filesystem-context

You can offload large tool outputs to a filesystem to prevent agent context window bloat. This skill enables agents to write oversized results to scratch pads and dynamically reload them, drastically reducing prompt token usage for long-running tasks.

Persisting agent plans and memory across context refreshes is achieved using file-based storage with simple JSON or YAML formats. This ensures progress is maintained in long-running, multi-step tasks even when the agent's context window is cleared.

You can share context and findings between multiple coordinating agents through a shared filesystem. The skill supports inter-agent memory and sub-agent workspaces, allowing isolated agents to write results and have a coordinator dynamically load them.

You do not need external libraries to implement dynamic context loading for agents. The skill relies on standard Python libraries to handle file-based memory, scratch pads, and plan persistence using straightforward JSON and YAML formats.

Filesystem-based context management should be used instead of keeping everything in the prompt during long-running data-gathering tasks. It prevents context limits from cutting off execution and allows dynamic discovery of offloaded results without bloating the prompt.