← Back to Skills Marketplace
hehkcaspar

LLM Loop Breaker

by hehkcaspar · GitHub ↗ · v1.0.0 · MIT-0
linux ⚠ suspicious
81
Downloads
0
Stars
0
Active Installs
1
Versions
Install in OpenClaw
/install llm-loop-breaker
Description
Two-layer defense for Openclaw gateway: kills hallucinating LLM streams via entropy analysis, and watchdogs host CPU/RAM/Disk to prevent resource exhaustion.
README (SKILL.md)

LLM Stream Guard

Two-layer runtime defense that protects the Openclaw gateway against hallucinating LLM streams and host resource exhaustion.

When to activate

Activate this skill once per fresh gateway deployment or whenever the gateway binary (openclaw.mjs) is rebuilt / updated.

What it does

Layer 1 -- Stream Entropy Breaker (Node.js)

Patches global.fetch to intercept every LLM streaming response (text/event-stream, application/x-ndjson, application/stream+json).

For each active stream it:

  1. Accumulates text chunks and measures compressed-vs-raw byte ratio using zlib deflate (compression ratio = uncompressed / compressed).
  2. Every 1024 bytes, schedules an entropy check via process.nextTick.
  3. Once 4000+ bytes have been received, evaluates kill conditions:
    • Hard kill: compression ratio > 10.0
    • Soft kill: compression ratio > 6.0 AND single-character dominance > 50%
  4. On kill: fires AbortController.abort() with reason HALLUCINATION_LOOP_DETECTED_BY_ENTROPY_BREAKER, severing the stream.

Bounded memory: accumulated text is truncated to the last 4096 characters when it exceeds 8192. A 30-second registry cleanup removes stale streams.

Layer 2 -- Host Resource Watchdog (Python)

Runs as a separate daemon outside the gateway process. Polls every 2 seconds via psutil.

Rule Trigger Action
Redline breach CPU >= 90% OR RAM >= 90% OR Disk >= 90% Log warning + audit entry
Poison pill Child process CPU > 95% with 0 I/O for 15 s Kill process tree + incident snapshot
Memory leak RSS monotonically grows > 100 MB over 60 s, CPU > 30% Kill process tree + incident snapshot
Crash loop Gateway PID changes 3+ times in 120 s Incident snapshot + 30 s cooldown

Incident snapshots capture journalctl and dmesg excerpts into ~/.openclaw/workspace/memory/core/incidents/.

Deployment

Run the deterministic deploy script. It is safe to run multiple times (idempotent).

export OPENCLAW_APP_DIR=/app          # path to openclaw installation root
bash deploy.sh

The script will:

  1. Verify python3, psutil, and $OPENCLAW_APP_DIR/openclaw.mjs exist.
  2. Copy runtime files to $OPENCLAW_APP_DIR/dist/llm_stream_guard/.
  3. Inject bootstrap code into openclaw.mjs (skipped if already present).
  4. Create a timestamped backup of openclaw.mjs before any modification.

After deployment, restart the Openclaw gateway to activate.

Verify

# Node.js layer tests (no network, no side effects)
node test-entropy-breaker.js

# Python layer tests (mocked psutil, no real process interaction)
python3 test-watchdog.py

Both test suites are deterministic and produce exit code 0 on success, 1 on failure.

Files

File Purpose
src/stream-entropy-breaker.cjs Fetch-patching + stream transform + abort logic
src/entropy-engine.cjs Zlib compression ratio calculator + repetition detector
host-resource-watchdog.py System resource monitor daemon
deploy.sh Deterministic deployment script
test-entropy-breaker.js JS test suite (16 tests)
test-watchdog.py Python test suite (14 tests)

Uninstall

Remove the injected block from openclaw.mjs (between the [LLM_STREAM_GUARD_START] and [LLM_STREAM_GUARD_END] markers), delete $OPENCLAW_APP_DIR/dist/llm_stream_guard/, and kill any running host-resource-watchdog.py process.

