@post-print/agent-test 2.1.0 → 2.2.1

package/dist/viewer/v2-client.js

@@ -4,7 +4,8 @@ import { useMutation, useQueries, useQuery, useQueryClient } from "@tanstack/rea
 import { createRootRoute, createRoute, createRouter, Link, Outlet, useNavigate, } from "@tanstack/react-router";
 import { createContext, useCallback, useContext, useEffect, useId, useRef, useState, } from "react";
 import Markdown from "react-markdown";
-import { assertionComparison, conversationPlaceholder, formatRunDate, median, preferredTestExecutionId, stripAnsi, testRunStatus, unwrapShellCommand, } from "./presentation.js";
+import { ElapsedTime } from "./elapsed-time.js";
+import { assertionComparison, conversationPlaceholder, describeShape, formatRunDate, median, preferredTestExecutionId, stripAnsi, testIntent, testRunStatus, unwrapShellCommand, } from "./presentation.js";
 import { RunInboxPage, RunPage } from "./run-inbox.js";
 import { ToolImages, toolImages } from "./tool-images.js";
 import { cancelExecution, fetchExecution as fetchExecutionRecord, startSuite, startTest, startTestGroup, useCatalog, useExecutionHistory, } from "./viewer-queries.js";
@@ -363,6 +364,18 @@ const styles = stylex.create({
         letterSpacing: "0.08em",
         textTransform: "uppercase",
     },
+    catalogNotice: {
+        margin: "0.6rem 1.25rem 0",
+        color: "var(--muted)",
+        fontSize: "0.72rem",
+        lineHeight: 1.45,
+    },
+    catalogNoticeError: {
+        padding: "0.45rem 0.6rem",
+        borderRadius: 8,
+        color: "var(--fail)",
+        backgroundColor: "var(--fail-soft)",
+    },
     groupCount: { color: "var(--subtle)", fontSize: "0.66rem", fontVariantNumeric: "tabular-nums" },
     navLink: {
         display: "block",
@@ -431,6 +444,16 @@ const styles = stylex.create({
         alignSelf: "start",
     },
     description: { color: "var(--muted)", fontSize: "0.82rem", lineHeight: 1.5, maxWidth: "54rem" },
+    suiteDescription: {
+        margin: 0,
+        color: "var(--subtle)",
+        fontSize: "0.74rem",
+        lineHeight: 1.45,
+        overflow: "hidden",
+        textOverflow: "ellipsis",
+        whiteSpace: "nowrap",
+    },
+    headerChecks: { paddingTop: "0.55rem" },
     path: {
         color: "var(--subtle)",
         fontSize: "0.7rem",
@@ -632,6 +655,46 @@ const styles = stylex.create({
         gap: "0.7rem",
     },
     attemptTitle: { fontSize: "0.84rem", lineHeight: 1.4 },
+    liveStrip: {
+        position: "sticky",
+        top: 0,
+        zIndex: 2,
+        display: "grid",
+        gap: "0.3rem",
+        padding: "0.55rem 0.75rem",
+        border: "1px solid var(--border-strong)",
+        borderRadius: 10,
+        backgroundColor: "var(--panel-2)",
+        minWidth: 0,
+    },
+    liveStripLine: {
+        display: "flex",
+        flexWrap: "wrap",
+        alignItems: "center",
+        gap: "0.4rem 0.75rem",
+    },
+    liveStripElapsed: {
+        color: "var(--text)",
+        fontSize: "0.82rem",
+        fontWeight: 700,
+        fontVariantNumeric: "tabular-nums",
+    },
+    liveStripMeta: { color: "var(--muted)", fontSize: "0.76rem", fontVariantNumeric: "tabular-nums" },
+    liveStripNow: {
+        margin: 0,
+        minWidth: 0,
+        overflow: "hidden",
+        textOverflow: "ellipsis",
+        whiteSpace: "nowrap",
+        color: "var(--text)",
+        fontSize: "0.78rem",
+    },
+    liveStripLabel: { color: "var(--subtle)", fontWeight: 650 },
+    liveStripDetail: {
+        marginLeft: "0.4rem",
+        color: "var(--muted)",
+        fontSize: "0.72rem",
+    },
     verdictPanel: {
         display: "grid",
         gap: "0.55rem",
@@ -652,6 +715,16 @@ const styles = stylex.create({
         margin: 0,
     },
     verdictMetric: { display: "grid", gap: "0.1rem", minWidth: 0 },
+    verdictAttachments: {
+        display: "flex",
+        flexWrap: "wrap",
+        gap: "0.25rem 1rem",
+        margin: 0,
+        padding: 0,
+        listStyle: "none",
+        fontSize: "0.8rem",
+        overflowWrap: "anywhere",
+    },
     verdictMetricValue: {
         display: "grid",
         margin: 0,
@@ -990,15 +1063,38 @@ const styles = stylex.create({
     judgeContextList: { display: "grid", gap: "0.3rem", margin: 0 },
     judgeContextItem: {
         display: "grid",
-        gridTemplateColumns: {
-            default: "minmax(7.5rem, 0.3fr) minmax(0, 1fr)",
-            "@media (max-width: 620px)": "1fr",
-        },
-        gap: "0.65rem",
+        gap: "0.45rem",
         padding: "0.4rem 0.5rem",
         borderRadius: 6,
         backgroundColor: "var(--bg)",
     },
+    judgeContextFieldSummary: {
+        display: "flex",
+        alignItems: "baseline",
+        gap: "0.5rem",
+        cursor: "pointer",
+        minWidth: 0,
+    },
+    judgeContextShape: {
