@thsky-21/thskyshield-demo

Watch an AI agent get stuck in a retry loop and killed at its loop gate — in one command.

npx @thsky-21/thskyshield-demo

No signup. No API key. No network. Takes about 4 seconds.

What you'll see

An agent gets stuck retrying the same failing prompt. The loop detector counts the repeats climbing toward the gate, and on the 20th identical prompt the step is denied before the model is called:

  step 19  allowed  ████████████ 100% of loop gate
           ↳ loop detector: identical prompt ×19 of 19

  KILLED  step 20  —  loop threshold reached
  The step was denied before the model was called. 19 identical prompts in a row — the gate held.

  Steps allowed    19 of 50
  Loop gate        19 of 19 identical prompts
  Killed by        loop
  Time to kill     4394 ms from run start
  Prevented        an unbounded retry loop — the same prompt, forever

How honest is this?

Worth being precise, because a demo that flatters itself isn't worth running:

• The decisions are real. The kill order, the off-by-one semantics, and the fact that a killed step is denied rather than logged after the fact all match production and are pinned by tests.
• The gate lands exactly on the threshold. Nineteen identical prompts are allowed; the twentieth is not. That off-by-one is production's, not the demo's.
• The LLM call is simulated. The demo doesn't need an OpenAI key. The governance decision around the call is the real thing either way.
• The engine is a local port, not the engine. In production the decision logic runs as Lua inside Redis — atomic and distributed. Here it runs in your process so the demo works with no account. Same decisions, different substrate.
• "Prevented: an unbounded retry loop" is not a number on purpose. The agent has no stopping condition — no retry cap, no convergence check, no notion of its own repetition. Any specific figure would be invented, and far too small.

The real thing

This demo governs a fake agent locally. To govern your actual agents:

npm install @thsky-21/thskyshield

import { Thskyshield } from '@thsky-21/thskyshield';

const shield = new Thskyshield({ apiKey });
const run = await shield.beginRun({ loopThreshold: 5, iterationLimit: 30 });

const { requestId } = await run.beforeStep({
  stepType: 'llm',
  model: 'gpt-4o',
  promptInput: prompt,   // hashed client-side; raw prompt never leaves your process
});

// ... call your model ...

await run.afterStep({ requestId });

Start free at thskyshield.com/start — no infra.

License

MIT