vitest-evals 0.13.0 → 0.14.0

package/dist/harness.mjs.map

@@ -1 +1 @@
-{"version":3,"sources":["../src/harness.ts"],"sourcesContent":["import {\n  assistantMessages,\n  failedSpans,\n  latestAssistantMessageContent,\n  messagesByRole,\n  spans,\n  spansByKind,\n  systemMessages,\n  toolCalls,\n  toolMessages,\n  userMessages,\n} from \"@vitest-evals/core\";\nimport type {\n  GenAiOperationName,\n  HarnessRun,\n  HarnessRunError,\n  JsonPrimitive,\n  JsonValue,\n  NormalizedMessage,\n  NormalizedSession,\n  NormalizedSpan,\n  NormalizedSpanAttributes,\n  NormalizedSpanEvent,\n  NormalizedTrace,\n  TimingSummary,\n  ToolCallRecord,\n  UsageSummary,\n} from \"@vitest-evals/core\";\n\nexport {\n  assistantMessages,\n  failedSpans,\n  latestAssistantMessageContent,\n  messagesByRole,\n  spans,\n  spansByKind,\n  systemMessages,\n  toolCalls,\n  toolMessages,\n  userMessages,\n} from \"@vitest-evals/core\";\nexport type {\n  GenAiOperationName,\n  GenAiOutputType,\n  GenAiProviderName,\n  GenAiSemanticAttributeKey,\n  GenAiSemanticAttributes,\n  GenAiTokenType,\n  GenAiToolType,\n  HarnessRun,\n  HarnessRunError,\n  JsonPrimitive,\n  JsonValue,\n  NormalizedMessage,\n  NormalizedSession,\n  NormalizedSpan,\n  NormalizedSpanAttributeKey,\n  NormalizedSpanAttributes,\n  NormalizedSpanEvent,\n  NormalizedTrace,\n  OpenTelemetrySemanticAttributeKey,\n  OpenTelemetrySemanticAttributes,\n  TimingSummary,\n  ToolCallRecord,\n  UsageSummary,\n} from \"@vitest-evals/core\";\n\n/** Options for converting normalized tool calls into trace spans. */\nexport type CreateToolCallSpansOptions = {\n  /** Trace id to attach to each generated tool span. */\n  traceId?: string;\n  /** Parent span id to attach to each generated tool span. */\n  parentId?: string;\n  /** Prefix used to create internal span ids instead of reusing tool-call ids. */\n  spanIdPrefix?: string;\n};\n\n/** Options for attaching a fallback run trace to a harness result. */\nexport type EnsureRunTraceOptions = {\n  /** Human-readable run or harness name. */\n  name: string;\n  /** Wall-clock start time for the harness run. */\n  startedAt: Date;\n  /** Wall-clock finish time for the harness run. */\n  finishedAt: Date;\n  /** Optional trace id. A generated id is used when omitted. */\n  id?: string;\n  /** GenAI operation name to place on the root run span. */\n  operationName?: GenAiOperationName;\n  /** Optional JSON-safe source marker for the trace metadata. */\n  source?: string;\n};\n\ntype OutputField<TOutput extends JsonValue | undefined> =\n  undefined extends TOutput ? { output?: TOutput } : { output: TOutput };\n\n/** Generic JSON-like metadata record used by normalized artifacts and reports. */\nexport type HarnessMetadata = Record<string, unknown>;\n\n/**\n * Runtime context passed from the eval fixture into a harness run.\n *\n * @example\n * ```ts\n * const harness: Harness<string> = {\n *   name: \"refund-agent\",\n *   async run(input, context) {\n *     context.setArtifact(\"inputLength\", input.length);\n *\n *     return {\n *       output: undefined,\n *       session: { messages: [{ role: \"user\", content: input }] },\n *       usage: {},\n *       errors: [],\n *     };\n *   },\n * };\n * ```\n */\nexport type HarnessContext = {\n  /** Abort signal from Vitest when available. */\n  signal?: AbortSignal;\n  /** Mutable JSON-safe artifact bag shared with the harness. */\n  artifacts: Record<string, JsonValue>;\n  /** Stores one JSON-safe artifact on the current run. */\n  setArtifact: (name: string, value: JsonValue) => void;\n};\n\n/**\n * Adapter that executes the system under test and returns a normalized run.\n *\n * @example\n * ```ts\n * const harness: Harness<string, { status: \"approved\" | \"denied\" }> = {\n *   name: \"refund-agent\",\n *   async run(input, context) {\n *     return normalizeHarnessRun(input, await runRefundFlow(input), context);\n *   },\n * };\n * ```\n */\nexport type Harness<\n  TInput = unknown,\n  TOutput extends JsonValue | undefined = JsonValue | undefined,\n> = {\n  /** Stable harness name used in reports. */\n  name: string;\n  /** Executes the system under test and returns a normalized run. */\n  run: (input: TInput, context: HarnessContext) => Promise<HarnessRun<TOutput>>;\n};\n\n/** Value or promise accepted by lightweight harness callbacks. */\nexport type MaybePromise<T> = T | Promise<T>;\n\n/** Lightweight tool-call record accepted by `createHarness(...)` results. */\nexport type SimpleToolCallRecord = Omit<\n  ToolCallRecord,\n  \"arguments\" | \"result\" | \"error\" | \"metadata\"\n> & {\n  /** Raw tool arguments accepted by `createHarness(...)` before normalization. */\n  arguments?: unknown;\n  /** Raw tool result accepted by `createHarness(...)` before normalization. */\n  result?: unknown;\n  /** Raw tool error accepted by `createHarness(...)` before normalization. */\n  error?: unknown;\n  /** Raw tool metadata accepted by `createHarness(...)` before normalization. */\n  metadata?: Record<string, unknown>;\n};\n\n/** Lightweight span event accepted by `createHarness(...)` results. */\nexport type SimpleSpanEvent = Omit<NormalizedSpanEvent, \"attributes\"> & {\n  /** Raw event attributes accepted by `createHarness(...)` before normalization. */\n  attributes?: Record<string, unknown>;\n};\n\n/** Lightweight span record accepted by `createHarness(...)` results. */\nexport type SimpleSpanRecord = Omit<\n  NormalizedSpan,\n  \"attributes\" | \"error\" | \"events\"\n> & {\n  /** Raw span attributes accepted by `createHarness(...)` before normalization. */\n  attributes?: Record<string, unknown>;\n  /** Raw span error accepted by `createHarness(...)` before normalization. */\n  error?: unknown;\n  /** Raw span events accepted by `createHarness(...)` before normalization. */\n  events?: SimpleSpanEvent[];\n};\n\n/** Lightweight trace record accepted by `createHarness(...)` results. */\nexport type SimpleTraceRecord = Omit<NormalizedTrace, \"metadata\" | \"spans\"> & {\n  /** Raw trace metadata accepted by `createHarness(...)` before normalization. */\n  metadata?: Record<string, unknown>;\n  /** Lightweight spans to normalize into the trace. */\n  spans: SimpleSpanRecord[];\n};\n\n/**\n * Lightweight result shape normalized by `createHarness(...)`.\n *\n * @example\n * ```ts\n * const result: SimpleHarnessResult<{ status: \"approved\" }> = {\n *   output: { status: \"approved\" },\n *   toolCalls: [{ name: \"lookupInvoice\", arguments: { invoiceId: \"inv_123\" } }],\n *   usage: { totalTokens: 260 },\n * };\n * ```\n */\nexport type SimpleHarnessResult<\n  TOutput extends JsonValue | undefined = JsonValue | undefined,\n> = OutputField<TOutput> & {\n  /** Pre-normalized transcript messages. When omitted, a default user/assistant transcript is created. */\n  messages?: NormalizedMessage[];\n  /** Lightweight tool-call records to normalize into the session. */\n  toolCalls?: SimpleToolCallRecord[];\n  /** Usage summary to attach to the run. */\n  usage?: UsageSummary;\n  /** Timing summary to attach to the run. */\n  timings?: TimingSummary;\n  /** Raw artifact values to normalize and merge into the run. */\n  artifacts?: Record<string, unknown>;\n  /** Lightweight traces and spans to normalize into the run. */\n  traces?: SimpleTraceRecord[];\n  /** Raw session metadata to normalize into the session. */\n  metadata?: Record<string, unknown>;\n  /** Raw errors to normalize into the run. */\n  errors?: unknown[];\n};\n\n/** Either a complete normalized run or a lightweight result to normalize. */\nexport type HarnessResultLike<\n  TOutput extends JsonValue | undefined = JsonValue | undefined,\n> = HarnessRun<TOutput> | SimpleHarnessResult<TOutput>;\n\n/** Arguments passed to the `createHarness(...)` convenience callback. */\nexport type CreateHarnessRunArgs<TInput> = {\n  /** Original input passed to `run(input)`. */\n  input: TInput;\n  /** Abort signal from Vitest when available. */\n  signal?: AbortSignal;\n  /** Mutable run artifact bag. */\n  artifacts: HarnessContext[\"artifacts\"];\n  /** Stores one JSON-safe artifact on the current run. */\n  setArtifact: HarnessContext[\"setArtifact\"];\n};\n\n/**\n * Options for creating a lightweight custom application harness.\n *\n * @example\n * ```ts\n * const options: CreateHarnessOptions<string, { status: \"approved\" }> = {\n *   name: \"refund-agent\",\n *   run: async ({ input }) => ({\n *     output: await classifyRefund(input),\n *   }),\n * };\n * ```\n */\nexport type CreateHarnessOptions<\n  TInput = unknown,\n  TOutput extends JsonValue | undefined = JsonValue | undefined,\n> = {\n  /** Stable harness name used in reports. */\n  name: string;\n  /** Executes application code and returns either a lightweight result or full `HarnessRun`. */\n  run: (\n    args: CreateHarnessRunArgs<TInput>,\n  ) => MaybePromise<HarnessResultLike<TOutput>>;\n};\n\nfunction isJsonPrimitive(value: unknown): value is JsonPrimitive {\n  return (\n    value === null ||\n    typeof value === \"string\" ||\n    typeof value === \"boolean\" ||\n    (typeof value === \"number\" && Number.isFinite(value))\n  );\n}\n\nfunction isJsonRecord(value: unknown): value is Record<string, unknown> {\n  return typeof value === \"object\" && value !== null && !Array.isArray(value);\n}\n\nfunction normalizeJsonArray(value: unknown[], seen: WeakSet<object>) {\n  if (seen.has(value)) {\n    return undefined;\n  }\n\n  seen.add(value);\n  const normalized = value.map((item) => {\n    const normalized = toJsonValueInternal(item, seen);\n    return normalized === undefined ? null : normalized;\n  });\n  seen.delete(value);\n\n  return normalized;\n}\n\nfunction normalizeJsonObject(\n  value: Record<string, unknown>,\n  seen: WeakSet<object>,\n): Record<string, JsonValue> {\n  const normalized: Record<string, JsonValue> = {};\n\n  if (seen.has(value)) {\n    return normalized;\n  }\n\n  seen.add(value);\n  try {\n    for (const [key, entryValue] of Object.entries(value)) {\n      const entry = toJsonValueInternal(entryValue, seen);\n      if (entry !== undefined) {\n        normalized[key] = entry;\n      }\n    }\n  } finally {\n    seen.delete(value);\n  }\n\n  return normalized;\n}\n\n/** Returns true when a value exposes a callable method with the given name. */\nexport function hasCallableMethod(value: unknown, methodName: string) {\n  return (\n    value !== null &&\n    (typeof value === \"object\" || typeof value === \"function\") &&\n    methodName in value &&\n    typeof (value as Record<string, unknown>)[methodName] === \"function\"\n  );\n}\n\n/** Normalizes an unknown value into the JSON-safe shape used by harness runs. */\nexport function toJsonValue(value: unknown): JsonValue | undefined {\n  return toJsonValueInternal(value, new WeakSet());\n}\n\nfunction toJsonValueInternal(\n  value: unknown,\n  seen: WeakSet<object>,\n): JsonValue | undefined {\n  if (isJsonPrimitive(value)) {\n    return value;\n  }\n\n  if (\n    value !== null &&\n    typeof value === \"object\" &&\n    seen.has(value as object)\n  ) {\n    return undefined;\n  }\n\n  if (Array.isArray(value)) {\n    return normalizeJsonArray(value, seen);\n  }\n\n  if (isJsonRecord(value)) {\n    return normalizeJsonObject(value, seen);\n  }\n\n  return undefined;\n}\n\n/** Drops non-JSON properties from a record while preserving valid values. */\nexport function normalizeRecord(\n  value: Record<string, unknown>,\n): Record<string, JsonValue> {\n  return normalizeJsonObject(value, new WeakSet());\n}\n\n/** Normalizes metadata and omits the field entirely when nothing survives. */\nexport function normalizeMetadata(\n  value: Record<string, unknown>,\n): Record<string, JsonValue> | undefined {\n  const normalized = normalizeRecord(value);\n  return Object.keys(normalized).length > 0 ? normalized : undefined;\n}\n\n/** Converts arbitrary content into the JSON-safe message content shape. */\nexport function normalizeContent(value: unknown): JsonValue {\n  const normalized = toJsonValue(value);\n  return normalized !== undefined ? normalized : String(value);\n}\n\n/**\n * Creates a harness from the common \"run app code and return output\" shape.\n *\n * @param options - Harness name plus the callback that executes app code.