Voice-call streaming transcription fixed

vincentkoc

·Apr 5, 2026·#61224fix(voice-call): use full config for realtime transcription

A regression in the voice-call plugin prevented realtime transcription providers from being discovered during streaming calls. The bug has been patched — calls should connect again.

Voice calls with realtime transcription had stopped working. A race condition in configuration passing meant the OpenAI realtime transcription provider could not be found, leaving the webhook server unable to resolve the plugin it needed to process streaming audio. The fix passes the complete OpenClaw configuration through to the provider resolution logic, ensuring the plugin registry is visible when the webhook server initializes. This was a beta release blocker. Voice-call users can now receive and process streaming calls as expected.

View Original GitHub Description

Summary

pass the full OpenClaw config into VoiceCallWebhookServer so realtime transcription provider resolution sees the same plugin registry/config surface as realtime voice
resolve streaming transcription providers from fullConfig instead of narrowing back to coreConfig
add a runtime regression test for the constructor wiring and record the user-facing fix in the changelog

Fixes #60936.

Related but separate: #61008 fixed Telegram DM voice-note transcription and does not address the voice-call regression here.

Test plan

pnpm exec vitest run --config vitest.config.ts extensions/voice-call/src/runtime.test.ts -t "passes fullConfig to the webhook server for streaming provider resolution"
pnpm exec vitest run --config vitest.config.ts extensions/voice-call/src/webhook.test.ts -t "auto-selects the first registered provider when streaming.provider is unset"
pnpm test:serial -- extensions/voice-call/src/runtime.test.ts -t "passes fullConfig to the webhook server for streaming provider resolution" was attempted, but the wrapper stalled after startup in this workspace