Make your coding agent prove each step before saying done
A developer built vstack, a small command-line guide for Claude Code and Codex; its README shows it stopping an agent that claimed 'done' without running the code.
Done with Claude Code
What they did
A developer liked an existing setup called pstack, where agents follow fixed steps and must show proof before calling work done. They built a smaller version, vstack, for Claude Code, Codex and Cursor. The agent gets one step at a time, such as plan, build and verify. To finish a step it must describe what it actually ran and saw, and the tool runs the project's tests and code checks.
How it went
The README shows a real small task. The agent first said 'it should work now' and was told to report what it observed. After it did, the tool still blocked it for a loose type, a swallowed error and a needless comment.
Worth knowing
The author notes vstack gives the agent far fewer instructions than pstack, and says this is a different design, not a better one. Pushing code still needs the person's permission.
Try it yourself with Claude Code
Work on [task, e.g. add CSV export] in [repo] in clear steps: goal, read the code, design, build, verify, review. Do not say a step is done until you show proof: quote the commands you ran and what you saw. Before marking verify done, run the tests, type checks and linter, and fix any loose types, empty error handlers or comments that just repeat the code. Show me the final summary and proposed commits, and wait for my approval before pushing anything.
Swap the highlighted [parts] for your own details.
Source: X · @valentinoswe · Oct 17, 2026
Written with AI assistance from the original post and checked against our editorial guidelines. Results are as reported by the source, not independently verified. Spotted a mistake? Tell us.
