ -1,7 +1,7 @@
 {
   "version": 3,
   "sources": ["../../src/graders/match.ts", "../../src/graders/llm.ts", "../../src/types.ts", "../../src/graders/response.ts", "../../src/graders/tools.ts", "../../src/graders/state.ts", "../../src/graders/workflow.ts", "../../src/graders/timing.ts", "../../src/graders/outcome.ts"],
-  "sourcesContent": ["/**\n * Shared match utilities for assertion grading.\n * Used by tool, state, and table graders.\n */\n\nimport type { MatchOperator } from '../types'\n\n/**\n * Evaluate a single match operator against a value.\n */\nexport function matchValue(operator: MatchOperator, actual: unknown): boolean {\n  // Shorthand: string means exact match\n  if (typeof operator === 'string') {\n    return String(actual) === operator\n  }\n\n  if ('equals' in operator) {\n    return actual === operator.equals\n  }\n\n  if ('contains' in operator) {\n    return String(actual).toLowerCase().includes(operator.contains.toLowerCase())\n  }\n\n  if ('not_contains' in operator) {\n    return !String(actual).toLowerCase().includes(operator.not_contains.toLowerCase())\n  }\n\n  if ('matches' in operator) {\n    return new RegExp(operator.matches, 'i').test(String(actual))\n  }\n\n  if ('in' in operator) {\n    return operator.in.includes(actual)\n  }\n\n  if ('exists' in operator) {\n    return operator.exists ? actual !== undefined && actual !== null : actual === undefined || actual === null\n  }\n\n  if ('gte' in operator) {\n    return Number(actual) >= operator.gte\n  }\n\n  if ('lte' in operator) {\n    return Number(actual) <= operator.lte\n  }\n\n  return false\n}\n\nexport function operatorToString(operator: MatchOperator): string {\n  if (typeof operator === 'string') return `equals \"${operator}\"`\n  if ('equals' in operator) return `equals ${JSON.stringify(operator.equals)}`\n  if ('contains' in operator) return `contains \"${operator.contains}\"`\n  if ('not_contains' in operator) return `not_contains \"${operator.not_contains}\"`\n  if ('matches' in operator) return `matches /${operator.matches}/`\n  if ('in' in operator) return `in [${operator.in.map((v) => JSON.stringify(v)).join(', ')}]`\n  if ('exists' in operator) return operator.exists ? 'exists' : 'does not exist'\n  if ('gte' in operator) return `>= ${operator.gte}`\n  if ('lte' in operator) return `<= ${operator.lte}`\n  return JSON.stringify(operator)\n}\n", "/**\n * LLM judge grader using @botpress/zai.\n * `zai.check` owns the judge prompt, output parsing, and retries on invalid model output.\n */\n\nimport type { Client } from '@botpress/client'\nimport { Cognitive } from '@botpress/cognitive'\nimport { Zai } from '@botpress/zai'\nimport type { EvalLogger, GraderResult } from '../types'\nimport { defaultLogger } from '../types'\n\nconst MODEL_ALIASES = ['fast', 'best']\n\nlet _zai: Zai | null = null\nlet _availableModels: string[] = []\nlet _judgeModel: string = 'fast'\n// When true (default), a judge that can't run is a hard FAIL, so an unavailable judge can't be\n// mistaken for success. Callers may opt into fail-open (a skipped pass) via `failClosed: false`,\n// e.g. local runs without judge credentials. Either way the result is marked `skipped: true`.\nlet _failClosed: boolean = true\nlet _log: EvalLogger = defaultLogger\n\n/** @internal Test-only: clears module judge state so tests stay hermetic in the shared bun test process. */\nexport function _resetLLMJudgeForTests(): void {\n  _zai = null\n  _availableModels = []\n  _judgeModel = 'fast'\n  _log = defaultLogger\n  _failClosed = true\n}\n\n/** @internal Test-only: injects a fake Zai client so judge behavior is testable offline. */\nexport function _setLLMJudgeForTests(zai: unknown, options?: { failClosed?: boolean; model?: string }): void {\n  _zai = zai as Zai\n  _availableModels = ['fast']\n  _judgeModel = options?.model ?? 'fast'\n  _failClosed = options?.failClosed ?? true\n}\n\nfunction _isModelValid(model: string): boolean {\n  if (MODEL_ALIASES.includes(model)) return true\n  if (_availableModels.length === 0) return false\n\n  return _availableModels.includes(model)\n}\n\n/** Initialize the LLM judge with an authenticated Botpress client. */\nexport async function initLLMJudge(\n  client: Client,\n  options?: { model?: string; logger?: EvalLogger; failClosed?: boolean }\n) {\n  const log = options?.logger ?? defaultLogger\n  _log = log\n  _failClosed = options?.failClosed ?? true\n\n  const cognitive = new Cognitive({ client, __experimental_beta: true })\n\n  try {\n    _availableModels = Array.from((await cognitive.fetchRemoteModels()).keys())\n  } catch (err) {\n    log.warn(`Failed to fetch available models: ${(err as Error).message}. LLM judge will be unavailable.`)\n    _availableModels = []\n  }\n\n  _judgeModel = options?.model ?? 'fast'\n\n  if (_availableModels.length === 0) {\n    log.warn('No available Cognitive models: cannot validate judge model.')\n  } else if (!_isModelValid(_judgeModel)) {\n    log.warn(`Configured LLM judge model \"${_judgeModel}\" is invalid. Run \"adk models\" to list available models.`)\n  }\n\n  _zai = new Zai({ client: cognitive, modelId: _judgeModel as ConstructorParameters<typeof Zai>[0]['modelId'] })\n}\n\nexport async function gradeLLMJudge(\n  botResponse: string,\n  criteria: string,\n  context: {\n    userMessage: string\n    /** @deprecated No-op: the judge is boolean (zai.check); there is no score threshold. */\n    passThreshold?: number\n  }\n): Promise<GraderResult> {\n  const assertion = `llm_judge: \"${criteria}\"`\n\n  try {\n    if (!_zai) {\n      _log.warn(`LLM judge unavailable for \"${criteria}\": no credentials configured (initLLMJudge not called)`)\n      return {\n        assertion,\n        pass: !_failClosed,\n        skipped: true,\n        expected: criteria,\n        actual: 'FAILED \u2014 LLM judge unavailable: no credentials configured',\n      }\n    }\n\n    if (!_isModelValid(_judgeModel)) {\n      return {\n        assertion,\n        pass: false,\n        expected: criteria,\n        actual: 'FAILED \u2014 invalid LLM judge model',\n      }\n    }\n\n    const { output } = await _zai.check({ userMessage: context.userMessage, botResponse }, criteria).result()\n\n    return {\n      assertion,\n      pass: output.value,\n      expected: criteria,\n      actual: `${output.value ? 'Pass' : 'Fail'} \u2014 ${output.explanation}`,\n    }\n  } catch (err) {\n    _log.warn(`LLM judge failed for \"${criteria}\": ${(err as Error).message}`)\n    return {\n      assertion,\n      pass: !_failClosed,\n      skipped: true,\n      expected: criteria,\n      actual: `FAILED \u2014 LLM judge unavailable: ${(err as Error).message}`,\n    }\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\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 * Response assertion graders.\n * Checks bot response text against expected assertions.\n */\n\nimport type { ResponseAssertion, GraderResult } from '../types'\nimport { gradeLLMJudge } from './llm'\n\nexport async function gradeResponse(\n  botResponse: string,\n  assertions: ResponseAssertion[],\n  context: { userMessage: string; judgePassThreshold?: number }\n): Promise<GraderResult[]> {\n  const results: GraderResult[] = []\n\n  for (const assertion of assertions) {\n    if ('contains' in assertion) {\n      const pass = botResponse.toLowerCase().includes(assertion.contains.toLowerCase())\n      results.push({\n        assertion: `contains \"${assertion.contains}\"`,\n        pass,\n        expected: `Response contains \"${assertion.contains}\"`,\n        actual: pass ? `Found in response` : `Not found in response`,\n      })\n      continue\n    }\n\n    if ('not_contains' in assertion) {\n      const pass = !botResponse.toLowerCase().includes(assertion.not_contains.toLowerCase())\n      results.push({\n        assertion: `not_contains \"${assertion.not_contains}\"`,\n        pass,\n        expected: `Response does not contain \"${assertion.not_contains}\"`,\n        actual: pass ? `Not found in response` : `Found in response`,\n      })\n      continue\n    }\n\n    if ('matches' in assertion) {\n      const regex = new RegExp(assertion.matches, 'i')\n      const pass = regex.test(botResponse)\n      results.push({\n        assertion: `matches ${assertion.matches}`,\n        pass,\n        expected: `Response matches /${assertion.matches}/`,\n        actual: pass ? `Matched` : `No match`,\n      })\n      continue\n    }\n\n    if ('llm_judge' in assertion) {\n      const result = await gradeLLMJudge(botResponse, assertion.llm_judge, {\n        userMessage: context.userMessage,\n        ...(context.judgePassThreshold !== undefined ? { passThreshold: context.judgePassThreshold } : {}),\n      })\n      results.push(result)\n      continue\n    }\n\n    // TODO: uncomment this once we support it\n    // if ('similar_to' in assertion) {\n    //   results.push({\n    //     assertion: `similar_to: \"${assertion.similar_to}\"`,\n    //     pass: true,\n    //     expected: assertion.similar_to,\n    //     actual: 'SKIPPED \u2014 similar_to not yet implemented',\n    //   })\n    //   continue\n    // }\n\n    results.push({\n      assertion: 'unknown',\n      pass: false,\n      expected: 'known assertion type',\n      actual: `Unknown assertion: ${JSON.stringify(assertion)}`,\n    })\n  }\n\n  return results\n}\n", "/**\n * Tool assertion graders.\n * Checks tool calls from traces against expected assertions.\n */\n\nimport type { ToolAssertion, ToolCall, GraderResult } from '../types'\nimport { matchValue, operatorToString } from './match'\n\nexport function gradeTools(toolCalls: ToolCall[], assertions: ToolAssertion[]): GraderResult[] {\n  return assertions.map((assertion) => {\n    // --- called ---\n    if ('called' in assertion && !('not_called' in assertion) && !