vitest-evals 0.13.1 → 0.14.0
This diff represents the content of publicly available package versions that have been released to one of the supported registries. The information contained in this diff is provided for informational purposes only and reflects changes between package versions as they appear in their respective public registries.
- package/README.md +53 -8
- package/dist/harness.d.mts +55 -39
- package/dist/harness.d.ts +55 -39
- package/dist/harness.js +34 -104
- package/dist/harness.js.map +1 -1
- package/dist/harness.mjs +37 -104
- package/dist/harness.mjs.map +1 -1
- package/dist/index.d.mts +5 -5
- package/dist/index.d.ts +5 -5
- package/dist/index.js +56 -117
- package/dist/index.js.map +1 -1
- package/dist/index.mjs +59 -117
- package/dist/index.mjs.map +1 -1
- package/dist/internal/scoring.d.mts +2 -2
- package/dist/internal/scoring.d.ts +2 -2
- package/dist/internal/scoring.js.map +1 -1
- package/dist/internal/toolCallScorer.js.map +1 -1
- package/dist/internal/toolCallScorer.mjs +4 -1
- package/dist/internal/toolCallScorer.mjs.map +1 -1
- package/dist/judges/factualityJudge.js.map +1 -1
- package/dist/judges/factualityJudge.mjs +4 -1
- package/dist/judges/factualityJudge.mjs.map +1 -1
- package/dist/judges/index.js +47 -110
- package/dist/judges/index.js.map +1 -1
- package/dist/judges/index.mjs +51 -111
- package/dist/judges/index.mjs.map +1 -1
- package/dist/judges/judgeHarness.js +47 -110
- package/dist/judges/judgeHarness.js.map +1 -1
- package/dist/judges/judgeHarness.mjs +51 -111
- package/dist/judges/judgeHarness.mjs.map +1 -1
- package/dist/judges/toolCallJudge.js.map +1 -1
- package/dist/judges/toolCallJudge.mjs +4 -1
- package/dist/judges/toolCallJudge.mjs.map +1 -1
- package/dist/judges/types.d.mts +2 -2
- package/dist/judges/types.d.ts +2 -2
- package/dist/judges/types.js.map +1 -1
- package/dist/legacy/scorers/index.js.map +1 -1
- package/dist/legacy/scorers/index.mjs +4 -1
- package/dist/legacy/scorers/index.mjs.map +1 -1
- package/dist/legacy/scorers/toolCallScorer.js.map +1 -1
- package/dist/legacy/scorers/toolCallScorer.mjs +4 -1
- package/dist/legacy/scorers/toolCallScorer.mjs.map +1 -1
- package/dist/legacy/shared.d.mts +1 -8
- package/dist/legacy/shared.d.ts +1 -8
- package/dist/legacy/shared.js.map +1 -1
- package/dist/legacy.js +15 -1
- package/dist/legacy.js.map +1 -1
- package/dist/legacy.mjs +19 -2
- package/dist/legacy.mjs.map +1 -1
- package/dist/reporter.d.mts +0 -3
- package/dist/reporter.d.ts +0 -3
- package/dist/reporter.js +10 -40
- package/dist/reporter.js.map +1 -1
- package/dist/reporter.mjs +14 -41
- package/dist/reporter.mjs.map +1 -1
- package/package.json +3 -3
package/dist/harness.js.map
CHANGED
|
@@ -1 +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)).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;AAAA;AAAA;AAAA;AAAA;AAAA;AAAA;AAAA;AAAA;AAAA;AAAA;AAAA;AAAA;AAAA;AAAA;AAAA;AAAA;AAAA;AAAA;AAAA;AAAA;AAAA;AAAA;AAAA;AAAA;AAAA;AAAA;AAAA;AAAA;AAAA;AAAA;AAAA;AAAA,kBAWO;AAkBP,IAAAA,eAWO;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,QAAMC,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,QAAMC,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,UAAMC,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,UAAI,mBAAM,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,wBAAoB,uBAAU,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":["import_core","normalized","toolCalls","spans","details","message","type"]}
|
|
1
|
+
