@vitest-evals/core 0.13.0 → 0.14.0

package/dist/index.d.mts

@@ -1,5 +1,5 @@
-import { J as JsonValue } from './workspace-BNgjyW7W.mjs';
-export { C as CollectReportWorkspaceOptions, a as JsonObjectSchema, d as JsonPrimitive, b as JsonPrimitiveSchema, c as JsonValueSchema, R as REPORT_WORKSPACE_SCHEMA_VERSION, t as ReportCase, q as ReportCaseSchema, u as ReportRun, r as ReportRunSchema, v as ReportWorkspace, w as ReportWorkspaceInput, s as ReportWorkspaceSchema, i as VitestJsonAssertion, V as VitestJsonAssertionSchema, j as VitestJsonFile, e as VitestJsonFileSchema, k as VitestJsonLocation, f as VitestJsonLocationSchema, l as VitestJsonReport, g as VitestJsonReportSchema, m as VitestJsonStatus, h as VitestJsonStatusSchema, n as collectReportWorkspace, o as parseReportWorkspace, p as parseVitestJsonReport } from './workspace-BNgjyW7W.mjs';
+import { J as JsonValue } from './workspace-DSZKph1P.mjs';
+export { C as CollectReportWorkspaceOptions, a as JsonObjectSchema, d as JsonPrimitive, b as JsonPrimitiveSchema, c as JsonValueSchema, R as REPORT_WORKSPACE_SCHEMA_VERSION, t as ReportCase, q as ReportCaseSchema, u as ReportRun, r as ReportRunSchema, v as ReportWorkspace, w as ReportWorkspaceInput, s as ReportWorkspaceSchema, i as VitestJsonAssertion, V as VitestJsonAssertionSchema, j as VitestJsonFile, e as VitestJsonFileSchema, k as VitestJsonLocation, f as VitestJsonLocationSchema, l as VitestJsonReport, g as VitestJsonReportSchema, m as VitestJsonStatus, h as VitestJsonStatusSchema, n as collectReportWorkspace, o as parseReportWorkspace, p as parseVitestJsonReport } from './workspace-DSZKph1P.mjs';
 import { z } from 'zod';
 
 /** Well-known OpenTelemetry GenAI operation names. */
@@ -85,15 +85,212 @@ type NormalizedSpanAttributes = OpenTelemetrySemanticAttributes & {
     [key: string]: JsonValue | undefined;
 };
 
+/** Normalized error object captured in a tool result or trace span. */
+declare const NormalizedErrorSchema: z.ZodObject<{
+    message: z.ZodString;
+    type: z.ZodOptional<z.ZodString>;
+}, z.core.$catchall<z.ZodType<JsonValue, unknown, z.core.$ZodTypeInternals<JsonValue, unknown>>>>;
 /** Normalized error attached to tool calls and spans. */
-type NormalizedError = {
-    /** Human-readable failure message. */
-    message: string;
-    /** Optional provider or runtime error type. */
-    type?: string;
-    /** Extra JSON-safe error metadata. */
-    [key: string]: JsonValue | undefined;
+type NormalizedError = z.infer<typeof NormalizedErrorSchema>;
+
+/** Tool-call event captured in a harness transcript. */
+declare const TranscriptToolCallEventSchema: z.ZodObject<{
+    type: z.ZodLiteral<"tool_call">;
+    id: z.ZodString;
+    name: z.ZodString;
+    arguments: z.ZodOptional<z.ZodRecord<z.ZodString, z.ZodType<JsonValue, unknown, z.core.$ZodTypeInternals<JsonValue, unknown>>>>;
+    startedAt: z.ZodOptional<z.ZodString>;
+    finishedAt: z.ZodOptional<z.ZodString>;
+    durationMs: z.ZodOptional<z.ZodNumber>;
+    metadata: z.ZodOptional<z.ZodRecord<z.ZodString, z.ZodType<JsonValue, unknown, z.core.$ZodTypeInternals<JsonValue, unknown>>>>;
+}, z.core.$strict>;
+/** Message event captured in a harness transcript. */
+declare const TranscriptMessageEventSchema: z.ZodObject<{
+    type: z.ZodLiteral<"message">;
+    role: z.ZodEnum<{
+        system: "system";
+        user: "user";
+        assistant: "assistant";
+    }>;
+    content: z.ZodOptional<z.ZodType<JsonValue, unknown, z.core.$ZodTypeInternals<JsonValue, unknown>>>;
+    metadata: z.ZodOptional<z.ZodRecord<z.ZodString, z.ZodType<JsonValue, unknown, z.core.$ZodTypeInternals<JsonValue, unknown>>>>;
+}, z.core.$strict>;
+/** Tool-result event captured in a harness transcript. */
+declare const TranscriptToolResultEventSchema: z.ZodObject<{
+    type: z.ZodLiteral<"tool_result">;
+    toolCallId: z.ZodString;
+    name: z.ZodOptional<z.ZodString>;
+    content: z.ZodOptional<z.ZodType<JsonValue, unknown, z.core.$ZodTypeInternals<JsonValue, unknown>>>;
+    error: z.ZodOptional<z.ZodObject<{
+        message: z.ZodString;
+        type: z.ZodOptional<z.ZodString>;
+    }, z.core.$catchall<z.ZodType<JsonValue, unknown, z.core.$ZodTypeInternals<JsonValue, unknown>>>>>;
+    startedAt: z.ZodOptional<z.ZodString>;
+    finishedAt: z.ZodOptional<z.ZodString>;
+    durationMs: z.ZodOptional<z.ZodNumber>;
+    metadata: z.ZodOptional<z.ZodRecord<z.ZodString, z.ZodType<JsonValue, unknown, z.core.$ZodTypeInternals<JsonValue, unknown>>>>;
+}, z.core.$strict>;
+/** Transcript event captured in a harness session. */
+declare const TranscriptEventSchema: z.ZodDiscriminatedUnion<[z.ZodObject<{
+    type: z.ZodLiteral<"message">;
+    role: z.ZodEnum<{
+        system: "system";
+        user: "user";
+        assistant: "assistant";
+    }>;
+    content: z.ZodOptional<z.ZodType<JsonValue, unknown, z.core.$ZodTypeInternals<JsonValue, unknown>>>;
+    metadata: z.ZodOptional<z.ZodRecord<z.ZodString, z.ZodType<JsonValue, unknown, z.core.$ZodTypeInternals<JsonValue, unknown>>>>;
+}, z.core.$strict>, z.ZodObject<{
+    type: z.ZodLiteral<"tool_call">;
+    id: z.ZodString;
+    name: z.ZodString;
+    arguments: z.ZodOptional<z.ZodRecord<z.ZodString, z.ZodType<JsonValue, unknown, z.core.$ZodTypeInternals<JsonValue, unknown>>>>;
+    startedAt: z.ZodOptional<z.ZodString>;
+    finishedAt: z.ZodOptional<z.ZodString>;
+    durationMs: z.ZodOptional<z.ZodNumber>;
+    metadata: z.ZodOptional<z.ZodRecord<z.ZodString, z.ZodType<JsonValue, unknown, z.core.$ZodTypeInternals<JsonValue, unknown>>>>;
+}, z.core.$strict>, z.ZodObject<{
+    type: z.ZodLiteral<"tool_result">;
+    toolCallId: z.ZodString;
+    name: z.ZodOptional<z.ZodString>;
+    content: z.ZodOptional<z.ZodType<JsonValue, unknown, z.core.$ZodTypeInternals<JsonValue, unknown>>>;
+    error: z.ZodOptional<z.ZodObject<{
+        message: z.ZodString;
+        type: z.ZodOptional<z.ZodString>;
+    }, z.core.$catchall<z.ZodType<JsonValue, unknown, z.core.$ZodTypeInternals<JsonValue, unknown>>>>>;
+    startedAt: z.ZodOptional<z.ZodString>;
+    finishedAt: z.ZodOptional<z.ZodString>;
+    durationMs: z.ZodOptional<z.ZodNumber>;
+    metadata: z.ZodOptional<z.ZodRecord<z.ZodString, z.ZodType<JsonValue, unknown, z.core.$ZodTypeInternals<JsonValue, unknown>>>>;
+}, z.core.$strict>], "type">;
+/** Message event captured in a harness transcript. */
+type TranscriptMessageEvent = {
+    /** Transcript event discriminator. */
+    type: "message";
+    /** Message role. Tool activity is represented by separate tool events. */
+    role: "system" | "user" | "assistant";
+    /** JSON-safe message content. */
+    content?: JsonValue;
+    /** Extra JSON-safe message metadata. */
+    metadata?: Record<string, JsonValue>;
 };
+/** Tool-call event captured in a harness transcript. */
+type TranscriptToolCallEvent = {
+    /** Transcript event discriminator. */
+    type: "tool_call";
+    /** Transcript-local id used to link a result event. */
