feat: add live_bidi_streaming_demo_app sample

[kazunori279]

Description

Adds a standalone FastAPI application demonstrating the full ADK bidirectional streaming lifecycle over a WebSocket, with a browser UI for text, voice, and camera input.

Unlike the other samples under contributing/samples/live/, this one does not run under adk web. It owns its own Runner, LiveRequestQueue, and run_live() loop, so it shows what an application built directly on ADK's streaming primitives looks like end to end:

1. Application initialization (once at startup) — Agent, SessionService, Runner
2. Session initialization (per connection) — Session, RunConfig, LiveRequestQueue, start run_live()
3. Bidi-streaming — concurrent upstream (client → queue) and downstream (events → client) tasks
4. Termination — close the LiveRequestQueue in a finally block

The bundled UI renders every raw ADK Event as it arrives, which makes it useful for inspecting Live API behavior directly.

Why here

This app is the reference implementation for the Gemini Live API Toolkit development guide, which walks through its source line by line with 36 pinned line-level references. It is currently published under google/adk-samples; moving it here keeps the guide's references alongside the ADK source they explain. The docs will be re-pointed at this path once this merges.

Only the minimum set needed to run and to satisfy those references is included. Deliberately excluded, since contributing/README.md scopes samples as "minimal and simplistic": Dockerfile, Agent Engine deployment scripts, pyproject.toml/uv.lock, and the test suite.

Testing Plan

Tested against ADK 2.5.0 on Vertex AI with gemini-live-2.5-flash-native-audio in us-east1.

Automated

An 8-test suite (kept in the source repo, not shipped here) spawns the real uvicorn server and talks to the real Live API. 8/8 pass, 33.9s.

Test	Coverage
T1	/ serves index.html; static JS/CSS return 200
T2	Text turn → audio parts + outputTranscription + turnComplete
T3	Real synthesized speech streamed as 20 ms PCM frames; asserts inputTranscription and a matching reply
T4	Image sent via {"type":"image"}, then a prompt about it
T5	Search-triggering prompt yields executableCode / functionCall / webSearchQueries / groundingChunks
T6	?proactivity=true&affective_dialog=true accepted and applied
T7	Reconnect with the same session_id reuses the session
T8	Disconnect closes LiveRequestQueue with no unhandled error

Browser

Chrome DevTools MCP: page loads, WebSocket auto-connects, a text turn produced the expected reply with streaming output transcription, token usage, and turn-complete events; with "Show audio" enabled the console showed audio/pcm response frames. 0 ERROR lines in the server log across the session; the only console error is a favicon.ico 404. Teardown logged the clean sequence (Client disconnected → Closing live_request_queue).

In place, from this path

Ran uvicorn main:app from contributing/samples/live/live_bidi_streaming_demo_app/app/ to confirm the google_search_agent import and static paths resolve here, then drove one live WebSocket turn against it — 3 audio parts returned, transcript "4" for "What is 2 plus 2?".

Lint

pre-commit run --files <all 13 files> — all hooks pass (isort, pyink, addlicense, ADK Compliance Checks, mdformat, codespell, whitespace).

Not covered

The Gemini API (non-Vertex) path is untested — no API key was available on the test machine. The code path differs only in which env vars load_dotenv() supplies to google-genai.

Notes for reviewers

• agent vs root_agent. Sibling live samples export root_agent because they load under adk web. This one is a standalone app that imports the symbol itself, and it exports agent to stay consistent with the published guide's code snippets. Happy to rename if you'd prefer uniformity — it would mean a follow-up docs update.
• Native audio only. The app always requests the AUDIO response modality with transcription in both directions. Half-cascade support was removed during testing because no half-cascade model was reachable (gemini-live-2.5-flash serves only from global; gemini-2.0-flash-live-001 only from us-east1).
• .env.example, not .env. The repo .gitignore blocks .env, so setup is cp .env.example app/.env.