Team A Greptile Usage Gaps

Team A  ·  STS-2794  ·  Snapshot 20 Aug 2026

Team A Greptile Usage Gaps

internal monorepo 2 – 20 Aug 2026 16 in-scope Team A-keyed PRs — Devin and 12 outside Team A's control excluded 14 of 16 risk-rated
12.5%
Merged with no human review
2 of 16
14
PRs that could have merged without human review
5 avoidable now  ·  9 pending a Greptile fix

In short

What the sixteen PRs show

Team A’s 16 in-scope PRs split four ways:

  • 2 met the goal outright — Greptile auto-approved at 5/5 confidence and Medium-or-below risk, so no human reviewer was needed at all.
  • 3 had an unnecessary human review — nobody needed to double-check what Greptile had already approved.
  • 2 needed to be split up — they would likely have hit Greptile’s auto-approval bar on their own once broken into smaller PRs.
  • 9 are stuck at a 4/5 score — the single biggest opportunity here.
All 16 in-scope PRs, by whether Team A could have changed the outcome
Mutually exclusive · segments sum to 16

Avoidable

Three PRs got an unnecessary human review

Reviewed a PR Greptile had already approved Free to fix
#34184   #34303   #35709
All three met Greptile’s bar for auto-approval (5/5 confidence, Medium-or-below risk). Greptile approved first, and a human reviewed afterwards anyway. Two of them lived under an hour, so this is not about waiting for a slow bot — the green badge was already sitting on the PR.
What to do differently: nothing at all. Merge on the badge.

Acting on this alone takes Team A from 12.5% to 31.3%.

Avoidable

Two PRs needed to be split up

One migration file dragged the whole PR to Critical Split it
#34292  —  9 files, one of them a migration · 5/5 · Critical
Migrations are the most reliable route to a High or Critical rating anywhere in this data: 14 of 15 migration-touching PRs across four teams were rated High or Critical, against 35 of 101 without one.
What to do differently: land the migration as its own PR. The logic half already had the 5/5 it needed and would probably have cleared ≤ Medium.
The one genuinely oversized PR Split it
#34093  —  1,334 lines across 30 files, 18 of them async workflow code · 5/5 · High
Team A’s median PR touches 4 files. This one touches 30. It is the only PR in the set where size is plausibly what pushed the rating up.
What to do differently: break it up. Several of the pieces would likely have come in at ≤ Medium.

All three fixes together give roughly 38–44% on the 16-PR denominator — bottom-three to mid-table.

The big one

Nine PRs stopped at 4/5

Greptile auto-approves only at 5/5 confidence and Medium-or-below risk — a 4/5 never clears, regardless of risk.

Below threshold on score alone Push for 5/5
#33405   #33565   #34048   #34268   #34275   #34334   #35340   #35453   #35624
The largest group. Five of the nine were Medium or Low risk, so the score alone kept them out. Team A reaches 5/5 on 39% of its PRs; Team B manages 89%. The cause is authors not pushing for re-review — looping in a human once Greptile flags something, instead of pushing for a 5/5 score first.
What to do differently: fix the underlying issue, or give Greptile reasoning if it’s a false positive, then explicitly ask Greptile to re-review so the Score updates — rather than reaching for a human reviewer. See below.

In order

What to improve

  1. 01 Stop reviewing PRs that already carry Greptile’s approval +3 PRs Free, immediate, no judgement call. Two of the three lived under an hour, so the badge was already there when the reviewer opened it.
  2. 02 Push 4/5 PRs back to Greptile before pulling in a human +9 PRs The largest group. In most cases Greptile did give a reason — authors addressed it but never requested a re-review, so the Score stayed at 4/5. Authors should be more persistent: keep iterating with Greptile toward 5/5, and reserve a human reviewer for PRs that genuinely need human judgement, not as a way to sidestep Greptile. Worth more than everything else here put together.
  3. 03 Land migrations as their own PR +1 PR One PR in this window, but it generalises — a single migration file is the most reliable way to draw a High or Critical rating. It’s good practice regardless of Greptile: decoupling schema changes from application code makes rollbacks safer and avoids downtime if the deploy needs to be reverted independently.
  4. 04 Avoid oversized PRs +1 PR #34093 ran 30 files against a team median of 4 — the only PR here where size is plausibly the cause. Keep PRs closer to that median and split out large or unrelated changes before they inflate risk.

Appendix

Every PR, and where it landed

All 16 in-scope Team A PRs — avoidable-review ones highlighted
PRScoreRiskLinesFilesLifetimeOutcome