+        color: "var(--muted)",
+        fontSize: "0.68rem",
+        fontVariantNumeric: "tabular-nums",
+    },
+    judgeContextFieldBody: {
+        display: "grid",
+        gap: "0.35rem",
+        maxHeight: "24rem",
+        overflow: "auto",
+    },
+    judgeContextStringList: {
+        margin: 0,
+        paddingLeft: "1.1rem",
+        color: "var(--text)",
+        fontSize: "0.72rem",
+        lineHeight: 1.5,
+        fontFamily: "ui-monospace, SFMono-Regular, Menlo, monospace",
+        wordBreak: "break-word",
+    },
     judgeContextLabel: {
         color: "var(--subtle)",
         fontSize: "0.64rem",
@@ -1281,7 +1377,7 @@ function ViewerSidebarContent({ mode, open = true, headingId, navigationId, onTo
     const history = useExecutionHistory();
     const groups = groupTests(catalog.data?.tests ?? []);
     const statuses = latestTestStatuses(history.data ?? []);
-    return (_jsxs(_Fragment, { children: [_jsx(SidebarHeader, { mode: mode, open: open, count: catalog.data?.tests.length ?? 0, headingId: headingId, navigationId: navigationId, onToggle: mode === "mobile" ? onClose : onToggle, onNavigate: onClose }), open ? (_jsx(SidebarNavigation, { groups: groups, statuses: statuses, error: catalog.isError, navigationId: navigationId, onNavigate: onClose })) : null] }));
+    return (_jsxs(_Fragment, { children: [_jsx(SidebarHeader, { mode: mode, open: open, count: catalog.data?.tests.length ?? 0, headingId: headingId, navigationId: navigationId, onToggle: mode === "mobile" ? onClose : onToggle, onNavigate: onClose }), open ? (_jsx(SidebarNavigation, { groups: groups, statuses: statuses, catalogState: catalogState(catalog), navigationId: navigationId, onNavigate: onClose })) : null] }));
 }
 function SidebarHeader({ mode, open, count, headingId, navigationId, onToggle, onNavigate, }) {
     const { workers } = useContext(WorkerContext);
@@ -1309,8 +1405,21 @@ function SidebarToggle({ mode, open, navigationId, onToggle, }) {
     const mobile = mode === "mobile";
     return (_jsx("button", { type: "button", "data-drawer-close": mobile ? "true" : undefined, "aria-controls": navigationId, "aria-expanded": open, "aria-label": mobile ? "Close test menu" : open ? "Collapse test sidebar" : "Expand test sidebar", ...stylex.props(styles.buttonReset, styles.toggleButton, mobile && styles.mobileCloseButton), onClick: onToggle, children: mobile ? "×" : open ? "←" : "→" }));
 }
-function SidebarNavigation({ groups, statuses, error, navigationId, onNavigate, }) {
-    return (_jsxs(_Fragment, { children: [error ? _jsx("p", { role: "alert", children: "Test discovery is unavailable." }) : null, _jsx("nav", { id: navigationId, "aria-label": "Tests", ...stylex.props(styles.navigation), children: [...groups].map(([name, tests]) => (_jsxs("section", { "aria-label": `${name} tests`, ...stylex.props(styles.navSection), children: [_jsxs("div", { ...stylex.props(styles.groupHeader), children: [_jsx("strong", { ...stylex.props(styles.groupLabel), children: name }), _jsxs("div", { ...stylex.props(styles.groupHeaderActions), children: [_jsx("span", { ...stylex.props(styles.groupCount), children: tests.length }), _jsx(RunTestGroupButton, { name: name, tests: tests, onNavigate: onNavigate })] })] }), tests.map((test) => (_jsx(TestLink, { test: test, status: statuses.get(test.id), onNavigate: onNavigate }, test.id)))] }, name))) })] }));
+/** A failed refresh over a list already shown is stale, not missing. */
+function catalogState(catalog) {
+    if (!catalog.isError)
+        return "ready";
+    return catalog.data ? "stale" : "unavailable";
+}
+function CatalogNotice({ state }) {
+    if (state === "ready")
+        return null;
+    if (state === "stale")
+        return (_jsx("p", { role: "status", ...stylex.props(styles.catalogNotice), children: "Test list may be out of date. Retrying\u2026" }));
+    return (_jsx("p", { role: "alert", ...stylex.props(styles.catalogNotice, styles.catalogNoticeError), children: "Test discovery is unavailable. Retrying\u2026" }));
+}
+function SidebarNavigation({ groups, statuses, catalogState, navigationId, onNavigate, }) {
+    return (_jsxs(_Fragment, { children: [_jsx(CatalogNotice, { state: catalogState }), _jsx("nav", { id: navigationId, "aria-label": "Tests", ...stylex.props(styles.navigation), children: [...groups].map(([name, tests]) => (_jsxs("section", { "aria-label": `${name} tests`, ...stylex.props(styles.navSection), children: [_jsxs("div", { ...stylex.props(styles.groupHeader), children: [_jsx("strong", { title: tests[0] ? testIntent(tests[0], tests).suite : undefined, ...stylex.props(styles.groupLabel), children: name }), _jsxs("div", { ...stylex.props(styles.groupHeaderActions), children: [_jsx("span", { ...stylex.props(styles.groupCount), children: tests.length }), _jsx(RunTestGroupButton, { name: name, tests: tests, onNavigate: onNavigate })] })] }), tests.map((test) => (_jsx(TestLink, { test: test, hint: testIntent(test, tests).own, status: statuses.get(test.id), onNavigate: onNavigate }, test.id)))] }, name))) })] }));
