Initial commit: homelab infrastructure wiki

- Full Obsidian vault content
- Host configs (ice, grizzley, ubuntu, proxmox, truenas, panda, hyte)
- Media stack documentation
- Traefik HA setup
- Automation scripts
- Bachelor party planning

homelab/entities/hermes-gateway.md

@@ -0,0 +1,71 @@
+---
+title: hermes-gateway
+created: 2026-04-28
+updated: 2026-04-29
+type: entity
+tags: [services, ai, gateway, watchdog]
+sources: []
+---
+
+# hermes-gateway
+
+**Role:** AI gateway — routes LLM requests across multiple providers
+**Hosts:** [[ice]] (primary), [[grizzley]] (secondary)
+**Runs on:** ice as systemd user service (`hermes-gateway.service`)
+
+## Overview
+
+hermes-gateway is the AI gateway that routes LLM requests (DeepSeek V4, OpenAI, Anthropic, OpenRouter, etc.) across multiple providers. It has a watchdog pattern deployed via system cron on both [[ice]] and [[grizzley]].
+
+## Providers
+
+| Provider | Model | Endpoint | Notes |
+|----------|-------|----------|-------|
+| DeepSeek | V4 | `https://api.deepseek.com/anthropic` | Anthropic format, 1M input / 384K output |
+| OpenAI | various | `https://api.openai.com` | |
+| Anthropic | various | `https://api.anthropic.com` | |
+| OpenRouter | various | `https://openrouter.ai/api` | |
+
+## Watchdog Pattern
+
+A shell script (`/home/bear/hermes-gateway-watchdog.sh`) runs via **system cron** on both ice and grizzley:
+
+1. Checks if hermes-gateway is responsive
+2. On failure: direct restart → tmux+OpenCode rescue if still down
+3. Sends Telegram notification on failure to topic **1033 "Cron Jobs"** in AigentZeroHermes (`-1003820156994`)
+
+**Telegram alert details:**
+- Bot token: `836803270:AAH-Ac5Y`
+- Chat ID: `-1003820156994` (AigentZeroHermes channel)
+- Topic ID: 1033 ("Cron Jobs")
+
+**Critical note:** On [[grizzley]], the systemd override for the watchdog is deployed directly to `/etc/systemd/system/` (not tracked in the homelab repo — it's a system unit).
+
+## DeepSeek V4 Provider
+
+Configured as: `https://api.deepseek.com/anthropic` (Anthropic format, not OpenAI).
+Context window: 1M input / 384K output.
+⚠️ Known bug: thinking mode passes `reasoning_content` back incorrectly — pass it back in multi-turn.
+
+## Access
+
+hermes-gateway runs as a user service. To check status:
+```bash
+# On ice (primary)
+ssh bear@192.168.50.197 "systemctl --user status hermes-gateway"
+journalctl --user -u hermes-gateway -f
+
+# On grizzley (secondary)
+ssh bear@192.168.50.84 "systemctl --user status hermes-gateway"
+```
+
+Watchdog logs (check cron output in syslog):
+```bash
+ssh bear@192.168.50.197 "grep hermes-gateway-watchdog /var/log/syslog"
+```
+
+## Related
+
+- [[ice]] — Primary host
+- [[grizzley]] — Secondary host with watchdog
+- [[authentik]] — SSO for gateway access (if applicable)