Voice TranscriptionCAUTION

A framework for building agentic apps

The instruction file, verbatim from the audited commit — this is the text the model reads, and the surface the audit's instruction layer examines. Quoted here so you can judge it without cloning anything.

---
name: voice-transcription
description: >-
  Framework-wide voice dictation in the agent sidebar composer. Use when
  changing composer microphone UX, the transcribe-voice route, or the
  Voice Transcription settings section. Covers transcription-source routing,
  cleanup routing, Google realtime gating, and the voice transcription
  application-state keys.
scope: dev
metadata:
  internal: true
---

# Voice Transcription

The microphone inside the sidebar composer offers two distinct paths:
editable dictation and an opt-in realtime speech-to-speech agent session.
Users configure dictation separately from AI cleanup. The source picker (Mac
Native, Google Realtime, Batch) is the Voice transcription row on Settings ›
Account › Preferences (`VoiceTranscriptionSection compact`, anchor `voice`);
cleanup and batch provider keys stay in the agent sidebar's Voice
Transcription section. Both write the per-user `voice-transcription-prefs`
application state. Both paths are available in every template that renders
the shared agent sidebar.

## UX rules

- **Always show the mic alongside the send button.** Cursor replaces send
  with mic when the composer is empty; their users complain. We keep both
  visible — Lovable does the same.
- **Ask before switching modes.** The mic opens a compact shadcn popover with
  **Real-time voice** as the primary action and **Dictate** as the secondary;
  never silently replace the established dictation behavior. Users can choose
  **Remember my preference** to skip the chooser on later clicks.
- **Dictation is click-to-toggle, not push-to-talk.** More forgiving in a sidebar, avoids
  host-app hotkey clashes. Keyboard shortcut is `Cmd/Ctrl+Shift+M` and
  `Escape` cancels mid-recording.
- **Transcript lands in the composer, editable, never auto-sent.** Insert at
  the caret via `editor.chain().focus().insertContent(text).run()`.
- **No CSS transitions for the recording state.** Framework rule; use static
  brand color (`#625DF5`) instead of pulses.
- **Icon:** Tabler `IconMicrophone` (idle) / `IconPlayerStopFilled` (recording).
  Never use a sparkle or robot icon.
- **Errors via inline alert or toast, never `window.alert`.**

## Realtime speech mode

Realtime speech mode is a full-duplex OpenAI Realtime WebRTC session, separate
from the dictation providers below. It prefers the authenticated user's
complete Builder connection and the Builder-managed metered gateway. A scoped
`OPENAI_API_KEY` is the fallback when Builder is not connected. The server owns
the `gpt-realtime-2.1` session configuration and keeps credentials out of
browser code. If neither provider is configured, show **Use Builder.io** as
the primary microphone-popover action and OpenAI key setup as the secondary
action before requesting microphone permission.

- Starting voice mode collapses the chat into a persistent bottom-end speech
  orb. The orb stays visible above the chat even when the chat opens
  automatically. Clicking it shows or hides chat without ending the session;
  opening chat slides the orb clear of the sidebar. The progressively disclosed
  controls contain only settings and the separate end-session action.
- The orb's compact waveform stays visible for the full session and reflects
  actual microphone or assistant audio activity. At silence it rests at its
  baseline instead of turning into a loading spinner. Once connected, a soft,
  slowly rotating blue glow marks the live session; brighten and speed up the
  glow while the agent is working, and keep it static under reduced motion.
  While connecting, show a distinct compact spinner; when the session becomes
  live, the assistant gives one brief spoken greeting so readiness is
  unambiguous. Do not apply a text-sized `max_output_tokens` cap to that audio
  greeting because audio tokens can truncate it mid-sentence.
- The orb settings cog opens in place without ending voice mode or navigating
  away. Nested pickers must not dismiss the settings popover. Language and
  intelligence update the active Realtime session; the microphone picker swaps
  the live WebRTC input track and remembers the browser-local device choice;
  an output-voice change updates immediately only before the assistant has
  emitted audio and otherwise applies to the next conversation, matching the
  Realtime API's voice immutability rule.
- Semantic VAD keeps listening, starts responses automatically, and supports
  barge-in while the agent is speaking.
- Function calls must cross the authenticated realtime tool bridge and enter
  `executeAgentToolCall`. Never call `ActionEntry.run` directly: that bypasses
  schema validation, approvals, audit/journal behavior, timeouts, redaction,
  and mutation refreshes.
- Preserve the active browser-tab id in request context so `set-url-path`,
  `set-search-params`, `view-screen`, and tab-scoped application state affect
  the app the user is actually speaking to. Realtime tool manifests are capped,
  so prioritize `navigate`, `set-url-path`, `set-search-params`, and
  `view-screen` before packing large template registries.
- Keep `tool-search` in the initial Realtime manifest. A successful specific
  search may expand the live session with matching action schemas, but the
  server must derive those schemas from its own registry, bind authorization
  to the authenticated voice session, and continue execution through
  `executeAgentToolCall`. Treat `session.update.tools` as a full replacement:
  preserve the pinned navigation/discovery tools, evict lower-priority entries
  within the manifest cap, and wait for `session.updated` before asking the
  model to continue with a newly discovered tool.
- Do not persist audio or interim transcript deltas. Append completed user and
  assistant utterances as ordinary text messages to the exact chat thread
  captured when the session starts. Input transcription completes
  asynchronously from assistant generation, so reserve provider conversation
  item order and publish only the contiguo

CAUTIONgrade B · trust 89/100 Install with care. The audit found things worth knowing before you trust its output.

What the source does

Filesystem

none-observed

Network

none-observed

Shell

none-observed

Dependencies

pinned

Secrets in source

none-found

Findings (5)

.agents/skills/an

.agents/skills/content-product-development

.agents/skills/design-exploration

.agents/skills/turn-into-app

.agents/skills/visual-edit

Gates applied: no_behavioural_pass.

[![ModelsAtlas audit: Voice Transcription](https://modelsatlas.com/skills/builderio/voice-transcription/badge.svg)](https://modelsatlas.com/skills/builderio/voice-transcription/)

Not your listing, or something out of date? Claim or update your listing.