Изоляция контекста фоновых агентов

orchestration

agent-context-isolation is a Claude Code skill that prevents background agent output from polluting the main conversation's context window through strict file-based coordination rules. Instead of calling TaskOutput — which dumps the full agent transcript (70,000+ tokens) into the main context — the skill requires launching agents with run_in_background=true and having them write results to .claude/cache/agents/<agent-type>/ files. It defines a multi-agent pipeline pattern (Research → Plan → Validate → Implement) where each agent reads the previous agent's file output rather than its TaskOutput. Agent progress is monitored via automatic system reminders and output file size checks; correctness is verified by running the test suite directly (bun test) rather than inspecting agent output. Essential for multi-agent orchestrations where preserving the main context budget and avoiding mid-conversation compaction are priorities.