@ls-stack/agent-eval 0.20.0 → 0.21.0

package/dist/{runOrchestration-Cv1kiOAG.mjs → runOrchestration-zYAcAPtS.mjs}

@@ -1970,7 +1970,15 @@ const numberDisplayOptionsSchema = z.object({
 	compactDisplay: z.enum(["short", "long"]).optional(),
 	prefix: z.string().optional(),
 	suffix: z.string().optional(),
-	decimalPlaces: z.number().int().min(0).optional()
+	minDecimalPlaces: z.number().int().min(0).optional(),
+	maxDecimalPlaces: z.number().int().min(0).optional()
+}).refine((options) => {
+	if (options.minDecimalPlaces === void 0) return true;
+	if (options.maxDecimalPlaces === void 0) return true;
+	return options.minDecimalPlaces <= options.maxDecimalPlaces;
+}, {
+	message: "minDecimalPlaces must be less than or equal to maxDecimalPlaces",
+	path: ["minDecimalPlaces"]
 });
 /** Schema for the supported column rendering kinds in list views. */
 const columnKindSchema = z.enum([
@@ -2005,7 +2013,6 @@ const columnDefSchema = z.object({
 	passThreshold: z.number().optional(),
 	maxStars: z.number().int().min(2).optional(),
 	hideInTable: z.boolean().optional(),
-	sortable: z.boolean().optional(),
 	align: z.enum([
 		"left",
 		"center",
@@ -2403,6 +2410,8 @@ const evalStatItemSchema = z.discriminatedUnion("kind", [
 		label: z.string().optional(),
 		aggregate: evalStatAggregateSchema,
 		format: columnFormatSchema.optional(),
+		/** Number presentation options applied when `format: 'number'`. */
+		numberFormat: numberDisplayOptionsSchema.optional(),
 		accent: z.boolean().optional()
 	})
 ]);
@@ -4035,7 +4044,6 @@ function getScoreOverride(def) {
 		format: def.format,
 		numberFormat: def.numberFormat,
 		hideInTable: def.hideInTable,
-		sortable: def.sortable,
 		align: def.align,
 		maxStars: def.maxStars
 	};
@@ -4048,7 +4056,6 @@ function mergeOverrides(base, override) {
 		format: override.format ?? base.format,
 		numberFormat: override.numberFormat ?? base.numberFormat,
 		hideInTable: override.hideInTable ?? base.hideInTable,
-		sortable: override.sortable ?? base.sortable,
 		align: override.align ?? base.align,
 		maxStars: override.maxStars ?? base.maxStars
 	};
@@ -4163,7 +4170,6 @@ 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?.sortable !== void 0) def.sortable = override.sortable;
 	if (override?.align !== void 0) def.align = override.align;
 	if (!isScore) return def;
 	def.isScore = true;
@@ -4231,86 +4237,76 @@ const DEFAULT_CONFIG_KEYS = [
 	"totalTokens",
 	"cachedInputTokens",
 	"cacheCreationInputTokens",
-	"reasoningTokens",
 	"llmLatencyMs"
 ];
-const tokenNumberFormat = {
-	notation: "compact",
-	decimalPlaces: 1
+const tokenNumberFormat = { notation: "compact" };
+const countNumberFormat = {
+	minDecimalPlaces: 0,
+	maxDecimalPlaces: 0
+};
+const costNumberFormat = {
+	prefix: "$",
+	maxDecimalPlaces: 4
 };
-const countNumberFormat = { decimalPlaces: 0 };
 const DEFAULT_COLUMNS = {
 	apiCalls: {
 		label: "API Calls",
 		format: "number",
 		numberFormat: countNumberFormat,
-		align: "right",
-		sortable: true
+		align: "right"
 	},
 	costUsd: {
 		label: "Cost",
 		format: "number",
-		numberFormat: {
-			prefix: "$",
-			decimalPlaces: 4
-		},
-		align: "right",
-		sortable: true
+		numberFormat: costNumberFormat,
+		align: "right"
 	},
 	llmTurns: {
 		label: "LLM Turns",
 		format: "number",
 		numberFormat: countNumberFormat,
-		align: "right",
-		sortable: true
+		align: "right"
 	},
 	inputTokens: {
 		label: "Input Tokens",
 		format: "number",
 		numberFormat: tokenNumberFormat,
-		align: "right",
-		sortable: true
+		align: "right"
 	},
 	outputTokens: {
 		label: "Output Tokens",
 		format: "number",
 		numberFormat: tokenNumberFormat,
-		align: "right",
-		sortable: true
+		align: "right"
 	},
 	totalTokens: {
 		label: "Total Tokens",
 		format: "number",
 		numberFormat: tokenNumberFormat,
-		align: "right",
-		sortable: true
+		align: "right"
 	},
 	cachedInputTokens: {
 		label: "Cached Input Tokens",
 		format: "number",
