agent-code-review

Deterministic discovery — candidate PRs found via gh on their own cadence, never involving the LLM; approved PRs are skipped

Eligibility holds — a quiet period stops reviews mid-rebase, a re-review cooldown gives the author room to respond to the last one

Durable DuckDB queue — candidates, positions and review history survive restarts, which is what powers Refreshed detection

Pluggable engine — codex or claude, same verdict contract and same review log, so switching is a one-key config change

Live review logs — the engine tees into the workspace, so an in-flight review streams via queue log -f or the dashboard

Usage floors — the loop pauses when the engine's rate-limit window drops below a threshold, and resumes when it refills

Every engine metered side by side — see whether the one you are not using has more headroom before switching

Per-review spend recorded — token counts and API-rate cost, so a budget comes from your own data rather than a guess

Author groups as policy — ignore, comment or approve, per repo, each carrying its own engine and extra instruction

Serve with a dashboard — always-on daemon, optionally on your tailnet; most config reloads live within ~30s

Everything is config — repos, groups, thresholds, cadence, prompts and rules; no handles or repos hardcoded

doctor checks every runtime assumption at once, and serve runs the same checks at boot