A model can describe a part. Can it edit one?
Models talk about printable parts fluently and misread their geometry. So my Hermes 3D-print skill measures before it edits, and says what it hasn't checked.
A strange personal engineering journal
I build systems I can trust.
Python engineer in Ukraine. Self-hosted infrastructure, dependable agents, and honest notes from everything that broke on the way.
— I build useful things, self-host most of them, and write down what breaks.
log/ — latest
all entries →A small model, a very confident mistake
I ran Laya, a small multilingual decision model, on a CPU-only HomeLab box. It was 0.948 sure that a message saying “No refund needed” wanted a refund.
Unflat it
You can't rotate a screenshot of a 3D model. So my agent now sends me a link to the exact revision, and I turn the actual geometry in the browser.
One night, one agent session, the rest of a weekly quota
- BROKE
- Codex weekly usage went from 22% to 100% between 21 and 25 Sep
- CAUSE
- Every permission prompt from a Claude Code worker woke the Hermes session that supervised it: 211 wake-ups, 808 API calls, ~240K context each
- FIX
- A supervisor script now answers routine prompts without a model and wakes the agent only for plans, questions, results and denials
- PROOF
- Unit tests for the supervisor; the old wait script kept for rollback
dmmeteo.dev, rebuilt as static field notes
The old page had a video background, analytics and a Twitter widget. Now it is static Hugo with no trackers and no runtime requests.
The deploy was green. The site was stale.
- BROKE
- GitHub Pages workflow green, live site unchanged
- CAUSE
- Hugo exits 0 even when it renders no index.html; the job never asserted its output
- FIX
- A smoke check in CI: index.html, canonical, sitemap, RSS and CNAME must exist
- PROOF
- The same check runs on every pull request, before merge
Hermes Drop is public
A self-hosted broker and Hermes plugin for handing an agent a secret without putting it in the chat. MIT, not security-audited.