('call_order' in assertion)) {\n      const matches = toolCalls.filter((tc) => tc.name === assertion.called)\n      const wasCalled = matches.length > 0\n\n      if (!wasCalled) {\n        return {\n          assertion: `tool called: ${assertion.called}`,\n          pass: false,\n          expected: `${assertion.called} was called`,\n          actual: `Not called. Tools called: [${toolCalls.map((tc) => tc.name).join(', ') || 'none'}]`,\n        }\n      }\n\n      // Check params if specified\n      if (assertion.params) {\n        const paramResults: { key: string; pass: boolean; detail: string }[] = []\n\n        for (const [key, operator] of Object.entries(assertion.params)) {\n          // Check across all matching calls \u2014 pass if any call matches\n          const anyMatch = matches.some((tc) => matchValue(operator, tc.input[key]))\n          paramResults.push({\n            key,\n            pass: anyMatch,\n            detail: anyMatch\n              ? `matched`\n              : `expected ${key} ${operatorToString(operator)}, got ${JSON.stringify(matches.map((tc) => tc.input[key]))}`,\n          })\n        }\n\n        const allParamsPass = paramResults.every((p) => p.pass)\n        const failedParams = paramResults.filter((p) => !p.pass)\n\n        return {\n          assertion: `tool called: ${assertion.called} with params`,\n          pass: allParamsPass,\n          expected: `${assertion.called} called with ${Object.entries(assertion.params)\n            .map(([k, v]) => `${k} ${operatorToString(v)}`)\n            .join(', ')}`,\n          actual: allParamsPass ? `Matched` : failedParams.map((p) => p.detail).join('; '),\n        }\n      }\n\n      return {\n        assertion: `tool called: ${assertion.called}`,\n        pass: true,\n        expected: `${assertion.called} was called`,\n        actual: `Called ${matches.length} time(s)`,\n      }\n    }\n\n    // --- not_called ---\n    if ('not_called' in assertion) {\n      const wasCalled = toolCalls.some((tc) => tc.name === assertion.not_called)\n      return {\n        assertion: `tool not_called: ${assertion.not_called}`,\n        pass: !wasCalled,\n        expected: `${assertion.not_called} was NOT called`,\n        actual: wasCalled ? `Was called` : `Not called`,\n      }\n    }\n\n    // --- call_order ---\n    if ('call_order' in assertion) {\n      const calledNames = toolCalls.map((tc) => tc.name)\n      const expectedOrder = assertion.call_order as string[]\n\n      // Check that the expected sequence appears in order (not necessarily contiguous)\n      let cursor = 0\n      for (const name of calledNames) {\n        if (cursor < expectedOrder.length && name === expectedOrder[cursor]) {\n          cursor++\n        }\n      }\n      const inOrder = cursor === expectedOrder.length\n\n      return {\n        assertion: `call_order: [${expectedOrder.join(' \u2192 ')}]`,\n        pass: inOrder,\n        expected: `Tools called in order: [${expectedOrder.join(' \u2192 ')}]`,\n        actual: `Actual order: [${calledNames.join(' \u2192 ') || 'none'}]`,\n      }\n    }\n\n    return {\n      assertion: 'unknown tool assertion',\n      pass: false,\n      expected: 'known assertion type',\n      actual: `Unknown: ${JSON.stringify(assertion)}`,\n    }\n  })\n}\n", "/**\n * State assertion graders.\n * Pure functions that grade state assertions against StateMutation[] from the transformer.\n * No API calls, no retries, no BpClient.\n */\n\nimport type { StateMutation, StateAssertion, GraderResult } from '../types'\n\n/**\n * Parse a state path like \"conversation.lastCity\" into its components.\n *   bot.*            -> type: \"bot\",          field: \"*\"\n *   user.*           -> type: \"user\",         field: \"*\"\n *   conversation.*   -> type: \"conversation\", field: \"*\"\n */\nfunction parseStatePath(path: string): { type: string; field: string } {\n  const dot = path.indexOf('.')\n  if (dot === -1) {\n    throw new Error(`Invalid state path \"${path}\" \u2014 expected \"type.field\" format`)\n  }\n\n  const type = path.slice(0, dot)\n  const field = path.slice(dot + 1)\n\n  if (type !== 'bot' && type !== 'user' && type !== 'conversation') {\n    throw new Error(`Unknown state type \"${type}\" in path \"${path}\" \u2014 expected bot, user, or conversation`)\n  }\n\n  return { type, field }\n}\n\n/**\n * Grade state assertions using StateMutation[] from the transformer.\n *\n * - `equals`: find the LAST mutation matching the state type, extract field from `current`, deep-compare\n * - `changed: true`: check if the field appears in ANY mutation's `changedKeys` for that state type\n * - `changed: false`: check that the field appears in NO mutation's `changedKeys`\n * - Swapped state: when `swappedToFile` is true, current/previous are undefined \u2014 report a clear message\n */\nexport function gradeState(mutations: StateMutation[], assertions: StateAssertion[]): GraderResult[] {\n  const results: GraderResult[] = []\n\n  for (const assertion of assertions) {\n    const parsed = parseStatePath(assertion.path)\n    const matchingMutations = mutations.filter((m) => m.type === parsed.type)\n\n    if (assertion.equals !== undefined) {\n      if (matchingMutations.length === 0) {\n        results.push({\n          assertion: `state: ${assertion.path} equals`,\n          pass: false,\n          expected: JSON.stringify(assertion.equals),\n          actual:\n            `No state.save span found for \"${parsed.type}\" \u2014 the bot did not mutate this state during the turn. ` +\n            'If this state was seeded via setup.state, assert on the tool/action that reads it, or use a conversation turn that triggers a state write.',\n        })\n        continue\n      }\n\n      const lastMutation = matchingMutations[matchingMutations.length - 1]!\n\n      if (lastMutation.swappedToFile) {\n        results.push({\n          assertion: `state: ${assertion.path} equals`,\n          pass: false,\n          expected: JSON.stringify(assertion.equals),\n          actual: 'State was too large for trace-based assertion (swapped to file)',\n        })\n        continue\n      }\n\n      const currentState = lastMutation.current as Record<string, unknown> | undefined\n      const actualValue = currentState ? currentState[parsed.field] : undefined\n      const pass = deepEqual(actualValue, assertion.equals)\n\n      results.push({\n        assertion: `state: ${assertion.path} equals`,\n        pass,\n        expected: JSON.stringify(assertion.equals),\n        actual: JSON.stringify(actualValue),\n      })\n    }\n\n    if (assertion.changed !== undefined) {\n      if (matchingMutations.length === 0) {\n        results.push({\n          assertion: `state: ${assertion.path} ${assertion.changed ? 'changed' : 'unchanged'}`,\n          pass: !assertion.changed,\n          expected: assertion.changed\n            ? `Field \"${parsed.field}\" changed in a \"${parsed.type}\" state mutation`\n            : `Field \"${parsed.field}\" unchanged`,\n          actual: 'No state mutation found for type \"' + parsed.type + '\"',\n        })\n        continue\n      }\n\n      // Check for swapped state in any matching mutation\n      const hasSwapped = matchingMutations.some((m) => m.swappedToFile)\n      if (hasSwapped) {\n        results.push({\n          assertion: `state: ${assertion.path} ${assertion.changed ? 'changed' : 'unchanged'}`,\n          pass: false,\n          expected: assertion.changed\n            ? `Field \"${parsed.field}\" changed in a \"${parsed.type}\" state mutation`\n            : `Field \"${parsed.field}\" unchanged`,\n          actual: 'State was too large for trace-based assertion (swapped to file)',\n        })\n        continue\n      }\n\n      const fieldChanged = matchingMutations.some((m) => m.changedKeys.includes(parsed.field))\n      const pass = assertion.changed ? fieldChanged : !fieldChanged\n\n      results.push({\n        assertion: `state: ${assertion.path} ${assertion.changed ? 'changed' : 'unchanged'}`,\n        pass,\n        expected: assertion.changed\n          ? `Field \"${parsed.field}\" changed in a \"${parsed.type}\" state mutation`\n          : `Field \"${parsed.field}\" unchanged`,\n        actual: fieldChanged ? `Field \"${parsed.field}\" was changed` : `Field \"${parsed.field}\" was not changed`,\n      })\n    }\n  }\n\n  return results\n}\n\nfunction deepEqual(a: unknown, b: unknown): boolean {\n  if (a === b) return true\n  if (a === null || b === null) return false\n  if (typeof a !== typeof b) return false\n\n  if (Array.isArray(a) && Array.isArray(b)) {\n    if (a.length !== b.length) return false\n    return a.every((v, i) => deepEqual(v, b[i]))\n  }\n\n  if (Array.isArray(a) !== Array.isArray(b)) return false\n\n  if (typeof a === 'object' && typeof b === 'object') {\n    const keysA = Object.keys(a as object)\n    const keysB = Object.keys(b as object)\n    if (keysA.length !== keysB.length) return false\n    return keysA.every((k) => deepEqual((a as Record<string, unknown>)[k], (b as Record<string, unknown>)[k]))\n  }\n\n  return false\n}\n", "/**\n * Workflow assertion graders.\n * Checks workflow execution via transformer-produced WorkflowSpan data.\n */\n\nimport type { WorkflowAssertion, WorkflowSpan, GraderResult } from '../types'\n\nexport function gradeWorkflows(workflows: WorkflowSpan[], assertions: WorkflowAssertion[]): GraderResult[] {\n  const results: GraderResult[] = []\n\n  for (const assertion of assertions) {\n    const matching = workflows.filter((wf) => wf.name === assertion.name)\n\n    // entered \u2014 was the workflow triggered at all?\n    if (assertion.entered !== undefined) {\n      const wasEntered = matching.length > 0\n      const pass = assertion.entered ? wasEntered : !wasEntered\n\n      results.push({\n        assertion: `workflow: ${assertion.name} ${assertion.entered ? 'entered' : 'not entered'}`,\n        pass,\n        expected: assertion.entered\n          ? `Workflow \"${assertion.name}\" was entered`\n          : `Workflow \"${assertion.name}\" was not entered`,\n        actual: wasEntered ? `Found ${matching.length} workflow span(s)` : `No workflow spans found`,\n      })\n    }\n\n    // completed \u2014 did the workflow complete?\n    if (assertion.completed !== undefined) {\n      const didComplete = matching.some(\n        (wf) => wf.statusFinal === 'completed' || wf.status === 'ok' || wf.status === 'error'\n      )\n      const pass = assertion.completed ? didComplete : !didComplete\n\n      results.push({\n        assertion: `workflow: ${assertion.name} ${assertion.completed ? 'completed' : 'not completed'}`,\n        pass,\n        expected: assertion.completed\n          ? `Workflow \"${assertion.name}\" completed`\n          : `Workflow \"${assertion.name}\" did not complete`,\n        actual: didComplete ? `Completed` : `Not completed`,\n      })\n    }\n  }\n\n  return results\n}\n", "/**\n * Timing assertion graders.\n * Checks bot response duration against expected thresholds.\n */\n\nimport type { TimingAssertion, GraderResult } from '../types'\nimport { matchValue, operatorToString } from './match'\n\nexport function gradeTiming(botDuration: number, assertions: TimingAssertion[]): GraderResult[] {\n  const results: GraderResult[] = []\n\n  for (const assertion of assertions) {\n    const pass = matchValue(assertion.response_time, botDuration)\n    const expected = operatorToString(assertion.response_time)\n\n    results.push({\n      assertion: `response_time ${expected}`,\n      pass,\n      expected: `Response time ${expected}`,\n      actual: `${botDuration}ms`,\n    })\n  }\n\n  return results\n}\n", "/**\n * Outcome assertion orchestrator.\n * Runs state and workflow graders after all conversation turns complete.\n */\n\nimport type { EvalDefinition, GraderResult, StateMutation, WorkflowSpan } from '../types'\nimport { gradeState } from './state'\nimport { gradeWorkflows } from './workflow'\n\n/**\n * Grade all outcome assertions after the conversation completes.\n * Takes transformer output (StateMutation[], WorkflowSpan[]) \u2014 no raw spans, no API calls.\n */\nexport function gradeOutcome(\n  mutations: StateMutation[],\n  evalDef: EvalDefinition,\n  workflowSpans: WorkflowSpan[]\n): GraderResult[] {\n  const outcome = evalDef.outcome\n  if (!outcome) return []\n\n  const results: GraderResult[] = []\n\n  if (outcome.state && outcome.state.length > 0) {\n    const stateResults = gradeState(mutations, outcome.state)\n    results.push(...stateResults)\n  }\n\n  if (outcome.workflow && outcome.workflow.length > 0) {\n    const workflowResults = gradeWorkflows(workflowSpans, outcome.workflow)\n    results.push(...workflowResults)\n  }\n\n  return results\n}\n"],
