Highlights
What this release added
- **RL-driven inference caching**: Redis adapters now support an epsilon-greedy cache TTL policy with telemetry surfaced on the admin ML dashboard, improving hit rate and latency for personalization and pattern endpoints.
- **Emotion-aware agentic mentoring**: Voice tone snapshots flow into the agentic orchestrator, injecting supportive steps and tone-aware summaries when frustration or low energy is detected, with Labs sandbox UI exposing the snapshot.
- **Pokee Deep Research resilience**: Credit exhaustion now returns a friendly fallback plan, optional retry guidance, and never triggers toast errors; attach hooks and API routes surface retry timing to Ops.
Fixes
What this release corrected
- `/api/research/pokee` responds with structured `credits_exhausted` payloads instead of 500s, and agentic polling translates 403s into grounded heuristics while logging retry hints.
- Added storage bridge so the latest high-confidence voice analysis is captured for agentic session requests, preventing stale emotion inputs.
- Admin ML metrics include RL policy summaries, ensuring ops can verify convergence and investigate slow endpoints quickly.
Known issues
What remained constrained
- RL cache policy state is in-memory today; cold restarts require a brief relearning period until telemetry warm-up completes.
- Pokee credit exhaustion still depends on manual credit replenishment; automated credit ledger syncing remains on the backlog.