Summary

• Accept OpenAI-compatible input_audio content parts on the final user message for /v1/chat/completions.
• Validate base64 payloads and supported Hermes STT audio formats, then transcribe through the existing STT pipeline before agent entry.
• Preserve mixed text/image/audio requests by replacing audio parts with transcript text while leaving text and image parts intact.
• Keep /v1/responses, assistant messages, and non-final user-message audio explicitly rejected with OpenAI-shaped errors.
• Document the API Server audio-input contract and supported formats.

Why

The API Server already accepts inline image content parts. This adds the audio side of the same OpenAI-compatible content-part surface without introducing a new provider or credential contract, because it reuses Hermes’ existing voice/STT implementation.

Verification

• scripts/run_tests.sh tests/gateway/test_api_server_multimodal.py tests/gateway/test_api_server.py
• Result: 155 passed, 91 warnings

Notes

• Inline audio still uses the API Server request body limit, so larger audio remains outside this payload path.
• The PR is draft while CI and maintainer review shake out any contract preferences.