Stop pasting repo links for free-tier CLIs until you have a fail-closed handoff card. A new protocol detailed on DEV.to argues that sharing a tool running on a sleeping free server or a wandering free model is a recipe for disaster. The solution is a strict JSON-based checker that bans the share if any evidence field is missing or unverified. This isn't about grading model quality; it's about answering one narrow question: can another person run one task and stop cleanly without you hovering over their shoulder?

The Gatekeeper Protocol

The core mechanism is a Python script that validates a JSON card against seven specific gates. You cannot proceed if you lack a same-day console screenshot for the model ceiling, a verified dry-run log that proves no files were written, or a rehearsed rollback command. The checker exits with code 2 if any field is blank, empty, or false. This hard failure is intentional; soft warnings train builders to ignore red flags. If you can't point to concrete evidence for each gate, the tool stays on your laptop. Hope is not a field in this JSON file.

Evidence Over Anecdote

Each gate requires pasteable proof stored directly in the card, not in a chat scroll or your memory. For the 'ceiling' gate, you must copy the exact line from the provider console on the same calendar day. A number from a blog post or yesterday's note is explicitly rejected. The 'dry-run' gate demands a clean git status log showing zero changes. If the tool wrote a file during the dry run, it fails. This prevents the common failure mode where a 'safe' preview actually mutates the worktree. The fixture gate requires a log of a known-bad input causing a refusal. If the tool exits with code 0 on garbage input, it's unsafe to share.

Scope and Cost Boundaries

This protocol is strictly for solo builders sharing small, single-task tools. It is not a release train for paying users or a gate for production data changes. The author budgets ninety minutes for creating a card; if it takes longer, the tool is too complex for this lightweight handoff. MonkeyCode is cited as an example provider for the free model lane and server, but the article stresses that live console data trumps any static blog claim. The card is a temporary contract, valid only until the free tier limits change. If the host sleeps or the ceiling line vanishes, the abandon rule triggers, and the share ends.

Key Takeaways

  • A blank evidence field is safer than a wrong number because it forces a hard ban rather than a silent failure.
  • The checker script is a template, not a hosted service; it only validates the JSON structure and strict boolean flags.
  • Rollback commands must be rehearsed locally before sharing; if only you know how to undo the change, the tool isn't ready.
  • Dry-run evidence must prove the tool refused to write files, not just that it didn't crash.

The Bottom Line

A pass means the card is complete, not that the diff is right. You still read the README change with your own eyes; the checker never opened your files and never will. If you cannot rehearse rollback in one command, keep the build on your laptop until undoing it is boring.