Exit codes & reliability

Exit codes

code	when	what an agent should do
0	yes / decided (or a batch where every input succeeded)	act
1	at least one no	take the other branch
3	unsure: in the band between no and yes, or a choice/score below min_confidence	check it yourself or ask the user
4	any error: no profile, missing $VAR, network, HTTP error after retries, invalid reply, bad usage	say the call failed; never treat it as a no

The rule the code enforces: an error exits 4, never 1. A dead endpoint cannot read as “no” in an if.

$ jevx is "Is this urgent?" --profile nope --in "x"; echo "exit $?"
jevx: no profile "nope" (have: jev, myjev)
exit 4

A batch exits 0, or 4 if any input failed; the failed line carries an "error" field instead of answers.

Retries

HTTP 429, 5xx and network errors are retried with backoff, retries 3 times by default, timeout_s 60 per request. Other HTTP errors (401, 400) fail at once with the status and the first 300 characters of the body.

Fail-closed validation

A reply is rejected, and the call is an error (exit 4), unless:

• it answers exactly the questions asked, with nothing extra;
• every probability is in [0, 1] and each distribution sums to 1 (±0.02);
• a choice names one of the offered keys.

A malformed or partial answer never becomes a verdict.

Hooks fail open

Plugins are the opposite on purpose: in act mode an endpoint error lets the tool call or the turn proceed (the hook exits 0), and shadow mode scores in a detached process so it cannot delay the agent. stop-judge never blocks the same turn twice. See Plugins.

Tests

go test ./... covers reply validation, retries, the Stop gate, the section parser, context merging and plugin conditions against a fake server.

There is no published accuracy benchmark for any question. The maintainer’s notes record that a small personal model answered several cookbook recipes wrongly, which is the honest summary: the verdicts are only as good as the model behind the profile.