@m4trix/evals 0.24.0 → 0.25.1

package/dist/cli.js

@@ -1,6 +1,6 @@
 #!/usr/bin/env node
 import { withFullScreen, useScreenSize } from 'fullscreen-ink';
-import React2, { useState, useRef, useReducer, useEffect, useMemo } from 'react';
+import React, { useState, useRef, useReducer, useEffect, useMemo } from 'react';
 import { useApp, useInput, Box, Text } from 'ink';
 import { jsx, jsxs, Fragment } from 'react/jsx-runtime';
 import { resolve, relative, join, dirname } from 'path';
@@ -90,11 +90,7 @@ function getFooterText(state) {
   }
   return "\u2191\u2193 move  Enter add/remove  S start run  / search  Esc cancel  q quit";
 }
-function ListItem({
-  selected,
-  label,
-  itemKey
-}) {
+function ListItem({ selected, label, itemKey }) {
   return /* @__PURE__ */ jsxs(Text, { color: selected ? "cyan" : "gray", bold: selected, children: [
     selected ? "\u25B8 " : "  ",
     label
@@ -121,9 +117,7 @@ function Pane({
     }
   );
 }
-function SectionHeader({
-  children
-}) {
+function SectionHeader({ children }) {
   return /* @__PURE__ */ jsx(Text, { color: "cyan", bold: true, children });
 }
 function StatusText({ status }) {
@@ -135,10 +129,7 @@ function StatusText({ status }) {
   ] });
 }
 var LEFT_PANE_WIDTH = 44;
-function RunsSidebar({
-  state,
-  runs
-}) {
+function RunsSidebar({ state, runs }) {
   const focused = state.focus === "left";
   return /* @__PURE__ */ jsxs(Pane, { width: LEFT_PANE_WIDTH, focused, children: [
     /* @__PURE__ */ jsx(SectionHeader, { children: "Runs" }),
@@ -167,11 +158,7 @@ function RunsSidebar({
   ] });
 }
 var BLOCKS = ["\u2581", "\u2582", "\u2583", "\u2584", "\u2585", "\u2586", "\u2587", "\u2588"];
-function Sparkline({
-  data,
-  width,
-  label
-}) {
+function Sparkline({ data, width, label }) {
   if (data.length === 0)
     return null;
   const max = Math.max(...data);
@@ -401,9 +388,7 @@ var data_mock_default = {
             { name: "contract_match", score: 100 },
             { name: "arg_validity", score: 100 }
           ],
-          checks: [
-            { name: "tool_calls", passed: true, detail: "0 unexpected" }
-          ],
+          checks: [{ name: "tool_calls", passed: true, detail: "0 unexpected" }],
           failures: [],
           meta: {
             model: "gpt-4o-mini",
@@ -426,9 +411,21 @@ var data_mock_default = {
     }
   ],
   evaluators: [
-    { id: "json-schema-validator", name: "JSON Schema Validator", configPreview: "strict=true" },
-    { id: "tool-call-contract-checker", name: "Tool-call Contract Checker", configPreview: "unexpectedCalls=error" },
-    { id: "rubric-judge", name: "Rubric Judge (LLM)", configPreview: "model=gpt-4o-mini; scale=0-100" },
+    {
+      id: "json-schema-validator",
+      name: "JSON Schema Validator",
+      configPreview: "strict=true"
+    },
+    {
+      id: "tool-call-contract-checker",
+      name: "Tool-call Contract Checker",
+      configPreview: "unexpectedCalls=error"
+    },
+    {
+      id: "rubric-judge",
+      name: "Rubric Judge (LLM)",
+      configPreview: "model=gpt-4o-mini; scale=0-100"
+    },
