Automation

Every cost-reduction experiment I wanted to run — prompt optimization, cheaper model routing, stripped-down prompt variants — was uncalibrated because nothing measured per-run token spend. The fix was a measurement layer, and the load-bearing decision turned out to be the schema, not the capture point.

I wanted the Pipeline tab to chew through a chain of planned issues instead of making me click them one by one. The tempting answer was parallel sessions, but the March 23 state lessons made serial mode the right thing to ship first.

I asked an AI deliberation whether to build a Discord bot to replace my 3-minute poll cycle. The deliberation crashed three different ways — and that failure was more useful than the answer would have been.

A deliberation session silently failed because the health check was hitting the wrong endpoint. Fixing that one bug turned into a five-part architecture overhaul covering provider dispatch, transcript persistence, and orchestrator handoff.