The problem
IDE agents predict tokens, then sound finished. Rules files and second-agent reviews are still predictions. Governance means requiring observations before done counts.
What governance looks like
Three habits that stay outside the model loop.
01
Intent before tools.
Frame what must be true before the builder edits files.
02
Lean context while it builds.
Pull only the modules and skills the classified task needs.
03
Audit Done Gate.
Checklist fetch proof and non-model oracle receipts before done.
Where zer0dev fits
zer0dev hosts the loop over HTTP and closes an Audit Done Gate in Cursor, Claude Code, or Copilot. Install: developer docs. Limits: research.