{"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;AAAA;AAAA;AAAA;AAAA;AAAA;AAAA;AAAA;AAAA;AAAA;AAAA;AAAA;AAAA;AAAA;AAAA;AAAA;AAAA;AAAA;AAAA;AAAA;AAAA;AAAA;AAAA;AAAA;AAAA;AAAA;AAAA;AAAA;AAAA;AAAA;AAAA;AAAA;AAAA,kBAcO;AA0BP,IAAAA,eAYO;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,kCAAsB,MAAM,KAAK,CAAC;AAAA,EACxE;AAEA,MAAI,cAAc,UAAU,MAAM,QAAQ,OAAO,QAAQ,GAAG;AAC1D,eAAO,wCAA2B,OAAO,QAAQ,EAAE;AAAA,MAAI,CAAC,UACtD,kCAAsB,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,QAAMC,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,UAAMC,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,UAAI,mBAAM,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,oCAAwB,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":["import_core","normalized","spans","details","message","type"]}
|
package/dist/harness.mjs
CHANGED
|
@@ -1,13 +1,16 @@
|
|
|
1
1
|
// src/harness.ts
|
|
2
2
|
import {
|
|
3
|
+
messagesToTranscriptEvents,
|
|
4
|
+
NormalizedSessionSchema,
|
|
3
5
|
spans,
|
|
4
|
-
|
|
6
|
+
TranscriptEventSchema
|
|
5
7
|
} from "@vitest-evals/core";
|
|
6
8
|
import {
|
|
7
9
|
assistantMessages as assistantMessages2,
|
|
8
10
|
failedSpans as failedSpans2,
|
|
9
11
|
latestAssistantMessageContent as latestAssistantMessageContent2,
|
|
10
12
|
messagesByRole as messagesByRole2,
|
|
13
|
+
messagesToTranscriptEvents as messagesToTranscriptEvents2,
|
|
11
14
|
spans as spans2,
|
|
12
15
|
spansByKind as spansByKind2,
|
|
13
16
|
systemMessages as systemMessages2,
|
|
@@ -139,14 +142,24 @@ function normalizeHarnessRun(input, result, context) {
|
|
|
139
142
|
}
|
|
140
143
|
return result;
|
|
141
144
|
}
|
|
145
|
+
if ("toolCalls" in result) {
|
|
146
|
+
throw new TypeError(
|
|
147
|
+
'createHarness results do not accept top-level toolCalls. Return ordered session events with type: "tool_call" and type: "tool_result" entries instead.'
|
|
148
|
+
);
|
|
149
|
+
}
|
|
142
150
|
const output = result.output;
|
|
143
|
-
const toolCalls3 = normalizeSimpleToolCalls(result.toolCalls);
|
|
144
151
|
const usage = result.usage ?? {};
|
|
145
|
-
const
|
|
146
|
-
|
|
147
|
-
|
|
148
|
-
|
|
149
|
-
|
|
152
|
+
const events = normalizeTranscriptInput(result);
|
|
153
|
+
if (!events) {
|
|
154
|
+
throw new TypeError(
|
|
155
|
+
"createHarness results must include ordered events or messages. Return a full HarnessRun or a lightweight result with events/messages."
|
|
156
|
+
);
|
|
157
|
+
}
|
|
158
|
+
if (events.length === 0) {
|
|
159
|
+
throw new TypeError(
|
|
160
|
+
"createHarness results must include at least one transcript event. Return ordered events or message transport inputs that normalize into events."
|
|
161
|
+
);
|
|
162
|
+
}
|
|
150
163
|
const metadata = result.metadata ? normalizeMetadata(result.metadata) : void 0;
|
|
151
164
|
const artifacts = normalizeMergedArtifacts(
|
|
152
165
|
context?.artifacts,
|
|
@@ -155,7 +168,7 @@ function normalizeHarnessRun(input, result, context) {
|
|
|
155
168
|
const traces = normalizeSimpleTraces(result.traces);
|
|
156
169
|
return {
|
|
157
170
|
session: {
|
|
158
|
-
|
|
171
|
+
events,
|
|
159
172
|
...usage.provider ? { provider: usage.provider } : {},
|
|
160
173
|
...usage.model ? { model: usage.model } : {},
|
|
161
174
|
...metadata ? { metadata } : {}
|
|
@@ -168,12 +181,24 @@ function normalizeHarnessRun(input, result, context) {
|
|
|
168
181
|
errors: normalizeSimpleErrors(result.errors)
|
|
169
182
|
};
|
|
170
183
|
}
|
|
184
|
+
function normalizeTranscriptInput(result) {
|
|
185
|
+
if ("events" in result && Array.isArray(result.events)) {
|
|
186