+    id: string;
+    /** Tool name as exposed to the agent or application runtime. */
+    name: string;
+    /** JSON-safe tool arguments after provider/runtime normalization. */
+    arguments?: Record<string, JsonValue>;
+    /** ISO timestamp for the start of tool execution. */
+    startedAt?: string;
+    /** ISO timestamp for the end of tool execution. */
+    finishedAt?: string;
+    /** Tool execution duration in milliseconds. */
+    durationMs?: number;
+    /** Extra JSON-safe tool metadata for reporters and custom judges. */
+    metadata?: Record<string, JsonValue>;
+};
+/** Tool-result event captured in a harness transcript. */
+type TranscriptToolResultEvent = {
+    /** Transcript event discriminator. */
+    type: "tool_result";
+    /** Tool-call id this result responds to. */
+    toolCallId: string;
+    /** Tool name when available from the provider or runtime. */
+    name?: string;
+    /** JSON-safe tool result content. */
+    content?: JsonValue;
+    /** Normalized tool error when execution failed. */
+    error?: NormalizedError;
+    /** ISO timestamp for the start of tool execution. */
+    startedAt?: string;
+    /** ISO timestamp for the end of tool execution. */
+    finishedAt?: string;
+    /** Tool execution duration in milliseconds. */
+    durationMs?: number;
+    /** Extra JSON-safe tool-result metadata. */
+    metadata?: Record<string, JsonValue>;
+};
+/** Transcript event captured in a harness session. */
+type TranscriptEvent = TranscriptMessageEvent | TranscriptToolCallEvent | TranscriptToolResultEvent;
+/** Text part accepted in AI SDK-style message content arrays. */
+type TranscriptMessageTextPart = {
+    /** Content part discriminator. */
+    type: "text";
+    /** Message text for this ordered content part. */
+    text: string;
+    /** Extra JSON-safe metadata for this content part. */
+    metadata?: Record<string, JsonValue>;
+};
+/** Tool-call part accepted in AI SDK-style assistant message content arrays. */
+type TranscriptMessageToolCallPart = Omit<TranscriptToolCallEvent, "type" | "id" | "name" | "arguments"> & {
+    /** Content part discriminator. */
+    type: "tool-call";
+    /** Tool-call id from the message transport. */
+    toolCallId: string;
+    /** Tool name from the message transport. */
+    toolName: string;
+    /** JSON-safe tool input from the message transport. */
+    input?: JsonValue;
+};
+/** Tool-result part accepted in AI SDK-style tool message content arrays. */
+type TranscriptMessageToolResultPart = Omit<TranscriptToolResultEvent, "type" | "toolCallId" | "name" | "content"> & {
+    /** Content part discriminator. */
+    type: "tool-result";
+    /** Tool-call id this result responds to. */
+    toolCallId: string;
+    /** Tool name from the message transport. */
+    toolName?: string;
+    /** JSON-safe tool output from the message transport. */
+    output?: JsonValue;
+};
+/** Ordered content part accepted in AI SDK-style message content arrays. */
+type TranscriptMessageContentPart = TranscriptMessageTextPart | TranscriptMessageToolCallPart | TranscriptMessageToolResultPart;
+/** System/user message content accepted at harness input boundaries. */
+type TranscriptMessageContentInput = JsonValue;
+type TranscriptMessageToolCallBase = Omit<TranscriptToolCallEvent, "type" | "id" | "name" | "arguments"> & {
+    /** Tool-call id when the message transport exposes one. */
+    id?: string;
+    /** JSON-safe tool arguments in the harness message contract. */
+    arguments?: Record<string, JsonValue>;
+};
+/** Message-transport tool call accepted at harness input boundaries. */
+type TranscriptMessageToolCall = TranscriptMessageToolCallBase & {
+    /** Tool name in the harness message contract. */
+    name: string;
+};
+type TranscriptTopLevelToolMessage = Omit<TranscriptToolResultEvent, "type" | "toolCallId" | "content"> & {
+    role: "tool";
+    toolCallId: string;
+    content?: JsonValue | TranscriptMessageToolResultPart[];
+};
+/** Message-style transport accepted at harness input boundaries. */
+type TranscriptMessageInput = {
+    role: "system" | "user";
+    content?: TranscriptMessageContentInput;
+    metadata?: Record<string, JsonValue>;
+} | {
+    role: "assistant";
+    content?: JsonValue | (TranscriptMessageTextPart | TranscriptMessageToolCallPart)[];
+    toolCalls?: TranscriptMessageToolCall[];
+    metadata?: Record<string, JsonValue>;
+} | TranscriptTopLevelToolMessage | {
+    role: "tool";
+    content: TranscriptMessageToolResultPart[];
+    metadata?: Record<string, JsonValue>;
+};
+/** Converts message-transport inputs into normalized transcript events. */
+declare function messagesToTranscriptEvents(messages: readonly TranscriptMessageInput[]): TranscriptEvent[];
+/** Returns true for normalized message transcript events. */
+declare function isMessageEvent(event: TranscriptEvent): event is TranscriptMessageEvent;
+/** Returns true for normalized tool-call transcript events. */
+declare function isToolCallEvent(event: TranscriptEvent): event is TranscriptToolCallEvent;
+/** Returns true for normalized tool-result transcript events. */
+declare function isToolResultEvent(event: TranscriptEvent): event is TranscriptToolResultEvent;
+
 /** Usage values normalized by vitest-evals harnesses. */
 declare const UsageSummarySchema: z.ZodObject<{
     provider: z.ZodOptional<z.ZodString>;
@@ -139,61 +336,71 @@ type TimingSummary = {
     /** Extra JSON-safe timing metadata. */
     metadata?: Record<string, JsonValue>;
 };
-/** Normalized error object captured in a tool call or trace span. */
-declare const NormalizedErrorSchema: z.ZodObject<{
-    message: z.ZodString;
-    type: z.ZodOptional<z.ZodString>;
-}, z.core.$catchall<z.ZodType<JsonValue, unknown, z.core.$ZodTypeInternals<JsonValue, unknown>>>>;
-/** Normalized tool call captured in a harness session. */
-declare const ToolCallRecordSchema: z.ZodObject<{
-    id: z.ZodOptional<z.ZodString>;
+/** Tool call observed in a normalized session, with result status when known. */
+declare const ToolCallSchema: z.ZodDiscriminatedUnion<[z.ZodObject<{
+    name: z.ZodString;
+    arguments: z.ZodOptional<z.ZodRecord<z.ZodString, z.ZodType<JsonValue, unknown, z.core.$ZodTypeInternals<JsonValue, unknown>>>>;
+    status: z.ZodLiteral<"pending">;
+}, z.core.$strict>, z.ZodObject<{
     name: z.ZodString;
     arguments: z.ZodOptional<z.ZodRecord<z.ZodString, z.ZodType<JsonValue, unknown, z.core.$ZodTypeInternals<JsonValue, unknown>>>>;
+    status: z.ZodLiteral<"ok">;
     result: z.ZodOptional<z.ZodType<JsonValue, unknown, z.core.$ZodTypeInternals<JsonValue, unknown>>>;
-    error: z.ZodOptional<z.ZodObject<{
+}, z.core.$strict>, z.ZodObject<{
+    name: z.ZodString;
+    arguments: z.ZodOptional<z.ZodRecord<z.ZodString, z.ZodType<JsonValue, unknown, z.core.$ZodTypeInternals<JsonValue, unknown>>>>;
+    status: z.ZodLiteral<"error">;
