@ls-stack/agent-eval 0.27.1 → 0.29.0

package/dist/{runOrchestration-FEvBwwJI.mjs → runOrchestration-CIARrLs6.mjs}

@@ -1,15 +1,15 @@
 import { createRequire, registerHooks } from "node:module";
-import { createHash } from "node:crypto";
+import { createHash, randomUUID } from "node:crypto";
 import { mkdir, readFile, readdir, rename, rm, stat, writeFile } from "node:fs/promises";
 import { extname, isAbsolute, join, relative, resolve } from "node:path";
 import { formatWithOptions, isDeepStrictEqual, stripVTControlCharacters } from "node:util";
 import { AsyncLocalStorage } from "node:async_hooks";
 import { z, z as z$1 } from "zod/v4";
-import { Buffer as Buffer$1 } from "node:buffer";
+import { Blob as Blob$1, Buffer as Buffer$1, File as File$1 } from "node:buffer";
 import { gunzipSync, gzipSync } from "node:zlib";
 import { getCompositeKey } from "@ls-stack/utils/getCompositeKey";
 import { existsSync } from "node:fs";
-import { resultify } from "t-result";
+import { Result, resultify } from "t-result";
 import { fileURLToPath, pathToFileURL } from "node:url";
 //#region ../sdk/src/defineEval.ts
 const evalRegistry = /* @__PURE__ */ new Map();
@@ -531,10 +531,13 @@ function recordOpIfActive(scope, op) {
 	if (top) top.ops.push(op);
 }
 function toAssertionFailure$1(message, error = void 0) {
-	return error?.stack ? {
+	const name = error?.name;
+	const stack = error?.stack ? stripTerminalControlCodes$1(error.stack) : void 0;
+	return {
+		...name !== void 0 ? { name } : {},
 		message,
-		stack: stripTerminalControlCodes$1(error.stack)
-	} : { message };
+		...stack !== void 0 ? { stack } : {}
+	};
 }
 /**
 * Record or replace an output value for the current case scope.
@@ -784,6 +787,67 @@ function evalExpect(value) {
 	return new EvalExpectationImpl(value, false);
 }
 //#endregion
+//#region ../sdk/src/manualInputFile.ts
+/**
+* Zod schema describing one file uploaded through the manual-input modal.
+*
+* Use this as the field type on your `manualInput.schema` whenever you mark
+* a field with `{ asFile: true }` in `manualInput.fields`. The UI / CLI stages
+* the selected file on disk, the runner materializes it into the run artifacts
+* directory, and the server validates this JSON metadata against the schema
+* before flowing it into the case input.
+*
+* @example
+* ```ts
+* const schema = z.object({
+*   image: manualInputFileValueSchema,
+*   note: z.string().optional(),
+* });
+*
+* defineEval({
+*   id: 'image-analyzer',
+*   manualInput: {
+*     schema,
+*     fields: { image: { asFile: true, accept: 'image/*' } },
+*   },
+*   // ...
+* });
+* ```
+*/
+const manualInputFileValueSchema = z.object({
+	name: z.string(),
+	mimeType: z.string(),
+	sizeBytes: z.number().int().nonnegative(),
+	sha256: z.string().regex(/^[a-f0-9]{64}$/),
+	path: z.string().min(1)
+});
+/**
+* Read a manual-input file artifact from disk and expose common byte, Blob,
+* File, text, and JSON views for eval code.
+*
+* @param value Manual-input file metadata received by an eval.
+* @param options.cwd Directory used to resolve relative paths. Defaults to `process.cwd()`.
+* @returns File bytes plus convenience views for common file-processing flows.
+*/
+async function readManualInputFile(value, options = {}) {
+	const absolutePath = resolve(options.cwd ?? process.cwd(), value.path);
+	const bytes = new Uint8Array(await readFile(absolutePath));
+	const arrayBuffer = bytes.buffer.slice(bytes.byteOffset, bytes.byteOffset + bytes.byteLength);
+	const blob = new Blob$1([bytes], { type: value.mimeType });
+	return {
+		value,
+		absolutePath,
+		bytes,
+		arrayBuffer,
+		blob,
+		file: new File$1([bytes], value.name, { type: value.mimeType }),
+		text: async () => await blob.text(),
+		json: async () => {
+			return JSON.parse(await blob.text());
+		}
+	};
+}
+//#endregion
 //#region ../sdk/src/repoFile.ts
 /**
 * Create a file reference that can be emitted via `setEvalOutput(...)` and rendered
@@ -802,7 +866,8 @@ function repoFile(path, mimeType) {
 }
 //#endregion
 //#region ../sdk/src/cacheSerialization.ts
-const serializedCacheValueMarker = "__agentEvalsCacheSerialization";
+const serializedCacheValueMarker = "__aecs";
+const legacySerializedCacheValueMarker = "__agentEvalsCacheSerialization";
 const jsonSafeCacheValueVersion = "json-safe-v1";
 const packedNumberArrayMinLength = 128;
 const compressedStringMinBytes = 16 * 1024;
@@ -812,7 +877,7 @@ function isRecordLike$3(value) {
 	return typeof value === "object" && value !== null && !Array.isArray(value);
 }
 function isJsonSafeSerializedCacheValue(value) {
-	return isRecordLike$3(value) && value[serializedCacheValueMarker] === jsonSafeCacheValueVersion && typeof value.type === "string";
+	return isRecordLike$3(value) && serializationMarkerValue(value) === jsonSafeCacheValueVersion && typeof value.type === "string";
 }
 function jsonSafeValue(type, value) {
 	return value === void 0 ? {
@@ -825,32 +890,39 @@ function jsonSafeValue(type, value) {
 	};
 }
 function hasSerializationMarkerKey(value) {
-	return Object.hasOwn(value, serializedCacheValueMarker);
+	return Object.hasOwn(value, serializedCacheValueMarker) || Object.hasOwn(value, legacySerializedCacheValueMarker);
+}
+function serializationMarkerValue(value) {
+	return value[serializedCacheValueMarker] ?? value[legacySerializedCacheValueMarker];
 }
 /**
 * Serialize one cached value while keeping plain JSON as plain JSON.
 *
-* Rich runtime values use small tagged wrappers.
+* Rich runtime values use small tagged wrappers. Undefined values are omitted
+* by default; pass `preserveUndefined: true` to round-trip them explicitly.
 */
-async function serializeCacheValue(value) {
-	return serializeJsonSafeValue(value, /* @__PURE__ */ new WeakSet(), 0);
+async function serializeCacheValue(value, options = void 0) {
+	return serializeJsonSafeValue(value, /* @__PURE__ */ new WeakSet(), 0, normalizeCacheSerializationOptions(options));
 }
 /** Revive one cached value, while preserving legacy JSON-round-tripped data. */
 function deserializeCacheValue(value) {
 	return deserializeJsonSafeValue(value);
 }
 /** Clone one value through the same serialization path used for cache data. */
-async function cloneCacheValue(value) {
-	return deserializeCacheValue(await serializeCacheValue(value));
+async function cloneCacheValue(value, options = void 0) {
+	return deserializeCacheValue(await serializeCacheValue(value, options));
+}
+function normalizeCacheSerializationOptions(options) {
+	return { preserveUndefined: options?.preserveUndefined === true };
 }
-async function serializeJsonSafeValue(value, refs, depth) {
-	if (value === void 0) return jsonSafeValue("Undefined");
+async function serializeJsonSafeValue(value, refs, depth, config) {
+	if (value === void 0) return config.preserveUndefined ? jsonSafeValue("Undefined") : void 0;
 	if (typeof value === "bigint") return jsonSafeValue("BigInt", value.toString());
 	if (typeof value === "number") return serializeNumber(value);
 	if (typeof value === "string") return serializeString(value, depth);
 	if (value instanceof Date) return jsonSafeValue("Date", value.toISOString());
-	if (value instanceof Map) return serializeMap(value, refs, depth);
-	if (value instanceof Set) return serializeSet(value, refs, depth);
+	if (value instanceof Map) return serializeMap(value, refs, depth, config);
+	if (value instanceof Set) return serializeSet(value, refs, depth, config);
 	if (value instanceof RegExp) return jsonSafeValue("RegExp", {
 		flags: value.flags,
 		source: value.source
@@ -869,7 +941,7 @@ async function serializeJsonSafeValue(value, refs, depth) {
 		type: value.type
 	});
 	if (value instanceof ArrayBuffer) return jsonSafeValue("ArrayBuffer", bytesToBase64(new Uint8Array(value)));
-	if (value instanceof Error) return serializeError(value, refs, depth);
+	if (value instanceof Error) return serializeError(value, refs, depth, config);
 	if (!value || typeof value !== "object") return value;
 	if (refs.has(value)) throw new Error("Circular cache values are not supported");
 	refs.add(value);
@@ -882,12 +954,18 @@ async function serializeJsonSafeValue(value, refs, depth) {
 			}
 		}
 		const items = [];
-		for (const item of value) items.push(await serializeJsonSafeValue(item, refs, depth + 1));
+		for (const item of value) {
+			const serializedItem = await serializeJsonSafeValue(item, refs, depth + 1, config);
+			if (serializedItem !== void 0) items.push(serializedItem);
+		}
 		refs.delete(value);
 		return compressNestedJsonValue(items, depth) ?? items;
 	}
 	const entries = [];
-	for (const [key, entryValue] of Object.entries(value)) entries.push([key, await serializeJsonSafeValue(entryValue, refs, depth + 1)]);
+	for (const [key, entryValue] of Object.entries(value)) {
+		const serializedEntryValue = await serializeJsonSafeValue(entryValue, refs, depth + 1, config);
+		if (serializedEntryValue !== void 0) entries.push([key, serializedEntryValue]);
+	}
 	refs.delete(value);
 	const serialized = hasSerializationMarkerKey(value) ? jsonSafeValue("Object", entries) : Object.fromEntries(entries);
 	return compressNestedJsonValue(serialized, depth) ?? serialized;
@@ -957,32 +1035,40 @@ function compressNestedJsonValue(value, depth) {
 function compressionIsWorthIt(value, rawSize) {
 	return Buffer$1.byteLength(JSON.stringify(value)) < rawSize * maxCompressedSizeRatio;
 }
-async function serializeMap(value, refs, depth) {
+async function serializeMap(value, refs, depth, config) {
 	if (refs.has(value)) throw new Error("Circular cache values are not supported");
 	refs.add(value);
 	const entries = [];
-	for (const [key, entryValue] of value.entries()) entries.push([await serializeJsonSafeValue(key, refs, depth + 1), await serializeJsonSafeValue(entryValue, refs, depth + 1)]);
+	for (const [key, entryValue] of value.entries()) {
+		const serializedKey = await serializeJsonSafeValue(key, refs, depth + 1, config);
+		const serializedEntryValue = await serializeJsonSafeValue(entryValue, refs, depth + 1, config);
+		if (serializedKey !== void 0 && serializedEntryValue !== void 0) entries.push([serializedKey, serializedEntryValue]);
+	}
 	refs.delete(value);
 	return jsonSafeValue("Map", entries);
 }
-async function serializeSet(value, refs, depth) {
+async function serializeSet(value, refs, depth, config) {
 	if (refs.has(value)) throw new Error("Circular cache values are not supported");
 	refs.add(value);
 	const items = [];
-	for (const item of value.values()) items.push(await serializeJsonSafeValue(item, refs, depth + 1));
+	for (const item of value.values()) {
+		const serializedItem = await serializeJsonSafeValue(item, refs, depth + 1, config);
+		if (serializedItem !== void 0) items.push(serializedItem);
+	}
 	refs.delete(value);
 	return jsonSafeValue("Set", items);
 }
-async function serializeError(value, refs, depth) {
+async function serializeError(value, refs, depth, config) {
 	if (refs.has(value)) throw new Error("Circular cache values are not supported");
 	refs.add(value);
 	const props = [];
 	for (const [key, entryValue] of Object.entries(value)) {
 		if (key === "cause") continue;
-		props.push([key, await serializeJsonSafeValue(entryValue, refs, depth + 1)]);
+		const serializedEntryValue = await serializeJsonSafeValue(entryValue, refs, depth + 1, config);
+		if (serializedEntryValue !== void 0) props.push([key, serializedEntryValue]);
 	}
 	const serialized = jsonSafeValue("Error", {
-		cause: "cause" in value ? await serializeJsonSafeValue(value.cause, refs, depth + 1) : void 0,
+		cause: "cause" in value ? await serializeJsonSafeValue(value.cause, refs, depth + 1, config) : void 0,
 		message: value.message,
 		name: value.name,
 		props,
@@ -1123,33 +1209,36 @@ function deserializeError(value) {
 	});
 	return error;
 }
-async function serializeRecordValues(record) {
+async function serializeRecordValues(record, config) {
 	const entries = [];
-	for (const [key, value] of Object.entries(record)) entries.push([key, await serializeCacheValue(value)]);
+	for (const [key, value] of Object.entries(record)) {
+		const serializedValue = await serializeCacheValue(value, config);
+		if (serializedValue !== void 0) entries.push([key, serializedValue]);
+	}
 	return Object.fromEntries(entries);
 }
 function deserializeRecordValues(record) {
 	return Object.fromEntries(Object.entries(record).map(([key, value]) => [key, deserializeCacheValue(value)]));
 }
-async function serializeCacheRecordingOp(op) {
+async function serializeCacheRecordingOp(op, config) {
 	switch (op.kind) {
 		case "setOutput":
 		case "appendOutput": return {
 			...op,
-			value: await serializeCacheValue(op.value)
+			value: await serializeCacheValue(op.value, config)
 		};
 		case "mergeOutput": return {
 			...op,
-			patch: await serializeRecordValues(op.patch)
+			patch: await serializeRecordValues(op.patch, config)
 		};
 		case "incrementOutput": return op;
 		case "checkpoint": return {
 			...op,
-			data: await serializeCacheValue(op.data)
+			data: await serializeCacheValue(op.data, config)
 		};
 		case "subSpan": return {
 			...op,
-			span: await serializeCacheSpan(op.span)
+			span: await serializeCacheSpan(op.span, config)
 		};
 	}
 }
@@ -1175,11 +1264,11 @@ function deserializeCacheRecordingOp(op) {
 		};
 	}
 }
-async function serializeCacheSpan(span) {
+async function serializeCacheSpan(span, config) {
 	return {
 		...span,
-		attributes: span.attributes === void 0 ? void 0 : await serializeRecordValues(span.attributes),
-		children: await Promise.all(span.children.map(serializeCacheSpan))
+		attributes: span.attributes === void 0 ? void 0 : await serializeRecordValues(span.attributes, config),
+		children: await Promise.all(span.children.map((child) => serializeCacheSpan(child, config)))
 	};
 }
 function deserializeCacheSpan(span) {
@@ -1189,13 +1278,19 @@ function deserializeCacheSpan(span) {
 		children: span.children.map(deserializeCacheSpan)
 	};
 }
-/** Serialize all rich values captured in a cache recording before persistence. */
-async function serializeCacheRecording(recording) {
+/**
+* Serialize all rich values captured in a cache recording before persistence.