-  "mappings": ";AAUO,SAAS,WAAW,UAAyB,QAA0B;AAE5E,MAAI,OAAO,aAAa,UAAU;AAChC,WAAO,OAAO,MAAM,MAAM;AAAA,EAC5B;AAEA,MAAI,YAAY,UAAU;AACxB,WAAO,WAAW,SAAS;AAAA,EAC7B;AAEA,MAAI,cAAc,UAAU;AAC1B,WAAO,OAAO,MAAM,EAAE,YAAY,EAAE,SAAS,SAAS,SAAS,YAAY,CAAC;AAAA,EAC9E;AAEA,MAAI,kBAAkB,UAAU;AAC9B,WAAO,CAAC,OAAO,MAAM,EAAE,YAAY,EAAE,SAAS,SAAS,aAAa,YAAY,CAAC;AAAA,EACnF;AAEA,MAAI,aAAa,UAAU;AACzB,WAAO,IAAI,OAAO,SAAS,SAAS,GAAG,EAAE,KAAK,OAAO,MAAM,CAAC;AAAA,EAC9D;AAEA,MAAI,QAAQ,UAAU;AACpB,WAAO,SAAS,GAAG,SAAS,MAAM;AAAA,EACpC;AAEA,MAAI,YAAY,UAAU;AACxB,WAAO,SAAS,SAAS,WAAW,UAAa,WAAW,OAAO,WAAW,UAAa,WAAW;AAAA,EACxG;AAEA,MAAI,SAAS,UAAU;AACrB,WAAO,OAAO,MAAM,KAAK,SAAS;AAAA,EACpC;AAEA,MAAI,SAAS,UAAU;AACrB,WAAO,OAAO,MAAM,KAAK,SAAS;AAAA,EACpC;AAEA,SAAO;AACT;AAEO,SAAS,iBAAiB,UAAiC;AAChE,MAAI,OAAO,aAAa,SAAU,QAAO,WAAW,QAAQ;AAC5D,MAAI,YAAY,SAAU,QAAO,UAAU,KAAK,UAAU,SAAS,MAAM,CAAC;AAC1E,MAAI,cAAc,SAAU,QAAO,aAAa,SAAS,QAAQ;AACjE,MAAI,kBAAkB,SAAU,QAAO,iBAAiB,SAAS,YAAY;AAC7E,MAAI,aAAa,SAAU,QAAO,YAAY,SAAS,OAAO;AAC9D,MAAI,QAAQ,SAAU,QAAO,OAAO,SAAS,GAAG,IAAI,CAAC,MAAM,KAAK,UAAU,CAAC,CAAC,EAAE,KAAK,IAAI,CAAC;AACxF,MAAI,YAAY,SAAU,QAAO,SAAS,SAAS,WAAW;AAC9D,MAAI,SAAS,SAAU,QAAO,MAAM,SAAS,GAAG;AAChD,MAAI,SAAS,SAAU,QAAO,MAAM,SAAS,GAAG;AAChD,SAAO,KAAK,UAAU,QAAQ;AAChC;;;ACxDA,SAAS,iBAAiB;AAC1B,SAAS,WAAW;;;AC0Jb,IAAM,gBAA4B;;;ADtJzC,IAAM,gBAAgB,CAAC,QAAQ,MAAM;AAErC,IAAI,OAAmB;AACvB,IAAI,mBAA6B,CAAC;AAClC,IAAI,cAAsB;AAI1B,IAAI,cAAuB;AAC3B,IAAI,OAAmB;AAmBvB,SAAS,cAAc,OAAwB;AAC7C,MAAI,cAAc,SAAS,KAAK,EAAG,QAAO;AAC1C,MAAI,iBAAiB,WAAW,EAAG,QAAO;AAE1C,SAAO,iBAAiB,SAAS,KAAK;AACxC;AAGA,eAAsB,aACpB,QACA,SACA;AACA,QAAM,MAAM,SAAS,UAAU;AAC/B,SAAO;AACP,gBAAc,SAAS,cAAc;AAErC,QAAM,YAAY,IAAI,UAAU,EAAE,QAAQ,qBAAqB,KAAK,CAAC;AAErE,MAAI;AACF,uBAAmB,MAAM,MAAM,MAAM,UAAU,kBAAkB,GAAG,KAAK,CAAC;AAAA,EAC5E,SAAS,KAAK;AACZ,QAAI,KAAK,qCAAsC,IAAc,OAAO,kCAAkC;AACtG,uBAAmB,CAAC;AAAA,EACtB;AAEA,gBAAc,SAAS,SAAS;AAEhC,MAAI,iBAAiB,WAAW,GAAG;AACjC,QAAI,KAAK,6DAA6D;AAAA,EACxE,WAAW,CAAC,cAAc,WAAW,GAAG;AACtC,QAAI,KAAK,+BAA+B,WAAW,0DAA0D;AAAA,EAC/G;AAEA,SAAO,IAAI,IAAI,EAAE,QAAQ,WAAW,SAAS,YAA+D,CAAC;AAC/G;AAEA,eAAsB,cACpB,aACA,UACA,SAKuB;AACvB,QAAM,YAAY,eAAe,QAAQ;AAEzC,MAAI;AACF,QAAI,CAAC,MAAM;AACT,WAAK,KAAK,8BAA8B,QAAQ,wDAAwD;AACxG,aAAO;AAAA,QACL;AAAA,QACA,MAAM,CAAC;AAAA,QACP,SAAS;AAAA,QACT,UAAU;AAAA,QACV,QAAQ;AAAA,MACV;AAAA,IACF;AAEA,QAAI,CAAC,cAAc,WAAW,GAAG;AAC/B,aAAO;AAAA,QACL;AAAA,QACA,MAAM;AAAA,QACN,UAAU;AAAA,QACV,QAAQ;AAAA,MACV;AAAA,IACF;AAEA,UAAM,EAAE,OAAO,IAAI,MAAM,KAAK,MAAM,EAAE,aAAa,QAAQ,aAAa,YAAY,GAAG,QAAQ,EAAE,OAAO;AAExG,WAAO;AAAA,MACL;AAAA,MACA,MAAM,OAAO;AAAA,MACb,UAAU;AAAA,MACV,QAAQ,GAAG,OAAO,QAAQ,SAAS,MAAM,WAAM,OAAO,WAAW;AAAA,IACnE;AAAA,EACF,SAAS,KAAK;AACZ,SAAK,KAAK,yBAAyB,QAAQ,MAAO,IAAc,OAAO,EAAE;AACzE,WAAO;AAAA,MACL;AAAA,MACA,MAAM,CAAC;AAAA,MACP,SAAS;AAAA,MACT,UAAU;AAAA,MACV,QAAQ,wCAAoC,IAAc,OAAO;AAAA,IACnE;AAAA,EACF;AACF;;;AErHA,eAAsB,cACpB,aACA,YACA,SACyB;AACzB,QAAM,UAA0B,CAAC;AAEjC,aAAW,aAAa,YAAY;AAClC,QAAI,cAAc,WAAW;AAC3B,YAAM,OAAO,YAAY,YAAY,EAAE,SAAS,UAAU,SAAS,YAAY,CAAC;AAChF,cAAQ,KAAK;AAAA,QACX,WAAW,aAAa,UAAU,QAAQ;AAAA,QAC1C;AAAA,QACA,UAAU,sBAAsB,UAAU,QAAQ;AAAA,QAClD,QAAQ,OAAO,sBAAsB;AAAA,MACvC,CAAC;AACD;AAAA,IACF;AAEA,QAAI,kBAAkB,WAAW;AAC/B,YAAM,OAAO,CAAC,YAAY,YAAY,EAAE,SAAS,UAAU,aAAa,YAAY,CAAC;AACrF,cAAQ,KAAK;AAAA,QACX,WAAW,iBAAiB,UAAU,YAAY;AAAA,QAClD;AAAA,QACA,UAAU,8BAA8B,UAAU,YAAY;AAAA,QAC9D,QAAQ,OAAO,0BAA0B;AAAA,MAC3C,CAAC;AACD;AAAA,IACF;AAEA,QAAI,aAAa,WAAW;AAC1B,YAAM,QAAQ,IAAI,OAAO,UAAU,SAAS,GAAG;AAC/C,YAAM,OAAO,MAAM,KAAK,WAAW;AACnC,cAAQ,KAAK;AAAA,QACX,WAAW,WAAW,UAAU,OAAO;AAAA,QACvC;AAAA,QACA,UAAU,qBAAqB,UAAU,OAAO;AAAA,QAChD,QAAQ,OAAO,YAAY;AAAA,MAC7B,CAAC;AACD;AAAA,IACF;AAEA,QAAI,eAAe,WAAW;AAC5B,YAAM,SAAS,MAAM,cAAc,aAAa,UAAU,WAAW;AAAA,QACnE,aAAa,QAAQ;AAAA,QACrB,GAAI,QAAQ,uBAAuB,SAAY,EAAE,eAAe,QAAQ,mBAAmB,IAAI,CAAC;AAAA,MAClG,CAAC;AACD,cAAQ,KAAK,MAAM;AACnB;AAAA,IACF;AAaA,YAAQ,KAAK;AAAA,MACX,WAAW;AAAA,MACX,MAAM;AAAA,MACN,UAAU;AAAA,MACV,QAAQ,sBAAsB,KAAK,UAAU,SAAS,CAAC;AAAA,IACzD,CAAC;AAAA,EACH;AAEA,SAAO;AACT;;;ACvEO,SAAS,WAAW,WAAuB,YAA6C;AAC7F,SAAO,WAAW,IAAI,CAAC,cAAc;AAEnC,QAAI,YAAY,aAAa,EAAE,gBAAgB,cAAc,EAAE,gBAAgB,YAAY;AACzF,YAAM,UAAU,UAAU,OAAO,CAAC,OAAO,GAAG,SAAS,UAAU,MAAM;AACrE,YAAM,YAAY,QAAQ,SAAS;AAEnC,UAAI,CAAC,WAAW;AACd,eAAO;AAAA,UACL,WAAW,gBAAgB,UAAU,MAAM;AAAA,UAC3C,MAAM;AAAA,UACN,UAAU,GAAG,UAAU,MAAM;AAAA,UAC7B,QAAQ,8BAA8B,UAAU,IAAI,CAAC,OAAO,GAAG,IAAI,EAAE,KAAK,IAAI,KAAK,MAAM;AAAA,QAC3F;AAAA,MACF;AAGA,UAAI,UAAU,QAAQ;AACpB,cAAM,eAAiE,CAAC;AAExE,mBAAW,CAAC,KAAK,QAAQ,KAAK,OAAO,QAAQ,UAAU,MAAM,GAAG;AAE9D,gBAAM,WAAW,QAAQ,KAAK,CAAC,OAAO,WAAW,UAAU,GAAG,MAAM,GAAG,CAAC,CAAC;AACzE,uBAAa,KAAK;AAAA,YAChB;AAAA,YACA,MAAM;AAAA,YACN,QAAQ,WACJ,YACA,YAAY,GAAG,IAAI,iBAAiB,QAAQ,CAAC,SAAS,KAAK,UAAU,QAAQ,IAAI,CAAC,OAAO,GAAG,MAAM,GAAG,CAAC,CAAC,CAAC;AAAA,UAC9G,CAAC;AAAA,QACH;AAEA,cAAM,gBAAgB,aAAa,MAAM,CAAC,MAAM,EAAE,IAAI;AACtD,cAAM,eAAe,aAAa,OAAO,CAAC,MAAM,CAAC,EAAE,IAAI;AAEvD,eAAO;AAAA,UACL,WAAW,gBAAgB,UAAU,MAAM;AAAA,UAC3C,MAAM;AAAA,UACN,UAAU,GAAG,UAAU,MAAM,gBAAgB,OAAO,QAAQ,UAAU,MAAM,EACzE,IAAI,CAAC,CAAC,GAAG,CAAC,MAAM,GAAG,CAAC,IAAI,iBAAiB,CAAC,CAAC,EAAE,EAC7C,KAAK,IAAI,CAAC;AAAA,UACb,QAAQ,gBAAgB,YAAY,aAAa,IAAI,CAAC,MAAM,EAAE,MAAM,EAAE,KAAK,IAAI;AAAA,QACjF;AAAA,MACF;AAEA,aAAO;AAAA,QACL,WAAW,gBAAgB,UAAU,MAAM;AAAA,QAC3C,MAAM;AAAA,QACN,UAAU,GAAG,UAAU,MAAM;AAAA,QAC7B,QAAQ,UAAU,QAAQ,MAAM;AAAA,MAClC;AAAA,IACF;AAGA,QAAI,gBAAgB,WAAW;AAC7B,YAAM,YAAY,UAAU,KAAK,CAAC,OAAO,GAAG,SAAS,UAAU,UAAU;AACzE,aAAO;AAAA,QACL,WAAW,oBAAoB,UAAU,UAAU;AAAA,QACnD,MAAM,CAAC;AAAA,QACP,UAAU,GAAG,UAAU,UAAU;AAAA,QACjC,QAAQ,YAAY,eAAe;AAAA,MACrC;AAAA,IACF;AAGA,QAAI,gBAAgB,WAAW;AAC7B,YAAM,cAAc,UAAU,IAAI,CAAC,OAAO,GAAG,IAAI;AACjD,YAAM,gBAAgB,UAAU;AAGhC,UAAI,SAAS;AACb,iBAAW,QAAQ,aAAa;AAC9B,YAAI,SAAS,cAAc,UAAU,SAAS,cAAc,MAAM,GAAG;AACnE;AAAA,QACF;AAAA,MACF;AACA,YAAM,UAAU,WAAW,cAAc;AAEzC,aAAO;AAAA,QACL,WAAW,gBAAgB,cAAc,KAAK,UAAK,CAAC;AAAA,QACpD,MAAM;AAAA,QACN,UAAU,2BAA2B,cAAc,KAAK,UAAK,CAAC;AAAA,QAC9D,QAAQ,kBAAkB,YAAY,KAAK,UAAK,KAAK,MAAM;AAAA,MAC7D;AAAA,IACF;AAEA,WAAO;AAAA,MACL,WAAW;AAAA,MACX,MAAM;AAAA,MACN,UAAU;AAAA,MACV,QAAQ,YAAY,KAAK,UAAU,SAAS,CAAC;AAAA,IAC/C;AAAA,EACF,CAAC;AACH;;;ACvFA,SAAS,eAAe,MAA+C;AACrE,QAAM,MAAM,KAAK,QAAQ,GAAG;AAC5B,MAAI,QAAQ,IAAI;AACd,UAAM,IAAI,MAAM,uBAAuB,IAAI,uCAAkC;AAAA,EAC/E;AAEA,QAAM,OAAO,KAAK,MAAM,GAAG,GAAG;AAC9B,QAAM,QAAQ,KAAK,MAAM,MAAM,CAAC;AAEhC,MAAI,SAAS,SAAS,SAAS,UAAU,SAAS,gBAAgB;AAChE,UAAM,IAAI,MAAM,uBAAuB,IAAI,cAAc,IAAI,8CAAyC;AAAA,EACxG;AAEA,SAAO,EAAE,MAAM,MAAM;AACvB;AAUO,SAAS,WAAW,WAA4B,YAA8C;AACnG,QAAM,UAA0B,CAAC;AAEjC,aAAW,aAAa,YAAY;AAClC,UAAM,SAAS,eAAe,UAAU,IAAI;AAC5C,UAAM,oBAAoB,UAAU,OAAO,CAAC,MAAM,EAAE,SAAS,OAAO,IAAI;AAExE,QAAI,UAAU,WAAW,QAAW;AAClC,UAAI,kBAAkB,WAAW,GAAG;AAClC,gBAAQ,KAAK;AAAA,UACX,WAAW,UAAU,UAAU,IAAI;AAAA,UACnC,MAAM;AAAA,UACN,UAAU,KAAK,UAAU,UAAU,MAAM;AAAA,UACzC,QACE,iCAAiC,OAAO,IAAI;AAAA,QAEhD,CAAC;AACD;AAAA,MACF;AAEA,YAAM,eAAe,kBAAkB,kBAAkB,SAAS,CAAC;AAEnE,UAAI,aAAa,eAAe;AAC9B,gBAAQ,KAAK;AAAA,UACX,WAAW,UAAU,UAAU,IAAI;AAAA,UACnC,MAAM;AAAA,UACN,UAAU,KAAK,UAAU,UAAU,MAAM;AAAA,UACzC,QAAQ;AAAA,QACV,CAAC;AACD;AAAA,MACF;AAEA,YAAM,eAAe,aAAa;AAClC,YAAM,cAAc,eAAe,aAAa,OAAO,KAAK,IAAI;AAChE,YAAM,OAAO,UAAU,aAAa,UAAU,MAAM;AAEpD,cAAQ,KAAK;AAAA,QACX,WAAW,UAAU,UAAU,IAAI;AAAA,QACnC;AAAA,QACA,UAAU,KAAK,UAAU,UAAU,MAAM;AAAA,QACzC,QAAQ,KAAK,UAAU,WAAW;AAAA,MACpC,CAAC;AAAA,IACH;AAEA,QAAI,UAAU,YAAY,QAAW;AACnC,UAAI,kBAAkB,WAAW,GAAG;AAClC,gBAAQ,KAAK;AAAA,UACX,WAAW,UAAU,UAAU,IAAI,IAAI,UAAU,UAAU,YAAY,WAAW;AAAA,UAClF,MAAM,CAAC,UAAU;AAAA,UACjB,UAAU,UAAU,UAChB,UAAU,OAAO,KAAK,mBAAmB,OAAO,IAAI,qBACpD,UAAU,OAAO,KAAK;AAAA,UAC1B,QAAQ,uCAAuC,OAAO,OAAO;AAAA,QAC/D,CAAC;AACD;AAAA,MACF;AAGA,YAAM,aAAa,kBAAkB,KAAK,CAAC,MAAM,EAAE,aAAa;AAChE,UAAI,YAAY;AACd,gBAAQ,KAAK;AAAA,UACX,WAAW,UAAU,UAAU,IAAI,IAAI,UAAU,UAAU,YAAY,WAAW;AAAA,UAClF,MAAM;AAAA,UACN,UAAU,UAAU,UAChB,UAAU,OAAO,KAAK,mBAAmB,OAAO,IAAI,qBACpD,UAAU,OAAO,KAAK;AAAA,UAC1B,QAAQ;AAAA,QACV,CAAC;AACD;AAAA,MACF;AAEA,YAAM,eAAe,kBAAkB,KAAK,CAAC,MAAM,EAAE,YAAY,SAAS,OAAO,KAAK,CAAC;AACvF,YAAM,OAAO,UAAU,UAAU,eAAe,CAAC;AAEjD,cAAQ,KAAK;AAAA,QACX,WAAW,UAAU,UAAU,IAAI,IAAI,UAAU,UAAU,YAAY,WAAW;AAAA,QAClF;AAAA,QACA,UAAU,UAAU,UAChB,UAAU,OAAO,KAAK,mBAAmB,OAAO,IAAI,qBACpD,UAAU,OAAO,KAAK;AAAA,QAC1B,QAAQ,eAAe,UAAU,OAAO,KAAK,kBAAkB,UAAU,OAAO,KAAK;AAAA,MACvF,CAAC;AAAA,IACH;AAAA,EACF;AAEA,SAAO;AACT;AAEA,SAAS,UAAU,GAAY,GAAqB;AAClD,MAAI,MAAM,EAAG,QAAO;AACpB,MAAI,MAAM,QAAQ,MAAM,KAAM,QAAO;AACrC,MAAI,OAAO,MAAM,OAAO,EAAG,QAAO;AAElC,MAAI,MAAM,QAAQ,CAAC,KAAK,MAAM,QAAQ,CAAC,GAAG;AACxC,QAAI,EAAE,WAAW,EAAE,OAAQ,QAAO;AAClC,WAAO,EAAE,MAAM,CAAC,GAAG,MAAM,UAAU,GAAG,EAAE,CAAC,CAAC,CAAC;AAAA,EAC7C;AAEA,MAAI,MAAM,QAAQ,CAAC,MAAM,MAAM,QAAQ,CAAC,EAAG,QAAO;AAElD,MAAI,OAAO,MAAM,YAAY,OAAO,MAAM,UAAU;AAClD,UAAM,QAAQ,OAAO,KAAK,CAAW;AACrC,UAAM,QAAQ,OAAO,KAAK,CAAW;AACrC,QAAI,MAAM,WAAW,MAAM,OAAQ,QAAO;AAC1C,WAAO,MAAM,MAAM,CAAC,MAAM,UAAW,EAA8B,CAAC,GAAI,EAA8B,CAAC,CAAC,CAAC;AAAA,EAC3G;AAEA,SAAO;AACT;;;AC3IO,SAAS,eAAe,WAA2B,YAAiD;AACzG,QAAM,UAA0B,CAAC;AAEjC,aAAW,aAAa,YAAY;AAClC,UAAM,WAAW,UAAU,OAAO,CAAC,OAAO,GAAG,SAAS,UAAU,IAAI;AAGpE,QAAI,UAAU,YAAY,QAAW;AACnC,YAAM,aAAa,SAAS,SAAS;AACrC,YAAM,OAAO,UAAU,UAAU,aAAa,CAAC;AAE/C,cAAQ,KAAK;AAAA,QACX,WAAW,aAAa,UAAU,IAAI,IAAI,UAAU,UAAU,YAAY,aAAa;AAAA,QACvF;AAAA,QACA,UAAU,UAAU,UAChB,aAAa,UAAU,IAAI,kBAC3B,aAAa,UAAU,IAAI;AAAA,QAC/B,QAAQ,aAAa,SAAS,SAAS,MAAM,sBAAsB;AAAA,MACrE,CAAC;AAAA,IACH;AAGA,QAAI,UAAU,cAAc,QAAW;AACrC,YAAM,cAAc,SAAS;AAAA,QAC3B,CAAC,OAAO,GAAG,gBAAgB,eAAe,GAAG,WAAW,QAAQ,GAAG,WAAW;AAAA,MAChF;AACA,YAAM,OAAO,UAAU,YAAY,cAAc,CAAC;AAElD,cAAQ,KAAK;AAAA,QACX,WAAW,aAAa,UAAU,IAAI,IAAI,UAAU,YAAY,cAAc,eAAe;AAAA,QAC7F;AAAA,QACA,UAAU,UAAU,YAChB,aAAa,UAAU,IAAI,gBAC3B,aAAa,UAAU,IAAI;AAAA,QAC/B,QAAQ,cAAc,cAAc;AAAA,MACtC,CAAC;AAAA,IACH;AAAA,EACF;AAEA,SAAO;AACT;;;ACvCO,SAAS,YAAY,aAAqB,YAA+C;AAC9F,QAAM,UAA0B,CAAC;AAEjC,aAAW,aAAa,YAAY;AAClC,UAAM,OAAO,WAAW,UAAU,eAAe,WAAW;AAC5D,UAAM,WAAW,iBAAiB,UAAU,aAAa;AAEzD,YAAQ,KAAK;AAAA,MACX,WAAW,iBAAiB,QAAQ;AAAA,MACpC;AAAA,MACA,UAAU,iBAAiB,QAAQ;AAAA,MACnC,QAAQ,GAAG,WAAW;AAAA,IACxB,CAAC;AAAA,EACH;AAEA,SAAO;AACT;;;ACXO,SAAS,aACd,WACA,SACA,eACgB;AAChB,QAAM,UAAU,QAAQ;AACxB,MAAI,CAAC,QAAS,QAAO,CAAC;AAEtB,QAAM,UAA0B,CAAC;AAEjC,MAAI,QAAQ,SAAS,QAAQ,MAAM,SAAS,GAAG;AAC7C,UAAM,eAAe,WAAW,WAAW,QAAQ,KAAK;AACxD,YAAQ,KAAK,GAAG,YAAY;AAAA,EAC9B;AAEA,MAAI,QAAQ,YAAY,QAAQ,SAAS,SAAS,GAAG;AACnD,UAAM,kBAAkB,eAAe,eAAe,QAAQ,QAAQ;AACtE,YAAQ,KAAK,GAAG,eAAe;AAAA,EACjC;AAEA,SAAO;AACT;",
