openclaw - 💡(How to fix) Fix [Bug]: Telegram voice messages not transcribed via Groq Whisper after multi-agent config (Docker) [1 participants]

Bug type

Regression (worked before, now fails)

Beta release blocker

No

Summary

Telegram voice messages (.ogg) are received and saved to media/inbound/ correctly, but Groq Whisper transcription is never triggered. Agent responds asking user to send text instead. No audio-related logs appear even with OPENCLAW_DIAGNOSTICS=telegram.,media. enabled.

Steps to reproduce

1. Deploy OpenClaw via Docker (ghcr.io/openclaw/openclaw:latest)
2. Configure multiple agents via agents.list (main, iris, sigma)
3. Enable Groq Whisper: tools.media.audio.enabled=true, models=[{provider:groq, model:whisper-large-v3}]
4. Set echoTranscript=true and scope.default=allow
5. Connect Telegram bot to main agent
6. Send a voice message via Telegram
7. Agent responds with "I can't process audio" — no transcription attempt logged

Expected behavior

Voice message should be transcribed by Groq Whisper and the transcript should be passed to the agent as text. With echoTranscript=true, a confirmation of the transcript should appear in Telegram before the agent response.

Actual behavior

Agent receives the voice message (.ogg file saved to media/inbound/) but transcription is never triggered. No audio or media-related logs appear. Agent responds asking user to send text instead. echoTranscript produces no output.

OpenClaw version

2026.4.2 (unknown)

Operating system

Ubuntu 24.04 (Hetzner CX23)

Install method

Docker

Model

ghcr.io/openclaw/openclaw:latest (version 2026.4.2)

Provider / routing chain

Anthropic / claude-haiku-4-5 → Telegram → main agent (default) Audio provider: Groq Whisper (whisper-large-v3) via GROQ_API_KEY env var

Additional provider/model setup details

• 3 agents configured in agents.list: main (default), iris, sigma
• Groq API accessible from container (curl to api.groq.com returns 200)
• GROQ_API_KEY loaded correctly (openclaw models status shows groq as available: env=gsk_...)
• tools.media.audio config confirmed correct via openclaw config get tools.media.audio
• .ogg files arrive in media/inbound/ confirming Telegram delivery works
• OPENCLAW_DIAGNOSTICS=telegram.,media. produces zero audio-related logs when voice message is sent
• applyMediaUnderstanding never appears in logs
• Issue appeared after Docker migration and multi-agent configuration
• Single agent setup (before adding iris and sigma) may have worked

Logs, screenshots, and evidence

Evidence collected during diagnosis:

1. .ogg files confirmed arriving in media/inbound/:
file_27---931e81c1.ogg (Apr 6 01:51)
file_26---e7982d2f.ogg (Apr 6 01:48)

2. Groq API accessible from container:
$ curl -H "Authorization: Bearer $GROQ_API_KEY" https://api.groq.com/openai/v1/models → 200 OK
$ curl .ogg file to Groq transcriptions endpoint → 200 OK

3. openclaw models status shows Groq available:
- groq effective=env:gsk_KFst...MZRDeGxZ | source=env: GROQ_API_KEY

4. tools.media.audio config:
{
  "enabled": true,
  "echoTranscript": true,
  "scope": { "default": "allow" },
  "models": [{ "provider": "groq", "model": "whisper-large-v3" }]
}

5. OPENCLAW_DIAGNOSTICS=telegram.*,media.* — zero audio-related logs when voice sent

6. docker compose logs after voice message sent — no entries generated

7. Agent response to voice message:
"Não consigo processar áudio. Se quiser que eu responda algo, manda em texto."

Related issues: #7899, #7460

Impact and severity

High — core feature (voice message transcription) completely non-functional in Docker multi-agent deployment. Workaround is manual text input only.

Additional information

• Deployment: Docker (ghcr.io/openclaw/openclaw:latest), Ubuntu 24.04, Hetzner CX23
• Transcription worked before Docker migration (npm install -g openclaw, version 2026.3.24)
• Issue persists after: docker compose down/up, gateway restart, config reset, scope.default=allow addition
• OPENCLAW_DIAGNOSTICS confirms zero media pipeline activity — applyMediaUnderstanding not being called
• Single agent config (before adding iris/sigma via agents.list) may not reproduce the issue
• Note: audio was never tested in Docker with single agent — unclear if issue is Docker-specific or multi-agent-specific