What it is
Prototype for reviewing what a coding agent did: it reads a trace and flags redundant reads, unjustified retries, loops, and claims of completion that come before the evidence.
What it does: Rules mode runs locally with no key; --judge jev asks Jev for semantic judgments with --max-calls / --max-input-bytes budgets and a local cache; failures and exhausted budgets become uncertain decisions. It can import a turn from your own Codex logs.
Why it belongs in the showcase: Stuck-loop and done-claim checks in one tool, with acceptance evidence that later file changes invalidate.
Caveats: Synthetic fixtures only; accuracy not validated; trace context leaves the machine in Jev mode.
Evidence
Public MIT README; 20 bundled synthetic traces with reference labels; Jev mode pins `jev-1.13.0`, a provisional 0.6 confidence threshold and TypeSafe SDK 0.6.0, with call and input-size budgets; author says 45 tests pass locally.
Caveats
- README says detection accuracy has not been validated and real-data labels have not had an independent second review
- Fixtures and rules were developed together; synthetic benchmark only
- Jev mode sends selected trace context to TypeSafe; imported Codex logs are not sanitized automatically
Unofficial case page. Stars / list position observed for heat; do not treat heat as a quality SLA. Review the upstream README before you depend on it.