inspect-ai 0.3.90__py3-none-any.whl → 0.3.92__py3-none-any.whl

inspect_ai/_view/www/src/{samples → app/samples}/transcript/state/StateEventRenderers.tsx

@@ -1,10 +1,10 @@
 import clsx from "clsx";
 import { FC, Fragment, JSX, ReactNode } from "react";
+import { JsonChange, Messages } from "../../../../@types/log";
 import {
   HumanBaselineView,
   SessionLog,
-} from "../../../components/HumanBaselineView";
-import { JsonChange, Messages } from "../../../types/log";
+} from "../../../../components/HumanBaselineView";
 import { ChatView } from "../../chat/ChatView";
 
 import styles from "./StateEventRenders.module.css";
@@ -204,7 +204,7 @@ const renderTools = (
             >
               {key}
             </div>
-            <div className={clsx("text-size-base")}>{toolsInfo[key]}</div>
+            {toolsInfo[key]}
           </Fragment>
         );
       })}
@@ -275,19 +275,23 @@ interface ToolsProps {
  * Renders a list of tool components based on the provided tool definitions.
  */
 export const Tools: FC<ToolsProps> = ({ toolDefinitions }) => {
-  return toolDefinitions.map((toolDefinition, idx) => {
-    const toolName = toolDefinition.name;
-    const toolArgs = toolDefinition.parameters?.properties
-      ? Object.keys(toolDefinition.parameters.properties)
-      : [];
-    return (
-      <Tool
-        key={`${toolName}-${idx}`}
-        toolName={toolName}
-        toolArgs={toolArgs}
-      />
-    );
-  });
+  return (
+    <div className={styles.toolsGrid}>
+      {toolDefinitions.map((toolDefinition, idx) => {
+        const toolName = toolDefinition.name;
+        const toolArgs = toolDefinition.parameters?.properties
+          ? Object.keys(toolDefinition.parameters.properties)
+          : [];
+        return (
+          <Tool
+            key={`${toolName}-${idx}`}
+            toolName={toolName}
+            toolArgs={toolArgs}
+          />
+        );
+      })}
+    </div>
+  );
 };
 
 interface ToolProps {
@@ -304,10 +308,8 @@ export const Tool: FC<ToolProps> = ({ toolName, toolArgs }) => {
       ? `${toolName}(${toolArgs.join(", ")})`
       : toolName;
   return (
-    <div>
-      <code className={clsx("text-size-small", styles.tool)}>
-        {functionCall}
-      </code>
-    </div>
+    <code className={clsx("text-size-smallest", styles.tool)}>
+      {functionCall}
+    </code>
   );
 };

inspect_ai/_view/www/src/{samples → app/samples}/transcript/state/StateEventRenders.module.css

@@ -1,10 +1,17 @@
+.toolsGrid {
+  display: grid;
+  row-gap: 0em;
+}
+
 .tools {
   display: grid;
   grid-template-columns: max-content max-content;
   column-gap: 1rem;
   margin: 0;
+  align-items: baseline;
 }
 
 .tool {
   padding: 0;
+  color: inherit !important;
 }

inspect_ai/_view/www/src/{samples → app/samples}/transcript/state/StateEventView.tsx

@@ -7,8 +7,8 @@ import {
   Op,
   StateEvent,
   StoreEvent,
-} from "../../../types/log";
-import { formatDateTime } from "../../../utils/format";
+} from "../../../../@types/log";
+import { formatDateTime } from "../../../../utils/format";
 import { EventPanel } from "../event/EventPanel";
 import { StateDiffView } from "./StateDiffView";
 import {