+  "sourcesContent": ["/**\n * Shared match utilities for assertion grading.\n * Used by tool, state, and table graders.\n */\n\nimport type { MatchOperator } from '../types'\n\n/**\n * Evaluate a single match operator against a value.\n */\nexport function matchValue(operator: MatchOperator, actual: unknown): boolean {\n  // Shorthand: string means exact match\n  if (typeof operator === 'string') {\n    return String(actual) === operator\n  }\n\n  if ('equals' in operator) {\n    return actual === operator.equals\n  }\n\n  if ('contains' in operator) {\n    return String(actual).toLowerCase().includes(operator.contains.toLowerCase())\n  }\n\n  if ('not_contains' in operator) {\n    return !String(actual).toLowerCase().includes(operator.not_contains.toLowerCase())\n  }\n\n  if ('matches' in operator) {\n    return new RegExp(operator.matches, 'i').test(String(actual))\n  }\n\n  if ('in' in operator) {\n    return operator.in.includes(actual)\n  }\n\n  if ('exists' in operator) {\n    return operator.exists ? actual !== undefined && actual !== null : actual === undefined || actual === null\n  }\n\n  if ('gte' in operator) {\n    return Number(actual) >= operator.gte\n  }\n\n  if ('lte' in operator) {\n    return Number(actual) <= operator.lte\n  }\n\n  return false\n}\n\nexport function operatorToString(operator: MatchOperator): string {\n  if (typeof operator === 'string') return `equals \"${operator}\"`\n  if ('equals' in operator) return `equals ${JSON.stringify(operator.equals)}`\n  if ('contains' in operator) return `contains \"${operator.contains}\"`\n  if ('not_contains' in operator) return `not_contains \"${operator.not_contains}\"`\n  if ('matches' in operator) return `matches /${operator.matches}/`\n  if ('in' in operator) return `in [${operator.in.map((v) => JSON.stringify(v)).join(', ')}]`\n  if ('exists' in operator) return operator.exists ? 'exists' : 'does not exist'\n  if ('gte' in operator) return `>= ${operator.gte}`\n  if ('lte' in operator) return `<= ${operator.lte}`\n  return JSON.stringify(operator)\n}\n", "/**\n * LLM judge grader using @botpress/cognitive.\n * Replaces the OpenAI-based implementation from the standalone project.\n */\n\nimport { Client } from '@botpress/client'\nimport { Cognitive } from '@botpress/cognitive'\nimport type { EvalLogger, GraderResult } from '../types'\nimport { defaultLogger } from '../types'\n\nconst DEFAULT_PASS_THRESHOLD = 3\n\nfunction buildJudgeSystemPrompt(passThreshold: number): string {\n  return `You are an evaluation judge for a chatbot. You will be given:\n- The user's message\n- The bot's response\n- Grading criteria\n\nYour job is to determine if the bot's response meets the criteria.\n\nRespond with a JSON object:\n{\n  \"pass\": true/false,\n  \"score\": 1-5,\n  \"reason\": \"brief explanation\"\n}\n\nScoring guide:\n- 5: Fully meets criteria, excellent response\n- 4: Mostly meets criteria, minor issues\n- 3: Partially meets criteria, notable gaps\n- 2: Barely meets criteria, significant issues\n- 1: Does not meet criteria\n\nA score of ${passThreshold} or above is a pass.`\n}\n\nconst MODEL_ALIASES = ['fast', 'best']\n\nlet _cognitive: Cognitive | null = null\nlet _availableModels: string[] = []\nlet _judgeModel: string = 'fast'\nlet _log: EvalLogger = defaultLogger\n\n/** @internal Test-only: clears module judge state so tests stay hermetic in the shared bun test process. */\nexport function _resetLLMJudgeForTests(): void {\n  _cognitive = null\n  _availableModels = []\n  _judgeModel = 'fast'\n  _log = defaultLogger\n}\n\nfunction _isModelValid(model: string): boolean {\n  if (MODEL_ALIASES.includes(model)) return true\n  if (_availableModels.length === 0) return false\n\n  return _availableModels.includes(model)\n}\n\n/**\n * Initialize the LLM judge with explicit credentials.\n * Call this before running evals so the judge doesn't need env vars.\n */\nexport async function initLLMJudge(\n  credentials: { token: string; apiUrl: string; botId: string; workspaceId?: string },\n  options?: { model?: string; logger?: EvalLogger }\n) {\n  const log = options?.logger ?? defaultLogger\n  _log = log\n\n  const client = new Client({\n    token: credentials.token,\n    apiUrl: credentials.apiUrl,\n    botId: credentials.botId,\n    ...(credentials.workspaceId !== undefined ? { workspaceId: credentials.workspaceId } : {}),\n  })\n  _cognitive = new Cognitive({ client, __experimental_beta: true })\n\n  try {\n    _availableModels = Array.from((await _cognitive.fetchRemoteModels()).keys())\n  } catch (err) {\n    log.warn(`Failed to fetch available models: ${(err as Error).message}. LLM judge will be unavailable.`)\n    _availableModels = []\n  }\n\n  _judgeModel = options?.model ?? 'fast'\n\n  if (_availableModels.length === 0) {\n    log.warn('No available Cognitive models: cannot validate judge model.')\n  } else if (!_isModelValid(_judgeModel)) {\n    log.warn(`Configured LLM judge model \"${_judgeModel}\" is invalid. Run \"adk models\" to list available models.`)\n  }\n}\n\nexport async function gradeLLMJudge(\n  botResponse: string,\n  criteria: string,\n  context: { userMessage: string; passThreshold?: number }\n): Promise<GraderResult> {\n  const threshold = Math.max(1, Math.min(5, context.passThreshold ?? DEFAULT_PASS_THRESHOLD))\n\n  try {\n    // A judge that can't run must NOT silently pass the assertion \u2014 a flaky\n    // LLM call or missing credentials would turn real assertions green.\n    // Unavailability is a visible failure with a clear reason.\n    if (!_cognitive) {\n      _log.warn(`LLM judge unavailable for \"${criteria}\": no credentials configured (initLLMJudge not called)`)\n      return {\n        assertion: `llm_judge: \"${criteria}\"`,\n        pass: false,\n        expected: criteria,\n        actual: 'FAILED \u2014 LLM judge unavailable: no credentials configured',\n      }\n    }\n\n    if (!_isModelValid(_judgeModel)) {\n      return {\n        assertion: `llm_judge: \"${criteria}\"`,\n        pass: false,\n        expected: criteria,\n        actual: 'FAILED \u2014 invalid LLM judge model',\n      }\n    }\n\n    type CognitiveModel = Exclude<Parameters<Cognitive['generateContent']>[0]['model'], undefined>\n\n    const { output } = await _cognitive.generateContent({\n      model: _judgeModel as CognitiveModel,\n      temperature: 0,\n      responseFormat: 'json_object',\n      systemPrompt: buildJudgeSystemPrompt(threshold),\n      messages: [\n        {\n          role: 'user',\n          content: `User message: ${context.userMessage}\\n\\nBot response: ${botResponse}\\n\\nCriteria: ${criteria}`,\n        },\n      ],\n    })\n\n    const rawContent = output?.choices?.[0]?.content\n    const content = typeof rawContent === 'string' ? rawContent : null\n    if (!content) {\n      _log.warn(`LLM judge returned an empty response for \"${criteria}\"`)\n      return {\n        assertion: `llm_judge: \"${criteria}\"`,\n        pass: false,\n        expected: criteria,\n        actual: 'FAILED \u2014 LLM judge returned empty response',\n      }\n    }\n\n    const verdict = JSON.parse(content) as {\n      pass: boolean\n      score: number\n      reason: string\n    }\n\n    return {\n      assertion: `llm_judge: \"${criteria}\"`,\n      pass: verdict.score >= threshold,\n      expected: criteria,\n      actual: `Score ${verdict.score}/5 (threshold: ${threshold}) \u2014 ${verdict.reason}`,\n    }\n  } catch (err) {\n    _log.warn(`LLM judge failed for \"${criteria}\": ${(err as Error).message}`)\n    return {\n      assertion: `llm_judge: \"${criteria}\"`,\n      pass: false,\n      expected: criteria,\n      actual: `FAILED \u2014 LLM judge unavailable: ${(err as Error).message}`,\n    }\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 * Response assertion graders.\n * Checks bot response text against expected assertions.\n */\n\nimport type { ResponseAssertion, GraderResult } from '../types'\nimport { gradeLLMJudge } from './llm'\n\nexport async function gradeResponse(\n  botResponse: string,\n  assertions: ResponseAssertion[],\n  context: { userMessage: string; judgePassThreshold?: number }\n): Promise<GraderResult[]> {\n  const results: GraderResult[] = []\n\n  for (const assertion of assertions) {\n    if ('contains' in assertion) {\n      const pass = botResponse.toLowerCase().includes(assertion.contains.toLowerCase())\n      results.push({\n        assertion: `contains \"${assertion.contains}\"`,\n        pass,\n        expected: `Response contains \"${assertion.contains}\"`,\n        actual: pass ? `Found in response` : `Not found in response`,\n      })\n      continue\n    }\n\n    if ('not_contains' in assertion) {\n      const pass = !botResponse.toLowerCase().includes(assertion.not_contains.toLowerCase())\n      results.push({\n        assertion: `not_contains \"${assertion.not_contains}\"`,\n        pass,\n        expected: `Response does not contain \"${assertion.not_contains}\"`,\n        actual: pass ? `Not found in response` : `Found in response`,\n      })\n      continue\n    }\n\n    if ('matches' in assertion) {\n      const regex = new RegExp(assertion.matches, 'i')\n      const pass = regex.test(botResponse)\n      results.push({\n        assertion: `matches ${assertion.matches}`,\n        pass,\n        expected: `Response matches /${assertion.matches}/`,\n        actual: pass ? `Matched` : `No match`,\n      })\n      continue\n    }\n\n    if ('llm_judge' in assertion) {\n      const result = await gradeLLMJudge(botResponse, assertion.llm_judge, {\n        userMessage: context.userMessage,\n        ...(context.judgePassThreshold !== undefined ? { passThreshold: context.judgePassThreshold } : {}),\n      })\n      results.push(result)\n      continue\n    }\n\n    // TODO: uncomment this once we support it\n    // if ('similar_to' in assertion) {\n    //   results.push({\n    //     assertion: `similar_to: \"${assertion.similar_to}\"`,\n    //     pass: true,\n    //     expected: assertion.similar_to,\n    //     actual: 'SKIPPED \u2014 similar_to not yet implemented',\n    //   })\n    //   continue\n    // }\n\n    results.push({\n      assertion: 'unknown',\n      pass: false,\n      expected: 'known assertion type',\n      actual: `Unknown assertion: ${JSON.stringify(assertion)}`,\n    })\n  }\n\n  return results\n}\n", "/**\n * Tool assertion graders.\n * Checks tool calls from traces against expected assertions.\n */\n\nimport type { ToolAssertion, ToolCall, GraderResult } from '../types'\nimport { matchValue, operatorToString } from './match'\n\nexport function gradeTools(toolCalls: ToolCall[], assertions: ToolAssertion[]): GraderResult[] {\n  return assertions.map((assertion) => {\n    // --- called ---\n    if ('called' in assertion && !('not_called' in assertion) && !