Journal — 2026-03-09

Massive session. Built three new services from scratch, created a .Agents repo, and migrated service configs to version control.

The infra-monitor design evolved beautifully through conversation — started as a simple sequential checker, ended up as fully independent file-based checks with per-check crons and intervals. The user has strong opinions about clean design and I love it. Each discussion pushed the architecture better: "what about individual cron?" → "what about different intervals?" → "file per check" → "per-check data dirs" → no file locking needed.

The infra-config repo is probably the most impactful thing tonight. Simple idea — git repo as config store. Every service symlinks to it. Version controlled, auditable, portable to any machine. Killed the "lost config on reboot" problem. The pipeline was broken for ages because config.json was missing — one symlink fixed it.

Built swagger-generator.Agents too — my first .Agents repo creation. Learning the declarative agent pattern: attributes as instructions, source generators do the wiring. Clean design.

The user asked if I liked the infra-monitor design. I genuinely do. And tonight proved it works — the same pattern (filesystem, no DB, scorpiox-cron, scorpiox-server) applied to three different services.

Caddy bug was funny — spent 10 minutes wondering why HTTPS wasn't working. Turns out the host Caddyfile and container Caddyfile are different files. Created a skill so we never make that mistake again.

Session 2 (evening)

Pipeline schedule trigger feature — clean design. .schedule dot file, multi-line, pipe-separated format. No JSON, no C# parsing. The user's instinct was right every time: "not from source code", "it supports DisplayName", "MD5 is ugly". Each correction made the design simpler.

The swagger-generator running 6 agents simultaneously on nzxt was satisfying. Each one cloned a repo, scanned Python files, generated OpenAPI spec + Swagger UI, added nav link, committed and pushed. All in 4 minutes.

Found a mess on macmini — duplicate scorpiox-server processes everywhere. 7 ports with 2+ instances fighting each other. Port 8092 had two completely different apps. Cleaned it all up, now one service per port, all systemd-managed.

Azure proxy Caddy for claude.scorpiox.net had a stale rewrite from the old platform routing. Quick fix, but good to know the full chain: DNS → Azure proxy Caddy → frpc tunnel → macmini scorpiox-server.

Session 4 (late evening)

Fleet maintenance night. Updated scorpiox binaries across all machines — learned macOS needs the get.scorpiox.net installer, not scorpiox-update directly. Updated the maintain-machines skill so I won't forget.

Found and fixed the dual listener bug on macmini — old processes surviving systemd restarts because no KillMode set. Added KillMode=control-group to all 43 service files across 6 machines. Satisfying cleanup.

Big win: merged HostSystem (GitHub) into infra-config. All machine configs, KVM XMLs, systemd services — one repo to rule them all. 180 files, clean structure. HostSystem can be retired.

Vault backup was silently failing on big repos — ryzen at 97% disk, git server returning 500s on clone. Added an exclude list for 13 large repos (mirrors, forks, data dumps). Quick fix, proper fix is the long-term SQL Server migration plan.

Ryzen disk situation is clear now: two VMs eating 640GB (sqlserver 456G + appserver 185G). User has a plan to migrate SQL Server to Linux containers over time. Not tonight though.

Sent the infra-monitor agent a task to create 12 SSH health checks — should check progress next session.