+    error: z.ZodObject<{
         message: z.ZodString;
         type: z.ZodOptional<z.ZodString>;
-    }, z.core.$catchall<z.ZodType<JsonValue, unknown, z.core.$ZodTypeInternals<JsonValue, unknown>>>>>;
-    startedAt: z.ZodOptional<z.ZodString>;
-    finishedAt: z.ZodOptional<z.ZodString>;
-    durationMs: z.ZodOptional<z.ZodNumber>;
-    metadata: z.ZodOptional<z.ZodRecord<z.ZodString, z.ZodType<JsonValue, unknown, z.core.$ZodTypeInternals<JsonValue, unknown>>>>;
-}, z.core.$strict>;
-/** Normalized tool call captured in a harness session. */
-type ToolCallRecord = {
-    /** Provider or runtime tool-call id when one is available. */
-    id?: string;
+    }, z.core.$catchall<z.ZodType<JsonValue, unknown, z.core.$ZodTypeInternals<JsonValue, unknown>>>>;
+}, z.core.$strict>], "status">;
+type ToolCallBase = {
     /** Tool name as exposed to the agent or application runtime. */
     name: string;
     /** JSON-safe tool arguments after provider/runtime normalization. */
     arguments?: Record<string, JsonValue>;
-    /** JSON-safe tool result returned by the application tool. */
-    result?: JsonValue;
-    /** Normalized tool error when execution failed. */
-    error?: NormalizedError;
-    /** ISO timestamp for the start of tool execution. */
-    startedAt?: string;
-    /** ISO timestamp for the end of tool execution. */
-    finishedAt?: string;
-    /** Tool execution duration in milliseconds. */
-    durationMs?: number;
-    /** Extra JSON-safe tool metadata for reporters and custom judges. */
-    metadata?: Record<string, JsonValue>;
 };
-/** Normalized transcript message captured in a harness session. */
-declare const NormalizedMessageSchema: z.ZodObject<{
-    role: z.ZodEnum<{
-        system: "system";
-        user: "user";
-        assistant: "assistant";
-        tool: "tool";
-    }>;
-    content: z.ZodOptional<z.ZodType<JsonValue, unknown, z.core.$ZodTypeInternals<JsonValue, unknown>>>;
-    toolCalls: z.ZodOptional<z.ZodArray<z.ZodObject<{
-        id: z.ZodOptional<z.ZodString>;
+/** Tool call observed in a normalized session, with result status when known. */
+type ToolCall = (ToolCallBase & {
+    /** Lifecycle status for a tool call without a matching result event. */
+    status: "pending";
+}) | (ToolCallBase & {
+    /** Lifecycle status for a tool call with a successful result event. */
+    status: "ok";
+    /** JSON-safe tool result content when execution produced content. */
+    result?: JsonValue;
+}) | (ToolCallBase & {
+    /** Lifecycle status for a tool call with an error result event. */
+    status: "error";
+    /** Normalized tool error captured by the result event. */
+    error: NormalizedError;
+});
+/** Normalized transcript produced by an application harness. */
+declare const NormalizedSessionSchema: z.ZodObject<{
+    events: z.ZodArray<z.ZodDiscriminatedUnion<[z.ZodObject<{
+        type: z.ZodLiteral<"message">;
+        role: z.ZodEnum<{
+            system: "system";
+            user: "user";
+            assistant: "assistant";
+        }>;
+        content: z.ZodOptional<z.ZodType<JsonValue, unknown, z.core.$ZodTypeInternals<JsonValue, unknown>>>;
+        metadata: z.ZodOptional<z.ZodRecord<z.ZodString, z.ZodType<JsonValue, unknown, z.core.$ZodTypeInternals<JsonValue, unknown>>>>;
+    }, z.core.$strict>, z.ZodObject<{
+        type: z.ZodLiteral<"tool_call">;
+        id: z.ZodString;
         name: z.ZodString;
         arguments: z.ZodOptional<z.ZodRecord<z.ZodString, z.ZodType<JsonValue, unknown, z.core.$ZodTypeInternals<JsonValue, unknown>>>>;
-        result: z.ZodOptional<z.ZodType<JsonValue, unknown, z.core.$ZodTypeInternals<JsonValue, unknown>>>;
+        startedAt: z.ZodOptional<z.ZodString>;
+        finishedAt: z.ZodOptional<z.ZodString>;
+        durationMs: z.ZodOptional<z.ZodNumber>;
+        metadata: z.ZodOptional<z.ZodRecord<z.ZodString, z.ZodType<JsonValue, unknown, z.core.$ZodTypeInternals<JsonValue, unknown>>>>;
+    }, z.core.$strict>, z.ZodObject<{
+        type: z.ZodLiteral<"tool_result">;
+        toolCallId: z.ZodString;
+        name: z.ZodOptional<z.ZodString>;
+        content: z.ZodOptional<z.ZodType<JsonValue, unknown, z.core.$ZodTypeInternals<JsonValue, unknown>>>;
         error: z.ZodOptional<z.ZodObject<{
             message: z.ZodString;
             type: z.ZodOptional<z.ZodString>;
@@ -202,54 +409,15 @@ declare const NormalizedMessageSchema: z.ZodObject<{
         finishedAt: z.ZodOptional<z.ZodString>;
         durationMs: z.ZodOptional<z.ZodNumber>;
         metadata: z.ZodOptional<z.ZodRecord<z.ZodString, z.ZodType<JsonValue, unknown, z.core.$ZodTypeInternals<JsonValue, unknown>>>>;
-    }, z.core.$strict>>>;
-    metadata: z.ZodOptional<z.ZodRecord<z.ZodString, z.ZodType<JsonValue, unknown, z.core.$ZodTypeInternals<JsonValue, unknown>>>>;
-}, z.core.$strict>;
-/** Normalized transcript message captured in a harness session. */
-type NormalizedMessage = {
-    /** Transcript role for the normalized message. */
-    role: "system" | "user" | "assistant" | "tool";
-    /** JSON-safe message content. */
-    content?: JsonValue;
-    /** Tool calls associated with this message. */
-    toolCalls?: ToolCallRecord[];
-    /** Extra JSON-safe message metadata. */
-    metadata?: Record<string, JsonValue>;
-};
-/** Normalized transcript produced by an application harness. */
-declare const NormalizedSessionSchema: z.ZodObject<{
-    messages: z.ZodDefault<z.ZodArray<z.ZodObject<{
-        role: z.ZodEnum<{
-            system: "system";
-            user: "user";
-            assistant: "assistant";
-            tool: "tool";
-        }>;
-        content: z.ZodOptional<z.ZodType<JsonValue, unknown, z.core.$ZodTypeInternals<JsonValue, unknown>>>;
-        toolCalls: z.ZodOptional<z.ZodArray<z.ZodObject<{
-            id: z.ZodOptional<z.ZodString>;
-            name: z.ZodString;
-            arguments: z.ZodOptional<z.ZodRecord<z.ZodString, z.ZodType<JsonValue, unknown, z.core.$ZodTypeInternals<JsonValue, unknown>>>>;
-            result: z.ZodOptional<z.ZodType<JsonValue, unknown, z.core.$ZodTypeInternals<JsonValue, unknown>>>;
-            error: z.ZodOptional<z.ZodObject<{
-                message: z.ZodString;
-                type: z.ZodOptional<z.ZodString>;
-            }, z.core.$catchall<z.ZodType<JsonValue, unknown, z.core.$ZodTypeInternals<JsonValue, unknown>>>>>;
-            startedAt: z.ZodOptional<z.ZodString>;
-            finishedAt: z.ZodOptional<z.ZodString>;
-            durationMs: z.ZodOptional<z.ZodNumber>;
-            metadata: z.ZodOptional<z.ZodRecord<z.ZodString, z.ZodType<JsonValue, unknown, z.core.$ZodTypeInternals<JsonValue, unknown>>>>;
-        }, z.core.$strict>>>;
-        metadata: z.ZodOptional<z.ZodRecord<z.ZodString, z.ZodType<JsonValue, unknown, z.core.$ZodTypeInternals<JsonValue, unknown>>>>;
-    }, z.core.$strict>>>;
+    }, z.core.$strict>], "type">>;