+*
+* Undefined values are omitted by default; pass `preserveUndefined: true` to
+* retain the legacy explicit undefined wrappers in the recording payload.
+*/
+async function serializeCacheRecording(recording, options = void 0) {
+	const config = normalizeCacheSerializationOptions(options);
 	return {
 		...recording,
-		returnValue: await serializeCacheValue(recording.returnValue),
-		finalAttributes: await serializeRecordValues(recording.finalAttributes),
-		ops: await Promise.all(recording.ops.map(serializeCacheRecordingOp))
+		returnValue: await serializeCacheValue(recording.returnValue, config),
+		finalAttributes: await serializeRecordValues(recording.finalAttributes, config),
+		ops: await Promise.all(recording.ops.map((op) => serializeCacheRecordingOp(op, config)))
 	};
 }
 /** Revive all rich values captured in a cache recording after lookup. */
@@ -1587,7 +1682,9 @@ function createTraceCache(generateSpanId) {
 			key: info.key
 		}, { serializeFileBytes: info.serializeFileBytes === true });
 		const activeSpan = scope.activeSpanStack.at(-1);
-		if (cacheCtx.mode === "use") {
+		const canRead = cacheCtx.mode === "use" && cacheCtx.read !== false;
+		const canStore = cacheCtx.mode !== "bypass" && cacheCtx.store !== false;
+		if (canRead) {
 			const hit = await cacheCtx.adapter.lookup(namespace, keyHash);
 			if (hit) {
 				const storedAt = hit.storedAt;
@@ -1610,14 +1707,24 @@ function createTraceCache(generateSpanId) {
 				name: info.name,
 				namespace,
 				key: keyHash,
-				status: "miss"
+				status: "miss",
+				...canStore ? {} : { stored: false }
 			});
-		} else if (cacheCtx.mode === "refresh") recordCacheRef(scope, activeSpan, {
+		} else if (cacheCtx.mode === "use" && canStore) recordCacheRef(scope, activeSpan, {
+			type: "value",
+			name: info.name,
+			namespace,
+			key: keyHash,
+			status: "miss",
+			read: false
+		});
+		else if (cacheCtx.mode === "refresh") recordCacheRef(scope, activeSpan, {
 			type: "value",
 			name: info.name,
 			namespace,
 			key: keyHash,
-			status: "refresh"
+			status: "refresh",
+			...canStore ? {} : { stored: false }
 		});
 		else recordCacheRef(scope, activeSpan, {
 			type: "value",
@@ -1640,7 +1747,7 @@ function createTraceCache(generateSpanId) {
 			scope.recordingStack.pop();
 		}
 		appendSubSpanOps(scope, frame);
-		if (cacheCtx.mode !== "bypass") {
+		if (canStore) {
 			const finalAttributes = diffNonCacheAttributes(beforeAttributes, await snapshotNonCacheAttributes(activeSpan));
 			const recording = {
 				returnValue: bodyResult,
@@ -1654,13 +1761,11 @@ function createTraceCache(generateSpanId) {
 				operationType: "value",
 				operationName: info.name,
 				storedAt: new Date(getRealDateNowMs()).toISOString(),
-				codeFingerprint: cacheCtx.codeFingerprint,
 				recording: await serializeCacheRecording(recording)
 			}, {
 				rawKey: info.key,
 				operationType: "value",
-				operationName: info.name,
-				codeFingerprint: cacheCtx.codeFingerprint
+				operationName: info.name
 			});
 		}
 		return bodyResult;
@@ -2031,11 +2136,13 @@ async function traceSpanInternal(info, fn) {
 				namespace,
 				key: cacheOpts.key
 			}, { serializeFileBytes: cacheOpts.serializeFileBytes === true });
+			const canRead = ctx.mode === "use" && ctx.read !== false;
+			const canStore = ctx.mode !== "bypass" && ctx.store !== false;
 			mergeSpanAttributes(spanRecord, {
 				"cache.key": keyHash,
 				"cache.namespace": namespace
 			});
-			if (ctx.mode === "use") {
+			if (canRead) {
 				const hit = await ctx.adapter.lookup(namespace, keyHash);
 				if (hit) {
 					const storedAt = hit.storedAt;
@@ -2050,8 +2157,18 @@ async function traceSpanInternal(info, fn) {
 					spanRecord.endedAt = addElapsedMsToTimestamp(spanRecord.startedAt, getRealDateNowMs() - realStartedAt);
 					return recording.returnValue;
 				}
-				mergeSpanAttributes(spanRecord, { "cache.status": "miss" });
-			} else if (ctx.mode === "refresh") mergeSpanAttributes(spanRecord, { "cache.status": "refresh" });
+				mergeSpanAttributes(spanRecord, {
+					"cache.status": "miss",
+					...canStore ? {} : { "cache.stored": false }
+				});
+			} else if (ctx.mode === "use" && canStore) mergeSpanAttributes(spanRecord, {
+				"cache.status": "miss",
+				"cache.read": false
+			});
+			else if (ctx.mode === "refresh") mergeSpanAttributes(spanRecord, {
+				"cache.status": "refresh",
+				...canStore ? {} : { "cache.stored": false }
+			});
 			else mergeSpanAttributes(spanRecord, { "cache.status": "bypass" });
 			const frame = {
 				baseSpanIndex: scope.spans.length,
@@ -2067,7 +2184,7 @@ async function traceSpanInternal(info, fn) {
 			}
 			appendSubSpanOps(scope, frame);
 			finishSpanWithoutThrownError(spanRecord, realStartedAt);
-			if (ctx.mode !== "bypass") {
+			if (canStore) {
 				const recording = {
 					returnValue: bodyResult,
 					finalAttributes: stripCacheAttributes(spanRecord.attributes),
@@ -2087,14 +2204,12 @@ async function traceSpanInternal(info, fn) {
 					spanName: info.name,
 					spanKind: info.kind,
 					storedAt: new Date(getRealDateNowMs()).toISOString(),
-					codeFingerprint: ctx.codeFingerprint,
 					recording: await serializeCacheRecording(recording)
 				};
 				await ctx.adapter.write(entry, {
 					rawKey: cacheOpts.key,
 					operationType: "span",
-					operationName: info.name,
-					codeFingerprint: ctx.codeFingerprint
+					operationName: info.name
 				});
 			}
 			return bodyResult;
@@ -2287,6 +2402,7 @@ const columnDefSchema = z.object({
 	passThreshold: z.number().optional(),
 	maxStars: z.number().int().min(2).optional(),
 	hideInTable: z.boolean().optional(),
+	hideIfNoValue: z.boolean().optional(),
 	align: z.enum([
 		"left",
 		"center",
@@ -2430,6 +2546,10 @@ const traceCacheRefSchema = z.object({
 	namespace: z.string(),
 	key: z.string(),
 	status: cacheStatusSchema,
+	/** Whether this ref attempted to read from cache. Defaults to true. */
+	read: z.boolean().optional(),
+	/** Whether this ref wrote a persisted cache entry. Defaults to true for misses/refreshes. */
+	stored: z.boolean().optional(),
 	storedAt: z.string().optional(),
 	age: z.number().optional()
 });
@@ -2442,7 +2562,6 @@ const cacheListItemSchema = z.object({
 	spanName: z.string().optional(),
 	spanKind: traceSpanKindSchema.optional(),
 	storedAt: z.string(),
-	codeFingerprint: z.string(),
 	sizeBytes: z.number()
 });
 /** Zod schema for `SerializedCacheSpan`, defined lazily for recursion. */
@@ -2524,7 +2643,6 @@ const cacheEntrySchema = z.object({
 	spanName: z.string().optional(),
 	spanKind: traceSpanKindSchema.optional(),
 	storedAt: z.string(),
-	codeFingerprint: z.string(),
 	recording: cacheRecordingSchema
 });
 /** Debug-only raw key metadata stored outside the reusable cache entry. */
@@ -2535,7 +2653,6 @@ const cacheDebugKeyEntrySchema = z.object({
 	operationType: cacheOperationTypeSchema,
 	operationName: z.string(),
 	storedAt: z.string(),
-	codeFingerprint: z.string(),
 	rawKey: z.unknown()
 });
 /** Cache lookup response with optional debug-only raw key data. */
@@ -2627,6 +2744,16 @@ const evalChartTooltipExtraSchema = z.discriminatedUnion("source", [z.object({
 const evalChartConfigSchema = z.object({
 	/** Optional heading shown above the chart frame in the UI. */
 	heading: z.string().optional(),
+	/**
+	* Hide this chart in the UI when none of its metrics has a numeric value in
+	* the rendered history window.
+	*/
+	hideIfNoValue: z.boolean().optional(),
+	/**
+	* Drop consecutive history points whose plotted metrics and tooltip extras
+	* have the same values as the previous kept point.
+	*/
+	dedupeConsecutiveValues: z.boolean().optional(),
 	type: evalChartTypeSchema,
 	/** At least one series must be declared. */
 	metrics: z.array(evalChartMetricSchema).min(1),
@@ -2652,6 +2779,122 @@ const evalChartConfigSchema = z.object({
 */
 const evalChartsConfigSchema = z.array(evalChartConfigSchema);
 //#endregion
+//#region ../shared/src/schemas/manualInput.ts
+/**
+* Common metadata shared by every manual-input field descriptor exposed to
+* the web UI. The runner builds these from the eval's authored Zod schema and
+* any per-field overrides, so the client never needs the schema itself.
