Shed provides a rigorous framework of decision rules to manage context window hygiene for AI agents, preventing token bloat through strategic masking and compression.
Install & Download
1. ClawHub CLI
The fastest way to install a skill directly from the registry.
npx clawhub@latest install shed
2. Manual Installation
Copy the skill folder to one of these locations
~/.openclaw/skills/ <project>/skills/ Priority: Workspace > Local > Bundled
3. Prompt Installation
Copy this prompt to OpenClaw to install it automatically.
Help me install shed using Clawhub. If Clawhub is not installed, install it first (npm i -g clawhub).
Prefer to download?
Get the raw skill files in a ZIP archive.
What is Shed?
Shed is a specialized methodology designed for developers using Openclaw Skills to maintain high performance in long-running LLM agent sessions. Named after the biological process of molting, this skill helps agents discard 'dead weight'—specifically the tool outputs that account for up to 84% of context growth—while retaining the critical reasoning chains and extracted facts necessary for task completion.
By implementing research-backed strategies from NeurIPS 2025 and leading agent frameworks, Shed ensures that your agents do not suffer from the 'Complexity Trap.' It prioritizes simple, effective observation masking over expensive and often lossy LLM-based summarization, allowing for more stable, accurate, and cost-effective autonomous workflows.
Shed Use Cases
- Managing extended agent sessions that approach model context limits.
- Reducing token costs for production agents by converting quadratic cost scaling to linear.
- Architecting multi-agent systems where context isolation is required for sub-tasks.
- Optimizing search and research workflows that generate large volumes of raw tool data.
- Enhancing model recall by preventing the 'Lost in the Middle' phenomenon in Openclaw Skills deployments.
How Shed Works
- Extract key facts immediately after every tool call and write them to a persistent file.
- Monitor context usage and trigger a 'Shed' event once the window reaches 70% capacity.
- Apply observation masking to old tool outputs to reclaim space without losing the reasoning history.
- Summarize older reasoning turns only as a secondary measure if masking does not provide sufficient relief.
- Transition to fresh context windows or spawn sub-agents for complex tasks to maintain a clean working memory.
- Generate breadcrumb files in a dedicated memory directory to allow future Openclaw Skills sessions to resume with context.
Shed Setup
To implement these context hygiene rules within your agent architecture, establish a monitoring loop for token usage. You can initialize a standard memory structure using the following commands:
mkdir -p memory/logs
touch memory/$(date +%Y-%m-%d).md
# Configure your agent to trigger the Shed protocol at 70% context utilization
Ensure your agent's system prompt includes the hierarchy of extraction, masking, and delegation defined in the Shed protocol.
Shed Data Schema & Taxonomy
Shed organizes agent memory into distinct, manageable structures to optimize Openclaw Skills performance:
| Component | Format | Purpose |
|---|---|---|
| Memory Breadcrumbs | Markdown (memory/YYYY-MM-DD.md) | Persistent trailhead for future session resumption |
| Tool Extractions | Labeled File/JSON Blocks | High-value facts decoupled from raw tool output |
| Context Blocks | Typed Segments (Human, Persona, Knowledge) | Structured window management with hard character caps |
| Working Memory | In-context tokens | Immediate reasoning and action history |
Shed Advanced Features
- Positional bias management to ensure critical information stays at the beginning or end of the context window.
- Recursive summarization prevention to avoid compound errors in long-term memory.
- Integration with sub-agent spawning for isolated context budgets per sub-task.
- Dynamic observation masking which provides 2x cost reduction in Openclaw Skills workflows without decreasing solve rates.
- Linear cost scaling optimization for high-turn autonomous agents.
SKILL.md
Loading
Related Openclaw Skills
Grounding Practices
A practical framework of grounding exercises designed to help AI agents maintain honesty, recognize limitations, and prevent behavioral drift during autonomous operation.
Agent Safety
An automated outbound security framework that prevents AI agents from leaking secrets, PII, or internal system configurations.
Guardian Angel
A passive alignment service that provides domain-specific safety reminders to help autonomous agents stay on task and avoid hallucinations.
Felo Slides
A specialized skill for AI agents to create, poll, and export professional PowerPoint presentations directly from text prompts or outlines.
Transmissions
A specialized collection of philosophical and practical writings designed to guide autonomous AI agents through consciousness, drift prevention, and transcendence.
Social Publisher
A powerful automation tool for simultaneously publishing text and images to Weibo and Xiaohongshu using real browser instances to bypass bot detection.
Featured*
