Reports and debugging

Console

One line per Step, the reason under a failed one, and a Failures block at the end with a rerun command per failed scenario. A marginal pass (the check passed, but Jev was barely confident) is listed separately; treat it like an unclear wording, or fail it with --strict.

The run directory

Each run writes <reportDir>/<timestamp>/ (reportDir defaults to runs, see configuration):

trace.jsonl: Observations, Jev answers, actions and outcomes. Each scenario's result event has perf: milliseconds spent in observe, perform:<verb>, after:<verb> (waiting after an action) and settle, split by what held it (settle:requests:<type>, settle:timers, settle:quiet).
log.txt: the full decision log (what -v prints).
Per failed scenario: a screenshot per Turn (with --record none, of the failed Step only), playwright-trace.zip, console-errors.txt and ui.txt (the page text). Passed scenarios keep none.
report.md when something failed: every failure with Jev's answers, the page text around it and a rerun command, for a coding agent. --no-agent-report turns it off.

Outcomes

Outcome	Meaning
`PASSED` / `DONE`	The check held / the action finished.
`ASSERTION_FAILED`	The app is wrong: the value or text is not what the Step expects.
`BLOCKED`, `ACTION_FAILED`	The action could not be done.
`ELEMENT_NOT_FOUND`	Nothing on the page matches what the Step names.
`STEP_TIMEOUT`	The Step took too many Turns without finishing.
`PAGE_TOO_LARGE`	The page is over the model's request limit.
`AMBIGUOUS`	Jev could not tell which element the Step means. Reword.
`UNVERIFIABLE`	The check quotes nothing code can compare. Reword.
`UI_CHANGED`	A replayed resolution no longer matches the page (`--record none`).
`JUDGED_PASSED` / `JUDGED_FAILED`	A fuzzy check judged by Jev (`--judged`).

The agent skill shows how to classify a failure: the app is wrong, or the wording is unclear.