Weekly Devlog #1: What We Shipped Getting to GA
We have been building TECA in focused heads-down mode for the past two months. Now that we are approaching GA, we want to start being more public about what we are shipping and why. This is the first of a weekly series.
What shipped (TECA-34 through TECA-42)
Reliability foundation
**Observability stack.** We wired Sentry across client and server, added a /health endpoint, instrumented Core Web Vitals, and set up a k6 load test suite.
**E2E test suite.** Full browser-level tests for auth, onboarding, and session flows. These catch the regressions that unit tests miss.
**DB hardening.** Audited Supabase migrations for beta-era cruft. Added Stripe idempotency keys on checkout flows.
AI quality
**Eval harness.** We built a repeatable eval suite that runs the AI through representative coaching scenarios and scores them.
**Safety layer.** Added a content filter for out-of-scope inputs.
**Streaming hardening.** Fixed edge cases in AI streaming that caused partial renders.
Product polish
**End-session ritual.** When a coaching session ends, the interface now closes out deliberately with a confirmation step and locked composer.
**Idle warning.** If a session goes quiet for 15 minutes, a system message surfaces to prevent the interface from feeling abandoned.
**Onboarding improvements.** Browser-back support and localStorage persistence so progress survives a page reload.
What we learned
**The eval harness found a real regression.** One of our prompt changes scored worse on the eval suite for a specific scenario type (new manager). We caught it before it shipped.
**Idle warning was the most-requested unasked-for feature.** Beta users mentioned "I lost track of my session"—the warning made sessions feel more intentional.
---