Summary

Upgrading from 2026.6.8 → 2026.6.9 silently breaks Groq voice-note transcription (STT). Every Telegram voice message arrives with an empty transcript, and openclaw infer audio transcribe fails with:

[media-understanding] audio: failed (0/1) reason=Media provider not available
Error: Media provider not available: groq

openclaw infer audio providers returns an empty list.

Root cause

In 2026.6.9, official providers were externalized into standalone npm packages ("Standalone official provider plugins", PR #93470). The Groq media-understanding provider moved to @openclaw/groq-provider.

Existing configs that reference the provider inline still work as before:

// ~/.openclaw/openclaw.json
"tools": { "media": { "audio": { "models": [
  { "provider": "groq", "model": "whisper-large-v3", "capabilities": ["audio"] }
] } } }

…but after the upgrade the core no longer bundles groq, so audio-provider resolution fails. Nothing during npm install -g [email protected] or at gateway startup tells the operator that the provider now needs to be installed separately.

Impact

• Silent failure: the agent receives an empty transcript. In a chat channel this is worse than an error — the model may hallucinate a reply to a voice note it never "heard", with no indication that STT failed.
• Re-registering the provider in models.providers does not help — the media-understanding registry is separate from the model/chat provider registry, so the only fix is installing the externalized plugin.

Reproduction

1. On 2026.6.8, configure Groq audio (tools.media.audio.models = [{ provider: "groq", model: "whisper-large-v3" }]) with a valid Groq API key. Send a Telegram voice note → it is transcribed and answered correctly.
2. npm install -g [email protected] and restart the gateway.
3. Send a Telegram voice note, or run:

   openclaw infer audio transcribe --file voice.ogg
   

4. Result: Error: Media provider not available: groq. openclaw infer audio providers is empty.

Environment

• OpenClaw 2026.6.9 (c645ec4)
• Linux, Node 24.16, gateway as a systemd user service
• Groq STT model: whisper-large-v3

Workaround / fix that works

Install and enable the externalized provider plugin, then restart:

openclaw plugins install @openclaw/groq-provider
openclaw plugins enable groq
systemctl --user restart openclaw-gateway.service

After this, openclaw infer audio providers lists groq (available:true, capabilities:["audio"]) and transcription works again — no config change and no downgrade required.

Suggestions

1. On upgrade, detect configs that reference a now-externalized provider and either auto-install it or print an actionable hint: "provider 'groq' moved to @openclaw/groq-provider; run openclaw plugins install @openclaw/groq-provider".
2. Make the media-understanding error name the missing package (e.g. "audio provider 'groq' not installed — install @openclaw/groq-provider") instead of the generic Media provider not available: groq.
3. Avoid handing an empty transcript to the agent as if it were valid input (related to the earlier closed report [Bug]: infer audio transcribe and infer image describe emit "No transcript/description returned for ... <path>" when no provider is configured, blaming the input instead of the missing config #73569).

Related

• PR fix(plugins): load externally-installed channel plugins at gateway startup #93470 — externalized official provider plugins (the change that introduced this)
• Historical (closed) Groq STT regressions: [Bug]: Groq Whisper STT completely broken since 2026.3.31 — "deprecate legacy provider compat subpaths" broke audio provider resolution #59875, Groq media understanding provider dropped from registry after first message — voice transcription works once then fails #60518, v2026.4.5: Telegram voice messages no longer transcribed (STT regression) #62205