|
+
return result.events.map((event) => TranscriptEventSchema.parse(event));
|
|
187
|
+
}
|
|
188
|
+
if ("messages" in result && Array.isArray(result.messages)) {
|
|
189
|
+
return messagesToTranscriptEvents(result.messages).map(
|
|
190
|
+
(event) => TranscriptEventSchema.parse(event)
|
|
191
|
+
);
|
|
192
|
+
}
|
|
193
|
+
return void 0;
|
|
194
|
+
}
|
|
171
195
|
function createFailedHarnessRun(input, error, options = {}) {
|
|
172
196
|
const artifacts = options.artifacts;
|
|
173
197
|
return {
|
|
174
198
|
session: {
|
|
175
|
-
|
|
199
|
+
events: [
|
|
176
200
|
{
|
|
201
|
+
type: "message",
|
|
177
202
|
role: "user",
|
|
178
203
|
content: normalizeContent(input)
|
|
179
204
|
}
|
|
@@ -184,67 +209,6 @@ function createFailedHarnessRun(input, error, options = {}) {
|
|
|
184
209
|
errors: [serializeError(error)]
|
|
185
210
|
};
|
|
186
211
|
}
|
|
187
|
-
function createDefaultSessionMessages({
|
|
188
|
-
input,
|
|
189
|
-
output,
|
|
190
|
-
toolCalls: normalizedToolCalls
|
|
191
|
-
}) {
|
|
192
|
-
const messages = [
|
|
193
|
-
{
|
|
194
|
-
role: "user",
|
|
195
|
-
content: normalizeContent(input)
|
|
196
|
-
}
|
|
197
|
-
];
|
|
198
|
-
if (output !== void 0 || normalizedToolCalls.length > 0) {
|
|
199
|
-
messages.push({
|
|
200
|
-
role: "assistant",
|
|
201
|
-
...output !== void 0 ? { content: normalizeContent(output) } : {},
|
|
202
|
-
...normalizedToolCalls.length > 0 ? { toolCalls: normalizedToolCalls } : {}
|
|
203
|
-
});
|
|
204
|
-
}
|
|
205
|
-
return messages;
|
|
206
|
-
}
|
|
207
|
-
function normalizeSimpleToolCalls(calls) {
|
|
208
|
-
return (calls ?? []).map((call) => {
|
|
209
|
-
const {
|
|
210
|
-
arguments: rawArguments,
|
|
211
|
-
result: rawResult,
|
|
212
|
-
error: rawError,
|
|
213
|
-
metadata: rawMetadata,
|
|
214
|
-
...toolCall
|
|
215
|
-
} = call;
|
|
216
|
-
const args = normalizeToolCallArguments(rawArguments);
|
|
217
|
-
const result = toJsonValue(rawResult);
|
|
218
|
-
const error = normalizeToolCallError(rawError);
|
|
219
|
-
const metadata = rawMetadata ? normalizeMetadata(rawMetadata) : void 0;
|
|
220
|
-
return {
|
|
221
|
-
...toolCall,
|
|
222
|
-
...args ? { arguments: args } : {},
|
|
223
|
-
...result !== void 0 ? { result } : {},
|
|
224
|
-
...error ? { error } : {},
|
|
225
|
-
...metadata ? { metadata } : {}
|
|
226
|
-
};
|
|
227
|
-
});
|
|
228
|
-
}
|
|
229
|
-
function normalizeToolCallArguments(value) {
|
|
230
|
-
if (value === void 0) {
|
|
231
|
-
return void 0;
|
|
232
|
-
}
|
|
233
|
-
const normalized = toJsonValue(value);
|
|
234
|
-
return normalized && typeof normalized === "object" && !Array.isArray(normalized) ? normalized : void 0;
|
|
235
|
-
}
|
|
236
|
-
function normalizeToolCallError(value) {
|
|
237
|
-
if (value === void 0) {
|
|
238
|
-
return void 0;
|
|
239
|
-
}
|
|
240
|
-
const serialized = serializeError(value);
|
|
241
|
-
const { message, type, ...details } = serialized;
|
|
242
|
-
return {
|
|
243
|
-
...details,
|
|
244
|
-
message: typeof message === "string" ? message : String(message),
|
|
245
|
-
...typeof type === "string" ? { type } : {}
|
|
246
|
-
};
|
|
247
|
-
}
|
|
248
212
|
function normalizeMergedArtifacts(contextArtifacts, resultArtifacts) {
|
|
249
213
|
const artifacts = {
|
|
250
214
|
...contextArtifacts ?? {},
|
|
@@ -370,32 +334,6 @@ function createGenAiUsageAttributes(usage, options = {}) {
|
|
|
370
334
|
"gen_ai.usage.reasoning.output_tokens": usage?.reasoningTokens
|
|
371
335
|
};
|
|
372
336
|
}
|
|
373
|
-
