perf: replace remark-based search with plain text indexOf

Manually ported from upstream 5c7f921e. Key changes:
- SessionSearcher: indexOf instead of remark AST, batch size 8→16
- conversationSlice: indexOf with MAX_SEARCH_MATCHES=500 cap
- Item-scoped store selectors (searchMatchItemIds Set) to skip re-renders
- Pre-filter in markdownTextSearch (skip parse if no raw match)
- SearchTextCache: 200→1000 entries
- ProjectScanner: 30s search project cache, batch 4→8

src/main/services/discovery/ProjectScanner.ts

@@ -71,6 +71,9 @@ import type { FileSystemProvider, FsDirent } from '../infrastructure/FileSystemP
 
 const logger = createLogger('Discovery:ProjectScanner');
 
+/** How long to reuse the cached project list for search (ms) */
+const SEARCH_PROJECT_CACHE_TTL_MS = 30_000;
+
 // IPC payload safety: session ID arrays can be extremely large for long-lived projects.
 // Keep counts accurate via totalSessions, but truncate ID lists to keep renderer responsive.
 // Keep this non-zero because parts of the renderer still rely on a (partial) sessionId list
@@ -147,6 +150,9 @@ export class ProjectScanner {
   private scanCache: { projects: Project[]; timestamp: number } | null = null;
   private static readonly SCAN_CACHE_TTL_MS = 2000;
 
+  /** Cached project list for search — avoids re-scanning disk on every query */
+  private searchProjectCache: { projects: Project[]; timestamp: number } | null = null;
+
   // Platform-aware batch sizes to avoid UV thread pool saturation on Windows
   private static readonly LOCAL_SESSION_BATCH = process.platform === 'win32' ? 16 : 64;
   private static readonly LOCAL_PROJECT_BATCH = process.platform === 'win32' ? 4 : 12;
@@ -1329,8 +1335,17 @@ export class ProjectScanner {
         return { results: [], totalMatches: 0, sessionsSearched: 0, query };
       }
 
-      // Get all projects
-      const projects = await this.scan();
+      // Use cached project list to avoid re-scanning disk on every keystroke
+      let projects: Project[];
+      if (
+        this.searchProjectCache &&
+        Date.now() - this.searchProjectCache.timestamp < SEARCH_PROJECT_CACHE_TTL_MS
+      ) {
+        projects = this.searchProjectCache.projects;
+      } else {
+        projects = await this.scan();
+        this.searchProjectCache = { projects, timestamp: Date.now() };
+      }
 
       if (projects.length === 0) {
         return { results: [], totalMatches: 0, sessionsSearched: 0, query };
@@ -1338,7 +1353,7 @@ export class ProjectScanner {
 
       // Search across all projects with bounded concurrency
       const allResults: SearchSessionsResult[] = [];
-      const searchBatchSize = this.fsProvider.type === 'ssh' ? 2 : 4;
+      const searchBatchSize = this.fsProvider.type === 'ssh' ? 2 : 8;
 
       for (let i = 0; i < projects.length; i += searchBatchSize) {
         const batch = projects.slice(i, i + searchBatchSize);

src/main/services/discovery/SearchTextCache.ts

@@ -21,7 +21,7 @@ export class SearchTextCache {
   private readonly cache = new Map<string, CacheEntry>();
   private readonly maxSize: number;
 
-  constructor(maxSize: number = 200) {
+  constructor(maxSize: number = 1000) {
     this.maxSize = maxSize;
   }
 

src/main/services/discovery/SessionSearcher.ts

@@ -15,10 +15,6 @@ import { LocalFileSystemProvider } from '@main/services/infrastructure/LocalFile
 import { parseJsonlFile } from '@main/utils/jsonl';
 import { extractBaseDir, extractSessionId } from '@main/utils/pathDecoder';
 import { createLogger } from '@shared/utils/logger';
-import {
-  extractMarkdownPlainText,
-  findMarkdownSearchMatches,
-} from '@shared/utils/markdownTextSearch';
 import * as path from 'path';
 
 import { startMainSpan } from '../../sentry';
@@ -112,7 +108,7 @@ export class SessionSearcher {
         sessionFiles.sort((a, b) => b.mtimeMs - a.mtimeMs);
 
         // Search session files with bounded concurrency and staged breadth in SSH mode.
-        const searchBatchSize = fastMode ? 3 : 8;
+        const searchBatchSize = fastMode ? 3 : 16;
         const stageBoundaries = fastMode
           ? this.buildFastSearchStageBoundaries(sessionFiles.length)
           : [sessionFiles.length];
@@ -236,6 +232,10 @@ export class SessionSearcher {
 
     const { entries, sessionTitle } = cached;
 
