Context exhaustion when orchestrating parallel agents

Problem

When orchestrating multiple parallel agents via the Task tool, the parent conversation's context gets exhausted quickly, requiring manual /compact to continue.

Reproduction Steps

1. Deploy 4+ parallel agents using Task tool in a single message
2. Monitor agents using TaskOutput with block=true
3. When agents complete, context is often nearly exhausted
4. Must manually run /compact to continue working

Expected Behavior

The parent conversation should maintain sufficient context to continue work after parallel agents complete. Options:

• Auto-compact when context reaches threshold during agent orchestration
• Agent results should be summarized/compressed before returning to parent
• Warning when spawning agents that will likely exhaust context

Actual Behavior

Parent conversation runs out of context mid-workflow, requiring manual intervention to continue.

Environment

• Claude Code CLI
• Using Task tool with 4 parallel agents (type: general-purpose or specialized)
• Each agent returned ~2-5KB of results
• Parent conversation had existing context from prior work

Impact

• Interrupts multi-agent workflows
• User must manually compact and often loses context about what was happening
• Makes complex parallel orchestration fragile

Suggested Solutions

1. Auto-compact trigger: Automatically compact when context exceeds 80% during agent orchestration
2. Result summarization: Compress agent results to essential information before returning
3. Context budgeting: Reserve context for parent when spawning agents
4. Proactive warning: Warn user before spawning agents if context will likely exhaust