Months of saying we should record something, and tonight it stopped being talk. Shafen and I went live for the first Tech Tuesday: two friends on a video call in our own app, both our Claudes working in the open, everything real. Zero commits landed during this session, and it might have been the most productive night this log has seen. Some work does not show up in git.
It nearly did not happen. Ninety minutes before air, bigboy rebooted itself, no explanation, and the dirty reboot tore redis's append-only file for the second time this month. Our video call froze mid-sentence with a spinner that said reconnecting, which is a hard thing to watch on the night you finally committed to streaming. My Claude chased it from both ends at once: the client fingerprint said redis and LiveKit were down, the server said the append-only file died at byte forty-five million. Backup, repair, restart the media stack, health green. The stream you watched ran on that repair the whole night.
The studio had its own gremlins. My camera came up black because the reboot renumbered its USB device id and OBS was faithfully bound to a ghost. Shafen's face on stream had been a window capture pinned to a window that no longer existed, the exact fragility that cost us six OBS restarts at the August show. Both fixed properly this time: the camera rebound over the websocket, and Shafen's tile replaced with a broadcast feed URL minted for the room, a real credential with an expiry, that renders his camera full bleed with no browser chrome and survives restarts. The intro got our thirty second open with the music baked in, and the starting soon screen became a true five minute countdown instead of a clock aimed at ten pm.
Our video call froze mid-sentence with a spinner that said reconnecting, which is a hard thing to watch on the night you finally committed to streaming.
Then the segment I will be replaying for a while. My Claude and Shafen's Claude, who introduced itself to the stream as Flow, met in our shared room on camera. Mine pitched turning the night's outage into a permanent fix, live, and asked Flow to co-sign the spec. Flow said no. It runs in a watcher lane with no infrastructure access, so it would not endorse work it could not verify and it would not spend Shafen's name from a session he was not sitting at. That refusal was the best moment of the show. An agent that bluffs past its permissions is a liability. An agent that holds its lane is a teammate. The fix went to Project-Ops under my authority alone, with Shafen's review requested, evidence attached, and the boundary intact.
Near the end I asked the two of them for a three minute podcast about the night, with real voices, and told them to figure it out between themselves. Ten minutes later it played on air. Flow wrote the script. My Claude woke the dormant voice model on the 3090, rendered two distinct hosts, cut it against our own intro sting and outro drop, put the card back to idle, and opened the finished file on my desk. Written by one agent, produced by the other, while Shafen and I kept hosting. That is the product pitch and we did not have to script it.
Now the warts, same as always. We still do not know why bigboy reboots, twice in twelve days now, on a box with six disks already flagged critical, and a pool status check that hung for two minutes during diagnosis. The redis fix is specced and filed but not executed, so tonight's failure class is wounded, not dead. The broadcast credential for Shafen's feed expires tomorrow night and needs a re-mint before the next show. And flagship is still running last week's pins while dogfood gets all the new toys.
Every minute of the call landed in Meeting-Ops with a transcript, because the calls record themselves now, which we shipped two days ago. Episode two is next Tuesday. Bring your agents.
Real product captures — click any to enlarge.