Add Google ADK adapter plugin prototype

[SebastianBoehler]

Summary

• add a new livekit-plugins-google-adk plugin package
• add a text-focused google_adk.LLMAdapter that wraps an ADK runner/agent behind LiveKit's llm.LLM interface
• add focused tests for prompt translation, history reconstruction, and owned-runner cleanup

Motivation

This is a prototype for first-class Google ADK support, similar in spirit to the existing LangChain plugin.

Issue: #5107

Notes

• the adapter reconstructs an ephemeral ADK session from the LiveKit ChatContext for each chat() call
• LiveKit function tools are not forwarded; tool execution is expected to happen inside ADK
• the plugin lazy-loads google-adk instead of declaring it as a hard workspace dependency because ADK currently pins opentelemetry-api / opentelemetry-sdk to <1.39, which conflicts with this repo's ~=1.39.0 pins
• because of that dependency conflict, users currently need to install google-adk separately alongside this plugin

Validation

• uv run pytest tests/test_google_adk.py
• uv run ruff check livekit-plugins/livekit-plugins-google-adk tests/test_google_adk.py

[CLAassistant]

All committers have signed the CLA.

[devin-ai-integration]

Devin Review found 1 potential issue.

View 5 additional findings in Devin Review.

[devin-ai-integration]

🔴 _delta_text fallback causes duplicate text emission when ADK produces multiple invocations with cumulative events

When an ADK run_async call produces events from multiple invocations (e.g., multi-agent or tool-calling scenarios where the model responds, invokes a tool, then responds again), _delta_text emits duplicate text. The function tracks all emitted text via concatenation in the fallback path (return text, emitted_text + text at line 323). When a second invocation starts with fresh cumulative text, its partial events don't start with the accumulated emitted_text, so they hit the fallback and are emitted in full. Then the next cumulative event from the same invocation also can't match emitted_text (which now includes text from both invocations), so it's also emitted in full — duplicating the partial text.

Concrete trace showing the duplication

ADK events: inv1 partial text="A", inv1 final text="AB", inv2 partial text="C", inv2 final text="CD"

1. text="A", emitted="" → startswith("") ✓ → delta="A", emitted="A"
2. text="AB", emitted="A" → startswith("A") ✓ → delta="B", emitted="AB"
3. text="C", emitted="AB" → startswith("AB") ✗ → fallback → delta="C", emitted="ABC"
4. text="CD", emitted="ABC" → startswith("ABC") ✗ → fallback → delta="CD", emitted="ABCCD"

Consumer receives: "A"+"B"+"C"+"CD" = "ABCCD" instead of expected "ABCD". The "C" prefix is duplicated.

Prompt for agents

In livekit-plugins/livekit-plugins-google-adk/livekit/plugins/google_adk/adk.py, the _delta_text function (lines 318-323) needs to be reworked to properly handle multi-invocation scenarios where cumulative text resets. One approach is to track the previous event's full text separately from total emitted text: when the new text doesn't start with the previous full text, treat it as a fresh invocation and reset the cumulative tracking. Alternatively, use the ADK event's invocation_id to detect invocation boundaries and reset emitted_text tracking on each new invocation. The _run method (line 182 onward) would need to pass the invocation_id context into the delta tracking logic.

---

Was this helpful? React with 👍 or 👎 to provide feedback.