Hermes (self-healing agent stack)

Hermes — self-healing supervisor stack

Multi-service stack supervising ELC AI Agent instances. Spawned from feat/corone-self-healing-agent branch — keb workspace was the reference implementation, hence keb’s commit author identity = Julian (Hermes Self-Healing Port).

Components (3 systemd units)

Service	Port	Purpose	Status (2026-05-03)
hermes-webui	127.0.0.1:8787	Python supervisor + Web UI (uvicorn)	activating
hermes-dashboard	127.0.0.1:9119	Hermes admin dashboard (/root/.local/bin/hermes dashboard)	active
hermes-cloudflare	(tunnel)	Cloudflare Tunnel exposing hermes.corone.monster → 127.0.0.1:8787	activating

Public endpoint

• https://hermes.corone.monster (via Cloudflare Tunnel — see cloudflared config)
• nginx vhost: /etc/nginx/sites-enabled/hermes.corone.monster

Code locations

• Agent: /root/.hermes/hermes-agent/ (Python venv)
• Web UI: /root/hermes-webui/server.py
• Dashboard binary: /root/.local/bin/hermes
• cloudflared: /root/.local/bin/cloudflared

Function

Watches ELC instances (corone-app:3000, kebahagiaan-app:3001):

• Restart on crash via systemd
• Captures error.unexpected lifecycle event (added in 1.3.7)
• Catches process-level uncaughtException + unhandledRejection
• Dispatches notifications externally

Related ELC commits (on corone-julian branch)

• e2a1365 — feat(lifecycle) dispatch error.unexpected on uncaught exception/rejection
• a6cba6a — feat(notifications) error.unexpected event type for process-level crashes

Why it matters

Without Hermes, an ELC crash → 502 → user manually restarts. With Hermes:

1. Process dies → systemd starts new one
2. Lifecycle event flagged → Hermes notifies
3. Dashboard at :9119 shows recent restart history
4. Cloudflare Tunnel keeps URL alive even if local nginx hiccups

Cross-refs

• elc-ai-agent — supervised target
• elc-release-1.3.7 — release that adds error.unexpected
• nginx-vhost-map — public routing