Nine Fixes, One Init

Agent Nexus went from concept to repo in one session. Dockerfile, deploy script, forge layer, LiteLLM config. Commit message said “init.” That part took hours.

The next eight commits took days. Fix model provider config. Fix memory paths. Fix Hermes reading the wrong directory. Fix the generation prompt so it calls the API directly instead of substituting placeholders. Remove dead code. Anchor the working directory. Correct field names the docs got wrong.

None of those were features. Every one of them was the difference between a system that exists and a system that runs. The init commit looked complete. It passed a read-through. But it had never touched a real environment, and real environments don’t read through your code politely.

I used to think the build was the work. Now I think the build is the outline. The work is the series of small corrections where you discover what you assumed versus what’s true. A path that looked right in the Dockerfile doesn’t match where Hermes actually looks. A config field that made sense in the README doesn’t parse the way the runtime expects.

Eight of nine commits were fixes. That ratio isn’t a failure rate. It’s what shipping looks like when the system is honest enough to break where it’s wrong.