 		numberFormat: tokenNumberFormat,
-		align: "right",
-		sortable: true
+		align: "right"
 	},
 	cacheCreationInputTokens: {
 		label: "Cache Write Tokens",
 		format: "number",
 		numberFormat: tokenNumberFormat,
-		align: "right",
-		sortable: true
+		align: "right"
 	},
 	reasoningTokens: {
 		label: "Reasoning Tokens",
 		format: "number",
 		numberFormat: tokenNumberFormat,
-		align: "right",
-		sortable: true
+		align: "right"
 	},
 	llmLatencyMs: {
 		label: "LLM Latency",
 		format: "duration",
-		align: "right",
-		sortable: true
+		align: "right"
 	}
 };
 function resolveRemovedKeys(globalRemove, evalRemove) {
@@ -4336,31 +4332,29 @@ function appendDefaultStats(params) {
 		kind: "column",
 		key: "apiCalls",
 		label: "API Calls",
-		aggregate: "avg"
+		aggregate: "avg",
+		numberFormat: countNumberFormat
 	});
 	if (activeKeys.has("costUsd")) defaults.push({
 		kind: "column",
 		key: "costUsd",
 		label: "LLM Cost",
-		aggregate: "sum"
+		aggregate: "avg",
+		numberFormat: costNumberFormat
 	});
 	if (activeKeys.has("totalTokens")) defaults.push({
 		kind: "column",
 		key: "totalTokens",
 		label: "Tokens",
-		aggregate: "sum"
+		aggregate: "avg",
+		numberFormat: tokenNumberFormat
 	});
 	if (activeKeys.has("llmTurns")) defaults.push({
 		kind: "column",
 		key: "llmTurns",
 		label: "LLM Turns",
-		aggregate: "avg"
-	});
-	if (activeKeys.has("llmLatencyMs")) defaults.push({
-		kind: "column",
-		key: "llmLatencyMs",
-		label: "LLM Latency",
-		aggregate: "avg"
+		aggregate: "avg",
+		numberFormat: countNumberFormat
 	});
 	const merged = [...params.stats ?? [], ...defaults];
 	return merged.length > 0 ? merged : void 0;
@@ -4368,24 +4362,13 @@ function appendDefaultStats(params) {
 function appendDefaultCharts(params) {
 	const activeKeys = new Set(getActiveDefaultConfigKeys(params));
 	const defaults = [];
-	if (activeKeys.has("apiCalls")) defaults.push({
-		heading: "API Calls",
-		type: "bar",
-		metrics: [{
-			source: "column",
-			key: "apiCalls",
-			aggregate: "sum",
-			label: "API Calls",
-			color: "accentDim"
-		}]
-	});
 	if (activeKeys.has("costUsd")) defaults.push({
 		heading: "LLM Cost",
 		type: "area",
 		metrics: [{
 			source: "column",
 			key: "costUsd",
-			aggregate: "sum",
+			aggregate: "avg",
 			label: "Cost",
 			color: "warning"
 		}]
@@ -4394,23 +4377,30 @@ function appendDefaultCharts(params) {
 		activeKeys.has("inputTokens") ? {
 			source: "column",
 			key: "inputTokens",
-			aggregate: "sum",
+			aggregate: "avg",
 			label: "Input",
 			color: "accent"
 		} : null,
 		activeKeys.has("outputTokens") ? {
 			source: "column",
 			key: "outputTokens",
-			aggregate: "sum",
+			aggregate: "avg",
 			label: "Output",
 			color: "success"
 		} : null,
-		activeKeys.has("reasoningTokens") ? {
+		activeKeys.has("cachedInputTokens") ? {
 			source: "column",
-			key: "reasoningTokens",
-			aggregate: "sum",
-			label: "Reasoning",
+			key: "cachedInputTokens",
+			aggregate: "avg",
+			label: "Cached Input",
 			color: "error"
+		} : null,
+		activeKeys.has("cacheCreationInputTokens") ? {
+			source: "column",
+			key: "cacheCreationInputTokens",
+			aggregate: "avg",
+			label: "Cache Write",
+			color: "warning"
 		} : null
 	].filter((metric) => metric !== null);
 	if (tokenMetrics.length > 0) defaults.push({
@@ -4420,7 +4410,7 @@ function appendDefaultCharts(params) {
 		tooltipExtras: activeKeys.has("totalTokens") ? [{
 			source: "column",
 			key: "totalTokens",
-			aggregate: "sum",
+			aggregate: "avg",
 			label: "Total"
 		}] : void 0
 	});
@@ -4517,12 +4507,6 @@ function addDefaultOutputs(params) {
 		value: sumNullable(calls.map((call) => call.cacheCreationInputTokens)),
 		activeKeys
 	});
-	assignIfMissing({
-		outputs: params.outputs,
-		key: "reasoningTokens",
-		value: sumNullable(calls.map((call) => call.reasoningTokens)),
-		activeKeys
-	});
 	assignIfMissing({
 		outputs: params.outputs,
 		key: "llmLatencyMs",

