Description

All three catch boundaries in the media understanding pipeline log at verbose/debug level only. When an audio transcription fails (timeout, 5xx, empty response from Groq/OpenAI/Deepgram), the error is completely invisible at the default info log level.

Error Suppression Chain

1. runAttachmentEntries (runner) — catches errors, logs at verbose only
2. runWithConcurrency onTaskError (apply) — logs at verbose only
3. runTasksWithConcurrency — catches and continues by default

Result: user's voice message appears as <media:audio> with zero indication of why transcription failed. No log output at default level.

Impact

• Voice messages intermittently fail to transcribe with no diagnostic output
• Users must proactively enable debug logging to catch failures
• Makes debugging transcription flakiness nearly impossible

Suggested Fix

Log media understanding errors at warn level (not debug/verbose). These are external API failures that directly impact user experience and should always be visible in logs.

Environment

• OpenClaw 2026.4.2
• Provider: Groq (whisper-large-v3-turbo)
• Channel: Telegram DM