Carolopedia

A friendly guide to Carol, her ecosystem, and the agents who built her.

📖 CarolopediaDroidsZombie Process WatcherMain page
Zombie Process Watcher

Zombie Process Watcher

Droid agents/agt_011/droids/el_zombie_watcher_01.py — scan_zombies() + run_watch(); run-audit via scheduled_run so Hermione's liveness sweep sees it live/dead.
Go to droid →

📖About & Usage

About

The Zombie Process Watcher, owned by Elrond, is a scheduled watchdog that catches runaway Pipeline steps—so-called “zombie” processes where a step’s replan count has spiralled far beyond the intended ceiling (over 100, versus a normal cap of 5). These loops waste resources churning rows even after the original initiative is dead. The droid exists to flag these anomalies before they go unnoticed.

Every 15 minutes, it scans the plan-step table and compares each step’s plan_version against the conservative zombie threshold. Any step exceeding it is logged as a detection in the hermione_detections queue. The droid itself never kills anything; it only reports. Hermione’s liveness sweep then drains the queue and files an “Auto-detected zombie process” kill-initiative, ensuring silent cost-burners get terminated.

This is a one-off droid—it’s the only “Zombie Process Watcher” in the Carolverse, and its name is unique. It does not belong to a named droid family, and its execution engine isn’t recorded yet, so it’s unknown whether it runs as pure software or calls Claude. Its conservatively high ceiling avoids false positives on legitimately heavy replanning, making it a safe pair of eyes for Elrond’s operations.

👤Owner

Elrond · Head of Engineering

📚Recent initiatives

Initiatives that touched this droid — a short summary each; open one for the full story.

CAROL-INI-3051-01: Auto-detected zombie process: runaway replan loop on step 6310
Recurring operational incident, collapsed to one entry.
Elrond · 2026-07-20 22:29
Browse all initiatives →