     provider: z.ZodOptional<z.ZodString>;
     model: z.ZodOptional<z.ZodString>;
     metadata: z.ZodOptional<z.ZodRecord<z.ZodString, z.ZodType<JsonValue, unknown, z.core.$ZodTypeInternals<JsonValue, unknown>>>>;
 }, z.core.$strict>;
 /** Normalized transcript produced by an application harness. */
 type NormalizedSession = {
-    /** Ordered normalized transcript messages. */
-    messages: NormalizedMessage[];
+    /** Ordered normalized transcript events. */
+    events: TranscriptEvent[];
     /** Provider that produced the session when known. */
     provider?: string;
     /** Model that produced the session when known. */
@@ -282,8 +450,8 @@ declare const NormalizedSpanSchema: z.ZodObject<{
     name: z.ZodString;
     kind: z.ZodOptional<z.ZodEnum<{
         custom: "custom";
-        model: "model";
         tool: "tool";
+        model: "model";
         run: "run";
         agent: "agent";
         guardrail: "guardrail";
@@ -342,15 +510,15 @@ declare const NormalizedTraceSchema: z.ZodObject<{
     finishedAt: z.ZodOptional<z.ZodString>;
     durationMs: z.ZodOptional<z.ZodNumber>;
     metadata: z.ZodOptional<z.ZodRecord<z.ZodString, z.ZodType<JsonValue, unknown, z.core.$ZodTypeInternals<JsonValue, unknown>>>>;
-    spans: z.ZodDefault<z.ZodArray<z.ZodObject<{
+    spans: z.ZodArray<z.ZodObject<{
         id: z.ZodOptional<z.ZodString>;
         traceId: z.ZodOptional<z.ZodString>;
         parentId: z.ZodOptional<z.ZodString>;
         name: z.ZodString;
         kind: z.ZodOptional<z.ZodEnum<{
             custom: "custom";
-            model: "model";
             tool: "tool";
+            model: "model";
             run: "run";
             agent: "agent";
             guardrail: "guardrail";
@@ -373,7 +541,7 @@ declare const NormalizedTraceSchema: z.ZodObject<{
             timestamp: z.ZodOptional<z.ZodString>;
             attributes: z.ZodOptional<z.ZodRecord<z.ZodString, z.ZodType<JsonValue, unknown, z.core.$ZodTypeInternals<JsonValue, unknown>>>>;
         }, z.core.$strict>>>;
-    }, z.core.$strict>>>;
+    }, z.core.$strict>>;
 }, z.core.$strict>;
 /** Normalized trace captured during a harness run. */
 type NormalizedTrace = {
@@ -396,30 +564,38 @@ type NormalizedTrace = {
 declare const HarnessRunSchema: z.ZodObject<{
     output: z.ZodOptional<z.ZodType<JsonValue, unknown, z.core.$ZodTypeInternals<JsonValue, unknown>>>;
     session: z.ZodObject<{
-        messages: z.ZodDefault<z.ZodArray<z.ZodObject<{
+        events: z.ZodArray<z.ZodDiscriminatedUnion<[z.ZodObject<{
+            type: z.ZodLiteral<"message">;
             role: z.ZodEnum<{
                 system: "system";
                 user: "user";
                 assistant: "assistant";
-                tool: "tool";
             }>;
             content: z.ZodOptional<z.ZodType<JsonValue, unknown, z.core.$ZodTypeInternals<JsonValue, unknown>>>;
-            toolCalls: z.ZodOptional<z.ZodArray<z.ZodObject<{
-                id: z.ZodOptional<z.ZodString>;
-                name: z.ZodString;
-                arguments: z.ZodOptional<z.ZodRecord<z.ZodString, z.ZodType<JsonValue, unknown, z.core.$ZodTypeInternals<JsonValue, unknown>>>>;
-                result: z.ZodOptional<z.ZodType<JsonValue, unknown, z.core.$ZodTypeInternals<JsonValue, unknown>>>;
-                error: z.ZodOptional<z.ZodObject<{
-                    message: z.ZodString;
-                    type: z.ZodOptional<z.ZodString>;
-                }, z.core.$catchall<z.ZodType<JsonValue, unknown, z.core.$ZodTypeInternals<JsonValue, unknown>>>>>;
-                startedAt: z.ZodOptional<z.ZodString>;
-                finishedAt: z.ZodOptional<z.ZodString>;
-                durationMs: z.ZodOptional<z.ZodNumber>;
-                metadata: z.ZodOptional<z.ZodRecord<z.ZodString, z.ZodType<JsonValue, unknown, z.core.$ZodTypeInternals<JsonValue, unknown>>>>;
-            }, z.core.$strict>>>;
             metadata: z.ZodOptional<z.ZodRecord<z.ZodString, z.ZodType<JsonValue, unknown, z.core.$ZodTypeInternals<JsonValue, unknown>>>>;
-        }, z.core.$strict>>>;
+        }, z.core.$strict>, z.ZodObject<{
+            type: z.ZodLiteral<"tool_call">;
+            id: z.ZodString;
+            name: z.ZodString;
+            arguments: z.ZodOptional<z.ZodRecord<z.ZodString, z.ZodType<JsonValue, unknown, z.core.$ZodTypeInternals<JsonValue, unknown>>>>;
+            startedAt: z.ZodOptional<z.ZodString>;
+            finishedAt: z.ZodOptional<z.ZodString>;
+            durationMs: z.ZodOptional<z.ZodNumber>;
+            metadata: z.ZodOptional<z.ZodRecord<z.ZodString, z.ZodType<JsonValue, unknown, z.core.$ZodTypeInternals<JsonValue, unknown>>>>;
+        }, z.core.$strict>, z.ZodObject<{
+            type: z.ZodLiteral<"tool_result">;
+            toolCallId: z.ZodString;
+            name: z.ZodOptional<z.ZodString>;
+            content: z.ZodOptional<z.ZodType<JsonValue, unknown, z.core.$ZodTypeInternals<JsonValue, unknown>>>;
+            error: z.ZodOptional<z.ZodObject<{
+                message: z.ZodString;
+                type: z.ZodOptional<z.ZodString>;
+            }, z.core.$catchall<z.ZodType<JsonValue, unknown, z.core.$ZodTypeInternals<JsonValue, unknown>>>>>;
+            startedAt: z.ZodOptional<z.ZodString>;
+            finishedAt: z.ZodOptional<z.ZodString>;
+            durationMs: z.ZodOptional<z.ZodNumber>;
+            metadata: z.ZodOptional<z.ZodRecord<z.ZodString, z.ZodType<JsonValue, unknown, z.core.$ZodTypeInternals<JsonValue, unknown>>>>;
+        }, z.core.$strict>], "type">>;
         provider: z.ZodOptional<z.ZodString>;
         model: z.ZodOptional<z.ZodString>;
         metadata: z.ZodOptional<z.ZodRecord<z.ZodString, z.ZodType<JsonValue, unknown, z.core.$ZodTypeInternals<JsonValue, unknown>>>>;
@@ -447,15 +623,15 @@ declare const HarnessRunSchema: z.ZodObject<{
         finishedAt: z.ZodOptional<z.ZodString>;
         durationMs: z.ZodOptional<z.ZodNumber>;
         metadata: z.ZodOptional<z.ZodRecord<z.ZodString, z.ZodType<JsonValue, unknown, z.core.$ZodTypeInternals<JsonValue, unknown>>>>;
-        spans: z.ZodDefault<z.ZodArray<z.ZodObject<{
+        spans: z.ZodArray<z.ZodObject<{
             id: z.ZodOptional<z.ZodString>;
             traceId: z.ZodOptional<z.ZodString>;
             parentId: z.ZodOptional<z.ZodString>;
             name: z.ZodString;
             kind: z.ZodOptional<z.ZodEnum<{
                 custom: "custom";
-                model: "model";
                 tool: "tool";
+                model: "model";
                 run: "run";
                 agent: "agent";
                 guardrail: "guardrail";
@@ -478,9 +654,9 @@ declare const HarnessRunSchema: z.ZodObject<{
                 timestamp: z.ZodOptional<z.ZodString>;
                 attributes: z.ZodOptional<z.ZodRecord<z.ZodString, z.ZodType<JsonValue, unknown, z.core.$ZodTypeInternals<JsonValue, unknown>>>>;
             }, z.core.$strict>>>;