\n *\n * @example\n * ```ts\n * import { createHarness } from \"vitest-evals\";\n *\n * export const refundHarness = createHarness<\n *   string,\n *   { status: \"approved\" | \"denied\" }\n * >({\n *   name: \"refund-agent\",\n *   run: async ({ input, setArtifact }) => {\n *     const result = await runRefundFlow(input);\n *     const output = { status: result.status };\n *\n *     setArtifact(\"case\", { invoiceId: result.invoiceId });\n *\n *     return {\n *       output,\n *       toolCalls: result.toolCalls,\n *       usage: { provider: \"openai\", model: \"gpt-4o-mini\" },\n *     };\n *   },\n * });\n * ```\n */\nexport function createHarness<\n  TInput = unknown,\n  TOutput extends JsonValue | undefined = JsonValue | undefined,\n>(options: CreateHarnessOptions<TInput, TOutput>): Harness<TInput, TOutput>;\nexport function createHarness<\n  TInput = unknown,\n  TOutput extends JsonValue | undefined = JsonValue | undefined,\n>(options: CreateHarnessOptions<TInput, TOutput>): Harness<TInput, TOutput> {\n  const harness: Harness<TInput, TOutput> = {\n    name: options.name,\n    run: async (input, context) => {\n      const startedAt = new Date();\n\n      try {\n        const result = await options.run({\n          input,\n          signal: context.signal,\n          artifacts: context.artifacts,\n          setArtifact: context.setArtifact,\n        });\n        const run = normalizeHarnessRun(input, result, context);\n        ensureRunTrace(run, {\n          name: options.name,\n          startedAt,\n          finishedAt: new Date(),\n        });\n\n        return run;\n      } catch (error) {\n        const partialRun = getHarnessRunFromError(error);\n        if (partialRun) {\n          if (\n            Object.keys(context.artifacts).length > 0 &&\n            !partialRun.artifacts\n          ) {\n            partialRun.artifacts = context.artifacts;\n          }\n          ensureRunTrace(partialRun, {\n            name: options.name,\n            startedAt,\n            finishedAt: new Date(),\n          });\n          throw attachHarnessRunToError(error, partialRun);\n        }\n\n        const failedRun = createFailedHarnessRun(input, error, {\n          artifacts: context.artifacts,\n        });\n        ensureRunTrace(failedRun, {\n          name: options.name,\n          startedAt,\n          finishedAt: new Date(),\n        });\n\n        throw attachHarnessRunToError(error, failedRun);\n      }\n    },\n  };\n\n  return harness;\n}\n\n/**\n * Normalizes a lightweight harness result into the reporter-facing run shape.\n *\n * @param input - Original input passed to the harness.\n * @param result - Lightweight result or pre-normalized harness run.\n * @param context - Optional per-run context used to merge artifacts.\n *\n * @example\n * ```ts\n * const run = normalizeHarnessRun(\"Refund invoice inv_123\", {\n *   output: { status: \"approved\" },\n *   toolCalls: [{ name: \"lookupInvoice\", arguments: { invoiceId: \"inv_123\" } }],\n *   usage: { provider: \"openai\", model: \"gpt-4o-mini\" },\n * });\n *\n * expect(toolCalls(run.session)).toHaveLength(1);\n * ```\n */\nexport function normalizeHarnessRun<\n  TInput = unknown,\n  TOutput extends JsonValue | undefined = JsonValue | undefined,\n>(\n  input: TInput,\n  result: HarnessResultLike<TOutput>,\n  context?: HarnessContext,\n): HarnessRun<TOutput> {\n  if (isHarnessRun(result)) {\n    if (\n      context &&\n      Object.keys(context.artifacts).length > 0 &&\n      !result.artifacts\n    ) {\n      return {\n        ...result,\n        artifacts: context.artifacts,\n      };\n    }\n\n    return result;\n  }\n\n  const output = result.output;\n  const toolCalls = normalizeSimpleToolCalls(result.toolCalls);\n  const usage = result.usage ?? {};\n  const messages =\n    result.messages ??\n    createDefaultSessionMessages({\n      input,\n      output,\n      toolCalls,\n    });\n  const metadata = result.metadata\n    ? normalizeMetadata(result.metadata)\n    : undefined;\n  const artifacts = normalizeMergedArtifacts(\n    context?.artifacts,\n    result.artifacts,\n  );\n  const traces = normalizeSimpleTraces(result.traces);\n\n  return {\n    session: {\n      messages,\n      ...(usage.provider ? { provider: usage.provider } : {}),\n      ...(usage.model ? { model: usage.model } : {}),\n      ...(metadata ? { metadata } : {}),\n    },\n    ...(output !== undefined ? { output } : {}),\n    usage,\n    ...(result.timings ? { timings: result.timings } : {}),\n    ...(artifacts ? { artifacts } : {}),\n    ...(traces ? { traces } : {}),\n    errors: normalizeSimpleErrors(result.errors),\n  } as HarnessRun<TOutput>;\n}\n\n/**\n * Builds a JSON-safe failed run for errors that happen before a harness can return.\n *\n * @param input - Original input passed to the harness.\n * @param error - Error thrown by setup or execution.\n * @param options - Optional artifacts to preserve on the failed run.\n */\nexport function createFailedHarnessRun(\n  input: unknown,\n  error: unknown,\n  options: { artifacts?: Record<string, JsonValue> } = {},\n): HarnessRun {\n  const artifacts = options.artifacts;\n\n  return {\n    session: {\n      messages: [\n        {\n          role: \"user\",\n          content: normalizeContent(input),\n        },\n      ],\n    },\n    usage: {},\n    ...(artifacts && Object.keys(artifacts).length > 0 ? { artifacts } : {}),\n    errors: [serializeError(error)],\n  };\n}\n\nfunction createDefaultSessionMessages<TInput>({\n  input,\n  output,\n  toolCalls: normalizedToolCalls,\n}: {\n  input: TInput;\n  output: JsonValue | undefined;\n  toolCalls: ToolCallRecord[];\n}): NormalizedMessage[] {\n  const messages: NormalizedMessage[] = [\n    {\n      role: \"user\",\n      content: normalizeContent(input),\n    },\n  ];\n\n  if (output !== undefined || normalizedToolCalls.length > 0) {\n    messages.push({\n      role: \"assistant\",\n      ...(output !== undefined ? { content: normalizeContent(output) } : {}),\n      ...(normalizedToolCalls.length > 0\n        ? { toolCalls: normalizedToolCalls }\n        : {}),\n    });\n  }\n\n  return messages;\n}\n\nfunction normalizeSimpleToolCalls(\n  calls: SimpleToolCallRecord[] | undefined,\n): ToolCallRecord[] {\n  return (calls ?? []).map((call) => {\n    const {\n      arguments: rawArguments,\n      result: rawResult,\n      error: rawError,\n      metadata: rawMetadata,\n      ...toolCall\n    } = call;\n    const args = normalizeToolCallArguments(rawArguments);\n    const result = toJsonValue(rawResult);\n    const error = normalizeToolCallError(rawError);\n    const metadata = rawMetadata ? normalizeMetadata(rawMetadata) : undefined;\n\n    return {\n      ...toolCall,\n      ...(args ? { arguments: args } : {}),\n      ...(result !== undefined ? { result } : {}),\n      ...(error ? { error } : {}),\n      ...(metadata ? { metadata } : {}),\n    };\n  });\n}\n\nfunction normalizeToolCallArguments(\n  value: unknown,\n): Record<string, JsonValue> | undefined {\n  if (value === undefined) {\n    return undefined;\n  }\n\n  const normalized = toJsonValue(value);\n  return normalized &&\n    typeof normalized === \"object\" &&\n    !Array.isArray(normalized)\n    ? normalized\n    : undefined;\n}\n\nfunction normalizeToolCallError(\n  value: unknown,\n): ToolCallRecord[\"error\"] | undefined {\n  if (value === undefined) {\n    return undefined;\n  }\n\n  const serialized = serializeError(value);\n  const { message, type, ...details } = serialized;\n\n  return {\n    ...details,\n    message: typeof message === \"string\" ? message : String(message),\n    ...(typeof type === \"string\" ? { type } : {}),\n  };\n}\n\nfunction normalizeMergedArtifacts(\n  contextArtifacts: Record<string, JsonValue> | undefined,\n  resultArtifacts: Record<string, unknown> | undefined,\n) {\n  const artifacts = {\n    ...(contextArtifacts ?? {}),\n    ...(resultArtifacts ? normalizeRecord(resultArtifacts) : {}),\n  };\n\n  return Object.keys(artifacts).length > 0 ? artifacts : undefined;\n}\n\nfunction normalizeSimpleErrors(\n  errors: unknown[] | undefined,\n): Array<Record<string, JsonValue>> {\n  return (errors ?? []).map((error) => {\n    const normalized = toJsonValue(error);\n\n    if (\n      normalized &&\n      typeof normalized === \"object\" &&\n      !Array.isArray(normalized) &&\n      Object.keys(normalized).length > 0\n    ) {\n      return normalized;\n    }\n\n    return serializeError(error);\n  });\n}\n\nfunction normalizeSimpleTraces(\n  traces: SimpleTraceRecord[] | undefined,\n): NormalizedTrace[] | undefined {\n  if (!Array.isArray(traces)) {\n    return undefined;\n  }\n\n  const normalized = traces\n    .map(normalizeSimpleTrace)\n    .filter((trace): trace is NormalizedTrace => Boolean(trace));\n\n  return normalized.length > 0 ? normalized : undefined;\n}\n\nfunction normalizeSimpleTrace(trace: unknown): NormalizedTrace | undefined {\n  if (!isJsonRecord(trace)) {\n    return undefined;\n  }\n\n  const {\n    metadata: rawMetadata,\n    spans: rawSpans,\n    ...traceFields\n  } = trace as Partial<SimpleTraceRecord>;\n  const spans = (Array.isArray(rawSpans) ? rawSpans : [])\n    .map((span) => normalizeSimpleSpan(span))\n    .filter((span): span is NormalizedSpan => Boolean(span));\n  const metadata = isJsonRecord(rawMetadata)\n    ? normalizeMetadata(rawMetadata)\n    : undefined;\n\n  if (spans.length === 0 && !traceFields.id && !traceFields.name) {\n    return undefined;\n  }\n\n  return {\n    ...traceFields,\n    ...(metadata ? { metadata } : {}),\n    spans,\n  };\n}\n\nfunction normalizeSimpleSpan(span: unknown): NormalizedSpan | undefined {\n  if (!isJsonRecord(span) || typeof span.name !== \"string\" || !span.name) {\n    return undefined;\n  }\n\n  const {\n    attributes: rawAttributes,\n    error: rawError,\n    events: rawEvents,\n    ...spanFields\n  } = span as Partial<SimpleSpanRecord> & { name: string };\n  const attributes = rawAttributes\n    ? isJsonRecord(rawAttributes)\n      ? normalizeMetadata(rawAttributes)\n      : undefined\n    : undefined;\n  const error = normalizeSpanError(rawError);\n  const events = normalizeSimpleSpanEvents(rawEvents);\n\n  return {\n    ...spanFields,\n    ...(attributes\n      ? { attributes: attributes as NormalizedSpanAttributes }\n      : {}),\n    ...(error ? { error } : {}),\n    ...(events ? { events } : {}),\n  };\n}\n\nfunction normalizeSimpleSpanEvents(\n  events: unknown,\n): NormalizedSpanEvent[] | undefined {\n  if (!Array.isArray(events)) {\n    return undefined;\n  }\n\n  const normalized = events\n    .map(normalizeSimpleSpanEvent)\n    .filter((event): event is NormalizedSpanEvent => Boolean(event));\n\n  return normalized.length > 0 ? normalized : undefined;\n}\n\nfunction normalizeSimpleSpanEvent(\n  event: unknown,\n): NormalizedSpanEvent | undefined {\n  if (!isJsonRecord(event) || typeof event.name !== \"string\" || !event.name) {\n    return undefined;\n  }\n\n  const { attributes: rawAttributes, ...eventFields } =\n    event as Partial<SimpleSpanEvent> & { name: string };\n  const attributes = rawAttributes\n    ? isJsonRecord(rawAttributes)\n      ? normalizeMetadata(rawAttributes)\n      : undefined\n    : undefined;\n\n  return {\n    ...eventFields,\n    ...(attributes\n      ? { attributes: attributes as NormalizedSpanAttributes }\n      : {}),\n  };\n}\n\n/** Normalizes arbitrary span errors while preserving object-shaped messages. */\nexport function normalizeSpanError(\n  error: unknown,\n): NormalizedSpan[\"error\"] | undefined {\n  if (error === undefined) {\n    return undefined;\n  }\n\n  if (error instanceof Error) {\n    const details = normalizeMetadata(\n      error as unknown as Record<string, unknown>,\n    );\n\n    return {\n      ...(details ?? {}),\n      type: error.name,\n      message: error.message,\n    };\n  }\n\n  if (\n    error &&\n    typeof error === \"object\" &&\n    !Array.isArray(error) &&\n    typeof (error as { message?: unknown }).message === \"string\"\n  ) {\n    const normalized = normalizeMetadata(error as Record<string, unknown>);\n    const { message, type, ...details } = normalized ?? {};\n\n    return {\n      ...details,\n      message: message as string,\n      ...(typeof type === \"string\" ? { type } : {}),\n    };\n  }\n\n  const serialized = serializeError(error);\n  const { message, type, ...details } = serialized;\n\n  return {\n    ...details,\n    message: typeof message === \"string\" ? message : String(message),\n    ...(typeof type === \"string\" ? { type } : {}),\n  };\n}\n\n/** Normalizes raw span attributes into the JSON-safe span attribute shape. */\nexport function normalizeSpanAttributes(\n  attributes: Record<string, unknown>,\n): NormalizedSpanAttributes | undefined {\n  return normalizeMetadata(attributes) as NormalizedSpanAttributes | undefined;\n}\n\n/** Builds common OpenTelemetry GenAI usage attributes from a usage summary. */\nexport function createGenAiUsageAttributes(\n  usage: UsageSummary | undefined,\n  options: { provider?: string } = {},\n) {\n  return {\n    \"gen_ai.provider.name\": usage?.provider ?? options.provider,\n    \"gen_ai.request.model\": usage?.model,\n    \"gen_ai.response.model\": usage?.model,\n    \"gen_ai.usage.input_tokens\": usage?.inputTokens,\n    \"gen_ai.usage.output_tokens\": usage?.outputTokens,\n    \"gen_ai.usage.reasoning.output_tokens\": usage?.reasoningTokens,\n  } satisfies Record<string, unknown>;\n}\n\n/**\n * Converts normalized tool-call records into trace spans.\n *\n * Tool-call ids are preserved as GenAI attributes. Pass `spanIdPrefix` when the\n * spans belong to a known trace so span ids stay internally unique.\n */\nexport function createToolCallSpans(\n  calls: ToolCallRecord[],\n  options: CreateToolCallSpansOptions = {},\n): NormalizedSpan[] {\n  return calls.map((call, index) => {\n    const spanError = call.error ? normalizeSpanError(call.error) : undefined;\n    const spanId = options.spanIdPrefix\n      ? `${options.spanIdPrefix}:${index + 1}`\n      : call.id;\n\n    return {\n      ...