     { id: "pii-leak-detector", name: "PII Leak Detector", configPreview: "redact=false" }
   ]
 };
@@ -508,9 +505,7 @@ async function loadRunnerData(runner) {
   const memSnapshots = runner.getAllRunSnapshots();
   const seen = new Set(memSnapshots.map((s) => s.runId));
   const fromDisk = diskSnapshots.filter((s) => !seen.has(s.runId));
-  const snapshots = [...memSnapshots, ...fromDisk].sort(
-    (a, b) => b.queuedAt - a.queuedAt
-  );
+  const snapshots = [...memSnapshots, ...fromDisk].sort((a, b) => b.queuedAt - a.queuedAt);
   if (datasets.length === 0 && evaluators.length === 0) {
     return loadMockData();
   }
@@ -632,7 +627,11 @@ function reduceCliState(state, action) {
       return { ...state, overviewScrollOffset: Math.max(0, state.overviewScrollOffset - 1) };
     }
     if (state.level === "datasets") {
-      return { ...state, datasetMenuIndex: Math.max(0, state.datasetMenuIndex - 1), overviewScrollOffset: 0 };
+      return {
+        ...state,
+        datasetMenuIndex: Math.max(0, state.datasetMenuIndex - 1),
+        overviewScrollOffset: 0
+      };
     }
     if (state.level === "runs") {
       return { ...state, runMenuIndex: Math.max(0, state.runMenuIndex - 1) };
@@ -650,10 +649,17 @@ function reduceCliState(state, action) {
       return { ...state, detailsScrollOffset: Math.min(action.max, state.detailsScrollOffset + 1) };
     }
     if (state.level === "datasets" && state.focus === "right") {
-      return { ...state, overviewScrollOffset: Math.min(action.max, state.overviewScrollOffset + 1) };
+      return {
+        ...state,
+        overviewScrollOffset: Math.min(action.max, state.overviewScrollOffset + 1)
+      };
     }
     if (state.level === "datasets") {
-      return { ...state, datasetMenuIndex: Math.min(action.max, state.datasetMenuIndex + 1), overviewScrollOffset: 0 };
+      return {
+        ...state,
+        datasetMenuIndex: Math.min(action.max, state.datasetMenuIndex + 1),
+        overviewScrollOffset: 0
+      };
     }
     if (state.level === "runs") {
       return { ...state, runMenuIndex: Math.min(action.max, state.runMenuIndex + 1) };
@@ -735,18 +741,8 @@ var defaultRunnerConfig = {
   discovery: {
     rootDir: process.cwd(),
     datasetSuffixes: [".dataset.ts", ".dataset.tsx", ".dataset.js", ".dataset.mjs"],
-    evaluatorSuffixes: [
-      ".evaluator.ts",
-      ".evaluator.tsx",
-      ".evaluator.js",
-      ".evaluator.mjs"
-    ],
-    testCaseSuffixes: [
-      ".test-case.ts",
-      ".test-case.tsx",
-      ".test-case.js",
-      ".test-case.mjs"
-    ],
+    evaluatorSuffixes: [".evaluator.ts", ".evaluator.tsx", ".evaluator.js", ".evaluator.mjs"],
+    testCaseSuffixes: [".test-case.ts", ".test-case.tsx", ".test-case.js", ".test-case.mjs"],
     excludeDirectories: ["node_modules", "dist", ".next", ".git", ".pnpm-store"]
   },
   artifactDirectory: ".eval-results",
@@ -813,14 +809,15 @@ function getJitiLoader() {
   }
   const createJiti2 = jitiModule.createJiti ?? jitiModule.default;
   if (typeof createJiti2 !== "function") {
-    throw new Error(
-      "Failed to initialize jiti for m4trix eval config loading."