('call_order' in assertion)) {\n      const matches = toolCalls.filter((tc) => tc.name === assertion.called)\n      const wasCalled = matches.length > 0\n\n      if (!wasCalled) {\n        return {\n          assertion: `tool called: ${assertion.called}`,\n          pass: false,\n          expected: `${assertion.called} was called`,\n          actual: `Not called. Tools called: [${toolCalls.map((tc) => tc.name).join(', ') || 'none'}]`,\n        }\n      }\n\n      // Check params if specified\n      if (assertion.params) {\n        const paramResults: { key: string; pass: boolean; detail: string }[] = []\n\n        for (const [key, operator] of Object.entries(assertion.params)) {\n          // Check across all matching calls \u2014 pass if any call matches\n          const anyMatch = matches.some((tc) => matchValue(operator, tc.input[key]))\n          paramResults.push({\n            key,\n            pass: anyMatch,\n            detail: anyMatch\n              ? `matched`\n              : `expected ${key} ${operatorToString(operator)}, got ${JSON.stringify(matches.map((tc) => tc.input[key]))}`,\n          })\n        }\n\n        const allParamsPass = paramResults.every((p) => p.pass)\n        const failedParams = paramResults.filter((p) => !p.pass)\n\n        return {\n          assertion: `tool called: ${assertion.called} with params`,\n          pass: allParamsPass,\n          expected: `${assertion.called} called with ${Object.entries(assertion.params)\n            .map(([k, v]) => `${k} ${operatorToString(v)}`)\n            .join(', ')}`,\n          actual: allParamsPass ? `Matched` : failedParams.map((p) => p.detail).join('; '),\n        }\n      }\n\n      return {\n        assertion: `tool called: ${assertion.called}`,\n        pass: true,\n        expected: `${assertion.called} was called`,\n        actual: `Called ${matches.length} time(s)`,\n      }\n    }\n\n    // --- not_called ---\n    if ('not_called' in assertion) {\n      const wasCalled = toolCalls.some((tc) => tc.name === assertion.not_called)\n      return {\n        assertion: `tool not_called: ${assertion.not_called}`,\n        pass: !wasCalled,\n        expected: `${assertion.not_called} was NOT called`,\n        actual: wasCalled ? `Was called` : `Not called`,\n      }\n    }\n\n    // --- call_order ---\n    if ('call_order' in assertion) {\n      const calledNames = toolCalls.map((tc) => tc.name)\n      const expectedOrder = assertion.call_order as string[]\n\n      // Check that the expected sequence appears in order (not necessarily contiguous)\n      let cursor = 0\n      for (const name of calledNames) {\n        if (cursor < expectedOrder.length && name === expectedOrder[cursor]) {\n          cursor++\n        }\n      }\n      const inOrder = cursor === expectedOrder.length\n\n      return {\n        assertion: `call_order: [${expectedOrder.join(' \u2192 ')}]`,\n        pass: inOrder,\n        expected: `Tools called in order: [${expectedOrder.join(' \u2192 ')}]`,\n        actual: `Actual order: [${calledNames.join(' \u2192 ') || 'none'}]`,\n      }\n    }\n\n    return {\n      assertion: 'unknown tool assertion',\n      pass: false,\n      expected: 'known assertion type',\n      actual: `Unknown: ${JSON.stringify(assertion)}`,\n    }\n  })\n}\n", "/**\n * State assertion graders.\n * Pure functions that grade state assertions against StateMutation[] from the transformer.\n * No API calls, no retries, no BpClient.\n */\n\nimport type { StateMutation, StateAssertion, GraderResult } from '../types'\n\n/**\n * Parse a state path like \"conversation.lastCity\" into its components.\n *   bot.*            -> type: \"bot\",          field: \"*\"\n *   user.*           -> type: \"user\",         field: \"*\"\n *   conversation.*   -> type: \"conversation\", field: \"*\"\n */\nfunction parseStatePath(path: string): { type: string; field: string } {\n  const dot = path.indexOf('.')\n  if (dot === -1) {\n    throw new Error(`Invalid state path \"${path}\" \u2014 expected \"type.field\" format`)\n  }\n\n  const type = path.slice(0, dot)\n  const field = path.slice(dot + 1)\n\n  if (type !== 'bot' && type !== 'user' && type !== 'conversation') {\n    throw new Error(`Unknown state type \"${type}\" in path \"${path}\" \u2014 expected bot, user, or conversation`)\n  }\n\n  return { type, field }\n}\n\n/**\n * Grade state assertions using StateMutation[] from the transformer.\n *\n * - `equals`: find the LAST mutation matching the state type, extract field from `current`, deep-compare\n * - `changed: true`: check if the field appears in ANY mutation's `changedKeys` for that state type\n * - `changed: false`: check that the field appears in NO mutation's `changedKeys`\n * - Swapped state: when `swappedToFile` is true, current/previous are undefined \u2014 report a clear message\n */\nexport function gradeState(mutations: StateMutation[], assertions: StateAssertion[]): GraderResult[] {\n  const results: GraderResult[] = []\n\n  for (const assertion of assertions) {\n    const parsed = parseStatePath(assertion.path)\n    const matchingMutations = mutations.filter((m) => m.type === parsed.type)\n\n    if (assertion.equals !== undefined) {\n      if (matchingMutations.length === 0) {\n        results.push({\n          assertion: `state: ${assertion.path} equals`,\n          pass: false,\n          expected: JSON.stringify(assertion.equals),\n          actual:\n            `No state.save span found for \"${parsed.type}\" \u2014 the bot did not mutate this state during the turn. ` +\n            'If this state was seeded via setup.state, assert on the tool/action that reads it, or use a conversation turn that triggers a state write.',\n        })\n        continue\n      }\n\n      const lastMutation = matchingMutations[matchingMutations.length - 1]!\n\n      if (lastMutation.swappedToFile) {\n        results.push({\n          assertion: `state: ${assertion.path} equals`,\n          pass: false,\n          expected: JSON.stringify(assertion.equals),\n          actual: 'State was too large for trace-based assertion (swapped to file)',\n        })\n        continue\n      }\n\n      const currentState = lastMutation.current as Record<string, unknown> | undefined\n      const actualValue = currentState ? currentState[parsed.field] : undefined\n      const pass = deepEqual(actualValue, assertion.equals)\n\n      results.push({\n        assertion: `state: ${assertion.path} equals`,\n        pass,\n        expected: JSON.stringify(assertion.equals),\n        actual: JSON.stringify(actualValue),\n      })\n    }\n\n    if (assertion.changed !== undefined) {\n      if (matchingMutations.length === 0) {\n        results.push({\n          assertion: `state: ${assertion.path} ${assertion.changed ? 'changed' : 'unchanged'}`,\n          pass: !assertion.changed,\n          expected: assertion.changed\n            ? `Field \"${parsed.field}\" changed in a \"${parsed.type}\" state mutation`\n            : `Field \"${parsed.field}\" unchanged`,\n          actual: 'No state mutation found for type \"' + parsed.type + '\"',\n        })\n        continue\n      }\n\n      // Check for swapped state in any matching mutation\n      const hasSwapped = matchingMutations.some((m) => m.swappedToFile)\n      if (hasSwapped) {\n        results.push({\n          assertion: `state: ${assertion.path} ${assertion.changed ? 'changed' : 'unchanged'}`,\n          pass: false,\n          expected: assertion.changed\n            ? `Field \"${parsed.field}\" changed in a \"${parsed.type}\" state mutation`\n            : `Field \"${parsed.field}\" unchanged`,\n          actual: 'State was too large for trace-based assertion (swapped to file)',\n        })\n        continue\n      }\n\n      const fieldChanged = matchingMutations.some((m) => m.changedKeys.includes(parsed.field))\n      const pass = assertion.changed ? fieldChanged : !fieldChanged\n\n      results.push({\n        assertion: `state: ${assertion.path} ${assertion.changed ? 'changed' : 'unchanged'}`,\n        pass,\n        expected: assertion.changed\n          ? `Field \"${parsed.field}\" changed in a \"${parsed.type}\" state mutation`\n          : `Field \"${parsed.field}\" unchanged`,\n        actual: fieldChanged ? `Field \"${parsed.field}\" was changed` : `Field \"${parsed.field}\" was not changed`,\n      })\n    }\n  }\n\n  return results\n}\n\nfunction deepEqual(a: unknown, b: unknown): boolean {\n  if (a === b) return true\n  if (a === null || b === null) return false\n  if (typeof a !== typeof b) return false\n\n  if (Array.isArray(a) && Array.isArray(b)) {\n    if (a.length !== b.length) return false\n    return a.every((v, i) => deepEqual(v, b[i]))\n  }\n\n  if (Array.isArray(a) !== Array.isArray(b)) return false\n\n  if (typeof a === 'object' && typeof b === 'object') {\n    const keysA = Object.keys(a as object)\n    const keysB = Object.keys(b as object)\n    if (keysA.length !== keysB.length) return false\n    return keysA.every((k) => deepEqual((a as Record<string, unknown>)[k], (b as Record<string, unknown>)[k]))\n  }\n\n  return false\n}\n", "/**\n * Workflow assertion graders.\n * Checks workflow execution via transformer-produced WorkflowSpan data.\n */\n\nimport type { WorkflowAssertion, WorkflowSpan, GraderResult } from '../types'\n\nexport function gradeWorkflows(workflows: WorkflowSpan[], assertions: WorkflowAssertion[]): GraderResult[] {\n  const results: GraderResult[] = []\n\n  for (const assertion of assertions) {\n    const matching = workflows.filter((wf) => wf.name === assertion.name)\n\n    // entered \u2014 was the workflow triggered at all?\n    if (assertion.entered !== undefined) {\n      const wasEntered = matching.length > 0\n      const pass = assertion.entered ? wasEntered : !wasEntered\n\n      results.push({\n        assertion: `workflow: ${assertion.name} ${assertion.entered ? 'entered' : 'not entered'}`,\n        pass,\n        expected: assertion.entered\n          ? `Workflow \"${assertion.name}\" was entered`\n          : `Workflow \"${assertion.name}\" was not entered`,\n        actual: wasEntered ? `Found ${matching.length} workflow span(s)` : `No workflow spans found`,\n      })\n    }\n\n    // completed \u2014 did the workflow complete?\n    if (assertion.completed !== undefined) {\n      const didComplete = matching.some(\n        (wf) => wf.statusFinal === 'completed' || wf.status === 'ok' || wf.status === 'error'\n      )\n      const pass = assertion.completed ? didComplete : !didComplete\n\n      results.push({\n        assertion: `workflow: ${assertion.name} ${assertion.completed ? 'completed' : 'not completed'}`,\n        pass,\n        expected: assertion.completed\n          ? `Workflow \"${assertion.name}\" completed`\n          : `Workflow \"${assertion.name}\" did not complete`,\n        actual: didComplete ? `Completed` : `Not completed`,\n      })\n    }\n  }\n\n  return results\n}\n", "/**\n * Timing assertion graders.\n * Checks bot response duration against expected thresholds.\n */\n\nimport type { TimingAssertion, GraderResult } from '../types'\nimport { matchValue, operatorToString } from './match'\n\nexport function gradeTiming(botDuration: number, assertions: TimingAssertion[]): GraderResult[] {\n  const results: GraderResult[] = []\n\n  for (const assertion of assertions) {\n    const pass = matchValue(assertion.response_time, botDuration)\n    const expected = operatorToString(assertion.response_time)\n\n    results.push({\n      assertion: `response_time ${expected}`,\n      pass,\n      expected: `Response time ${expected}`,\n      actual: `${botDuration}ms`,\n    })\n  }\n\n  return results\n}\n", "/**\n * Outcome assertion orchestrator.\n * Runs state and workflow graders after all conversation turns complete.\n */\n\nimport type { EvalDefinition, GraderResult, StateMutation, WorkflowSpan } from '../types'\nimport { gradeState } from './state'\nimport { gradeWorkflows } from './workflow'\n\n/**\n * Grade all outcome assertions after the conversation completes.\n * Takes transformer output (StateMutation[], WorkflowSpan[]) \u2014 no raw spans, no API calls.\n */\nexport function gradeOutcome(\n  mutations: StateMutation[],\n  evalDef: EvalDefinition,\n  workflowSpans: WorkflowSpan[]\n): GraderResult[] {\n  const outcome = evalDef.outcome\n  if (!outcome) return []\n\n  const results: GraderResult[] = []\n\n  if (outcome.state && outcome.state.length > 0) {\n    const stateResults = gradeState(mutations, outcome.state)\n    results.push(...stateResults)\n  }\n\n  if (outcome.workflow && outcome.workflow.length > 0) {\n    const workflowResults = gradeWorkflows(workflowSpans, outcome.workflow)\n    results.push(...workflowResults)\n  }\n\n  return results\n}\n"],
