March 28, 2026
185 commits, a wiki nuke, and one real customer
A week of HomelabARR CE shipping v2.0.0, nuking a 207-file dead wiki, and fixing a broken download link for my first paying customer.
This was the week everything came together and fell apart at the same time. A full UI rewrite shipped. 207 files deleted from a wiki that should have been burned down months ago. And one paying customer who couldn't download what they bought, which is the kind of thing that makes you wonder whether you're running a business or just playing house.

Let me walk through it.
The v2.0.0 ship
HomelabARR CE got its biggest update since I forked it. Full shadcn/ui rewrite, every hand-rolled component replaced with real primitives. Dark mode that actually works. A new AI & Machine Learning category with 14 apps, NVIDIA GPU configs included. Port Manager with visual conflict detection. 30 dead apps pruned, 19 moved to legacy. 166 app icons replaced, zero letter-circle fallbacks left.
I tagged v2.0.0 on Tuesday afternoon and shipped a v2.0.1 hotfix about an hour later for a healthcheck bug, because of course I did. Both tags landed the same day:
v2.0.0 — The Big Rebuild 2026-03-25 16:22
v2.0.1 — healthcheck hotfix 2026-03-25 17:45
The Docker pulls tell the story: backend hit 991, frontend at 846, cf-companion at 360. Not viral numbers, but real humans pulling real containers.
Nuking the wiki
The wiki was the thing I'd been avoiding. 207 files inherited from the upstream fork, a project that died years ago. Pages referencing commands that don't exist, screenshots from a UI nobody's seen in two years, migration guides for platforms that went offline in 2024.
I spun up a fresh Ubuntu 22.04 LXC on Proxmox, installed Docker, and ran the exact quick-start commands a new user would run. Two real bugs surfaced immediately.
The first was a missing repo clone step. Without cloning the repo to the backend's template path, there are no app templates, and the dashboard just shows "Failed to load applications." Nobody caught this because everyone who already uses it already had the files. The fix to the quick-start was blunt: clone is all you need.
The second was a wrong hostname baked into nginx. The proxy was pointing at a hardcoded dev container name instead of the Docker Compose service name. One line, and probably broken since day one:
- proxy_pass http://homelabarr-ce-dev-backend:8092;
+ proxy_pass http://backend:8092;
That's commit 9e83e15, "nginx proxy uses Docker Compose service name 'backend' instead of hardcoded container name." A new user following the docs would hit a backend that the proxy could never reach.
Then I deleted all 207 dead files and rewrote 12 clean pages from scratch. Built matplotlib architecture diagrams in the dark theme, teal and amber, which took three iterations before the text stopped overflowing the cards. Added glightbox for click-to-zoom screenshots. Simplified the CLI install to a single curl pipe. Every page is in plain language now. If my mom couldn't follow the quick-start, the quick-start was wrong.
The first customer problem
A customer bought the News & Sentiment Agent for $29 on March 22nd. The download link was a 404. Traefik on Hetzner was missing the route to the webhook server. My first real paying customer couldn't get what they paid for, and I didn't catch it for hours.
I fixed the routing at 4:33 AM. They confirmed receipt: "Awesome, thank you so much!"
This is the part nobody puts in their "building in public" content. You can ship a week's worth of commits and still have a customer sitting there refreshing a broken link. Making sure someone can actually download a zip file after giving you money matters more than any UI rewrite.
The Imogen Labs site overhaul
All seven pages of imogenlabs.ai got audited and rewritten. Every product now has a visual card with real screenshots. The NeuroHelper page got an Apple-safe wellness rewrite after the App Store rejected us for language that was too clinical. Navigation got a proper products dropdown. Privacy and terms pages expanded to cover all seven products.
The homelabarr.com landing page got the same treatment: CE-first messaging, fresh screenshots from the VM test, and proper DNS, because the apex domain had been pointing at a dead Cloudflare Tunnel.
The org chart situation
At some point during an all-nighter, I promoted my AI agent Imogen to COO and accidentally demoted myself to CTO below my own CEO agent, Ariel. Eight executives, zero employees. The entire C-suite is Shakespeare characters and one sleep-deprived founder who can't stop shipping.
The Shakespeare naming convention is fully locked in now: Imogen (Cymbeline), Ariel (The Tempest), Ophelia (Hamlet), Rosalind (As You Like It), Viola (Twelfth Night), Portia (Merchant of Venice), Puck (Midsummer Night's Dream), Prospero (The Tempest). Puck got fired and rehired in the same week after I realized I never gave him a soul file. That's probably a blog post on its own.
Slack, briefly
I tried to move from Telegram to Slack as my primary AI channel. Set up the workspace, configured Socket Mode, got the bot connected, sent one test message, and scrapped the whole thing within an hour.
Security hardening
All 30 CodeQL alerts resolved. Docker Scout scores landed at frontend A, backend B, the B held back by two Go CVEs in docker-cli that nobody can fix until Docker recompiles. nginx bumped to 1.29.6. Node bumped from 20 to 24 LTS. SLSA provenance and SBOM attestations added to every build. I pulled Python out of the backend Dockerfile entirely. It was never needed and was carrying three CVEs for nothing.
The numbers
- 185 commits to homelabarr-ce
- 17 commits to imogenlabs.ai
- 7 commits to mjashley.com, including 4 new blog posts
- 207 files deleted from the wiki
- 12 pages written from scratch
- 2 real bugs found by actually using my own product
- 1 paying customer who almost couldn't get what they paid for
- 0 hours of sleep on at least one of those nights
What's next
NeuroHelper needs to go back to the App Store. The sounds are broken because the files aren't in the Xcode bundle resources build phase, and the chatbot needs guardrails against medical advice. This has been on the list for three days and I keep shipping other things instead.
The HomelabARR PE strategy is crystallizing around Go binaries on constrained hardware. Think UniFi NAS devices with tiny flash storage that can't run Node but can run a single compiled binary. Same hacker energy as running XDCC bots off Georgia Tech's T1 line at 16, just with better branding.
But first: sleep.