The distinction that matters
A coding agent can finish speaking without finishing the work. Conversely, a process can exit unsuccessfully after producing useful partial work. Neither a confident final message nor an exit code independently establishes that the requested result is correct.
Commandem separates the agent turn, the verification gate and the review decision. This makes a failure inspectable instead of translating every stop into success. It does not turn an arbitrary shell command into a complete specification.
Write the acceptance contract first
A useful gate names an observable property and the command that checks it. A file-existence check proves a file exists. It cannot prove that the file contains a sound design. A browser screenshot proves that pixels were produced; without additional checks, it does not prove that the page has no errors or works with a keyboard.
The verified-run guide deliberately starts with a bounded structural check, then asks you to replace it with your project’s real acceptance tests. The optional browser verifier combines HTTP success, browser diagnostics and a screenshot. Its evidence still has a scope: it is not an exhaustive interaction, accessibility or security audit.
Evidence should explain its own limits
Keep the requested change, selected account, verification command and observed outcome together. A receipt is a record of what ran and what was observed. It does not certify that all unstated requirements passed. Inspect the diff before landing, preserve a rejected gate result and distinguish a retry from the original attempt.
For unattended work, explicitly select isolation and budget limits. An example using --worktree is not a promise that every launch mode enables isolation by default. Review the actual mode and generated plan.
Where this leaves the product
The current pre-release provides verification and receipt paths. Professional launch documentation must make their boundaries visible, especially where providers differ. Our launch checklist therefore tests negative paths as well as happy paths: failed clipboard writes must not say “Copied,” and a failed registry must not look like an empty, successfully loaded catalog.
Read the provider capability matrix before relying on a feature across accounts or agents. This article describes the reviewed source, not an announcement of a public release.