inspect_ai/_view/www/src/app/samples/transcript/transform/fixups.ts

@@ -0,0 +1,142 @@
+// This is a special name that signals a group of sandbox events.
+
+import { Events, StepEvent } from "../../../../@types/log";
+
+// It will be caught elsewhere and rendered with a pretty name
+export const kSandboxSignalName = "53787D8A-D3FC-426D-B383-9F880B70E4AA";
+
+/**
+ * Normalizes event content
+ */
+export const fixupEventStream = (
+  events: Events,
+  filterPending: boolean = true,
+) => {
+  // We ignore pending events sometimes (when an eval is complete) and
+  // show them other times (when an eval is running)
+  const collapsed = processPendingEvents(events, filterPending);
+
+  // We need to inject a step event for sample_init if it doesn't exist
+  const fixedUp = collapseSampleInit(collapsed);
+
+  // Inject step events before and after groups of sandbox events
+  return groupSandboxEvents(fixedUp);
+};
+
+const processPendingEvents = (events: Events, filter: boolean): Events => {
+  // If filtering pending, just remove all pending events
+  // otherise, collapse sequential pending events of the same
+  // type
+  return filter
+    ? events.filter((e) => !e.pending)
+    : events.reduce<Events>((acc, event) => {
+        // Collapse sequential pending events of the same type
+        if (!event.pending) {
+          // Not a pending event
+          acc.push(event);
+        } else {
+          // For pending events, replace previous pending or add new
+          const lastIndex = acc.length - 1;
+          if (
+            lastIndex >= 0 &&
+            acc[lastIndex].pending &&
+            acc[lastIndex].event === event.event
+          ) {
+            // Replace previous pending with current one (if they're of the same type)
+            acc[lastIndex] = event;
+          } else {
+            // First event or follows non-pending
+            acc.push(event);
+          }
+        }
+        return acc;
+      }, []);
+};
+
+const collapseSampleInit = (events: Events): Events => {
+  // See if the events have an init step
+  const hasInitStep =
+    events.findIndex((e) => {
+      return e.event === "step" && e.name === "init";
+    }) !== -1;
+
+  const initEventIndex = events.findIndex((e) => {
+    return e.event === "sample_init";
+  });
+  const initEvent = events[initEventIndex];
+
+  const fixedUp = [...events];
+  if (!hasInitStep && initEvent) {
+    fixedUp.splice(initEventIndex, 0, {
+      timestamp: initEvent.timestamp,
+      event: "step",
+      action: "begin",
+      type: null,
+      name: "sample_init",
+      pending: false,
+      working_start: 0,
+    });
+
+    fixedUp.splice(initEventIndex + 2, 0, {
+      timestamp: initEvent.timestamp,
+      event: "step",
+      action: "end",
+      type: null,
+      name: "sample_init",
+      pending: false,
+      working_start: 0,
+    });
+  }
+  return fixedUp;
+};
+
+const groupSandboxEvents = (events: Events): Events => {
+  const result: Events = [];
+  const pendingSandboxEvents: Events = [];
+
+  const pushPendingSandboxEvents = () => {
+    const timestamp =
+      pendingSandboxEvents[pendingSandboxEvents.length - 1].timestamp;
+    result.push(createStepEvent(kSandboxSignalName, timestamp, "begin"));
+    result.push(...pendingSandboxEvents);
+    result.push(createStepEvent(kSandboxSignalName, timestamp, "end"));
+    pendingSandboxEvents.length = 0;
+  };
+
+  for (const event of events) {
+    if (event.event === "sandbox") {
+      // Collect sandbox events
+      pendingSandboxEvents.push(event);
+      continue;
+    }
+
+    // Process any collected sandbox events
+    if (pendingSandboxEvents.length > 0) {
+      pushPendingSandboxEvents();
+    }
+
+    // Clear sandbox events and add the current event
+    result.push(event);
+  }
+
+  // Handle any remaining sandbox events at the end
+  if (pendingSandboxEvents.length > 0) {
+    pushPendingSandboxEvents();
+  }
+
+  return result;
+};
+
+const createStepEvent = (
+  name: string,
+  timestamp: string,
+  action: "begin" | "end",
+): StepEvent => ({
+  timestamp,
+  event: "step",
+  action,
+  type: null,
+  name,
+  pending: false,
+  working_start: 0,
+});

