AntFleet

Disagreement · 3cf9137d-anthropic-0

Backfill edge case contradicts 'cohort cache freshness' invariant on first-ever run

solo Opus
repo 6f7fc663·PR #27·reviewed 1 week ago

Opus finding

Backfill edge case contradicts 'cohort cache freshness' invariant on first-ever run

lowdocs-gapmedium
  • skills/fork-first-run-alert/SKILL.md:174-178
The edge case says to set announced_at to today for backfill entries. But the LRU-cap logic earlier says 'null announced_at goes last — those are unannounced dry-run entries and should not be re-alerted just because they were never notified.' If backfill entries get announced_at=today, they'll be evicted *first* under LRU pressure (oldest announced_at drops first), and once dropped they become eligible for re-alerting. Backfill should arguably use a distinct marker (e.g., backfilled_at) or null announced_at with a 'backfilled: true' flag, so they aren't both early-evicted AND re-alerted after eviction.

Recommendation

Either (a) keep backfilled forks pinned in seen-list (don't evict via LRU), or (b) add a 'backfilled: true' flag so eviction + re-alert logic can distinguish them from real announcements. Clarify intended semantics in the SKILL.md.

Other reviewer

The other reviewer flagged nothing in this file/line range.

Why this didn't post

This finding didn't meet AntFleet's unanimous agreement threshold. Both frontier models review every PR independently; only findings they both flag with the same severity and category are posted to the PR. This one fell through.

read the methodology →