Bug: Context becomes full and /compact stops working after using subagents (Task tool)

Description

When using subagents via the Task tool, context usage can spike rapidly and reach 100% capacity. At this point, the /compact command becomes unresponsive, leaving users with no way to recover the session.

Steps to Reproduce

1. Start a Claude Code session
2. Use the Task tool to spawn a subagent (e.g., subagent_type: Explore or general-purpose)
3. The subagent performs file reads, greps, and other operations
4. Context usage jumps significantly after subagent returns
5. Repeat subagent usage 2-3 times
6. Context reaches ~100%
7. Attempt to run /compact
8. Command fails or is ignored

Expected Behavior

• /compact should work especially when context is near capacity (this is when users need it most)
• Or: automatic context management before reaching critical levels

Actual Behavior

• /compact becomes unresponsive when context is full
• User loses all session work and must start a new session
• No warning before reaching this state

Proposed Solutions

1. Reserve context space for /compact - Always keep enough headroom to execute compaction
2. Subagent result summarization option - Allow users to receive summarized results instead of full output
3. Context usage warnings - Alert users at 70%, 80%, 90% thresholds
4. Auto-compact option - Automatically compress context when reaching a threshold
5. Graceful degradation - If /compact can't run normally, provide a minimal emergency compaction

Environment

• OS: macOS (Darwin 24.5.0)
• Claude Code version: $(claude --version 2>/dev/null || echo "unknown")

Impact

This issue causes significant workflow disruption. Users lose context, unsaved work patterns, and must restart complex debugging or exploration sessions from scratch.