inspect-ai 0.3.81__py3-none-any.whl → 0.3.83__py3-none-any.whl

inspect_ai/_view/www/src/types/log.d.ts

@@ -13,6 +13,7 @@ export type Task = string;
 export type TaskId = string;
 export type TaskVersion = number;
 export type TaskFile = string | null;
+export type TaskRegistryName = string | null;
 export type Solver = string | null;
 export type SolverArgs = {} | null;
 export type Tags = string[] | null;
@@ -44,6 +45,7 @@ export type SandboxCleanup = boolean | null;
 export type LogSamples = boolean | null;
 export type LogImages = boolean | null;
 export type LogBuffer = number | null;
+export type LogShared = number | null;
 export type ScoreDisplay = boolean | null;
 export type Type1 = "git";
 export type Origin = string;
@@ -160,6 +162,7 @@ export type Content =
     )[];
 export type Type3 = "text";
 export type Text = string;
+export type Refusal = boolean | null;
 export type Type4 = "reasoning";
 export type Reasoning = string;
 export type Signature = string | null;
@@ -203,12 +206,11 @@ export type Role2 = "assistant";
 export type ToolCalls = ToolCall[] | null;
 export type Id4 = string;
 export type Function = string;
-export type Type8 = string;
-export type InternalName = string | null;
 export type ParseError = string | null;
 export type Title = string | null;
 export type Format2 = "text" | "markdown";
 export type Content3 = string;
+export type Model1 = string | null;
 export type Id5 = string | null;
 export type Content4 =
   | string
@@ -223,8 +225,7 @@ export type Source3 = ("input" | "generate") | null;
 export type Role3 = "tool";
 export type ToolCallId1 = string | null;
 export type Function1 = string | null;
-export type InternalName1 = string | null;
-export type Type9 =
+export type Type8 =
   | "parsing"
   | "timeout"
   | "unicode_decode"
@@ -245,7 +246,7 @@ export type Messages = (
   | ChatMessageAssistant
   | ChatMessageTool
 )[];
-export type Model1 = string;
+export type Model2 = string;
 export type StopReason =
   | "stop"
   | "max_tokens"
@@ -304,7 +305,7 @@ export type Timestamp1 = string;
 export type WorkingStart1 = number;
 export type Pending1 = boolean | null;
 export type Event1 = "sample_limit";
-export type Type10 =
+export type Type9 =
   | "message"
   | "time"
   | "working"
@@ -344,7 +345,7 @@ export type Timestamp5 = string;
 export type WorkingStart5 = number;
 export type Pending5 = boolean | null;
 export type Event5 = "model";
-export type Model2 = string;
+export type Model3 = string;
 export type Input3 = (
   | ChatMessageSystem
   | ChatMessageUser
@@ -353,7 +354,7 @@ export type Input3 = (
 )[];
 export type Name8 = string;
 export type Description2 = string;
-export type Type11 = "object";
+export type Type10 = "object";
 export type Required1 = string[];
 export type Additionalproperties1 = boolean;
 export type Tools1 = ToolInfo[];
@@ -368,10 +369,9 @@ export type Timestamp6 = string;
 export type WorkingStart6 = number;
 export type Pending6 = boolean | null;
 export type Event6 = "tool";
-export type Type12 = "function";
+export type Type11 = "function";
 export type Id7 = string;
 export type Function2 = string;
-export type InternalName2 = string | null;
 export type Result1 =
   | string
   | number
@@ -447,14 +447,14 @@ export type WorkingStart13 = number;
 export type Pending13 = boolean | null;
 export type Event13 = "step";
 export type Action1 = "begin" | "end";
-export type Type13 = string | null;
+export type Type12 = string | null;
 export type Name11 = string;
 export type Timestamp14 = string;
 export type WorkingStart14 = number;
 export type Pending14 = boolean | null;
 export type Event14 = "subtask";
 export type Name12 = string;
-export type Type14 = string | null;
+export type Type13 = string | null;
 export type Events2 = (
   | SampleInitEvent
   | SampleLimitEvent
@@ -493,6 +493,8 @@ export type Events1 = (
 )[];
 export type Completed3 = string | null;
 export type WorkingTime2 = number | null;
+export type Agent = string | null;
+export type Failed = boolean | null;
 export type Events = (
   | SampleInitEvent
