aozyildirim/agena/review-backlog-killer

Review Backlog Killer

Pull requests get lost in review queues; team velocity drops silently. Sprint planning shows it as "I dunno, sometimes the reviewers are slow". This pattern surfaces stuck PRs deterministically and sends the reviewer a nudge through whichever channel they actually read.

How to apply this pattern

1. Define open: PR is open when `merged_at IS NULL AND closed_at IS

NULL AND its provider status is in {open, opened, pending, review_required, in_review, active}. PRs with provider status in {abandoned, declined, closed, rejected}` drop off the backlog.

1. Configure thresholds per workspace (chip-pick is best UX —

6h / 12h / 1d / 2d / 3d / 1w / custom). Defaults: 24h warn, 48h critical, 6h between auto-nudges to the same reviewer.

1. Poller every 30 minutes updates age_hours and severity

(info < warn ≤ warning < critical ≤ critical) on every open PR. Merged or closed PRs auto-resolve out of the backlog.

1. Pick a notify channel per workspace:

• slack_dm — direct, most attention
• slack_channel — visible to team
• email — for async-first teams
• `pr_comment` — most attention-grabbing because it lands where

the reviewer already lives. The comment text is locale-aware so non-English teams don't get an English nudge.

• manual — dashboard-only, no auto-notify

1. Auto-escalate once when a PR crosses the critical threshold:

set escalated_at, mark severity, surface above the fold to tech leads. Fire it once per PR, not on every poll cycle.

1. Manual nudge button on every row — fires the configured channel

immediately, overrides the interval. Useful for a tech lead doing a one-shot push before standup.

Example PR comment template

⏱️ AGENA Review Backlog
 
This PR has been waiting for review for **47 hours** (severity: critical).
Nudge #2.
 
Configure thresholds at /dashboard/review-backlog.

When the Reviews module is also on, the message includes AGENA review score: 84 so reviewers see the time investment before clicking.

Notes

• Idempotency on PR comment: before posting, check whether a

prior AGENA comment already exists on that PR — re-post would be spammy. If the comment was deleted by a human, treat the row as un-nudged so the next cycle is allowed to comment again.

• Provider-aware status refresh: PRs marked active in your local DB

may have been abandoned upstream; refresh status before deciding to nudge so you don't ping reviewers on dead PRs.

• Exempt repos are essential — legacy / maintenance repos shouldn't

trigger nudges. Comma-separated repo-mapping ids in a workspace setting is the simplest UX.

• The poller scales linearly with open PR count; 1000 open PRs ≈

1 second of polling per cycle.