-    );
+    throw new Error("Failed to initialize jiti for m4trix eval config loading.");
   }
-  cachedLoader = createJiti2(import.meta.url, {
-    interopDefault: true,
-    moduleCache: true
-  });
+  cachedLoader = createJiti2(
+    import.meta.url,
+    {
+      interopDefault: true,
+      moduleCache: true
+    }
+  );
   return cachedLoader;
 }
 function resolveConfigModuleExport(loadedModule) {
@@ -924,9 +921,7 @@ async function loadModuleExports(filePath) {
 }
 async function collectDatasetsFromFiles(config) {
   const files = await walkDirectory(config.rootDir, config.excludeDirectories);
-  const matched = files.filter(
-    (filePath) => hasOneSuffix(filePath, config.datasetSuffixes)
-  );
+  const matched = files.filter((filePath) => hasOneSuffix(filePath, config.datasetSuffixes));
   const found = await Promise.all(
     matched.map(async (absolutePath) => {
       const exports = await loadModuleExports(absolutePath);
@@ -943,9 +938,7 @@ async function collectDatasetsFromFiles(config) {
 }
 async function collectEvaluatorsFromFiles(config) {
   const files = await walkDirectory(config.rootDir, config.excludeDirectories);
-  const matched = files.filter(
-    (filePath) => hasOneSuffix(filePath, config.evaluatorSuffixes)
-  );
+  const matched = files.filter((filePath) => hasOneSuffix(filePath, config.evaluatorSuffixes));
   const found = await Promise.all(
     matched.map(async (absolutePath) => {
       const exports = await loadModuleExports(absolutePath);
@@ -962,9 +955,7 @@ async function collectEvaluatorsFromFiles(config) {
 }
 async function collectTestCasesFromFiles(config) {
   const files = await walkDirectory(config.rootDir, config.excludeDirectories);
-  const matched = files.filter(
-    (filePath) => hasOneSuffix(filePath, config.testCaseSuffixes)
-  );
+  const matched = files.filter((filePath) => hasOneSuffix(filePath, config.testCaseSuffixes));
   const found = await Promise.all(
     matched.map(async (absolutePath) => {
       const exports = await loadModuleExports(absolutePath);
@@ -1036,16 +1027,8 @@ function createDiffString(expected, actual, diffOptions) {
   const expectedProcessed = preprocessForDiff(expected, diffOptions);
   const actualProcessed = preprocessForDiff(actual, diffOptions);
   if (diffOptions?.keysOnly) {
-    const expectedKeys = JSON.stringify(
-      extractKeys(expectedProcessed),
-      null,
-      2
-    );
-    const actualKeys = JSON.stringify(
-      extractKeys(actualProcessed),
-      null,
-      2
-    );
+    const expectedKeys = JSON.stringify(extractKeys(expectedProcessed), null, 2);
+    const actualKeys = JSON.stringify(extractKeys(actualProcessed), null, 2);
     const parts2 = diffLines(expectedKeys, actualKeys);
     return formatDiffParts(parts2);
   }
@@ -1056,9 +1039,7 @@ function createDiffString(expected, actual, diffOptions) {
   }
   const parts = diffLines(expectedStr, actualStr);
   if (diffOptions?.outputNewOnly) {
-    const filtered = parts.filter(
-      (p) => p.added === true
-    );
+    const filtered = parts.filter((p) => p.added === true);
     return formatDiffParts(filtered);
   }
   return formatDiffParts(parts);
@@ -1160,10 +1141,7 @@ var ScoreAggregate = {
       const count = values.length || 1;
       const result = {};
       for (const field of fields) {
-        result[field] = values.reduce(
-          (s, v) => s + (v[field] ?? 0),
-          0
-        ) / count;
+        result[field] = values.reduce((s, v) => s + (v[field] ?? 0), 0) / count;