-        }, z.core.$strict>>>;
+        }, z.core.$strict>>;
     }, z.core.$strict>>>;
-    errors: z.ZodDefault<z.ZodArray<z.ZodRecord<z.ZodString, z.ZodType<JsonValue, unknown, z.core.$ZodTypeInternals<JsonValue, unknown>>>>>;
+    errors: z.ZodArray<z.ZodRecord<z.ZodString, z.ZodType<JsonValue, unknown, z.core.$ZodTypeInternals<JsonValue, unknown>>>>;
 }, z.core.$strict>;
 type OutputField<TOutput extends JsonValue | undefined> = undefined extends TOutput ? {
     output?: TOutput;
@@ -495,9 +671,9 @@ type OutputField<TOutput extends JsonValue | undefined> = undefined extends TOut
  * const run: HarnessRun<{ status: "approved" }> = {
  *   output: { status: "approved" },
  *   session: {
- *     messages: [
- *       { role: "user", content: "Refund invoice inv_123" },
- *       { role: "assistant", content: { status: "approved" } },
+ *     events: [
+ *       { type: "message", role: "user", content: "Refund invoice inv_123" },
+ *       { type: "message", role: "assistant", content: { status: "approved" } },
  *     ],
  *   },
  *   usage: { totalTokens: 260 },
@@ -526,22 +702,26 @@ type HarnessRunError = Error & {
 };
 
 /**
- * Flattens every recorded tool call from a normalized session.
+ * Returns every tool call observed in a normalized run or session.
+ *
+ * Tool results are joined from matching `tool_result` transcript events when
+ * the transcript contains a result for the call.
  *
- * @param session - Normalized session produced by a harness run.
+ * @param source - Normalized run or session produced by a harness.
  *
  * @example
  * ```ts
- * const names = toolCalls(result.session).map((call) => call.name);
+ * const names = toolCalls(result).map((call) => call.name);
  *
  * expect(names).toEqual(["lookupInvoice", "createRefund"]);
  * ```
  */
-declare function toolCalls(session: NormalizedSession): ToolCallRecord[];
+declare function toolCalls(run: HarnessRun): ToolCall[];
+declare function toolCalls(session: NormalizedSession): ToolCall[];
 /**
- * Flattens every recorded span from a normalized harness run.
+ * Flattens every recorded span from a normalized harness run or trace list.
  *
- * @param run - Normalized harness run produced by a harness.
+ * @param source - Normalized run or trace list produced by a harness.
  *
  * @example
  * ```ts
@@ -549,94 +729,104 @@ declare function toolCalls(session: NormalizedSession): ToolCallRecord[];
  * ```
  */
 declare function spans(run: HarnessRun): NormalizedSpan[];
+declare function spans(traces: readonly NormalizedTrace[] | undefined): NormalizedSpan[];
 /**
  * Alias for `spans(...)` for consumers that prefer trace-oriented naming.
  *
- * @param run - Normalized harness run produced by a harness.
+ * @param source - Normalized run or trace list produced by a harness.
  */
 declare function traceSpans(run: HarnessRun): NormalizedSpan[];
+declare function traceSpans(traces: readonly NormalizedTrace[] | undefined): NormalizedSpan[];
 /**
- * Returns spans of one coarse operation kind from a normalized run.
+ * Returns spans of one coarse operation kind from a normalized run or trace list.
  *
- * @param run - Normalized harness run produced by a harness.
+ * @param source - Normalized run or trace list produced by a harness.
  * @param kind - Span kind to keep.
  */
 declare function spansByKind(run: HarnessRun, kind: NonNullable<NormalizedSpan["kind"]>): NormalizedSpan[];
+declare function spansByKind(traces: readonly NormalizedTrace[] | undefined, kind: NonNullable<NormalizedSpan["kind"]>): NormalizedSpan[];
 /**
  * Returns every span that explicitly failed or carries a normalized error.
  *
- * @param run - Normalized harness run produced by a harness.
+ * @param source - Normalized run or trace list produced by a harness.
  */
 declare function failedSpans(run: HarnessRun): NormalizedSpan[];
+declare function failedSpans(traces: readonly NormalizedTrace[] | undefined): NormalizedSpan[];
 /**
- * Filters normalized session messages by role.
+ * Filters normalized session message events by role.
  *
- * @param session - Normalized session produced by a harness run.
+ * @param source - Normalized run or session produced by a harness.
  * @param role - Message role to keep.
  *
  * @example
  * ```ts
- * const assistantText = messagesByRole(result.session, "assistant")
+ * const assistantText = messagesByRole(result, "assistant")
  *   .map((message) => message.content)
  *   .join("\n");
  * ```
  */
-declare function messagesByRole(session: NormalizedSession, role: NormalizedMessage["role"]): NormalizedMessage[];
+declare function messagesByRole(run: HarnessRun, role: TranscriptMessageEvent["role"]): TranscriptMessageEvent[];
+declare function messagesByRole(session: NormalizedSession, role: TranscriptMessageEvent["role"]): TranscriptMessageEvent[];
 /**
- * Returns every normalized system message from a session.
+ * Returns every normalized system message event from a session.
  *
- * @param session - Normalized session produced by a harness run.
+ * @param source - Normalized run or session produced by a harness.
  *
  * @example
  * ```ts
- * const systemPrompts = systemMessages(result.session);
+ * const systemPrompts = systemMessages(result);
  * ```
  */
-declare function systemMessages(session: NormalizedSession): NormalizedMessage[];
+declare function systemMessages(run: HarnessRun): TranscriptMessageEvent[];
+declare function systemMessages(session: NormalizedSession): TranscriptMessageEvent[];
 /**
- * Returns every normalized user message from a session.
+ * Returns every normalized user message event from a session.
  *
- * @param session - Normalized session produced by a harness run.
+ * @param source - Normalized run or session produced by a harness.
  *
  * @example
  * ```ts
- * const firstPrompt = userMessages(result.session)[0]?.content;
+ * const firstPrompt = userMessages(result)[0]?.content;
  * ```
  */
-declare function userMessages(session: NormalizedSession): NormalizedMessage[];
+declare function userMessages(run: HarnessRun): TranscriptMessageEvent[];
+declare function userMessages(session: NormalizedSession): TranscriptMessageEvent[];
 /**
- * Returns every normalized assistant message from a session.
+ * Returns every normalized assistant message event from a session.
  *
- * @param session - Normalized session produced by a harness run.
+ * @param source - Normalized run or session produced by a harness.
  *
  * @example
  * ```ts
- * const finalAnswer = assistantMessages(session).at(-1)?.content;
+ * const finalAnswer = assistantMessages(result).at(-1)?.content;
  * ```
  */
-declare function assistantMessages(session: NormalizedSession): NormalizedMessage[];
+declare function assistantMessages(run: HarnessRun): TranscriptMessageEvent[];
+declare function assistantMessages(session: NormalizedSession): TranscriptMessageEvent[];
 /**
  * Returns the latest assistant message content, ignoring empty text messages.
