feat(desktop): realtime-as-hub voice path (Phase 1 BYOK + Phase 2 ephemeral)

[vendz]

Realtime-as-hub voice path for the desktop floating bar — one realtime model does in-session STT + reasoning + routing (via tool choice) + spoken reply, replacing the STT → Haiku-router → Claude → TTS cascade on the push-to-talk path.

⚠️ Stacked on #7957. This branch sits on top of the prompt-cache / QueryTracer work in #7957 — the realtime code instruments QueryTracer and calls the prompt-cached /v2/chat/completions, so those commits are a prerequisite and can't be dropped without breaking the build. GitHub can only base this on main, so until #7957 merges the bottom ~12 commits of this diff belong to #7957 — review only the realtime commits on top; they drop out of the diff automatically once #7957 lands. Merge #7957 first.

Reliability + voice-UX pass (latest commits)

Hardening after live testing surfaced drops, cutoffs, and routing issues on the Gemini path:

• Barge-in is now in-session — interrupt the in-flight reply with a fresh activityStart instead of tearing down + reconnecting the socket (the reconnect lost conversation context and dropped the next turn). A per-turn reply gate stops an interrupted turn's trailing audio / bookkeeping turnComplete from leaking into the next.
• WS keepalive ping so the Gemini socket doesn't idle-close (~2.5 min) and silently fall back to STT; managed-user auto-re-warm fixed (was gated on an isActive that's false once the session is nil, so it never recovered).
• Energy-first speech gate — loud/clear speech always passes (Silero was intermittently dropping real speech); dropped turns now log peak/RMS/device for real diagnosis instead of guesswork.
• Screenshot is sent as a realtimeInput.video frame — the Live model rejects mid-session clientContent with close 1007, which was killing the socket.
• Context-window compression enabled so long sessions don't degrade/stall.
• get_tasks local read tool (speaks your real tasks, no background agent); routing rewritten — answer creative/general/long-form yourself, escalate to ask_higher_model only on user pushback or precise-fact needs, and spawn_agent must actually emit the call for actions (it was narrating instead of acting).
• Voice follow-up to agents — a mic button on a finished pill captures via the hub STT and continues that agent's session (continueAgent); finished agents keep their session alive for follow-ups (capped at 8, oldest trimmed).

Verified (live)

• BYOK + managed/ephemeral connect and run full turns on OpenAI and Gemini.
• get_tasks speaks the user's real tasks; voice follow-up routes the transcript into the agent's session; mint route 401 unauthed / 200 + token authed.
• Clean release build (arm64).

⚠️ Known limitation

Gemini seems to bug out for no apparent reason. sometimes it works flawlessly but sometimes it just doesn't. Currently trying to pinpoint the issue

🤖 Generated with Claude Code

[ThomsenDrake]

ClawSweeper local pilot review

Recommendation: keep open, but this broad draft is not merge-ready.

What it found:

• This PR is explicitly stacked on perf(desktop): floating-bar latency — Anthropic prompt caching, router/screenshot fast-paths, and a query tracer #7957 and should wait for that prerequisite to land or be closed first.
• The realtime voice-hub direction has an auth/cost-boundary blocker around minted realtime tokens.
• The branch needs a narrower reviewable diff, redacted end-to-end proof, and docs for the new desktop voice path.

Suggested next step: land or close #7957 first, then rebase this branch; gate minted realtime tokens to managed paid users only, preserve BYOK direct mode, and add proof/docs before merge.

Posted from a local report-only ClawSweeper pilot by request; no labels, closes, repairs, or merges were performed.