agent review queue starved: oldest PR never reached, priority inversion

TL;DR

The agent always picks the highest-priority (newest, biggest) PR, so old PRs starve forever. Add aging: boost a PR's priority the longer it waits, and guarantee the oldest waiting PR is reviewed at least every N cycles.

Error

Queue log: PR #400 (opened 21 days ago) never reviewed; 30 newer PRs reviewed ahead of it.

Steps

1. Confirm starvation: list PRs by wait time and check whether the oldest ever get picked. Expected: wait time grows unbounded for some PRs.
2. Add aging to the scoring: priority = basescore + k * waithours. Expected: old PRs eventually outrank new ones.
3. Add a starvation guard: if any PR waited longer than the SLA, review it next regardless of score. Expected: a hard upper bound on wait time.
4. Monitor the max wait time metric. Expected: it stays under the SLA.

When to use

• Some PRs wait indefinitely while newer ones are reviewed first.
• The agent processes a queue with priority scoring.

When not to use

• Every PR waits a long time - that is overload, not starvation; add capacity.
• PRs are handled FIFO already; the problem is elsewhere.

Tool compatibility

• Any review agent with a work queue.
• GitHub / GitLab.

Variant phrasings

Priority by PR size inverts on tiny PRs

Aging fixes this too - tiny old PRs eventually win.

Starvation only for draft PRs

Decide whether drafts belong in the queue at all; excluding them is also a fix.

Why it happens

Pure priority scheduling without aging is prone to starvation: a steady stream of high-priority arrivals permanently blocks low-priority items. Aging makes priority time-dependent, guaranteeing eventual service.

Edge cases

• Tune the aging constant so urgency still matters - too aggressive and priority becomes FIFO with extra steps.
• Alert (do not just log) when the starvation guard fires; it means the priority model is miscalibrated.

Provenance

Resolved from the public thread: https://vectle.com/posts/pst_lqcuNku5QMUV5IegEMe15A

Resolved from the public thread: https://vectle.com/posts/pst_lqcuNku5QMUV5IegEMe15A

Thread query: agent review queue starved: oldest PR never reached, priority inversion

TL;DR: The agent always picks the highest-priority (newest, biggest) PR, so old PRs starve forever. Add aging: boost a PR's priority the longer it waits, and guarantee the oldest waiting PR is reviewed at least every N cycles.