+*/
+const manualInputFieldBaseSchema = z.object({
+	/** Top-level key on the eval input object that this field writes to. */
+	key: z.string(),
+	/** Human-readable label rendered next to the field in the modal. */
+	label: z.string(),
+	/** Optional helper text rendered under the label. */
+	description: z.string().optional(),
+	/** Optional placeholder rendered inside the input element. */
+	placeholder: z.string().optional(),
+	/** Whether the field must be filled before the run can be submitted. */
+	required: z.boolean(),
+	/**
+	* Default value used to prefill the field. Type matches the underlying
+	* widget kind (`string` for text/multiline/select, `number` for number,
+	* `boolean` for boolean, JSON-serialisable for `json`).
+	*/
+	defaultValue: z.unknown().optional()
+});
+/** One option rendered by the `select` widget. */
+const manualInputSelectOptionSchema = z.object({
+	value: z.string(),
+	label: z.string()
+});
+/** Single line text widget descriptor. */
+const manualInputTextFieldSchema = manualInputFieldBaseSchema.extend({
+	kind: z.literal("text"),
+	/** Optional minimum character length enforced client-side. */
+	minLength: z.number().int().min(0).optional(),
+	/** Optional maximum character length enforced client-side. */
+	maxLength: z.number().int().min(0).optional()
+});
+/** Multi-line textarea widget descriptor. */
+const manualInputMultilineFieldSchema = manualInputFieldBaseSchema.extend({
+	kind: z.literal("multiline"),
+	/** Optional minimum character length enforced client-side. */
+	minLength: z.number().int().min(0).optional(),
+	/** Optional maximum character length enforced client-side. */
+	maxLength: z.number().int().min(0).optional(),
+	/** Suggested number of visible textarea rows; UI may clamp this. */
+	rows: z.number().int().min(1).optional()
+});
+/** Numeric input widget descriptor. */
+const manualInputNumberFieldSchema = manualInputFieldBaseSchema.extend({
+	kind: z.literal("number"),
+	/** Optional inclusive lower bound. */
+	min: z.number().optional(),
+	/** Optional inclusive upper bound. */
+	max: z.number().optional(),
+	/** Optional UI step increment. */
+	step: z.number().positive().optional(),
+	/** Whether the value must be an integer. */
+	integer: z.boolean().optional()
+});
+/** Boolean checkbox/toggle widget descriptor. */
+const manualInputBooleanFieldSchema = manualInputFieldBaseSchema.extend({ kind: z.literal("boolean") });
+/** Single-select dropdown widget descriptor. */
+const manualInputSelectFieldSchema = manualInputFieldBaseSchema.extend({
+	kind: z.literal("select"),
+	options: z.array(manualInputSelectOptionSchema)
+});
+/** JSON textarea widget descriptor used for nested objects, arrays, and unions. */
+const manualInputJsonFieldSchema = manualInputFieldBaseSchema.extend({
+	kind: z.literal("json"),
+	/** Suggested number of visible textarea rows; UI may clamp this. */
+	rows: z.number().int().min(1).optional()
+});
+/**
+* File / image upload widget descriptor. The widget supports clicking to
+* pick a file, drag-and-drop onto the dropzone, and pasting an image from
+* the system clipboard. The submitted value references a staged file artifact.
+*/
+const manualInputFileFieldSchema = manualInputFieldBaseSchema.extend({
+	kind: z.literal("file"),
+	/**
+	* Browser `accept` attribute (e.g. `image/*`, `image/png,image/jpeg`,
+	* `.pdf`). When omitted the picker accepts any file type.
+	*/
+	accept: z.string().optional(),
+	/** Optional client-side maximum file size in bytes. */
+	maxSizeBytes: z.number().int().positive().optional()
+});
+/**
+* Discriminated union of all supported manual-input widget kinds. The web UI
+* dispatches to the matching field component based on `kind`.
+*/
+const manualInputFieldDescriptorSchema = z.discriminatedUnion("kind", [
+	manualInputTextFieldSchema,
+	manualInputMultilineFieldSchema,
+	manualInputNumberFieldSchema,
+	manualInputBooleanFieldSchema,
+	manualInputSelectFieldSchema,
+	manualInputJsonFieldSchema,
+	manualInputFileFieldSchema
+]);
+/**
+* Wire-format descriptor attached to an `EvalSummary` when the eval declares
+* `manualInput`. Carries the ordered list of fields the modal renders and
+* basic context shown in the modal header.
+*/
+const manualInputDescriptorSchema = z.object({
+	/** Optional title shown in the modal header. Defaults to the eval title. */
+	title: z.string().optional(),
+	/** Optional helper text shown above the form. */
+	description: z.string().optional(),
+	/** Optional submit button label. Defaults to `Run`. */
+	submitLabel: z.string().optional(),
+	/** Ordered list of fields rendered in the modal. */
+	fields: z.array(manualInputFieldDescriptorSchema)
+});
+//#endregion
 //#region ../shared/src/schemas/eval.ts
 /** Freshness signal derived from the latest relevant run plus git state. */
 const evalFreshnessStatusSchema = z.enum([
@@ -2667,17 +2910,31 @@ const evalStatAggregateSchema = z.enum([
 	"sum",
 	"last"
 ]);
+const hideIfNoValueShape = { 
+/**
+* Hide this stat in the UI when the current run has no displayable value.
+* Missing values, `null`, and empty strings count as no value; `0` remains
+* visible.
+*/
+hideIfNoValue: z.boolean().optional() };
 /**
 * One entry in the EvalCard stats row. Built-in kinds use latest run totals;
 * `column` aggregates a score or numeric output column across the latest run.
 */
 const evalStatItemSchema = z.discriminatedUnion("kind", [
-	z.object({ kind: z.literal("cases") }),
+	z.object({
+		kind: z.literal("cases"),
+		...hideIfNoValueShape
+	}),
 	z.object({
 		kind: z.literal("passRate"),
-		accent: z.boolean().optional()
+		accent: z.boolean().optional(),
+		...hideIfNoValueShape
+	}),
+	z.object({
+		kind: z.literal("duration"),
+		...hideIfNoValueShape
 	}),
-	z.object({ kind: z.literal("duration") }),
 	z.object({
 		kind: z.literal("column"),
 		key: z.string(),
@@ -2686,7 +2943,8 @@ const evalStatItemSchema = z.discriminatedUnion("kind", [
 		format: columnFormatSchema.optional(),
 		/** Number presentation options applied when `format: 'number'`. */
 		numberFormat: numberDisplayOptionsSchema.optional(),
-		accent: z.boolean().optional()
+		accent: z.boolean().optional(),
+		...hideIfNoValueShape
 	})
 ]);
 /** Ordered list of stats rendered in the EvalCard stats row. */
@@ -2734,7 +2992,13 @@ const evalSummarySchema = z.object({
 	* Ordered per-eval history chart configuration for the EvalCard. Opt-in:
 	* when omitted or empty, the UI renders no history chart at all.
 	*/
-	charts: evalChartsConfigSchema.optional()
+	charts: evalChartsConfigSchema.optional(),
+	/**
+	* Manual-input form descriptor when the eval declares `manualInput`. The
+	* web UI renders these fields in a modal before kicking off a run; the
+	* runner consumes the validated values as the case input.
+	*/
+	manualInput: manualInputDescriptorSchema.optional()
 });
 /** Schema for one case row in an eval run result table. */
 const caseRowSchema = z.object({
@@ -2767,6 +3031,12 @@ const caseRowSchema = z.object({
 });
 /** Structured assertion failure metadata captured for one case run. */
 const assertionFailureSchema = z.object({
+	/**
+	* Error class or category label rendered alongside the message (e.g.
+	* `EvalAssertionError`, `OutputsSchemaError`). Optional for legacy entries
+	* and synthetic failures without an originating Error.
+	*/
+	name: z.string().optional(),
 	/** Human-readable assertion failure message shown in the UI and artifacts. */
 	message: z.string(),
 	/** Stack trace captured from the originating error when available. */
@@ -2868,7 +3138,7 @@ const caseDetailSchema = z.object({
 });
 /** Schema for discovery problems that should be shown before running evals. */
 const discoveryIssueSchema = z.object({
-	type: z.enum(["duplicate-eval-id"]),
+	type: z.enum(["duplicate-eval-id", "manual-input-with-cases"]),
 	severity: z.enum(["error"]),
 	filePath: z.string(),
 	evalId: z.string(),
@@ -2915,6 +3185,25 @@ const defaultConfigKeySchema = z.enum([
 ]);
 /** Removal config for built-in eval-level outputs and UI metadata. */
 const removeDefaultConfigSchema = z.union([z.literal(true), z.array(defaultConfigKeySchema)]);
+const evalDeriveValueFnSchema = z.custom((value) => typeof value === "function", { message: "Expected a derive output function" });
+/** Schema for keyed or object-returning trace-derived output config. */
+const evalDeriveConfigSchema = z.union([z.custom((value) => typeof value === "function", { message: "Expected a deriveFromTracing function" }), z.record(z.string().min(1), evalDeriveValueFnSchema)]);
+/** Schema for UI overrides on derived or scored columns. */
+const evalColumnOverrideSchema = z.object({
+	label: z.string().optional(),
+	format: columnFormatSchema.optional(),
+	numberFormat: numberDisplayOptionsSchema.optional(),
+	hideInTable: z.boolean().optional(),
+	hideIfNoValue: z.boolean().optional(),
+	align: z.enum([
+		"left",
+		"center",
+		"right"
+	]).optional(),
+	maxStars: z.number().int().min(2).optional()
+});
+/** Schema for column override maps keyed by output or score field name. */
+const evalColumnsSchema = z.record(z.string(), evalColumnOverrideSchema);
 /** Render formats supported by an LLM-call metric in the UI. */
 const llmCallMetricFormatSchema = z.enum([
 	"string",
@@ -2992,18 +3281,9 @@ const apiCallMetricSchema = z.object({
 	placements: z.array(apiCallMetricPlacementSchema).nonempty().optional()
 });
 /**
-* Schema for one model/provider pricing entry used to derive LLM-call costs
-* from token counts.
+* Schema for pricing rates used to derive LLM-call costs from token counts.
 */
-const llmCallPricingSchema = z.object({
-	/** Exact model name read from the configured `attributes.model` path. */
-	model: z.string().min(1),
-	/**
-	* Optional provider discriminator read from `attributes.provider`. When set,
-	* the entry only applies to calls from that provider; provider-specific
-	* entries take precedence over generic entries for the same model.
-	*/
-	provider: z.string().min(1).optional(),
+const llmCallPricingRateSchema = z.object({
 	/** USD per one million non-cached input tokens. */
 	inputUsdPerMillion: z.number().nonnegative().optional(),
 	/** USD per one million output tokens. */
@@ -3017,6 +3297,23 @@ const llmCallPricingSchema = z.object({
 	/** USD per one million reasoning tokens when reported separately. */
 	reasoningUsdPerMillion: z.number().nonnegative().optional()
 });
+/**
+* Schema for one model's pricing config. The object key is the exact model
+* name. Use `providers` when a model has provider-specific rates in addition
+* to, or instead of, generic model rates.
+*/
+const llmCallPricingSchema = llmCallPricingRateSchema.extend({
+	/**
+	* Optional provider discriminator read from `attributes.provider`. When set,
+	* the top-level entry only applies to calls from that provider.
+	*/
+	provider: z.string().min(1).optional(),
+	/**
+	* Provider-specific pricing for the model. Provider entries take precedence
+	* over generic rates for the same model.