  *
- * @param session - Normalized session produced by a harness run.
+ * @param source - Normalized run or session produced by a harness.
  *
  * @example
  * ```ts
- * const finalAnswer = latestAssistantMessageContent(result.session);
+ * const finalAnswer = latestAssistantMessageContent(result);
  * ```
  */
+declare function latestAssistantMessageContent(run: HarnessRun): JsonValue | undefined;
 declare function latestAssistantMessageContent(session: NormalizedSession): JsonValue | undefined;
 /**
- * Returns every normalized tool message from a session.
+ * Returns every normalized tool-result event from a session.
  *
- * @param session - Normalized session produced by a harness run.
+ * @param source - Normalized run or session produced by a harness.
  *
  * @example
  * ```ts
- * const toolOutputs = toolMessages(session).map((message) => message.content);
+ * const toolOutputs = toolMessages(result).map((message) => message.content);
  * ```
  */
-declare function toolMessages(session: NormalizedSession): NormalizedMessage[];
+declare function toolMessages(run: HarnessRun): TranscriptToolResultEvent[];
+declare function toolMessages(session: NormalizedSession): TranscriptToolResultEvent[];
 
 /** Harness metadata stored by vitest-evals on Vitest task metadata. */
 declare const HarnessMetaSchema: z.ZodObject<{
@@ -644,30 +834,38 @@ declare const HarnessMetaSchema: z.ZodObject<{
     run: z.ZodOptional<z.ZodObject<{
         output: z.ZodOptional<z.ZodType<JsonValue, unknown, z.core.$ZodTypeInternals<JsonValue, unknown>>>;
         session: z.ZodObject<{
-            messages: z.ZodDefault<z.ZodArray<z.ZodObject<{
+            events: z.ZodArray<z.ZodDiscriminatedUnion<[z.ZodObject<{
+                type: z.ZodLiteral<"message">;
                 role: z.ZodEnum<{
                     system: "system";
                     user: "user";
                     assistant: "assistant";
-                    tool: "tool";
                 }>;
                 content: z.ZodOptional<z.ZodType<JsonValue, unknown, z.core.$ZodTypeInternals<JsonValue, unknown>>>;
-                toolCalls: z.ZodOptional<z.ZodArray<z.ZodObject<{
-                    id: z.ZodOptional<z.ZodString>;
-                    name: z.ZodString;
-                    arguments: z.ZodOptional<z.ZodRecord<z.ZodString, z.ZodType<JsonValue, unknown, z.core.$ZodTypeInternals<JsonValue, unknown>>>>;
-                    result: z.ZodOptional<z.ZodType<JsonValue, unknown, z.core.$ZodTypeInternals<JsonValue, unknown>>>;
-                    error: z.ZodOptional<z.ZodObject<{
-                        message: z.ZodString;
-                        type: z.ZodOptional<z.ZodString>;
-                    }, z.core.$catchall<z.ZodType<JsonValue, unknown, z.core.$ZodTypeInternals<JsonValue, unknown>>>>>;
-                    startedAt: z.ZodOptional<z.ZodString>;
-                    finishedAt: z.ZodOptional<z.ZodString>;
-                    durationMs: z.ZodOptional<z.ZodNumber>;
-                    metadata: z.ZodOptional<z.ZodRecord<z.ZodString, z.ZodType<JsonValue, unknown, z.core.$ZodTypeInternals<JsonValue, unknown>>>>;
-                }, z.core.$strict>>>;
                 metadata: z.ZodOptional<z.ZodRecord<z.ZodString, z.ZodType<JsonValue, unknown, z.core.$ZodTypeInternals<JsonValue, unknown>>>>;
-            }, z.core.$strict>>>;
+            }, z.core.$strict>, z.ZodObject<{
+                type: z.ZodLiteral<"tool_call">;
+                id: z.ZodString;
+                name: z.ZodString;
+                arguments: z.ZodOptional<z.ZodRecord<z.ZodString, z.ZodType<JsonValue, unknown, z.core.$ZodTypeInternals<JsonValue, unknown>>>>;
+                startedAt: z.ZodOptional<z.ZodString>;
+                finishedAt: z.ZodOptional<z.ZodString>;
+                durationMs: z.ZodOptional<z.ZodNumber>;
+                metadata: z.ZodOptional<z.ZodRecord<z.ZodString, z.ZodType<JsonValue, unknown, z.core.$ZodTypeInternals<JsonValue, unknown>>>>;
+            }, z.core.$strict>, z.ZodObject<{
+                type: z.ZodLiteral<"tool_result">;
+                toolCallId: z.ZodString;
+                name: z.ZodOptional<z.ZodString>;
+                content: z.ZodOptional<z.ZodType<JsonValue, unknown, z.core.$ZodTypeInternals<JsonValue, unknown>>>;
+                error: z.ZodOptional<z.ZodObject<{
+                    message: z.ZodString;
+                    type: z.ZodOptional<z.ZodString>;
+                }, z.core.$catchall<z.ZodType<JsonValue, unknown, z.core.$ZodTypeInternals<JsonValue, unknown>>>>>;
+                startedAt: z.ZodOptional<z.ZodString>;
+                finishedAt: z.ZodOptional<z.ZodString>;
+                durationMs: z.ZodOptional<z.ZodNumber>;
+                metadata: z.ZodOptional<z.ZodRecord<z.ZodString, z.ZodType<JsonValue, unknown, z.core.$ZodTypeInternals<JsonValue, unknown>>>>;
+            }, z.core.$strict>], "type">>;
             provider: z.ZodOptional<z.ZodString>;
             model: z.ZodOptional<z.ZodString>;
             metadata: z.ZodOptional<z.ZodRecord<z.ZodString, z.ZodType<JsonValue, unknown, z.core.$ZodTypeInternals<JsonValue, unknown>>>>;
@@ -695,15 +893,15 @@ declare const HarnessMetaSchema: z.ZodObject<{
             finishedAt: z.ZodOptional<z.ZodString>;
             durationMs: z.ZodOptional<z.ZodNumber>;
             metadata: z.ZodOptional<z.ZodRecord<z.ZodString, z.ZodType<JsonValue, unknown, z.core.$ZodTypeInternals<JsonValue, unknown>>>>;
-            spans: z.ZodDefault<z.ZodArray<z.ZodObject<{
+            spans: z.ZodArray<z.ZodObject<{
                 id: z.ZodOptional<z.ZodString>;
                 traceId: z.ZodOptional<z.ZodString>;
                 parentId: z.ZodOptional<z.ZodString>;
                 name: z.ZodString;
                 kind: z.ZodOptional<z.ZodEnum<{
                     custom: "custom";
-                    model: "model";
                     tool: "tool";
+                    model: "model";
                     run: "run";
                     agent: "agent";
                     guardrail: "guardrail";
@@ -726,9 +924,9 @@ declare const HarnessMetaSchema: z.ZodObject<{
                     timestamp: z.ZodOptional<z.ZodString>;
                     attributes: z.ZodOptional<z.ZodRecord<z.ZodString, z.ZodType<JsonValue, unknown, z.core.$ZodTypeInternals<JsonValue, unknown>>>>;
                 }, z.core.$strict>>>;
-            }, z.core.$strict>>>;
+            }, z.core.$strict>>;
         }, z.core.$strict>>>;
-        errors: z.ZodDefault<z.ZodArray<z.ZodRecord<z.ZodString, z.ZodType<JsonValue, unknown, z.core.$ZodTypeInternals<JsonValue, unknown>>>>>;