(spanId ? { id: spanId } : {}),\n      ...(options.traceId ? { traceId: options.traceId } : {}),\n      ...(options.parentId ? { parentId: options.parentId } : {}),\n      name: call.name,\n      kind: \"tool\",\n      ...(call.startedAt ? { startedAt: call.startedAt } : {}),\n      ...(call.finishedAt ? { finishedAt: call.finishedAt } : {}),\n      ...(call.durationMs !== undefined ? { durationMs: call.durationMs } : {}),\n      status: spanError ? \"error\" : \"ok\",\n      ...(spanError ? { error: spanError } : {}),\n      attributes: normalizeSpanAttributes({\n        \"gen_ai.operation.name\": \"execute_tool\",\n        \"gen_ai.tool.name\": call.name,\n        \"gen_ai.tool.type\": \"function\",\n        ...(call.id ? { \"gen_ai.tool.call.id\": call.id } : {}),\n        ...(call.arguments !== undefined\n          ? { \"gen_ai.tool.call.arguments\": call.arguments }\n          : {}),\n        ...(call.result !== undefined\n          ? { \"gen_ai.tool.call.result\": call.result }\n          : {}),\n      }),\n    } satisfies NormalizedSpan;\n  });\n}\n\n/**\n * Attaches a fallback run trace when a harness result does not already contain spans.\n *\n * This keeps custom harnesses inspectable while first-party harness packages\n * remain free to attach richer native traces.\n */\nexport function ensureRunTrace(\n  run: HarnessRun,\n  options: EnsureRunTraceOptions,\n): NormalizedTrace | undefined {\n  if (spans(run).length > 0) {\n    return undefined;\n  }\n\n  const traceId = options.id ?? createGeneratedTraceId();\n  const rootSpanId = `${traceId}:run`;\n  const durationMs = options.finishedAt.getTime() - options.startedAt.getTime();\n  const rootError =\n    run.errors.length > 0 ? normalizeSpanError(run.errors[0]) : undefined;\n  const runSpan: NormalizedSpan = {\n    id: rootSpanId,\n    traceId,\n    name: options.name,\n    kind: \"run\",\n    startedAt: options.startedAt.toISOString(),\n    finishedAt: options.finishedAt.toISOString(),\n    durationMs,\n    status: rootError ? \"error\" : \"ok\",\n    ...(rootError ? { error: rootError } : {}),\n    attributes: normalizeSpanAttributes({\n      \"gen_ai.operation.name\": options.operationName ?? \"invoke_workflow\",\n      \"gen_ai.workflow.name\": options.name,\n      ...createGenAiUsageAttributes(run.usage),\n    }),\n  };\n  const toolSpans = createToolCallSpans(toolCalls(run.session), {\n    traceId,\n    parentId: rootSpanId,\n    spanIdPrefix: `${traceId}:tool`,\n  });\n  const trace: NormalizedTrace = {\n    id: traceId,\n    name: options.name,\n    startedAt: options.startedAt.toISOString(),\n    finishedAt: options.finishedAt.toISOString(),\n    durationMs,\n    ...(options.source ? { metadata: { source: options.source } } : {}),\n    spans: [runSpan, ...toolSpans],\n  };\n\n  run.traces = [trace];\n  return trace;\n}\n\nlet nextGeneratedTraceId = 0;\n\nfunction createGeneratedTraceId() {\n  nextGeneratedTraceId += 1;\n  return `trace_${nextGeneratedTraceId}`;\n}\n\n/**\n * Attaches a partial or complete harness run to an arbitrary thrown error.\n *\n * @param error - Thrown value to wrap.\n * @param run - Partial or complete normalized harness run to preserve.\n *\n * @example\n * ```ts\n * try {\n *   return await runAgent(input);\n * } catch (error) {\n *   throw attachHarnessRunToError(error, partialRun);\n * }\n * ```\n */\nexport function attachHarnessRunToError(\n  error: unknown,\n  run: HarnessRun,\n): HarnessRunError {\n  const baseError =\n    error instanceof Error\n      ? error\n      : new Error(String(error ?? \"Unknown error\"));\n  return Object.assign(baseError, {\n    vitestEvalsRun: run,\n  });\n}\n\n/**\n * Reads an attached harness run back off a previously wrapped error value.\n *\n * @param error - Unknown thrown value that may contain a harness run.\n *\n * @example\n * ```ts\n * const partialRun = getHarnessRunFromError(error);\n *\n * if (partialRun) {\n *   console.log(toolCalls(partialRun.session));\n * }\n * ```\n */\nexport function getHarnessRunFromError(error: unknown): HarnessRun | undefined {\n  if (\n    error &&\n    typeof error === \"object\" &&\n    \"vitestEvalsRun\" in error &&\n    isHarnessRun((error as { vitestEvalsRun?: unknown }).vitestEvalsRun)\n  ) {\n    return (error as { vitestEvalsRun: HarnessRun }).vitestEvalsRun;\n  }\n\n  return undefined;\n}\n\n/** Returns true when a value matches the normalized `HarnessRun` contract. */\nexport function isHarnessRun(value: unknown): value is HarnessRun {\n  if (!value || typeof value !== \"object\") {\n    return false;\n  }\n\n  const candidate = value as {\n    session?: unknown;\n    usage?: unknown;\n    errors?: unknown;\n  };\n\n  return (\n    isNormalizedSession(candidate.session) &&\n    Boolean(candidate.usage) &&\n    typeof candidate.usage === \"object\" &&\n    !Array.isArray(candidate.usage) &&\n    Array.isArray(candidate.errors)\n  );\n}\n\n/** Returns true when a value matches the normalized session contract. */\nexport function isNormalizedSession(\n  value: unknown,\n): value is NormalizedSession {\n  return (\n    Boolean(value) &&\n    typeof value === \"object\" &&\n    value !== null &&\n    \"messages\" in value &&\n    Array.isArray((value as { messages?: unknown }).messages)\n  );\n}\n\n/** Reuses pre-normalized harness errors when a runtime already returns them. */\nexport function resolveHarnessRunErrors(\n  result: unknown,\n): Array<Record<string, JsonValue>> {\n  if (\n    result &&\n    typeof result === \"object\" &&\n    Array.isArray((result as Record<string, unknown>).errors)\n  ) {\n    return (result as { errors: Array<Record<string, JsonValue>> }).errors;\n  }\n\n  return [];\n}\n\n/** Serializes an arbitrary thrown value into the normalized error shape. */\nexport function serializeError(error: unknown): Record<string, JsonValue> {\n  if (error instanceof Error) {\n    return {\n      type: error.name,\n      message: error.message,\n    };\n  }\n\n  return {\n    type: \"Error\",\n    message: String(error),\n  };\n}\n"],"mappings":";AAAA;AAAA,EAKE;AAAA,EAGA;AAAA,OAGK;AAkBP;AAAA,EACE,qBAAAA;AAAA,EACA,eAAAC;AAAA,EACA,iCAAAC;AAAA,EACA,kBAAAC;AAAA,EACA,SAAAC;AAAA,EACA,eAAAC;AAAA,EACA,kBAAAC;AAAA,EACA,aAAAC;AAAA,EACA,gBAAAC;AAAA,EACA,gBAAAC;AAAA,OACK;AAuOP,SAAS,gBAAgB,OAAwC;AAC/D,SACE,UAAU,QACV,OAAO,UAAU,YACjB,OAAO,UAAU,aAChB,OAAO,UAAU,YAAY,OAAO,SAAS,KAAK;AAEvD;AAEA,SAAS,aAAa,OAAkD;AACtE,SAAO,OAAO,UAAU,YAAY,UAAU,QAAQ,CAAC,MAAM,QAAQ,KAAK;AAC5E;AAEA,SAAS,mBAAmB,OAAkB,MAAuB;AACnE,MAAI,KAAK,IAAI,KAAK,GAAG;AACnB,WAAO;AAAA,EACT;AAEA,OAAK,IAAI,KAAK;AACd,QAAM,aAAa,MAAM,IAAI,CAAC,SAAS;AACrC,UAAMC,cAAa,oBAAoB,MAAM,IAAI;AACjD,WAAOA,gBAAe,SAAY,OAAOA;AAAA,EAC3C,CAAC;AACD,OAAK,OAAO,KAAK;AAEjB,SAAO;AACT;AAEA,SAAS,oBACP,OACA,MAC2B;AAC3B,QAAM,aAAwC,CAAC;AAE/C,MAAI,KAAK,IAAI,KAAK,GAAG;AACnB,WAAO;AAAA,EACT;AAEA,OAAK,IAAI,KAAK;AACd,MAAI;AACF,eAAW,CAAC,KAAK,UAAU,KAAK,OAAO,QAAQ,KAAK,GAAG;AACrD,YAAM,QAAQ,oBAAoB,YAAY,IAAI;AAClD,UAAI,UAAU,QAAW;AACvB,mBAAW,GAAG,IAAI;AAAA,MACpB;AAAA,IACF;AAAA,EACF,UAAE;AACA,SAAK,OAAO,KAAK;AAAA,EACnB;AAEA,SAAO;AACT;AAGO,SAAS,kBAAkB,OAAgB,YAAoB;AACpE,SACE,UAAU,SACT,OAAO,UAAU,YAAY,OAAO,UAAU,eAC/C,cAAc,SACd,OAAQ,MAAkC,UAAU,MAAM;AAE9D;AAGO,SAAS,YAAY,OAAuC;AACjE,SAAO,oBAAoB,OAAO,oBAAI,QAAQ,CAAC;AACjD;AAEA,SAAS,oBACP,OACA,MACuB;AACvB,MAAI,gBAAgB,KAAK,GAAG;AAC1B,WAAO;AAAA,EACT;AAEA,MACE,UAAU,QACV,OAAO,UAAU,YACjB,KAAK,IAAI,KAAe,GACxB;AACA,WAAO;AAAA,EACT;AAEA,MAAI,MAAM,QAAQ,KAAK,GAAG;AACxB,WAAO,mBAAmB,OAAO,IAAI;AAAA,EACvC;AAEA,MAAI,aAAa,KAAK,GAAG;AACvB,WAAO,oBAAoB,OAAO,IAAI;AAAA,EACxC;AAEA,SAAO;AACT;AAGO,SAAS,gBACd,OAC2B;AAC3B,SAAO,oBAAoB,OAAO,oBAAI,QAAQ,CAAC;AACjD;AAGO,SAAS,kBACd,OACuC;AACvC,QAAM,aAAa,gBAAgB,KAAK;AACxC,SAAO,OAAO,KAAK,UAAU,EAAE,SAAS,IAAI,aAAa;AAC3D;AAGO,SAAS,iBAAiB,OAA2B;AAC1D,QAAM,aAAa,YAAY,KAAK;AACpC,SAAO,eAAe,SAAY,aAAa,OAAO,KAAK;AAC7D;AAmCO,SAAS,cAGd,SAA0E;AAC1E,QAAM,UAAoC;AAAA,IACxC,MAAM,QAAQ;AAAA,IACd,KAAK,OAAO,OAAO,YAAY;AAC7B,YAAM,YAAY,oBAAI,KAAK;AAE3B,UAAI;AACF,cAAM,SAAS,MAAM,QAAQ,IAAI;AAAA,UAC/B;AAAA,UACA,QAAQ,QAAQ;AAAA,UAChB,WAAW,QAAQ;AAAA,UACnB,aAAa,QAAQ;AAAA,QACvB,CAAC;AACD,cAAM,MAAM,oBAAoB,OAAO,QAAQ,OAAO;AACtD,uBAAe,KAAK;AAAA,UAClB,MAAM,QAAQ;AAAA,UACd;AAAA,UACA,YAAY,oBAAI,KAAK;AAAA,QACvB,CAAC;AAED,eAAO;AAAA,MACT,SAAS,OAAO;AACd,cAAM,aAAa,uBAAuB,KAAK;AAC/C,YAAI,YAAY;AACd,cACE,OAAO,KAAK,QAAQ,SAAS,EAAE,SAAS,KACxC,CAAC,WAAW,WACZ;AACA,uBAAW,YAAY,QAAQ;AAAA,UACjC;AACA,yBAAe,YAAY;AAAA,YACzB,MAAM,QAAQ;AAAA,YACd;AAAA,YACA,YAAY,oBAAI,KAAK;AAAA,UACvB,CAAC;AACD,gBAAM,wBAAwB,OAAO,UAAU;AAAA,QACjD;AAEA,cAAM,YAAY,uBAAuB,OAAO,OAAO;AAAA,UACrD,WAAW,QAAQ;AAAA,QACrB,CAAC;AACD,uBAAe,WAAW;AAAA,UACxB,MAAM,QAAQ;AAAA,UACd;AAAA,UACA,YAAY,oBAAI,KAAK;AAAA,QACvB,CAAC;AAED,cAAM,wBAAwB,OAAO,SAAS;AAAA,MAChD;AAAA,IACF;AAAA,EACF;AAEA,SAAO;AACT;AAoBO,SAAS,oBAId,OACA,QACA,SACqB;AACrB,MAAI,aAAa,MAAM,GAAG;AACxB,QACE,WACA,OAAO,KAAK,QAAQ,SAAS,EAAE,SAAS,KACxC,CAAC,OAAO,WACR;AACA,aAAO;AAAA,QACL,GAAG;AAAA,QACH,WAAW,QAAQ;AAAA,MACrB;AAAA,IACF;AAEA,WAAO;AAAA,EACT;AAEA,QAAM,SAAS,OAAO;AACtB,QAAMH,aAAY,yBAAyB,OAAO,SAAS;AAC3D,QAAM,QAAQ,OAAO,SAAS,CAAC;AAC/B,QAAM,WACJ,OAAO,YACP,6BAA6B;AAAA,IAC3B;AAAA,IACA;AAAA,IACA,WAAAA;AAAA,EACF,CAAC;AACH,QAAM,WAAW,OAAO,WACpB,kBAAkB,OAAO,QAAQ,IACjC;AACJ,QAAM,YAAY;AAAA,IAChB,SAAS;AAAA,IACT,OAAO;AAAA,EACT;AACA,QAAM,SAAS,sBAAsB,OAAO,MAAM;AAElD,SAAO;AAAA,IACL,SAAS;AAAA,MACP;AAAA,MACA,GAAI,MAAM,WAAW,EAAE,UAAU,MAAM,SAAS,IAAI,CAAC;AAAA,MACrD,GAAI,MAAM,QAAQ,EAAE,OAAO,MAAM,MAAM,IAAI,CAAC;AAAA,MAC5C,GAAI,WAAW,EAAE,SAAS,IAAI,CAAC;AAAA,IACjC;AAAA,IACA,GAAI,WAAW,SAAY,EAAE,OAAO,IAAI,CAAC;AAAA,IACzC;AAAA,IACA,GAAI,OAAO,UAAU,EAAE,SAAS,OAAO,QAAQ,IAAI,CAAC;AAAA,IACpD,GAAI,YAAY,EAAE,UAAU,IAAI,CAAC;AAAA,IACjC,GAAI,SAAS,EAAE,OAAO,IAAI,CAAC;AAAA,IAC3B,QAAQ,sBAAsB,OAAO,MAAM;AAAA,EAC7C;AACF;AASO,SAAS,uBACd,OACA,OACA,UAAqD,CAAC,GAC1C;AACZ,QAAM,YAAY,QAAQ;AAE1B,SAAO;AAAA,IACL,SAAS;AAAA,MACP,UAAU;AAAA,QACR;AAAA,UACE,MAAM;AAAA,UACN,SAAS,iBAAiB,KAAK;AAAA,QACjC;AAAA,MACF;AAAA,IACF;AAAA,IACA,OAAO,CAAC;AAAA,IACR,GAAI,aAAa,OAAO,KAAK,SAAS,EAAE,SAAS,IAAI,EAAE,UAAU,IAAI,CAAC;AAAA,IACtE,QAAQ,CAAC,eAAe,KAAK,CAAC;AAAA,EAChC;AACF;AAEA,SAAS,6BAAqC;AAAA,EAC5C;AAAA,EACA;AAAA,EACA,WAAW;AACb,GAIwB;AACtB,QAAM,WAAgC;AAAA,IACpC;AAAA,MACE,MAAM;AAAA,MACN,SAAS,iBAAiB,KAAK;AAAA,IACjC;AAAA,EACF;AAEA,MAAI,WAAW,UAAa,oBAAoB,SAAS,GAAG;AAC1D,aAAS,KAAK;AAAA,MACZ,MAAM;AAAA,MACN,GAAI,WAAW,SAAY,EAAE,SAAS,iBAAiB,MAAM,EAAE,IAAI,CAAC;AAAA,MACpE,GAAI,oBAAoB,SAAS,IAC7B,EAAE,WAAW,oBAAoB,IACjC,CAAC;AAAA,IACP,CAAC;AAAA,EACH;AAEA,SAAO;AACT;AAEA,SAAS,yBACP,OACkB;AAClB,UAAQ,SAAS,CAAC,GAAG,IAAI,CAAC,SAAS;AACjC,UAAM;AAAA,MACJ,WAAW;AAAA,MACX,QAAQ;AAAA,MACR,OAAO;AAAA,MACP,UAAU;AAAA,MACV,GAAG;AAAA,IACL,IAAI;AACJ,UAAM,OAAO,2BAA2B,YAAY;AACpD,UAAM,SAAS,YAAY,SAAS;AACpC,UAAM,QAAQ,uBAAuB,QAAQ;AAC7C,UAAM,WAAW,cAAc,kBAAkB,WAAW,IAAI;AAEhE,WAAO;AAAA,MACL,GAAG;AAAA,MACH,GAAI,OAAO,EAAE,WAAW,KAAK,IAAI,CAAC;AAAA,MAClC,GAAI,WAAW,SAAY,EAAE,OAAO,IAAI,CAAC;AAAA,MACzC,GAAI,QAAQ,EAAE,MAAM,IAAI,CAAC;AAAA,MACzB,GAAI,WAAW,EAAE,SAAS,IAAI,CAAC;AAAA,IACjC;AAAA,EACF,CAAC;AACH;AAEA,SAAS,2BACP,OACuC;AACvC,MAAI,UAAU,QAAW;AACvB,WAAO;AAAA,EACT;AAEA,QAAM,aAAa,YAAY,KAAK;AACpC,SAAO,cACL,OAAO,eAAe,YACtB,CAAC,MAAM,QAAQ,UAAU,IACvB,aACA;AACN;AAEA,SAAS,uBACP,OACqC;AACrC,MAAI,UAAU,QAAW;AACvB,WAAO;AAAA,EACT;AAEA,QAAM,aAAa,eAAe,KAAK;AACvC,QAAM,EAAE,SAAS,MAAM,GAAG,QAAQ,IAAI;AAEtC,SAAO;AAAA,IACL,GAAG;AAAA,IACH,SAAS,OAAO,YAAY,WAAW,UAAU,OAAO,OAAO;AAAA,IAC/D,GAAI,OAAO,SAAS,WAAW,EAAE,KAAK,IAAI,CAAC;AAAA,EAC7C;AACF;AAEA,SAAS,yBACP,kBACA,iBACA;AACA,QAAM,YAAY;AAAA,IAChB,GAAI,oBAAoB,CAAC;AAAA,IACzB,GAAI,kBAAkB,gBAAgB,eAAe,IAAI,CAAC;AAAA,EAC5D;AAEA,SAAO,OAAO,KAAK,SAAS,EAAE,SAAS,IAAI,YAAY;AACzD;AAEA,SAAS,sBACP,QACkC;AAClC,UAAQ,UAAU,CAAC,GAAG,IAAI,CAAC,UAAU;AACnC,UAAM,aAAa,YAAY,KAAK;AAEpC,QACE,cACA,OAAO,eAAe,YACtB,CAAC,MAAM,QAAQ,UAAU,KACzB,OAAO,KAAK,UAAU,EAAE,SAAS,GACjC;AACA,aAAO;AAAA,IACT;AAEA,WAAO,eAAe,KAAK;AAAA,EAC7B,CAAC;AACH;AAEA,SAAS,sBACP,QAC+B;AAC/B,MAAI,CAAC,MAAM,QAAQ,MAAM,GAAG;AAC1B,WAAO;AAAA,EACT;AAEA,QAAM,aAAa,OAChB,IAAI,oBAAoB,EACxB,OAAO,CAAC,UAAoC,QAAQ,KAAK,CAAC;AAE7D,SAAO,WAAW,SAAS,IAAI,aAAa;AAC9C;AAEA,SAAS,qBAAqB,OAA6C;AACzE,MAAI,CAAC,aAAa,KAAK,GAAG;AACxB,WAAO;AAAA,EACT;AAEA,QAAM;AAAA,IACJ,UAAU;AAAA,IACV,OAAO;AAAA,IACP,GAAG;AAAA,EACL,IAAI;AACJ,QAAMH,UAAS,MAAM,QAAQ,QAAQ,IAAI,WAAW,CAAC,GAClD,IAAI,CAAC,SAAS,oBAAoB,IAAI,CAAC,EACvC,OAAO,CAAC,SAAiC,QAAQ,IAAI,CAAC;AACzD,QAAM,WAAW,aAAa,WAAW,IACrC,kBAAkB,WAAW,IAC7B;AAEJ,MAAIA,OAAM,WAAW,KAAK,CAAC,YAAY,MAAM,CAAC,YAAY,MAAM;AAC9D,WAAO;AAAA,EACT;AAEA,SAAO;AAAA,IACL,GAAG;AAAA,IACH,GAAI,WAAW,EAAE,SAAS,IAAI,CAAC;AAAA,IAC/B,OAAAA;AAAA,EACF;AACF;AAEA,SAAS,oBAAoB,MAA2C;AACtE,MAAI,CAAC,aAAa,IAAI,KAAK,OAAO,KAAK,SAAS,YAAY,CAAC,KAAK,MAAM;AACtE,WAAO;AAAA,EACT;AAEA,QAAM;AAAA,IACJ,YAAY;AAAA,IACZ,OAAO;AAAA,IACP,QAAQ;AAAA,IACR,GAAG;AAAA,EACL,IAAI;AACJ,QAAM,aAAa,gBACf,aAAa,aAAa,IACxB,kBAAkB,aAAa,IAC/B,SACF;AACJ,QAAM,QAAQ,mBAAmB,QAAQ;AACzC,QAAM,SAAS,0BAA0B,SAAS;AAElD,SAAO;AAAA,IACL,GAAG;AAAA,IACH,GAAI,aACA,EAAE,WAAmD,IACrD,CAAC;AAAA,IACL,GAAI,QAAQ,EAAE,MAAM,IAAI,CAAC;AAAA,IACzB,GAAI,SAAS,EAAE,OAAO,IAAI,CAAC;AAAA,EAC7B;AACF;AAEA,SAAS,0BACP,QACmC;AACnC,MAAI,CAAC,MAAM,QAAQ,MAAM,GAAG;AAC1B,WAAO;AAAA,EACT;AAEA,QAAM,aAAa,OAChB,IAAI,wBAAwB,EAC5B,OAAO,CAAC,UAAwC,QAAQ,KAAK,CAAC;AAEjE,SAAO,WAAW,SAAS,IAAI,aAAa;AAC9C;AAEA,SAAS,yBACP,OACiC;AACjC,MAAI,CAAC,aAAa,KAAK,KAAK,OAAO,MAAM,SAAS,YAAY,CAAC,MAAM,MAAM;AACzE,WAAO;AAAA,EACT;AAEA,QAAM,EAAE,YAAY,eAAe,GAAG,YAAY,IAChD;AACF,QAAM,aAAa,gBACf,aAAa,aAAa,IACxB,kBAAkB,aAAa,IAC/B,SACF;AAEJ,SAAO;AAAA,IACL,GAAG;AAAA,IACH,GAAI,aACA,EAAE,WAAmD,IACrD,CAAC;AAAA,EACP;AACF;AAGO,SAAS,mBACd,OACqC;AACrC,MAAI,UAAU,QAAW;AACvB,WAAO;AAAA,EACT;AAEA,MAAI,iBAAiB,OAAO;AAC1B,UAAMO,WAAU;AAAA,MACd;AAAA,IACF;AAEA,WAAO;AAAA,MACL,GAAIA,YAAW,CAAC;AAAA,MAChB,MAAM,MAAM;AAAA,MACZ,SAAS,MAAM;AAAA,IACjB;AAAA,EACF;AAEA,MACE,SACA,OAAO,UAAU,YACjB,CAAC,MAAM,QAAQ,KAAK,KACpB,OAAQ,MAAgC,YAAY,UACpD;AACA,UAAM,aAAa,kBAAkB,KAAgC;AACrE,UAAM,EAAE,SAAAC,UAAS,MAAAC,OAAM,GAAGF,SAAQ,IAAI,cAAc,CAAC;AAErD,WAAO;AAAA,MACL,GAAGA;AAAA,MACH,SAASC;AAAA,MACT,GAAI,OAAOC,UAAS,WAAW,EAAE,MAAAA,MAAK,IAAI,CAAC;AAAA,IAC7C;AAAA,EACF;AAEA,QAAM,aAAa,eAAe,KAAK;AACvC,QAAM,EAAE,SAAS,MAAM,GAAG,QAAQ,IAAI;AAEtC,SAAO;AAAA,IACL,GAAG;AAAA,IACH,SAAS,OAAO,YAAY,WAAW,UAAU,OAAO,OAAO;AAAA,IAC/D,GAAI,OAAO,SAAS,WAAW,EAAE,KAAK,IAAI,CAAC;AAAA,EAC7C;AACF;AAGO,SAAS,wBACd,YACsC;AACtC,SAAO,kBAAkB,UAAU;AACrC;AAGO,SAAS,2BACd,OACA,UAAiC,CAAC,GAClC;AACA,SAAO;AAAA,IACL,wBAAwB,OAAO,YAAY,QAAQ;AAAA,IACnD,wBAAwB,OAAO;AAAA,IAC/B,yBAAyB,OAAO;AAAA,IAChC,6BAA6B,OAAO;AAAA,IACpC,8BAA8B,OAAO;AAAA,IACrC,wCAAwC,OAAO;AAAA,EACjD;AACF;AAQO,SAAS,oBACd,OACA,UAAsC,CAAC,GACrB;AAClB,SAAO,MAAM,IAAI,CAAC,MAAM,UAAU;AAChC,UAAM,YAAY,KAAK,QAAQ,mBAAmB,KAAK,KAAK,IAAI;AAChE,UAAM,SAAS,QAAQ,eACnB,GAAG,QAAQ,YAAY,IAAI,QAAQ,CAAC,KACpC,KAAK;AAET,WAAO;AAAA,MACL,GAAI,SAAS,EAAE,IAAI,OAAO,IAAI,CAAC;AAAA,MAC/B,GAAI,QAAQ,UAAU,EAAE,SAAS,QAAQ,QAAQ,IAAI,CAAC;AAAA,MACtD,GAAI,QAAQ,WAAW,EAAE,UAAU,QAAQ,SAAS,IAAI,CAAC;AAAA,MACzD,MAAM,KAAK;AAAA,MACX,MAAM;AAAA,MACN,GAAI,KAAK,YAAY,EAAE,WAAW,KAAK,UAAU,IAAI,CAAC;AAAA,MACtD,GAAI,KAAK,aAAa,EAAE,YAAY,KAAK,WAAW,IAAI,CAAC;AAAA,MACzD,GAAI,KAAK,eAAe,SAAY,EAAE,YAAY,KAAK,WAAW,IAAI,CAAC;AAAA,MACvE,QAAQ,YAAY,UAAU;AAAA,MAC9B,GAAI,YAAY,EAAE,OAAO,UAAU,IAAI,CAAC;AAAA,MACxC,YAAY,wBAAwB;AAAA,QAClC,yBAAyB;AAAA,QACzB,oBAAoB,KAAK;AAAA,QACzB,oBAAoB;AAAA,QACpB,GAAI,KAAK,KAAK,EAAE,uBAAuB,KAAK,GAAG,IAAI,CAAC;AAAA,QACpD,GAAI,KAAK,cAAc,SACnB,EAAE,8BAA8B,KAAK,UAAU,IAC/C,CAAC;AAAA,QACL,GAAI,KAAK,WAAW,SAChB,EAAE,2BAA2B,KAAK,OAAO,IACzC,CAAC;AAAA,MACP,CAAC;AAAA,IACH;AAAA,EACF,CAAC;AACH;AAQO,SAAS,eACd,KACA,SAC6B;AAC7B,MAAI,MAAM,GAAG,EAAE,SAAS,GAAG;AACzB,WAAO;AAAA,EACT;AAEA,QAAM,UAAU,QAAQ,MAAM,uBAAuB;AACrD,QAAM,aAAa,GAAG,OAAO;AAC7B,QAAM,aAAa,QAAQ,WAAW,QAAQ,IAAI,QAAQ,UAAU,QAAQ;AAC5E,QAAM,YACJ,IAAI,OAAO,SAAS,IAAI,mBAAmB,IAAI,OAAO,CAAC,CAAC,IAAI;AAC9D,QAAM,UAA0B;AAAA,IAC9B,IAAI;AAAA,IACJ;AAAA,IACA,MAAM,QAAQ;AAAA,IACd,MAAM;AAAA,IACN,WAAW,QAAQ,UAAU,YAAY;AAAA,IACzC,YAAY,QAAQ,WAAW,YAAY;AAAA,IAC3C;AAAA,IACA,QAAQ,YAAY,UAAU;AAAA,IAC9B,GAAI,YAAY,EAAE,OAAO,UAAU,IAAI,CAAC;AAAA,IACxC,YAAY,wBAAwB;AAAA,MAClC,yBAAyB,QAAQ,iBAAiB;AAAA,MAClD,wBAAwB,QAAQ;AAAA,MAChC,GAAG,2BAA2B,IAAI,KAAK;AAAA,IACzC,CAAC;AAAA,EACH;AACA,QAAM,YAAY,oBAAoB,UAAU,IAAI,OAAO,GAAG;AAAA,IAC5D;AAAA,IACA,UAAU;AAAA,IACV,cAAc,GAAG,OAAO;AAAA,EAC1B,CAAC;AACD,QAAM,QAAyB;AAAA,IAC7B,IAAI;AAAA,IACJ,MAAM,QAAQ;AAAA,IACd,WAAW,QAAQ,UAAU,YAAY;AAAA,IACzC,YAAY,QAAQ,WAAW,YAAY;AAAA,IAC3C;AAAA,IACA,GAAI,QAAQ,SAAS,EAAE,UAAU,EAAE,QAAQ,QAAQ,OAAO,EAAE,IAAI,CAAC;AAAA,IACjE,OAAO,CAAC,SAAS,GAAG,SAAS;AAAA,EAC/B;AAEA,MAAI,SAAS,CAAC,KAAK;AACnB,SAAO;AACT;AAEA,IAAI,uBAAuB;AAE3B,SAAS,yBAAyB;AAChC,0BAAwB;AACxB,SAAO,SAAS,oBAAoB;AACtC;AAiBO,SAAS,wBACd,OACA,KACiB;AACjB,QAAM,YACJ,iBAAiB,QACb,QACA,IAAI,MAAM,OAAO,SAAS,eAAe,CAAC;AAChD,SAAO,OAAO,OAAO,WAAW;AAAA,IAC9B,gBAAgB;AAAA,EAClB,CAAC;AACH;AAgBO,SAAS,uBAAuB,OAAwC;AAC7E,MACE,SACA,OAAO,UAAU,YACjB,oBAAoB,SACpB,aAAc,MAAuC,cAAc,GACnE;AACA,WAAQ,MAAyC;AAAA,EACnD;AAEA,SAAO;AACT;AAGO,SAAS,aAAa,OAAqC;AAChE,MAAI,CAAC,SAAS,OAAO,UAAU,UAAU;AACvC,WAAO;AAAA,EACT;AAEA,QAAM,YAAY;AAMlB,SACE,oBAAoB,UAAU,OAAO,KACrC,QAAQ,UAAU,KAAK,KACvB,OAAO,UAAU,UAAU,YAC3B,CAAC,MAAM,QAAQ,UAAU,KAAK,KAC9B,MAAM,QAAQ,UAAU,MAAM;AAElC;AAGO,SAAS,oBACd,OAC4B;AAC5B,SACE,QAAQ,KAAK,KACb,OAAO,UAAU,YACjB,UAAU,QACV,cAAc,SACd,MAAM,QAAS,MAAiC,QAAQ;AAE5D;AAGO,SAAS,wBACd,QACkC;AAClC,MACE,UACA,OAAO,WAAW,YAClB,MAAM,QAAS,OAAmC,MAAM,GACxD;AACA,WAAQ,OAAwD;AAAA,EAClE;AAEA,SAAO,CAAC;AACV;AAGO,SAAS,eAAe,OAA2C;AACxE,MAAI,iBAAiB,OAAO;AAC1B,WAAO;AAAA,MACL,MAAM,MAAM;AAAA,MACZ,SAAS,MAAM;AAAA,IACjB;AAAA,EACF;AAEA,SAAO;AAAA,IACL,MAAM;AAAA,IACN,SAAS,OAAO,KAAK;AAAA,EACvB;AACF;","names":["assistantMessages","failedSpans","latestAssistantMessageContent","messagesByRole","spans","spansByKind","systemMessages","toolCalls","toolMessages","userMessages","normalized","details","message","type"]}
+{"version":3,"sources":["../src/harness.ts"],"sourcesContent":["import {\n  assistantMessages,\n  failedSpans,\n  latestAssistantMessageContent,\n  messagesByRole,\n  messagesToTranscriptEvents,\n  NormalizedSessionSchema,\n  spans,\n  spansByKind,\n  systemMessages,\n  TranscriptEventSchema,\n  toolCalls,\n  toolMessages,\n  userMessages,\n} from \"@vitest-evals/core\";\nimport type {\n  GenAiOperationName,\n  HarnessRun,\n  HarnessRunError,\n  NormalizedError,\n  JsonPrimitive,\n  JsonValue,\n  TranscriptMessageEvent,\n  NormalizedSession,\n  NormalizedSpan,\n  NormalizedSpanAttributes,\n  NormalizedSpanEvent,\n  TranscriptEvent,\n  NormalizedTrace,\n  TimingSummary,\n  ToolCall,\n  TranscriptMessageInput,\n  TranscriptMessageContentPart,\n  TranscriptMessageTextPart,\n  TranscriptMessageToolCallPart,\n  TranscriptMessageToolCall,\n  TranscriptMessageToolResultPart,\n  UsageSummary,\n} from \"@vitest-evals/core\";\n\nexport {\n  assistantMessages,\n  failedSpans,\n  latestAssistantMessageContent,\n  messagesByRole,\n  messagesToTranscriptEvents,\n  spans,\n  spansByKind,\n  systemMessages,\n  toolCalls,\n  toolMessages,\n  userMessages,\n} from \"@vitest-evals/core\";\nexport type {\n  GenAiOperationName,\n  GenAiOutputType,\n  GenAiProviderName,\n  GenAiSemanticAttributeKey,\n  GenAiSemanticAttributes,\n  GenAiTokenType,\n  GenAiToolType,\n  HarnessRun,\n  HarnessRunError,\n  NormalizedError,\n  JsonPrimitive,\n  JsonValue,\n  TranscriptMessageEvent,\n  NormalizedSession,\n  NormalizedSpan,\n  NormalizedSpanAttributeKey,\n  NormalizedSpanAttributes,\n  NormalizedSpanEvent,\n  TranscriptToolCallEvent,\n  TranscriptToolResultEvent,\n  TranscriptEvent,\n  NormalizedTrace,\n  OpenTelemetrySemanticAttributeKey,\n  OpenTelemetrySemanticAttributes,\n  TimingSummary,\n  ToolCall,\n  TranscriptMessageInput,\n  TranscriptMessageContentPart,\n  TranscriptMessageTextPart,\n  TranscriptMessageToolCallPart,\n  TranscriptMessageToolCall,\n  TranscriptMessageToolResultPart,\n  UsageSummary,\n} from \"@vitest-evals/core\";\n\n/** Options for attaching a fallback run trace to a harness result. */\nexport type EnsureRunTraceOptions = {\n  /** Human-readable run or harness name. */\n  name: string;\n  /** Wall-clock start time for the harness run. */\n  startedAt: Date;\n  /** Wall-clock finish time for the harness run. */\n  finishedAt: Date;\n  /** Optional trace id. A generated id is used when omitted. */\n  id?: string;\n  /** GenAI operation name to place on the root run span. */\n  operationName?: GenAiOperationName;\n  /** Optional JSON-safe source marker for the trace metadata. */\n  source?: string;\n};\n\ntype OutputField<TOutput extends JsonValue | undefined> =\n  undefined extends TOutput ? { output?: TOutput } : { output: TOutput };\n\n/** Generic JSON-like metadata record used by normalized artifacts and reports. */\nexport type HarnessMetadata = Record<string, unknown>;\n\n/**\n * Runtime context passed from the eval fixture into a harness run.\n *\n * @example\n * ```ts\n * const harness: Harness<string> = {\n *   name: \"refund-agent\",\n *   async run(input, context) {\n *     context.setArtifact(\"inputLength\", input.length);\n *\n *     return {\n *       output: undefined,\n *       session: { events: [{ type: \"message\", role: \"user\", content: input }] },\n *       usage: {},\n *       errors: [],\n *     };\n *   },\n * };\n * ```\n */\nexport type HarnessContext = {\n  /** Abort signal from Vitest when available. */\n  signal?: AbortSignal;\n  /** Mutable JSON-safe artifact bag shared with the harness. */\n  artifacts: Record<string, JsonValue>;\n  /** Stores one JSON-safe artifact on the current run. */\n  setArtifact: (name: string, value: JsonValue) => void;\n};\n\n/**\n * Adapter that executes the system under test and returns a normalized run.\n *\n * @example\n * ```ts\n * const harness: Harness<string, { status: \"approved\" | \"denied\" }> = {\n *   name: \"refund-agent\",\n *   async run(input, context) {\n *     return normalizeHarnessRun(input, await runRefundFlow(input), context);\n *   },\n * };\n * ```\n */\nexport type Harness<\n  TInput = unknown,\n  TOutput extends JsonValue | undefined = JsonValue | undefined,\n> = {\n  /** Stable harness name used in reports. */\n  name: string;\n  /** Executes the system under test and returns a normalized run. */\n  run: (input: TInput, context: HarnessContext) => Promise<HarnessRun<TOutput>>;\n};\n\n/** Value or promise accepted by lightweight harness callbacks. */\nexport type MaybePromise<T> = T | Promise<T>;\n\n/** Lightweight span event accepted by `createHarness(...)` results. */\nexport type SimpleSpanEvent = Omit<NormalizedSpanEvent, \"attributes\"> & {\n  /** Raw event attributes accepted by `createHarness(...)` before normalization. */\n  attributes?: Record<string, unknown>;\n};\n\n/** Lightweight span record accepted by `createHarness(...)` results. */\nexport type SimpleSpanRecord = Omit<\n  NormalizedSpan,\n  \"attributes\" | \"error\" | \"events\"\n> & {\n  /** Raw span attributes accepted by `createHarness(...)` before normalization. */\n  attributes?: Record<string, unknown>;\n  /** Raw span error accepted by `createHarness(...)` before normalization. */\n  error?: unknown;\n  /** Raw span events accepted by `createHarness(...)` before normalization. */\n  events?: SimpleSpanEvent[];\n};\n\n/** Lightweight trace record accepted by `createHarness(...)` results. */\nexport type SimpleTraceRecord = Omit<NormalizedTrace, \"metadata\" | \"spans\"> & {\n  /** Raw trace metadata accepted by `createHarness(...)` before normalization. */\n  metadata?: Record<string, unknown>;\n  /** Lightweight spans to normalize into the trace. */\n  spans: SimpleSpanRecord[];\n};\n\n/** Lightweight transcript input accepted by `createHarness(...)` results. */\nexport type SimpleTranscriptInput =\n  | {\n      /** Ordered normalized transcript events for the application run. */\n      events: TranscriptEvent[];\n      messages?: never;\n    }\n  | {\n      /** Strict camelCase message transport normalized into transcript events. */\n      messages: TranscriptMessageInput[];\n      events?: never;\n    };\n\n/**\n * Lightweight result shape normalized by `createHarness(...)`.\n *\n * @example\n * ```ts\n * const result: SimpleHarnessResult<{ status: \"approved\" }> = {\n *   output: { status: \"approved\" },\n *   events: [\n *     { type: \"message\", role: \"user\", content: \"Refund invoice inv_123\" },\n *     { type: \"message\", role: \"assistant\", content: { status: \"approved\" } },\n *   ],\n *   usage: { totalTokens: 260 },\n * };\n * ```\n */\nexport type SimpleHarnessResult<\n  TOutput extends JsonValue | undefined = JsonValue | undefined,\n> = OutputField<TOutput> &\n  SimpleTranscriptInput & {\n    /** Usage summary to attach to the run. */\n    usage?: UsageSummary;\n    /** Timing summary to attach to the run. */\n    timings?: TimingSummary;\n    /** Raw artifact values to normalize and merge into the run. */\n    artifacts?: Record<string, unknown>;\n    /** Lightweight traces and spans to normalize into the run. */\n    traces?: SimpleTraceRecord[];\n    /** Raw session metadata to normalize into the session. */\n    metadata?: Record<string, unknown>;\n    /** Raw errors to normalize into the run. */\n    errors?: unknown[];\n  };\n\n/** Either a complete normalized run or a lightweight result to normalize. */\nexport type HarnessResultLike<\n  TOutput extends JsonValue | undefined = JsonValue | undefined,\n> = HarnessRun<TOutput> | SimpleHarnessResult<TOutput>;\n\n/** Arguments passed to the `createHarness(...)` convenience callback. */\nexport type CreateHarnessRunArgs<TInput> = {\n  /** Original input passed to `run(input)`. */\n  input: TInput;\n  /** Abort signal from Vitest when available. */\n  signal?: AbortSignal;\n  /** Mutable run artifact bag. */\n  artifacts: HarnessContext[\"artifacts\"];\n  /** Stores one JSON-safe artifact on the current run. */\n  setArtifact: HarnessContext[\"setArtifact\"];\n};\n\n/**\n * Options for creating a lightweight custom application harness.\n *\n * Prefer this helper for custom harnesses. Implement `Harness` directly only\n * when the callback already returns a full `HarnessRun` with canonical\n * `session.events`.