Feature: In-flight task/sub-agent status in status line

Feature Request

Problem

The status line currently only updates after each assistant message completes (debounced at 300ms). It has no visibility into in-flight operations — running sub-agents, tool executions, or extended thinking phases. When Claude Code spawns multiple sub-agents or runs long tool calls, the user has no real-time feedback on what's happening.

Current Behavior

The status line JSON payload contains only cumulative session metrics:

• Model name, session cost, duration
• Context window usage
• Lines added/removed
• Vim mode state

It does not include:

• Active sub-agent count or status
• Which tool is currently executing
• Extended thinking state
• Task progress (e.g., "3/5 sub-agents complete")

Updates only fire between assistant messages, so the status line goes stale during long operations.

Proposed Behavior

Expose in-flight operational state to the status line script, for example:

{
  "active_tasks": [
    {"id": "abc123", "type": "Explore", "description": "Searching codebase", "started_at": "..."}
  ],
  "active_task_count": 1,
  "tool_in_progress": "Grep",
  "thinking": true
}

Key capabilities:

1. Active sub-agent tracking — count, type, and short description of running sub-agents
2. Tool execution state — which tool is currently running
3. Thinking indicator — whether Claude is in an extended thinking phase
4. Mid-execution updates — fire status line updates during long operations, not only after message completion

This would let users build status lines like:

⏳ 2 agents running | 🔍 Grep | ctx: 45% | $0.12

Workarounds Considered

• Hooks (SubagentStart/SubagentStop) can log events externally but can't feed back into the status line in real time
• Transcript tailing works but is outside the Claude Code UI
• Running sub-agents in the foreground loses parallelism benefits

Use Case

Power users running complex multi-step tasks benefit from knowing what Claude Code is doing at any moment — especially when multiple sub-agents are spawned in parallel or a single tool call takes a long time. This reduces uncertainty and helps users decide whether to wait or interrupt.