SPIKE: Upgrade to Latest Voice Live SDK + Evaluate Multi-Agent Orchestration #13
Description
Summary
Upgrade the codebase to the latest Voice Live SDK and run a focused spike to validate:
• Multi-agent orchestration patterns (handoff/swap during a live call).
• Connection/session model for routing between “live” agents.
• Runtime-mutable attributes (agent, thread/session, model, tools, params) without tearing down the call.
Goals
1. Bump to the newest Voice Live SDK and resolve breaking changes.
2. Prototype in-call agent switching and agent handoff patterns.
3. Determine which attributes are safe to mutate at runtime (agent, model, thread/context, tool config, temperature, VAD, TTS voice, etc.) and what requires a new connection.
4. Produce latency/quality and stability metrics; document limits and recommended patterns.
OUT-OF-SCOPE
• Full production rollout.
• Broad UX changes beyond the minimal controls required for the spike.
Acceptance Criteria
• Upgrade builds/CI green; no lint/type errors; e2e sanity passes.
• Agent swap in the same connection keeps audio alive; no more than +150 ms added gap on swap.
• Handoff to another agent (new connection) completes in < 1.5 s with preserved context (intent + short transcript).
• Runtime changes verified:
• agent switch: works w/o reconnect (or documented otherwise).
• model swap: either hot-swappable or documented as requires reconnect.
• thread/session switch: supported pattern for context rebind or fallback approach is documented.
• tools/params (temperature, stop, VAD, voice): updated mid-call reflect in subsequent turns.