Контроль контекста в многоагентных пайплайнах

general

no-task-output is a Claude Code skill that prohibits the TaskOutput tool to prevent the main context window from being flooded with agent transcripts of 70,000+ tokens. When TaskOutput is called, it reads the full agent transcript directly into context, triggering mid-conversation compaction and undermining the entire point of agent context isolation. The skill enforces a single alternative pattern: use the Task tool in synchronous mode (run_in_background=false), which runs the agent and returns only a concise summary instead of a raw transcript dump. It is aimed at developers building multi-agent pipelines in Claude Code who encounter unexpected context exhaustion or performance degradation during agent orchestration.