AI-472 Add replay-safe Google ADK metrics sample

[xumaple]

This PR adds a Google ADK sample that exports ADK OpenTelemetry metrics through ReplaySafeMeterProvider. A deterministic local model keeps the sample API-key-free, and the Workflow includes a post-metrics replay boundary so duplicate observations are detectable.

Note: The sample pins the SDK revision that introduces ReplaySafeMeterProvider until that API is available in a release.

[DABH]

Reviewed at 0b92811. The core thesis holds and I verified it rather than taking the README's word for it — I probed _skip_recording directly and got 6 recordings live, 6 suppressed under Replayer, counts_after == counts_before — and the happy path works end to end against a real dev server. Switching this commit to Replayer and relative equality was the right response to the macOS failures; it's one indentation away from green.

One item that isn't tied to a line in this diff: metrics is missing from the suite scenario table. google_adk_agents/README.md:39-46 lists all six siblings and nothing in the repo links to google_adk_agents/metrics/README.md, so the sample is unreachable from any index.

| [metrics](./metrics/README.md) | Google ADK's OpenTelemetry metrics exported to a local Prometheus endpoint, with `ReplaySafeMeterProvider` keeping replay from double-counting observations. |

That row is only accurate once the nested pyproject.toml goes away — the suite's generic run block at :48-53 doesn't work for this scenario today. See the pyproject.toml:8 comment.

[DABH]

This is why CI is red on all eight jobs. The async with await WorkflowEnvironment.start_time_skipping() opened at :23 exits at :41, which shuts the ephemeral server down, and handle.fetch_history() at :50 then RPCs a dead address — RPCError: tcp connect error after nine gRPC retries against 127.0.0.1:43563 in job 98639038931. Deterministic on every platform, not a macOS timing thing this time.

Fetch the history while the server is still up. I applied exactly this and the test goes green in 0.85s:

Suggested change

	result = await handle.result()
	result = await handle.result()
	history = await handle.fetch_history()

Upstream does the same — sdk-python/tests/contrib/google_adk_agents/test_replay_metrics.py:296 fetches history as the last statement inside async with Worker(...). Note this disappears on its own if you take the test-placement suggestion at :10, since the session-scoped env fixture outlives the test.

[DABH]

Companion to the :40 suggestion — use the history fetched inside the environment block.

Suggested change

	await handle.fetch_history()
	history

[DABH]

This is the only real test in the repo outside tests/ — find . \( -name 'test_*.py' -o -name '*_test.py' \) | grep -v '^./tests/' returns two paths and the other (polling/test_service.py) has no test functions, it just matches pytest's glob. All six ADK siblings live at tests/google_adk_agents/<scenario>_test.py with the signature async def test_basic(client: Client, monkeypatch: pytest.MonkeyPatch).

The cost isn't just tidiness. tests/conftest.py doesn't apply out here, so this can't use the session-scoped env/client fixtures at :40-57, it ignores --workflow-environment (:18-23) so both CI passes run an identical path and each start a redundant server, and it hand-rolls the environment whose premature teardown is the failure at :50. It also collects first among the ADK tests purely because it sorts before tests/ — in the a058b20 macOS log it starts at 19:35:18.26 against 19:35:30.83+ for the siblings, so it alone paid ADK's cold in-workflow import anthropic (google/adk/flows/llm_flows/contents.py:62), which is what blew the 2s deadlock budget on 3/3 macOS runners.

Moving it to tests/google_adk_agents/metrics_test.py taking client: Client, dropping the inline WorkflowEnvironment, and registering LocalMetricsModel through monkeypatch the way tests/openai_agents/_mock_model.py:59 does would fix the blocker and the macOS flakiness together.

[DABH]

start_time_skipping takes plugins directly (temporalio/testing/_workflow.py:241), so the client rebuild isn't needed. Also type(environment.client)(**config) where all six siblings just write Client(**config).

Suggested change

	async with await WorkflowEnvironment.start_time_skipping() as environment:
	plugin = GoogleAdkPlugin()
	config = environment.client.config()
	config["plugins"] = [*config["plugins"], plugin]
	client = type(environment.client)(**config)
	async with await WorkflowEnvironment.start_time_skipping(
	plugins=[GoogleAdkPlugin()]
	) as environment:
	client = environment.client
	task_queue = f"google-adk-agents-metrics-{uuid.uuid4()}"

If you take this, plugin is no longer in scope for the Replayer(...) call at :49 — construct it once above the async with and pass the same instance to both.

