Agent Context Isolation

Keeps agent outputs from flooding your main context window by enforcing file-based coordination instead of TaskOutput calls. The core insight is that reading agent transcripts with TaskOutput can dump 70k+ tokens into your conversation, triggering compaction and losing track of the original request. Instead, run agents in background mode and have them write results to files in .claude/cache/agents, then verify completion by checking for those files or running tests directly. The skill includes patterns for chaining agents through file outputs and detecting when background agents get stuck by monitoring system reminders and file growth.