What problem does it solve?
The filesystem-context skill solves the problem of limited model context windows by offloading large or persistent agent state to files so agents can read relevant information on demand instead of carrying bulky data in every prompt. It prevents token bloat from long tool outputs, preserves full tool results for later inspection, and enables sub-agents to share high-fidelity findings without repeated summarization.
Core Features & Use Cases
- Scratch pad offloading: Write large tool outputs to timestamped files and return compact references and summaries to keep the active context small.
- Plan persistence: Store structured plans on disk so agents can re-read objectives and progress across long-horizon workflows.
- Sub-agent workspaces: Use per-agent files for findings, status, and logs so coordinators read raw outputs directly.
- Dynamic skill loading: Keep only skill names and short descriptions in static context and load full SKILL.md files on demand.
- Terminal and log persistence: Persist terminal session outputs as searchable files and grep for targeted retrieval.
- Self-modification guardrails: Validate and bound agent-written preferences and provide cleanup strategies for scratch files.
- Use cases include large web search results, multi-agent research coordination, long-running devops tasks, and incremental code or test outputs that would otherwise overflow the context window.
Quick Start
Save the large tool output to a scratch file, return a one-line summary and the saved file path for later grep-based retrieval.