+    // Fast pre-filter: skip sessions where no entry contains the query in raw text
+    const hasAnyMatch = entries.some((entry) => entry.text.toLowerCase().includes(query));
+    if (!hasAnyMatch) return results;
+
     for (const entry of entries) {
       if (results.length >= maxResults) break;
 
@@ -262,31 +262,20 @@ export class SessionSearcher {
     sessionId: string,
     sessionTitle?: string
   ): void {
-    const mdMatches = findMarkdownSearchMatches(entry.text, query);
-    if (mdMatches.length === 0) return;
+    // Plain indexOf search — no markdown/remark parsing
+    const lowerText = entry.text.toLowerCase();
+    if (!lowerText.includes(query)) return;
 
-    // Build plain text once for context snippet extraction
-    const plainText = extractMarkdownPlainText(entry.text);
-    const lowerPlain = plainText.toLowerCase();
-
-    for (const mdMatch of mdMatches) {
+    // Use raw text directly for context snippets
+    let pos = 0;
+    let matchIndex = 0;
+    while ((pos = lowerText.indexOf(query, pos)) !== -1) {
       if (results.length >= maxResults) return;
 
-      // Find approximate position in plain text for context extraction
-      let pos = 0;
-      for (let i = 0; i < mdMatch.matchIndexInItem; i++) {
-        const idx = lowerPlain.indexOf(query, pos);
-        if (idx === -1) break;
-        pos = idx + query.length;
-      }
-      const matchPos = lowerPlain.indexOf(query, pos);
-      const effectivePos = matchPos >= 0 ? matchPos : 0;
-
-      const contextStart = Math.max(0, effectivePos - 50);
-      const contextEnd = Math.min(plainText.length, effectivePos + query.length + 50);
-      const context = plainText.slice(contextStart, contextEnd);
-      const matchedText =
-        matchPos >= 0 ? plainText.slice(matchPos, matchPos + query.length) : query;
+      const contextStart = Math.max(0, pos - 50);
+      const contextEnd = Math.min(entry.text.length, pos + query.length + 50);
+      const context = entry.text.slice(contextStart, contextEnd);
+      const matchedText = entry.text.slice(pos, pos + query.length);
 
       results.push({
         sessionId,
@@ -294,15 +283,20 @@ export class SessionSearcher {
         sessionTitle: sessionTitle ?? 'Untitled Session',
         matchedText,
         context:
-          (contextStart > 0 ? '...' : '') + context + (contextEnd < plainText.length ? '...' : ''),
+          (contextStart > 0 ? '...' : '') +
+          context +
+          (contextEnd < entry.text.length ? '...' : ''),
         messageType: entry.messageType,
         timestamp: entry.timestamp,
         groupId: entry.groupId,
         itemType: entry.itemType,
-        matchIndexInItem: mdMatch.matchIndexInItem,
-        matchStartOffset: effectivePos,
+        matchIndexInItem: matchIndex,
+        matchStartOffset: pos,
         messageUuid: entry.messageUuid,
       });
+
+      matchIndex++;
+      pos += query.length;
     }
   }
 

src/renderer/components/chat/LastOutputDisplay.tsx

@@ -11,7 +11,7 @@ import { CopyButton } from '../common/CopyButton';
 import { OngoingBanner } from '../common/OngoingIndicator';
 
 import { createMarkdownComponents, markdownComponents } from './markdownComponents';
-import { createSearchContext } from './searchHighlightUtils';
+import { createSearchContext, EMPTY_SEARCH_MATCHES } from './searchHighlightUtils';
 
 import type { AIGroupLastOutput } from '@renderer/types/groups';
 
