SuperAPI AI Directory

filesystem-context

General & other

View repository

Updated 10-16 05:48

About this skill

📂 Filesystem as Extended Context

This skill allows using the local file system as a persistent and extended memory layer. The goal is to offload verbose or static information (logs, detailed plans, massive documentation) outside the active context window, keeping only high-signal references.

Mental Model: The file system is the agent's "hard drive," while the context is the "RAM." Optimizing the flow between both maximizes reasoning capacity without saturating attention.

🚩 Fragility and Autonomy

  • Fragility: Low. The file system is a stable and reliable medium.
  • Freedom: Very High in the organization of directories and persistence formats.

🚀 When to Activate

  • In debugging sessions with massive terminal outputs.
  • When coordinating complex tasks among multiple sub-agents.
  • To maintain long-running execution plans that survive session restarts.
  • When implementing self-modification systems or preference learning.

🧠 Core Concepts

1. The Scratchpad Pattern

Use temporary files (.scratch, plans/) to jot down ideas, code drafts, or research results before integrating them into the final solution.

2. Offloading Observations

Move verbose tool outputs (e.g., npm test with 1000 lines) to a file and provide the model with a summary and the path.

3. Plan Persistence (Stateful Agents)

Save the current state of the plan in a YAML/Markdown file so that the agent (or a new one) can resume exactly where it left off.

🛠️ Implementation in Gemini Elite Core

Pattern: Structured Working Directory

Define a clear hierarchy so the agent knows where to look for each type of "memory."

.gemini/
├── scratch/          # Archivos efímeros de trabajo
├── plans/            # Planes de tareas activos
├── agents/           # Workspaces de sub-agentes
└── preferences.yaml  # Preferencias aprendidas

Pattern: Dynamic Skill Loading

Do not load all skills at the start. Load the SKILL.md in full only when the agent detects that it is relevant to the current task.

<OFFLOADED_CONTEXT>
- Skill: advanced-evaluation (Path: skills/advanced-evaluation/SKILL.md)
- Reason: User requested LLM-as-a-Judge implementation.
</OFFLOADED_CONTEXT>

🔗 Integrations

  • context-optimization: The filesystem is the main tool for context "compaction."
  • multi-agent-patterns: Use of shared workspaces for asynchronous communication.
  • project-development: Tracking of milestones and progress in project files.

📚 Internal References

  • references/implementation-patterns.md: Library of classes and scripts for ScratchPad, PlanPersistence, and Workspaces.