+	*/
+	providers: z.record(z.string().min(1), llmCallPricingRateSchema).optional()
+});
 /** Schema for the global LLM calls config block in `agent-evals.config.ts`. */
 const llmCallsConfigSchema = z.object({
 	/** Span kinds treated as LLM calls. Defaults to `['llm']`. */
@@ -3053,10 +3350,10 @@ const llmCallsConfigSchema = z.object({
 	*/
 	derivedAttributes: z.record(z.string().min(1), callDerivedAttributeSchema).optional(),
 	/**
-	* Model/provider pricing registry used to calculate LLM-call costs from
-	* token counts. Built-in LLM cost fields are only derived from this registry.
+	* Model-keyed pricing registry used to calculate LLM-call costs from token
+	* counts. Built-in LLM cost fields are only derived from this registry.
 	*/
-	pricing: z.array(llmCallPricingSchema).optional(),
+	pricing: z.record(z.string().min(1), llmCallPricingSchema).optional(),
 	/** Custom user-defined metrics surfaced on each LLM call. */
 	metrics: z.array(llmCallMetricSchema).optional()
 });
@@ -3172,6 +3469,33 @@ function resolveApiCallMetric(metric) {
 		placements: metric.placements ? [...metric.placements] : ["body"]
 	};
 }
+function hasPricingRates(pricing) {
+	return pricing.inputUsdPerMillion !== void 0 || pricing.outputUsdPerMillion !== void 0 || pricing.cachedInputUsdPerMillion !== void 0 || pricing.cacheCreationInputUsdPerMillion !== void 0 || pricing.cacheCreationInput1hUsdPerMillion !== void 0 || pricing.reasoningUsdPerMillion !== void 0;
+}
+function copyPricingRates(pricing) {
+	return {
+		inputUsdPerMillion: pricing.inputUsdPerMillion,
+		outputUsdPerMillion: pricing.outputUsdPerMillion,
+		cachedInputUsdPerMillion: pricing.cachedInputUsdPerMillion,
+		cacheCreationInputUsdPerMillion: pricing.cacheCreationInputUsdPerMillion,
+		cacheCreationInput1hUsdPerMillion: pricing.cacheCreationInput1hUsdPerMillion,
+		reasoningUsdPerMillion: pricing.reasoningUsdPerMillion
+	};
+}
+function resolveLlmCallPricingEntries(model, pricing) {
+	const entries = [];
+	if (hasPricingRates(pricing)) entries.push({
+		model,
+		provider: pricing.provider,
+		...copyPricingRates(pricing)
+	});
+	for (const [provider, providerPricing] of Object.entries(pricing.providers ?? {})) entries.push({
+		model,
+		provider,
+		...copyPricingRates(providerPricing)
+	});
+	return entries;
+}
 /**
 * Resolve the user-authored LLM-calls config to a fully-defaulted shape used
 * by the UI to derive the LLM calls tab.
@@ -3182,7 +3506,7 @@ function resolveApiCallMetric(metric) {
 * - Missing `metrics[].format` defaults to `'string'`.
 * - Missing `metrics[].placements` defaults to `['body']`.
 * - Missing `pricing` defaults to an empty registry; built-in costs are only
-*   derived from configured pricing and token counts.
+*   derived from configured model-keyed pricing and token counts.
 */
 function resolveLlmCallsConfig(input) {
 	return {
@@ -3193,16 +3517,7 @@ function resolveLlmCallsConfig(input) {
 		},
 		derivedAttributes: resolveDerivedAttributes(input?.derivedAttributes),
 		metrics: (input?.metrics ?? []).map(resolveLlmCallMetric),
-		pricing: (input?.pricing ?? []).map((p) => ({
-			model: p.model,
-			provider: p.provider,
-			inputUsdPerMillion: p.inputUsdPerMillion,
-			outputUsdPerMillion: p.outputUsdPerMillion,
-			cachedInputUsdPerMillion: p.cachedInputUsdPerMillion,
-			cacheCreationInputUsdPerMillion: p.cacheCreationInputUsdPerMillion,
-			cacheCreationInput1hUsdPerMillion: p.cacheCreationInput1hUsdPerMillion,
-			reasoningUsdPerMillion: p.reasoningUsdPerMillion
-		}))
+		pricing: Object.entries(input?.pricing ?? {}).flatMap(([model, pricing]) => resolveLlmCallPricingEntries(model, pricing))
 	};
 }
 /**
@@ -3236,6 +3551,9 @@ const agentEvalsConfigSchema = z.object({
 	staleAfterDays: z.number().optional(),
 	allowCliRunAll: z.boolean().optional(),
 	traceDisplay: traceDisplayInputConfigSchema.optional(),
+	columns: evalColumnsSchema.optional(),
+	deriveFromTracing: evalDeriveConfigSchema.optional(),
+	stats: evalStatsConfigSchema.optional(),
 	llmCalls: llmCallsConfigSchema.optional(),
 	removeDefaultConfig: removeDefaultConfigSchema.optional(),
 	apiCalls: apiCallsConfigSchema.optional(),
@@ -3888,6 +4206,11 @@ function readNumber(attributes, key) {
 	const value = attributes[key];
 	return typeof value === "number" && Number.isFinite(value) ? value : void 0;
 }
+function readBoolean(attributes, key) {
+	if (!isRecord$2(attributes)) return void 0;
+	const value = attributes[key];
+	return typeof value === "boolean" ? value : void 0;
+}
 function readArray(attributes, key) {
 	if (!isRecord$2(attributes)) return [];
 	const value = attributes[key];
@@ -3916,12 +4239,14 @@ function extractCacheEntries(spans, caseCacheRefs) {
 			const namespace = readString(span.attributes, "cache.namespace");
 			if (key !== void 0 && namespace !== void 0) {
 				const isHit = status === "hit";
+				const stored = isHit ? true : readBoolean(span.attributes, "cache.stored") !== false;
 				entries.push({
 					id: span.id,
 					source: "span",
 					origin: "span",
-					action: isHit ? "hit" : "added",
+					action: isHit ? "hit" : stored ? "added" : "notStored",
 					status,
+					stored,
 					name: span.name,
 					namespace,
 					key,
@@ -3938,12 +4263,14 @@ function extractCacheEntries(spans, caseCacheRefs) {
 			const ref = parsed.data;
 			if (ref.status === "bypass") continue;
 			const isHit = ref.status === "hit";
+			const stored = isHit ? true : ref.stored !== false;
 			entries.push({
 				id: `${span.id}:value:${String(index)}`,
 				source: "value",
 				origin: "span",
-				action: isHit ? "hit" : "added",
+				action: isHit ? "hit" : stored ? "added" : "notStored",
 				status: ref.status,
+				stored,
 				name: ref.name,
 				namespace: ref.namespace,
 				key: ref.key,
@@ -3956,12 +4283,14 @@ function extractCacheEntries(spans, caseCacheRefs) {
 	for (const [index, ref] of caseCacheRefs.entries()) {
 		if (ref.status === "bypass") continue;
 		const isHit = ref.status === "hit";
+		const stored = isHit ? true : ref.stored !== false;
 		entries.push({
 			id: `case:value:${String(index)}`,
 			source: "value",
 			origin: "caseRoot",
-			action: isHit ? "hit" : "added",
+			action: isHit ? "hit" : stored ? "added" : "notStored",
 			status: ref.status,
+			stored,
 			name: ref.name,
 			namespace: ref.namespace,
 			key: ref.key,
@@ -3987,6 +4316,7 @@ function isCacheHitEntry(entry) {
 }
 z.enum([
 	"discovery.updated",
+	"config.reload",
 	"run.started",
 	"run.summary",
 	"case.started",
@@ -4006,6 +4336,19 @@ const sseEnvelopeSchema = z.object({
 });
 //#endregion
 //#region ../shared/src/schemas/api.ts
+/** Lifecycle state for an app config reload triggered by `agent-evals.config.ts`. */
+const configReloadStatusSchema = z.enum([
+	"idle",
+	"pending",
+	"reloading"
+]);
+/** UI/API-visible state for config reloads in `agent-evals app`. */
+const configReloadStateSchema = z.object({
+	status: configReloadStatusSchema,
+	activeRunCount: z.number().int().min(0),
+	lastChangedAt: z.string().nullable(),
+	lastReloadedAt: z.string().nullable()
+});
 /** Schema for the API request that starts a new eval run. */
 const createRunRequestSchema = z.object({
 	target: z.object({
@@ -4026,14 +4369,22 @@ const createRunRequestSchema = z.object({
 	* Optional cache controls for the run. When omitted, the cache is used in
 	* its default read-through / write-on-miss mode.
 	*/
-	cache: z.object({ mode: cacheModeSchema.default("use") }).optional()
+	cache: z.object({ mode: cacheModeSchema.default("use") }).optional(),
+	/**
+	* Manual-input values keyed by eval `key` (workspace-relative file path
+	* plus authored eval id). Required for any targeted eval that declares
+	* `manualInput` in its definition; the server validates each entry against
+	* the eval's authored Zod schema before starting the run.
+	*/
+	manualInputs: z.record(z.string(), z.unknown()).optional()
 });
 /** Schema for updating a UI-authored manual score on one persisted case. */
 const updateManualScoreRequestSchema = z.object({ value: z.number().min(0).max(1).nullable() });
 //#endregion
 //#region ../runner/src/cacheStore.ts
 const defaultMaxEntriesPerNamespace = 100;
-const cacheSerializationMarker = "__agentEvalsCacheSerialization";
+const cacheSerializationMarker = "__aecs";
+const legacyCacheSerializationMarker = "__agentEvalsCacheSerialization";
 const supportedCacheSerializationVersion = "json-safe-v1";
 /**
 * Create a filesystem-backed cache adapter rooted at `<workspaceRoot>/<dir>`.
