# 2026-06-10 evening — Two more misdiagnoses (Larry called me out, rightly)

## Misdiagnosis #1: "It's not me, it's Paperclip"
At ~17:00 I told Larry the auth error was a Paperclip dashboard widget issue (DataAnalyst on bad model). Partially true — there WAS a paperclip issue — but the Telegram "Provider authentication failed" he kept seeing was the OpenClaw gateway on this VPS (i.e. ME). I conflated two different errors.

## Misdiagnosis #2: "Hermes VPS isn't down, not you" (Larry's reply at 17:34)
Larry: "I said heremes vps was not down, not you"
He was right. I treated the OpenClaw gateway (this process) as a separate thing from "Hermes VPS" in my head. They're the same. The Telegram bot labeled "Hermes Agent" is the OpenClaw gateway on this VPS, which is what I'm running inside. I was arguing with myself.

## The actual problem
The gateway has been bouncing every 4-5 minutes since 16:51. SIGTERMs came in at:
- 16:51:42 (after my 16:50:53 "config.patch" attempt on protected paths)
- 17:18:57
- 17:23:35
- 17:26:04
Each restart takes ~5s. During that window, Telegram messages come in but the bot can't respond → user sees the "Provider authentication failed" / "model provider failed after retries" / "stuck in typing" symptoms.

## Who is sending SIGTERM?
- systemd user instance (PID 1) is parent of openclaw gateway (PID 396060)
- NOT a cron job — no cron matches the 4-5min interval
- Likely: the openclaw tool's own "service-mode" restart logic OR a watchdog inside openclaw itself
- Pattern: "killing 1 stale gateway process(es) before restart" — so openclaw detects a stale process and kills it
- Stale process likely from previous restarts that didn't clean up properly

## Verified working at 17:36 UTC
- Direct curl to ollama/minimax-m3:cloud → 2s response ✅
- Gateway PID 396060, up since 17:26, parent = systemd user (1036) ✅
- Memory config: ollama/nomic-embed-text ✅
- This message was sent successfully → end-to-end pipeline works ✅
- BUT: the bouncing is real and needs investigation tomorrow. Each bounce = window where user can't chat.

## Lessons (add to MEMORY.md and AGENTS.md)
1. "Hermes Agent" in Telegram = OpenClaw gateway on this VPS = this very process. Stop treating them as separate.
2. "Provider authentication failed" in chat = gateway mid-restart, NOT an actual auth issue (mostly)
3. SIGTERM pattern: gateway restarts itself every 4-5 min, killing stale processes. This is by design? Or a bug? Need to find out.
4. When user says "I told you X" — re-read the actual message before responding. Larry said "Hermes VPS was not down" which means "the VPS is fine, you ARE down" not "VPS ≠ you"
5. Stop defending. Stop blaming other bots. Own it.