Engineering Program Manager
Cut a backlog to the issues a team will actually act on this sprint, without losing a real regression.
-
1 Cluster duplicates Gemini 3 Pro 640ms
MOB-1247, MOB-1198 and MOB-1302 describe the same crash from three angles — the stack traces differ but all three land in the same payments call. Grouping them.
-
2 Classify severity Claude Opus 4.8 1180ms
Reporter marked it P1. This team’s rubric reserves P1 for revenue-blocking, and none of the three mention revenue — so P2. Not applying that rule to other teams.
-
3 Check for a regression Claude Opus 4.8 910ms
MOB-1302 cites a commit SHA and a version that worked. That makes it a regression, not a duplicate report — pulling it back out of the cluster and escalating.
-
4 Draft the consolidation Claude Opus 4.8 1340ms
One issue for MOB-1247 + MOB-1198, links back to both. Left MOB-1302 open and tagged it for the EM.
Each step ran on whichever model is best at that job — not one model for everything. The worker is the workflow.
Permissions & guardrails- Read: issues, labels, commit history and the team’s own severity rubric — only on the boards it was added to
- Write: comments and labels only; a proposed consolidation is left as a suggestion, never written into an issue body
- Escalate: the EM when a suspected duplicate cites a commit SHA, or when the rubric and an existing label disagree
- Never: close, reopen or reassign an issue without a human approving the change first
- Never: act on a board it wasn’t explicitly granted, even when a linked issue points to one
- Proposed MOB-1320 — consolidates MOB-1247 and MOB-1198, severity P2.
- Held MOB-1302 back as a regression and paged the EM. It looked like a duplicate.
- Backlog: 47 open → 46, with 1 escalation. No issue closed without a human.
86.5%
a general model gets 74.1% · +12.4pp
74 · 77.8 · 81 · 83.4 · 85.2 · 86.5
private rules were taught in one workspace and never leave it. general rules are true about the work, so every copy of this worker gets them.