@@ -41,12 +41,16 @@ export const LastOutputDisplay = ({
   isLastGroup = false,
   isSessionOngoing = false,
 }: Readonly<LastOutputDisplayProps>): React.JSX.Element | null => {
+  // Only re-render if THIS AI group has search matches
   const { searchQuery, searchMatches, currentSearchIndex } = useStore(
-    useShallow((s) => ({
-      searchQuery: s.searchQuery,
-      searchMatches: s.searchMatches,
-      currentSearchIndex: s.currentSearchIndex,
-    }))
+    useShallow((s) => {
+      const hasMatch = s.searchMatchItemIds.has(aiGroupId);
+      return {
+        searchQuery: hasMatch ? s.searchQuery : '',
+        searchMatches: hasMatch ? s.searchMatches : EMPTY_SEARCH_MATCHES,
+        currentSearchIndex: hasMatch ? s.currentSearchIndex : -1,
+      };
+    })
   );
   const isTextOutput = lastOutput?.type === 'text' && Boolean(lastOutput.text);
 

src/renderer/components/chat/UserChatGroup.tsx

@@ -21,6 +21,7 @@ import { CopyButton } from '../common/CopyButton';
 
 import {
   createSearchContext,
+  EMPTY_SEARCH_MATCHES,
   highlightSearchInChildren,
   type SearchContext,
 } from './searchHighlightUtils';
@@ -401,13 +402,16 @@ const UserChatGroupInner = ({ userGroup }: Readonly<UserChatGroupProps>): React.
     [teams]
   );
 
-  // Get search state for highlighting
+  // Get search state for highlighting — only re-render if THIS item has matches
   const { searchQuery, searchMatches, currentSearchIndex } = useStore(
-    useShallow((s) => ({
-      searchQuery: s.searchQuery,
-      searchMatches: s.searchMatches,
-      currentSearchIndex: s.currentSearchIndex,
-    }))
+    useShallow((s) => {
+      const hasMatch = s.searchMatchItemIds.has(groupId);
+      return {
+        searchQuery: hasMatch ? s.searchQuery : '',
+        searchMatches: hasMatch ? s.searchMatches : EMPTY_SEARCH_MATCHES,
+        currentSearchIndex: hasMatch ? s.currentSearchIndex : -1,
+      };
+    })
   );
 
   const hasImages = content.images.length > 0;

src/renderer/components/chat/searchHighlightUtils.ts

@@ -8,6 +8,9 @@ import React from 'react';
 
 import type { SearchMatch } from '@renderer/store/types';
 
+/** Stable empty array for item-scoped search selectors (avoids re-renders) */
+export const EMPTY_SEARCH_MATCHES: SearchMatch[] = [];
+
 // Highlight styles matching SearchHighlight.tsx
 const baseStyles: React.CSSProperties = {
   borderRadius: '0.125rem',

src/renderer/components/chat/viewers/MarkdownViewer.tsx

@@ -36,6 +36,7 @@ import { useShallow } from 'zustand/react/shallow';
 
 import {
   createSearchContext,
+  EMPTY_SEARCH_MATCHES,
   highlightSearchInChildren,
   type SearchContext,
 } from '../searchHighlightUtils';
@@ -44,8 +45,6 @@ import { highlightLine } from '../viewers/syntaxHighlighter';
 import { FileLink, isRelativeUrl } from './FileLink';
 import { MermaidDiagram } from './MermaidDiagram';
 
-import type { SearchMatch } from '@renderer/store/types';
-
 // =============================================================================
 // Types
 // =============================================================================
@@ -73,7 +72,6 @@ interface MarkdownViewerProps {
 
 const EMPTY_TEAMS: { teamName?: string; displayName?: string; color?: string }[] = [];
 const EMPTY_TEAM_COLOR_MAP = new Map<string, string>();
-const EMPTY_SEARCH_MATCHES: SearchMatch[] = [];
 const NOOP_TEAM_CLICK = (): void => undefined;
 
 // =============================================================================
@@ -714,13 +712,16 @@ export const MarkdownViewer: React.FC<MarkdownViewerProps> = ({
   const isTooLarge = content.length > MAX_MARKDOWN_CHARS;
   const disableHighlight = content.length > DISABLE_HIGHLIGHT_CHARS;
 
-  // Only subscribe to search store when itemId is provided
+  // Only re-render if THIS item has search matches
   const { searchQuery, searchMatches, currentSearchIndex } = useStore(
-    useShallow((s) => ({
-      searchQuery: itemId ? s.searchQuery : '',
-      searchMatches: itemId ? s.searchMatches : EMPTY_SEARCH_MATCHES,
-      currentSearchIndex: itemId ? s.currentSearchIndex : -1,
-    }))
+    useShallow((s) => {
+      const hasMatch = itemId ? s.searchMatchItemIds.has(itemId) : false;
+      return {
+        searchQuery: hasMatch ? s.searchQuery : '',
+        searchMatches: hasMatch ? s.searchMatches : EMPTY_SEARCH_MATCHES,
+        currentSearchIndex: hasMatch ? s.currentSearchIndex : -1,
+      };
+    })
   );
 
