bro self judge
A calibrated decision judge for agent loops — typed questions in, typed answers with confidence out.
The judge answers typed questions (choice / score / noul) over a
compact state and returns typed answers with confidence — sub-second,
priced in fractions of a cent. The resolved chain is a primary backend
plus judge.fallback escalation on unanswered questions and
low-confidence answers; providers come from the
registry via judge.provider/judge.model,
with the legacy judge.{baseUrl,apiKeyEnv} + judge.llm shape still
accepted (synthesized into anonymous api entries, with a deprecation
warning).
| Command | What it does |
|---|---|
bro self judge decide --state <file|-> --questions <file> [--connector N] [--json] | One decide() over the resolved chain — answers, decidedBy, model, latency, usage. The smoke path: exercises the connector regardless of judge.mode |
bro self judge stats [--since ISO] [--json] [--replay] | The verdict-journal report: agreement matrix (judge action vs recorded outcome), calibration buckets, p50/p95 latency, cost per provider+model |
bro self judge replay [--pr N…] [--merged-since ISO] [--limit N] [--connector N] [--json] | Re-judge archived review threads from merged PRs on the live chain, journal replay verdicts with inferred outcomes, print the replay-scoped stats. Idempotent — a replayed subject is never re-judged |
judge.mode governs consumers, not explicit invocation: shadow
annotates bro review act threads with a judge: … line under each unresolved
row (journaled so stats can score judge-vs-outcome agreement);
off (default) decides nothing on its own. See
judge config.