ADR 0031 — Voice as a channel-layer capability (v53-F6)
Date: 2026-07-17 · Status: accepted
Context
Skep had zero voice; the operator wants to TALK to it. v51 cut TTS as a "channel-layer concern" — this un-rejects that cut with the scoped rationale the cut itself implied: voice lives in the channel layer (web UI + outbound delivery), never on the Queen's tool surface. The model never sees or produces audio: STT feeds text in before the model, TTS renders text out after the turn.
Decision
-
Web-first. The browser's speech APIs cover "talk to skep" with zero backend: a mic button (recognition → composer text) and a spoken-replies toggle (synthesis over the reply). Honesty over mechanics: Chrome's recognition is CLOUD-BACKED (audio goes to Google's speech service) and the mic tooltip says exactly that — "the browser does all the work" would be a lie of omission.
-
Server-side TTS is config-gated, default
none, providers labeled by their egress.piperis the LOCAL story.edgeis Microsoft's cloud service — "free, no API key" does NOT mean no network: every rendered reply's text egresses when chosen (the draft plan claimed edge was local; the v53 review corrected it).openailikewise. Theset_tts_providercard and the setting descriptions carry these labels; the tool result records the egress truth in the transcript. -
Config-gated, not operator-policy-gated — recorded boundary. TTS/STT egress is channel infrastructure, the same trust class as the Discord API calls that deliver the text itself: enabled by explicit operator configuration, not per-call policy decisions. If voice ever grows per-content decisions (e.g. "never speak approval details"), that becomes an operator-policy scope.
-
Delivery bounds. Voice messages ride Discord (one multipart send on the existing messages endpoint), best-effort after the text lands — a render or upload failure is logged and swallowed (the outbound posture). Telegram/Slack delivery is demand-driven. Messenger voice-message STT is DEFERRED with a named trigger: it needs inbound attachment machinery per channel and no field test has ever sent one.
-
No voice approval path. Approvals remain text/card. A spoken "yes" is not a confirmation surface.
Consequences
uv pip install 'skep[voice]'pulls the optional providers; without them every path degrades to a clean logged skip.- Reply text spoken aloud is capped (1,500 chars) — a message, not a podcast.