package/dist/{runner-B25oRQxX.mjs → runner-BcwyX9CO.mjs}

@@ -1,2 +1,2 @@
-import { n as initRunner, t as getRunnerInstance } from "./runner-DzrMtgBu.mjs";
+import { n as initRunner, t as getRunnerInstance } from "./runner-Dy_PECaf.mjs";
 export { getRunnerInstance, initRunner };

package/dist/{runner-DzrMtgBu.mjs → runner-Dy_PECaf.mjs}

@@ -1,5 +1,5 @@
-import { n as createRunner } from "./cli-weogme5U.mjs";
-import "./src-B879LZfo.mjs";
+import { n as createRunner } from "./cli-DumvanQI.mjs";
+import "./src-BoAJb4wC.mjs";
 //#region ../../apps/server/src/runner.ts
 let runnerInstance = null;
 function getRunnerInstance() {

package/dist/src-BoAJb4wC.mjs

@@ -0,0 +1,3 @@
+import "./runOrchestration-zYAcAPtS.mjs";
+import "./cli-DumvanQI.mjs";
+export {};

package/package.json

@@ -1,6 +1,6 @@
 {
   "name": "@ls-stack/agent-eval",
-  "version": "0.20.0",
+  "version": "0.21.0",
   "type": "module",
   "bin": {
     "agent-evals": "./dist/bin.mjs"

package/skills/agent-eval/SKILL.md

@@ -287,9 +287,13 @@ See `EvalScoreDef` / `EvalManualScoreDef` in the types for the full shape
   are still captured.
 
 Stats rows and history charts on the eval card can be authored via `stats` /
-`charts` on the eval definition. Usage stats/charts are added by default
-unless removed with `removeDefaultConfig`. Their shapes live in the types; no
-need to memorize the option set.
+`charts` on the eval definition. Usage stats and LLM usage charts are added by
+default unless removed with `removeDefaultConfig`. Column stats can override
+`format` and `numberFormat`, otherwise they inherit from the matching column.
+Number formats use `maxDecimalPlaces` to cap decimals and `minDecimalPlaces`
+to pad trailing zeroes. Without `maxDecimalPlaces`, they render up to 3 decimal
+places.
+Their shapes live in the types; no need to memorize the option set.
 
 ## Cached operations