   | SampleLimitEvent
@@ -513,7 +515,7 @@ export type Events = (
 export type TotalTime = number | null;
 export type WorkingTime3 = number | null;
 export type Uuid = string | null;
-export type Type15 =
+export type Type14 =
   | "context"
   | "time"
   | "working"
@@ -565,6 +567,7 @@ export interface EvalSpec {
   task_id: TaskId;
   task_version: TaskVersion;
   task_file: TaskFile;
+  task_registry_name: TaskRegistryName;
   task_attribs: TaskAttribs;
   task_args: TaskArgs;
   solver: Solver;
@@ -627,6 +630,7 @@ export interface EvalConfig {
   log_samples: LogSamples;
   log_images: LogImages;
   log_buffer: LogBuffer;
+  log_shared: LogShared;
   score_display: ScoreDisplay;
 }
 export interface ApprovalPolicyConfig {
@@ -845,6 +849,7 @@ export interface ChatMessageSystem {
   id: Id1;
   content: Content;
   source: Source;
+  internal: unknown;
   role: Role;
 }
 /**
@@ -853,6 +858,7 @@ export interface ChatMessageSystem {
 export interface ContentText {
   type: Type3;
   text: Text;
+  refusal: Refusal;
 }
 /**
  * Reasoning content.
@@ -896,6 +902,7 @@ export interface ChatMessageUser {
   id: Id2;
   content: Content1;
   source: Source1;
+  internal: unknown;
   role: Role1;
   tool_call_id: ToolCallId;
 }
@@ -906,15 +913,16 @@ export interface ChatMessageAssistant {
   id: Id3;
   content: Content2;
   source: Source2;
+  internal: unknown;
   role: Role2;
   tool_calls: ToolCalls;
+  model: Model1;
 }
 export interface ToolCall {
   id: Id4;
   function: Function;
   arguments: Arguments;
-  type: Type8;
-  internal_name: InternalName;
+  internal: unknown;
   parse_error: ParseError;
   view: ToolCallContent | null;
 }
@@ -934,21 +942,21 @@ export interface ChatMessageTool {
   id: Id5;
   content: Content4;
   source: Source3;
+  internal: unknown;
   role: Role3;
   tool_call_id: ToolCallId1;
   function: Function1;
-  internal_name: InternalName1;
   error: ToolCallError | null;
 }
 export interface ToolCallError {
-  type: Type9;
+  type: Type8;
   message: Message1;
 }
 /**
  * Output from model generation.
  */
 export interface ModelOutput {
-  model: Model1;
+  model: Model2;
   choices: Choices1;
   usage: ModelUsage1 | null;
   time: Time;
@@ -1029,7 +1037,7 @@ export interface SampleLimitEvent {
   working_start: WorkingStart1;
   pending: Pending1;
   event: Event1;
-  type: Type10;
+  type: Type9;
   message: Message2;
   limit: Limit1;
 }
@@ -1092,7 +1100,7 @@ export interface ModelEvent {
   working_start: WorkingStart5;
   pending: Pending5;
   event: Event5;
-  model: Model2;
+  model: Model3;
   input: Input3;
   tools: Tools1;
   tool_choice: ToolChoice;
@@ -1139,7 +1147,7 @@ export interface ToolInfo {
  * Description of tool parameters object in JSON Schema format.
  */
 export interface ToolParams {
-  type: Type11;
+  type: Type10;
   properties: Properties1;
   required: Required1;
   additionalProperties: Additionalproperties1;
@@ -1202,11 +1210,11 @@ export interface ToolEvent {
   working_start: WorkingStart6;
   pending: Pending6;
   event: Event6;
-  type: Type12;
+  type: Type11;
   id: Id7;
   function: Function2;
   arguments: Arguments1;
-  internal_name: InternalName2;
+  internal: unknown;
   view: ToolCallContent | null;
   result: Result1;
   truncated: Truncated;
@@ -1214,6 +1222,8 @@ export interface ToolEvent {
   events: Events1;
   completed: Completed3;
   working_time: WorkingTime2;
+  agent: Agent;
+  failed: Failed;
 }
 export interface Arguments1 {
   [k: string]: JsonValue;
@@ -1322,7 +1332,7 @@ export interface StepEvent {
   pending: Pending13;
   event: Event13;
   action: Action1;
-  type: Type13;
+  type: Type12;
   name: Name11;
 }
 /**
@@ -1334,7 +1344,7 @@ export interface SubtaskEvent {
   pending: Pending14;
   event: Event14;
   name: Name12;
-  type: Type14;
+  type: Type13;
   input: Input5;
   result: Result2;
   events: Events2;
@@ -1355,7 +1365,7 @@ export interface Attachments {
  * Limit encontered by sample.
  */
 export interface EvalSampleLimit {
-  type: Type15;
+  type: Type14;
   limit: Limit2;
 }
 /**