\n *\n * @example\n * ```ts\n * const options: CreateHarnessOptions<string, { status: \"approved\" }> = {\n *   name: \"refund-agent\",\n *   run: async ({ input }) => ({\n *     output: await classifyRefund(input),\n *     events: [{ type: \"message\", role: \"user\", content: input }],\n *   }),\n * };\n * ```\n */\nexport type CreateHarnessOptions<\n  TInput = unknown,\n  TOutput extends JsonValue | undefined = JsonValue | undefined,\n> = {\n  /** Stable harness name used in reports. */\n  name: string;\n  /** Executes application code and returns either a lightweight result or full `HarnessRun`. */\n  run: (\n    args: CreateHarnessRunArgs<TInput>,\n  ) => MaybePromise<HarnessResultLike<TOutput>>;\n};\n\nfunction isJsonPrimitive(value: unknown): value is JsonPrimitive {\n  return (\n    value === null ||\n    typeof value === \"string\" ||\n    typeof value === \"boolean\" ||\n    (typeof value === \"number\" && Number.isFinite(value))\n  );\n}\n\nfunction isJsonRecord(value: unknown): value is Record<string, unknown> {\n  return typeof value === \"object\" && value !== null && !Array.isArray(value);\n}\n\nfunction normalizeJsonArray(value: unknown[], seen: WeakSet<object>) {\n  if (seen.has(value)) {\n    return undefined;\n  }\n\n  seen.add(value);\n  const normalized = value.map((item) => {\n    const normalized = toJsonValueInternal(item, seen);\n    return normalized === undefined ? null : normalized;\n  });\n  seen.delete(value);\n\n  return normalized;\n}\n\nfunction normalizeJsonObject(\n  value: Record<string, unknown>,\n  seen: WeakSet<object>,\n): Record<string, JsonValue> {\n  const normalized: Record<string, JsonValue> = {};\n\n  if (seen.has(value)) {\n    return normalized;\n  }\n\n  seen.add(value);\n  try {\n    for (const [key, entryValue] of Object.entries(value)) {\n      const entry = toJsonValueInternal(entryValue, seen);\n      if (entry !== undefined) {\n        normalized[key] = entry;\n      }\n    }\n  } finally {\n    seen.delete(value);\n  }\n\n  return normalized;\n}\n\n/** Returns true when a value exposes a callable method with the given name. */\nexport function hasCallableMethod(value: unknown, methodName: string) {\n  return (\n    value !== null &&\n    (typeof value === \"object\" || typeof value === \"function\") &&\n    methodName in value &&\n    typeof (value as Record<string, unknown>)[methodName] === \"function\"\n  );\n}\n\n/** Normalizes an unknown value into the JSON-safe shape used by harness runs. */\nexport function toJsonValue(value: unknown): JsonValue | undefined {\n  return toJsonValueInternal(value, new WeakSet());\n}\n\nfunction toJsonValueInternal(\n  value: unknown,\n  seen: WeakSet<object>,\n): JsonValue | undefined {\n  if (isJsonPrimitive(value)) {\n    return value;\n  }\n\n  if (\n    value !== null &&\n    typeof value === \"object\" &&\n    seen.has(value as object)\n  ) {\n    return undefined;\n  }\n\n  if (Array.isArray(value)) {\n    return normalizeJsonArray(value, seen);\n  }\n\n  if (isJsonRecord(value)) {\n    return normalizeJsonObject(value, seen);\n  }\n\n  return undefined;\n}\n\n/** Drops non-JSON properties from a record while preserving valid values. */\nexport function normalizeRecord(\n  value: Record<string, unknown>,\n): Record<string, JsonValue> {\n  return normalizeJsonObject(value, new WeakSet());\n}\n\n/** Normalizes metadata and omits the field entirely when nothing survives. */\nexport function normalizeMetadata(\n  value: Record<string, unknown>,\n): Record<string, JsonValue> | undefined {\n  const normalized = normalizeRecord(value);\n  return Object.keys(normalized).length > 0 ? normalized : undefined;\n}\n\n/** Converts arbitrary content into the JSON-safe message content shape. */\nexport function normalizeContent(value: unknown): JsonValue {\n  const normalized = toJsonValue(value);\n  return normalized !== undefined ? normalized : String(value);\n}\n\n/**\n * Creates a harness from the common \"run app code and return output\" shape.\n *\n * @param options - Harness name plus the callback that executes app code.\n *\n * @example\n * ```ts\n * import { createHarness } from \"vitest-evals\";\n *\n * export const refundHarness = createHarness<\n *   string,\n *   { status: \"approved\" | \"denied\" }\n * >({\n *   name: \"refund-agent\",\n *   run: async ({ input, setArtifact }) => {\n *     const result = await runRefundFlow(input);\n *     const output = { status: result.status };\n *\n *     setArtifact(\"case\", { invoiceId: result.invoiceId });\n *\n *     return {\n *       output,\n *       events: [\n *         { type: \"message\", role: \"user\", content: input },\n *         {\n *           type: \"tool_call\",\n *           id: \"call_lookup\",\n *           name: \"lookupInvoice\",\n *           arguments: { invoiceId: result.invoiceId },\n *         },\n *         {\n *           type: \"tool_result\",\n *           toolCallId: \"call_lookup\",\n *           name: \"lookupInvoice\",\n *           content: { refundable: result.refundable },\n *         },\n *         { type: \"message\", role: \"assistant\", content: output },\n *       ],\n *       usage: { provider: \"openai\", model: \"gpt-4o-mini\" },\n *     };\n *   },\n * });\n * ```\n */\nexport function createHarness<\n  TInput = unknown,\n  TOutput extends JsonValue | undefined = JsonValue | undefined,\n>(options: CreateHarnessOptions<TInput, TOutput>): Harness<TInput, TOutput>;\nexport function createHarness<\n  TInput = unknown,\n  TOutput extends JsonValue | undefined = JsonValue | undefined,\n>(options: CreateHarnessOptions<TInput, TOutput>): Harness<TInput, TOutput> {\n  const harness: Harness<TInput, TOutput> = {\n    name: options.name,\n    run: async (input, context) => {\n      const startedAt = new Date();\n\n      try {\n        const result = await options.run({\n          input,\n          signal: context.signal,\n          artifacts: context.artifacts,\n          setArtifact: context.setArtifact,\n        });\n        const run = normalizeHarnessRun(input, result, context);\n        ensureRunTrace(run, {\n          name: options.name,\n          startedAt,\n          finishedAt: new Date(),\n        });\n\n        return run;\n      } catch (error) {\n        const partialRun = getHarnessRunFromError(error);\n        if (partialRun) {\n          if (\n            Object.keys(context.artifacts).length > 0 &&\n            !partialRun.artifacts\n          ) {\n            partialRun.artifacts = context.artifacts;\n          }\n          ensureRunTrace(partialRun, {\n            name: options.name,\n            startedAt,\n            finishedAt: new Date(),\n          });\n          throw attachHarnessRunToError(error, partialRun);\n        }\n\n        const failedRun = createFailedHarnessRun(input, error, {\n          artifacts: context.artifacts,\n        });\n        ensureRunTrace(failedRun, {\n          name: options.name,\n          startedAt,\n          finishedAt: new Date(),\n        });\n\n        throw attachHarnessRunToError(error, failedRun);\n      }\n    },\n  };\n\n  return harness;\n}\n\n/**\n * Normalizes a lightweight harness result into the reporter-facing run shape.\n *\n * @param input - Original input passed to the harness.\n * @param result - Lightweight result or pre-normalized harness run.\n * @param context - Optional per-run context used to merge artifacts.\n *\n * @example\n * ```ts\n * const run = normalizeHarnessRun(\"Refund invoice inv_123\", {\n *   output: { status: \"approved\" },\n *   events: [\n *     { type: \"message\", role: \"user\", content: \"Refund invoice inv_123\" },\n *     {\n *       type: \"tool_call\",\n *       id: \"call_lookup\",\n *       name: \"lookupInvoice\",\n *       arguments: { invoiceId: \"inv_123\" },\n *     },\n *     {\n *       type: \"tool_result\",\n *       toolCallId: \"call_lookup\",\n *       name: \"lookupInvoice\",\n *       content: { refundable: true },\n *     },\n *   ],\n *   usage: { provider: \"openai\", model: \"gpt-4o-mini\" },\n * });\n *\n * expect(toolCalls(run)).toHaveLength(1);\n * ```\n */\nexport function normalizeHarnessRun<\n  TInput = unknown,\n  TOutput extends JsonValue | undefined = JsonValue | undefined,\n>(\n  input: TInput,\n  result: HarnessResultLike<TOutput>,\n  context?: HarnessContext,\n): HarnessRun<TOutput> {\n  if (isHarnessRun(result)) {\n    if (\n      context &&\n      Object.keys(context.artifacts).length > 0 &&\n      !result.artifacts\n    ) {\n      return {\n        ...result,\n        artifacts: context.artifacts,\n      };\n    }\n\n    return result;\n  }\n\n  if (\"toolCalls\" in result) {\n    throw new TypeError(\n      'createHarness results do not accept top-level toolCalls. Return ordered session events with type: \"tool_call\" and type: \"tool_result\" entries instead.',\n    );\n  }\n\n  const output = result.output;\n  const usage = result.usage ?? {};\n  const events = normalizeTranscriptInput(result);\n  if (!events) {\n    throw new TypeError(\n      \"createHarness results must include ordered events or messages. Return a full HarnessRun or a lightweight result with events/messages.\",\n    );\n  }\n  if (events.length === 0) {\n    throw new TypeError(\n      \"createHarness results must include at least one transcript event. Return ordered events or message transport inputs that normalize into events.\",\n    );\n  }\n  const metadata = result.metadata\n    ? normalizeMetadata(result.metadata)\n    : undefined;\n  const artifacts = normalizeMergedArtifacts(\n    context?.artifacts,\n    result.artifacts,\n  );\n  const traces = normalizeSimpleTraces(result.traces);\n\n  return {\n    session: {\n      events,\n      ...(usage.provider ? { provider: usage.provider } : {}),\n      ...(usage.model ? { model: usage.model } : {}),\n      ...(metadata ? { metadata } : {}),\n    },\n    ...(output !== undefined ? { output } : {}),\n    usage,\n    ...(result.timings ? { timings: result.timings } : {}),\n    ...(artifacts ? { artifacts } : {}),\n    ...(traces ? { traces } : {}),\n    errors: normalizeSimpleErrors(result.errors),\n  } as HarnessRun<TOutput>;\n}\n\nfunction normalizeTranscriptInput(\n  result: HarnessResultLike,\n): TranscriptEvent[] | undefined {\n  if (\"events\" in result && Array.isArray(result.events)) {\n    return result.events.map((event) => TranscriptEventSchema.parse(event));\n  }\n\n  if (\"messages\" in result && Array.isArray(result.messages)) {\n    return messagesToTranscriptEvents(result.messages).map((event) =>\n      TranscriptEventSchema.parse(event),\n    );\n  }\n\n  return undefined;\n}\n\n/**\n * Builds a JSON-safe failed run for errors that happen before a harness can return.\n *\n * @param input - Original input passed to the harness.\n * @param error - Error thrown by setup or execution.\n * @param options - Optional artifacts to preserve on the failed run.\n */\nexport function createFailedHarnessRun(\n  input: unknown,\n  error: unknown,\n  options: { artifacts?: Record<string, JsonValue> } = {},\n): HarnessRun {\n  const artifacts = options.artifacts;\n\n  return {\n    session: {\n      events: [\n        {\n          type: \"message\",\n          role: \"user\",\n          content: normalizeContent(input),\n        },\n      ],\n    },\n    usage: {},\n    ...(artifacts && Object.keys(artifacts).length > 0 ? { artifacts } : {}),\n    errors: [serializeError(error)],\n  };\n}\n\nfunction normalizeMergedArtifacts(\n  contextArtifacts: Record<string, JsonValue> | undefined,\n  resultArtifacts: Record<string, unknown> | undefined,\n) {\n  const artifacts = {\n    ...(contextArtifacts ?? {}),\n    ...(resultArtifacts ? normalizeRecord(resultArtifacts) : {}),\n  };\n\n  return Object.keys(artifacts).length > 0 ? artifacts : undefined;\n}\n\nfunction normalizeSimpleErrors(\n  errors: unknown[] | undefined,\n): Array<Record<string, JsonValue>> {\n  return (errors ?? []).map((error) => {\n    const normalized = toJsonValue(error);\n\n    if (\n      normalized &&\n      typeof normalized === \"object\" &&\n      !Array.isArray(normalized) &&\n      Object.keys(normalized).length > 0\n    ) {\n      return normalized;\n    }\n\n    return serializeError(error);\n  });\n}\n\nfunction normalizeSimpleTraces(\n  traces: SimpleTraceRecord[] | undefined,\n): NormalizedTrace[] | undefined {\n  if (!Array.isArray(traces)) {\n    return undefined;\n  }\n\n  const normalized = traces\n    .map(normalizeSimpleTrace)\n    .filter((trace): trace is NormalizedTrace => Boolean(trace));\n\n  return normalized.length > 0 ? normalized : undefined;\n}\n\nfunction normalizeSimpleTrace(trace: unknown): NormalizedTrace | undefined {\n  if (!isJsonRecord(trace)) {\n    return undefined;\n  }\n\n  const {\n    metadata: rawMetadata,\n    spans: rawSpans,\n    ...traceFields\n  } = trace as Partial<SimpleTraceRecord>;\n  const spans = (Array.isArray(rawSpans) ? rawSpans : [])\n    .map((span) => normalizeSimpleSpan(span))\n    .filter((span): span is NormalizedSpan => Boolean(span));\n  const metadata = isJsonRecord(rawMetadata)\n    ? normalizeMetadata(rawMetadata)\n    : undefined;\n\n  if (spans.length === 0 && !traceFields.id && !traceFields.name) {\n    return undefined;\n  }\n\n  return {\n    ...traceFields,\n    ...(metadata ? { metadata } : {}),\n    spans,\n  };\n}\n\nfunction normalizeSimpleSpan(span: unknown): NormalizedSpan | undefined {\n  if (!isJsonRecord(span) || typeof span.name !