Catch unverified claims
Look for statements such as “all tests pass” when the session provides no matching evidence.
AGENT RELIABILITY / GITHUB
“All tests pass.” Did they actually run?
Check an AI coding assistant’s final summary against the actions and evidence in its session.
THE PROJECT, IN PRACTICE
What you can do with it.
Make uncertainty explicit
Flag weasel words that soften a claim without explaining what was checked.
Surface omitted failures
Detect summaries that leave out a failed command or unfinished step while presenting the overall work as complete.
Ask for a grounded correction
When a summary fails the check, the assistant receives specific feedback and can verify the work or correct its report.
FROM IDEA TO USE
How it works.
- 01
Configure Jev
Install the plugin for your assistant and supply a TypeSafe API key through its documented configuration.
- 02
Work normally
The check runs when the assistant produces its final summary.
- 03
Review the correction
Flagged claims are returned to the assistant with reasons tied to the session evidence.
Assistant: “All tests are passing.”
Session evidence: no test command was run.
Feedback: Verify the claim against recorded actions,
or state plainly that tests were not run.Illustrative example. A model-based check is not a guarantee that every incorrect claim will be detected.
TAKE IT FOR A SPIN
Start here.
Choose the installation instructions for Claude Code, Codex, pi, or OpenCode. Configure the key in the plugin’s own protected configuration file and approve the hook where your host requires it.
STATUS & REQUIREMENTS
Supports Claude Code, Codex, pi, and OpenCode. Requires Python 3.10+ and TypeSafe service access. Host-specific behavior and retry limits are documented in the repository.