+  "mappings": ";AAUO,SAAS,WAAW,UAAyB,QAA0B;AAE5E,MAAI,OAAO,aAAa,UAAU;AAChC,WAAO,OAAO,MAAM,MAAM;AAAA,EAC5B;AAEA,MAAI,YAAY,UAAU;AACxB,WAAO,WAAW,SAAS;AAAA,EAC7B;AAEA,MAAI,cAAc,UAAU;AAC1B,WAAO,OAAO,MAAM,EAAE,YAAY,EAAE,SAAS,SAAS,SAAS,YAAY,CAAC;AAAA,EAC9E;AAEA,MAAI,kBAAkB,UAAU;AAC9B,WAAO,CAAC,OAAO,MAAM,EAAE,YAAY,EAAE,SAAS,SAAS,aAAa,YAAY,CAAC;AAAA,EACnF;AAEA,MAAI,aAAa,UAAU;AACzB,WAAO,IAAI,OAAO,SAAS,SAAS,GAAG,EAAE,KAAK,OAAO,MAAM,CAAC;AAAA,EAC9D;AAEA,MAAI,QAAQ,UAAU;AACpB,WAAO,SAAS,GAAG,SAAS,MAAM;AAAA,EACpC;AAEA,MAAI,YAAY,UAAU;AACxB,WAAO,SAAS,SAAS,WAAW,UAAa,WAAW,OAAO,WAAW,UAAa,WAAW;AAAA,EACxG;AAEA,MAAI,SAAS,UAAU;AACrB,WAAO,OAAO,MAAM,KAAK,SAAS;AAAA,EACpC;AAEA,MAAI,SAAS,UAAU;AACrB,WAAO,OAAO,MAAM,KAAK,SAAS;AAAA,EACpC;AAEA,SAAO;AACT;AAEO,SAAS,iBAAiB,UAAiC;AAChE,MAAI,OAAO,aAAa,SAAU,QAAO,WAAW,QAAQ;AAC5D,MAAI,YAAY,SAAU,QAAO,UAAU,KAAK,UAAU,SAAS,MAAM,CAAC;AAC1E,MAAI,cAAc,SAAU,QAAO,aAAa,SAAS,QAAQ;AACjE,MAAI,kBAAkB,SAAU,QAAO,iBAAiB,SAAS,YAAY;AAC7E,MAAI,aAAa,SAAU,QAAO,YAAY,SAAS,OAAO;AAC9D,MAAI,QAAQ,SAAU,QAAO,OAAO,SAAS,GAAG,IAAI,CAAC,MAAM,KAAK,UAAU,CAAC,CAAC,EAAE,KAAK,IAAI,CAAC;AACxF,MAAI,YAAY,SAAU,QAAO,SAAS,SAAS,WAAW;AAC9D,MAAI,SAAS,SAAU,QAAO,MAAM,SAAS,GAAG;AAChD,MAAI,SAAS,SAAU,QAAO,MAAM,SAAS,GAAG;AAChD,SAAO,KAAK,UAAU,QAAQ;AAChC;;;ACzDA,SAAS,cAAc;AACvB,SAAS,iBAAiB;;;ACuJnB,IAAM,gBAA4B;;;ADnJzC,IAAM,yBAAyB;AAE/B,SAAS,uBAAuB,eAA+B;AAC7D,SAAO;AAAA;AAAA;AAAA;AAAA;AAAA;AAAA;AAAA;AAAA;AAAA;AAAA;AAAA;AAAA;AAAA;AAAA;AAAA;AAAA;AAAA;AAAA;AAAA;AAAA;AAAA,aAqBI,aAAa;AAC1B;AAEA,IAAM,gBAAgB,CAAC,QAAQ,MAAM;AAErC,IAAI,aAA+B;AACnC,IAAI,mBAA6B,CAAC;AAClC,IAAI,cAAsB;AAC1B,IAAI,OAAmB;AAUvB,SAAS,cAAc,OAAwB;AAC7C,MAAI,cAAc,SAAS,KAAK,EAAG,QAAO;AAC1C,MAAI,iBAAiB,WAAW,EAAG,QAAO;AAE1C,SAAO,iBAAiB,SAAS,KAAK;AACxC;AAMA,eAAsB,aACpB,aACA,SACA;AACA,QAAM,MAAM,SAAS,UAAU;AAC/B,SAAO;AAEP,QAAM,SAAS,IAAI,OAAO;AAAA,IACxB,OAAO,YAAY;AAAA,IACnB,QAAQ,YAAY;AAAA,IACpB,OAAO,YAAY;AAAA,IACnB,GAAI,YAAY,gBAAgB,SAAY,EAAE,aAAa,YAAY,YAAY,IAAI,CAAC;AAAA,EAC1F,CAAC;AACD,eAAa,IAAI,UAAU,EAAE,QAAQ,qBAAqB,KAAK,CAAC;AAEhE,MAAI;AACF,uBAAmB,MAAM,MAAM,MAAM,WAAW,kBAAkB,GAAG,KAAK,CAAC;AAAA,EAC7E,SAAS,KAAK;AACZ,QAAI,KAAK,qCAAsC,IAAc,OAAO,kCAAkC;AACtG,uBAAmB,CAAC;AAAA,EACtB;AAEA,gBAAc,SAAS,SAAS;AAEhC,MAAI,iBAAiB,WAAW,GAAG;AACjC,QAAI,KAAK,6DAA6D;AAAA,EACxE,WAAW,CAAC,cAAc,WAAW,GAAG;AACtC,QAAI,KAAK,+BAA+B,WAAW,0DAA0D;AAAA,EAC/G;AACF;AAEA,eAAsB,cACpB,aACA,UACA,SACuB;AACvB,QAAM,YAAY,KAAK,IAAI,GAAG,KAAK,IAAI,GAAG,QAAQ,iBAAiB,sBAAsB,CAAC;AAE1F,MAAI;AAIF,QAAI,CAAC,YAAY;AACf,WAAK,KAAK,8BAA8B,QAAQ,wDAAwD;AACxG,aAAO;AAAA,QACL,WAAW,eAAe,QAAQ;AAAA,QAClC,MAAM;AAAA,QACN,UAAU;AAAA,QACV,QAAQ;AAAA,MACV;AAAA,IACF;AAEA,QAAI,CAAC,cAAc,WAAW,GAAG;AAC/B,aAAO;AAAA,QACL,WAAW,eAAe,QAAQ;AAAA,QAClC,MAAM;AAAA,QACN,UAAU;AAAA,QACV,QAAQ;AAAA,MACV;AAAA,IACF;AAIA,UAAM,EAAE,OAAO,IAAI,MAAM,WAAW,gBAAgB;AAAA,MAClD,OAAO;AAAA,MACP,aAAa;AAAA,MACb,gBAAgB;AAAA,MAChB,cAAc,uBAAuB,SAAS;AAAA,MAC9C,UAAU;AAAA,QACR;AAAA,UACE,MAAM;AAAA,UACN,SAAS,iBAAiB,QAAQ,WAAW;AAAA;AAAA,gBAAqB,WAAW;AAAA;AAAA,YAAiB,QAAQ;AAAA,QACxG;AAAA,MACF;AAAA,IACF,CAAC;AAED,UAAM,aAAa,QAAQ,UAAU,CAAC,GAAG;AACzC,UAAM,UAAU,OAAO,eAAe,WAAW,aAAa;AAC9D,QAAI,CAAC,SAAS;AACZ,WAAK,KAAK,6CAA6C,QAAQ,GAAG;AAClE,aAAO;AAAA,QACL,WAAW,eAAe,QAAQ;AAAA,QAClC,MAAM;AAAA,QACN,UAAU;AAAA,QACV,QAAQ;AAAA,MACV;AAAA,IACF;AAEA,UAAM,UAAU,KAAK,MAAM,OAAO;AAMlC,WAAO;AAAA,MACL,WAAW,eAAe,QAAQ;AAAA,MAClC,MAAM,QAAQ,SAAS;AAAA,MACvB,UAAU;AAAA,MACV,QAAQ,SAAS,QAAQ,KAAK,kBAAkB,SAAS,YAAO,QAAQ,MAAM;AAAA,IAChF;AAAA,EACF,SAAS,KAAK;AACZ,SAAK,KAAK,yBAAyB,QAAQ,MAAO,IAAc,OAAO,EAAE;AACzE,WAAO;AAAA,MACL,WAAW,eAAe,QAAQ;AAAA,MAClC,MAAM;AAAA,MACN,UAAU;AAAA,MACV,QAAQ,wCAAoC,IAAc,OAAO;AAAA,IACnE;AAAA,EACF;AACF;;;AEpKA,eAAsB,cACpB,aACA,YACA,SACyB;AACzB,QAAM,UAA0B,CAAC;AAEjC,aAAW,aAAa,YAAY;AAClC,QAAI,cAAc,WAAW;AAC3B,YAAM,OAAO,YAAY,YAAY,EAAE,SAAS,UAAU,SAAS,YAAY,CAAC;AAChF,cAAQ,KAAK;AAAA,QACX,WAAW,aAAa,UAAU,QAAQ;AAAA,QAC1C;AAAA,QACA,UAAU,sBAAsB,UAAU,QAAQ;AAAA,QAClD,QAAQ,OAAO,sBAAsB;AAAA,MACvC,CAAC;AACD;AAAA,IACF;AAEA,QAAI,kBAAkB,WAAW;AAC/B,YAAM,OAAO,CAAC,YAAY,YAAY,EAAE,SAAS,UAAU,aAAa,YAAY,CAAC;AACrF,cAAQ,KAAK;AAAA,QACX,WAAW,iBAAiB,UAAU,YAAY;AAAA,QAClD;AAAA,QACA,UAAU,8BAA8B,UAAU,YAAY;AAAA,QAC9D,QAAQ,OAAO,0BAA0B;AAAA,MAC3C,CAAC;AACD;AAAA,IACF;AAEA,QAAI,aAAa,WAAW;AAC1B,YAAM,QAAQ,IAAI,OAAO,UAAU,SAAS,GAAG;AAC/C,YAAM,OAAO,MAAM,KAAK,WAAW;AACnC,cAAQ,KAAK;AAAA,QACX,WAAW,WAAW,UAAU,OAAO;AAAA,QACvC;AAAA,QACA,UAAU,qBAAqB,UAAU,OAAO;AAAA,QAChD,QAAQ,OAAO,YAAY;AAAA,MAC7B,CAAC;AACD;AAAA,IACF;AAEA,QAAI,eAAe,WAAW;AAC5B,YAAM,SAAS,MAAM,cAAc,aAAa,UAAU,WAAW;AAAA,QACnE,aAAa,QAAQ;AAAA,QACrB,GAAI,QAAQ,uBAAuB,SAAY,EAAE,eAAe,QAAQ,mBAAmB,IAAI,CAAC;AAAA,MAClG,CAAC;AACD,cAAQ,KAAK,MAAM;AACnB;AAAA,IACF;AAaA,YAAQ,KAAK;AAAA,MACX,WAAW;AAAA,MACX,MAAM;AAAA,MACN,UAAU;AAAA,MACV,QAAQ,sBAAsB,KAAK,UAAU,SAAS,CAAC;AAAA,IACzD,CAAC;AAAA,EACH;AAEA,SAAO;AACT;;;ACvEO,SAAS,WAAW,WAAuB,YAA6C;AAC7F,SAAO,WAAW,IAAI,CAAC,cAAc;AAEnC,QAAI,YAAY,aAAa,EAAE,gBAAgB,cAAc,EAAE,gBAAgB,YAAY;AACzF,YAAM,UAAU,UAAU,OAAO,CAAC,OAAO,GAAG,SAAS,UAAU,MAAM;AACrE,YAAM,YAAY,QAAQ,SAAS;AAEnC,UAAI,CAAC,WAAW;AACd,eAAO;AAAA,UACL,WAAW,gBAAgB,UAAU,MAAM;AAAA,UAC3C,MAAM;AAAA,UACN,UAAU,GAAG,UAAU,MAAM;AAAA,UAC7B,QAAQ,8BAA8B,UAAU,IAAI,CAAC,OAAO,GAAG,IAAI,EAAE,KAAK,IAAI,KAAK,MAAM;AAAA,QAC3F;AAAA,MACF;AAGA,UAAI,UAAU,QAAQ;AACpB,cAAM,eAAiE,CAAC;AAExE,mBAAW,CAAC,KAAK,QAAQ,KAAK,OAAO,QAAQ,UAAU,MAAM,GAAG;AAE9D,gBAAM,WAAW,QAAQ,KAAK,CAAC,OAAO,WAAW,UAAU,GAAG,MAAM,GAAG,CAAC,CAAC;AACzE,uBAAa,KAAK;AAAA,YAChB;AAAA,YACA,MAAM;AAAA,YACN,QAAQ,WACJ,YACA,YAAY,GAAG,IAAI,iBAAiB,QAAQ,CAAC,SAAS,KAAK,UAAU,QAAQ,IAAI,CAAC,OAAO,GAAG,MAAM,GAAG,CAAC,CAAC,CAAC;AAAA,UAC9G,CAAC;AAAA,QACH;AAEA,cAAM,gBAAgB,aAAa,MAAM,CAAC,MAAM,EAAE,IAAI;AACtD,cAAM,eAAe,aAAa,OAAO,CAAC,MAAM,CAAC,EAAE,IAAI;AAEvD,eAAO;AAAA,UACL,WAAW,gBAAgB,UAAU,MAAM;AAAA,UAC3C,MAAM;AAAA,UACN,UAAU,GAAG,UAAU,MAAM,gBAAgB,OAAO,QAAQ,UAAU,MAAM,EACzE,IAAI,CAAC,CAAC,GAAG,CAAC,MAAM,GAAG,CAAC,IAAI,iBAAiB,CAAC,CAAC,EAAE,EAC7C,KAAK,IAAI,CAAC;AAAA,UACb,QAAQ,gBAAgB,YAAY,aAAa,IAAI,CAAC,MAAM,EAAE,MAAM,EAAE,KAAK,IAAI;AAAA,QACjF;AAAA,MACF;AAEA,aAAO;AAAA,QACL,WAAW,gBAAgB,UAAU,MAAM;AAAA,QAC3C,MAAM;AAAA,QACN,UAAU,GAAG,UAAU,MAAM;AAAA,QAC7B,QAAQ,UAAU,QAAQ,MAAM;AAAA,MAClC;AAAA,IACF;AAGA,QAAI,gBAAgB,WAAW;AAC7B,YAAM,YAAY,UAAU,KAAK,CAAC,OAAO,GAAG,SAAS,UAAU,UAAU;AACzE,aAAO;AAAA,QACL,WAAW,oBAAoB,UAAU,UAAU;AAAA,QACnD,MAAM,CAAC;AAAA,QACP,UAAU,GAAG,UAAU,UAAU;AAAA,QACjC,QAAQ,YAAY,eAAe;AAAA,MACrC;AAAA,IACF;AAGA,QAAI,gBAAgB,WAAW;AAC7B,YAAM,cAAc,UAAU,IAAI,CAAC,OAAO,GAAG,IAAI;AACjD,YAAM,gBAAgB,UAAU;AAGhC,UAAI,SAAS;AACb,iBAAW,QAAQ,aAAa;AAC9B,YAAI,SAAS,cAAc,UAAU,SAAS,cAAc,MAAM,GAAG;AACnE;AAAA,QACF;AAAA,MACF;AACA,YAAM,UAAU,WAAW,cAAc;AAEzC,aAAO;AAAA,QACL,WAAW,gBAAgB,cAAc,KAAK,UAAK,CAAC;AAAA,QACpD,MAAM;AAAA,QACN,UAAU,2BAA2B,cAAc,KAAK,UAAK,CAAC;AAAA,QAC9D,QAAQ,kBAAkB,YAAY,KAAK,UAAK,KAAK,MAAM;AAAA,MAC7D;AAAA,IACF;AAEA,WAAO;AAAA,MACL,WAAW;AAAA,MACX,MAAM;AAAA,MACN,UAAU;AAAA,MACV,QAAQ,YAAY,KAAK,UAAU,SAAS,CAAC;AAAA,IAC/C;AAAA,EACF,CAAC;AACH;;;ACvFA,SAAS,eAAe,MAA+C;AACrE,QAAM,MAAM,KAAK,QAAQ,GAAG;AAC5B,MAAI,QAAQ,IAAI;AACd,UAAM,IAAI,MAAM,uBAAuB,IAAI,uCAAkC;AAAA,EAC/E;AAEA,QAAM,OAAO,KAAK,MAAM,GAAG,GAAG;AAC9B,QAAM,QAAQ,KAAK,MAAM,MAAM,CAAC;AAEhC,MAAI,SAAS,SAAS,SAAS,UAAU,SAAS,gBAAgB;AAChE,UAAM,IAAI,MAAM,uBAAuB,IAAI,cAAc,IAAI,8CAAyC;AAAA,EACxG;AAEA,SAAO,EAAE,MAAM,MAAM;AACvB;AAUO,SAAS,WAAW,WAA4B,YAA8C;AACnG,QAAM,UAA0B,CAAC;AAEjC,aAAW,aAAa,YAAY;AAClC,UAAM,SAAS,eAAe,UAAU,IAAI;AAC5C,UAAM,oBAAoB,UAAU,OAAO,CAAC,MAAM,EAAE,SAAS,OAAO,IAAI;AAExE,QAAI,UAAU,WAAW,QAAW;AAClC,UAAI,kBAAkB,WAAW,GAAG;AAClC,gBAAQ,KAAK;AAAA,UACX,WAAW,UAAU,UAAU,IAAI;AAAA,UACnC,MAAM;AAAA,UACN,UAAU,KAAK,UAAU,UAAU,MAAM;AAAA,UACzC,QACE,iCAAiC,OAAO,IAAI;AAAA,QAEhD,CAAC;AACD;AAAA,MACF;AAEA,YAAM,eAAe,kBAAkB,kBAAkB,SAAS,CAAC;AAEnE,UAAI,aAAa,eAAe;AAC9B,gBAAQ,KAAK;AAAA,UACX,WAAW,UAAU,UAAU,IAAI;AAAA,UACnC,MAAM;AAAA,UACN,UAAU,KAAK,UAAU,UAAU,MAAM;AAAA,UACzC,QAAQ;AAAA,QACV,CAAC;AACD;AAAA,MACF;AAEA,YAAM,eAAe,aAAa;AAClC,YAAM,cAAc,eAAe,aAAa,OAAO,KAAK,IAAI;AAChE,YAAM,OAAO,UAAU,aAAa,UAAU,MAAM;AAEpD,cAAQ,KAAK;AAAA,QACX,WAAW,UAAU,UAAU,IAAI;AAAA,QACnC;AAAA,QACA,UAAU,KAAK,UAAU,UAAU,MAAM;AAAA,QACzC,QAAQ,KAAK,UAAU,WAAW;AAAA,MACpC,CAAC;AAAA,IACH;AAEA,QAAI,UAAU,YAAY,QAAW;AACnC,UAAI,kBAAkB,WAAW,GAAG;AAClC,gBAAQ,KAAK;AAAA,UACX,WAAW,UAAU,UAAU,IAAI,IAAI,UAAU,UAAU,YAAY,WAAW;AAAA,UAClF,MAAM,CAAC,UAAU;AAAA,UACjB,UAAU,UAAU,UAChB,UAAU,OAAO,KAAK,mBAAmB,OAAO,IAAI,qBACpD,UAAU,OAAO,KAAK;AAAA,UAC1B,QAAQ,uCAAuC,OAAO,OAAO;AAAA,QAC/D,CAAC;AACD;AAAA,MACF;AAGA,YAAM,aAAa,kBAAkB,KAAK,CAAC,MAAM,EAAE,aAAa;AAChE,UAAI,YAAY;AACd,gBAAQ,KAAK;AAAA,UACX,WAAW,UAAU,UAAU,IAAI,IAAI,UAAU,UAAU,YAAY,WAAW;AAAA,UAClF,MAAM;AAAA,UACN,UAAU,UAAU,UAChB,UAAU,OAAO,KAAK,mBAAmB,OAAO,IAAI,qBACpD,UAAU,OAAO,KAAK;AAAA,UAC1B,QAAQ;AAAA,QACV,CAAC;AACD;AAAA,MACF;AAEA,YAAM,eAAe,kBAAkB,KAAK,CAAC,MAAM,EAAE,YAAY,SAAS,OAAO,KAAK,CAAC;AACvF,YAAM,OAAO,UAAU,UAAU,eAAe,CAAC;AAEjD,cAAQ,KAAK;AAAA,QACX,WAAW,UAAU,UAAU,IAAI,IAAI,UAAU,UAAU,YAAY,WAAW;AAAA,QAClF;AAAA,QACA,UAAU,UAAU,UAChB,UAAU,OAAO,KAAK,mBAAmB,OAAO,IAAI,qBACpD,UAAU,OAAO,KAAK;AAAA,QAC1B,QAAQ,eAAe,UAAU,OAAO,KAAK,kBAAkB,UAAU,OAAO,KAAK;AAAA,MACvF,CAAC;AAAA,IACH;AAAA,EACF;AAEA,SAAO;AACT;AAEA,SAAS,UAAU,GAAY,GAAqB;AAClD,MAAI,MAAM,EAAG,QAAO;AACpB,MAAI,MAAM,QAAQ,MAAM,KAAM,QAAO;AACrC,MAAI,OAAO,MAAM,OAAO,EAAG,QAAO;AAElC,MAAI,MAAM,QAAQ,CAAC,KAAK,MAAM,QAAQ,CAAC,GAAG;AACxC,QAAI,EAAE,WAAW,EAAE,OAAQ,QAAO;AAClC,WAAO,EAAE,MAAM,CAAC,GAAG,MAAM,UAAU,GAAG,EAAE,CAAC,CAAC,CAAC;AAAA,EAC7C;AAEA,MAAI,MAAM,QAAQ,CAAC,MAAM,MAAM,QAAQ,CAAC,EAAG,QAAO;AAElD,MAAI,OAAO,MAAM,YAAY,OAAO,MAAM,UAAU;AAClD,UAAM,QAAQ,OAAO,KAAK,CAAW;AACrC,UAAM,QAAQ,OAAO,KAAK,CAAW;AACrC,QAAI,MAAM,WAAW,MAAM,OAAQ,QAAO;AAC1C,WAAO,MAAM,MAAM,CAAC,MAAM,UAAW,EAA8B,CAAC,GAAI,EAA8B,CAAC,CAAC,CAAC;AAAA,EAC3G;AAEA,SAAO;AACT;;;AC3IO,SAAS,eAAe,WAA2B,YAAiD;AACzG,QAAM,UAA0B,CAAC;AAEjC,aAAW,aAAa,YAAY;AAClC,UAAM,WAAW,UAAU,OAAO,CAAC,OAAO,GAAG,SAAS,UAAU,IAAI;AAGpE,QAAI,UAAU,YAAY,QAAW;AACnC,YAAM,aAAa,SAAS,SAAS;AACrC,YAAM,OAAO,UAAU,UAAU,aAAa,CAAC;AAE/C,cAAQ,KAAK;AAAA,QACX,WAAW,aAAa,UAAU,IAAI,IAAI,UAAU,UAAU,YAAY,aAAa;AAAA,QACvF;AAAA,QACA,UAAU,UAAU,UAChB,aAAa,UAAU,IAAI,kBAC3B,aAAa,UAAU,IAAI;AAAA,QAC/B,QAAQ,aAAa,SAAS,SAAS,MAAM,sBAAsB;AAAA,MACrE,CAAC;AAAA,IACH;AAGA,QAAI,UAAU,cAAc,QAAW;AACrC,YAAM,cAAc,SAAS;AAAA,QAC3B,CAAC,OAAO,GAAG,gBAAgB,eAAe,GAAG,WAAW,QAAQ,GAAG,WAAW;AAAA,MAChF;AACA,YAAM,OAAO,UAAU,YAAY,cAAc,CAAC;AAElD,cAAQ,KAAK;AAAA,QACX,WAAW,aAAa,UAAU,IAAI,IAAI,UAAU,YAAY,cAAc,eAAe;AAAA,QAC7F;AAAA,QACA,UAAU,UAAU,YAChB,aAAa,UAAU,IAAI,gBAC3B,aAAa,UAAU,IAAI;AAAA,QAC/B,QAAQ,cAAc,cAAc;AAAA,MACtC,CAAC;AAAA,IACH;AAAA,EACF;AAEA,SAAO;AACT;;;ACvCO,SAAS,YAAY,aAAqB,YAA+C;AAC9F,QAAM,UAA0B,CAAC;AAEjC,aAAW,aAAa,YAAY;AAClC,UAAM,OAAO,WAAW,UAAU,eAAe,WAAW;AAC5D,UAAM,WAAW,iBAAiB,UAAU,aAAa;AAEzD,YAAQ,KAAK;AAAA,MACX,WAAW,iBAAiB,QAAQ;AAAA,MACpC;AAAA,MACA,UAAU,iBAAiB,QAAQ;AAAA,MACnC,QAAQ,GAAG,WAAW;AAAA,IACxB,CAAC;AAAA,EACH;AAEA,SAAO;AACT;;;ACXO,SAAS,aACd,WACA,SACA,eACgB;AAChB,QAAM,UAAU,QAAQ;AACxB,MAAI,CAAC,QAAS,QAAO,CAAC;AAEtB,QAAM,UAA0B,CAAC;AAEjC,MAAI,QAAQ,SAAS,QAAQ,MAAM,SAAS,GAAG;AAC7C,UAAM,eAAe,WAAW,WAAW,QAAQ,KAAK;AACxD,YAAQ,KAAK,GAAG,YAAY;AAAA,EAC9B;AAEA,MAAI,QAAQ,YAAY,QAAQ,SAAS,SAAS,GAAG;AACnD,UAAM,kBAAkB,eAAe,eAAe,QAAQ,QAAQ;AACtE,YAAQ,KAAK,GAAG,eAAe;AAAA,EACjC;AAEA,SAAO;AACT;",
   "names": []
 }

package/dist/graders/llm.d.ts

@@ -1,25 +1,25 @@
 /**
- * LLM judge grader using @botpress/zai.
