ASR is a capability, not a device

A watch recorder needs speech recognition, but speech recognition should not belong to the watch integration. The same file may come from a phone, a meeting recorder or an agent workflow. Requiring each source to choose its own model, authentication and transcript format repeats the hardest part of the experience.

ConnectOnion 1.9.3rc23 gives these sources one capability: co asr recording.m4a. Codex is the default; the existing co configuration can select ConnectOnion's own transcription instead. An explicit provider override affects one call. Provider failure never changes where an audio file is sent.

We considered installing the separate Rust codex-asr CLI, using the Codex app-server realtime protocol, or requiring a local Whisper model. The owner chose an integrated Python adapter, informed by Wangnov/codex-asr, so no extra ASR command, runtime or local proxy needs installation. Existing transcribe() users keep their API; the new provider selector does not silently change their behavior.

The Codex adapter uploads the selected audio to an unofficial Desktop dictation endpoint using the local file-backed login. It is cloud recognition, and a successful local adapter does not establish permanent account entitlement or unlimited use. One synthesized short English sentence completed a live request; long recordings, mixed languages and recognition quality still need evaluation. An original recording must survive every processing failure.

A file-size guard also turned out to be the wrong product boundary: a three-hour meeting should not become unusable because one HTTP request has a budget. Local FFmpeg windows make that a request concern. Each completed part is saved before the next starts, so a failed request does not restart the whole meeting. We also removed a tempting text-similarity merge: two matching phrases can be genuine repeated speech. Raw words remain in order, including possible boundary overlap; any later cleaned-up account is a separate version.

The PyPI preview and GitHub prerelease are public, with matching artifact hashes and a verified clean installation. The ASR guide records setup, recovery and limitations. Watch pairing, the background Host and automatic REM ingestion remain separate work.