MergeQueue Triage — Jev-powered PR triage gate for agent-flooded repos
The problem
Coding agents write PRs faster than humans can review them. Faros AI: high-AI-adoption teams merged 98% more PRs but review time rose 91%; 2026 follow-up: median time-in-review up 400%. LinearB 2026: AI PR 30-day merge rate 32.7% vs 84.4% unassisted. A May 2026 arXiv study of 33,596 agent PRs found 61.38% got zero review activity.
The idea
Don't generate more review prose — triage. A GitHub Action asks Jev a small battery over deterministic signals (CI status, diff size, files touched vs ticket scope, coverage delta, bot author): safe to fast-track? (noul), risk level (score), routing (choice: auto-merge / normal / deep review / block). Safe diffs auto-merge; risky PRs get routed to the right human with a one-line reason. Thresholds in code, every decision logged with probabilities.
Why now
Prose-generating reviewers were tried and measured: an ICSE 2025 study across 4,335 PRs found they increased average closure time from 5h52m to 8h20m. The missing piece is calibrated low-latency judgment — y0usaf/pi-jev already demonstrates the pattern: four Jev questions in one ~300ms round trip with shadow mode, caching, and fail-open paths.
Build it with
- Jev (TypeSafe AI)choice/score/noul triage judgments; batch the battery into one request like pi-jev does.
- GitHub Actions + Checks APITrigger on PR events; deterministic signals (tests, lint, coverage) stay in code.
- pi-jev patternShadow mode first (log what would have auto-merged), enforce later; per-repo threshold config.
Evidence
- AI code review bottleneck (dev.to)
- Review agent-written PRs in 10 minutes (AI Tech Connect)
- y0usaf/pi-jev
Get the week's best AI launches, plus 3 ideas worth building
One email every Saturday. Ranked by traction, not hype. Free.