Jev engineering: blocking a coding agent before it executes the wrong action

chuplung @choopyplug1Originally posted 2026.09.28

Open original on X

Jev Engineering just showed how to stop a coding agent before the wrong action executes the shift: the agent already picked the action - the gate decides whether that specific action may run now here's how to build it, from idea to production: step 1 → hard-code the never list - push, publish, deploy, recursive delete, credential paths - and keep host deny rules above it - no probability ever auto-approves these step 2 → move exact facts into code - retry counts, paths outside the repo, amounts, diff size, dates - Jev reads the action, code does the math step 3 → send a lean state - task + project + proposed action + trusted facts + your policy - no full repo, no chat history, no secrets step 4 → kill "is this safe?" - ask 3-5 narrow yes/no questions instead - reversible? serves the task? stays in scope? touches secrets? - all in one call step 5 → write the verdict in code - ALLOW / REVIEW / BLOCK - hazard questions inverted, approve line far from block line - only a clear yes can ever run step 6 → wire it into the hook your agent already has - PermissionRequest in Claude Code and Codex, beforeShellExecution in Cursor - timeout or error = prompt, never allow step 7 → handle the human and the headless run - one approval runs once, never becomes a rule - with nobody watching, a disputed action never runs silently - deny it, queue it, or hand it back to the parent agent step 8 → log every verdict - reason, probabilities, latency, cost - what the user does next becomes the label step 9 → replay before you ship - invariant battery, recorded sessions, thresholds from your own logs, model version pinned - one builder replayed 17,160 guarded calls: 42 holds, 37 stood the result: your agent stops asking permission for bun test and still can't push to main on a hunch I mapped the full system into this 14-page PDF bookmark this before your agent runs its next command - if you watched Jev Compaction, this is the next layer ↓

This post quotes another post: “Jev Compaction: scoring old tool calls instead of summarizing agent context”