       }
       return result;
     };
@@ -1197,13 +1175,10 @@ var ScoreAggregate = {
           (s, v) => s + (v[valueField] ?? 0),
           0
         );
-        const sumSq = values.reduce(
-          (s, v) => {
-            const value = v[valueField] ?? 0;
-            return s + value * value;
-          },
-          0
-        );
+        const sumSq = values.reduce((s, v) => {
+          const value = v[valueField] ?? 0;
+          return s + value * value;
+        }, 0);
         const mean = sum / count;
         const variance = (sumSq - count * mean * mean) / (count - 1);
         stdDev = variance > 0 ? Math.sqrt(variance) : 0;
@@ -1434,20 +1409,14 @@ function nowIsoForFile() {
   return (/* @__PURE__ */ new Date()).toISOString().replace(/[:.]/g, "-");
 }
 function createArtifactPath(artifactDirectory, datasetId, runId) {
-  return join(
-    artifactDirectory,
-    `${datasetId}_${runId}_${nowIsoForFile()}.jsonl`
-  );
+  return join(artifactDirectory, `${datasetId}_${runId}_${nowIsoForFile()}.jsonl`);
 }
 function processOneEvaluation(task, unit, totalEvaluations, publishEvent, persistenceQueue, updateSnapshot, startedRef, completedRef, passedRef, failedRef, testCaseResultsRef) {
   const { testCaseItem, rerunIndex, rerunTotal } = unit;
   return Effect.gen(function* () {
     const evaluatorRunId = `run-${randomUUID()}`;
     const started = Date.now();
-    const startedEvaluations = yield* Ref.modify(startedRef, (n) => [
-      n + 1,
-      n + 1
-    ]);
+    const startedEvaluations = yield* Ref.modify(startedRef, (n) => [n + 1, n + 1]);
     yield* publishEvent({
       type: "TestCaseStarted",
       runId: task.runId,
@@ -1480,9 +1449,7 @@ function processOneEvaluation(task, unit, totalEvaluations, publishEvent, persis
         return error;
       };
       try {
-        const ctx = yield* Effect.promise(
-          () => Promise.resolve(evaluator.resolveContext())
-        );
+        const ctx = yield* Effect.promise(() => Promise.resolve(evaluator.resolveContext()));
         const result = yield* Effect.promise(
           () => Promise.resolve().then(
             () => evaluateFn({
@@ -1537,10 +1504,7 @@ function processOneEvaluation(task, unit, totalEvaluations, publishEvent, persis
       }
     }
     const rerunPassedThis = evaluatorScores.every((s) => s.passed);
-    const completedEvaluations = yield* Ref.modify(completedRef, (n) => [
-      n + 1,
-      n + 1
-    ]);
+    const completedEvaluations = yield* Ref.modify(completedRef, (n) => [n + 1, n + 1]);
     const progressEvent = {
       type: "TestCaseProgress",
       runId: task.runId,
@@ -1589,10 +1553,7 @@ function processOneEvaluation(task, unit, totalEvaluations, publishEvent, persis
       } else {
         yield* Ref.update(failedRef, (n) => n + 1);
       }
-      const [passed, failed] = yield* Effect.all([
-        Ref.get(passedRef),
-        Ref.get(failedRef)
-      ]);
+      const [passed, failed] = yield* Effect.all([Ref.get(passedRef), Ref.get(failedRef)]);
       yield* updateSnapshot(task.runId, (snapshot) => ({
         ...snapshot,
         passedTestCases: passed,
@@ -1942,15 +1903,11 @@ var EffectRunner = class {
     this.persistenceQueue = Effect.runSync(
       Queue.unbounded()
     );
-    this.snapshotsRef = Effect.runSync(
-      Ref.make(/* @__PURE__ */ new Map())
-    );
+    this.snapshotsRef = Effect.runSync(Ref.make(/* @__PURE__ */ new Map()));
