Voicebox documentation

Live provider fixes — 2026-09-21

OpenAI tool contract (voicebox-beads-g1l)

Reproduced before edits on public main a9bd404db339daa92ea9090c12486e5d7edecf39, not inferred from the reviewer’s older cd27c8e drive. The actual OpenAI adapter ignored response.function_call_arguments.done and declared neither tools nor instructions. A subsequent transcript arrived but the requested file and reply did not. Baseline characterization.

The adapter now advertises the shared command catalogue, emits the host’s tool-call event, and sends conversation.item.create / function_call_output with the original call_id. Invalid JSON/non-object arguments return invalid-tool-call; a missing correlation ID emits a named page error instead of invoking an action. Executor refusals retain their original names. response.create waits for generation and all outstanding tool results, including either completion order, to avoid starting an OpenAI response while another one is active.

Validation uses the actual page, native Chromium fake microphone, server, executor, audit and provider. Only vendor WebSocket destinations redirect to owned loopback; synthetic keys, ephemeral ports and removable private workspace/host directories. No paid vendor/model call, real microphone, audible speech or deployed acceptance is claimed. Screenshots were captured, not visually reviewed.

The review’s other findings (failed-response status, empty close reason, malformed audio diagnostics and a provider throwing during PCM) are outside these two fixes.

Settings-selected rate (voicebox-beads-94c)

Also reproduced on unmodified a9bd404: environment Gemini plus settings OpenAI actually dialed OpenAI but declared Gemini/16000. The native capture ran at 16000 and the correct comparison reported no contradiction with that false declaration. Baseline.

beginSession now snapshots the settings provider and its model once. That provider supplies the rate, actual dial and later session metadata. Changing settings affects the next session, not the running one’s model label. The contradiction comparison is untouched. The library still supports its VOICEBOX_LIVE_PROVIDER (old name LIVE_PROVIDER) fallback when no explicit provider is supplied; the server supplies the settings provider explicitly. Invalid settings refuse at their existing API admission boundary. The frame tests now exercise that actual boundary rather than an unused environment override, and own/remove their host directories instead of creating tokens in the checkout.

The live half, driven for real (journal-6g0, 2026-09-25)

Everything above this section either redirected the vendor or declined to claim a paid call. This one is the bead’s own live half, and the claim is the other way round: a REAL, AUTHENTICATED OpenAI Realtime session, real speech in, real audio out, and machine witnesses for both directions of the rate contract.

The drive (evidence/live-fixes/6g0/drive.mjs, run under node v24.21.0) builds the session exactly as the host does — createLiveSession({ provider: "openai" }), so the dial is the 7ade86e probe-as-dial with Authorization: Bearer attached — waits for session.updated, then speaks a TTS-synthesized question (“What is two plus two”) INTO the session as 24 kHz pcm16 frames, followed by 1.4 s of room tone. The room tone is not decoration: the vendor’s server VAD end-points on silence IN THE AUDIO TIMELINE, and a buffer that merely stops appending never advances past the last word — measured: speech_started, then no commit and no response for a full timeout. Pacing the send has the opposite failure: perceived silence between chunks end-points the utterance early and the vendor answers a fragment (measured twice). Burst speech, then real silence.

What the passing receipt proves, check by check:

What stays open, stated plainly: the vendor emitted no output_audio_transcript.delta this provider sees (audio + whisper prove the turn regardless), and the provider’s session.update configures no input_audio_transcription, so there is no input-side ASR receipt — the input witness is the model’s own correct answer. Both are provider observations, not rate failures. Pitch and speed are correct BY RATE AGREEMENT (declared = sent = played = 24000, duration natural) plus the whisper transcription; no human ear was on the call.

Paid-call accounting: one TTS render, one realtime turn, and one whisper transcription per attempt; two attempts total for the clean receipt (attempt 1 is recorded in the receipt with its fragment answer).