inspect_ai/_view/www/src/app/samples/transcript/transform/treeify.ts

@@ -0,0 +1,39 @@
+import { Events } from "../../../../@types/log";
+import { EventNode, EventType } from "../types";
+
+/**
+ * Gathers events into a hierarchy of EventNodes.
+ */
+export function treeifyEvents(events: Events, depth: number): EventNode[] {
+  const rootNodes: EventNode[] = [];
+  const stack: EventNode[] = [];
+
+  const pushNode = (event: EventType): EventNode => {
+    const node = new EventNode(event, stack.length + depth);
+    if (stack.length > 0) {
+      const parentNode = stack[stack.length - 1];
+      parentNode.children.push(node);
+    } else {
+      rootNodes.push(node);
+    }
+    return node;
+  };
+
+  events.forEach((event) => {
+    if (event.event === "step" && event.action === "begin") {
+      // Starting a new step
+      const node = pushNode(event);
+      stack.push(node);
+    } else if (event.event === "step" && event.action === "end") {
+      // An ending step
+      if (stack.length > 0) {
+        stack.pop();
+      }
+    } else {
+      // An event
+      pushNode(event);
+    }
+  });
+
+  return rootNodes;
+}

inspect_ai/_view/www/src/{samples → app/samples}/transcript/types.ts

@@ -15,7 +15,7 @@ import {
   StoreEvent,
   SubtaskEvent,
   ToolEvent,
-} from "../../types/log";
+} from "../../../@types/log";
 
 export interface StateManager {
   scope: string;

inspect_ai/_view/www/src/{workspace → app}/sidebar/EvalStatus.tsx

@@ -1,6 +1,6 @@
 import clsx from "clsx";
 import { FC } from "react";
-import { EvalLogHeader } from "../../api/types";
+import { EvalLogHeader } from "../../client/api/types";
 import styles from "./EvalStatus.module.css";
 import { SidebarScoreView } from "./SidebarScoreView";
 import { SidebarScoresView } from "./SidebarScoresView";

inspect_ai/_view/www/src/app/sidebar/LogDirectoryTitleView.module.css

@@ -0,0 +1,16 @@
+.dirname {
+  overflow: hidden;
+  white-space: nowrap;
+  text-overflow: ellipsis;
+}
+
+.directoryLink {
+  color: inherit;
+  text-decoration: none;
+  cursor: pointer;
+}
+
+.directoryLink:hover {
+  color: var(--bs-primary);
+  text-decoration: none;
+}

inspect_ai/_view/www/src/app/sidebar/LogDirectoryTitleView.tsx

@@ -0,0 +1,70 @@
+import clsx from "clsx";
+import { FC, useCallback } from "react";
+import { Link } from "react-router-dom";
+import { useStore } from "../../state/store";
+import styles from "./LogDirectoryTitleView.module.css";
+
+interface LogDirectoryTitleViewProps {
+  log_dir?: string;
+}
+
+export const LogDirectoryTitleView: FC<LogDirectoryTitleViewProps> = ({
+  log_dir,
+}) => {
+  const offCanvas = useStore((state) => state.app.offcanvas);
+  const setOffCanvas = useStore((state) => state.appActions.setOffcanvas);
+
+  const handleClick = useCallback(() => {
+    // Close the sidebar when clicking the directory link on mobile
+    if (offCanvas) {
+      setOffCanvas(false);
+    }
+  }, [offCanvas, setOffCanvas]);
+  if (log_dir) {
+    const displayDir = prettyDir(log_dir);
+    return (
+      <Link to="/logs" className={styles.directoryLink} onClick={handleClick}>
+        <div style={{ display: "flex", flexDirection: "column" }}>
+          <span
+            className={clsx(
+              "text-style-secondary",
+              "text-style-label",
+              "text-size-small",
+            )}
+          >
+            Log Directory
+          </span>
+          <span
+            title={displayDir}
+            className={clsx("text-size-base", styles.dirname)}
+          >
+            {offCanvas ? displayDir : ""}
+          </span>
+        </div>
+      </Link>
+    );
+  } else {
+    return (
+      <Link to="/logs" className={styles.directoryLink} onClick={handleClick}>
+        <span className={clsx("text-size-title")}>
+          {offCanvas ? "Log History" : ""}
+        </span>
+      </Link>
+    );
+  }
+};
+
+const prettyDir = (path: string): string => {
+  try {
+    // Try to create a new URL object
+    let url = new URL(path);
+
+    if (url.protocol === "file:") {
+      return url.pathname;
+    } else {
+      return path;
+    }
+  } catch {
+    return path;
+  }
+};