inspect_ai/_view/www/src/types/markdown-it-katex.d.ts

@@ -0,0 +1,21 @@
+declare module "markdown-it-katex" {
+  import MarkdownIt from "markdown-it";
+
+  interface KatexOptions {
+    throwOnError?: boolean;
+    errorColor?: string;
+    macros?: Record<string, string>;
+    fleqn?: boolean;
+    trust?: boolean;
+    output?: "html" | "htmlAndMathml" | "mathml";
+    minRuleThickness?: number;
+    colorIsTextColor?: boolean;
+    maxSize?: number;
+    maxExpand?: number;
+    strict?: boolean | string | Function;
+  }
+
+  const markdownItKatex: (md: MarkdownIt, options?: KatexOptions) => void;
+
+  export default markdownItKatex;
+}

inspect_ai/_view/www/src/types.ts

@@ -1,49 +1,105 @@
+import { StateSnapshot } from "react-virtuoso";
 import {
+  AttachmentData,
   EvalLogHeader,
   EvalSummary,
+  EventData,
   LogFiles,
+  PendingSamples,
   SampleSummary,
 } from "./api/types";
-import { ContentImage, ContentText, EvalSample } from "./types/log";
-
-export interface ApplicationState {
-  logs?: LogFiles;
-  selectedLogIndex?: number;
-  logHeaders?: Record<string, EvalLogHeader>;
-  headersLoading?: boolean;
-  selectedLog?: CurrentLog;
-  selectedWorkspaceTab?: string;
-  selectedSampleIndex?: number;
-  selectedSample?: EvalSample;
-  sampleStatus?: "loading" | "ok" | "error";
-  sampleError?: Error;
-  selectedSampleTab?: string;
-  sampleScrollPosition?: number;
-  showingSampleDialog?: boolean;
-  status?: AppStatus;
-  offcanvas?: boolean;
-  showFind?: boolean;
-  filter?: ScoreFilter;
-  epoch?: string;
-  sort?: string;
-  scores?: ScoreLabel[];
+import { ScorerInfo } from "./scoring/utils";
+import {
+  ApprovalEvent,
+  ContentImage,
+  ContentText,
+  EvalSample,
+  InfoEvent,
+  LoggerEvent,
+  ModelEvent,
+  SampleInitEvent,
+  SampleLimitEvent,
+  SandboxEvent,
+  ScoreEvent,
+  StateEvent,
+  StepEvent,
+  StoreEvent,
+  SubtaskEvent,
+  ToolEvent,
+} from "./types/log";
+
+export interface AppState {
+  status: AppStatus;
+  offcanvas: boolean;
+  showFind: boolean;
+  tabs: {
+    workspace: string;
+    sample: string;
+  };
+  dialogs: {
+    sample: boolean;
+  };
+  scrollPositions: Record<string, number>;
+  listPositions: Record<string, StateSnapshot>;
+  collapsed: Record<string, boolean>;
+  messages: Record<string, boolean>;
+  propertyBags: Record<string, Record<string, unknown>>;
+}
+
+export interface LogsState {
+  logs: LogFiles;
+  logHeaders: Record<string, EvalLogHeader>;
+  headersLoading: boolean;
+  selectedLogIndex: number;
+}
+
+export interface LogState {
+  loadedLog?: string;
+
+  selectedSampleIndex: number;
+  selectedLogSummary?: EvalSummary;
+  pendingSampleSummaries?: PendingSamples;
+
+  filter: ScoreFilter;
+  epoch: string;
+  sort: string;
   score?: ScoreLabel;
-  filteredSamples?: SampleSummary[];
-  groupBy?: "none" | "epoch" | "sample";
-  groupByOrder?: "asc" | "desc";
-  workspaceTabScrollPosition?: Record<string, number>;
+  scores?: ScorerInfo[];
 }
 