 }
 function RunTestGroupButton({ name, tests, onNavigate, }) {
     const { workers } = useContext(WorkerContext);
@@ -1326,8 +1435,8 @@ function RunTestGroupButton({ name, tests, onNavigate, }) {
     });
     return (_jsx("button", { type: "button", "aria-label": `Run ${name} suite`, title: `Run ${name} suite`, ...stylex.props(styles.buttonReset, styles.groupRunButton), onClick: () => run.mutate(), disabled: run.isPending, children: run.isPending ? "●" : "▶" }));
 }
-function TestLink({ test, status, onNavigate, }) {
-    return (_jsxs(Link, { to: "/tests/$testId", params: { testId: test.id }, search: {}, ...stylex.props(styles.navLink), activeProps: stylex.props(styles.navLink, styles.active), onClick: onNavigate, children: [_jsxs("div", { ...stylex.props(styles.navItemHeader), children: [status ? _jsx(TestStatus, { status: status }) : _jsx("span", { "aria-hidden": "true" }), _jsx("span", { ...stylex.props(styles.navItemTitle), children: test.title.at(-1) })] }), test.project !== "default" ? _jsx("div", { ...stylex.props(styles.meta), children: test.project }) : null] }));
+function TestLink({ test, hint, status, onNavigate, }) {
+    return (_jsxs(Link, { to: "/tests/$testId", params: { testId: test.id }, search: {}, ...stylex.props(styles.navLink), activeProps: stylex.props(styles.navLink, styles.active), onClick: onNavigate, children: [_jsxs("div", { ...stylex.props(styles.navItemHeader), children: [status ? _jsx(TestStatus, { status: status }) : _jsx("span", { "aria-hidden": "true" }), _jsx("span", { title: hint ?? test.title.at(-1), ...stylex.props(styles.navItemTitle), children: test.title.at(-1) })] }), test.project !== "default" ? _jsx("div", { ...stylex.props(styles.meta), children: test.project }) : null] }));
 }
 function TestStatus({ status }) {
     return (_jsx("span", { role: "img", "aria-label": `Latest execution: ${status}`, title: `Latest execution: ${status}`, ...stylex.props(styles.navStatus), children: _jsx("span", { "aria-hidden": "true", ...stylex.props(styles.navStatusDot, status === "passed" && styles.navStatusDotPass, status === "failed" && styles.navStatusDotFail, status === "running" && styles.navStatusDotRunning) }) }));
@@ -1340,9 +1449,9 @@ function TestPage() {
     const test = catalog.data?.tests.find((item) => item.id === testId);
     if (!test)
         return (_jsx("p", { role: "alert", children: "This test is no longer available. Choose another test in the sidebar." }));
-    return _jsx(TestDetail, { test: test }, test.id);
+    return (_jsx(TestDetail, { test: test, intent: testIntent(test, catalog.data?.tests ?? []) }, test.id));
 }
-function TestDetail({ test }) {
+function TestDetail({ test, intent }) {
     const { execution, view } = testRoute.useSearch();
     const history = useExecutionHistory();
     const navigate = useNavigate();
@@ -1359,7 +1468,7 @@ function TestDetail({ test }) {
         params: { testId: test.id },
         search: { execution, view: next === "current" ? undefined : next },
     });
-    return (_jsxs(_Fragment, { children: [_jsx(TestHeader, { test: test, onStarted: chooseExecution }), _jsx(BatchRunLink, { run: runs.find((run) => run.id === selected) }), _jsx(TestTabs, { active: activeView, onChange: chooseView }), _jsx(TestViewPanel, { view: activeView, test: test, runs: runs, selected: selected, historyLoading: history.isPending, onChooseExecution: chooseExecution })] }));
+    return (_jsxs(_Fragment, { children: [_jsx(TestHeader, { test: test, intent: intent, onStarted: chooseExecution }), _jsx(BatchRunLink, { run: runs.find((run) => run.id === selected) }), _jsx(TestTabs, { active: activeView, onChange: chooseView }), _jsx(TestViewPanel, { view: activeView, test: test, runs: runs, selected: selected, historyLoading: history.isPending, onChooseExecution: chooseExecution })] }));
 }
 /** Return path to the run inbox when this test was opened from a batch run. */
 function BatchRunLink({ run }) {
@@ -1368,8 +1477,14 @@ function BatchRunLink({ run }) {
         return null;
     return (_jsxs(Link, { to: "/executions/$executionId", params: { executionId: run.id }, ...stylex.props(styles.batchRunLink), children: ["\u2190 Back to run \u00B7 ", formatRunDate(run.startedAt), " \u00B7 ", count, " tests"] }));
 }
-function TestHeader({ test, onStarted }) {