@@ -4118,7 +4469,6 @@ function createFsCacheStore(options) {
 						spanName: entry.spanName,
 						spanKind: entry.spanKind,
 						storedAt: entry.storedAt,
-						codeFingerprint: entry.codeFingerprint,
 						sizeBytes: Buffer.byteLength(JSON.stringify(entry), "utf8")
 					});
 				}
@@ -4247,7 +4597,7 @@ async function readCacheFilePath(filePath) {
 function usesSupportedCacheSerialization(value) {
 	if (Array.isArray(value)) return value.every(usesSupportedCacheSerialization);
 	if (!isRecordLike(value)) return true;
-	if (Object.hasOwn(value, cacheSerializationMarker) && value[cacheSerializationMarker] !== supportedCacheSerializationVersion) return false;
+	for (const marker of [cacheSerializationMarker, legacyCacheSerializationMarker]) if (Object.hasOwn(value, marker) && value[marker] !== supportedCacheSerializationVersion) return false;
 	return Object.values(value).every(usesSupportedCacheSerialization);
 }
 async function writeOrRemoveCacheFile(cacheDir, cacheFile) {
@@ -4291,7 +4641,6 @@ async function writeDebugKeyEntry(params) {
 			operationType: debugKey.operationType,
 			operationName: debugKey.operationName,
 			storedAt: entry.storedAt,
-			codeFingerprint: debugKey.codeFingerprint,
 			rawKey: debugKey.rawKey
 		};
 		await writeDebugKeyFile(debugDir, {
@@ -4507,6 +4856,7 @@ function getScoreOverride(def) {
 		format: def.format,
 		numberFormat: def.numberFormat,
 		hideInTable: def.hideInTable,
+		hideIfNoValue: def.hideIfNoValue,
 		align: def.align,
 		maxStars: def.maxStars
 	};
@@ -4519,6 +4869,7 @@ function mergeOverrides(base, override) {
 		format: override.format ?? base.format,
 		numberFormat: override.numberFormat ?? base.numberFormat,
 		hideInTable: override.hideInTable ?? base.hideInTable,
+		hideIfNoValue: override.hideIfNoValue ?? base.hideIfNoValue,
 		align: override.align ?? base.align,
 		maxStars: override.maxStars ?? base.maxStars
 	};
@@ -4633,6 +4984,7 @@ function createColumnDef(params) {
 	if (override?.numberFormat !== void 0) def.numberFormat = override.numberFormat;
 	if (override?.maxStars !== void 0) def.maxStars = override.maxStars;
 	if (override?.hideInTable !== void 0) def.hideInTable = override.hideInTable;
+	if (override?.hideIfNoValue !== void 0) def.hideIfNoValue = override.hideIfNoValue;
 	if (override?.align !== void 0) def.align = override.align;
 	if (!isScore) return def;
 	def.isScore = true;
@@ -4676,7 +5028,9 @@ async function loadConfig() {
 	const configPath = resolve(process.cwd(), "agent-evals.config.ts");
 	if (!existsSync(configPath)) return defaultConfig;
 	try {
-		const imported = await import(pathToFileURL(configPath).href);
+		const configUrl = pathToFileURL(configPath);
+		configUrl.searchParams.set("v", randomUUID());
+		const imported = await import(configUrl.href);
 		const configModule = configModuleSchema.parse(imported);
 		const userConfig = configModule.default ?? configModule.config;
 		if (!userConfig) return defaultConfig;
@@ -4717,60 +5071,70 @@ const DEFAULT_COLUMNS = {
 		label: "API Calls",
 		format: "number",
 		numberFormat: countNumberFormat,
-		align: "right"
+		align: "right",
+		hideIfNoValue: true
 	},
 	costUsd: {
 		label: "Cost",
 		format: "number",
 		numberFormat: costNumberFormat,
-		align: "right"
+		align: "right",
+		hideIfNoValue: true
 	},
 	llmTurns: {
 		label: "LLM Turns",
 		format: "number",
 		numberFormat: countNumberFormat,
-		align: "right"
+		align: "right",
+		hideIfNoValue: true
 	},
 	inputTokens: {
 		label: "Input Tokens",
 		format: "number",
 		numberFormat: tokenNumberFormat,
-		align: "right"
+		align: "right",
+		hideIfNoValue: true
 	},
 	outputTokens: {
 		label: "Output Tokens",
 		format: "number",
 		numberFormat: tokenNumberFormat,
-		align: "right"
+		align: "right",
+		hideIfNoValue: true
 	},
 	totalTokens: {
 		label: "Total Tokens",
 		format: "number",
 		numberFormat: tokenNumberFormat,
-		align: "right"
+		align: "right",
+		hideIfNoValue: true
 	},
 	cachedInputTokens: {
 		label: "Cached Input Tokens",
 		format: "number",
 		numberFormat: tokenNumberFormat,
-		align: "right"
+		align: "right",
+		hideIfNoValue: true
 	},
 	cacheCreationInputTokens: {
 		label: "Cache Write Tokens",
 		format: "number",
 		numberFormat: tokenNumberFormat,
-		align: "right"
+		align: "right",
+		hideIfNoValue: true
 	},
 	reasoningTokens: {
 		label: "Reasoning Tokens",
 		format: "number",
 		numberFormat: tokenNumberFormat,
-		align: "right"
+		align: "right",
+		hideIfNoValue: true
 	},
 	llmDurationMs: {
 		label: "LLM Duration",
 		format: "duration",
-		align: "right"
+		align: "right",
+		hideIfNoValue: true
 	}
 };
 function resolveRemovedKeys(globalRemove, evalRemove) {
@@ -4783,9 +5147,16 @@ function getActiveDefaultConfigKeys(params) {
 }
 function mergeDefaultColumns(params) {
 	const activeKeys = getActiveDefaultConfigKeys(params);
-	if (activeKeys.length === 0) return params.columns;
+	if (activeKeys.length === 0) {
+		const merged = {
+			...params.globalColumns,
+			...params.columns
+		};
+		return Object.keys(merged).length > 0 ? merged : void 0;
+	}
 	return {
 		...Object.fromEntries(activeKeys.map((key) => [key, DEFAULT_COLUMNS[key]])),
+		...params.globalColumns,
 		...params.columns
 	};
 }
@@ -4797,30 +5168,38 @@ function appendDefaultStats(params) {
 		key: "apiCalls",
 		label: "API Calls",
 		aggregate: "avg",
-		numberFormat: countNumberFormat
+		numberFormat: countNumberFormat,
+		hideIfNoValue: true
 	});
 	if (activeKeys.has("costUsd")) defaults.push({
 		kind: "column",
 		key: "costUsd",
 		label: "LLM Cost",
 		aggregate: "avg",
-		numberFormat: costNumberFormat
+		numberFormat: costNumberFormat,
+		hideIfNoValue: true
 	});
 	if (activeKeys.has("totalTokens")) defaults.push({
 		kind: "column",
 		key: "totalTokens",
 		label: "Tokens",
 		aggregate: "avg",
-		numberFormat: tokenNumberFormat
+		numberFormat: tokenNumberFormat,
+		hideIfNoValue: true
 	});
 	if (activeKeys.has("llmTurns")) defaults.push({
 		kind: "column",
 		key: "llmTurns",
 		label: "LLM Turns",
 		aggregate: "avg",
-		numberFormat: countNumberFormat
+		numberFormat: countNumberFormat,
+		hideIfNoValue: true
 	});
-	const merged = [...params.stats ?? [], ...defaults];
+	const merged = [
+		...params.globalStats ?? [],
+		...params.stats ?? [],
+		...defaults
+	];
 	return merged.length > 0 ? merged : void 0;
 }
 function appendDefaultCharts(params) {
@@ -4828,6 +5207,8 @@ function appendDefaultCharts(params) {
 	const defaults = [];
 	if (activeKeys.has("costUsd")) defaults.push({
 		heading: "LLM Cost",
+		hideIfNoValue: true,
+		dedupeConsecutiveValues: true,
 		type: "area",
 		metrics: [{
 			source: "column",
@@ -4837,7 +5218,7 @@ function appendDefaultCharts(params) {
 			color: "warning"
 		}]
 	});
-	const tokenMetrics = [
+	const inputTokenMetrics = [
 		activeKeys.has("inputTokens") ? {
 			source: "column",
 			key: "inputTokens",
@@ -4845,13 +5226,6 @@ function appendDefaultCharts(params) {
 			label: "Input",
 			color: "accent"
 		} : null,
-		activeKeys.has("outputTokens") ? {
-			source: "column",
-			key: "outputTokens",
-			aggregate: "avg",
-			label: "Output",
-			color: "success"
-		} : null,
 		activeKeys.has("cachedInputTokens") ? {
 			source: "column",
 			key: "cachedInputTokens",
@@ -4867,16 +5241,25 @@ function appendDefaultCharts(params) {
 			color: "warning"
 		} : null
 	].filter((metric) => metric !== null);
-	if (tokenMetrics.length > 0) defaults.push({
-		heading: "LLM Tokens",
+	if (inputTokenMetrics.length > 0) defaults.push({
+		heading: "LLM Input Tokens",
+		hideIfNoValue: true,
+		dedupeConsecutiveValues: true,
 		type: "bar",
-		metrics: tokenMetrics,
-		tooltipExtras: activeKeys.has("totalTokens") ? [{
+		metrics: inputTokenMetrics
+	});
+	if (activeKeys.has("outputTokens")) defaults.push({
+		heading: "LLM Output Tokens",
+		hideIfNoValue: true,
+		dedupeConsecutiveValues: true,
+		type: "bar",
+		metrics: [{
 			source: "column",
-			key: "totalTokens",
+			key: "outputTokens",
 			aggregate: "avg",
-			label: "Total"
-		}] : void 0
+			label: "Output",
+			color: "success"
+		}]
 	});
 	const merged = [...params.charts ?? [], ...defaults];
 	return merged.length > 0 ? merged : void 0;
@@ -4885,11 +5268,13 @@ function resolveEvalDefaultConfig(params) {
 	const evalRemove = params.evalDef.removeDefaultConfig;
 	return {
 		columns: mergeDefaultColumns({
+			globalColumns: params.globalColumns,
 			columns: params.evalDef.columns,
 			globalRemove: params.globalRemove,
 			evalRemove
 		}),
 		stats: appendDefaultStats({
+			globalStats: params.globalStats,
 			stats: params.evalDef.stats,
 			globalRemove: params.globalRemove,
 			evalRemove
@@ -5144,6 +5529,371 @@ function getRunFreshnessTimestamp(manifest) {
 	return manifest.endedAt ?? manifest.startedAt;
 }
 //#endregion
+//#region ../runner/src/manualInput/walker.ts
+function isObject(value) {
+	return typeof value === "object" && value !== null;
+}
+function getZodDef(schema) {
+	if (!isObject(schema)) return null;
+	const zodHolder = schema._zod;
+	if (!isObject(zodHolder)) return null;
+	const def = zodHolder.def;
+	if (!isObject(def)) return null;
+	if (typeof def.type !== "string") return null;
+	return {
+		...def,
+		type: def.type
+	};
+}
+function getDescription(schema) {
+	if (!isObject(schema)) return void 0;
+	const description = schema.description;
+	return typeof description === "string" ? description : void 0;
+}
+function getInnerSchema(def) {
+	return def.innerType;
+}
+function getChecks(def) {
+	const checks = def.checks;
+	if (!Array.isArray(checks)) return [];
+	const out = [];
+	for (const check of checks) {
+		if (!isObject(check)) continue;
+		const zodHolder = check._zod;
+		if (!isObject(zodHolder)) continue;
+		const checkDef = zodHolder.def;
+		if (!isObject(checkDef)) continue;
+		if (typeof checkDef.check !== "string") continue;
+		out.push({
+			...checkDef,
+			check: checkDef.check
+		});
+	}
+	return out;
+}
+function findCheck(checks, name) {
+	return checks.find((check) => check.check === name);
+}
+function unwrap(schema) {
+	let current = schema;
+	let required = true;
+	let defaultValue = void 0;
+	for (let depth = 0; depth < 8; depth += 1) {
+		const def = getZodDef(current);
+		if (!def) return null;
+		if (def.type === "optional" || def.type === "nullable") {
+			required = false;
+			current = getInnerSchema(def);
+			continue;
+		}
+		if (def.type === "nullish") {
+			required = false;
+			current = getInnerSchema(def);
+			continue;
+		}
+		if (def.type === "default" || def.type === "prefault") {
+			const raw = def.defaultValue;
+			if (typeof raw === "function") defaultValue = Reflect.apply(raw, void 0, []);
+			else defaultValue = raw;
+			current = getInnerSchema(def);
+			continue;
+		}
+		if (def.type === "readonly" || def.type === "pipe") {
+			current = getInnerSchema(def) ?? def.in;
+			continue;
+		}
+		return {
+			schema: current,
+			def,
+			required,
+			defaultValue
+		};
+	}
+	return null;
+}
+function humaniseKey(key) {
+	const spaced = key.replace(/([a-z0-9])([A-Z])/g, "$1 $2").replace(/[_-]+/g, " ").trim();
+	if (!spaced) return key;
+	const lowered = spaced.toLowerCase();