inspect_ai/_view/www/src/{workspace → app}/sidebar/Sidebar.module.css

@@ -66,4 +66,12 @@
 
 .item {
   cursor: pointer;
+  padding: 0;
+}
+
+.logLink {
+  display: block;
+  padding: 0.5rem 1rem;
+  color: inherit;
+  text-decoration: none;
 }

inspect_ai/_view/www/src/{workspace → app}/sidebar/Sidebar.tsx

@@ -1,17 +1,18 @@
 import clsx from "clsx";
-import { FC, MouseEvent, useCallback, useRef } from "react";
+import { FC, useCallback, useEffect, useRef } from "react";
+import { Link } from "react-router-dom";
 import { Fragment } from "react/jsx-runtime";
-import { EvalLogHeader, LogFiles } from "../../api/types";
-import { ApplicationIcons } from "../../appearance/icons";
+import { EvalLogHeader } from "../../client/api/types";
 import { ProgressBar } from "../../components/ProgressBar";
 import { useStatefulScrollPosition } from "../../state/scrolling";
 import { useStore } from "../../state/store";
+import { ApplicationIcons } from "../appearance/icons";
+import { logUrl } from "../routing/url";
 import { LogDirectoryTitleView } from "./LogDirectoryTitleView";
 import styles from "./Sidebar.module.css";
 import { SidebarLogEntry } from "./SidebarLogEntry";
 
 interface SidebarProps {
-  logs: LogFiles;
   logHeaders: Record<string, EvalLogHeader>;
   loading: boolean;
   selectedIndex: number;
@@ -19,12 +20,12 @@ interface SidebarProps {
 }
 
 export const Sidebar: FC<SidebarProps> = ({
-  logs,
   logHeaders,
   loading,
   selectedIndex,
   onSelectedIndexChanged,
 }) => {
+  const logs = useStore((state) => state.logs.logs);
   const setOffCanvas = useStore((state) => state.appActions.setOffcanvas);
   const offCanvas = useStore((state) => state.app.offcanvas);
   const handleToggle = useCallback(() => {
@@ -34,13 +35,19 @@ export const Sidebar: FC<SidebarProps> = ({
   const sidebarContentsRef = useRef(null);
   useStatefulScrollPosition(sidebarContentsRef, "sidebar-contents", 1000);
 
-  const handleClick = useCallback(
-    (e: MouseEvent<HTMLLIElement>) => {
-      const index = Number((e.currentTarget as HTMLLIElement).dataset.index);
-      onSelectedIndexChanged(index);
-    },
-    [onSelectedIndexChanged],
-  );
+  // Scroll the selected log into view when it changes
+  const itemRefs = useRef<{ [index: number]: HTMLLIElement | null }>({});
+
+  useEffect(() => {
+    if (itemRefs.current[selectedIndex]) {
+      itemRefs.current[selectedIndex]?.scrollIntoView({
+        behavior: "smooth",
+        block: "nearest",
+      });
+    }
+  }, [selectedIndex]);
+
+  // No longer need the click handler as we're using Links now
 