Usage Guidance
This skill's code is coherent with its description, but it is invasive: it injects code into your gateway binary and launches a watchdog that can kill processes and collects system logs. Before installing: (1) review the injected code and the full watchdog script (you have a backup step in deploy.sh — keep that backup), (2) run it in a staging environment and exercise the test suites, (3) confirm you accept local log/incident collection (journalctl/dmesg) and the watchdog's kill policies, and (4) ensure the user running the gateway has appropriate permissions (journalctl/dmesg access may require elevated rights). If anything is unclear, ask the author for provenance or run the deployment in an isolated environment first.
Capability Analysis
Type: OpenClaw Skill Name: llm-loop-breaker Version: 1.0.0 The skill implements a defensive watchdog and entropy monitor but employs high-risk behaviors, specifically patching the primary application binary (openclaw.mjs) and executing system-level commands to monitor and kill process trees. It also collects sensitive system information, including kernel logs via dmesg and system logs via journalctl, for local incident reports. While these actions align with the stated purpose of preventing resource exhaustion and hallucinations, the invasive installation method and broad process-management capabilities represent a significant security risk.
Capability Assessment
✓ Purpose & Capability
The name/description (entropy-based stream breaker + host watchdog) match the required binaries (node, python3, bash), the files present (Node.js breaker and Python watchdog), and the install actions. OPENCLAW_APP_DIR is used as the target installation path; nothing requested appears unrelated to the stated purpose.
ℹ Instruction Scope
SKILL.md and deploy.sh explicitly instruct the agent to copy files into the Openclaw installation and inject a code block into openclaw.mjs which patches global.fetch and starts a Python daemon. Those instructions are invasive (modify the gateway binary) but are directly tied to the described functionality; they do not reference unrelated secrets or remote endpoints. The watchdog captures local system logs (journalctl, dmesg) into the workspace for incident snapshots — this is expected for forensic snapshots but worth noting as sensitive.
✓ Install Mechanism
No external downloads or unknown URLs: deploy.sh copies local files into $OPENCLAW_APP_DIR and installs psutil via apt-get or pip if needed. The install is deterministic and idempotent as claimed; no archive extraction from untrusted hosts detected.
ℹ Credentials
The only required env var is OPENCLAW_APP_DIR (used as an installation root) and the watchdog optionally uses OPENCLAW_WORKSPACE. No unrelated credentials are requested. However the skill writes into the Openclaw installation, uses HOME to place logs (~/.openclaw), creates an agents sessions directory, and collects system logs — these are legitimate for a supervisor but may expose sensitive local/system data, so confirm you’re comfortable with that scope.
⚠ Persistence & Privilege
The deploy script injects code into openclaw.mjs (modifies an existing gateway binary) and the injected code will start a background watchdog daemon on gateway startup that can kill gateway child processes based on heuristics and write forensic snapshots. This is consistent with the skill’s purpose but is a high-impact, persistent change: it can cause service disruption if heuristics misfire. The skill is not marked always:true, but it does gain ongoing privileges by modifying the gateway.
How to Use
  1. Make sure OpenClaw is installed (local or Docker)
  2. Run the install command in chat: /install llm-loop-breaker
  3. After installation, invoke the skill by name or use /llm-loop-breaker
  4. Provide required inputs per the skill's parameter spec and get structured output
Version History
v1.0.0
- Initial release of LLM Stream Guard with two-layer protection for the Openclaw gateway. - Adds Node.js-based Stream Entropy Breaker to detect and abort hallucinating LLM streams. - Introduces Python-based Host Resource Watchdog daemon to monitor CPU, RAM, disk, and process health. - Deterministic deploy script for safe, idempotent installation and easy verification via automated test suites. - Comprehensive uninstall instructions included.
Metadata
Slug llm-loop-breaker
Version 1.0.0
License MIT-0
All-time Installs 0
Active Installs 0
Total Versions 1
Frequently Asked Questions

What is LLM Loop Breaker?

Two-layer defense for Openclaw gateway: kills hallucinating LLM streams via entropy analysis, and watchdogs host CPU/RAM/Disk to prevent resource exhaustion. It is an AI Agent Skill for Claude Code / OpenClaw, with 81 downloads so far.

How do I install LLM Loop Breaker?

Run "/install llm-loop-breaker" in the OpenClaw or Claude Code chat to install it in one step — no extra setup required.

Is LLM Loop Breaker free?

Yes, LLM Loop Breaker is completely free, licensed under MIT-0. You can download, install and use it at no cost.

Which platforms does LLM Loop Breaker support?

LLM Loop Breaker is cross-platform and runs anywhere OpenClaw / Claude Code is available (linux).

Who created LLM Loop Breaker?

It is built and maintained by hehkcaspar (@hehkcaspar); the current version is v1.0.0.

More Skills
self-improving agent
Captures learnings, errors, and corrections to enable continuous improvement. Use when: (1) A command or operation fails unexpectedly, (2) User corrects Clau...
⬇ 463283 · by pskoett
Skill Vetter
Security-first skill vetting for AI agents. Use before installing any skill from ClawdHub, GitHub, or other sources. Checks for red flags, permission scope, and suspicious patterns.
⬇ 259410 · by spclaudehome
Polymarket
Query Polymarket prediction markets. Check odds, find trending markets, search events, track price movements.
⬇ 232503 · by joelchance
Self-Improving + Proactive Agent
Self-reflection + Self-criticism + Self-learning + Self-organizing memory. Agent evaluates its own work, catches mistakes, and improves permanently. Use when...
⬇ 200423 · by ivangdavila
Github
Interact with GitHub using the `gh` CLI. Use `gh issue`, `gh pr`, `gh run`, and `gh api` for issues, PRs, CI runs, and advanced queries.
⬇ 191115 · by steipete
ontology
Typed knowledge graph for structured agent memory and composable skills. Use when creating/querying entities (Person, Project, Task, Event, Document), linkin...
⬇ 190156 · by oswalpalash

💬 Comments