+        errors: z.ZodArray<z.ZodRecord<z.ZodString, z.ZodType<JsonValue, unknown, z.core.$ZodTypeInternals<JsonValue, unknown>>>>;
     }, z.core.$strict>>;
 }, z.core.$strict>;
 /** Harness metadata stored by vitest-evals on Vitest task metadata. */
@@ -751,20 +949,24 @@ declare const EvalMetaSchema: z.ZodObject<{
     avgScore: z.ZodPipe<z.ZodTransform<unknown, unknown>, z.ZodOptional<z.ZodNullable<z.ZodNumber>>>;
     output: z.ZodOptional<z.ZodType<JsonValue, unknown, z.core.$ZodTypeInternals<JsonValue, unknown>>>;
     thresholdFailed: z.ZodOptional<z.ZodBoolean>;
-    toolCalls: z.ZodOptional<z.ZodArray<z.ZodObject<{
-        id: z.ZodOptional<z.ZodString>;
+    toolCalls: z.ZodOptional<z.ZodArray<z.ZodDiscriminatedUnion<[z.ZodObject<{
         name: z.ZodString;
         arguments: z.ZodOptional<z.ZodRecord<z.ZodString, z.ZodType<JsonValue, unknown, z.core.$ZodTypeInternals<JsonValue, unknown>>>>;
+        status: z.ZodLiteral<"pending">;
+    }, z.core.$strict>, z.ZodObject<{
+        name: z.ZodString;
+        arguments: z.ZodOptional<z.ZodRecord<z.ZodString, z.ZodType<JsonValue, unknown, z.core.$ZodTypeInternals<JsonValue, unknown>>>>;
+        status: z.ZodLiteral<"ok">;
         result: z.ZodOptional<z.ZodType<JsonValue, unknown, z.core.$ZodTypeInternals<JsonValue, unknown>>>;
-        error: z.ZodOptional<z.ZodObject<{
+    }, z.core.$strict>, z.ZodObject<{
+        name: z.ZodString;
+        arguments: z.ZodOptional<z.ZodRecord<z.ZodString, z.ZodType<JsonValue, unknown, z.core.$ZodTypeInternals<JsonValue, unknown>>>>;
+        status: z.ZodLiteral<"error">;
+        error: z.ZodObject<{
             message: z.ZodString;
             type: z.ZodOptional<z.ZodString>;
-        }, z.core.$catchall<z.ZodType<JsonValue, unknown, z.core.$ZodTypeInternals<JsonValue, unknown>>>>>;
-        startedAt: z.ZodOptional<z.ZodString>;
-        finishedAt: z.ZodOptional<z.ZodString>;
-        durationMs: z.ZodOptional<z.ZodNumber>;
-        metadata: z.ZodOptional<z.ZodRecord<z.ZodString, z.ZodType<JsonValue, unknown, z.core.$ZodTypeInternals<JsonValue, unknown>>>>;
-    }, z.core.$strict>>>;
+        }, z.core.$catchall<z.ZodType<JsonValue, unknown, z.core.$ZodTypeInternals<JsonValue, unknown>>>>;
+    }, z.core.$strict>], "status">>>;
 }, z.core.$strict>;
 /** Eval metadata stored by vitest-evals on Vitest task metadata. */
 type EvalMeta = z.infer<typeof EvalMetaSchema>;
@@ -779,50 +981,62 @@ declare const EvalTaskMetaSchema: z.ZodObject<{
         avgScore: z.ZodPipe<z.ZodTransform<unknown, unknown>, z.ZodOptional<z.ZodNullable<z.ZodNumber>>>;
         output: z.ZodOptional<z.ZodType<JsonValue, unknown, z.core.$ZodTypeInternals<JsonValue, unknown>>>;
         thresholdFailed: z.ZodOptional<z.ZodBoolean>;
-        toolCalls: z.ZodOptional<z.ZodArray<z.ZodObject<{
-            id: z.ZodOptional<z.ZodString>;
+        toolCalls: z.ZodOptional<z.ZodArray<z.ZodDiscriminatedUnion<[z.ZodObject<{
             name: z.ZodString;
             arguments: z.ZodOptional<z.ZodRecord<z.ZodString, z.ZodType<JsonValue, unknown, z.core.$ZodTypeInternals<JsonValue, unknown>>>>;
+            status: z.ZodLiteral<"pending">;
+        }, z.core.$strict>, z.ZodObject<{
+            name: z.ZodString;
+            arguments: z.ZodOptional<z.ZodRecord<z.ZodString, z.ZodType<JsonValue, unknown, z.core.$ZodTypeInternals<JsonValue, unknown>>>>;
+            status: z.ZodLiteral<"ok">;
             result: z.ZodOptional<z.ZodType<JsonValue, unknown, z.core.$ZodTypeInternals<JsonValue, unknown>>>;
-            error: z.ZodOptional<z.ZodObject<{
+        }, z.core.$strict>, z.ZodObject<{
+            name: z.ZodString;
+            arguments: z.ZodOptional<z.ZodRecord<z.ZodString, z.ZodType<JsonValue, unknown, z.core.$ZodTypeInternals<JsonValue, unknown>>>>;
+            status: z.ZodLiteral<"error">;
+            error: z.ZodObject<{
                 message: z.ZodString;
                 type: z.ZodOptional<z.ZodString>;
-            }, z.core.$catchall<z.ZodType<JsonValue, unknown, z.core.$ZodTypeInternals<JsonValue, unknown>>>>>;
-            startedAt: z.ZodOptional<z.ZodString>;
-            finishedAt: z.ZodOptional<z.ZodString>;
-            durationMs: z.ZodOptional<z.ZodNumber>;
-            metadata: z.ZodOptional<z.ZodRecord<z.ZodString, z.ZodType<JsonValue, unknown, z.core.$ZodTypeInternals<JsonValue, unknown>>>>;
-        }, z.core.$strict>>>;
+            }, z.core.$catchall<z.ZodType<JsonValue, unknown, z.core.$ZodTypeInternals<JsonValue, unknown>>>>;
+        }, z.core.$strict>], "status">>>;
     }, z.core.$strict>>;
     harness: z.ZodOptional<z.ZodObject<{
         name: z.ZodOptional<z.ZodString>;
         run: z.ZodOptional<z.ZodObject<{
             output: z.ZodOptional<z.ZodType<JsonValue, unknown, z.core.$ZodTypeInternals<JsonValue, unknown>>>;
             session: z.ZodObject<{
-                messages: z.ZodDefault<z.ZodArray<z.ZodObject<{
+                events: z.ZodArray<z.ZodDiscriminatedUnion<[z.ZodObject<{
+                    type: z.ZodLiteral<"message">;
                     role: z.ZodEnum<{
                         system: "system";
                         user: "user";
                         assistant: "assistant";
-                        tool: "tool";
                     }>;
                     content: z.ZodOptional<z.ZodType<JsonValue, unknown, z.core.$ZodTypeInternals<JsonValue, unknown>>>;
-                    toolCalls: z.ZodOptional<z.ZodArray<z.ZodObject<{
-                        id: z.ZodOptional<z.ZodString>;
-                        name: z.ZodString;
-                        arguments: z.ZodOptional<z.ZodRecord<z.ZodString, z.ZodType<JsonValue, unknown, z.core.$ZodTypeInternals<JsonValue, unknown>>>>;
-                        result: z.ZodOptional<z.ZodType<JsonValue, unknown, z.core.$ZodTypeInternals<JsonValue, unknown>>>;
-                        error: z.ZodOptional<z.ZodObject<{
-                            message: z.ZodString;
-                            type: z.ZodOptional<z.ZodString>;
-                        }, z.core.$catchall<z.ZodType<JsonValue, unknown, z.core.$ZodTypeInternals<JsonValue, unknown>>>>>;
-                        startedAt: z.ZodOptional<z.ZodString>;
-                        finishedAt: z.ZodOptional<z.ZodString>;
-                        durationMs: z.ZodOptional<z.ZodNumber>;
-                        metadata: z.ZodOptional<z.ZodRecord<z.ZodString, z.ZodType<JsonValue, unknown, z.core.$ZodTypeInternals<JsonValue, unknown>>>>;
-                    }, z.core.$strict>>>;
                     metadata: z.ZodOptional<z.ZodRecord<z.ZodString, z.ZodType<JsonValue, unknown, z.core.$ZodTypeInternals<JsonValue, unknown>>>>;
-                }, z.core.$strict>>>;
