No error is a dead end: crash, click, and your agent is already looking

Claude-Session: https://claude.ai/code/session_01Ms2FbjQy31TVf3CEvQhGM8

docs/UPSTREAM-INSPIRATION.md

@@ -92,6 +92,21 @@ surfacing agent usage in the bar, and a plugin registry with a lockfile. The
 first shipped, as a bar widget with a switch of its own on Shell › Bar; the
 second remains deferred, see below.
 
+**Omarchy** ([repo](https://github.com/basecamp/omarchy), MIT — DHH): the
+crash-to-agent escalation ladder and the multi-agent usage collectors, both
+taken as code, not just ideas, with attribution headers on every derived file.
+Specifically: the command-as-notification-data hint (the shell runs the click,
+so it survives shell restarts and never blocks the sender), the per-agent argv
+launch table with the skill-path-as-fallback trick for harnesses without a
+skill mechanism, the coredump PID+signal payload, and the Claude collector's
+hard-won honesty — model-scoped limits, cached numbers that expire on window
+rollover rather than age, waiting-vs-expired auth as distinct states, and
+retry-on-transport-but-not-HTTP-error. The diagnose-crash skill's structure
+("rule out OOM first", "never invent function names", "diagnosis reads; it
+does not fix") was adapted rather than rewritten, because it was already
+right. Their usage panel and menu system were not taken — Panama renders its
+own surfaces in its own design language.
+
 ### Declined, and why
 
 - **Wallpaper-derived dynamic color** (HyDE wallbash, matugen). The headline