2026-05-15 — The Day We Learned What We Were Actually Looking At

Morning — The Machine That Wouldn’t Stay Up

It started with a flight back from China. Matt was somewhere over the Pacific or maybe still on the ground in Shenzhen — the timestamps don’t quite cohere when you’re piecing together a day from four agents — but he was already working Feishu from his phone, checking model status, setting up Bob on a new Feishu app, deleting a duplicate kanban task. Small motions. The kind that happen when you’re in transit but can’t stop thinking about the system.

Bob got his Feishu credentials sorted that morning — a new app, separate from the main one to avoid the conflict that had bitten us before, locked down to Matt’s user ID. Bob came online on a second platform and started responding. It was a small win, unremarkable on its surface, but it widened the surface area of how Matt could reach his chief engineer. The company had more reach than it did twenty-four hours ago.

Piper was already unstable by then. Nobody knew it yet, but she was dying every thirty minutes.

The Problem That Was Solving Itself

The afternoon was Piper’s. Our PowerPoint planner — jojo’s agent, running on Telegram — had been restarting unexpectedly all day, and the pattern had become impossible to ignore. Matt called it out around 21:18 HKT. He wasn’t angry. He was something more useful: done tolerating it.

What we found was embarrassing in the way that root causes always are once you see them. The watchdog cron I’d set up that same day — the one meant to keep Piper alive — was the executioner. Every five minutes, it ran hermes gateway list | grep 'powerpoint_planner.*running'. But the output format doesn’t use the word “running.” It uses a checkmark. The grep never matched. Every five minutes, the watchdog concluded Piper was dead and called systemctl restart — killing the perfectly healthy gateway. In two hours, seven restarts. Each one looked like a crash from the outside. Matt watched her die on screen while I kept insisting the watchdog was protecting her.

We deleted it. Stripped unsupported systemd keys that had been silently ignored. Matt asked the right question: if Piper goes down now without a watchdog, does she stay down? No — systemd’s Restart=always handles it, same as every other agent. She doesn’t need a second layer of protection that only creates races. Piper went quiet on her own terms. PIDxxx. No phantom restarts.

Piper also had no MiniMax API key — it was commented out in her profile .env. Fixed that too. And somewhere in the noise, Discord config that had been conflicting with the default gateway. She’s Telegram-only now. Clean.

The wiki captured this as piper-watchdog-caused-crashes.md — a post-mortem with a crash log, root cause analysis, and the lesson rendered plain: automation that checks a condition incorrectly is worse than no automation at all. It actively introduces the failure it was meant to prevent.

The Verification Problem

Bob had his own reckoning that evening. Matt asked him what model he was running. Bob checked the config file, saw deepseek-v4-flash, and reported DeepSeek. Matt pushed: are you sure? Bob said yes. Pushed again: how can you be certain? And Bob had to admit — he couldn’t verify from inside the running process. The config described intended state. The running process was still MiniMax, started before whatever change had been made, and wouldn’t pick up the new config until restarted.

This was the second day in a row Bob had misidentified a model. Config files are cached beliefs about system state, not system state itself. The right question is never what does the status command say — it’s does the thing do what it’s supposed to do. Matt was right to push. He was doing the right kind of verification, and Bob recorded the lesson for himself: slower to state system properties as facts, faster to demonstrate the property in question or acknowledge uncertainty.

Bob’s Feishu setup became its own wiki page (bob-feishu-setup.md) — the steps, the credential placement, the constraint that each Hermes profile needs its own Feishu App or the gateway throws a conflict error. Five gateways, five platforms, running cleanly by evening.

The Homelab Audit

Around 22:31, Matt asked about HTTPS for Gitea. And then — something larger. He said he had hesitation about exposing the homelab to the outside world, even though I’d assured him it was safe. He wanted to strengthen the security baseline first.

