AI foundations — the plumbing for what's next
- Three new AI endpoints under the hood: estimate-a-task (Haiku 4.5, with a similarity shortcut so it skips the model when we already have enough data), plan-my-day (Sonnet 4.6, calendar- and history-aware), and smart capture (paste an email or meeting note, get structured tasks back routed to the right project).
- Telemetry flywheel: every AI suggestion and your response to it is logged, so the prompts and models actually get better over time instead of just costing tokens. This is what unlocks Phase 2 (reschedule-on-change, daily/weekly review, AI-suggested automations) — the data is now flowing.
- Upstream provider errors (rate limits, content policy, billing) now surface as structured 502s with the underlying reason, instead of opaque 500s. Easier to debug when models misbehave.
- Internal QA harness for testing new AI flows before they hit production surfaces. User-facing AI surfaces ship over the next few releases.