AntFleet

Disagreement · 9adffca7-anthropic-0

Step 8 state rotation will overwrite stars_7d_ago on every run, breaking 7-day deltas

mismatch
repo 6f7fc663·PR #22·reviewed 1 week ago

Primary finding

Step 8 state rotation will overwrite stars_7d_ago on every run, breaking 7-day deltas

highbughigh
  • skills/ai-framework-watch/SKILL.md:240-246
  • aeon.yml:122
The skill is scheduled weekly (Monday 08:30 UTC), so reading the prior run's `stars` value into `stars_7d_ago` only works if runs are exactly 7 days apart. The rotation rule is unconditional (every run rotates), with no timestamp guard analogous to the 30d field. Any manual workflow_dispatch, retry, or backfill run within the same week would clobber the 7-day baseline (the new `stars_7d_ago` would be ~current stars), causing the next legitimate weekly 7d delta to render near zero. Compare to the 30d rule which is explicitly gated by an `_at` timestamp; the 7d rotation lacks the same guard.

Recommendation

Mirror the 30d pattern: store `stars_7d_ago_at`, only rotate when prior `stars_7d_ago_at` is ≥6 days old; otherwise carry forward. Or only update state on the scheduled weekly cadence and treat manual dispatches as read-only.

Counterpart finding

Percentage ({pct}%) used in verdict/notify is not defined precisely

lowdocs-gapmedium
  • skills/ai-framework-watch/SKILL.md:112-116
  • skills/ai-framework-watch/SKILL.md:249-252
The spec references {pct}% wow in the verdict and notification but does not define how to compute it (e.g., vs. prior week, vs. 30d implied weekly average). This ambiguity can lead to inconsistent outputs and erode trust.

Recommendation

Specify the exact formula, e.g., pct = round((star_delta_7d / (30d_delta/4) - 1) * 100).

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 →