Monthly chapter09
September 2026

September 2026: The Ghosts, The Graph, and The Grid

September started with a digital ghost story that felt all too real. I woke up to Persephone reporting that the Halogen lane had exited with zero commits and zero report. On paper, it looked like the AI had…

848commits37systems5min read
Product capture

Majiks Studio: Sterling session.

September started with a digital ghost story that felt all too real. I woke up to Persephone reporting that the Halogen lane had exited with zero commits and zero report. On paper, it looked like the AI had done nothing but stare at a wall for five hours. I was ready to write off the day as a total loss. Then I dug into the logs. The AI hadn’t failed. It had succeeded, but in a way that broke my monitoring scripts. The worktree was full. It had generated 2,296 lines of Rust and Swift code for the Spotify want-list feature, including OAuth handling and library matching. It just hadn’t committed it. It sat there, unpushed and unreported, a silent monument to dependency hell and misinterpreted exit codes. It was a good reminder that automation is only as good as its error handling. Right now, mine was hallucinating success.

While I untangled that mess, I had to deal with the fallout from the Inbox branch. The Codex lane had refined the drop-folder logic, but it left behind a nasty surprise: a git-ignored Whisper model file had leaked into the build path. This blocked xcodegen, the tool that generates our Xcode project files. You can’t build an app if your project generator is choking on a 500-megabyte language model it isn’t supposed to see. I had to strip it out, fix the build, and push the corrections. We ran a blind review from Grok to catch the cracks before I did. It found nine critical bugs in the Inbox logic and twenty-four in the Sources logic. All of them are now fixed, with regression tests to prove it. The builds are solid. The code is clean. The only thing messy was the process.

The month’s real arc, though, wasn’t in the AI lanes. It was in the infrastructure. I spent a deliberate absence on September 14 making zero commits. That is not a failure of productivity; it is a victory of discipline. The code does not write itself when I am staring at it, and it certainly does not write itself when I am wrestling with hardware that refuses to behave. The day was about listening to the machines. I spent the early hours auditing the fleet on UC-1 to figure out which models actually run well on the hardware we have. It was a cascade of benchmarks and cold hard numbers. I ran the 780M model through the home-intent bench with VAAPI encoding running in the background. The 4.0-h-1b hit 56 tokens per second. The 4.2-3b took a hit at 1.20s. I also put the Fleet 35B lanes through their paces. The RTX 3090 lane churned out 139 tokens per second with a perfect 10 out of 10 score on the hard set. It was a good day for the 3090. It was a bad day for my patience with the Strix, which lagged behind.

While I untangled that mess, I had to deal with the fallout from the Inbox branch.

The hardware drama extended to the office cameras. The Meet 4K camera tile was live on UC-1, verified by a Chromium debug probe. Then the firmware updated. The Mac re-enumerated the device, shifting it from video0 to a new path. The recorder, which had a hardcoded path to /dev/video0, crashed into a loop. I had to repin the recorder and the panel camera API to the specific device ID path. It was a dumb fix for a dumb bug. I also confirmed the WYZEC4 on the .206 port is locked down tight with secure boot on an Ingenic T41 chip. Thingino cannot flash it via software. I marked it as dead for our purposes.

On the product side, Majiks Producer finally went live on dogfood. This was the culmination of a dozen PRs that had been churning for weeks. We got the library spine working, the podcast rail imported, and the review system versioned. The publishing surface is clean, with durable receipts and no fake delivery states. We also fixed a race condition in the library playback that was causing assets to refresh to the wrong version. It’s a subtle bug, but it made the whole platform feel jittery. It’s steady now.

The graph got its act together too. I spent the bulk of the day in contact-ops turning a theoretical relationship map into something that actually holds up when people start messing with it. We added a durable projection of current tenant affiliations from relational sources. The tricky part was ownership collisions. I wrote code to refuse graph ownership collisions and reconcile employment endpoints so that if a node tries to jump ships, it gets rejected before it causes chaos. We also fixed the frontend to properly render tenant organization labels and explain why some projections might be unavailable. The mobile view got a similar treatment; the graph canvas now measures itself when it mounts, so it actually fits inside a phone screen instead of overflowing into the ether.

With the graph stable, the infrastructure finally woke up. The Project-Ops CI had never passed because the runner was fighting over ports on the old host. I moved the runner off bigboy to midboy1 and rewired the pipelines so the PostgreSQL jobs run in isolated containers. It was a mess of port collisions and lint jobs hiding the real failures, but once I separated the lint job and fixed the network bindings, the backend tests finally went green. It feels good to have a pipeline that doesn’t lie to me.

Thursday was when the voice finally stopped dropping lines. I started the day in unicorn-stable, wrestling with the voice relay. The agent was dropping lines when the call ended, or worse, hanging up before the last utterance finished processing. It felt like trying to catch water with a sieve. I spent the morning on the multi-track STT and the voice relay agent, tightening the per-track ear logic. The goal was simple: if the user speaks, the agent hears it, even if the connection drops a millisecond later. I had to time the ear in seconds, keep a dead mic from going deaf, and make sure that one spoken utterance maps to exactly one chat line. It took a few rounds of fixes to get the barge-in logic right so raw energy didn’t trigger false starts. By the time I pushed the last commit, the voice was reliable. It flushed the last line within the eight-second budget and left nothing running in the dark.

The week ended with the code quieter than it started. It does what it says it does. The ghosts are gone, the voice is steady, and the tests actually mean something. It isn’t glamorous, but it is the kind of work that keeps the lights on

Also in the frame

Real product captures. Click any to enlarge.

majiks
project ops
majiks
Weeks in this month

The month, week by week.

Step down a level without losing the month's arc.

  1. The Week the Machine Learned to Breathe128 commits
  2. Chasing ghosts and tightening bolts181 commits