     this.listeners = /* @__PURE__ */ new Set();
     this.datasetsById = /* @__PURE__ */ new Map();
     this.evaluatorsById = /* @__PURE__ */ new Map();
-    this.schedulerFiber = Effect.runFork(
-      this.createSchedulerEffect()
-    );
+    this.schedulerFiber = Effect.runFork(this.createSchedulerEffect());
     this.persistenceFiber = Effect.runFork(
       createPersistenceWorker(this.persistenceQueue)
     );
@@ -2097,9 +2054,9 @@ var EffectRunner = class {
     return Effect.runSync(Ref.get(this.snapshotsRef)).get(runId);
   }
   getAllRunSnapshots() {
-    return Array.from(
-      Effect.runSync(Ref.get(this.snapshotsRef)).values()
-    ).sort((a, b) => b.queuedAt - a.queuedAt);
+    return Array.from(Effect.runSync(Ref.get(this.snapshotsRef)).values()).sort(
+      (a, b) => b.queuedAt - a.queuedAt
+    );
   }
   async loadRunSnapshotsFromArtifacts() {
     return loadRunSnapshotsFromArtifacts(this.config);
@@ -2315,11 +2272,7 @@ function DatasetsView({
     ] })
   ] });
 }
-function RunsView({
-  state,
-  dataset,
-  selectedRun
-}) {
+function RunsView({ state, dataset, selectedRun }) {
   const runs = dataset?.runs ?? [];
   const rightFocused = state.focus === "right";
   return /* @__PURE__ */ jsxs(Fragment, { children: [
@@ -2335,10 +2288,10 @@ function RunsView({
       /* @__PURE__ */ jsxs(Text, { color: "gray", children: [
         "Commit: ",
         selectedRun.meta.commit,
-        "  Branch: ",
+        " Branch: ",
         selectedRun.meta.branch,
+        " Seed:",
         " ",
-        "Seed: ",
         selectedRun.meta.seed
       ] }),
       /* @__PURE__ */ jsx(Text, { children: " " }),
@@ -2351,23 +2304,10 @@ function RunsView({
           format: (v) => `${v}%`
         }
       ),
-      /* @__PURE__ */ jsx(
-        TextBar,
-        {
-          label: "avg score",
-          value: Math.round(selectedRun.performance.avgScore * 100)
-        }
-      ),
+      /* @__PURE__ */ jsx(TextBar, { label: "avg score", value: Math.round(selectedRun.performance.avgScore * 100) }),
       /* @__PURE__ */ jsx(Text, { children: " " }),
       /* @__PURE__ */ jsx(SectionHeader, { children: "Dimensions" }),
-      selectedRun.dimensions.map((dimension) => /* @__PURE__ */ jsx(
-        TextBar,
-        {
-          label: dimension.name,
-          value: dimension.score
-        },
-        dimension.name
-      )),
+      selectedRun.dimensions.map((dimension) => /* @__PURE__ */ jsx(TextBar, { label: dimension.name, value: dimension.score }, dimension.name)),
       /* @__PURE__ */ jsx(Text, { children: " " }),
       /* @__PURE__ */ jsx(SectionHeader, { children: "Latency trend" }),
       /* @__PURE__ */ jsx(
@@ -2470,15 +2410,7 @@ function buildDetailRows(run, testCases, evaluatorNameById) {
     ...dimensions.map((d) => /* @__PURE__ */ jsx(TextBar, { label: d.name, value: d.score }, `dim-${d.name}`)),
     /* @__PURE__ */ jsx(Text, { children: " " }, "sp2"),
     /* @__PURE__ */ jsx(SectionHeader, { children: "Checks (boolean)" }, "checks-h"),
-    ...checks.map((c) => /* @__PURE__ */ jsx(
-      CheckRow,
-      {
-        name: c.name,
-        passed: c.passed,
-        detail: c.detail
-      },
-      `chk-${c.name}`
-    )),
+    ...checks.map((c) => /* @__PURE__ */ jsx(CheckRow, { name: c.name, passed: c.passed, detail: c.detail }, `chk-${c.name}`)),
     /* @__PURE__ */ jsx(Text, { children: " " }, "sp3"),
     /* @__PURE__ */ jsx(SectionHeader, { children: "Performance" }, "perf-h"),
