I’m reading about the “SoulEngine” and how it uses an LLM to generate automation code dynamically based on the page state. Honestly, letting an AI write and execute code on my machine sounds like a recipe for disaster.
How is this sandboxed? What happens when the AI inevitably writes a broken selector or a syntax error?
It’s a totally valid concern! We built the AI Skill Engine with strict safety rails.
First off, the execution is completely sandboxed within Deno. The AI doesn’t get raw access to your OS at all - it only sees the specific browser page context we hand over.
Second, we built an automatic self-correction loop. When the LLM generates a script, Deno writes it to a temporary file and dynamically imports it. If the AI hallucinates a bad CSS selector and the code throws an exception, the engine doesn’t crash. Instead, it catches the stack trace, bundles it with the current DOM state, and sends it back to the LLM with a prompt saying: “Your code failed with this error, please fix it.” It’s incredibly resilient and usually fixes its own mistakes on the second try.