shed

What This Skill Does

Shed provides a rigorous framework and set of operational procedures for context window hygiene in long-running LLM agents. By treating context as a finite, precious resource, Shed allows agents to manage their own memory by aggressively pruning low-value tool outputs, summarizing reasoning chains before they degrade, and delegating tasks to sub-agents with clean state. It implements industry-standard research findings—such as avoiding the 'Lost in the Middle' phenomenon and minimizing the 84% context bloat caused by raw tool outputs—to ensure your agents remain performant, coherent, and cost-effective over multi-hour sessions.

Installation

To integrate Shed into your agent environment, run the following command in your terminal: clawhub install openclaw/skills/skills/compass-soul/shed

Use Cases

• Long-term Coding Projects: Maintain clarity during massive refactoring tasks where file logs and search results would otherwise consume your context window.
• Autonomous Research Agents: Automatically compress large web-scraping results into concise facts, leaving the 'reasoning' intact without the raw data bloat.
• Agent Orchestration: Use Shed rules to enforce context isolation when spawning sub-agents, preventing the 'inherited bloat' that cripples multi-agent systems.
• Debugging Sessions: Clean up redundant stack traces and file read outputs after an issue is identified to focus the agent on the fix.

Example Prompts

1. "We've hit 70% of our context limit. Trigger the Shed protocol: extract the critical findings from these logs, save them to memory/logs.md, and mask the raw output to free up space."
2. "I am delegating the unit testing sub-task. Create a new sub-agent, use Shed to strip my current context of non-essential data, and provide the child agent with only the relevant module spec."
3. "Summarize our current reasoning path and archive the completed task details to the memory directory. Clear the remaining tool outputs to prepare for the next project milestone."

Tips & Limitations

• Never Recursive Summarization: Shed warns against summarizing summaries. Always store the original reasoning or save to a file before compressing to prevent information drift.
• Positional Awareness: Remember that models underperform on data placed in the middle of the context window. Use Shed to move critical 'bread-crumbs' to the very top or bottom of your context buffer.
• Tool Efficiency: The primary value of this skill lies in the discipline of extraction. Treat raw tool output as disposable the moment a summary is written to a file.
• Scope: Shed is designed to manage context state, not perform the actual LLM calls. It assumes you are implementing the condensation logic alongside the agent's core decision-making loop.