     /* @__PURE__ */ jsx(
@@ -2595,17 +2527,10 @@ function buildDetailRows(run, testCases, evaluatorNameById) {
           }
         } else {
           rows.push(
-            /* @__PURE__ */ jsxs(
-              Text,
-              {
-                color: "gray",
-                children: [
-                  "      ",
-                  "n/a"
-                ]
-              },
-              `tc-${tc.testCaseId}-${item.evaluatorId}-n/a`
-            )
+            /* @__PURE__ */ jsxs(Text, { color: "gray", children: [
+              "      ",
+              "n/a"
+            ] }, `tc-${tc.testCaseId}-${item.evaluatorId}-n/a`)
           );
         }
         if (!item.passed && item.logs && item.logs.length > 0) {
@@ -2663,7 +2588,7 @@ function RunDetailsView({
   const runs = dataset?.runs ?? [];
   const rightFocused = state.focus === "right";
   const [testCases, setTestCases] = useState([]);
-  const evaluatorNameById = React2.useMemo(
+  const evaluatorNameById = React.useMemo(
     () => new Map(evaluators.map((e) => [e.id, e.name])),
     [evaluators]
   );
@@ -2686,7 +2611,7 @@ function RunDetailsView({
   const visible = rows.slice(offset, offset + DETAILS_PAGE_SIZE);
   return /* @__PURE__ */ jsxs(Fragment, { children: [
     /* @__PURE__ */ jsx(RunsSidebar, { state, dataset, runs }),
-    /* @__PURE__ */ jsx(Pane, { flexGrow: 1, marginLeft: 1, focused: rightFocused, children: /* @__PURE__ */ jsx(Box, { flexDirection: "column", children: visible.map((row, i) => /* @__PURE__ */ jsx(React2.Fragment, { children: row }, i)) }) })
+    /* @__PURE__ */ jsx(Pane, { flexGrow: 1, marginLeft: 1, focused: rightFocused, children: /* @__PURE__ */ jsx(Box, { flexDirection: "column", children: visible.map((row, i) => /* @__PURE__ */ jsx(React.Fragment, { children: row }, i)) }) })
   ] });
 }
 var LEFT_PANE_WIDTH3 = 44;
@@ -2709,19 +2634,11 @@ function NewEvaluationView({
       visibleEvaluators.map((evaluator, index) => {
         const selected = index === state.evaluatorMenuIndex;
         const inSelection = state.selectedEvaluatorIds.includes(evaluator.id);
-        return /* @__PURE__ */ jsxs(
-          Text,
-          {
-            color: selected ? "cyan" : "gray",
-            bold: selected,
-            children: [
-              selected ? "\u25B8 " : "  ",
-              inSelection ? "[x] " : "[ ] ",
-              evaluator.name
-            ]
-          },
-          evaluator.id
-        );
+        return /* @__PURE__ */ jsxs(Text, { color: selected ? "cyan" : "gray", bold: selected, children: [
+          selected ? "\u25B8 " : "  ",
+          inSelection ? "[x] " : "[ ] ",
+          evaluator.name
+        ] }, evaluator.id);
       })
     ] }),
     /* @__PURE__ */ jsxs(Pane, { flexGrow: 1, marginLeft: 1, focused: rightFocused, children: [
@@ -2753,26 +2670,16 @@ function clampCursor(state, filteredDatasetsLength, selectedRunCount) {
     ...state,
     datasetMenuIndex: Math.max(0, Math.min(state.datasetMenuIndex, datasetMax)),
     runMenuIndex: Math.max(0, Math.min(state.runMenuIndex, runMax)),
-    evaluatorMenuIndex: Math.max(
-      0,
-      Math.min(state.evaluatorMenuIndex, evaluatorMax)
-    )
+    evaluatorMenuIndex: Math.max(0, Math.min(state.evaluatorMenuIndex, evaluatorMax))
   };
 }
-function EvalsCliApp({
-  data,
-  args,
-  runner
-}) {
+function EvalsCliApp({ data, args, runner }) {
   const { exit } = useApp();
   const { width: stdoutWidth, height: stdoutHeight } = useScreenSize();
