I let a 322M local model judge every shell command my coding agent runs

Description and use

<h2> laya-guard: A local safety net for coding agents </h2> <p>Coding agents run shell commands all the time. Most are harmless: <code>ls</code>, <code>git diff</code>, <code>npm test</code>. But every now and then, an agent runs <code>rm -rf</code> in the wrong directory or copies a <code>curl ... | sh</code> command from a README without thinking twice.</p> <p>The usual options aren't great. You either approve commands one by one or turn approvals off and hope for the best.</p> <p>I wanted something in between. Something that checks every tool call before it runs, stays on my machine, and do

Problem

<h2> laya-guard: A local safety net for coding agents </h2> <p>Coding agents run shell commands all the time. Most are harmless: <code>ls</code>, <code>git diff</code>, <code>npm test</code>. But every now and then, an agent runs <code>rm -rf</code> in the wrong directory or copies a <code>curl ... | sh</code> command from a README without thinking twice.</p> <p>The usual options aren't great. You either approve commands one by one or turn approvals off and hope for the best.</p> <p>I wanted something in between. Something that checks every tool call before it runs, stays on my machine, and do

Outcome

<h2> laya-guard: A local safety net for coding agents </h2> <p>Coding agents run shell commands all the time. Most are harmless: <code>ls</code>, <code>git diff</code>, <code>npm test</code>. But every now and then, an agent runs <code>rm -rf</code> in the wrong directory or copies a <code>curl ... | sh</code> command from a README without thinking twice.</p> <p>The usual options aren't great. You either approve commands one by one or turn approvals off and hope for the best.</p> <p>I wanted something in between. Something that checks every tool call before it runs, stays on my machine, and do

Evidence

I let a 322M local model judge every shell command my coding agent runs

Source