   // Guard: very large markdown can freeze the renderer (remark/rehype + highlighting).

src/renderer/store/slices/conversationSlice.ts

@@ -4,7 +4,6 @@
 
 import { findLastOutput } from '@renderer/utils/aiGroupEnhancer';
 import { stripAgentBlocks } from '@shared/constants/agentBlocks';
-import { findMarkdownSearchMatches } from '@shared/utils/markdownTextSearch';
 
 import type { AppState, SearchMatch } from '../types';
 import type { AIGroupExpansionLevel } from '@renderer/types/groups';
@@ -17,6 +16,9 @@ import type { StateCreator } from 'zustand';
 
 type DetailItemType = 'thinking' | 'text' | 'linked-tool' | 'subagent';
 
+/** Maximum number of search matches to track. Beyond this, results are capped. */
+const MAX_SEARCH_MATCHES = 500;
+
 const isSearchDebugEnabled = (): boolean => {
   if (typeof window === 'undefined') return false;
   try {
@@ -57,6 +59,10 @@ export interface ConversationSlice {
   searchResultCount: number;
   currentSearchIndex: number;
   searchMatches: SearchMatch[];
+  /** True when total matches exceeded the cap and results were truncated */
+  searchResultsCapped: boolean;
+  /** Item IDs that contain at least one search match — used by components to skip re-renders */
+  searchMatchItemIds: Set<string>;
 
   // Auto-expand state for search results
   /** AI group IDs that should be expanded to show search results */
@@ -126,6 +132,8 @@ export const createConversationSlice: StateCreator<AppState, [], [], Conversatio
   searchResultCount: 0,
   currentSearchIndex: -1,
   searchMatches: [],
+  searchResultsCapped: false,
+  searchMatchItemIds: new Set(),
 
   // Auto-expand state for search results (initial values)
   searchExpandedAIGroupIds: new Set(),
@@ -222,58 +230,66 @@ export const createConversationSlice: StateCreator<AppState, [], [], Conversatio
         searchResultCount: 0,
         currentSearchIndex: -1,
         searchMatches: [],
+        searchResultsCapped: false,
+        searchMatchItemIds: new Set(),
         searchCurrentDisplayItemId: null,
         searchCurrentSubagentItemId: null,
       });
       return;
     }
 
-    // Build search matches by scanning conversation
-    // ONLY searches: user message text and AI lastOutput text (not tool results)
-    // Uses remark-based markdown parsing to extract visible text segments,
-    // ensuring match counts align with what ReactMarkdown renders.
+    // Build search matches by scanning conversation.
+    // Plain indexOf search — no markdown parsing. Match counts may differ
+    // slightly from rendered highlights; syncSearchMatchesWithRendered corrects this.
     const matches: SearchMatch[] = [];
     const lowerQuery = query.toLowerCase();
     let globalIndex = 0;
+    let capped = false;
 
-    // Helper to find markdown-aware matches and add to matches array
-    const addMarkdownMatches = (
+    // Count occurrences of lowerQuery in text using plain indexOf
+    const addPlainTextMatches = (
       text: string,
       itemId: string,
       itemType: 'user' | 'ai',
       displayItemId?: string
     ): void => {
-      const mdMatches = findMarkdownSearchMatches(text, lowerQuery);
-      for (const mdMatch of mdMatches) {
+      if (capped) return;
+      const lowerText = text.toLowerCase();
+      let pos = 0;
+      let matchIndexInItem = 0;
+      while ((pos = lowerText.indexOf(lowerQuery, pos)) !== -1) {
+        if (matches.length >= MAX_SEARCH_MATCHES) {
+          capped = true;
+          return;
+        }
         matches.push({
           itemId,
           itemType,
-          matchIndexInItem: mdMatch.matchIndexInItem,
+          matchIndexInItem,
           globalIndex,
           displayItemId,
         });
+        matchIndexInItem++;
         globalIndex++;
+        pos += lowerQuery.length;
       }
     };
 
     for (const item of conversation.items) {
+      if (capped) break;
       if (item.type === 'user') {
         const raw = item.group.content.rawText ?? item.group.content.text ?? '';
         const text = stripAgentBlocks(raw);
-        addMarkdownMatches(text, item.group.id, 'user');
+        addPlainTextMatches(text, item.group.id, 'user');
       } else if (item.type === 'ai') {
-        // For AI items: ONLY search lastOutput text (not tool results, thinking, or subagents)
         const aiGroup = item.group;
         const itemId = aiGroup.id;
         const lastOutput = findLastOutput(aiGroup.steps, aiGroup.isOngoing ?? false);
 