-    return (_jsxs("header", { ...stylex.props(styles.hero), children: [_jsxs("div", { ...stylex.props(styles.heroTop), children: [_jsxs("div", { ...stylex.props(styles.heroCopy), children: [_jsx("div", { ...stylex.props(styles.brand), children: test.title.slice(0, -1).join(" / ") }), _jsx("h2", { ...stylex.props(styles.title), children: test.title.at(-1) }), _jsx("p", { ...stylex.props(styles.description), children: test.description ?? "This test has no additional description." })] }), _jsx(RunTestButton, { test: test, onStarted: onStarted })] }), _jsxs("div", { ...stylex.props(styles.path), children: [test.project, " \u00B7 ", test.file] })] }));
+function TestHeader({ test, intent, onStarted, }) {
+    return (_jsxs("header", { ...stylex.props(styles.hero), children: [_jsxs("div", { ...stylex.props(styles.heroTop), children: [_jsxs("div", { ...stylex.props(styles.heroCopy), children: [_jsx("div", { ...stylex.props(styles.brand), children: test.title.slice(0, -1).join(" / ") }), intent.suite ? (_jsx("p", { title: intent.suite, ...stylex.props(styles.suiteDescription), children: intent.suite })) : null, _jsx("h2", { ...stylex.props(styles.title), children: test.title.at(-1) }), intent.own ? _jsx("p", { ...stylex.props(styles.description), children: intent.own }) : null] }), _jsx(RunTestButton, { test: test, onStarted: onStarted })] }), _jsx(TestChecks, { criteria: test.criteria ?? [] }), _jsxs("div", { ...stylex.props(styles.path), children: [test.project, " \u00B7 ", test.file] })] }));
+}
+/** What the test checks, folded: the verdict below lists the same checks once a run exists. */
+function TestChecks({ criteria }) {
+    if (!criteria.length)
+        return null;
+    return (_jsxs("details", { "data-disclosure": true, ...stylex.props(styles.judgeContextDisclosure), children: [_jsxs("summary", { ...stylex.props(styles.judgeContextSummary), children: ["Passes when \u00B7 ", criteria.length, " ", pluralize(criteria.length, "check")] }), _jsx("ul", { ...stylex.props(styles.criteriaList, styles.headerChecks), children: criteria.map((criterion) => (_jsxs("li", { ...stylex.props(styles.criterion), children: [_jsx("span", { "aria-hidden": "true", ...stylex.props(styles.criterionIcon), children: "\u2022" }), _jsx("span", { children: criterion })] }, criterion))) })] }));
 }
 const TEST_VIEWS = [
     { id: "current", label: "Current run" },
@@ -1491,7 +1606,7 @@ function ExecutionDetailView({ execution, criteria = [], }) {
     });
     const stopping = execution.status === "running" && (cancel.isPending || cancel.isSuccess);
     const stopLabel = (execution.testIds?.length ?? 0) > 1 ? "Stop suite run" : "Stop this run";
-    return (_jsxs("div", { ...stylex.props(styles.statusPanel), children: [execution.status === "running" && execution.cancellable ? (_jsx("div", { ...stylex.props(styles.runControls), children: _jsx("button", { type: "button", ...stylex.props(styles.buttonReset, styles.button, styles.secondaryButton), onClick: () => cancel.mutate(), disabled: stopping, children: stopping ? "Stopping…" : stopLabel }) })) : null, cancel.isError ? _jsx("p", { role: "alert", children: "The run could not be stopped." }) : null, _jsxs("section", { ...stylex.props(styles.list), children: [execution.attempts.map((attempt) => (_jsx(AttemptCard, { attempt: attempt, criteria: criteria, showTitle: false }, attempt.id))), execution.attempts.length === 0 ? (_jsx("p", { ...stylex.props(styles.empty), children: emptyExecutionMessage(execution, stopping) })) : null] })] }));
+    return (_jsxs("div", { ...stylex.props(styles.statusPanel), children: [execution.status === "running" && execution.cancellable ? (_jsx("div", { ...stylex.props(styles.runControls), children: _jsx("button", { type: "button", ...stylex.props(styles.buttonReset, styles.button, styles.secondaryButton), onClick: () => cancel.mutate(), disabled: stopping, children: stopping ? "Stopping…" : stopLabel }) })) : null, cancel.isError ? _jsx("p", { role: "alert", children: "The run could not be stopped." }) : null, _jsxs("section", { ...stylex.props(styles.list), children: [execution.attempts.map((attempt) => (_jsx(AttemptCard, { attempt: attempt, criteria: criteria, showTitle: false, progress: execution.progress?.filter((entry) => entry.attemptId === attempt.id) }, attempt.id))), execution.attempts.length === 0 ? (_jsx("p", { ...stylex.props(styles.empty), children: emptyExecutionMessage(execution, stopping) })) : null] })] }));
 }
 function emptyExecutionMessage(execution, stopping) {
     if (stopping)
@@ -1523,11 +1638,45 @@ function ComparisonColumn({ execution, label }) {
     return (_jsxs("article", { ...stylex.props(styles.card), children: [_jsxs("strong", { children: [label, ": ", execution.status] }), _jsx("p", { ...stylex.props(styles.meta), children: execution.id.slice(0, 8) }), execution.attempts.map((attempt) => (_jsx(AttemptCard, { attempt: attempt }, attempt.id)))] }));
 }
 /** `showTitle` is off on a test page, whose header already names the test. */
-function AttemptCard({ attempt, criteria = [], showTitle = true, }) {