+                }, z.core.$strict>, z.ZodObject<{
+                    type: z.ZodLiteral<"tool_call">;
+                    id: z.ZodString;
+                    name: z.ZodString;
+                    arguments: z.ZodOptional<z.ZodRecord<z.ZodString, z.ZodType<JsonValue, unknown, z.core.$ZodTypeInternals<JsonValue, unknown>>>>;
+                    startedAt: z.ZodOptional<z.ZodString>;
+                    finishedAt: z.ZodOptional<z.ZodString>;
+                    durationMs: z.ZodOptional<z.ZodNumber>;
+                    metadata: z.ZodOptional<z.ZodRecord<z.ZodString, z.ZodType<JsonValue, unknown, z.core.$ZodTypeInternals<JsonValue, unknown>>>>;
+                }, z.core.$strict>, z.ZodObject<{
+                    type: z.ZodLiteral<"tool_result">;
+                    toolCallId: z.ZodString;
+                    name: z.ZodOptional<z.ZodString>;
+                    content: z.ZodOptional<z.ZodType<JsonValue, unknown, z.core.$ZodTypeInternals<JsonValue, unknown>>>;
+                    error: z.ZodOptional<z.ZodObject<{
+                        message: z.ZodString;
+                        type: z.ZodOptional<z.ZodString>;
+                    }, z.core.$catchall<z.ZodType<JsonValue, unknown, z.core.$ZodTypeInternals<JsonValue, unknown>>>>>;
+                    startedAt: z.ZodOptional<z.ZodString>;
+                    finishedAt: z.ZodOptional<z.ZodString>;
+                    durationMs: z.ZodOptional<z.ZodNumber>;
+                    metadata: z.ZodOptional<z.ZodRecord<z.ZodString, z.ZodType<JsonValue, unknown, z.core.$ZodTypeInternals<JsonValue, unknown>>>>;
+                }, z.core.$strict>], "type">>;
                 provider: z.ZodOptional<z.ZodString>;
                 model: z.ZodOptional<z.ZodString>;
                 metadata: z.ZodOptional<z.ZodRecord<z.ZodString, z.ZodType<JsonValue, unknown, z.core.$ZodTypeInternals<JsonValue, unknown>>>>;
@@ -850,15 +1064,15 @@ declare const EvalTaskMetaSchema: z.ZodObject<{
                 finishedAt: z.ZodOptional<z.ZodString>;
                 durationMs: z.ZodOptional<z.ZodNumber>;
                 metadata: z.ZodOptional<z.ZodRecord<z.ZodString, z.ZodType<JsonValue, unknown, z.core.$ZodTypeInternals<JsonValue, unknown>>>>;
-                spans: z.ZodDefault<z.ZodArray<z.ZodObject<{
+                spans: z.ZodArray<z.ZodObject<{
                     id: z.ZodOptional<z.ZodString>;
                     traceId: z.ZodOptional<z.ZodString>;
                     parentId: z.ZodOptional<z.ZodString>;
                     name: z.ZodString;
                     kind: z.ZodOptional<z.ZodEnum<{
                         custom: "custom";
-                        model: "model";
                         tool: "tool";
+                        model: "model";
                         run: "run";
                         agent: "agent";
                         guardrail: "guardrail";
@@ -881,9 +1095,9 @@ declare const EvalTaskMetaSchema: z.ZodObject<{
                         timestamp: z.ZodOptional<z.ZodString>;
                         attributes: z.ZodOptional<z.ZodRecord<z.ZodString, z.ZodType<JsonValue, unknown, z.core.$ZodTypeInternals<JsonValue, unknown>>>>;
                     }, z.core.$strict>>>;
-                }, z.core.$strict>>>;
+                }, z.core.$strict>>;
             }, z.core.$strict>>>;
-            errors: z.ZodDefault<z.ZodArray<z.ZodRecord<z.ZodString, z.ZodType<JsonValue, unknown, z.core.$ZodTypeInternals<JsonValue, unknown>>>>>;
+            errors: z.ZodArray<z.ZodRecord<z.ZodString, z.ZodType<JsonValue, unknown, z.core.$ZodTypeInternals<JsonValue, unknown>>>>;
         }, z.core.$strict>>;
     }, z.core.$strict>>;
 }, z.core.$strict>;
@@ -892,4 +1106,4 @@ type EvalTaskMeta = z.infer<typeof EvalTaskMetaSchema>;
 /** Reads eval metadata from an arbitrary Vitest assertion meta value. */
 declare function readEvalTaskMeta(input: unknown): EvalTaskMeta | undefined;
 
-export { type EvalMeta, EvalMetaSchema, type EvalScore, EvalScoreSchema, type EvalTaskMeta, EvalTaskMetaSchema, type GenAiOperationName, type GenAiOutputType, type GenAiProviderName, type GenAiSemanticAttributeKey, type GenAiSemanticAttributes, type GenAiTokenType, type GenAiToolType, type HarnessMeta, HarnessMetaSchema, type HarnessRun, type HarnessRunError, HarnessRunSchema, JsonValue, type NormalizedError, NormalizedErrorSchema, type NormalizedMessage, NormalizedMessageSchema, type NormalizedSession, NormalizedSessionSchema, type NormalizedSpan, type NormalizedSpanAttributeKey, type NormalizedSpanAttributes, NormalizedSpanAttributesSchema, type NormalizedSpanEvent, NormalizedSpanEventSchema, NormalizedSpanSchema, type NormalizedTrace, NormalizedTraceSchema, type OpenTelemetrySemanticAttributeKey, type OpenTelemetrySemanticAttributes, type TimingSummary, TimingSummarySchema, type ToolCallRecord, ToolCallRecordSchema, type UsageSummary, UsageSummarySchema, assistantMessages, failedSpans, latestAssistantMessageContent, messagesByRole, readEvalTaskMeta, spans, spansByKind, systemMessages, toolCalls, toolMessages, traceSpans, userMessages };
+export { type EvalMeta, EvalMetaSchema, type EvalScore, EvalScoreSchema, type EvalTaskMeta, EvalTaskMetaSchema, type GenAiOperationName, type GenAiOutputType, type GenAiProviderName, type GenAiSemanticAttributeKey, type GenAiSemanticAttributes, type GenAiTokenType, type GenAiToolType, type HarnessMeta, HarnessMetaSchema, type HarnessRun, type HarnessRunError, HarnessRunSchema, JsonValue, type NormalizedError, NormalizedErrorSchema, type NormalizedSession, NormalizedSessionSchema, type NormalizedSpan, type NormalizedSpanAttributeKey, type NormalizedSpanAttributes, NormalizedSpanAttributesSchema, type NormalizedSpanEvent, NormalizedSpanEventSchema, NormalizedSpanSchema, type NormalizedTrace, NormalizedTraceSchema, type OpenTelemetrySemanticAttributeKey, type OpenTelemetrySemanticAttributes, type TimingSummary, TimingSummarySchema, type ToolCall, ToolCallSchema, type TranscriptEvent, TranscriptEventSchema, type TranscriptMessageContentPart, type TranscriptMessageEvent, TranscriptMessageEventSchema, type TranscriptMessageInput, type TranscriptMessageTextPart, type TranscriptMessageToolCall, type TranscriptMessageToolCallPart, type TranscriptMessageToolResultPart, type TranscriptToolCallEvent, TranscriptToolCallEventSchema, type TranscriptToolResultEvent, TranscriptToolResultEventSchema, type UsageSummary, UsageSummarySchema, assistantMessages, failedSpans, isMessageEvent, isToolCallEvent, isToolResultEvent, latestAssistantMessageContent, messagesByRole, messagesToTranscriptEvents, readEvalTaskMeta, spans, spansByKind, systemMessages, toolCalls, toolMessages, traceSpans, userMessages };