Bug type

Behavior bug (incorrect output/state without crash)

Beta release blocker

No

Summary

With tools.media.audio.enabled: true and an OpenAI provider entry whose API key verifiably works against /v1/audio/transcriptions, both the Telegram voice-note path and the openclaw infer audio transcribe CLI silently fail to produce any transcript for valid OGG/Opus files, with no error or warning surfaced in the default gateway log.

Steps to reproduce

1. Set the following in ~/.openclaw/openclaw.json:
   • tools.media.audio.enabled: true
   • tools.media.audio.models[0]: { "provider": "openai", "model": "gpt-4o-mini-transcribe" }
   • models.providers.openai: { "baseUrl": "https://api.openai.com/v1", "models": [], "apiKey": "sk-proj-...", "auth": "api-key" }
2. openclaw config validate → Config valid: ~/.openclaw/openclaw.json.
3. systemctl --user restart openclaw-gateway.service; gateway log reports ready (6 plugins, ~2s).
4. Run: openclaw infer audio transcribe --file /home/maoa/.openclaw/media/inbound/file_0.ogg --model openai/gpt-4o-mini-transcribe --language es (Source file is a valid Telegram voice note, audio/ogg; codecs=opus, 120 KB, Spanish speech, previously downloaded by the gateway Telegram channel into media/inbound/.)
5. Observe the CLI output below in Actual behavior.

Expected behavior

1. Set the following in ~/.openclaw/openclaw.json:
   • tools.media.audio.enabled: true
   • tools.media.audio.models[0]: { "provider": "openai", "model": "gpt-4o-mini-transcribe" }
   • models.providers.openai: { "baseUrl": "https://api.openai.com/v1", "models": [], "apiKey": "sk-proj-...", "auth": "api-key" }
2. openclaw config validate → Config valid: ~/.openclaw/openclaw.json.
3. systemctl --user restart openclaw-gateway.service; gateway log reports ready (6 plugins, ~2s).
4. Run: openclaw infer audio transcribe --file /home/maoa/.openclaw/media/inbound/file_0.ogg --model openai/gpt-4o-mini-transcribe --language es (Source file is a valid Telegram voice note, audio/ogg; codecs=opus, 120 KB, Spanish speech, previously downloaded by the gateway Telegram channel into media/inbound/.)
5. Observe the CLI output below in Actual behavior.

Actual behavior

CLI path. The command

openclaw infer audio transcribe --file /home/maoa/.openclaw/media/inbound/file_0.ogg --model openai/gpt-4o-mini-transcribe --language es

returns (redacted real path):

Error: No transcript returned for audio: /home/maoa/.openclaw/media/inbound/file_0.ogg

The string No transcript returned for audio: originates in the bundled capability-cli-C8QLmK_t.js:656, which throws when the resolved transcript text is empty or missing.

ACP / Telegram path. On Telegram voice-note delivery, the gateway downloads the file to ~/.openclaw/media/inbound/file_X---<uuid>.ogg, then delivers the user message to the agent session with the raw template intact:

[media attached: /home/maoa/.openclaw/media/inbound/file_0---<uuid>.ogg (audio/ogg; codecs=opus) | .../file_0---<uuid>.ogg]
...
<media:audio>

No [Audio] block replaces Body, {{Transcript}} is not populated, and no error or warn line appears in journalctl --user -u openclaw-gateway.service or in the gateway log file under /tmp/openclaw/openclaw-*.log.

Counter-evidence that the file and key are valid. The exact same file and the exact same API key return HTTP 200 with the correct transcript when posted directly to OpenAI, outside OpenClaw:

$ curl -s -o /tmp/oai.json -w "%{http_code}\n" \
    -H "Authorization: Bearer $OPENAI_API_KEY" \
    -F "file=@/home/maoa/.openclaw/media/inbound/file_0.ogg" \
    -F "model=gpt-4o-mini-transcribe" \
    https://api.openai.com/v1/audio/transcriptions
200
$ jq -r .text /tmp/oai.json
Es decir que me entiendes si escribo un mensaje así de voz? Confírmame si me entiendes.

The same request with type=audio/ogg, type=audio/ogg; codecs=opus, and type=application/octet-stream all return HTTP 200 with the same transcript; MIME type is not the differentiator.

OpenClaw version

2026.4.9 (0512059)

Operating system

Ubuntu 24.04.4 LTS under WSL2 (kernel 6.6.87.2-microsoft-standard-WSL2)

Install method

npm global — installed under ~/.npm-global/lib/node_modules/openclaw ; gateway runs as systemd --user service (openclaw-gateway.service) on loopback bind

Model

openai/gpt-4o-mini-transcribe (audio transcription; model under test)

Provider / routing chain

openclaw -> api.openai.com/v1/audio/transcriptions (direct, no proxy, no gateway-in-front)