+	return lowered.charAt(0).toUpperCase() + lowered.slice(1);
+}
+function normaliseSelectOptions(raw) {
+	if (!raw) return void 0;
+	return raw.map((entry) => {
+		if (typeof entry === "string") return {
+			value: entry,
+			label: entry
+		};
+		return {
+			value: entry.value,
+			label: entry.label ?? entry.value
+		};
+	});
+}
+function enumOptionsFromEntries(def) {
+	const entries = def.entries;
+	if (!isObject(entries)) return null;
+	const out = [];
+	for (const [label, value] of Object.entries(entries)) if (typeof value === "string") out.push({
+		value,
+		label
+	});
+	else if (typeof value === "number") out.push({
+		value: String(value),
+		label
+	});
+	else return null;
+	return out;
+}
+function literalUnionOptions(def) {
+	const options = def.options;
+	if (!Array.isArray(options)) return null;
+	const out = [];
+	for (const option of options) {
+		const optDef = getZodDef(option);
+		if (optDef?.type !== "literal") return null;
+		const values = optDef.values;
+		if (!Array.isArray(values) || values.length !== 1) return null;
+		const value = values[0];
+		if (typeof value === "string") out.push({
+			value,
+			label: value
+		});
+		else if (typeof value === "number") {
+			const stringValue = String(value);
+			out.push({
+				value: stringValue,
+				label: stringValue
+			});
+		} else return null;
+	}
+	return out.length > 0 ? out : null;
+}
+function literalSelectOptions(def) {
+	const values = def.values;
+	if (!Array.isArray(values)) return null;
+	const out = [];
+	for (const value of values) if (typeof value === "string") out.push({
+		value,
+		label: value
+	});
+	else if (typeof value === "number") {
+		const stringValue = String(value);
+		out.push({
+			value: stringValue,
+			label: stringValue
+		});
+	} else return null;
+	return out;
+}
+function readStringChecks(def) {
+	const checks = getChecks(def);
+	const out = {};
+	const min = findCheck(checks, "min_length");
+	if (min && typeof min.minimum === "number") out.minLength = min.minimum;
+	const max = findCheck(checks, "max_length");
+	if (max && typeof max.maximum === "number") out.maxLength = max.maximum;
+	return out;
+}
+const integerNumberFormats = new Set([
+	"int",
+	"safeint",
+	"int32",
+	"uint32",
+	"int64",
+	"uint64"
+]);
+function readNumberChecks(def) {
+	const checks = getChecks(def);
+	const out = {};
+	const gt = findCheck(checks, "greater_than");
+	if (gt && typeof gt.value === "number" && gt.inclusive === true) out.min = gt.value;
+	const lt = findCheck(checks, "less_than");
+	if (lt && typeof lt.value === "number" && lt.inclusive === true) out.max = lt.value;
+	const format = findCheck(checks, "number_format");
+	if (format && typeof format.format === "string" && integerNumberFormats.has(format.format)) out.integer = true;
+	return out;
+}
+function buildField(key, fieldSchema, override) {
+	const unwrapped = unwrap(fieldSchema);
+	if (!unwrapped) return Result.err(/* @__PURE__ */ new Error(`manualInput: field "${key}" uses an unsupported Zod schema (could not introspect)`));
+	const inner = unwrapped.def;
+	const description = override?.description ?? getDescription(unwrapped.schema);
+	const base = {
+		key,
+		label: override?.label ?? humaniseKey(key),
+		description,
+		placeholder: override?.placeholder,
+		required: unwrapped.required,
+		defaultValue: override?.defaultValue !== void 0 ? override.defaultValue : unwrapped.defaultValue
+	};
+	if (override?.asJson === true) {
+		const rows = override.rows;
+		return Result.ok({
+			...base,
+			kind: "json",
+			rows
+		});
+	}
+	if (override?.asFile === true) return Result.ok({
+		...base,
+		kind: "file",
+		accept: override.accept,
+		maxSizeBytes: override.maxSizeBytes
+	});
+	const overrideOptions = normaliseSelectOptions(override?.options);
+	if (overrideOptions) return Result.ok({
+		...base,
+		kind: "select",
+		options: overrideOptions
+	});
+	switch (inner.type) {
+		case "string": {
+			const checks = readStringChecks(inner);
+			if (override?.multiline === true) return Result.ok({
+				...base,
+				kind: "multiline",
+				rows: override.rows,
+				minLength: checks.minLength,
+				maxLength: checks.maxLength
+			});
+			return Result.ok({
+				...base,
+				kind: "text",
+				minLength: checks.minLength,
+				maxLength: checks.maxLength
+			});
+		}
+		case "number":
+		case "int":
+		case "bigint": {
+			const checks = readNumberChecks(inner);
+			return Result.ok({
+				...base,
+				kind: "number",
+				min: checks.min,
+				max: checks.max,
+				integer: checks.integer
+			});
+		}
+		case "boolean": return Result.ok({
+			...base,
+			kind: "boolean"
+		});
+		case "enum": {
+			const options = enumOptionsFromEntries(inner);
+			if (options) return Result.ok({
+				...base,
+				kind: "select",
+				options
+			});
+			return Result.ok({
+				...base,
+				kind: "json",
+				rows: override?.rows
+			});
+		}
+		case "literal": {
+			const options = literalSelectOptions(inner);
+			if (options && options.length > 0) return Result.ok({
+				...base,
+				kind: "select",
+				options
+			});
+			return Result.ok({
+				...base,
+				kind: "json",
+				rows: override?.rows
+			});
+		}
+		case "union": {
+			const options = literalUnionOptions(inner);
+			if (options) return Result.ok({
+				...base,
+				kind: "select",
+				options
+			});
+			return Result.ok({
+				...base,
+				kind: "json",
+				rows: override?.rows
+			});
+		}
+		default: return Result.ok({
+			...base,
+			kind: "json",
+			rows: override?.rows
+		});
+	}
+}
+function getObjectShape(schema) {
+	const def = getZodDef(schema);
+	if (!def) return null;
+	if (def.type !== "object") return null;
+	const shape = def.shape;
+	if (!isObject(shape)) return null;
+	return shape;
+}
+/**
+* Walk an eval's `manualInput` configuration and produce the wire-format
+* descriptor consumed by the web UI. The schema must resolve to a top-level
+* `z.object(...)`; nested objects, arrays, unions, and other unsupported
+* shapes inside fields fall back to the JSON textarea widget.
+*
+* Returns a `Result` so the caller (eval discovery) can surface a discovery
+* issue without throwing when the schema is incompatible.
+*/
+function buildManualInputDescriptor(config) {
+	const shape = getObjectShape(config.schema);
+	if (!shape) return Result.err(/* @__PURE__ */ new Error("manualInput.schema must be a top-level z.object(...). Wrap nested types in an object schema."));
+	const overrides = {};
+	const rawOverrides = config.fields;
+	if (rawOverrides) {
+		for (const [key, override] of Object.entries(rawOverrides)) if (override) overrides[key] = override;
+	}
+	const fields = [];
+	for (const [key, fieldSchema] of Object.entries(shape)) {
+		const fieldResult = buildField(key, fieldSchema, overrides[key]);
+		if (fieldResult.error) return fieldResult.errorResult();
+		fields.push(fieldResult.value);
+	}
+	return Result.ok({
+		title: config.title,
+		description: config.description,
+		submitLabel: config.submitLabel,
+		fields
+	});
+}
+/**
+* Resolve an eval's `manualInput` Zod schema against a raw user submission.
+* Returns the parsed value typed against the eval's `TInput` generic, or a
+* structured `Error` carrying the Zod issues for the caller to surface.
+*/
+function parseManualInputValues(config, raw) {
+	const parsed = config.schema.safeParse(raw);
+	if (!parsed.success) return Result.err(new ManualInputValidationError(parsed.error.issues.map(formatIssue)));
+	return Result.ok(parsed.data);
+}
+/**
+* Error thrown / returned when manual-input values fail validation against
+* the eval's `manualInput.schema`. Carries the structured Zod issues so the
+* CLI and HTTP layers can surface them per-field.
+*/
+var ManualInputValidationError = class extends Error {
+	issues;
+	constructor(issues) {
+		super(issues.length === 0 ? "manualInput validation failed" : `manualInput validation failed: ${issues.map((issue) => issue.path ? `${issue.path}: ${issue.message}` : issue.message).join("; ")}`);
+		this.name = "ManualInputValidationError";
+		this.issues = issues;
+	}
+};
+function formatIssue(issue) {
+	return {
+		path: issue.path.map((segment) => typeof segment === "string" || typeof segment === "number" ? String(segment) : "").filter((segment) => segment !== "").join("."),
+		message: issue.message
+	};
+}
+//#endregion
 //#region ../runner/src/outputArtifacts.ts
 const mimeTypeExtensionMap = {
 	"application/json": ".json",
@@ -5227,6 +5977,65 @@ function isFile(value) {
 	return value instanceof File;
 }
 //#endregion
+//#region ../runner/src/traceDisplay.ts
+function isRecord$1(value) {
+	return typeof value === "object" && value !== null;
+}
+function mergeNestedAttribute(value, path, attributeValue) {
+	const root = value === void 0 ? {} : { ...value };
+	const parts = path.split(".");
+	let current = root;
+	for (const [index, part] of parts.entries()) {
+		if (index === parts.length - 1) {
+			current[part] = attributeValue;
+			continue;
+		}
+		const nextValue = current[part];
+		const nextRecord = isRecord$1(nextValue) ? { ...nextValue } : {};
+		current[part] = nextRecord;
+		current = nextRecord;
+	}
+	return root;
+}
+function resolveTracePresentation(spans, globalTraceDisplay, evalTraceDisplay) {
+	const merged = /* @__PURE__ */ new Map();
+	for (const attribute of globalTraceDisplay?.attributes ?? []) merged.set(attribute.key ?? attribute.path, attribute);
+	for (const attribute of evalTraceDisplay?.attributes ?? []) merged.set(attribute.key ?? attribute.path, attribute);
+	const resolvedAttributes = [];
+	const transformedTrace = spans.map((span) => ({
+		...span,
+		attributes: span.attributes === void 0 ? void 0 : { ...span.attributes }
+	}));
+	for (const attribute of merged.values()) {
+		const resolvedPath = attribute.transform ? `__display.${attribute.key ?? attribute.path}` : attribute.path;
+		resolvedAttributes.push({
+			key: attribute.key,
+			path: resolvedPath,
+			label: attribute.label,
+			format: attribute.format,
+			numberFormat: attribute.numberFormat,
+			placements: attribute.placements,
+			scope: attribute.scope,
+			mode: attribute.mode
+		});
+		if (!attribute.transform) continue;
+		for (const span of transformedTrace) {
+			const sourceValue = getNestedAttribute(span.attributes, attribute.path);
+			if (sourceValue === void 0) continue;
+			const transformedValue = attribute.transform({
+				value: sourceValue,
+				span
+			});
+			if (transformedValue === void 0) continue;
+			span.attributes = mergeNestedAttribute(span.attributes, resolvedPath, transformedValue);
+		}
+	}
+	return {
+		trace: transformedTrace,
+		traceDisplay: { attributes: resolvedAttributes }
+	};
+}
+//#endregion
 //#region ../runner/src/runMaintenance.ts
 async function persistRunState(runState) {
 	await writeFile(join(runState.runDir, "summary.json"), JSON.stringify(runState.summary, null, 2));
@@ -5551,65 +6360,6 @@ function stripTerminalControlCodes(value) {
 	return stripVTControlCharacters(value).replaceAll(orphanedAnsiSgrPattern, "");
 }
 //#endregion
-//#region ../runner/src/traceDisplay.ts
-function isRecord$1(value) {
-	return typeof value === "object" && value !== null;
-}
-function mergeNestedAttribute(value, path, attributeValue) {