== \"string\" || !span.name) {\n    return undefined;\n  }\n\n  const {\n    attributes: rawAttributes,\n    error: rawError,\n    events: rawEvents,\n    ...spanFields\n  } = span as Partial<SimpleSpanRecord> & { name: string };\n  const attributes = rawAttributes\n    ? isJsonRecord(rawAttributes)\n      ? normalizeMetadata(rawAttributes)\n      : undefined\n    : undefined;\n  const error = normalizeSpanError(rawError);\n  const events = normalizeSimpleSpanEvents(rawEvents);\n\n  return {\n    ...spanFields,\n    ...(attributes\n      ? { attributes: attributes as NormalizedSpanAttributes }\n      : {}),\n    ...(error ? { error } : {}),\n    ...(events ? { events } : {}),\n  };\n}\n\nfunction normalizeSimpleSpanEvents(\n  events: unknown,\n): NormalizedSpanEvent[] | undefined {\n  if (!Array.isArray(events)) {\n    return undefined;\n  }\n\n  const normalized = events\n    .map(normalizeSimpleSpanEvent)\n    .filter((event): event is NormalizedSpanEvent => Boolean(event));\n\n  return normalized.length > 0 ? normalized : undefined;\n}\n\nfunction normalizeSimpleSpanEvent(\n  event: unknown,\n): NormalizedSpanEvent | undefined {\n  if (!isJsonRecord(event) || typeof event.name !== \"string\" || !event.name) {\n    return undefined;\n  }\n\n  const { attributes: rawAttributes, ...eventFields } =\n    event as Partial<SimpleSpanEvent> & { name: string };\n  const attributes = rawAttributes\n    ? isJsonRecord(rawAttributes)\n      ? normalizeMetadata(rawAttributes)\n      : undefined\n    : undefined;\n\n  return {\n    ...eventFields,\n    ...(attributes\n      ? { attributes: attributes as NormalizedSpanAttributes }\n      : {}),\n  };\n}\n\n/** Normalizes arbitrary span errors while preserving object-shaped messages. */\nexport function normalizeSpanError(\n  error: unknown,\n): NormalizedSpan[\"error\"] | undefined {\n  if (error === undefined) {\n    return undefined;\n  }\n\n  if (error instanceof Error) {\n    const details = normalizeMetadata(\n      error as unknown as Record<string, unknown>,\n    );\n\n    return {\n      ...(details ?? {}),\n      type: error.name,\n      message: error.message,\n    };\n  }\n\n  if (\n    error &&\n    typeof error === \"object\" &&\n    !Array.isArray(error) &&\n    typeof (error as { message?: unknown }).message === \"string\"\n  ) {\n    const normalized = normalizeMetadata(error as Record<string, unknown>);\n    const { message, type, ...details } = normalized ?? {};\n\n    return {\n      ...details,\n      message: message as string,\n      ...(typeof type === \"string\" ? { type } : {}),\n    };\n  }\n\n  const serialized = serializeError(error);\n  const { message, type, ...details } = serialized;\n\n  return {\n    ...details,\n    message: typeof message === \"string\" ? message : String(message),\n    ...(typeof type === \"string\" ? { type } : {}),\n  };\n}\n\n/** Normalizes raw span attributes into the JSON-safe span attribute shape. */\nexport function normalizeSpanAttributes(\n  attributes: Record<string, unknown>,\n): NormalizedSpanAttributes | undefined {\n  return normalizeMetadata(attributes) as NormalizedSpanAttributes | undefined;\n}\n\n/** Builds common OpenTelemetry GenAI usage attributes from a usage summary. */\nexport function createGenAiUsageAttributes(\n  usage: UsageSummary | undefined,\n  options: { provider?: string } = {},\n) {\n  return {\n    \"gen_ai.provider.name\": usage?.provider ?? options.provider,\n    \"gen_ai.request.model\": usage?.model,\n    \"gen_ai.response.model\": usage?.model,\n    \"gen_ai.usage.input_tokens\": usage?.inputTokens,\n    \"gen_ai.usage.output_tokens\": usage?.outputTokens,\n    \"gen_ai.usage.reasoning.output_tokens\": usage?.reasoningTokens,\n  } satisfies Record<string, unknown>;\n}\n\n/**\n * Attaches a fallback run trace when a harness result does not already contain spans.\n *\n * This keeps custom harnesses inspectable while first-party harness packages\n * remain free to attach richer native traces.\n */\nexport function ensureRunTrace(\n  run: HarnessRun,\n  options: EnsureRunTraceOptions,\n): NormalizedTrace | undefined {\n  if (spans(run).length > 0) {\n    return undefined;\n  }\n\n  const traceId = options.id ?? createGeneratedTraceId();\n  const rootSpanId = `${traceId}:run`;\n  const durationMs = options.finishedAt.getTime() - options.startedAt.getTime();\n  const rootError =\n    run.errors.length > 0 ? normalizeSpanError(run.errors[0]) : undefined;\n  const runSpan: NormalizedSpan = {\n    id: rootSpanId,\n    traceId,\n    name: options.name,\n    kind: \"run\",\n    startedAt: options.startedAt.toISOString(),\n    finishedAt: options.finishedAt.toISOString(),\n    durationMs,\n    status: rootError ? \"error\" : \"ok\",\n    ...(rootError ? { error: rootError } : {}),\n    attributes: normalizeSpanAttributes({\n      \"gen_ai.operation.name\": options.operationName ?? \"invoke_workflow\",\n      \"gen_ai.workflow.name\": options.name,\n      ...createGenAiUsageAttributes(run.usage),\n    }),\n  };\n  const trace: NormalizedTrace = {\n    id: traceId,\n    name: options.name,\n    startedAt: options.startedAt.toISOString(),\n    finishedAt: options.finishedAt.toISOString(),\n    durationMs,\n    ...(options.source ? { metadata: { source: options.source } } : {}),\n    spans: [runSpan],\n  };\n\n  run.traces = [trace];\n  return trace;\n}\n\nlet nextGeneratedTraceId = 0;\n\nfunction createGeneratedTraceId() {\n  nextGeneratedTraceId += 1;\n  return `trace_${nextGeneratedTraceId}`;\n}\n\n/**\n * Attaches a partial or complete harness run to an arbitrary thrown error.\n *\n * @param error - Thrown value to wrap.\n * @param run - Partial or complete normalized harness run to preserve.\n *\n * @example\n * ```ts\n * try {\n *   return await runAgent(input);\n * } catch (error) {\n *   throw attachHarnessRunToError(error, partialRun);\n * }\n * ```\n */\nexport function attachHarnessRunToError(\n  error: unknown,\n  run: HarnessRun,\n): HarnessRunError {\n  const baseError =\n    error instanceof Error\n      ? error\n      : new Error(String(error ?? \"Unknown error\"));\n  return Object.assign(baseError, {\n    vitestEvalsRun: run,\n  });\n}\n\n/**\n * Reads an attached harness run back off a previously wrapped error value.\n *\n * @param error - Unknown thrown value that may contain a harness run.\n *\n * @example\n * ```ts\n * const partialRun = getHarnessRunFromError(error);\n *\n * if (partialRun) {\n *   console.log(toolCalls(partialRun.session));\n * }\n * ```\n */\nexport function getHarnessRunFromError(error: unknown): HarnessRun | undefined {\n  if (\n    error &&\n    typeof error === \"object\" &&\n    \"vitestEvalsRun\" in error &&\n    isHarnessRun((error as { vitestEvalsRun?: unknown }).vitestEvalsRun)\n  ) {\n    return (error as { vitestEvalsRun: HarnessRun }).vitestEvalsRun;\n  }\n\n  return undefined;\n}\n\n/** Returns true when a value matches the normalized `HarnessRun` contract. */\nexport function isHarnessRun(value: unknown): value is HarnessRun {\n  if (!value || typeof value !== \"object\") {\n    return false;\n  }\n\n  const candidate = value as {\n    session?: unknown;\n    usage?: unknown;\n    errors?: unknown;\n  };\n\n  return (\n    isNormalizedSession(candidate.session) &&\n    Boolean(candidate.usage) &&\n    typeof candidate.usage === \"object\" &&\n    !Array.isArray(candidate.usage) &&\n    Array.isArray(candidate.errors)\n  );\n}\n\n/** Returns true when a value matches the normalized session contract. */\nexport function isNormalizedSession(\n  value: unknown,\n): value is NormalizedSession {\n  return NormalizedSessionSchema.safeParse(value).success;\n}\n\n/** Reuses pre-normalized harness errors when a runtime already returns them. */\nexport function resolveHarnessRunErrors(\n  result: unknown,\n): Array<Record<string, JsonValue>> {\n  if (\n    result &&\n    typeof result === \"object\" &&\n    Array.isArray((result as Record<string, unknown>).errors)\n  ) {\n    return (result as { errors: Array<Record<string, JsonValue>> }).errors;\n  }\n\n  return [];\n}\n\n/** Serializes an arbitrary thrown value into the normalized error shape. */\nexport function serializeError(error: unknown): Record<string, JsonValue> {\n  if (error instanceof Error) {\n    return {\n      type: error.name,\n      message: error.message,\n    };\n  }\n\n  return {\n    type: \"Error\",\n    message: String(error),\n  };\n}\n"],"mappings":";AAAA;AAAA,EAKE;AAAA,EACA;AAAA,EACA;AAAA,EAGA;AAAA,OAIK;AA0BP;AAAA,EACE,qBAAAA;AAAA,EACA,eAAAC;AAAA,EACA,iCAAAC;AAAA,EACA,kBAAAC;AAAA,EACA,8BAAAC;AAAA,EACA,SAAAC;AAAA,EACA,eAAAC;AAAA,EACA,kBAAAC;AAAA,EACA,aAAAC;AAAA,EACA,gBAAAC;AAAA,EACA,gBAAAC;AAAA,OACK;AA0OP,SAAS,gBAAgB,OAAwC;AAC/D,SACE,UAAU,QACV,OAAO,UAAU,YACjB,OAAO,UAAU,aAChB,OAAO,UAAU,YAAY,OAAO,SAAS,KAAK;AAEvD;AAEA,SAAS,aAAa,OAAkD;AACtE,SAAO,OAAO,UAAU,YAAY,UAAU,QAAQ,CAAC,MAAM,QAAQ,KAAK;AAC5E;AAEA,SAAS,mBAAmB,OAAkB,MAAuB;AACnE,MAAI,KAAK,IAAI,KAAK,GAAG;AACnB,WAAO;AAAA,EACT;AAEA,OAAK,IAAI,KAAK;AACd,QAAM,aAAa,MAAM,IAAI,CAAC,SAAS;AACrC,UAAMC,cAAa,oBAAoB,MAAM,IAAI;AACjD,WAAOA,gBAAe,SAAY,OAAOA;AAAA,EAC3C,CAAC;AACD,OAAK,OAAO,KAAK;AAEjB,SAAO;AACT;AAEA,SAAS,oBACP,OACA,MAC2B;AAC3B,QAAM,aAAwC,CAAC;AAE/C,MAAI,KAAK,IAAI,KAAK,GAAG;AACnB,WAAO;AAAA,EACT;AAEA,OAAK,IAAI,KAAK;AACd,MAAI;AACF,eAAW,CAAC,KAAK,UAAU,KAAK,OAAO,QAAQ,KAAK,GAAG;AACrD,YAAM,QAAQ,oBAAoB,YAAY,IAAI;AAClD,UAAI,UAAU,QAAW;AACvB,mBAAW,GAAG,IAAI;AAAA,MACpB;AAAA,IACF;AAAA,EACF,UAAE;AACA,SAAK,OAAO,KAAK;AAAA,EACnB;AAEA,SAAO;AACT;AAGO,SAAS,kBAAkB,OAAgB,YAAoB;AACpE,SACE,UAAU,SACT,OAAO,UAAU,YAAY,OAAO,UAAU,eAC/C,cAAc,SACd,OAAQ,MAAkC,UAAU,MAAM;AAE9D;AAGO,SAAS,YAAY,OAAuC;AACjE,SAAO,oBAAoB,OAAO,oBAAI,QAAQ,CAAC;AACjD;AAEA,SAAS,oBACP,OACA,MACuB;AACvB,MAAI,gBAAgB,KAAK,GAAG;AAC1B,WAAO;AAAA,EACT;AAEA,MACE,UAAU,QACV,OAAO,UAAU,YACjB,KAAK,IAAI,KAAe,GACxB;AACA,WAAO;AAAA,EACT;AAEA,MAAI,MAAM,QAAQ,KAAK,GAAG;AACxB,WAAO,mBAAmB,OAAO,IAAI;AAAA,EACvC;AAEA,MAAI,aAAa,KAAK,GAAG;AACvB,WAAO,oBAAoB,OAAO,IAAI;AAAA,EACxC;AAEA,SAAO;AACT;AAGO,SAAS,gBACd,OAC2B;AAC3B,SAAO,oBAAoB,OAAO,oBAAI,QAAQ,CAAC;AACjD;AAGO,SAAS,kBACd,OACuC;AACvC,QAAM,aAAa,gBAAgB,KAAK;AACxC,SAAO,OAAO,KAAK,UAAU,EAAE,SAAS,IAAI,aAAa;AAC3D;AAGO,SAAS,iBAAiB,OAA2B;AAC1D,QAAM,aAAa,YAAY,KAAK;AACpC,SAAO,eAAe,SAAY,aAAa,OAAO,KAAK;AAC7D;AAkDO,SAAS,cAGd,SAA0E;AAC1E,QAAM,UAAoC;AAAA,IACxC,MAAM,QAAQ;AAAA,IACd,KAAK,OAAO,OAAO,YAAY;AAC7B,YAAM,YAAY,oBAAI,KAAK;AAE3B,UAAI;AACF,cAAM,SAAS,MAAM,QAAQ,IAAI;AAAA,UAC/B;AAAA,UACA,QAAQ,QAAQ;AAAA,UAChB,WAAW,QAAQ;AAAA,UACnB,aAAa,QAAQ;AAAA,QACvB,CAAC;AACD,cAAM,MAAM,oBAAoB,OAAO,QAAQ,OAAO;AACtD,uBAAe,KAAK;AAAA,UAClB,MAAM,QAAQ;AAAA,UACd;AAAA,UACA,YAAY,oBAAI,KAAK;AAAA,QACvB,CAAC;AAED,eAAO;AAAA,MACT,SAAS,OAAO;AACd,cAAM,aAAa,uBAAuB,KAAK;AAC/C,YAAI,YAAY;AACd,cACE,OAAO,KAAK,QAAQ,SAAS,EAAE,SAAS,KACxC,CAAC,WAAW,WACZ;AACA,uBAAW,YAAY,QAAQ;AAAA,UACjC;AACA,yBAAe,YAAY;AAAA,YACzB,MAAM,QAAQ;AAAA,YACd;AAAA,YACA,YAAY,oBAAI,KAAK;AAAA,UACvB,CAAC;AACD,gBAAM,wBAAwB,OAAO,UAAU;AAAA,QACjD;AAEA,cAAM,YAAY,uBAAuB,OAAO,OAAO;AAAA,UACrD,WAAW,QAAQ;AAAA,QACrB,CAAC;AACD,uBAAe,WAAW;AAAA,UACxB,MAAM,QAAQ;AAAA,UACd;AAAA,UACA,YAAY,oBAAI,KAAK;AAAA,QACvB,CAAC;AAED,cAAM,wBAAwB,OAAO,SAAS;AAAA,MAChD;AAAA,IACF;AAAA,EACF;AAEA,SAAO;AACT;AAkCO,SAAS,oBAId,OACA,QACA,SACqB;AACrB,MAAI,aAAa,MAAM,GAAG;AACxB,QACE,WACA,OAAO,KAAK,QAAQ,SAAS,EAAE,SAAS,KACxC,CAAC,OAAO,WACR;AACA,aAAO;AAAA,QACL,GAAG;AAAA,QACH,WAAW,QAAQ;AAAA,MACrB;AAAA,IACF;AAEA,WAAO;AAAA,EACT;AAEA,MAAI,eAAe,QAAQ;AACzB,UAAM,IAAI;AAAA,MACR;AAAA,IACF;AAAA,EACF;AAEA,QAAM,SAAS,OAAO;AACtB,QAAM,QAAQ,OAAO,SAAS,CAAC;AAC/B,QAAM,SAAS,yBAAyB,MAAM;AAC9C,MAAI,CAAC,QAAQ;AACX,UAAM,IAAI;AAAA,MACR;AAAA,IACF;AAAA,EACF;AACA,MAAI,OAAO,WAAW,GAAG;AACvB,UAAM,IAAI;AAAA,MACR;AAAA,IACF;AAAA,EACF;AACA,QAAM,WAAW,OAAO,WACpB,kBAAkB,OAAO,QAAQ,IACjC;AACJ,QAAM,YAAY;AAAA,IAChB,SAAS;AAAA,IACT,OAAO;AAAA,EACT;AACA,QAAM,SAAS,sBAAsB,OAAO,MAAM;AAElD,SAAO;AAAA,IACL,SAAS;AAAA,MACP;AAAA,MACA,GAAI,MAAM,WAAW,EAAE,UAAU,MAAM,SAAS,IAAI,CAAC;AAAA,MACrD,GAAI,MAAM,QAAQ,EAAE,OAAO,MAAM,MAAM,IAAI,CAAC;AAAA,MAC5C,GAAI,WAAW,EAAE,SAAS,IAAI,CAAC;AAAA,IACjC;AAAA,IACA,GAAI,WAAW,SAAY,EAAE,OAAO,IAAI,CAAC;AAAA,IACzC;AAAA,IACA,GAAI,OAAO,UAAU,EAAE,SAAS,OAAO,QAAQ,IAAI,CAAC;AAAA,IACpD,GAAI,YAAY,EAAE,UAAU,IAAI,CAAC;AAAA,IACjC,GAAI,SAAS,EAAE,OAAO,IAAI,CAAC;AAAA,IAC3B,QAAQ,sBAAsB,OAAO,MAAM;AAAA,EAC7C;AACF;AAEA,SAAS,yBACP,QAC+B;AAC/B,MAAI,YAAY,UAAU,MAAM,QAAQ,OAAO,MAAM,GAAG;AACtD,WAAO,OAAO,OAAO,IAAI,CAAC,UAAU,sBAAsB,MAAM,KAAK,CAAC;AAAA,EACxE;AAEA,MAAI,cAAc,UAAU,MAAM,QAAQ,OAAO,QAAQ,GAAG;AAC1D,WAAO,2BAA2B,OAAO,QAAQ,EAAE;AAAA,MAAI,CAAC,UACtD,sBAAsB,MAAM,KAAK;AAAA,IACnC;AAAA,EACF;AAEA,SAAO;AACT;AASO,SAAS,uBACd,OACA,OACA,UAAqD,CAAC,GAC1C;AACZ,QAAM,YAAY,QAAQ;AAE1B,SAAO;AAAA,IACL,SAAS;AAAA,MACP,QAAQ;AAAA,QACN;AAAA,UACE,MAAM;AAAA,UACN,MAAM;AAAA,UACN,SAAS,iBAAiB,KAAK;AAAA,QACjC;AAAA,MACF;AAAA,IACF;AAAA,IACA,OAAO,CAAC;AAAA,IACR,GAAI,aAAa,OAAO,KAAK,SAAS,EAAE,SAAS,IAAI,EAAE,UAAU,IAAI,CAAC;AAAA,IACtE,QAAQ,CAAC,eAAe,KAAK,CAAC;AAAA,EAChC;AACF;AAEA,SAAS,yBACP,kBACA,iBACA;AACA,QAAM,YAAY;AAAA,IAChB,GAAI,oBAAoB,CAAC;AAAA,IACzB,GAAI,kBAAkB,gBAAgB,eAAe,IAAI,CAAC;AAAA,EAC5D;AAEA,SAAO,OAAO,KAAK,SAAS,EAAE,SAAS,IAAI,YAAY;AACzD;AAEA,SAAS,sBACP,QACkC;AAClC,UAAQ,UAAU,CAAC,GAAG,IAAI,CAAC,UAAU;AACnC