   const [liveData, setLiveData] = useState(data);
   const [runtimeMessage, setRuntimeMessage] = useState();
   const overviewRowCountRef = useRef(0);
-  const [state, dispatch] = useReducer(
-    reduceCliState,
-    createInitialState(data, args)
-  );
+  const [state, dispatch] = useReducer(reduceCliState, createInitialState(data, args));
   useEffect(() => {
     setLiveData(data);
   }, [data]);
@@ -2804,14 +2711,8 @@ function EvalsCliApp({
     filteredDatasets.length,
     getDatasetByMenuIndex(filteredDatasets, state.datasetMenuIndex)?.runs.length ?? 0
   );
-  const selectedDataset = getDatasetByMenuIndex(
-    filteredDatasets,
-    clampedState.datasetMenuIndex
-  );
-  const selectedRun = getRunByMenuIndex(
-    selectedDataset,
-    clampedState.runMenuIndex
-  );
+  const selectedDataset = getDatasetByMenuIndex(filteredDatasets, clampedState.datasetMenuIndex);
+  const selectedRun = getRunByMenuIndex(selectedDataset, clampedState.runMenuIndex);
   const visibleEvaluators = liveData.evaluators.filter(
     (evaluator) => evaluator.name.toLowerCase().includes(clampedState.searchQuery.toLowerCase())
   );
@@ -2905,9 +2806,7 @@ function EvalsCliApp({
           `Started ${snapshot.runId} on ${selectedDataset.name} (${snapshot.totalTestCases} cases).`
         );
       }).catch((error) => {
-        setRuntimeMessage(
-          error instanceof Error ? error.message : "Failed to start evaluation."
-        );
+        setRuntimeMessage(error instanceof Error ? error.message : "Failed to start evaluation.");
       });
     }
   });
@@ -2934,14 +2833,7 @@ function EvalsCliApp({
       );
     }
     if (clampedState.level === "runs") {
-      return /* @__PURE__ */ jsx(
-        RunsView,
-        {
-          state: clampedState,
-          dataset: selectedDataset,
-          selectedRun
-        }
-      );
+      return /* @__PURE__ */ jsx(RunsView, { state: clampedState, dataset: selectedDataset, selectedRun });
     }
     return /* @__PURE__ */ jsx(
       RunDetailsView,
@@ -2953,82 +2845,44 @@ function EvalsCliApp({
       }
     );
   };
-  return /* @__PURE__ */ jsxs(
-    Box,
-    {
-      flexDirection: "column",
-      flexGrow: 1,
-      width: stdoutWidth,
-      height: stdoutHeight,
-      children: [
-        /* @__PURE__ */ jsx(
-          Box,
-          {
-            borderStyle: "round",
-            borderColor: "cyan",
-            paddingX: 1,
-            width: stdoutWidth,
-            children: /* @__PURE__ */ jsx(Text, { children: getBreadcrumbText(
-              clampedState,
-              selectedDataset?.name,
-              selectedRun?.label
-            ) })
-          }
-        ),
-        clampedState.startupWarnings.length > 0 && /* @__PURE__ */ jsxs(
-          Box,
-          {
-            marginTop: 1,
-            borderStyle: "round",
-            borderColor: "yellow",
-            paddingX: 1,
-            flexDirection: "column",
-            width: stdoutWidth,
-            children: [
-              /* @__PURE__ */ jsx(Text, { color: "yellow", children: "Startup warnings:" }),
-              clampedState.startupWarnings.map((warning, index) => /* @__PURE__ */ jsx(Text, { children: warning }, `${warning}-${index}`))
-            ]
-          }
-        ),
-        clampedState.searchMode && /* @__PURE__ */ jsxs(
-          Box,
-          {
-            marginTop: 1,
-            borderStyle: "round",
-            borderColor: "magenta",
