ten_ai_base

Development setup

Install uv, task, and tman, then prepare the platform-specific TEN runtime and Python environment:

task prepare

The command installs TEN packages into .ten and creates or updates the local dev/.venv from dev/pyproject.toml. Run checks through the same managed environment with task lint and task test.

The patched TEN Python runtime is installed into dev/.venv as a local development dependency, including its native libraries. After task prepare, run package tests with uv run --project dev pytest tests; no PYTHONPATH configuration is needed.

The root pyproject.toml declares the reusable package only. Local sources and standalone development tools belong to the separate dev project. A parent uv workspace can use ten_ai_base as a member and configure its own TEN runtime source without requiring this repository's .ten directory.

TTS request tracing

AsyncTTS2BaseExtension uses the host's OpenTelemetry provider through the ten_ai_base.tts2 tracer. It creates one tts.request span per accepted request, starting with the first text chunk and including queue wait. The first chunk's parent context and baggage are retained across queued and interleaved chunks. The request context is also used by TTS output helpers called from background provider tasks.

Spans end at finish_request, processing failure, flush, or shutdown. Attributes include tts.request_id and tts.request.outcome (completed, error, or interrupted); synthesis errors set the span's error status. Text and audio are not added to span attributes.

The host configures the SDK provider and exporters. Without a provider, valid incoming context still propagates, but no request spans are recorded. Tracing exceptions are suppressed at instrumentation boundaries so they do not block callbacks or message delivery. Application and transport exceptions retain their existing behavior. Cancellation and failures still follow the existing TTS request state machine; tracing does not change request scheduling.