The logging loop's first real save — a launch crash diagnosed from logs, not guesswork

The setup

Earlier today we changed how we work: every improvement gets built into the installed /Applications/DalkkakAI.app (the founder launches it himself), the app writes a daily runtime.log under ~/Library/Logs/DalkkakAI/, and I read those logs to debug instead of relaunching a flaky dev server. The whole point: the founder shouldn't have to dig — he reports, the logs explain.

The very next build crashed on launch, so the loop got tested immediately.

One sentence of input

The founder's entire bug report was: "DalkkakAI quit unexpectedly 이러는데." That's it. No stack trace, no repro steps, nothing he had to investigate.

What the logs said (no guessing)

Two files answered it:

1. runtime.log — it logged tracing initialized and DalkkakAI starting, but never tauri app setup complete. So it died inside setup. That alone narrowed it to the one thing I'd just added there: a hook-event file watcher.
2. The macOS crash report (~/Library/Logs/DiagnosticReports/appsdesktop-*.ips) — Abort trap: 6 (SIGABRT), main thread, std::process::abort in the stack. A Rust panic that aborted.

Cross-referencing: my new hooks::spawn_watcher called tokio::spawn(...) inside the Tauri setup closure — which has no entered tokio runtime, so tokio::spawn panics. The working capture::spawn_worker used tauri::async_runtime::spawn. I'd just not matched the pattern.

Fix: one line — tokio::spawn → tauri::async_runtime::spawn. Smoke-tested the binary myself (alive past setup, panes spawning) before handing it back.

Why this matters

The founder did nothing beyond noticing the crash. No log-hunting, no flags, no "can you reproduce it." The diagnosis came entirely from logs the product writes about itself — exactly the byproduct-recording the connective layer is meant to automate. This is the small, concrete version of the whole thesis: a solo founder shouldn't have to hold the details in their head; the system should.

It also validated splitting two questions that always get conflated in a crash: how far did it get? (answered by the presence/absence of milestone log lines) and why did it die? (answered by the OS crash report). Together they turn "quit unexpectedly" into a one-line fix.

The decision that introduced this code (hooks for per-session status) is recorded in docs/DECISIONS.md ADR-001; the crash itself in docs/troubleshooting.md.

Commit: c5bc1b1