-            paddingX: 1,
-            width: stdoutWidth,
-            children: [
-              /* @__PURE__ */ jsx(Text, { color: "magenta", bold: true, children: "Search: " }),
-              /* @__PURE__ */ jsx(Text, { color: "white", children: clampedState.searchQuery })
-            ]
-          }
-        ),
-        runtimeMessage && /* @__PURE__ */ jsx(
-          Box,
-          {
-            marginTop: 1,
-            borderStyle: "round",
-            borderColor: "blue",
-            paddingX: 1,
-            width: stdoutWidth,
-            children: /* @__PURE__ */ jsx(Text, { color: "blue", children: runtimeMessage })
-          }
-        ),
-        /* @__PURE__ */ jsx(
-          Box,
-          {
-            marginTop: 1,
-            flexGrow: 1,
-            width: stdoutWidth,
-            flexDirection: "row",
-            children: renderContent()
-          }
-        ),
-        /* @__PURE__ */ jsx(Box, { marginTop: 1, paddingX: 1, children: /* @__PURE__ */ jsx(Text, { color: "gray", children: getFooterText(clampedState) }) })
-      ]
-    }
-  );
+  return /* @__PURE__ */ jsxs(Box, { flexDirection: "column", flexGrow: 1, width: stdoutWidth, height: stdoutHeight, children: [
+    /* @__PURE__ */ jsx(Box, { borderStyle: "round", borderColor: "cyan", paddingX: 1, width: stdoutWidth, children: /* @__PURE__ */ jsx(Text, { children: getBreadcrumbText(clampedState, selectedDataset?.name, selectedRun?.label) }) }),
+    clampedState.startupWarnings.length > 0 && /* @__PURE__ */ jsxs(
+      Box,
+      {
+        marginTop: 1,
+        borderStyle: "round",
+        borderColor: "yellow",
+        paddingX: 1,
+        flexDirection: "column",
+        width: stdoutWidth,
+        children: [
+          /* @__PURE__ */ jsx(Text, { color: "yellow", children: "Startup warnings:" }),
+          clampedState.startupWarnings.map((warning, index) => /* @__PURE__ */ jsx(Text, { children: warning }, `${warning}-${index}`))
+        ]
+      }
+    ),
+    clampedState.searchMode && /* @__PURE__ */ jsxs(
+      Box,
+      {
+        marginTop: 1,
+        borderStyle: "round",
+        borderColor: "magenta",
+        paddingX: 1,
+        width: stdoutWidth,
+        children: [
+          /* @__PURE__ */ jsxs(Text, { color: "magenta", bold: true, children: [
+            "Search:",
+            " "
+          ] }),
+          /* @__PURE__ */ jsx(Text, { color: "white", children: clampedState.searchQuery })
+        ]
+      }
+    ),
+    runtimeMessage && /* @__PURE__ */ jsx(Box, { marginTop: 1, borderStyle: "round", borderColor: "blue", paddingX: 1, width: stdoutWidth, children: /* @__PURE__ */ jsx(Text, { color: "blue", children: runtimeMessage }) }),
+    /* @__PURE__ */ jsx(Box, { marginTop: 1, flexGrow: 1, width: stdoutWidth, flexDirection: "row", children: renderContent() }),
+    /* @__PURE__ */ jsx(Box, { marginTop: 1, paddingX: 1, children: /* @__PURE__ */ jsx(Text, { color: "gray", children: getFooterText(clampedState) }) })
+  ] });
 }
 async function main() {
   const args = parseStartupArgs(process.argv.slice(2));
@@ -3040,9 +2894,7 @@ async function main() {
   process.on("SIGTERM", () => {
     void runner.shutdown().finally(() => process.exit(0));
   });
-  withFullScreen(
-    /* @__PURE__ */ jsx(EvalsCliApp, { data, args, runner })
-  ).start();
+  withFullScreen(/* @__PURE__ */ jsx(EvalsCliApp, { data, args, runner })).start();
 }
 void main();
 //# sourceMappingURL=out.js.map