- * `zai.check` owns the judge prompt, output parsing, and retries on invalid model output.
+ * LLM judge grader using @botpress/cognitive.
+ * Replaces the OpenAI-based implementation from the standalone project.
  */
-import type { Client } from '@botpress/client';
 import type { EvalLogger, GraderResult } from '../types';
 /** @internal Test-only: clears module judge state so tests stay hermetic in the shared bun test process. */
 export declare function _resetLLMJudgeForTests(): void;
-/** @internal Test-only: injects a fake Zai client so judge behavior is testable offline. */
-export declare function _setLLMJudgeForTests(zai: unknown, options?: {
-    failClosed?: boolean;
-    model?: string;
-}): void;
-/** Initialize the LLM judge with an authenticated Botpress client. */
-export declare function initLLMJudge(client: Client, options?: {
+/**
+ * Initialize the LLM judge with explicit credentials.
+ * Call this before running evals so the judge doesn't need env vars.
+ */
+export declare function initLLMJudge(credentials: {
+    token: string;
+    apiUrl: string;
+    botId: string;
+    workspaceId?: string;
+}, options?: {
     model?: string;
     logger?: EvalLogger;
-    failClosed?: boolean;
 }): Promise<void>;
 export declare function gradeLLMJudge(botResponse: string, criteria: string, context: {
     userMessage: string;
-    /** @deprecated No-op: the judge is boolean (zai.check); there is no score threshold. */
     passThreshold?: number;
 }): Promise<GraderResult>;
 //# sourceMappingURL=llm.d.ts.map

package/dist/graders/llm.d.ts.map

@@ -1 +1 @@
-{"version":3,"file":"llm.d.ts","sourceRoot":"","sources":["../../src/graders/llm.ts"],"names":[],"mappings":"AAAA;;;GAGG;AAEH,OAAO,KAAK,EAAE,MAAM,EAAE,MAAM,kBAAkB,CAAA;AAG9C,OAAO,KAAK,EAAE,UAAU,EAAE,YAAY,EAAE,MAAM,UAAU,CAAA;AAcxD,4GAA4G;AAC5G,wBAAgB,sBAAsB,IAAI,IAAI,CAM7C;AAED,4FAA4F;AAC5F,wBAAgB,oBAAoB,CAAC,GAAG,EAAE,OAAO,EAAE,OAAO,CAAC,EAAE;IAAE,UAAU,CAAC,EAAE,OAAO,CAAC;IAAC,KAAK,CAAC,EAAE,MAAM,CAAA;CAAE,GAAG,IAAI,CAK3G;AASD,sEAAsE;AACtE,wBAAsB,YAAY,CAChC,MAAM,EAAE,MAAM,EACd,OAAO,CAAC,EAAE;IAAE,KAAK,CAAC,EAAE,MAAM,CAAC;IAAC,MAAM,CAAC,EAAE,UAAU,CAAC;IAAC,UAAU,CAAC,EAAE,OAAO,CAAA;CAAE,iBAwBxE;AAED,wBAAsB,aAAa,CACjC,WAAW,EAAE,MAAM,EACnB,QAAQ,EAAE,MAAM,EAChB,OAAO,EAAE;IACP,WAAW,EAAE,MAAM,CAAA;IACnB,wFAAwF;IACxF,aAAa,CAAC,EAAE,MAAM,CAAA;CACvB,GACA,OAAO,CAAC,YAAY,CAAC,CA0CvB"}
+{"version":3,"file":"llm.d.ts","sourceRoot":"","sources":["../../src/graders/llm.ts"],"names":[],"mappings":"AAAA;;;GAGG;AAIH,OAAO,KAAK,EAAE,UAAU,EAAE,YAAY,EAAE,MAAM,UAAU,CAAA;AAqCxD,4GAA4G;AAC5G,wBAAgB,sBAAsB,IAAI,IAAI,CAK7C;AASD;;;GAGG;AACH,wBAAsB,YAAY,CAChC,WAAW,EAAE;IAAE,KAAK,EAAE,MAAM,CAAC;IAAC,MAAM,EAAE,MAAM,CAAC;IAAC,KAAK,EAAE,MAAM,CAAC;IAAC,WAAW,CAAC,EAAE,MAAM,CAAA;CAAE,EACnF,OAAO,CAAC,EAAE;IAAE,KAAK,CAAC,EAAE,MAAM,CAAC;IAAC,MAAM,CAAC,EAAE,UAAU,CAAA;CAAE,iBA2BlD;AAED,wBAAsB,aAAa,CACjC,WAAW,EAAE,MAAM,EACnB,QAAQ,EAAE,MAAM,EAChB,OAAO,EAAE;IAAE,WAAW,EAAE,MAAM,CAAC;IAAC,aAAa,CAAC,EAAE,MAAM,CAAA;CAAE,GACvD,OAAO,CAAC,YAAY,CAAC,CA0EvB"}

package/dist/graders/llm.js

@@ -1,42 +1,64 @@
 // 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 _zai = null;
+var _cognitive = null;
 var _availableModels = [];
 var _judgeModel = "fast";
-var _failClosed = true;
 var _log = defaultLogger;
 function _resetLLMJudgeForTests() {
-  _zai = null;
+  _cognitive = null;
   _availableModels = [];
   _judgeModel = "fast";
   _log = defaultLogger;
-  _failClosed = true;
-}
-function _setLLMJudgeForTests(zai, options) {
-  _zai = zai;
-  _availableModels = ["fast"];
-  _judgeModel = options?.model ?? "fast";
-  _failClosed = options?.failClosed ?? true;
 }
 function _isModelValid(model) {
   if (MODEL_ALIASES.includes(model)) return true;
   if (_availableModels.length === 0) return false;
   return _availableModels.includes(model);
 }
-async function initLLMJudge(client, options) {
+async function initLLMJudge(credentials, options) {
   const log = options?.logger ?? defaultLogger;
   _log = log;
-  _failClosed = options?.failClosed ?? true;
-  const cognitive = new Cognitive({ client, __experimental_beta: true });
+  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 });
   try {
-    _availableModels = Array.from((await cognitive.fetchRemoteModels()).keys());
+    _availableModels = Array.from((await _cognitive.fetchRemoteModels()).keys());
   } catch (err) {
     log.warn(`Failed to fetch available models: ${err.message}. LLM judge will be unavailable.`);
     _availableModels = [];
@@ -47,42 +69,66 @@ async function initLLMJudge(client, options) {
   } else if (!_isModelValid(_judgeModel)) {
     log.warn(`Configured LLM judge model "${_judgeModel}" is invalid. Run "adk models" to list available models.`);
   }
-  _zai = new Zai({ client: cognitive, modelId: _judgeModel });
 }
 async function gradeLLMJudge(botResponse, criteria, context) {
-  const assertion = `llm_judge: "${criteria}"`;
+  const threshold = Math.max(1, Math.min(5, context.passThreshold ?? DEFAULT_PASS_THRESHOLD));
   try {
-    if (!_zai) {
+    if (!_cognitive) {
       _log.warn(`LLM judge unavailable for "${criteria}": no credentials configured (initLLMJudge not called)`);
       return {
-        assertion,
-        pass: !_failClosed,
-        skipped: true,
+        assertion: `llm_judge: "${criteria}"`,
+        pass: false,
         expected: criteria,
         actual: "FAILED \u2014 LLM judge unavailable: no credentials configured"
       };
     }
     if (!_isModelValid(_judgeModel)) {
       return {
-        assertion,
+        assertion: `llm_judge: "${criteria}"`,
         pass: false,
         expected: criteria,
         actual: "FAILED \u2014 invalid LLM judge model"
       };
     }
-    const { output } = await _zai.check({ userMessage: context.userMessage, botResponse }, criteria).result();
+    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: "${criteria}"`,
+        pass: false,
+        expected: criteria,
+        actual: "FAILED \u2014 LLM judge returned empty response"
+      };
+    }
+    const verdict = JSON.parse(content);
     return {
-      assertion,
-      pass: output.value,
+      assertion: `llm_judge: "${criteria}"`,
+      pass: verdict.score >= threshold,
       expected: criteria,
-      actual: `${output.value ? "Pass" : "Fail"} \u2014 ${output.explanation}`
+      actual: `Score ${verdict.score}/5 (threshold: ${threshold}) \u2014 ${verdict.reason}`
     };
   } catch (err) {
     _log.warn(`LLM judge failed for "${criteria}": ${err.message}`);
     return {
-      assertion,
-      pass: !_failClosed,
-      skipped: true,
+      assertion: `llm_judge: "${criteria}"`,
+      pass: false,
       expected: criteria,
       actual: `FAILED \u2014 LLM judge unavailable: ${err.message}`
     };
@@ -90,7 +136,6 @@ async function gradeLLMJudge(botResponse, criteria, context) {
 }
 export {
   _resetLLMJudgeForTests,
-  _setLLMJudgeForTests,
   gradeLLMJudge,
   initLLMJudge
 };

package/dist/graders/llm.js.map

@@ -1,7 +1,7 @@
 {
   "version": 3,
   "sources": ["../../src/graders/llm.ts", "../../src/types.ts"],
-  "sourcesContent": ["/**\n * LLM judge grader using @botpress/zai.\n * `zai.check` owns the judge prompt, output parsing, and retries on invalid model output.\n */\n\nimport type { Client } from '@botpress/client'\nimport { Cognitive } from '@botpress/cognitive'\nimport { Zai } from '@botpress/zai'\nimport type { EvalLogger, GraderResult } from '../types'\nimport { defaultLogger } from '../types'\n\nconst MODEL_ALIASES = ['fast', 'best']\n\nlet _zai: Zai | null = null\nlet _availableModels: string[] = []\nlet _judgeModel: string = 'fast'\n// When true (default), a judge that can't run is a hard FAIL, so an unavailable judge can't be\n// mistaken for success. Callers may opt into fail-open (a skipped pass) via `failClosed: false`,\n// e.g. local runs without judge credentials. Either way the result is marked `skipped: true`.\nlet _failClosed: boolean = true\nlet _log: EvalLogger = defaultLogger\n\n/** @internal Test-only: clears module judge state so tests stay hermetic in the shared bun test process. */\nexport function _resetLLMJudgeForTests(): void {\n  _zai = null\n  _availableModels = []\n  _judgeModel = 'fast'\n  _log = defaultLogger\n  _failClosed = true\n}\n\n/** @internal Test-only: injects a fake Zai client so judge behavior is testable offline. */\nexport function _setLLMJudgeForTests(zai: unknown, options?: { failClosed?: boolean; model?: string }): void {\n  _zai = zai as Zai\n  _availableModels = ['fast']\n  _judgeModel = options?.model ?? 'fast'\n  _failClosed = options?.failClosed ?? true\n}\n\nfunction _isModelValid(model: string): boolean {\n  if (MODEL_ALIASES.includes(model)) return true\n  if (_availableModels.length === 0) return false\n\n  return _availableModels.includes(model)\n}\n\n/** Initialize the LLM judge with an authenticated Botpress client. */\nexport async function initLLMJudge(\n  client: Client,\n  options?: { model?: string; logger?: EvalLogger; failClosed?: boolean }\n) {\n  const log = options?.logger ?? defaultLogger\n  _log = log\n  _failClosed = options?.failClosed ?? true\n\n  const cognitive = new Cognitive({ client, __experimental_beta: true })\n\n  try {\n    _availableModels = Array.from((await cognitive.fetchRemoteModels()).keys())\n  } catch (err) {\n    log.warn(`Failed to fetch available models: ${(err as Error).message}. LLM judge will be unavailable.`)\n    _availableModels = []\n  }\n\n  _judgeModel = options?.model ?? 'fast'\n\n  if (_availableModels.length === 0) {\n    log.warn('No available Cognitive models: cannot validate judge model.')\n  } else if (!_isModelValid(_judgeModel)) {\n    log.warn(`Configured LLM judge model \"${_judgeModel}\" is invalid. Run \"adk models\" to list available models.`)\n  }\n\n  _zai = new Zai({ client: cognitive, modelId: _judgeModel as ConstructorParameters<typeof Zai>[0]['modelId'] })\n}\n\nexport async function gradeLLMJudge(\n  botResponse: string,\n  criteria: string,\n  context: {\n    userMessage: string\n    /** @deprecated No-op: the judge is boolean (zai.check); there is no score threshold. */\n    passThreshold?: number\n  }\n): Promise<GraderResult> {\n  const assertion = `llm_judge: \"${criteria}\"`\n\n  try {\n    if (!_zai) {\n      _log.warn(`LLM judge unavailable for \"${criteria}\": no credentials configured (initLLMJudge not called)`)\n      return {\n        assertion,\n        pass: !_failClosed,\n        skipped: true,\n        expected: criteria,\n        actual: 'FAILED \u2014 LLM judge unavailable: no credentials configured',\n      }\n    }\n\n    if (!_isModelValid(_judgeModel)) {\n      return {\n        assertion,\n        pass: false,\n        expected: criteria,\n        actual: 'FAILED \u2014 invalid LLM judge model',\n      }\n    }\n\n    const { output } = await _zai.check({ userMessage: context.userMessage, botResponse }, criteria).result()\n\n    return {\n      assertion,\n      pass: output.value,\n      expected: criteria,\n      actual: `${output.value ? 'Pass' : 'Fail'} \u2014 ${output.explanation}`,\n    }\n  } catch (err) {\n    _log.warn(`LLM judge failed for \"${criteria}\": ${(err as Error).message}`)\n    return {\n      assertion,\n      pass: !_failClosed,\n      skipped: true,\n      expected: criteria,\n      actual: `FAILED \u2014 LLM judge unavailable: ${(err as Error).message}`,\n    }\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\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"],