,UAAM,aAAa,YAAY,KAAK;AAEpC,QACE,cACA,OAAO,eAAe,YACtB,CAAC,MAAM,QAAQ,UAAU,KACzB,OAAO,KAAK,UAAU,EAAE,SAAS,GACjC;AACA,aAAO;AAAA,IACT;AAEA,WAAO,eAAe,KAAK;AAAA,EAC7B,CAAC;AACH;AAEA,SAAS,sBACP,QAC+B;AAC/B,MAAI,CAAC,MAAM,QAAQ,MAAM,GAAG;AAC1B,WAAO;AAAA,EACT;AAEA,QAAM,aAAa,OAChB,IAAI,oBAAoB,EACxB,OAAO,CAAC,UAAoC,QAAQ,KAAK,CAAC;AAE7D,SAAO,WAAW,SAAS,IAAI,aAAa;AAC9C;AAEA,SAAS,qBAAqB,OAA6C;AACzE,MAAI,CAAC,aAAa,KAAK,GAAG;AACxB,WAAO;AAAA,EACT;AAEA,QAAM;AAAA,IACJ,UAAU;AAAA,IACV,OAAO;AAAA,IACP,GAAG;AAAA,EACL,IAAI;AACJ,QAAMN,UAAS,MAAM,QAAQ,QAAQ,IAAI,WAAW,CAAC,GAClD,IAAI,CAAC,SAAS,oBAAoB,IAAI,CAAC,EACvC,OAAO,CAAC,SAAiC,QAAQ,IAAI,CAAC;AACzD,QAAM,WAAW,aAAa,WAAW,IACrC,kBAAkB,WAAW,IAC7B;AAEJ,MAAIA,OAAM,WAAW,KAAK,CAAC,YAAY,MAAM,CAAC,YAAY,MAAM;AAC9D,WAAO;AAAA,EACT;AAEA,SAAO;AAAA,IACL,GAAG;AAAA,IACH,GAAI,WAAW,EAAE,SAAS,IAAI,CAAC;AAAA,IAC/B,OAAAA;AAAA,EACF;AACF;AAEA,SAAS,oBAAoB,MAA2C;AACtE,MAAI,CAAC,aAAa,IAAI,KAAK,OAAO,KAAK,SAAS,YAAY,CAAC,KAAK,MAAM;AACtE,WAAO;AAAA,EACT;AAEA,QAAM;AAAA,IACJ,YAAY;AAAA,IACZ,OAAO;AAAA,IACP,QAAQ;AAAA,IACR,GAAG;AAAA,EACL,IAAI;AACJ,QAAM,aAAa,gBACf,aAAa,aAAa,IACxB,kBAAkB,aAAa,IAC/B,SACF;AACJ,QAAM,QAAQ,mBAAmB,QAAQ;AACzC,QAAM,SAAS,0BAA0B,SAAS;AAElD,SAAO;AAAA,IACL,GAAG;AAAA,IACH,GAAI,aACA,EAAE,WAAmD,IACrD,CAAC;AAAA,IACL,GAAI,QAAQ,EAAE,MAAM,IAAI,CAAC;AAAA,IACzB,GAAI,SAAS,EAAE,OAAO,IAAI,CAAC;AAAA,EAC7B;AACF;AAEA,SAAS,0BACP,QACmC;AACnC,MAAI,CAAC,MAAM,QAAQ,MAAM,GAAG;AAC1B,WAAO;AAAA,EACT;AAEA,QAAM,aAAa,OAChB,IAAI,wBAAwB,EAC5B,OAAO,CAAC,UAAwC,QAAQ,KAAK,CAAC;AAEjE,SAAO,WAAW,SAAS,IAAI,aAAa;AAC9C;AAEA,SAAS,yBACP,OACiC;AACjC,MAAI,CAAC,aAAa,KAAK,KAAK,OAAO,MAAM,SAAS,YAAY,CAAC,MAAM,MAAM;AACzE,WAAO;AAAA,EACT;AAEA,QAAM,EAAE,YAAY,eAAe,GAAG,YAAY,IAChD;AACF,QAAM,aAAa,gBACf,aAAa,aAAa,IACxB,kBAAkB,aAAa,IAC/B,SACF;AAEJ,SAAO;AAAA,IACL,GAAG;AAAA,IACH,GAAI,aACA,EAAE,WAAmD,IACrD,CAAC;AAAA,EACP;AACF;AAGO,SAAS,mBACd,OACqC;AACrC,MAAI,UAAU,QAAW;AACvB,WAAO;AAAA,EACT;AAEA,MAAI,iBAAiB,OAAO;AAC1B,UAAMO,WAAU;AAAA,MACd;AAAA,IACF;AAEA,WAAO;AAAA,MACL,GAAIA,YAAW,CAAC;AAAA,MAChB,MAAM,MAAM;AAAA,MACZ,SAAS,MAAM;AAAA,IACjB;AAAA,EACF;AAEA,MACE,SACA,OAAO,UAAU,YACjB,CAAC,MAAM,QAAQ,KAAK,KACpB,OAAQ,MAAgC,YAAY,UACpD;AACA,UAAM,aAAa,kBAAkB,KAAgC;AACrE,UAAM,EAAE,SAAAC,UAAS,MAAAC,OAAM,GAAGF,SAAQ,IAAI,cAAc,CAAC;AAErD,WAAO;AAAA,MACL,GAAGA;AAAA,MACH,SAASC;AAAA,MACT,GAAI,OAAOC,UAAS,WAAW,EAAE,MAAAA,MAAK,IAAI,CAAC;AAAA,IAC7C;AAAA,EACF;AAEA,QAAM,aAAa,eAAe,KAAK;AACvC,QAAM,EAAE,SAAS,MAAM,GAAG,QAAQ,IAAI;AAEtC,SAAO;AAAA,IACL,GAAG;AAAA,IACH,SAAS,OAAO,YAAY,WAAW,UAAU,OAAO,OAAO;AAAA,IAC/D,GAAI,OAAO,SAAS,WAAW,EAAE,KAAK,IAAI,CAAC;AAAA,EAC7C;AACF;AAGO,SAAS,wBACd,YACsC;AACtC,SAAO,kBAAkB,UAAU;AACrC;AAGO,SAAS,2BACd,OACA,UAAiC,CAAC,GAClC;AACA,SAAO;AAAA,IACL,wBAAwB,OAAO,YAAY,QAAQ;AAAA,IACnD,wBAAwB,OAAO;AAAA,IAC/B,yBAAyB,OAAO;AAAA,IAChC,6BAA6B,OAAO;AAAA,IACpC,8BAA8B,OAAO;AAAA,IACrC,wCAAwC,OAAO;AAAA,EACjD;AACF;AAQO,SAAS,eACd,KACA,SAC6B;AAC7B,MAAI,MAAM,GAAG,EAAE,SAAS,GAAG;AACzB,WAAO;AAAA,EACT;AAEA,QAAM,UAAU,QAAQ,MAAM,uBAAuB;AACrD,QAAM,aAAa,GAAG,OAAO;AAC7B,QAAM,aAAa,QAAQ,WAAW,QAAQ,IAAI,QAAQ,UAAU,QAAQ;AAC5E,QAAM,YACJ,IAAI,OAAO,SAAS,IAAI,mBAAmB,IAAI,OAAO,CAAC,CAAC,IAAI;AAC9D,QAAM,UAA0B;AAAA,IAC9B,IAAI;AAAA,IACJ;AAAA,IACA,MAAM,QAAQ;AAAA,IACd,MAAM;AAAA,IACN,WAAW,QAAQ,UAAU,YAAY;AAAA,IACzC,YAAY,QAAQ,WAAW,YAAY;AAAA,IAC3C;AAAA,IACA,QAAQ,YAAY,UAAU;AAAA,IAC9B,GAAI,YAAY,EAAE,OAAO,UAAU,IAAI,CAAC;AAAA,IACxC,YAAY,wBAAwB;AAAA,MAClC,yBAAyB,QAAQ,iBAAiB;AAAA,MAClD,wBAAwB,QAAQ;AAAA,MAChC,GAAG,2BAA2B,IAAI,KAAK;AAAA,IACzC,CAAC;AAAA,EACH;AACA,QAAM,QAAyB;AAAA,IAC7B,IAAI;AAAA,IACJ,MAAM,QAAQ;AAAA,IACd,WAAW,QAAQ,UAAU,YAAY;AAAA,IACzC,YAAY,QAAQ,WAAW,YAAY;AAAA,IAC3C;AAAA,IACA,GAAI,QAAQ,SAAS,EAAE,UAAU,EAAE,QAAQ,QAAQ,OAAO,EAAE,IAAI,CAAC;AAAA,IACjE,OAAO,CAAC,OAAO;AAAA,EACjB;AAEA,MAAI,SAAS,CAAC,KAAK;AACnB,SAAO;AACT;AAEA,IAAI,uBAAuB;AAE3B,SAAS,yBAAyB;AAChC,0BAAwB;AACxB,SAAO,SAAS,oBAAoB;AACtC;AAiBO,SAAS,wBACd,OACA,KACiB;AACjB,QAAM,YACJ,iBAAiB,QACb,QACA,IAAI,MAAM,OAAO,SAAS,eAAe,CAAC;AAChD,SAAO,OAAO,OAAO,WAAW;AAAA,IAC9B,gBAAgB;AAAA,EAClB,CAAC;AACH;AAgBO,SAAS,uBAAuB,OAAwC;AAC7E,MACE,SACA,OAAO,UAAU,YACjB,oBAAoB,SACpB,aAAc,MAAuC,cAAc,GACnE;AACA,WAAQ,MAAyC;AAAA,EACnD;AAEA,SAAO;AACT;AAGO,SAAS,aAAa,OAAqC;AAChE,MAAI,CAAC,SAAS,OAAO,UAAU,UAAU;AACvC,WAAO;AAAA,EACT;AAEA,QAAM,YAAY;AAMlB,SACE,oBAAoB,UAAU,OAAO,KACrC,QAAQ,UAAU,KAAK,KACvB,OAAO,UAAU,UAAU,YAC3B,CAAC,MAAM,QAAQ,UAAU,KAAK,KAC9B,MAAM,QAAQ,UAAU,MAAM;AAElC;AAGO,SAAS,oBACd,OAC4B;AAC5B,SAAO,wBAAwB,UAAU,KAAK,EAAE;AAClD;AAGO,SAAS,wBACd,QACkC;AAClC,MACE,UACA,OAAO,WAAW,YAClB,MAAM,QAAS,OAAmC,MAAM,GACxD;AACA,WAAQ,OAAwD;AAAA,EAClE;AAEA,SAAO,CAAC;AACV;AAGO,SAAS,eAAe,OAA2C;AACxE,MAAI,iBAAiB,OAAO;AAC1B,WAAO;AAAA,MACL,MAAM,MAAM;AAAA,MACZ,SAAS,MAAM;AAAA,IACjB;AAAA,EACF;AAEA,SAAO;AAAA,IACL,MAAM;AAAA,IACN,SAAS,OAAO,KAAK;AAAA,EACvB;AACF;","names":["assistantMessages","failedSpans","latestAssistantMessageContent","messagesByRole","messagesToTranscriptEvents","spans","spansByKind","systemMessages","toolCalls","toolMessages","userMessages","normalized","details","message","type"]}

package/dist/index.d.mts

@@ -1,7 +1,7 @@
 import * as vitest from 'vitest';
 import { TestAPI } from 'vitest';
 import { Harness } from './harness.mjs';
-export { CreateHarnessOptions, CreateHarnessRunArgs, CreateToolCallSpansOptions, EnsureRunTraceOptions, HarnessContext, HarnessMetadata, HarnessResultLike, MaybePromise, SimpleHarnessResult, SimpleSpanEvent, SimpleSpanRecord, SimpleToolCallRecord, SimpleTraceRecord, attachHarnessRunToError, createFailedHarnessRun, createGenAiUsageAttributes, createHarness, createToolCallSpans, ensureRunTrace, getHarnessRunFromError, normalizeHarnessRun, normalizeSpanAttributes, normalizeSpanError, toJsonValue } from './harness.mjs';
+export { CreateHarnessOptions, CreateHarnessRunArgs, EnsureRunTraceOptions, HarnessContext, HarnessMetadata, HarnessResultLike, MaybePromise, SimpleHarnessResult, SimpleSpanEvent, SimpleSpanRecord, SimpleTraceRecord, SimpleTranscriptInput, attachHarnessRunToError, createFailedHarnessRun, createGenAiUsageAttributes, createHarness, ensureRunTrace, getHarnessRunFromError, normalizeHarnessRun, normalizeSpanAttributes, normalizeSpanError, toJsonValue } from './harness.mjs';
 import { JudgeContext, Judge, JudgeResult, JudgeAssessFn, JudgeOptions, JudgeAssessor, JudgeAssessWithAssessorFn } from './judges/types.mjs';
 export { BoundJudgeAssessor, JudgeAssessorOptions } from './judges/types.mjs';
 import { JudgeHarness } from './judges/judgeHarness.mjs';
@@ -11,8 +11,8 @@ export { FactualityJudge, FactualityJudgeChoice, FactualityJudgeConfig, Factuali
 export { StructuredOutputJudge, StructuredOutputJudgeConfig, StructuredOutputJudgeExpected, StructuredOutputJudgeOptions } from './judges/structuredOutputJudge.mjs';
 export { ToolCallJudge, ToolCallJudgeConfig, ToolCallJudgeExpectedTool, ToolCallJudgeOptions } from './judges/toolCallJudge.mjs';
 export { BaseMatcherConfig, FuzzyMatchOptions, MatchStrategy } from './internal/matchers.mjs';
-import { JsonValue, HarnessRun, NormalizedSession, ToolCallRecord } from '@vitest-evals/core';
-export { GenAiOperationName, GenAiOutputType, GenAiProviderName, GenAiSemanticAttributeKey, GenAiSemanticAttributes, GenAiTokenType, GenAiToolType, HarnessRun, HarnessRunError, JsonPrimitive, JsonValue, NormalizedMessage, NormalizedSession, NormalizedSpan, NormalizedSpanAttributeKey, NormalizedSpanAttributes, NormalizedSpanEvent, NormalizedTrace, OpenTelemetrySemanticAttributeKey, OpenTelemetrySemanticAttributes, TimingSummary, ToolCallRecord, UsageSummary, assistantMessages, failedSpans, latestAssistantMessageContent, messagesByRole, spans, spansByKind, systemMessages, toolCalls, toolMessages, userMessages } from '@vitest-evals/core';
+import { JsonValue, HarnessRun, NormalizedSession, ToolCall } from '@vitest-evals/core';
+export { GenAiOperationName, GenAiOutputType, GenAiProviderName, GenAiSemanticAttributeKey, GenAiSemanticAttributes, GenAiTokenType, GenAiToolType, HarnessRun, HarnessRunError, JsonPrimitive, JsonValue, NormalizedSession, NormalizedSpan, NormalizedSpanAttributeKey, NormalizedSpanAttributes, NormalizedSpanEvent, NormalizedTrace, OpenTelemetrySemanticAttributeKey, OpenTelemetrySemanticAttributes, TimingSummary, ToolCall, TranscriptEvent, TranscriptMessageContentPart, TranscriptMessageEvent, TranscriptMessageInput, TranscriptMessageTextPart, TranscriptMessageToolCall, TranscriptMessageToolCallPart, TranscriptMessageToolResultPart, TranscriptToolCallEvent, TranscriptToolResultEvent, UsageSummary, assistantMessages, failedSpans, latestAssistantMessageContent, messagesByRole, messagesToTranscriptEvents, spans, spansByKind, systemMessages, toolCalls, toolMessages, userMessages } from '@vitest-evals/core';
 import './internal/structuredOutputScorer.mjs';
 import './internal/scoring.mjs';
 import './internal/toolCallScorer.mjs';
@@ -24,7 +24,7 @@ type EvalTaskMeta = {
         })[];
         avgScore: number;
         output?: unknown;
-        toolCalls?: ToolCallRecord[];
+        toolCalls?: ToolCall[];
         thresholdFailed?: boolean;
     };
     harness?: {
@@ -149,7 +149,7 @@ type JudgeAssertionOptions<TJudgeOptions extends JudgeContext<any, any, any> = J
     /** Override or provide the app-facing output for the judge. */
     output?: JudgeAssertionOutput<TJudgeOptions>;
     /** Override or provide flattened tool calls for the judge. */
-    toolCalls?: ToolCallRecord[];
+    toolCalls?: ToolCall[];
     /** Override or provide the complete normalized harness run. */
     run?: HarnessRun<JudgeAssertionOutput<TJudgeOptions>>;
     /** Override or provide the normalized session transcript. */
@@ -219,7 +219,7 @@ declare module "vitest" {
  *     const result = await run("Refund invoice inv_123");
  *
  *     expect(result.output).toMatchObject({ status: "approved" });
- *     expect(toolCalls(result.session)).toHaveLength(2);
+ *     expect(toolCalls(result)).toHaveLength(2);
  *     await expect(result).toSatisfyJudge(FactualityJudge(), {
  *       expected: "Invoice inv_123 should be refunded.",
  *       threshold: 0.6,

package/dist/index.d.ts

@@ -1,7 +1,7 @@
 import * as vitest from 'vitest';
 import { TestAPI } from 'vitest';
 import { Harness } from './harness.js';
-export { CreateHarnessOptions, CreateHarnessRunArgs, CreateToolCallSpansOptions, EnsureRunTraceOptions, HarnessContext, HarnessMetadata, HarnessResultLike, MaybePromise, SimpleHarnessResult, SimpleSpanEvent, SimpleSpanRecord, SimpleToolCallRecord, SimpleTraceRecord, attachHarnessRunToError, createFailedHarnessRun, createGenAiUsageAttributes, createHarness, createToolCallSpans, ensureRunTrace, getHarnessRunFromError, normalizeHarnessRun, normalizeSpanAttributes, normalizeSpanError, toJsonValue } from './harness.js';
+export { CreateHarnessOptions, CreateHarnessRunArgs, EnsureRunTraceOptions, HarnessContext, HarnessMetadata, HarnessResultLike, MaybePromise, SimpleHarnessResult, SimpleSpanEvent, SimpleSpanRecord, SimpleTraceRecord, SimpleTranscriptInput, attachHarnessRunToError, createFailedHarnessRun, createGenAiUsageAttributes, createHarness, ensureRunTrace, getHarnessRunFromError, normalizeHarnessRun, normalizeSpanAttributes, normalizeSpanError, toJsonValue } from './harness.js';
 import { JudgeContext, Judge, JudgeResult, JudgeAssessFn, JudgeOptions, JudgeAssessor, JudgeAssessWithAssessorFn } from './judges/types.js';
 export { BoundJudgeAssessor, JudgeAssessorOptions } from './judges/types.js';
 import { JudgeHarness } from './judges/judgeHarness.js';
@@ -11,8 +11,8 @@ export { FactualityJudge, FactualityJudgeChoice, FactualityJudgeConfig, Factuali
 export { StructuredOutputJudge, StructuredOutputJudgeConfig, StructuredOutputJudgeExpected, StructuredOutputJudgeOptions } from './judges/structuredOutputJudge.js';
 export { ToolCallJudge, ToolCallJudgeConfig, ToolCallJudgeExpectedTool, ToolCallJudgeOptions } from './judges/toolCallJudge.js';
 export { BaseMatcherConfig, FuzzyMatchOptions, MatchStrategy } from './internal/matchers.js';
-import { JsonValue, HarnessRun, NormalizedSession, ToolCallRecord } from '@vitest-evals/core';
-export { GenAiOperationName, GenAiOutputType, GenAiProviderName, GenAiSemanticAttributeKey, GenAiSemanticAttributes, GenAiTokenType, GenAiToolType, HarnessRun, HarnessRunError, JsonPrimitive, JsonValue, NormalizedMessage, NormalizedSession, NormalizedSpan, NormalizedSpanAttributeKey, NormalizedSpanAttributes, NormalizedSpanEvent, NormalizedTrace, OpenTelemetrySemanticAttributeKey, OpenTelemetrySemanticAttributes, TimingSummary, ToolCallRecord, UsageSummary, assistantMessages, failedSpans, latestAssistantMessageContent, messagesByRole, spans, spansByKind, systemMessages, toolCalls, toolMessages, userMessages } from '@vitest-evals/core';
+import { JsonValue, HarnessRun, NormalizedSession, ToolCall } from '@vitest-evals/core';
+export { GenAiOperationName, GenAiOutputType, GenAiProviderName, GenAiSemanticAttributeKey, GenAiSemanticAttributes, GenAiTokenType, GenAiToolType, HarnessRun, HarnessRunError, JsonPrimitive, JsonValue, NormalizedSession, NormalizedSpan, NormalizedSpanAttributeKey, NormalizedSpanAttributes, NormalizedSpanEvent, NormalizedTrace, OpenTelemetrySemanticAttributeKey, OpenTelemetrySemanticAttributes, TimingSummary, ToolCall, TranscriptEvent, TranscriptMessageContentPart, TranscriptMessageEvent, TranscriptMessageInput, TranscriptMessageTextPart, TranscriptMessageToolCall, TranscriptMessageToolCallPart, TranscriptMessageToolResultPart, TranscriptToolCallEvent, TranscriptToolResultEvent, UsageSummary, assistantMessages, failedSpans, latestAssistantMessageContent, messagesByRole, messagesToTranscriptEvents, spans, spansByKind, systemMessages, toolCalls, toolMessages, userMessages } from '@vitest-evals/core';