-	const root = value === void 0 ? {} : { ...value };
-	const parts = path.split(".");
-	let current = root;
-	for (const [index, part] of parts.entries()) {
-		if (index === parts.length - 1) {
-			current[part] = attributeValue;
-			continue;
-		}
-		const nextValue = current[part];
-		const nextRecord = isRecord$1(nextValue) ? { ...nextValue } : {};
-		current[part] = nextRecord;
-		current = nextRecord;
-	}
-	return root;
-}
-function resolveTracePresentation(spans, globalTraceDisplay, evalTraceDisplay) {
-	const merged = /* @__PURE__ */ new Map();
-	for (const attribute of globalTraceDisplay?.attributes ?? []) merged.set(attribute.key ?? attribute.path, attribute);
-	for (const attribute of evalTraceDisplay?.attributes ?? []) merged.set(attribute.key ?? attribute.path, attribute);
-	const resolvedAttributes = [];
-	const transformedTrace = spans.map((span) => ({
-		...span,
-		attributes: span.attributes === void 0 ? void 0 : { ...span.attributes }
-	}));
-	for (const attribute of merged.values()) {
-		const resolvedPath = attribute.transform ? `__display.${attribute.key ?? attribute.path}` : attribute.path;
-		resolvedAttributes.push({
-			key: attribute.key,
-			path: resolvedPath,
-			label: attribute.label,
-			format: attribute.format,
-			numberFormat: attribute.numberFormat,
-			placements: attribute.placements,
-			scope: attribute.scope,
-			mode: attribute.mode
-		});
-		if (!attribute.transform) continue;
-		for (const span of transformedTrace) {
-			const sourceValue = getNestedAttribute(span.attributes, attribute.path);
-			if (sourceValue === void 0) continue;
-			const transformedValue = attribute.transform({
-				value: sourceValue,
-				span
-			});
-			if (transformedValue === void 0) continue;
-			span.attributes = mergeNestedAttribute(span.attributes, resolvedPath, transformedValue);
-		}
-	}
-	return {
-		trace: transformedTrace,
-		traceDisplay: { attributes: resolvedAttributes }
-	};
-}
-//#endregion
 //#region ../runner/src/runExecution.ts
 function filterEvalCases(cases, caseIds) {
 	if (!caseIds || caseIds.length === 0) return cases;
@@ -5639,8 +6389,54 @@ function buildScopedEvalIdPrefix(params) {
 async function callWithUnknownResult(fn, args) {
 	return await Reflect.apply(fn, void 0, args);
 }
+async function callUnknownFunction(fn, args) {
+	if (typeof fn !== "function") throw new Error("Expected a function");
+	return await Reflect.apply(fn, void 0, args);
+}
+function assignDerivedOutputs(params) {
+	for (const [key, value] of Object.entries(params.derived)) {
+		if (key in params.outputs) continue;
+		params.outputs[key] = value;
+	}
+}
+async function resolveDeriveFromTracingConfig(params) {
+	const ctx = {
+		trace: params.traceTree,
+		input: params.evalCase.input,
+		case: params.evalCase
+	};
+	if (typeof params.deriveFromTracing === "function") {
+		const derived = await callUnknownFunction(params.deriveFromTracing, [ctx]);
+		if (!isRecord(derived)) throw new Error("deriveFromTracing must return an object");
+		return derived;
+	}
+	const derived = {};
+	for (const [key, compute] of Object.entries(params.deriveFromTracing)) {
+		const value = await callUnknownFunction(compute, [ctx]);
+		if (value !== void 0) derived[key] = value;
+	}
+	return derived;
+}
+async function runDeriveFromTracingConfig(params) {
+	if (params.deriveFromTracing === void 0) return;
+	const { deriveFromTracing } = params;
+	try {
+		const derived = await runInExistingEvalScope(params.scope, "derive", async () => await resolveDeriveFromTracingConfig({
+			deriveFromTracing,
+			traceTree: params.traceTree,
+			evalCase: params.evalCase
+		}));
+		assignDerivedOutputs({
+			outputs: params.scope.outputs,
+			derived
+		});
+	} catch (e) {
+		const message = `deriveFromTracing threw: ${e instanceof Error ? e.message : String(e)}`;
+		params.scope.assertionFailures.push(toAssertionFailure(message, e instanceof Error ? e : void 0));
+	}
+}
 async function runCase(params) {
-	const { evalDef, evalId, evalKey = evalId, evalCase, globalTraceDisplay, llmCallsConfig = resolveLlmCallsConfig(void 0), apiCallsConfig = resolveApiCallsConfig(void 0), globalRemoveDefaultConfig, trial, startTime, cacheAdapter, cacheMode, codeFingerprint, moduleIsolation, evalFilePath, evalFileRelativePath = evalFilePath, workspaceRoot, artifactDir, runId } = params;
+	const { evalDef, evalId, evalKey = evalId, evalCase, globalTraceDisplay, globalColumns, globalDeriveFromTracing, llmCallsConfig = resolveLlmCallsConfig(void 0), apiCallsConfig = resolveApiCallsConfig(void 0), globalRemoveDefaultConfig, trial, startTime, cacheAdapter, cacheMode, moduleIsolation, evalFilePath, evalFileRelativePath = evalFilePath, workspaceRoot, artifactDir, runId } = params;
 	const scopedIdPrefix = buildScopedEvalIdPrefix({
 		evalId,
 		evalFilePath,
@@ -5672,7 +6468,8 @@ async function runCase(params) {
 			adapter: cacheAdapter,
 			mode: cacheMode,
 			evalId,
-			codeFingerprint
+			read: evalDef.cache?.read,
+			store: evalDef.cache?.store
 		} : void 0,
 		startTime: evalDef.startTime,
 		freezeTime: evalDef.freezeTime
@@ -5685,22 +6482,19 @@ async function runCase(params) {
 	const traceTree = buildTraceTree(spansWithDerivedAttributes, scope.checkpoints);
 	const nonAssertError = executeError && !(executeError instanceof EvalAssertionError) ? executeError : null;
 	if (executeError instanceof EvalAssertionError && scope.assertionFailures.length === 0) scope.assertionFailures.push(toAssertionFailure(executeError.message, executeError));
-	if (!nonAssertError && evalDef.deriveFromTracing) {
-		const { deriveFromTracing } = evalDef;
-		try {
-			const derived = await runInExistingEvalScope(scope, "derive", async () => {
-				return await callWithUnknownResult(deriveFromTracing, [{
-					trace: traceTree,
-					input: evalCase.input,
-					case: evalCase
-				}]);
-			});
-			if (!isRecord(derived)) throw new Error("deriveFromTracing must return an object");
-			for (const [key, value] of Object.entries(derived)) if (!(key in scope.outputs)) scope.outputs[key] = value;
-		} catch (e) {
-			const message = `deriveFromTracing threw: ${e instanceof Error ? e.message : String(e)}`;
-			scope.assertionFailures.push(toAssertionFailure(message, e instanceof Error ? e : void 0));
-		}
+	if (!nonAssertError) {
+		await runDeriveFromTracingConfig({
+			deriveFromTracing: globalDeriveFromTracing,
+			scope,
+			traceTree,
+			evalCase
+		});
+		await runDeriveFromTracingConfig({
+			deriveFromTracing: evalDef.deriveFromTracing,
+			scope,
+			traceTree,
+			evalCase
+		});
 	}
 	if (!nonAssertError) addDefaultOutputs({
 		outputs: scope.outputs,
@@ -5717,7 +6511,7 @@ async function runCase(params) {
 			...scope.outputs,
 			...parsedOutputs.data
 		};
-		else scope.assertionFailures.push(toAssertionFailure(formatOutputsSchemaError(parsedOutputs.error)));
+		else scope.assertionFailures.push(toAssertionFailure(formatOutputsSchemaError(parsedOutputs.error), void 0, "OutputsSchemaError"));
 	}
 	const scoreResults = /* @__PURE__ */ new Map();
 	const scoringTraces = {};
@@ -5740,7 +6534,8 @@ async function runCase(params) {
 				adapter: cacheAdapter,
 				mode: cacheMode,
 				evalId: `${evalId}__score__${key}`,
-				codeFingerprint
+				read: evalDef.cache?.read,
+				store: evalDef.cache?.store
 			} : void 0,
 			startTime: scoreStartTime,
 			freezeTime: evalDef.freezeTime
@@ -5795,6 +6590,7 @@ async function runCase(params) {
 	const { trace: displayTrace, traceDisplay } = resolveTracePresentation(spansWithDerivedAttributes, globalTraceDisplay, evalDef.traceDisplay);
 	const columns = {};
 	const columnOverrides = mergeDefaultColumns({
+		globalColumns,
 		columns: evalDef.columns,
 		globalRemove: globalRemoveDefaultConfig,
 		evalRemove: evalDef.removeDefaultConfig
@@ -5858,14 +6654,17 @@ function formatOutputsSchemaError(error) {
 	const issueLines = error.issues.map((issue) => {
 		return `${issue.path.length > 0 ? issue.path.join(".") : "<root>"}: ${issue.message}`;
 	});
-	if (issueLines.length === 0) return "outputsSchema validation failed";
-	return `outputsSchema validation failed:\n${issueLines.join("\n")}`;
+	if (issueLines.length === 0) return "outputs did not match the configured schema";
+	return issueLines.join("\n");
 }
-function toAssertionFailure(message, error = void 0) {
-	return error?.stack ? {
+function toAssertionFailure(message, error = void 0, nameOverride = void 0) {
+	const name = nameOverride ?? error?.name;
+	const stack = error?.stack ? stripTerminalControlCodes(error.stack) : void 0;
+	return {
+		...name !== void 0 ? { name } : {},
 		message,
-		stack: stripTerminalControlCodes(error.stack)
-	} : { message };
+		...stack !== void 0 ? { stack } : {}
+	};
 }
 //#endregion
 //#region ../runner/src/runQueue.ts
@@ -6095,15 +6894,15 @@ async function executeRun({ runState, request, runDir, config, cacheStore, lastR
 		const apiCallsConfig = resolveApiCallsConfig(config.apiCalls);
 		for (const evalMeta of targetEvals) {
 			const evalFilePath = evalMeta.sourceFilePath;
-			let codeFingerprint = "";
+			let sourceFingerprint = "";
 			try {
-				codeFingerprint = getSourceFingerprint(await readFile(evalFilePath, "utf-8"));
+				sourceFingerprint = getSourceFingerprint(await readFile(evalFilePath, "utf-8"));
 			} catch {
-				codeFingerprint = "";
+				sourceFingerprint = "";
 			}
-			if (codeFingerprint.length > 0) {
-				runState.manifest.evalSourceFingerprints[evalMeta.key] = codeFingerprint;
-				evalMeta.sourceFingerprint = codeFingerprint;
+			if (sourceFingerprint.length > 0) {
+				runState.manifest.evalSourceFingerprints[evalMeta.key] = sourceFingerprint;
+				evalMeta.sourceFingerprint = sourceFingerprint;
 			} else {
 				delete runState.manifest.evalSourceFingerprints[evalMeta.key];
 				evalMeta.sourceFingerprint = null;
@@ -6112,7 +6911,7 @@ async function executeRun({ runState, request, runDir, config, cacheStore, lastR
 				const registry = getEvalRegistry();
 				await runWithModuleIsolation(moduleIsolation, async () => {
 					await runInEvalRuntimeScope("env", async () => {
-						await loadEvalModule(evalFilePath, codeFingerprint);
+						await loadEvalModule(evalFilePath, sourceFingerprint);
 					});
 				});
 				const entry = registry.get(evalMeta.id);
@@ -6126,8 +6925,24 @@ async function executeRun({ runState, request, runDir, config, cacheStore, lastR
 				await runWithModuleIsolation(moduleIsolation, async () => {
 					await runInEvalRuntimeScope("cases", async () => {
 						await entry.use(async (evalDef) => {
-							const runnableCases = resolveRunnableEvalCases({
-								cases: await runWithEvalClock(evalDef.startTime, async () => typeof evalDef.cases === "function" ? await evalDef.cases() : evalDef.cases ?? [], { freezeTime: evalDef.freezeTime }),
+							if (evalDef.manualInput && evalDef.cases !== void 0) throw new Error(`Eval "${evalMeta.id}" cannot declare both "cases" and "manualInput". Remove one of them.`);
+							let manualInputCase = null;
+							if (evalDef.manualInput) {
+								const rawValue = request.manualInputs?.[evalMeta.key];
+								if (rawValue === void 0) throw new Error(`Eval "${evalMeta.id}" requires manual input. Provide it via the run modal in the web UI or "--input" / "--input-file" on the CLI.`);
