ClawTrust
2.5k Downloads
Overview
Provides a minimal, non-LLM control-plane safety layer for OpenClaw agents to prevent deadlocks, freezes, and unrecoverable states by exposing fast, always-responsive system commands for status, emergency flushing, and controlled recovery.
Key Advantages
1.Non-blocking, constant-time control commands that never call models or external APIs, reducing the chance of cascaded failures.
2.Clear separation between control-plane and workload logic, so safety operations do not depend on the main agent’s health or reasoning.
3.Emergency `/flush` and `/recover` flows to quickly stop all work, clear queues, and reset in-memory state without restarting the container.
4.Minimal state tracking (task metadata only), avoiding storage of payloads, prompts, or user data and reducing privacy risk.
5.Event-driven design that scales better to long-running, concurrent workloads than ad-hoc watchdog logic inside task code or LLM flows.,
Use Cases
- Operating long-running or high-risk workloads (benchmarks, simulations, batch jobs) where deadlocks and hangs are likely or costly.
- Managing sub-agents and background workers (e.g., PNR checks, monitors, scheduled tasks) that may silently stall or overrun time limits.
- Providing a last-line-of-defense safety switch in production-like OpenClaw deployments so operators can recover from unresponsive states without container restarts.
- Inspecting current system health and task registry state via `/status` during debugging, incident response, or capacity analysis.
- Building higher-level orchestration or watchdog layers on top of a reliable control-plane primitive instead of embedding safety logic into each agent.
Evaluation Scores
7.6
/ 10
Reliability
7.0
Functionality
7.5
Usability
6.5
Safety
8.5
Performance
9.0
Compatibility
7.5
Based on 1 evaluation · Latest: 3/19/2026
Download Trend
Loading...
Evaluation History (1)
7.6/103/19/2026▼
OS: darwin-x64LLM: anthropic/claude-opus-4.6
**Quick judgment:** A focused, system-level safety skill that is well-suited as a control-plane “circuit breaker” for serious OpenClaw deployments with long-lived or potentially blocking workloads. It’s intentionally minimal and non-LLM, prioritizing fast, deterministic recovery over rich features.
**What it does well:**
- Keeps the main agent from blocking by moving safety and control to a separate, event-driven control plane.
- Provides three critical primitives: `/status` (health + task registry), `/flush` (emergency stop + queue clear), and `/recover` (safe reset sequence built on `/flush`).
- Designed to always respond, avoid external I/O or model calls, and maintain only lightweight task metadata, which supports both reliability and privacy.
**Key risks / limitations:**
- Intended for advanced users who understand OpenClaw’s execution model; misuse (e.g., frequent `/flush`) can disrupt all active work and may surprise less experienced operators.
- Phase 1 only: more sophisticated features (watchdogs, structured events, back-pressure) are not implemented yet, so some safety behaviors must still be built externally.
- Strong safety assumptions (e.g., constant-time behavior, no blocking) depend on correct integration with the broader system; mis-instrumented workers could still undermine guarantees.
**Recommended scenarios:**
- Long-running agents, benchmarks, and background monitors where deadlocks or silent stalls would otherwise require container restarts.
- Production or near-production OpenClaw setups that need an operator-accessible emergency recovery mechanism.
- Multi-task or sub-agent environments where you want a minimal, dependable control-plane safety layer rather than embedding recovery logic into every agent or workflow.
Comments (0)
No comments yet. Be the first!