 import './internal/structuredOutputScorer.js';
 import './internal/scoring.js';
 import './internal/toolCallScorer.js';
@@ -24,7 +24,7 @@ type EvalTaskMeta = {
         })[];
         avgScore: number;
         output?: unknown;
-        toolCalls?: ToolCallRecord[];
+        toolCalls?: ToolCall[];
         thresholdFailed?: boolean;
     };
     harness?: {
@@ -149,7 +149,7 @@ type JudgeAssertionOptions<TJudgeOptions extends JudgeContext<any, any, any> = J
     /** Override or provide the app-facing output for the judge. */
     output?: JudgeAssertionOutput<TJudgeOptions>;
     /** Override or provide flattened tool calls for the judge. */
-    toolCalls?: ToolCallRecord[];
+    toolCalls?: ToolCall[];
     /** Override or provide the complete normalized harness run. */
     run?: HarnessRun<JudgeAssertionOutput<TJudgeOptions>>;
     /** Override or provide the normalized session transcript. */
@@ -219,7 +219,7 @@ declare module "vitest" {
  *     const result = await run("Refund invoice inv_123");
  *
  *     expect(result.output).toMatchObject({ status: "approved" });
- *     expect(toolCalls(result.session)).toHaveLength(2);
+ *     expect(toolCalls(result)).toHaveLength(2);
  *     await expect(result).toSatisfyJudge(FactualityJudge(), {
  *       expected: "Invoice inv_123 should be refunded.",
  *       threshold: 0.6,

package/dist/index.js

@@ -30,7 +30,6 @@ __export(index_exports, {
   createHarness: () => createHarness,
   createJudge: () => createJudge,
   createJudgeHarness: () => createJudgeHarness,
-  createToolCallSpans: () => createToolCallSpans,
   describeEval: () => describeEval,
   ensureRunTrace: () => ensureRunTrace,
   failedSpans: () => import_core2.failedSpans,
@@ -38,6 +37,7 @@ __export(index_exports, {
   getHarnessRunFromError: () => getHarnessRunFromError,
   latestAssistantMessageContent: () => import_core2.latestAssistantMessageContent,
   messagesByRole: () => import_core2.messagesByRole,
+  messagesToTranscriptEvents: () => import_core2.messagesToTranscriptEvents,
   normalizeHarnessRun: () => normalizeHarnessRun,
   normalizeSpanAttributes: () => normalizeSpanAttributes,
   normalizeSpanError: () => normalizeSpanError,
@@ -179,14 +179,24 @@ function normalizeHarnessRun(input, result, context) {
     }
     return result;
   }
+  if ("toolCalls" in result) {
+    throw new TypeError(
+      'createHarness results do not accept top-level toolCalls. Return ordered session events with type: "tool_call" and type: "tool_result" entries instead.'
+    );
+  }
   const output = result.output;
-  const toolCalls3 = normalizeSimpleToolCalls(result.toolCalls);
   const usage = result.usage ?? {};
-  const messages = result.messages ?? createDefaultSessionMessages({
-    input,
-    output,
-    toolCalls: toolCalls3
-  });
+  const events = normalizeTranscriptInput(result);
+  if (!events) {
+    throw new TypeError(
+      "createHarness results must include ordered events or messages. Return a full HarnessRun or a lightweight result with events/messages."
+    );
+  }
+  if (events.length === 0) {
+    throw new TypeError(
+      "createHarness results must include at least one transcript event. Return ordered events or message transport inputs that normalize into events."
+    );
+  }
   const metadata = result.metadata ? normalizeMetadata(result.metadata) : void 0;
   const artifacts = normalizeMergedArtifacts(
     context?.artifacts,
@@ -195,7 +205,7 @@ function normalizeHarnessRun(input, result, context) {
   const traces = normalizeSimpleTraces(result.traces);
   return {
     session: {
-      messages,
+      events,
       ...usage.provider ? { provider: usage.provider } : {},
       ...usage.model ? { model: usage.model } : {},
       ...metadata ? { metadata } : {}
@@ -208,12 +218,24 @@ function normalizeHarnessRun(input, result, context) {
     errors: normalizeSimpleErrors(result.errors)
   };
 }
+function normalizeTranscriptInput(result) {
+  if ("events" in result && Array.isArray(result.events)) {
+    return result.events.map((event) => import_core.TranscriptEventSchema.parse(event));
+  }
+  if ("messages" in result && Array.isArray(result.messages)) {
+    return (0, import_core.messagesToTranscriptEvents)(result.messages).map(
+      (event) => import_core.TranscriptEventSchema.parse(event)
+    );
+  }
+  return void 0;
+}
 function createFailedHarnessRun(input, error, options = {}) {
   const artifacts = options.artifacts;
   return {
     session: {
-      messages: [
+      events: [
         {
+          type: "message",
           role: "user",
           content: normalizeContent(input)
         }
@@ -224,67 +246,6 @@ function createFailedHarnessRun(input, error, options = {}) {
     errors: [serializeError(error)]
   };
 }
-function createDefaultSessionMessages({
-  input,
-  output,
-  toolCalls: normalizedToolCalls
-}) {
-  const messages = [
-    {
-      role: "user",
-      content: normalizeContent(input)
-    }
-  ];
-  if (output !== void 0 || normalizedToolCalls.length > 0) {
-    messages.push({
-      role: "assistant",
-      ...output !== void 0 ? { content: normalizeContent(output) } : {},
-      ...normalizedToolCalls.length > 0 ? { toolCalls: normalizedToolCalls } : {}
-    });
-  }
-  return messages;
-}
-function normalizeSimpleToolCalls(calls) {
-  return (calls ?? []).map((call) => {
-    const {
-      arguments: rawArguments,
-      result: rawResult,
-      error: rawError,
-      metadata: rawMetadata,
-      ...toolCall
-    } = call;
-    const args = normalizeToolCallArguments(rawArguments);
-    const result = toJsonValue(rawResult);
-    const error = normalizeToolCallError(rawError);
-    const metadata = rawMetadata ? normalizeMetadata(rawMetadata) : void 0;
-    return {
-      ...toolCall,
-      ...args ? { arguments: args } : {},
-      ...result !== void 0 ? { result } : {},
-      ...error ? { error } : {},
-      ...metadata ? { metadata } : {}
-    };
-  });
-}
-function normalizeToolCallArguments(value) {
-  if (value === void 0) {
-    return void 0;
-  }
-  const normalized = toJsonValue(value);
-  return normalized && typeof normalized === "object" && !Array.isArray(normalized) ? normalized : void 0;
-}
-function normalizeToolCallError(value) {
-  if (value === void 0) {
-    return void 0;
-  }
-  const serialized = serializeError(value);
-  const { message, type, ...details } = serialized;
-  return {
-    ...details,
-    message: typeof message === "string" ? message : String(message),
-    ...typeof type === "string" ? { type } : {}
-  };
-}
 function normalizeMergedArtifacts(contextArtifacts, resultArtifacts) {
   const artifacts = {
     ...contextArtifacts ?? {},
@@ -410,32 +371,6 @@ function createGenAiUsageAttributes(usage, options = {}) {
     "gen_ai.usage.reasoning.output_tokens": usage?.reasoningTokens
   };
 }
-function createToolCallSpans(calls, options = {}) {
-  return calls.map((call, index) => {
-    const spanError = call.error ? normalizeSpanError(call.error) : void 0;
-    const spanId = options.spanIdPrefix ? `${options.spanIdPrefix}:${index + 1}` : call.id;
-    return {
-      ...spanId ? { id: spanId } : {},
-      ...options.traceId ? { traceId: options.traceId } : {},
-      ...options.parentId ? { parentId: options.parentId } : {},
-      name: call.name,
-      kind: "tool",
-      ...call.startedAt ? { startedAt: call.startedAt } : {},
-      ...call.finishedAt ? { finishedAt: call.finishedAt } : {},
-      ...call.durationMs !== void 0 ? { durationMs: call.durationMs } : {},
-      status: spanError ? "error" : "ok",
-      ...spanError ? { error: spanError } : {},
-      attributes: normalizeSpanAttributes({
-        "gen_ai.operation.name": "execute_tool",
-        "gen_ai.tool.name": call.name,
-        "gen_ai.tool.type": "function",
-        ...call.id ? { "gen_ai.tool.call.id": call.id } : {},
-        ...call.arguments !== void 0 ? { "gen_ai.tool.call.arguments": call.arguments } : {},
-        ...call.result !== void 0 ? { "gen_ai.tool.call.result": call.result } : {}
-      })
-    };
-  });
-}
 function ensureRunTrace(run, options) {
   if ((0, import_core.spans)(run).length > 0) {
     return void 0;
@@ -460,11 +395,6 @@ function ensureRunTrace(run, options) {
       ...createGenAiUsageAttributes(run.usage)
     })
   };
-  const toolSpans = createToolCallSpans((0, import_core.toolCalls)(run.session), {
-    traceId,
-    parentId: rootSpanId,
-    spanIdPrefix: `${traceId}:tool`
-  });
   const trace = {
     id: traceId,
     name: options.name,
@@ -472,7 +402,7 @@ function ensureRunTrace(run, options) {
     finishedAt: options.finishedAt.toISOString(),
     durationMs,
     ...options.source ? { metadata: { source: options.source } } : {},
-    spans: [runSpan, ...toolSpans]
+    spans: [runSpan]
   };
   run.traces = [trace];
   return trace;
@@ -502,7 +432,7 @@ function isHarnessRun(value) {
   return isNormalizedSession(candidate.session) && Boolean(candidate.usage) && typeof candidate.usage === "object" && !Array.isArray(candidate.usage) && Array.isArray(candidate.errors);
 }
 function isNormalizedSession(value) {
-  return Boolean(value) && typeof value === "object" && value !== null && "messages" in value && Array.isArray(value.messages);
+  return import_core.NormalizedSessionSchema.safeParse(value).success;
 }
 function serializeError(error) {
   if (error instanceof Error) {
@@ -522,7 +452,10 @@ function createJudgeHarness(options) {
   return createHarness({
     name: options.name ?? "judge-harness",
     run: async ({ input, signal }) => {
-      return normalizeJudgeHarnessResult(await options.run(input, { signal }));
+      return normalizeJudgeHarnessResult(
+        input,
+        await options.run(input, { signal })
+      );
     }
   });
 }
@@ -545,17 +478,14 @@ function createRunJudge(judgeHarness, signal) {
     signal: options?.signal ?? signal
   });
 }
-function normalizeJudgeHarnessResult(result) {
+function normalizeJudgeHarnessResult(input, result) {
   if (isHarnessRun(result)) {
     return result;
   }
-  if (hasOutputField(result)) {
-    return {
-      output: normalizeJudgeHarnessOutput(result.output)
-    };
-  }
+  const output = hasOutputField(result) ? normalizeJudgeHarnessOutput(result.output) : normalizeJudgeHarnessOutput(result);
   return {
-    output: normalizeJudgeHarnessOutput(result)
+    output,
+    messages: createJudgeHarnessMessages(input, output)
   };
 }
 function hasOutputField(value) {
@@ -567,6 +497,13 @@ function normalizeJudgeHarnessOutput(value) {
   }
   return normalizeContent(value);
 }
+function createJudgeHarnessMessages(input, output) {
+  return [
+    ...input.system ? [{ role: "system", content: input.system }] : [],
+    { role: "user", content: input.prompt },
+    ...output !== void 0 ? [{ role: "assistant", content: output }] : []
+  ];
+}
 function resolveJudgeHarnessAssistantOutput(run) {
   return (0, import_core2.latestAssistantMessageContent)(run.session) ?? "";
 }
@@ -1694,23 +1631,25 @@ function resolveJudgeAssertionOutput(received, run, explicitOutput) {
   return normalizeJudgeJsonValue(received);
 }
 function createSyntheticJudgeSession(received, options) {
-  const messages = [];
+  const events = [];
   const userContent = normalizeJudgeJsonValue(options.input);
   if (userContent !== void 0) {
-    messages.push({
+    events.push({
+      type: "message",
       role: "user",
       content: userContent
     });
   }
   const assistantContent = normalizeJudgeJsonValue(received);
   if (assistantContent !== void 0) {
-    messages.push({
+    events.push({
+      type: "message",
       role: "assistant",
       content: assistantContent
     });
   }
   return {
-    messages
+    events
   };
 }
 function inferJudgeOutputValue(received, session) {
@@ -1718,7 +1657,7 @@ function inferJudgeOutputValue(received, session) {
     return received.output;
   }
   if (isNormalizedSession(received)) {
-    return resolveAssistantOutput(session) ?? normalizeJudgeJsonValue(received.messages);
+    return resolveAssistantOutput(session) ?? normalizeJudgeJsonValue(received.events);
   }
   return normalizeJudgeJsonValue(received);
 }
@@ -1792,7 +1731,6 @@ function createJudge(nameOrConfig, assessOrAssessor, assess) {
   createHarness,
   createJudge,
   createJudgeHarness,
-  createToolCallSpans,
   describeEval,
   ensureRunTrace,
   failedSpans,
@@ -1800,6 +1738,7 @@ function createJudge(nameOrConfig, assessOrAssessor, assess) {
   getHarnessRunFromError,
   latestAssistantMessageContent,
   messagesByRole,
+  messagesToTranscriptEvents,
   normalizeHarnessRun,
   normalizeSpanAttributes,
   normalizeSpanError,