fix: correct Task tool description extraction for subagent window titles

- Add parseTaskDescriptionsFromTerminalData() to parse raw terminal data in
  interactive sessions (not just one-shot mode)
- Strip ANSI codes before regex matching since terminal output has embedded
  codes like [1mExplore[0m
- Fix getTaskDescriptionForSubagent() to check ALL sessions with matching
  project hash (not just the first one found) since multiple sessions can
  share the same working directory
- Add tool_result event forwarding for subagent SSE broadcasts

The pattern Explore(Description) in terminal output is now correctly parsed
and used as the subagent window title instead of extracting from the prompt.

Co-Authored-By: Claude Opus 4.5 <noreply@anthropic.com>

src/session.ts

@@ -69,6 +69,11 @@ const FOCUS_ESCAPE_FILTER = /\x1b\[\?1004[hl]|\x1b\[[IO]/g;
 // - Single-char escapes: ESC = or ESC >
 const ANSI_ESCAPE_PATTERN = /\x1b(?:\[[0-9;?]*[A-Za-z]|\][^\x07\x1b]*(?:\x07|\x1b\\)|[=>])/g;
 const TOKEN_PATTERN = /(\d+(?:\.\d+)?)\s*([kKmM])?\s*tokens/;
+// Pattern to match Task tool invocations in terminal output
+// Matches: "Explore(Description)", "Task(Description)", "Bash(Description)", etc.
+// The prefix characters vary (●, ·, ✶, etc.) so we don't require them
+// We look for the tool name followed by (description)
+const TASK_TOOL_PATTERN = /\b(Explore|Task|Bash|Plan|general-purpose)\(([^)]+)\)/g;
 
 // ============================================================================
 // Claude CLI PATH Resolution
@@ -425,6 +430,11 @@ export class Session extends EventEmitter {
     toolsUpdate: (tools: ActiveBashTool[]) => void;
   } | null = null;
 
+  // Task descriptions parsed from terminal output (e.g., "Explore(Description)")
+  // Used to correlate with SubagentWatcher discoveries for better window titles
+  private _recentTaskDescriptions: Map<number, string> = new Map(); // timestamp -> description
+  private static readonly TASK_DESCRIPTION_MAX_AGE_MS = 30000; // Keep descriptions for 30 seconds
+
   constructor(config: Partial<SessionConfig> & {
     workingDir: string;
     mode?: SessionMode;
@@ -893,6 +903,10 @@ export class Session extends EventEmitter {
       // Parse token count from status line (e.g., "123.4k tokens" or "5234 tokens")
       this.parseTokensFromStatusLine(data);
 
+      // Parse task descriptions from terminal output (e.g., "Explore(Check files)")
+      // This enables correlating subagent windows with their short descriptions
+      this.parseTaskDescriptionsFromTerminalData(data);
+
       // Detect if Claude is working or at prompt
       // The prompt line contains "❯" when waiting for input
       if (data.includes('❯') || data.includes('\u276f')) {
@@ -1305,10 +1319,108 @@ export class Session extends EventEmitter {
       } else if (trimmed) {
         this._textOutput.append(line + '\n');
       }
+
+      // Parse task descriptions from terminal output (e.g., "Explore(Description)")
+      // This captures the short description from Claude Code's Task tool output
+      this.parseTaskDescriptionsFromLine(cleanLine);
     }
     // Note: BufferAccumulator auto-trims when max size exceeded
   }
 
+  /**
+   * Parse task descriptions from raw terminal data (may contain multiple lines).
+   * Called from interactive mode's onData handler.
+   */
+  private parseTaskDescriptionsFromTerminalData(data: string): void {
+    // Quick pre-check: skip if no parentheses present
+    if (!data.includes('(') || !data.includes(')')) return;
+
+    // Split by newlines and process each line
+    const lines = data.split(/\r?\n/);
+    for (const line of lines) {
+      this.parseTaskDescriptionsFromLine(line);
+    }
+  }
+
+  /**
+   * Parse task descriptions from terminal output.
+   * Claude Code outputs Task tool calls as "ToolName(Description)" in the terminal.
+   * We capture these descriptions to use as window titles for subagents.
+   */
+  private parseTaskDescriptionsFromLine(line: string): void {
+    // Quick pre-check: skip expensive regex if no common tool patterns present
+    if (!line.includes('(') || !line.includes(')')) return;
+
+    // Strip ANSI codes before matching - terminal output has embedded codes like [1mExplore[0m
+    const cleanLine = line.replace(ANSI_ESCAPE_PATTERN, '');
+
+    // Reset regex lastIndex for global pattern
+    TASK_TOOL_PATTERN.lastIndex = 0;
+
+    let match;
+    while ((match = TASK_TOOL_PATTERN.exec(cleanLine)) !== null) {
+      const description = match[2].trim();
+      if (description && description.length > 0) {
+        const now = Date.now();
+        this._recentTaskDescriptions.set(now, description);
+
+        // Cleanup old entries
+        this.cleanupOldTaskDescriptions();
+      }
+    }
+  }
+
+  /**
+   * Remove task descriptions older than TASK_DESCRIPTION_MAX_AGE_MS.
+   */
+  private cleanupOldTaskDescriptions(): void {
+    const cutoff = Date.now() - Session.TASK_DESCRIPTION_MAX_AGE_MS;
+    for (const [timestamp] of this._recentTaskDescriptions) {
+      if (timestamp < cutoff) {
+        this._recentTaskDescriptions.delete(timestamp);
+      }
+    }
+  }
+
+  /**
+   * Get recent task descriptions parsed from terminal output.
+   * Returns descriptions sorted by timestamp (most recent first).
+   */
+  getRecentTaskDescriptions(): Array<{ timestamp: number; description: string }> {
+    this.cleanupOldTaskDescriptions();
+    const results: Array<{ timestamp: number; description: string }> = [];
+    for (const [timestamp, description] of this._recentTaskDescriptions) {
+      results.push({ timestamp, description });
+    }
+    return results.sort((a, b) => b.timestamp - a.timestamp);
+  }
+
+  /**
+   * Find a task description that was parsed close to a given timestamp.
+   * Used to correlate with SubagentWatcher discoveries.
+   *
+   * @param subagentStartTime - The timestamp when the subagent was discovered
+   * @param maxAgeMs - Maximum age difference to consider (default 10 seconds)
+   * @returns The matching description or undefined
+   */
+  findTaskDescriptionNear(subagentStartTime: number, maxAgeMs: number = 10000): string | undefined {
+    this.cleanupOldTaskDescriptions();
+
+    // Find the most recent description that was parsed before or around the subagent start time
+    let bestMatch: { timestamp: number; description: string } | undefined;
+    let bestDiff = Infinity;
+
+    for (const [timestamp, description] of this._recentTaskDescriptions) {
+      const diff = Math.abs(subagentStartTime - timestamp);
+      if (diff < maxAgeMs && diff < bestDiff) {
+        bestMatch = { timestamp, description };
+        bestDiff = diff;
+      }
+    }
+
+    return bestMatch?.description;
+  }
+
   // Parse token count from Claude's status line in interactive mode
   // Matches patterns like "123.4k tokens", "5234 tokens", "1.2M tokens"
   private parseTokensFromStatusLine(data: string): void {