+export type SampleStatus = "ok" | "loading" | "streaming" | "error";
+
+export interface SampleState {
+  selectedSample: EvalSample | undefined;
+  sampleStatus: SampleStatus;
+  sampleError: Error | undefined;
+
+  // Events and attachments
+  runningEvents: Event[];
+}
+
+export type Event =
+  | SampleInitEvent
+  | SampleLimitEvent
+  | SandboxEvent
+  | StateEvent
+  | StoreEvent
+  | ModelEvent
+  | ToolEvent
+  | ApprovalEvent
+  | InputEvent
+  | ScoreEvent
+  | ErrorEvent
+  | LoggerEvent
+  | InfoEvent
+  | StepEvent
+  | SubtaskEvent;
+
 export interface AppStatus {
   loading: boolean;
   error?: Error;
 }
 
-export interface Capabilities {
-  downloadFiles: boolean;
-  webWorkers: boolean;
-}
-
 export interface CurrentLog {
   name: string;
   contents: EvalSummary;
@@ -69,3 +125,9 @@ export interface ContentTool {
   type: "tool";
   content: (ContentImage | ContentText)[];
 }
+
+export interface RunningSampleData {
+  events: Map<string, EventData>;
+  attachments: Map<string, AttachmentData>;
+  summary?: SampleSummary;
+}

inspect_ai/_view/www/src/utils/attachments.ts

@@ -1,40 +1,75 @@
-/**
- * Resolves individual value by replacing protocol references with attachment content
- */
-export const resolveAttachments = (
-  value: any,
+export const resolveAttachments = <T>(
+  value: T,
   attachments: Record<string, string>,
-): any => {
-  const kContentProtocol = "tc://";
-  const kAttachmentProtocol = "attachment://";
+): T => {
+  const CONTENT_PROTOCOL = "tc://";
+  const ATTACHMENT_PROTOCOL = "attachment://";
+
+  // Handle null or undefined early
+  if (value === null || value === undefined) {
+    return value;
+  }
 
   // Handle arrays recursively
   if (Array.isArray(value)) {
-    return value.map((v) => resolveAttachments(v, attachments));
+    let hasChanged = false;
+    const resolvedArray = value.map((v) => {
+      const resolved = resolveAttachments(v, attachments);
+      if (resolved !== v) hasChanged = true;
+      return resolved;
+    });
+
+    // Only return the new array if something actually changed
+    return hasChanged ? (resolvedArray as unknown as T) : value;
   }
 
-  // Handle objects recursively
-  if (value && typeof value === "object") {
+  // Handle objects recursively, but skip Date instances and other special object types
+  if (
+    typeof value === "object" &&
+    !(value instanceof Date) &&
+    !(value instanceof RegExp)
+  ) {
+    let hasChanged = false;
     const resolvedObject: Record<string, unknown> = {};
-    for (const key of Object.keys(value)) {
-      resolvedObject[key] = resolveAttachments(value[key], attachments);
+
+    for (const [key, val] of Object.entries(value)) {
+      const resolved = resolveAttachments(val, attachments);
+      resolvedObject[key] = resolved;
+
+      // Track if anything changed
+      if (resolved !== val) hasChanged = true;
     }
-    return resolvedObject;
+
+    // Only return the new object if something actually changed
+    return hasChanged ? (resolvedObject as unknown as T) : value;
   }
 
   // Handle string values with protocol references
   if (typeof value === "string") {
-    let resolvedValue = value;
-    if (resolvedValue.startsWith(kContentProtocol)) {
-      resolvedValue = resolvedValue.replace(
-        kContentProtocol,
-        kAttachmentProtocol,
-      );
+    // Check if the string starts with the content protocol
+    if (value.startsWith(CONTENT_PROTOCOL)) {
+      const updatedValue = value.replace(CONTENT_PROTOCOL, ATTACHMENT_PROTOCOL);
+
+      // Now check if it's an attachment reference
+      if (updatedValue.startsWith(ATTACHMENT_PROTOCOL)) {
+        const attachmentId = updatedValue.slice(ATTACHMENT_PROTOCOL.length);
+        const attachment = attachments[attachmentId];
+
+        // Return the attachment content if it exists, otherwise return the original string
+        return (attachment !== undefined ? attachment : value) as unknown as T;
+      }
+
+      return updatedValue as unknown as T;
     }
-    if (resolvedValue.startsWith(kAttachmentProtocol)) {
-      return attachments[resolvedValue.replace(kAttachmentProtocol, "")];
+
+    // Check if it's directly an attachment reference
+    if (value.startsWith(ATTACHMENT_PROTOCOL)) {
+      const attachmentId = value.slice(ATTACHMENT_PROTOCOL.length);
+      const attachment = attachments[attachmentId];
+
+      // Return the attachment content if it exists, otherwise return the original string
+      return (attachment !== undefined ? attachment : value) as unknown as T;
     }
-    return resolvedValue;
   }
 