Additional provider/model setup details

• Agent loop model (not under test for this bug, included for completeness): openai-codex/gpt-5.4 via ChatGPT OAuth, configured under agents.defaults.model.primary. This is a different auth path than the transcription call and is functioning correctly for the text-only agent loop.
• Transcription auth path used for this bug: models.providers.openai.apiKey = direct OpenAI API key (sk-proj-...), independent from the Codex OAuth profile. The key has Audio + Embeddings scopes in the OpenAI platform dashboard.
• The same API key is also duplicated under skills.entries.openai-whisper-api.apiKey (where openclaw skills check openai-whisper-api reports ✓ Ready, Environment: ✓ OPENAI_API_KEY) and under plugins.entries.memory-lancedb.config.embedding.apiKey (where the memory-lancedb plugin loads and successfully produces text-embedding-3-small embeddings — confirming the same key works for other OpenAI endpoints from within OpenClaw). Only the tools.media.audio transcription path fails.
• openclaw infer audio providers reports for openai: {"available": true, "configured": true, "selected": false, "id": "openai", "capabilities": ["image", "audio"]}. NOT_ENOUGH_INFO on whether selected: false is the idle state or the symptom; the maintainer can answer this directly.
• Telegram channel config: dmPolicy: pairing, groupPolicy: allowlist, single-user setup with 2 paired senders. Voice notes arrive via the standard bot polling provider.

Logs, screenshots, and evidence

**Gateway restart log (redacted, the relevant lines only):**

    Apr 11 11:09:29 [gateway] loading configuration…
    Apr 11 11:09:29 [gateway] resolving authentication…
    Apr 11 11:09:31 [gateway] agent model: openai-codex/gpt-5.4
    Apr 11 11:09:31 [gateway] ready (6 plugins, 2.2s)
    Apr 11 11:09:31 [gateway] starting channels and sidecars...
    Apr 11 11:09:31 [telegram] [default] starting provider (@<redacted>_bot)

No `error` or `warn` line contains `audio`, `transcri`, `ogg`, `whisper`, or `media` between gateway start and the failed CLI invocation. Verified with:

    grep -iE 'audio|transcri|ogg|whisper|media' /tmp/openclaw/openclaw-2026-04-11.log

which returns no matches relevant to the failure path, only the earlier unrelated `[warn] memory: sqlite-vec unavailable` line from the `memory-core` subsystem, which is not related to `tools.media.audio`.

**Session JSONL excerpt showing the unreplaced body on the ACP path (redacted):**

    {"type":"message","message":{"role":"user","content":[{"type":"text","text":"[media attached: /home/maoa/.openclaw/media/inbound/file_0---<uuid>.ogg (audio/ogg; codecs=opus) | .../file_0---<uuid>.ogg]\n...\n<media:audio>"}]}}

No downstream message in the same session contains `[Audio]` or a `Transcript` field derived from this attachment.

**Only code-path observation (not speculation):** `dispatch-acp-rBcmOCzP.js:858` in the bundled `dist/` wraps `applyMediaUnderstanding` in `try/catch` where the error branch only calls `logVerbose(...)`; without `--verbose` the swallowed error is invisible in the default gateway log. This is an observation about the default diagnostic path, not a claim about the root cause. NOT_ENOUGH_INFO on the exact error being swallowed, because capturing `--verbose` on a live `systemd --user` gateway would require restarting the unit with different arguments and interrupting the active Telegram channel; happy to capture it if the maintainer can point me at a scoped verbose flag for `tools.media.audio` alone.

Impact and severity

• Affected: single-user personal-assistant deployment, Telegram channel with voice-note ingestion; the CLI openclaw infer audio transcribe path is also affected.
• Severity: blocks workflow for the audio understanding modality; text-only agent workflow is unaffected.
• Frequency: always. 2/2 distinct voice notes (120 KB Spanish speech, 50 KB short silence) fail via ACP; 2/2 invocations of openclaw infer audio transcribe fail via CLI. 0/0 successes on the OpenClaw path; 2/2 successes against the same files through direct curl to /v1/audio/transcriptions.
• Consequence: voice notes from users are silently dropped into the agent session as raw media markers without transcripts, so the agent cannot act on voice content; the text-only workflow continues to function.

Additional information

• This bug is isolated to the tools.media.audio transcription pipeline. All other OpenClaw subsystems exercised in this deployment are functional:
  • PDF/XML/ZIP invoice ingestion via a workspace skill
  • memory-lancedb with OpenAI embeddings (text-embedding-3-small)
  • plugins.entries.brave.config.webSearch (Brave Search API)
  • browser plugin with Google Chrome 147.0.7727.55 headless (JavaScript-heavy sites scraped successfully)
  • Telegram channel polling and dmPolicy: pairing enforcement
• NOT_ENOUGH_INFO on first known bad version or last known good version: this is the first OpenClaw install in this environment and audio was never observed working. This is therefore not asserted as a regression.
• Workaround: tools.media.audio.enabled: false and text-only user input. Functional, non-disruptive to the rest of the stack.

Report prepared by Claude (Anthropic) via Claude Code, working as Maelo's technical operator inside his live WSL environment. All evidence in this issue is grounded in direct observation of logs, configs, and CLI output from that environment; nothing has been paraphrased from third parties or inferred from other setups. Where evidence is missing, the literal marker NOT_ENOUGH_INFO appears in place of speculation.