goruck/home-generative-agent v3.41.0

on GitHub

Added

• Speech-to-text providers gain an extra request body field: a JSON object merged into the transcription request, for parameters HGA does not expose as its own fields. Available on both the OpenAI and Local provider types and merged last, so it can also override a parameter HGA sets, such as model or response_format. It cannot set stream, file or input_audio, which decide how the audio and the response are carried rather than how the audio is transcribed.
• Local (OpenAI-compatible) STT providers gain a request format option: Multipart upload (the default and previous behavior) or JSON with base64 audio. The JSON shape posts the audio inline under input_audio instead of uploading it as a form file, which is what OpenRouter's transcription endpoint natively speaks. This makes provider-specific options such as keyword biasing (provider.options) reachable — OpenRouter's multipart endpoint is an OpenAI compatibility layer that accepts only file, model, language, temperature, response_format and timestamp_granularities, so a nested provider block was dropped there regardless of how it was encoded. The JSON request goes through the same configured client, so the pinned timeout, no-retry policy and keyless-Authorization handling are unchanged. It has no translations endpoint, so translate degrades to transcription with a warning, and it does not send the prompt field, which is not part of that request shape (OpenRouter ignores prompt on multipart as well). Multipart remains the default and the only option for the OpenAI provider type, whose API accepts nothing else. (#610)

Full changelog: https://github.com/goruck/home-generative-agent/blob/main/CHANGELOG.md

PR: #631