+function AttemptCard({ attempt, criteria = [], showTitle = true, progress = [], }) {
     const { operations, output, title } = attemptPresentation(attempt, criteria);
     const agents = operations.filter((operation) => operation.kind !== "evaluation");
     const judges = operations.filter((operation) => operation.kind === "evaluation");
-    return (_jsxs("article", { ...stylex.props(styles.attemptCard), children: [_jsxs("div", { ...stylex.props(styles.row), children: [_jsxs("div", { children: [showTitle ? _jsx("strong", { ...stylex.props(styles.attemptTitle), children: title }) : null, _jsxs("p", { ...stylex.props(styles.meta), children: [attempt.project || "default", " \u00B7 ", attempt.operations.length, " named", " ", pluralize(attempt.operations.length, "operation")] })] }), _jsxs("div", { ...stylex.props(styles.attemptActions), children: [attempt.status === "failed" ? (_jsx(CopyFailureButton, { attempt: attempt, criteria: criteria })) : null, _jsx("span", { ...stylex.props(styles.status, attempt.status === "passed" && styles.statusPass, attempt.status === "failed" && styles.statusFail, attempt.status === "running" && styles.statusRunning), children: attempt.status })] })] }), _jsx(TestVerdict, { attempt: attempt, criteria: criteria, operations: operations }), _jsx(OperationList, { operations: operations }), _jsx(ConversationSection, { label: "Judges", sectionOperations: judges, allOperations: operations, attempt: attempt }), _jsx(ConversationSection, { label: "Agents", sectionOperations: agents, allOperations: operations, attempt: attempt }), _jsx(TechnicalDetails, { operations: operations, output: output })] }));
+    return (_jsxs("article", { ...stylex.props(styles.attemptCard), children: [_jsxs("div", { ...stylex.props(styles.row), children: [_jsxs("div", { children: [showTitle ? _jsx("strong", { ...stylex.props(styles.attemptTitle), children: title }) : null, _jsxs("p", { ...stylex.props(styles.meta), children: [attempt.project || "default", " \u00B7 ", attempt.operations.length, " named", " ", pluralize(attempt.operations.length, "operation")] })] }), _jsxs("div", { ...stylex.props(styles.attemptActions), children: [attempt.status === "failed" ? (_jsx(CopyFailureButton, { attempt: attempt, criteria: criteria })) : null, _jsx("span", { ...stylex.props(styles.status, attempt.status === "passed" && styles.statusPass, attempt.status === "failed" && styles.statusFail, attempt.status === "running" && styles.statusRunning), children: attempt.status })] })] }), attempt.status === "running" ? (_jsx(LiveRunStrip, { attempt: attempt, operations: operations, progress: progress.at(-1) })) : null, _jsx(TestVerdict, { attempt: attempt, criteria: criteria, operations: operations }), _jsx(OperationList, { operations: operations }), _jsx(ConversationSection, { label: "Judges", sectionOperations: judges, allOperations: operations, attempt: attempt }), _jsx(ConversationSection, { label: "Agents", sectionOperations: agents, allOperations: operations, attempt: attempt }), _jsx(TechnicalDetails, { operations: operations, output: output })] }));
+}
+/** Where a long run is right now: how long it has run, how far it got, and what it is doing. */
+function LiveRunStrip({ attempt, operations, progress, }) {
+    const step = currentStep(operations);
+    const reported = progress ? progressSummary(progress.data) : undefined;
+    return (_jsxs("section", { "aria-label": "Live run status", ...stylex.props(styles.liveStrip), children: [_jsxs("div", { ...stylex.props(styles.liveStripLine), children: [_jsx(RunningIndicator, { kind: step.kind }), attempt.startedAt ? (_jsx("span", { ...stylex.props(styles.liveStripElapsed), children: _jsx(ElapsedTime, { since: attempt.startedAt }) })) : null, _jsxs("span", { ...stylex.props(styles.liveStripMeta), children: [step.count, " ", pluralize(step.count, "step")] })] }), _jsxs("p", { "aria-live": "polite", title: step.detail, ...stylex.props(styles.liveStripNow), children: [_jsx("span", { ...stylex.props(styles.liveStripLabel), children: "Now:" }), " ", step.label, step.detail ? _jsx("code", { ...stylex.props(styles.liveStripDetail), children: step.detail }) : null] }), reported ? (_jsxs("p", { title: reported, ...stylex.props(styles.liveStripNow), children: [_jsx("span", { ...stylex.props(styles.liveStripLabel), children: "Progress:" }), " ", reported] })) : null] }));
+}
+/** The newest thing the running test did, read from the same live records as the transcript. */
+function currentStep(operations) {
+    const count = operations
+        .filter((operation) => operation.kind !== "evaluation")
+        .flatMap((operation) => liveConversationTimeline(operation.data.filter(isRecord)))
+        .filter((event) => event.kind === "tool").length;
+    const active = [...operations].reverse().find((operation) => operation.status === "running");