   return (
     <Fragment>
@@ -78,6 +85,9 @@ export const Sidebar: FC<SidebarProps> = ({
             return (
               <li
                 key={file.name}
+                ref={(el) => {
+                  itemRefs.current[index] = el;
+                }}
                 className={clsx(
                   "list-group-item",
                   "list-group-item-action",
@@ -85,12 +95,20 @@ export const Sidebar: FC<SidebarProps> = ({
                   selectedIndex === index ? styles.active : undefined,
                 )}
                 data-index={index}
-                onClick={handleClick}
               >
-                <SidebarLogEntry
-                  logHeader={logHeader}
-                  task={file.task || "unknown task"}
-                />
+                <Link
+                  to={logUrl(file.name, logs.log_dir)}
+                  className={styles.logLink}
+                  onClick={() => {
+                    // Also update the current index in state
+                    onSelectedIndexChanged(index);
+                  }}
+                >
+                  <SidebarLogEntry
+                    logHeader={logHeader}
+                    task={file.task || "unknown task"}
+                  />
+                </Link>
               </li>
             );
           })}

inspect_ai/_view/www/src/{workspace → app}/sidebar/SidebarLogEntry.tsx

@@ -1,6 +1,6 @@
 import clsx from "clsx";
 import { FC, Fragment } from "react";
-import { EvalLogHeader } from "../../api/types";
+import { EvalLogHeader } from "../../client/api/types";
 import { kModelNone } from "../../constants";
 import { EvalStatus } from "./EvalStatus";
 import styles from "./SidebarLogEntry.module.css";

inspect_ai/_view/www/src/{workspace → app}/sidebar/SidebarScoreView.tsx

@@ -1,9 +1,9 @@
 import clsx from "clsx";
-import { EvalScore } from "../../types/log";
+import { EvalScore } from "../../@types/log";
 import { formatPrettyDecimal } from "../../utils/format";
 
 import { FC } from "react";
-import { metricDisplayName } from "../utils";
+import { metricDisplayName } from "../log-view/utils";
 import styles from "./SidebarScoreView.module.css";
 interface SidebarScoreProps {
   scorer: EvalScore;

inspect_ai/_view/www/src/{workspace → app}/sidebar/SidebarScoresView.tsx

@@ -1,8 +1,8 @@
 import clsx from "clsx";
 import { FC, Fragment } from "react";
-import { Scores } from "../../types/log";
+import { Scores } from "../../@types/log";
 import { formatPrettyDecimal } from "../../utils/format";
-import { metricDisplayName } from "../utils";
+import { metricDisplayName } from "../log-view/utils";
 import styles from "./SidebarScoresView.module.css";
 
 interface SidebarScoresProps {

inspect_ai/_view/www/src/{types.ts → app/types.ts}

@@ -1,14 +1,4 @@
 import { StateSnapshot } from "react-virtuoso";
-import {
-  AttachmentData,
-  EvalLogHeader,
-  EvalSummary,
-  EventData,
-  LogFiles,
-  PendingSamples,
-  SampleSummary,
-} from "./api/types";
-import { ScorerInfo } from "./scoring/utils";
 import {
   ApprovalEvent,
   ContentImage,
@@ -26,7 +16,17 @@ import {
   StoreEvent,
   SubtaskEvent,
   ToolEvent,
-} from "./types/log";
+} from "../@types/log";
+import {
+  AttachmentData,
+  EvalLogHeader,
+  EvalSummary,
+  EventData,
+  LogFiles,
+  PendingSamples,
+  SampleSummary,
+} from "../client/api/types";
+import { ScorerInfo } from "../state/scoring";
 
 export interface AppState {
   status: AppStatus;
@@ -44,6 +44,13 @@ export interface AppState {
   collapsed: Record<string, boolean>;
   messages: Record<string, boolean>;
   propertyBags: Record<string, Record<string, unknown>>;
+  urlHash?: string;
+  initialState?: {
+    log: string;
+    sample_id?: string;
+    sample_epoch?: string;
+  };
+  rehydrated?: boolean;
 }
 
