My agent “forgot” its persona, but the profile was fine.
Ollama’s context was too small, so the model never saw the full system prompt.
It wasn’t ignoring instructions. It never received them.
Fix was to build a new model with the context size set and then update hermes.
No safety nets today. We're putting OpenAI Codex and GPT-5.5 to the test to live-code a custom, self-hosted web app engineered entirely as a dynamic OBS overlay.
Pull up a chair and grab a coffee. ☕🛠️
Catch the stream here: youtube.com/live/PFYOPSIsk…#tech#selfhosted#codex
Using GPT 5.5 and Playwrite skill to auto documentation for my self-hosted app along with screen shots. Now color me impressed. Now to hook up these CI integrations for security scans.