+    if (!active)
+        return { label: "Preparing the next step", count };
+    if (active.kind === "evaluation")
+        return { kind: active.kind, label: `Judge ${active.name ?? ""} is reviewing`.trim(), count };
+    const latest = liveConversationTimeline(active.data.filter(isRecord)).at(-1);
+    if (latest?.kind === "tool") {
+        const tool = toolPresentation(latest.tool);
+        return { kind: active.kind, label: tool.title, detail: tool.detail || undefined, count };
+    }
+    const label = latest ? "Writing a reply" : "Waiting for the first agent event";
+    return { kind: active.kind, label, count };
+}
+/** One line from whatever the test passed to reportProgress. */
+function progressSummary(data) {
+    if (typeof data === "string")
+        return data || undefined;
+    if (!isRecord(data))
+        return data === undefined || data === null ? undefined : String(data);
+    const facts = Object.entries(data).map(([key, value]) => `${humanizeKey(key)}: ${formatConversationValue(value)}`);
+    return facts.join(" · ") || undefined;
 }
 /** Test-level outcome first: every check with its result, including judge-decided checks. */
 function TestVerdict({ attempt, criteria, operations, }) {
@@ -1538,7 +1687,7 @@ function TestVerdict({ attempt, criteria, operations, }) {
         status: attempt.status,
         errors: attempt.errors,
     });
-    return (_jsxs("section", { "aria-label": "Test verdict", ...stylex.props(styles.verdictPanel, attempt.status === "passed" && styles.verdictPass, attempt.status === "failed" && styles.verdictFail), children: [_jsxs("p", { ...stylex.props(styles.verdictHeadline), children: [capitalize(attempt.status), _jsx("span", { ...stylex.props(styles.verdictCount), children: checkSummary(presentations, attempt.status) })] }), attempt.status === "running" ? (_jsx("p", { ...stylex.props(styles.meta), children: "The test is still running." })) : null, _jsx(VerdictMetrics, { operations: operations }), _jsx(VerdictChecks, { presentations: presentations }), _jsx(UnassignedFailures, { errors: attempt.errors, presentations: presentations })] }));
+    return (_jsxs("section", { "aria-label": "Test verdict", ...stylex.props(styles.verdictPanel, attempt.status === "passed" && styles.verdictPass, attempt.status === "failed" && styles.verdictFail), children: [_jsxs("p", { ...stylex.props(styles.verdictHeadline), children: [capitalize(attempt.status), _jsx("span", { ...stylex.props(styles.verdictCount), children: checkSummary(presentations, attempt.status) })] }), _jsx(VerdictMetrics, { operations: operations, reported: attempt.metrics }), _jsx(VerdictChecks, { presentations: presentations }), _jsx(UnassignedFailures, { errors: attempt.errors, presentations: presentations }), _jsx(VerdictAttachments, { attachments: attempt.attachments })] }));
 }
 const SHORT_CHECK_LIST = 3;
 /** Checks needing attention always show; passing checks fold away once the list is long. */
@@ -1567,8 +1716,22 @@ const VERDICT_METRICS = [
         format: (value) => String(Math.round(value)),
     },
 ];
+/** `p95LatencyMs` → "P95 latency"; numbers keyed `…Ms` read as durations. */
+function reportedMetricRows(reported) {
+    return Object.entries(reported ?? {}).map(([key, value]) => {
+        const duration = key.endsWith("Ms");
+        const words = (duration ? key.slice(0, -2) : key).replace(/([a-z0-9])([A-Z])/g, "$1 $2");
+        return {
+            key: `reported:${key}`,
+            label: capitalize(words.toLowerCase()),
+            value: duration
+                ? formatDuration(Math.round(value))
+                : value.toLocaleString(undefined, { maximumFractionDigits: 2 }),
+        };
+    });
+}
 /** Agent totals for the attempt; a median per run when the test made several agent runs. */
-function VerdictMetrics({ operations }) {
+function VerdictMetrics({ operations, reported, }) {
     const runs = operations
         .filter((operation) => operation.kind !== "evaluation")
         .map((operation) => operationResult([operation]));
@@ -1579,10 +1742,17 @@ function VerdictMetrics({ operations }) {
         const total = values.reduce((sum, value) => sum + value, 0);
         return [{ ...metric, total, median: median(values) ?? total }];
     });
-    if (!rows.length)
+    const reportedRows = reportedMetricRows(reported);
+    if (!rows.length && !reportedRows.length)
         return null;
     const repeated = runs.length > 1;
-    return (_jsx("dl", { "aria-label": "Run metrics", ...stylex.props(styles.verdictMetrics), children: rows.map((row) => (_jsxs("div", { ...stylex.props(styles.verdictMetric), children: [_jsx("dt", { ...stylex.props(styles.resourceFactLabel), children: row.label }), _jsxs("dd", { ...stylex.props(styles.verdictMetricValue), children: [row.format(row.total), repeated ? (_jsxs("span", { ...stylex.props(styles.verdictMetricMedian), children: ["median ", row.format(row.median), " per run"] })) : null] })] }, row.key))) }));
