Step 5: Chrome-button move + per-session button retirement

[heavygee]

Goal

The atomic "voice moves up a tier" switch: relocate the realtime conversational surface to application chrome (attached to the Overseer), and in the same move replace the per-session conversational voice button with lightweight dictation (composer mic) + on-demand read-back. One coherent moment — never an interim state with two ways to talk in a session.

Spec

• docs/plans/2026-06-03-overseer-build-sequence.md Step 5 (primary)
• docs/plans/2026-06-03-overseer-framing.md "The chrome-button insight", "Why voice belongs at the Overseer tier, not the worker tier"
• docs/plans/2026-06-03-overseer-prioritization.md §9 (attention budget)
• Upstream alignment: tiann in discussion Voice roadmap: picker -> provider abstraction -> local stack -> fleet command centre tiann/hapi#691. Builds on feat(voice): pluggable voice backend with Gemini Live & Qwen Realtime tiann/hapi#692 + the STT/TTS plumbing in Voice provider abstraction: expose standalone STT + TTS endpoints #29.

The atomic switch (both halves fire together)

• Chrome / Overseer half: the conversational, realtime, fleet-aware voice surface relocates to chrome, attached to the persistent Overseer session (Step 3: Read-only Overseer wired to voice #25); survives page nav + device hop (Step 1: Voice persistence + receiving-session indicator #21 persistence; the voiceFocus kind re-targets session → overseer, not a rewrite).
• Per-session half: the per-session conversational button is removed and replaced by (a) composer-mic dictation (STT → composer, manual send, transcription reviewable before send — the ChatGPT dictate pattern) and (b) on-demand "summarize & read back" TTS of the last turn / important result. Speech transport only; no realtime turn-loop in a worker session.

Why atomic: keeping the heavy per-session conversation alive until this moment preserves the habit-teacher (#21) and avoids any conversation-less gap on the daily driver; doing the per-session swap at the same instant avoids an interim "two ways to talk" UX. Conceptual basis: cross-session conversation is meaningful because work sessions are things-to-do with completion states (an inbox to arbitrate), unlike independent chat topics — so conversation belongs at the fleet/Overseer tier, and the session keeps only speech transport.

Acceptance

• Chrome conversational voice surface attached to the Overseer session (Step 3: Read-only Overseer wired to voice #25); primary on web, mobile-web, webXR (on the feat(voice): pluggable voice backend with Gemini Live & Qwen Realtime tiann/hapi#692 abstraction).
• Conversational surface survives page nav + device hop (Step 1: Voice persistence + receiving-session indicator #21).
• Per-session conversational button removed.
• Per-session composer-mic dictation: speak → transcription fills the composer → manual send (no auto-send) → transcription reviewable before send.
• Per-session on-demand "summarize & read back" TTS.
• Both per-session affordances use the operator's chosen provider via the Voice provider abstraction: expose standalone STT + TTS endpoints #29 STT/TTS endpoints; no cross-provider mixing.
• PTT remains for fast dictation.
• Mobile-web smoke tests pass — the "shower-speaker via mobile-web + bluetooth" use case (Vision: hands-free shower-speaker agent direction over Bluetooth call #16) becomes live from here.

Out of scope

• Provider STT/TTS endpoint plumbing — Voice provider abstraction: expose standalone STT + TTS endpoints #29 (prerequisite, ships earlier).
• Standing-order autonomy, channels, native mobile app, salience learning, webXR rendering polish — Step 6+.

Dependencies

• Blocked by: Step 4: Disagreement-capable Overseer + voice dispatch with confirm #26 (MVP acceptance bar), Step 3: Read-only Overseer wired to voice #25 (the Overseer conversational entity the chrome button attaches to), Voice provider abstraction: expose standalone STT + TTS endpoints #29 (STT/TTS endpoints for the per-session affordances)
• Relates to: Step 1: Voice persistence + receiving-session indicator #21 (voice persistence / habit-teacher)
• Part of: Fleet Overseer architecture — umbrella #19

Suggested PR breakdown

1-2 PRs: chrome Overseer voice button; per-session swap (remove conversational button + add composer-mic dictation + read-back); mobile-web smoke tests.

Risks

• The always-on inbox failure mode (persistent low-amplitude anxiety): aggressive bounded-deferral defaults (§9), no sounds for routine items, clear "Overseer wants you" vs "Overseer is processing" distinction. Failure mode to avoid: "Clippy got into cocaine".
• Auto-send hazard for dictation — a per-session message is an instruction a worker acts on; manual-send-after-readback is required, not optional.