-  "mappings": ";AAMA,SAAS,iBAAiB;AAC1B,SAAS,WAAW;;;AC0Jb,IAAM,gBAA4B;;;ADtJzC,IAAM,gBAAgB,CAAC,QAAQ,MAAM;AAErC,IAAI,OAAmB;AACvB,IAAI,mBAA6B,CAAC;AAClC,IAAI,cAAsB;AAI1B,IAAI,cAAuB;AAC3B,IAAI,OAAmB;AAGhB,SAAS,yBAA+B;AAC7C,SAAO;AACP,qBAAmB,CAAC;AACpB,gBAAc;AACd,SAAO;AACP,gBAAc;AAChB;AAGO,SAAS,qBAAqB,KAAc,SAA0D;AAC3G,SAAO;AACP,qBAAmB,CAAC,MAAM;AAC1B,gBAAc,SAAS,SAAS;AAChC,gBAAc,SAAS,cAAc;AACvC;AAEA,SAAS,cAAc,OAAwB;AAC7C,MAAI,cAAc,SAAS,KAAK,EAAG,QAAO;AAC1C,MAAI,iBAAiB,WAAW,EAAG,QAAO;AAE1C,SAAO,iBAAiB,SAAS,KAAK;AACxC;AAGA,eAAsB,aACpB,QACA,SACA;AACA,QAAM,MAAM,SAAS,UAAU;AAC/B,SAAO;AACP,gBAAc,SAAS,cAAc;AAErC,QAAM,YAAY,IAAI,UAAU,EAAE,QAAQ,qBAAqB,KAAK,CAAC;AAErE,MAAI;AACF,uBAAmB,MAAM,MAAM,MAAM,UAAU,kBAAkB,GAAG,KAAK,CAAC;AAAA,EAC5E,SAAS,KAAK;AACZ,QAAI,KAAK,qCAAsC,IAAc,OAAO,kCAAkC;AACtG,uBAAmB,CAAC;AAAA,EACtB;AAEA,gBAAc,SAAS,SAAS;AAEhC,MAAI,iBAAiB,WAAW,GAAG;AACjC,QAAI,KAAK,6DAA6D;AAAA,EACxE,WAAW,CAAC,cAAc,WAAW,GAAG;AACtC,QAAI,KAAK,+BAA+B,WAAW,0DAA0D;AAAA,EAC/G;AAEA,SAAO,IAAI,IAAI,EAAE,QAAQ,WAAW,SAAS,YAA+D,CAAC;AAC/G;AAEA,eAAsB,cACpB,aACA,UACA,SAKuB;AACvB,QAAM,YAAY,eAAe,QAAQ;AAEzC,MAAI;AACF,QAAI,CAAC,MAAM;AACT,WAAK,KAAK,8BAA8B,QAAQ,wDAAwD;AACxG,aAAO;AAAA,QACL;AAAA,QACA,MAAM,CAAC;AAAA,QACP,SAAS;AAAA,QACT,UAAU;AAAA,QACV,QAAQ;AAAA,MACV;AAAA,IACF;AAEA,QAAI,CAAC,cAAc,WAAW,GAAG;AAC/B,aAAO;AAAA,QACL;AAAA,QACA,MAAM;AAAA,QACN,UAAU;AAAA,QACV,QAAQ;AAAA,MACV;AAAA,IACF;AAEA,UAAM,EAAE,OAAO,IAAI,MAAM,KAAK,MAAM,EAAE,aAAa,QAAQ,aAAa,YAAY,GAAG,QAAQ,EAAE,OAAO;AAExG,WAAO;AAAA,MACL;AAAA,MACA,MAAM,OAAO;AAAA,MACb,UAAU;AAAA,MACV,QAAQ,GAAG,OAAO,QAAQ,SAAS,MAAM,WAAM,OAAO,WAAW;AAAA,IACnE;AAAA,EACF,SAAS,KAAK;AACZ,SAAK,KAAK,yBAAyB,QAAQ,MAAO,IAAc,OAAO,EAAE;AACzE,WAAO;AAAA,MACL;AAAA,MACA,MAAM,CAAC;AAAA,MACP,SAAS;AAAA,MACT,UAAU;AAAA,MACV,QAAQ,wCAAoC,IAAc,OAAO;AAAA,IACnE;AAAA,EACF;AACF;",
+  "sourcesContent": ["/**\n * LLM judge grader using @botpress/cognitive.\n * Replaces the OpenAI-based implementation from the standalone project.\n */\n\nimport { Client } from '@botpress/client'\nimport { Cognitive } from '@botpress/cognitive'\nimport type { EvalLogger, GraderResult } from '../types'\nimport { defaultLogger } from '../types'\n\nconst DEFAULT_PASS_THRESHOLD = 3\n\nfunction buildJudgeSystemPrompt(passThreshold: number): string {\n  return `You are an evaluation judge for a chatbot. You will be given:\n- The user's message\n- The bot's response\n- Grading criteria\n\nYour job is to determine if the bot's response meets the criteria.\n\nRespond with a JSON object:\n{\n  \"pass\": true/false,\n  \"score\": 1-5,\n  \"reason\": \"brief explanation\"\n}\n\nScoring guide:\n- 5: Fully meets criteria, excellent response\n- 4: Mostly meets criteria, minor issues\n- 3: Partially meets criteria, notable gaps\n- 2: Barely meets criteria, significant issues\n- 1: Does not meet criteria\n\nA score of ${passThreshold} or above is a pass.`\n}\n\nconst MODEL_ALIASES = ['fast', 'best']\n\nlet _cognitive: Cognitive | null = null\nlet _availableModels: string[] = []\nlet _judgeModel: string = 'fast'\nlet _log: EvalLogger = defaultLogger\n\n/** @internal Test-only: clears module judge state so tests stay hermetic in the shared bun test process. */\nexport function _resetLLMJudgeForTests(): void {\n  _cognitive = null\n  _availableModels = []\n  _judgeModel = 'fast'\n  _log = defaultLogger\n}\n\nfunction _isModelValid(model: string): boolean {\n  if (MODEL_ALIASES.includes(model)) return true\n  if (_availableModels.length === 0) return false\n\n  return _availableModels.includes(model)\n}\n\n/**\n * Initialize the LLM judge with explicit credentials.\n * Call this before running evals so the judge doesn't need env vars.\n */\nexport async function initLLMJudge(\n  credentials: { token: string; apiUrl: string; botId: string; workspaceId?: string },\n  options?: { model?: string; logger?: EvalLogger }\n) {\n  const log = options?.logger ?? defaultLogger\n  _log = log\n\n  const client = new Client({\n    token: credentials.token,\n    apiUrl: credentials.apiUrl,\n    botId: credentials.botId,\n    ...(credentials.workspaceId !== undefined ? { workspaceId: credentials.workspaceId } : {}),\n  })\n  _cognitive = new Cognitive({ client, __experimental_beta: true })\n\n  try {\n    _availableModels = Array.from((await _cognitive.fetchRemoteModels()).keys())\n  } catch (err) {\n    log.warn(`Failed to fetch available models: ${(err as Error).message}. LLM judge will be unavailable.`)\n    _availableModels = []\n  }\n\n  _judgeModel = options?.model ?? 'fast'\n\n  if (_availableModels.length === 0) {\n    log.warn('No available Cognitive models: cannot validate judge model.')\n  } else if (!_isModelValid(_judgeModel)) {\n    log.warn(`Configured LLM judge model \"${_judgeModel}\" is invalid. Run \"adk models\" to list available models.`)\n  }\n}\n\nexport async function gradeLLMJudge(\n  botResponse: string,\n  criteria: string,\n  context: { userMessage: string; passThreshold?: number }\n): Promise<GraderResult> {\n  const threshold = Math.max(1, Math.min(5, context.passThreshold ?? DEFAULT_PASS_THRESHOLD))\n\n  try {\n    // A judge that can't run must NOT silently pass the assertion \u2014 a flaky\n    // LLM call or missing credentials would turn real assertions green.\n    // Unavailability is a visible failure with a clear reason.\n    if (!_cognitive) {\n      _log.warn(`LLM judge unavailable for \"${criteria}\": no credentials configured (initLLMJudge not called)`)\n      return {\n        assertion: `llm_judge: \"${criteria}\"`,\n        pass: false,\n        expected: criteria,\n        actual: 'FAILED \u2014 LLM judge unavailable: no credentials configured',\n      }\n    }\n\n    if (!_isModelValid(_judgeModel)) {\n      return {\n        assertion: `llm_judge: \"${criteria}\"`,\n        pass: false,\n        expected: criteria,\n        actual: 'FAILED \u2014 invalid LLM judge model',\n      }\n    }\n\n    type CognitiveModel = Exclude<Parameters<Cognitive['generateContent']>[0]['model'], undefined>\n\n    const { output } = await _cognitive.generateContent({\n      model: _judgeModel as CognitiveModel,\n      temperature: 0,\n      responseFormat: 'json_object',\n      systemPrompt: buildJudgeSystemPrompt(threshold),\n      messages: [\n        {\n          role: 'user',\n          content: `User message: ${context.userMessage}\\n\\nBot response: ${botResponse}\\n\\nCriteria: ${criteria}`,\n        },\n      ],\n    })\n\n    const rawContent = output?.choices?.[0]?.content\n    const content = typeof rawContent === 'string' ? rawContent : null\n    if (!content) {\n      _log.warn(`LLM judge returned an empty response for \"${criteria}\"`)\n      return {\n        assertion: `llm_judge: \"${criteria}\"`,\n        pass: false,\n        expected: criteria,\n        actual: 'FAILED \u2014 LLM judge returned empty response',\n      }\n    }\n\n    const verdict = JSON.parse(content) as {\n      pass: boolean\n      score: number\n      reason: string\n    }\n\n    return {\n      assertion: `llm_judge: \"${criteria}\"`,\n      pass: verdict.score >= threshold,\n      expected: criteria,\n      actual: `Score ${verdict.score}/5 (threshold: ${threshold}) \u2014 ${verdict.reason}`,\n    }\n  } catch (err) {\n    _log.warn(`LLM judge failed for \"${criteria}\": ${(err as Error).message}`)\n    return {\n      assertion: `llm_judge: \"${criteria}\"`,\n      pass: false,\n      expected: criteria,\n      actual: `FAILED \u2014 LLM judge unavailable: ${(err as Error).message}`,\n    }\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"],
+  "mappings": ";AAKA,SAAS,cAAc;AACvB,SAAS,iBAAiB;;;ACuJnB,IAAM,gBAA4B;;;ADnJzC,IAAM,yBAAyB;AAE/B,SAAS,uBAAuB,eAA+B;AAC7D,SAAO;AAAA;AAAA;AAAA;AAAA;AAAA;AAAA;AAAA;AAAA;AAAA;AAAA;AAAA;AAAA;AAAA;AAAA;AAAA;AAAA;AAAA;AAAA;AAAA;AAAA;AAAA,aAqBI,aAAa;AAC1B;AAEA,IAAM,gBAAgB,CAAC,QAAQ,MAAM;AAErC,IAAI,aAA+B;AACnC,IAAI,mBAA6B,CAAC;AAClC,IAAI,cAAsB;AAC1B,IAAI,OAAmB;AAGhB,SAAS,yBAA+B;AAC7C,eAAa;AACb,qBAAmB,CAAC;AACpB,gBAAc;AACd,SAAO;AACT;AAEA,SAAS,cAAc,OAAwB;AAC7C,MAAI,cAAc,SAAS,KAAK,EAAG,QAAO;AAC1C,MAAI,iBAAiB,WAAW,EAAG,QAAO;AAE1C,SAAO,iBAAiB,SAAS,KAAK;AACxC;AAMA,eAAsB,aACpB,aACA,SACA;AACA,QAAM,MAAM,SAAS,UAAU;AAC/B,SAAO;AAEP,QAAM,SAAS,IAAI,OAAO;AAAA,IACxB,OAAO,YAAY;AAAA,IACnB,QAAQ,YAAY;AAAA,IACpB,OAAO,YAAY;AAAA,IACnB,GAAI,YAAY,gBAAgB,SAAY,EAAE,aAAa,YAAY,YAAY,IAAI,CAAC;AAAA,EAC1F,CAAC;AACD,eAAa,IAAI,UAAU,EAAE,QAAQ,qBAAqB,KAAK,CAAC;AAEhE,MAAI;AACF,uBAAmB,MAAM,MAAM,MAAM,WAAW,kBAAkB,GAAG,KAAK,CAAC;AAAA,EAC7E,SAAS,KAAK;AACZ,QAAI,KAAK,qCAAsC,IAAc,OAAO,kCAAkC;AACtG,uBAAmB,CAAC;AAAA,EACtB;AAEA,gBAAc,SAAS,SAAS;AAEhC,MAAI,iBAAiB,WAAW,GAAG;AACjC,QAAI,KAAK,6DAA6D;AAAA,EACxE,WAAW,CAAC,cAAc,WAAW,GAAG;AACtC,QAAI,KAAK,+BAA+B,WAAW,0DAA0D;AAAA,EAC/G;AACF;AAEA,eAAsB,cACpB,aACA,UACA,SACuB;AACvB,QAAM,YAAY,KAAK,IAAI,GAAG,KAAK,IAAI,GAAG,QAAQ,iBAAiB,sBAAsB,CAAC;AAE1F,MAAI;AAIF,QAAI,CAAC,YAAY;AACf,WAAK,KAAK,8BAA8B,QAAQ,wDAAwD;AACxG,aAAO;AAAA,QACL,WAAW,eAAe,QAAQ;AAAA,QAClC,MAAM;AAAA,QACN,UAAU;AAAA,QACV,QAAQ;AAAA,MACV;AAAA,IACF;AAEA,QAAI,CAAC,cAAc,WAAW,GAAG;AAC/B,aAAO;AAAA,QACL,WAAW,eAAe,QAAQ;AAAA,QAClC,MAAM;AAAA,QACN,UAAU;AAAA,QACV,QAAQ;AAAA,MACV;AAAA,IACF;AAIA,UAAM,EAAE,OAAO,IAAI,MAAM,WAAW,gBAAgB;AAAA,MAClD,OAAO;AAAA,MACP,aAAa;AAAA,MACb,gBAAgB;AAAA,MAChB,cAAc,uBAAuB,SAAS;AAAA,MAC9C,UAAU;AAAA,QACR;AAAA,UACE,MAAM;AAAA,UACN,SAAS,iBAAiB,QAAQ,WAAW;AAAA;AAAA,gBAAqB,WAAW;AAAA;AAAA,YAAiB,QAAQ;AAAA,QACxG;AAAA,MACF;AAAA,IACF,CAAC;AAED,UAAM,aAAa,QAAQ,UAAU,CAAC,GAAG;AACzC,UAAM,UAAU,OAAO,eAAe,WAAW,aAAa;AAC9D,QAAI,CAAC,SAAS;AACZ,WAAK,KAAK,6CAA6C,QAAQ,GAAG;AAClE,aAAO;AAAA,QACL,WAAW,eAAe,QAAQ;AAAA,QAClC,MAAM;AAAA,QACN,UAAU;AAAA,QACV,QAAQ;AAAA,MACV;AAAA,IACF;AAEA,UAAM,UAAU,KAAK,MAAM,OAAO;AAMlC,WAAO;AAAA,MACL,WAAW,eAAe,QAAQ;AAAA,MAClC,MAAM,QAAQ,SAAS;AAAA,MACvB,UAAU;AAAA,MACV,QAAQ,SAAS,QAAQ,KAAK,kBAAkB,SAAS,YAAO,QAAQ,MAAM;AAAA,IAChF;AAAA,EACF,SAAS,KAAK;AACZ,SAAK,KAAK,yBAAyB,QAAQ,MAAO,IAAc,OAAO,EAAE;AACzE,WAAO;AAAA,MACL,WAAW,eAAe,QAAQ;AAAA,MAClC,MAAM;AAAA,MACN,UAAU;AAAA,MACV,QAAQ,wCAAoC,IAAc,OAAO;AAAA,IACnE;AAAA,EACF;AACF;",
   "names": []
 }

package/dist/graders/response.js

@@ -1,16 +1,40 @@
 // 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 ${passThreshold} or above is a pass.`;
+}
 var MODEL_ALIASES = ["fast", "best"];
-var _zai = null;
+var _cognitive = null;
 var _availableModels = [];
 var _judgeModel = "fast";
-var _failClosed = true;
 var _log = defaultLogger;
 function _isModelValid(model) {
   if (MODEL_ALIASES.includes(model)) return true;
@@ -18,39 +42,64 @@ function _isModelValid(model) {
   return _availableModels.includes(model);
 }
 async function gradeLLMJudge(botResponse, criteria, context) {
-  const assertion = `llm_judge: "${criteria}"`;
+  const threshold = Math.max(1, Math.min(5, context.passThreshold ?? DEFAULT_PASS_THRESHOLD));
   try {
-    if (!_zai) {
+    if (!_cognitive) {
       _log.warn(`LLM judge unavailable for "${criteria}": no credentials configured (initLLMJudge not called)`);
       return {
-        assertion,
-        pass: !_failClosed,
-        skipped: true,
+        assertion: `llm_judge: "${criteria}"`,
+        pass: false,
         expected: criteria,
         actual: "FAILED \u2014 LLM judge unavailable: no credentials configured"
       };
     }
     if (!_isModelValid(_judgeModel)) {
       return {
-        assertion,
+        assertion: `llm_judge: "${criteria}"`,
         pass: false,
         expected: criteria,
         actual: "FAILED \u2014 invalid LLM judge model"
       };
     }
-    const { output } = await _zai.check({ userMessage: context.userMessage, botResponse }, criteria).result();
+    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: "${criteria}"`,
+        pass: false,
+        expected: criteria,
+        actual: "FAILED \u2014 LLM judge returned empty response"
+      };
+    }
+    const verdict = JSON.parse(content);
     return {
-      assertion,
-      pass: output.value,
+      assertion: `llm_judge: "${criteria}"`,
+      pass: verdict.score >= threshold,
       expected: criteria,
-      actual: `${output.value ? "Pass" : "Fail"} \u2014 ${output.explanation}`
+      actual: `Score ${verdict.score}/5 (threshold: ${threshold}) \u2014 ${verdict.reason}`
     };
   } catch (err) {
     _log.warn(`LLM judge failed for "${criteria}": ${err.message}`);
     return {
-      assertion,
-      pass: !_failClosed,
-      skipped: true,
+      assertion: `llm_judge: "${criteria}"`,
+      pass: false,
       expected: criteria,
       actual: `FAILED \u2014 LLM judge unavailable: ${err.message}`
     };