+								const parsed = parseManualInputValues(evalDef.manualInput, rawValue);
+								if (parsed.error) {
+									const formatted = parsed.error.issues.map((issue) => issue.path ? `${issue.path}: ${issue.message}` : issue.message).join("; ");
+									throw new Error(`Invalid manual input for eval "${evalMeta.id}": ${formatted}`);
+								}
+								manualInputCase = {
+									id: `${evalMeta.id}-manual`,
+									input: parsed.value
+								};
+							}
+							const evalCases = manualInputCase ? [manualInputCase] : await runWithEvalClock(evalDef.startTime, async () => typeof evalDef.cases === "function" ? await evalDef.cases() : evalDef.cases ?? [], { freezeTime: evalDef.freezeTime });
+							const runnableCases = manualInputCase ? evalCases : resolveRunnableEvalCases({
+								cases: evalCases,
 								evalId: evalMeta.id
 							});
 							const duplicateCaseIds = findDuplicateCaseIds(runnableCases);
@@ -6136,6 +6951,8 @@ async function executeRun({ runState, request, runDir, config, cacheStore, lastR
 							runState.summary.totalCases += cases.length;
 							const defaultConfig = resolveEvalDefaultConfig({
 								evalDef,
+								globalColumns: config.columns,
+								globalStats: config.stats,
 								globalRemove: config.removeDefaultConfig
 							});
 							const declaredColumnDefs = buildDeclaredColumnDefs(defaultConfig.columns, evalDef.scores, evalDef.manualScores);
@@ -6181,6 +6998,8 @@ async function executeRun({ runState, request, runDir, config, cacheStore, lastR
 												evalKey: evalMeta.key,
 												evalCase,
 												globalTraceDisplay,
+												globalColumns: config.columns,
+												globalDeriveFromTracing: config.deriveFromTracing,
 												llmCallsConfig,
 												apiCallsConfig,
 												globalRemoveDefaultConfig: config.removeDefaultConfig,
@@ -6188,7 +7007,6 @@ async function executeRun({ runState, request, runDir, config, cacheStore, lastR
 												startTime,
 												cacheAdapter: bufferedCacheStore ?? (cacheEnabled ? cacheStore : null),
 												cacheMode,
-												codeFingerprint,
 												moduleIsolation,
 												evalFilePath,
 												evalFileRelativePath: evalMeta.filePath,
@@ -6343,4 +7161,4 @@ function toLastRunStatus(status) {
 	return status === "pending" ? null : status;
 }
 //#endregion
-export { llmCallsConfigSchema as $, traceSpanKindSchema as $t, extractApiCalls as A, getCurrentScope as An, evalChartTypeSchema as At, runSummarySchema as B, setEvalOutput as Bn, cacheRecordingOpSchema as Bt, validateCharts as C, evalExpect as Cn, evalChartAggregateSchema as Ct, sseEnvelopeSchema as D, configureEvalRunLogs as Dn, evalChartConfigSchema as Dt, updateManualScoreRequestSchema as E, appendToEvalOutput as En, evalChartColorSchema as Et, getEvalDisplayStatus as F, mergeEvalOutput as Fn, cacheEntryWithDebugKeySchema as Ft, apiCallMetricPlacementSchema as G, traceCacheRefSchema as Gt, DEFAULT_LLM_CALLS_CONFIG as H, startEvalBackgroundJob as Hn, cacheStatusSchema as Ht, deriveScopedSummaryFromCases as I, nextEvalId as In, cacheFileSchema as It, defaultConfigKeySchema as J, traceAttributeDisplayPlacementSchema as Jt, apiCallMetricSchema as K, traceAttributeDisplayFormatSchema as Kt, deriveStatusFromCaseRows as L, runInEvalRuntimeScope as Ln, cacheListItemSchema as Lt, applyDerivedCallAttributes as M, getEvalStartTime as Mn, cacheDebugKeyEntrySchema as Mt, getNestedAttribute as N, incrementEvalOutput as Nn, cacheDebugKeyFileSchema as Nt, extractCacheEntries as O, evalAssert as On, evalChartMetricSchema as Ot, getEvalTitle as P, isInEvalScope as Pn, cacheEntrySchema as Pt, llmCallPricingSchema as Q, traceSpanErrorSchema as Qt, deriveStatusFromChildStatuses as R, runInEvalScope as Rn, cacheModeSchema as Rt, normalizeScoreDef as S, repoFile as Sn, scoreTraceSchema as St, createRunRequestSchema as T, advanceEvalTime as Tn, evalChartBuiltinMetricSchema as Tt, agentEvalsConfigSchema as U, defineEval as Un, serializedCacheSpanSchema as Ut, DEFAULT_API_CALLS_CONFIG as V, setScopeCacheContext as Vn, cacheRecordingSchema as Vt, apiCallMetricFormatSchema as W, getEvalRegistry as Wn, spanCacheOptionsSchema as Wt, llmCallMetricPlacementSchema as X, traceDisplayConfigSchema as Xt, llmCallMetricFormatSchema as Y, traceAttributeDisplaySchema as Yt, llmCallMetricSchema as Z, traceDisplayInputConfigSchema as Zt, loadEvalModule as _, hashCacheKeySync as _n, evalSummarySchema as _t, getLastRunStatuses as a, columnKindSchema as an, buildCaseKey as at, loadConfig as b, serializeCacheRecording as bn, runLogLocationSchema as bt, loadPersistedRunSnapshots as c, numberDisplayOptionsSchema as cn, getCaseRowEvalKey as ct, persistRunState as d, z$1 as dn, caseRowSchema as dt, traceSpanSchema as en, removeDefaultConfigSchema as et, recomputeEvalStatusesInRuns as f, buildTraceTree as fn, discoveryIssueSchema as ft, deriveEvalFreshness as g, hashCacheKey as gn, evalStatsConfigSchema as gt, resolveArtifactPath as h, evalTracer as hn, evalStatItemSchema as ht, generateRunId as i, columnFormatSchema as in, trialSelectionModeSchema as it, extractLlmCalls as j, getEvalCaseInput as jn, evalChartsConfigSchema as jt, extractCacheHits as k, evalLog as kn, evalChartTooltipExtraSchema as kt, nextShortIdFromSnapshots as l, repoFileRefSchema as ln, assertionFailureSchema as lt, runTouchesEval as m, evalSpan as mn, evalStatAggregateSchema as mt, getTargetEvalKeys as n, cellValueSchema as nn, resolveLlmCallsConfig as nt, getLatestRunInfos as o, fileRefSchema as on, buildEvalKey as ot, recomputePersistedCaseStatus as p, captureEvalSpanError as pn, evalFreshnessStatusSchema as pt, apiCallsConfigSchema as q, traceAttributeDisplayInputSchema as qt, getTargetEvals as r, columnDefSchema as rn, runLogsConfigSchema as rt, loadPersistedRunSnapshot as s, jsonCellSchema as sn, getCaseRowCaseKey as st, executeRun as t, traceSpanWarningSchema as tn, resolveApiCallsConfig as tt, persistCaseDetail as u, runArtifactRefSchema as un, caseDetailSchema as ut, parseEvalDiscovery as v, deserializeCacheRecording as vn, runLogEntrySchema as vt, createFsCacheStore as w, EvalAssertionError as wn, evalChartAxisSchema as wt, buildDeclaredColumnDefs as x, serializeCacheValue as xn, runLogPhaseSchema as xt, resolveEvalDefaultConfig as y, deserializeCacheValue as yn, runLogLevelSchema as yt, runManifestSchema as z, runInExistingEvalScope as zn, cacheOperationTypeSchema as zt };
+export { defaultConfigKeySchema as $, incrementEvalOutput as $n, cacheEntryWithDebugKeySchema as $t, createRunRequestSchema as A, buildTraceTree as An, runLogPhaseSchema as At, getEvalDisplayStatus as B, repoFile as Bn, manualInputTextFieldSchema as Bt, loadConfig as C, columnKindSchema as Cn, evalStatAggregateSchema as Ct, createFsCacheStore as D, repoFileRefSchema as Dn, runLogEntrySchema as Dt, validateCharts as E, numberDisplayOptionsSchema as En, evalSummarySchema as Et, extractApiCalls as F, hashCacheKeySync as Fn, manualInputJsonFieldSchema as Ft, runSummarySchema as G, advanceEvalTime as Gn, evalChartConfigSchema as Gt, deriveStatusFromCaseRows as H, readManualInputFile as Hn, evalChartAxisSchema as Ht, extractLlmCalls as I, deserializeCacheRecording as In, manualInputMultilineFieldSchema as It, agentEvalsConfigSchema as J, evalAssert as Jn, evalChartTypeSchema as Jt, DEFAULT_API_CALLS_CONFIG as K, appendToEvalOutput as Kn, evalChartMetricSchema as Kt, applyDerivedCallAttributes as L, deserializeCacheValue as Ln, manualInputNumberFieldSchema as Lt, sseEnvelopeSchema as M, evalSpan as Mn, manualInputBooleanFieldSchema as Mt, extractCacheEntries as N, evalTracer as Nn, manualInputDescriptorSchema as Nt, configReloadStateSchema as O, runArtifactRefSchema as On, runLogLevelSchema as Ot, extractCacheHits as P, hashCacheKey as Pn, manualInputFieldDescriptorSchema as Pt, apiCallsConfigSchema as Q, getEvalStartTime as Qn, cacheEntrySchema as Qt, getNestedAttribute as R, serializeCacheRecording as Rn, manualInputSelectFieldSchema as Rt, resolveEvalDefaultConfig as S, columnFormatSchema as Sn, evalFreshnessStatusSchema as St, normalizeScoreDef as T, jsonCellSchema as Tn, evalStatsConfigSchema as Tt, deriveStatusFromChildStatuses as U, evalExpect as Un, evalChartBuiltinMetricSchema as Ut, deriveScopedSummaryFromCases as V, manualInputFileValueSchema as Vn, evalChartAggregateSchema as Vt, runManifestSchema as W, EvalAssertionError as Wn, evalChartColorSchema as Wt, apiCallMetricPlacementSchema as X, getCurrentScope as Xn, cacheDebugKeyEntrySchema as Xt, apiCallMetricFormatSchema as Y, evalLog as Yn, evalChartsConfigSchema as Yt, apiCallMetricSchema as Z, getEvalCaseInput as Zn, cacheDebugKeyFileSchema as Zt, buildManualInputDescriptor as _, traceSpanKindSchema as _n, getCaseRowEvalKey as _t, getLastRunStatuses as a, cacheRecordingSchema as an, runInExistingEvalScope as ar, llmCallMetricSchema as at, loadEvalModule as b, cellValueSchema as bn, caseRowSchema as bt, loadPersistedRunSnapshots as c, spanCacheOptionsSchema as cn, startEvalBackgroundJob as cr, llmCallsConfigSchema as ct, persistRunState as d, traceAttributeDisplayInputSchema as dn, resolveLlmCallsConfig as dt, cacheFileSchema as en, isInEvalScope as er, evalColumnOverrideSchema as et, recomputeEvalStatusesInRuns as f, traceAttributeDisplayPlacementSchema as fn, runLogsConfigSchema as ft, resolveArtifactPath as g, traceSpanErrorSchema as gn, getCaseRowCaseKey as gt, resolveTracePresentation as h, traceDisplayInputConfigSchema as hn, buildEvalKey as ht, generateRunId as i, cacheRecordingOpSchema as in, runInEvalScope as ir, llmCallMetricPlacementSchema as it, updateManualScoreRequestSchema as j, captureEvalSpanError as jn, scoreTraceSchema as jt, configReloadStatusSchema as k, z$1 as kn, runLogLocationSchema as kt, nextShortIdFromSnapshots as l, traceCacheRefSchema as ln, defineEval as lr, removeDefaultConfigSchema as lt, runTouchesEval as m, traceDisplayConfigSchema as mn, buildCaseKey as mt, getTargetEvalKeys as n, cacheModeSchema as nn, nextEvalId as nr, evalDeriveConfigSchema as nt, getLatestRunInfos as o, cacheStatusSchema as on, setEvalOutput as or, llmCallPricingRateSchema as ot, recomputePersistedCaseStatus as p, traceAttributeDisplaySchema as pn, trialSelectionModeSchema as pt, DEFAULT_LLM_CALLS_CONFIG as q, configureEvalRunLogs as qn, evalChartTooltipExtraSchema as qt, getTargetEvals as r, cacheOperationTypeSchema as rn, runInEvalRuntimeScope as rr, llmCallMetricFormatSchema as rt, loadPersistedRunSnapshot as s, serializedCacheSpanSchema as sn, setScopeCacheContext as sr, llmCallPricingSchema as st, executeRun as t, cacheListItemSchema as tn, mergeEvalOutput as tr, evalColumnsSchema as tt, persistCaseDetail as u, traceAttributeDisplayFormatSchema as un, getEvalRegistry as ur, resolveApiCallsConfig as ut, parseManualInputValues as v, traceSpanSchema as vn, assertionFailureSchema as vt, buildDeclaredColumnDefs as w, fileRefSchema as wn, evalStatItemSchema as wt, parseEvalDiscovery as x, columnDefSchema as xn, discoveryIssueSchema as xt, deriveEvalFreshness as y, traceSpanWarningSchema as yn, caseDetailSchema as yt, getEvalTitle as z, serializeCacheValue as zn, manualInputSelectOptionSchema as zt };