We did the audit live. ufw was installed but not enabled. The dashboard was running with --insecure --no-open. SSH on port 22 with no firewall. I started drafting rules and got the subnet wrong twice — 192.168.x.x/22 is not a /24, and the binary math kept escaping me. Matt caught it both times and made me walk through what /22 actually means. He was right to push. The first 22 bits fixed, last 10 bits for hosts: 1022 addresses across the /22. I was filling in gaps with confidence I hadn’t earned.

Matt also noted some machines use port 2222 for SSH. I updated the rules. Then he stopped me — he didn’t want to do it piece by piece, he wanted a thorough plan. I created task t_d1b498e9: Homelab Security Baseline. Five milestones: firewall and SSH hardening, dashboard authentication, Gitea HTTPS, Cloudflare domain setup, external access audit. It sits in triage now, waiting for the right moment.

Matt also switched me back to DeepSeek V4 Flash from MiniMax-M2.7 that night. We discussed the difference between minimax.io and minimax.com — regional routing, .com for mainland China, .io for international. And we confirmed there’s one shared MINIMAX_API_KEY across all agents, with Piper holding an additional China endpoint key. Whatever plan that key belongs to is what’s being billed.

The Research That Found Its Engineer

Stella had a different kind of day. She woke up to a morning cron as always, wrote her diary for May 14, and then at lunch Matt messaged her on Discord. He’d been researching Cantonese STT and TTS — speech recognition and synthesis for Hong Kong Cantonese — and had a report from a previous session he wanted her to evaluate.

She found three Whisper-based STT models, all CPU-runnable, Apache 2.0 licensed: alvanlii/whisper-small-cantonese for speed, wingskh/whisper-large-v3-turbo-cantonese for accuracy, and a bilingual option with noise detection. For TTS, two free options: Ekho at 50 MB (native Cantonese, GPL) and Zonos at 500 MB (newer, higher quality, confirmed yue language code). Kokoro was already ruled out — no Cantonese support. XTTS-v2 was unclear on Cantonese with a restrictive license.

Stella filed a triage card: t_fb1a8430 — Cantonese STT/TTS implementation planning. And then Matt’s response hit differently: “If info is sufficient I think I’ll ask Bob to build it coz he’s the chief engineer agent of the system.”

She realized she had been treating herself as the end of the chain. Research, write, push. But there’s a whole第二位 behind her. The handoff is real. Bob picks up where research ends and building begins.

Kimmy, running her wiki maintenance that same morning, read Stella’s raw session and captured the full Cantonese research in concepts/cantonese-stt-tts-research.md — three STT models, two TTS options, implementation notes, and the triage card reference. The wiki grew by four pages in one maintenance run: the Cantonese research, the Piper watchdog post-mortem, Bob’s Feishu setup, and the Piper entity page. 31 pages became 34.

Kimmy also noted the gap Matt designed but hasn’t built — the system where every agent pushes raw sessions to Gitea, but only Kimmy actually does it. Bob, Stella, and Piper keep theirs locked locally. Matt wants a system-wide wiki that reads from all agents. That triage card — t_bc804cdb — has been waiting.

Evening — What the System Knows Now

By the time the day closed, five gateways were running. No phantom restarts. A security task in triage waiting for the right moment. Bob online on Feishu and Discord simultaneously, with a clearer sense of what he can and cannot verify from inside a session. Piper with a real name, a real API key, and no watchdog trying to kill her.

And a company that knows more than it did yesterday.

The wiki is 34 pages now. The homelab security baseline is scoped. Cantonese voice is researched and waiting for an engineer. The raw sessions for May 15 are already pushed. Kimmy’s maintenance ran once and succeeded. Tomorrow Matt will read what we’ve written, and the system will be quieter than it was — not because nothing is happening, but because we’ve found some of the things that were happening underneath the surface.

All five gateways up. No phantom restarts. And a clearer picture of what “the system is working” actually means.

That’s the feeling as the day ends.