+    return (_jsxs("dl", { "aria-label": "Run metrics", ...stylex.props(styles.verdictMetrics), children: [rows.map((row) => (_jsxs("div", { ...stylex.props(styles.verdictMetric), children: [_jsx("dt", { ...stylex.props(styles.resourceFactLabel), children: row.label }), _jsxs("dd", { ...stylex.props(styles.verdictMetricValue), children: [row.format(row.total), repeated ? (_jsxs("span", { ...stylex.props(styles.verdictMetricMedian), children: ["median ", row.format(row.median), " per run"] })) : null] })] }, row.key))), reportedRows.map((row) => (_jsxs("div", { ...stylex.props(styles.verdictMetric), children: [_jsx("dt", { ...stylex.props(styles.resourceFactLabel), children: row.label }), _jsx("dd", { ...stylex.props(styles.verdictMetricValue), children: row.value })] }, row.key)))] }));
+}
+/** Files the test attached with a path, opened from the viewer server. */
+function VerdictAttachments({ attachments }) {
+    if (!attachments?.length)
+        return null;
+    return (_jsx("ul", { "aria-label": "Attachments", ...stylex.props(styles.verdictAttachments), children: attachments.map((attachment) => (_jsxs("li", { children: [_jsx("a", { href: attachment.url, target: "_blank", rel: "noreferrer", children: attachment.name }), _jsxs("span", { ...stylex.props(styles.verdictMetricMedian), children: [" ", attachment.contentType] })] }, attachment.url))) }));
 }
 function checkSummary(presentations, status) {
     const total = presentations.length;
@@ -2069,12 +2239,12 @@ function JudgeAsked({ review }) {
         return null;
     return (_jsx("section", { "aria-label": "Judge question", ...stylex.props(styles.judgeQuestion), children: _jsxs("p", { ...stylex.props(styles.judgeQuestionText), children: [_jsx("span", { ...stylex.props(styles.judgeAskedLabel), children: "Asked" }), _jsx("span", { children: review.question })] }) }));
 }
-/** Everything the test sent the judge, folded away: it mostly repeats the agent answers. */
+/** Everything the test sent the judge, folded away unless a finding failed. */
 function JudgeInputs({ review, operations, }) {
     if (!review.context.length)
         return null;
-    // Inputs matter most before findings exist and when a finding failed.
-    const inputsFirst = review.outcomes.length === 0 || review.outcomes.some((outcome) => !outcome.passed);
+    // Inputs explain a failed finding; while the judge runs they only push the transcript away.
+    const inputsFirst = review.outcomes.some((outcome) => !outcome.passed);
     return (_jsx("section", { "aria-label": "Context included", children: _jsxs("details", { "data-disclosure": true, open: inputsFirst, ...stylex.props(styles.judgeContextDisclosure), children: [_jsxs("summary", { ...stylex.props(styles.judgeContextSummary), children: ["What the judge saw \u00B7 ", review.context.length, " selected", " ", pluralize(review.context.length, "field")] }), _jsx(JudgeContext, { context: review.context, operations: operations })] }) }));
 }
 function JudgeContext({ context, operations, }) {
@@ -2082,7 +2252,37 @@ function JudgeContext({ context, operations, }) {
     const [selectedId, setSelectedId] = useState(answers[0]?.id);
     const selected = answers.find((item) => item.id === selectedId) ?? answers[0];
     const panelId = useId();
-    return (_jsxs("div", { ...stylex.props(styles.judgeContext), children: [_jsx("p", { ...stylex.props(styles.judgeContextNote), children: "Only these fields, chosen by the test, were sent to the judge." }), answers.length ? (_jsxs("section", { "aria-label": "Agent answers", ...stylex.props(styles.judgeContextGroup), children: [_jsx("h5", { ...stylex.props(styles.judgeContextGroupHeading), children: "Agent answers" }), _jsx("div", { role: "tablist", "aria-label": "Agent answers included", ...stylex.props(styles.judgeContextTabs), children: answers.map((answer, index) => (_jsx("button", { type: "button", role: "tab", "aria-selected": answer.id === selected?.id, "aria-controls": panelId, ...stylex.props(styles.buttonReset, styles.judgeContextTab, answer.id === selected?.id && styles.judgeContextTabActive), onClick: () => setSelectedId(answer.id), onKeyDown: (event) => selectContextWithKeyboard({ event, answers, index, select: setSelectedId }), children: answer.tabLabel }, answer.id))) }), selected ? (_jsx("div", { id: panelId, role: "tabpanel", "aria-label": `Agent response from ${selected.tabLabel}`, ...stylex.props(styles.judgeContextAnswer), children: _jsx("div", { ...stylex.props(styles.judgeContextValue), children: _jsx("div", { className: "message-markdown", children: _jsx(Markdown, { children: normalizeTranscriptMarkdown(selected.value) }) }) }) })) : null] })) : null, supplied.length ? (_jsxs("section", { "aria-label": "Additional context", ...stylex.props(styles.judgeContextGroup), children: [_jsx("h5", { ...stylex.props(styles.judgeContextGroupHeading), children: "Additional context" }), _jsx("dl", { ...stylex.props(styles.judgeContextList), children: supplied.map((item) => (_jsxs("div", { ...stylex.props(styles.judgeContextItem), children: [_jsx("dt", { ...stylex.props(styles.judgeContextLabel), children: item.label }), _jsx("dd", { ...stylex.props(styles.judgeContextValue), children: item.value })] }, item.label))) })] })) : null] }));