function createToolCallSpans(calls, options = {}) {
|
|
374
|
-
return calls.map((call, index) => {
|
|
375
|
-
const spanError = call.error ? normalizeSpanError(call.error) : void 0;
|
|
376
|
-
const spanId = options.spanIdPrefix ? `${options.spanIdPrefix}:${index + 1}` : call.id;
|
|
377
|
-
return {
|
|
378
|
-
...spanId ? { id: spanId } : {},
|
|
379
|
-
...options.traceId ? { traceId: options.traceId } : {},
|
|
380
|
-
...options.parentId ? { parentId: options.parentId } : {},
|
|
381
|
-
name: call.name,
|
|
382
|
-
kind: "tool",
|
|
383
|
-
...call.startedAt ? { startedAt: call.startedAt } : {},
|
|
384
|
-
...call.finishedAt ? { finishedAt: call.finishedAt } : {},
|
|
385
|
-
...call.durationMs !== void 0 ? { durationMs: call.durationMs } : {},
|
|
386
|
-
status: spanError ? "error" : "ok",
|
|
387
|
-
...spanError ? { error: spanError } : {},
|
|
388
|
-
attributes: normalizeSpanAttributes({
|
|
389
|
-
"gen_ai.operation.name": "execute_tool",
|
|
390
|
-
"gen_ai.tool.name": call.name,
|
|
391
|
-
"gen_ai.tool.type": "function",
|
|
392
|
-
...call.id ? { "gen_ai.tool.call.id": call.id } : {},
|
|
393
|
-
...call.arguments !== void 0 ? { "gen_ai.tool.call.arguments": call.arguments } : {},
|
|
394
|
-
...call.result !== void 0 ? { "gen_ai.tool.call.result": call.result } : {}
|
|
395
|
-
})
|
|
396
|
-
};
|
|
397
|
-
});
|
|
398
|
-
}
|
|
399
337
|
function ensureRunTrace(run, options) {
|
|
400
338
|
if (spans(run).length > 0) {
|
|
401
339
|
return void 0;
|
|
@@ -420,11 +358,6 @@ function ensureRunTrace(run, options) {
|
|
|
420
358
|
...createGenAiUsageAttributes(run.usage)
|
|
421
359
|
})
|
|
422
360
|
};
|
|
423
|
-
const toolSpans = createToolCallSpans(toolCalls(run.session), {
|
|
424
|
-
traceId,
|
|
425
|
-
parentId: rootSpanId,
|
|
426
|
-
spanIdPrefix: `${traceId}:tool`
|
|
427
|
-
});
|
|
428
361
|
const trace = {
|
|
429
362
|
id: traceId,
|
|
430
363
|
name: options.name,
|
|
@@ -432,7 +365,7 @@ function ensureRunTrace(run, options) {
|
|
|
432
365
|
finishedAt: options.finishedAt.toISOString(),
|
|
433
366
|
durationMs,
|
|
434
367
|
...options.source ? { metadata: { source: options.source } } : {},
|
|
435
|
-
spans: [runSpan
|
|
368
|
+
spans: [runSpan]
|
|
436
369
|
};
|
|
437
370
|
run.traces = [trace];
|
|
438
371
|
return trace;
|
|
@@ -462,7 +395,7 @@ function isHarnessRun(value) {
|
|
|
462
395
|
return isNormalizedSession(candidate.session) && Boolean(candidate.usage) && typeof candidate.usage === "object" && !Array.isArray(candidate.usage) && Array.isArray(candidate.errors);
|
|
463
396
|
}
|
|
464
397
|
function isNormalizedSession(value) {
|
|
465
|
-
return
|
|
398
|
+
return NormalizedSessionSchema.safeParse(value).success;
|
|
466
399
|
}
|
|
467
400
|
function resolveHarnessRunErrors(result) {
|
|
468
401
|
if (result && typeof result === "object" && Array.isArray(result.errors)) {
|
|
@@ -488,7 +421,6 @@ export {
|
|
|
488
421
|
createFailedHarnessRun,
|
|
489
422
|
createGenAiUsageAttributes,
|
|
490
423
|
createHarness,
|
|
491
|
-
createToolCallSpans,
|
|
492
424
|
ensureRunTrace,
|
|
493
425
|
failedSpans2 as failedSpans,
|
|
494
426
|
getHarnessRunFromError,
|
|
@@ -497,6 +429,7 @@ export {
|
|
|
497
429
|
isNormalizedSession,
|
|
498
430
|
latestAssistantMessageContent2 as latestAssistantMessageContent,
|
|
499
431
|
messagesByRole2 as messagesByRole,
|
|
432
|
+
messagesToTranscriptEvents2 as messagesToTranscriptEvents,
|
|
500
433
|
normalizeContent,
|
|
501
434
|
normalizeHarnessRun,
|
|
502
435
|
normalizeMetadata,
|