   // Return unchanged for other types

inspect_ai/_view/www/src/utils/json-worker.ts

@@ -1,43 +1,110 @@
 export const asyncJsonParse = async (text: string): Promise<any> => {
+  // Encode the input text
   const encoder = new TextEncoder();
   const encodedText = encoder.encode(text);
+
+  // Create a worker from the inline script
   const blob = new Blob([kWorkerCode], { type: "application/javascript" });
   const blobURL = URL.createObjectURL(blob);
   const worker = new Worker(blobURL);
+
   try {
     const result = new Promise((resolve, reject) => {
       worker.onmessage = function (e) {
         if (e.data.success) {
-          resolve(e.data.result);
+          if (e.data.serialized) {
+            // Deserialize the result if it was sent as a transferable
+            const decoder = new TextDecoder();
+            const resultString = decoder.decode(e.data.result);
+            resolve(JSON.parse(resultString));
+          } else {
+            resolve(e.data.result);
+          }
         } else {
-          reject(new Error(e.data.error));
+          const error = new Error(e.data.error);
+          if (e.data.stack) {
+            error.stack = e.data.stack;
+          }
+          reject(error);
         }
       };
+
       worker.onerror = function (error) {
-        reject(new Error(error.message));
+        reject(new Error(`Worker error: ${error.message}`));
       };
     });
-    worker.postMessage({ scriptContent: kJson5ScriptBase64, encodedText }, [
-      encodedText.buffer,
-    ]);
+
+    // Transfer the encoded text buffer to the worker
+    worker.postMessage(
+      {
+        scriptContent: kJson5ScriptBase64,
+        encodedText,
+      },
+      [encodedText.buffer],
+    );
+
     return await result;
   } finally {
+    // Clean up resources
     worker.terminate();
     URL.revokeObjectURL(blobURL);
   }
 };
 
 const kWorkerCode = `
+// Store the JSON5 parser once loaded
+let JSON5 = null;
+
 self.onmessage = function (e) {
-  eval(atob(e.data.scriptContent));
-  const { encodedText } = e.data;
-  const decoder = new TextDecoder();
-  const text = decoder.decode(encodedText);
+  const { encodedText, scriptContent } = e.data;
+  
   try {
+    // Only load the JSON5 script if we haven't done so yet
+    if (!JSON5) {
+      const script = atob(scriptContent);
+
+      new Function(script)();
+      // Verify it was loaded properly
+      if (typeof self.JSON5 !== 'object' || typeof self.JSON5.parse !== 'function') {
+        throw new Error('Failed to initialize JSON5 parser');
+      }
+      JSON5 = self.JSON5;
+    }
+    
+    // Decode the text using TextDecoder
+    const decoder = new TextDecoder();
+    const text = decoder.decode(encodedText);
+    
+    // Parse with JSON5
     const result = JSON5.parse(text);
-    postMessage({ success: true, result });
+    
+    if (result && typeof result === 'object' && 
+        (Array.isArray(result) ? result.length > 10000 : Object.keys(result).length > 10000)) {
+      
+      // Large result, use transferrable object
+      const resultString = JSON.stringify(result);
+      const encoder = new TextEncoder();
+      const serialized = encoder.encode(resultString);
+      
+      postMessage({
+        success: true, 
+        serialized: true,
+        result: serialized
+      }, [serialized.buffer]);
+    } else {
+      // Small results, send directly
+      postMessage({ 
+        success: true, 
+        serialized: false, 
+        result: result 
+      });
+    }
   } catch (err) {
-    postMessage({ success: false, error: err.message });
+    postMessage({ 
+      success: false, 
+      error: err.message,
+      stack: err.stack || ''
+    });
   }
 };`;