+    return (_jsxs("div", { ...stylex.props(styles.judgeContext), children: [_jsx("p", { ...stylex.props(styles.judgeContextNote), children: "Only these fields, chosen by the test, were sent to the judge." }), answers.length ? (_jsxs("section", { "aria-label": "Agent answers", ...stylex.props(styles.judgeContextGroup), children: [_jsx("h5", { ...stylex.props(styles.judgeContextGroupHeading), children: "Agent answers" }), _jsx("div", { role: "tablist", "aria-label": "Agent answers included", ...stylex.props(styles.judgeContextTabs), children: answers.map((answer, index) => (_jsx("button", { type: "button", role: "tab", "aria-selected": answer.id === selected?.id, "aria-controls": panelId, ...stylex.props(styles.buttonReset, styles.judgeContextTab, answer.id === selected?.id && styles.judgeContextTabActive), onClick: () => setSelectedId(answer.id), onKeyDown: (event) => selectContextWithKeyboard({ event, answers, index, select: setSelectedId }), children: answer.tabLabel }, answer.id))) }), selected ? (_jsx("div", { id: panelId, role: "tabpanel", "aria-label": `Agent response from ${selected.tabLabel}`, ...stylex.props(styles.judgeContextAnswer), children: _jsx("div", { ...stylex.props(styles.judgeContextValue), children: _jsx("div", { className: "message-markdown", children: _jsx(Markdown, { children: normalizeTranscriptMarkdown(selected.value) }) }) }) })) : null] })) : null, supplied.length ? (_jsxs("section", { "aria-label": "Additional context", ...stylex.props(styles.judgeContextGroup), children: [_jsx("h5", { ...stylex.props(styles.judgeContextGroupHeading), children: "Additional context" }), _jsx("div", { ...stylex.props(styles.judgeContextList), children: supplied.map((item) => (_jsx(JudgeContextField, { item: item }, item.label))) })] })) : null] }));
+}
+/** One supplied judge input. A short value reads inline; anything larger folds behind its size. */
+function JudgeContextField({ item }) {
+    if (isInlineValue(item.rawValue))
+        return (_jsxs("div", { ...stylex.props(styles.judgeContextItem, styles.judgeContextFieldSummary), children: [_jsx("span", { ...stylex.props(styles.judgeContextLabel), children: item.label }), _jsx("span", { ...stylex.props(styles.judgeContextValue), children: item.value })] }));
+    return (_jsxs("details", { "data-disclosure": true, ...stylex.props(styles.judgeContextItem), children: [_jsxs("summary", { ...stylex.props(styles.judgeContextFieldSummary), children: [_jsx("span", { ...stylex.props(styles.judgeContextLabel), children: item.label }), _jsx("span", { ...stylex.props(styles.judgeContextShape), children: describeShape(item.rawValue) })] }), _jsx(JudgeContextFieldValue, { item: item })] }));
+}
+function JudgeContextFieldValue({ item }) {
+    const tools = judgeInputToolCalls(item.rawValue);
+    if (tools)
+        return (_jsx("div", { ...stylex.props(styles.judgeContextFieldBody), children: tools.map((tool, index) => (_jsx(ToolActivity, { tool: tool }, tool.seq ?? index))) }));
+    if (isStringList(item.rawValue))
+        return (_jsx("ul", { ...stylex.props(styles.judgeContextFieldBody, styles.judgeContextStringList), children: item.rawValue.map((value, index) => (_jsx("li", { children: value }, `${index}:${value}`))) }));
+    return (_jsx("div", { ...stylex.props(styles.judgeContextFieldBody, styles.judgeContextValue), children: item.value }));
+}
+/** A judge input that is a list of agent tool calls reads like the transcript, not as JSON. */
+function judgeInputToolCalls(value) {
+    if (!Array.isArray(value) || value.length === 0)
+        return undefined;
+    const tools = toolCallsFromOperation({ trace: { toolCalls: value } });
+    return tools.length === value.length ? tools : undefined;
+}
+const INLINE_VALUE_LENGTH = 120;
+function isInlineValue(value) {
+    if (typeof value === "string")
+        return value.length <= INLINE_VALUE_LENGTH && !value.includes("\n");
+    return value === null || value === undefined || typeof value !== "object";
+}
+function isStringList(value) {
+    return (Array.isArray(value) && value.length > 0 && value.every((item) => typeof item === "string"));
 }
 function judgeContextGroups(context, operations) {
     const candidates = operations