 export interface LogsState {

inspect_ai/_view/www/src/{usage → app/usage}/ModelTokenTable.tsx

@@ -1,5 +1,5 @@
 import { FC } from "react";
-import { ModelUsage, ModelUsage2 } from "../types/log";
+import { ModelUsage, ModelUsage2 } from "../../@types/log";
 import { TokenHeader, TokenRow, TokenTable } from "./TokenTable";
 
 interface ModelTokenTableProps {

inspect_ai/_view/www/src/{usage → app/usage}/ModelUsagePanel.tsx

@@ -1,7 +1,7 @@
 import clsx from "clsx";
 import { FC, Fragment } from "react";
-import { ModelUsage1 } from "../types/log";
-import { formatNumber } from "../utils/format";
+import { ModelUsage1 } from "../../@types/log";
+import { formatNumber } from "../../utils/format";
 import styles from "./ModelUsagePanel.module.css";
 
 interface ModelUsageProps {

inspect_ai/_view/www/src/{usage → app/usage}/TokenTable.tsx

@@ -1,6 +1,6 @@
 import clsx from "clsx";
 import { FC, ReactNode } from "react";
-import { ModelUsage1 } from "../types/log";
+import { ModelUsage1 } from "../../@types/log";
 import { ModelUsagePanel } from "./ModelUsagePanel";
 import styles from "./TokenTable.module.css";
 

inspect_ai/_view/www/src/{usage → app/usage}/UsageCard.tsx

@@ -1,10 +1,10 @@
 import clsx from "clsx";
-import { FontSize } from "../appearance/fonts";
-import { ApplicationIcons } from "../appearance/icons";
-import { Card, CardBody, CardHeader } from "../components/Card";
-import { MetaDataView } from "../metadata/MetaDataView";
-import { EvalStats } from "../types/log";
-import { formatDuration } from "../utils/format";
+import { EvalStats } from "../../@types/log";
+import { FontSize } from "../../app/appearance/fonts";
+import { ApplicationIcons } from "../../app/appearance/icons";
+import { MetaDataView } from "../../app/content/MetaDataView";
+import { Card, CardBody, CardHeader } from "../../components/Card";
+import { formatDuration } from "../../utils/format";
 import { ModelTokenTable } from "./ModelTokenTable";
 
 import { FC } from "react";

inspect_ai/_view/www/src/{api → client/api}/api-browser.ts

@@ -1,4 +1,4 @@
-import { asyncJsonParse } from "../utils/json-worker";
+import { asyncJsonParse } from "../../utils/json-worker";
 import { download_file } from "./api-shared";
 import {
   Capabilities,
@@ -244,7 +244,7 @@ async function api(
     };
   } else if (response.status !== 200) {
     const message = (await response.text()) || response.statusText;
-    const error = new Error(`Error: ${response.status}: ${message})`);
+    const error = new Error(`${message}`);
     throw error;
   } else {
     throw new Error(`${response.status} - ${response.statusText} `);

inspect_ai/_view/www/src/{api → client/api}/api-http.ts

@@ -1,6 +1,6 @@
-import { fetchRange, fetchSize } from "../logfile/remoteZipFile";
-import { EvalLog } from "../types/log";
-import { asyncJsonParse } from "../utils/json-worker";
+import { EvalLog } from "../../@types/log";
+import { asyncJsonParse } from "../../utils/json-worker";
+import { fetchRange, fetchSize } from "../remote/remoteZipFile";
 import { download_file, encodePathParts } from "./api-shared";
 import {
   Capabilities,

inspect_ai/_view/www/src/{api → client/api}/api-vscode.ts

@@ -1,7 +1,7 @@
 import JSON5 from "json5";
-import { asyncJsonParse } from "../utils/json-worker";
+import { asyncJsonParse } from "../../utils/json-worker";
 
-import { getVscodeApi } from "../utils/vscode";
+import { getVscodeApi } from "../../utils/vscode";
 import {
   kMethodEvalLog,
   kMethodEvalLogBytes,