         if (lastOutput?.type === 'text' && lastOutput.text) {
-          // Last output text - displayItemId indicates this is lastOutput content
-          addMarkdownMatches(lastOutput.text, itemId, 'ai', 'lastOutput');
+          addPlainTextMatches(lastOutput.text, itemId, 'ai', 'lastOutput');
         }
-        // Skip tool_result type - only searching text output
       }
-      // Skip system items entirely
     }
 
     if (isSearchDebugEnabled()) {
@@ -293,11 +309,19 @@ export const createConversationSlice: StateCreator<AppState, [], [], Conversatio
       console.info('[search] sample', sample);
     }
 
+    // Build set of item IDs that have matches — components use this to skip re-renders
+    const matchItemIds = new Set<string>();
+    for (const match of matches) {
+      matchItemIds.add(match.itemId);
+    }
+
     set({
       searchQuery: query,
       searchResultCount: matches.length,
       currentSearchIndex: matches.length > 0 ? 0 : -1,
       searchMatches: matches,
+      searchResultsCapped: capped,
+      searchMatchItemIds: matchItemIds,
     });
   },
 
@@ -406,6 +430,8 @@ export const createConversationSlice: StateCreator<AppState, [], [], Conversatio
       searchResultCount: 0,
       currentSearchIndex: -1,
       searchMatches: [],
+      searchResultsCapped: false,
+      searchMatchItemIds: new Set(),
       searchExpandedAIGroupIds: new Set(),
       searchExpandedSubagentIds: new Set(),
       searchCurrentDisplayItemId: null,

src/shared/utils/markdownTextSearch.ts

@@ -49,7 +49,7 @@ function parseMarkdown(text: string): MdastRoot {
 // Segment cache (parse once, search many times per query keystroke)
 // ---------------------------------------------------------------------------
 
-const MAX_CACHE_SIZE = 200;
+const MAX_CACHE_SIZE = 1000;
 const segmentCache = new Map<string, string[]>();
 
 function getCachedSegments(markdown: string): string[] {
@@ -154,6 +154,9 @@ export interface MarkdownSearchMatch {
 export function findMarkdownSearchMatches(markdown: string, query: string): MarkdownSearchMatch[] {
   if (!query || !markdown) return [];
 
+  // Fast pre-filter: skip expensive markdown parsing if query doesn't appear in raw text
+  if (!markdown.toLowerCase().includes(query.toLowerCase())) return [];
+
   const segments = getCachedSegments(markdown);
   const lowerQuery = query.toLowerCase();
   const matches: MarkdownSearchMatch[] = [];
@@ -178,6 +181,9 @@ export function findMarkdownSearchMatches(markdown: string, query: string): Mark
 export function countMarkdownSearchMatches(markdown: string, query: string): number {
   if (!query || !markdown) return 0;
 
+  // Fast pre-filter: skip expensive markdown parsing if query doesn't appear in raw text
+  if (!markdown.toLowerCase().includes(query.toLowerCase())) return 0;
+
   const segments = getCachedSegments(markdown);
   const lowerQuery = query.toLowerCase();
   let count = 0;

test/main/services/discovery/SessionSearcher.test.ts

@@ -76,7 +76,7 @@ describe('SessionSearcher', () => {
     ).toBe(true);
   });
 
-  it('does not produce phantom matches for code fence language identifiers', async () => {
+  it('matches text in code fences with plain text search', async () => {
     const projectsDir = fs.mkdtempSync(path.join(os.tmpdir(), 'session-searcher-md-'));
     tempDirs.push(projectsDir);
 
@@ -110,13 +110,11 @@ describe('SessionSearcher', () => {
     const searcher = new SessionSearcher(projectsDir);
     const result = await searcher.searchSessions(projectId, 'tsx', 50);
 
-    // "tsx" should match in user text ("Show me tsx code") but NOT in the
-    // code fence language identifier (```tsx). It should also not match in
-    // the code block content since "const x = 1;" doesn't contain "tsx".
+    // Plain text search: "tsx" matches in user text AND in the code fence identifier
     const userResults = result.results.filter((r) => r.itemType === 'user');
     const aiResults = result.results.filter((r) => r.itemType === 'ai');
 
     expect(userResults).toHaveLength(1);
-    expect(aiResults).toHaveLength(0);
+    expect(aiResults).toHaveLength(1);
   });
 });