[DABH]

getattr(point, "count", 1) reads as 1 per attribute set for a NumberDataPoint, which has value and no count — so a Counter that had doubled would still compare equal. Dead branch today since every ADK instrument here is a histogram, but it's silently defeating the assertion this test exists for. Better to fail loudly on anything unexpected:

Suggested change

	getattr(point, "count", 1) for point in metric.data.data_points
	point.count for point in metric.data.data_points

[DABH]

Undocumented scaffolding — no comment in a file with no comments, no README mention, no sibling precedent; the rationale ("a post-metrics replay boundary so duplicate observations are detectable") lives only in the PR description. At head it changes nothing measurable either way (6/6), and it only ever mattered under the max_cached_workflows=0 that's being removed.

Suggested change

	await workflow.sleep(timedelta(milliseconds=1))
	return final_text

timedelta on :1 becomes unused with it. Note root ruff runs only --select I, so an orphan import won't be flagged.

[DABH]

Companion to the pyproject.toml:8 comment — once the nested project is gone, --project has nothing to point at, and the suite README already documents the house form at :48-53.

Suggested change

	```shell
	uv run --project google_adk_agents/metrics python -m google_adk_agents.metrics.run_worker
	```
	Then run the Workflow:
	```shell
	uv run --project google_adk_agents/metrics python -m google_adk_agents.metrics.run_metrics_workflow
	```
	```shell
	uv run python -m google_adk_agents.metrics.run_worker
	```
	Then run the Workflow:
	```shell
	uv run python -m google_adk_agents.metrics.run_metrics_workflow
	```

[DABH]

rg is the only ripgrep invocation in any markdown file in the repo (grep -rln '| rg ' --include='*.md' . gives one hit) and it isn't a documented prerequisite.

Suggested change

	curl http://127.0.0.1:9464/metrics | rg 'gen_ai'
	curl -s http://127.0.0.1:9464/metrics | grep gen_ai

[DABH]

Two corrections here.

The no-inflation sentence is true — I verified it — but it's the README's only statement about recording semantics and it omits the half the SDK spells out at _meter_provider.py:221-223: recordings are first-execution-only, and a Workflow task retry re-executes live and can record again. The guard is in_workflow() and is_replaying_history_events(), and a failed WFT appends WorkflowTaskFailed without advancing the replay boundary — which is exactly how a058b20 produced assert 2 == 1 on macOS after a TMPRL1101 deadlock. For a sample whose entire subject is metric accuracy, a reader will otherwise take these as exactly-once.

Second, the instrumentation scope isn't observable in the output — the Prometheus exporter emits no otel_scope_* labels, and the real line is the munged gen_ai_invoke_agent_duration_seconds_count{gen_ai_agent_name="metrics_agent"} 1.0. And the max_cached_workflows sentence goes away with the run_worker.py:28 change.

Suggested change

	The output includes `gen_ai.invoke_agent` metrics, `gen_ai.client.operation.duration`, and `gen_ai.client.token.usage` from instrumentation scope `gcp.vertex.agent`. The worker sets `max_cached_workflows=0`, forcing replay between Workflow tasks. `ReplaySafeMeterProvider` drops observations made while replaying, so replay does not multiply the recorded counts.
	The output includes `gen_ai.invoke_agent` metrics, `gen_ai.client.operation.duration`, and `gen_ai.client.token.usage`, exported with dots munged to underscores — for example `gen_ai_invoke_agent_duration_seconds_count{gen_ai_agent_name="metrics_agent"} 1.0`. `ReplaySafeMeterProvider` drops observations made while replaying, so replay does not multiply the recorded counts.
	Recordings are first-execution-only rather than exactly-once: replay is suppressed, but a Workflow task retry re-executes live and can record again. Treat these counters as at-least-once — aggregate with rates and percentiles rather than relying on exact counts.

[DABH]

stream=True silently yields a single non-partial response rather than streaming, so a reader who points the streaming scenario at this model gets quietly wrong behavior instead of an error.

Suggested change

	async def generate_content_async(
	self, llm_request: LlmRequest, stream: bool = False
	) -> AsyncGenerator[LlmResponse, None]:
	async def generate_content_async(
	self, llm_request: LlmRequest, stream: bool = False
	) -> AsyncGenerator[LlmResponse, None]:
	if stream:
	raise NotImplementedError(
	"LocalMetricsModel does not implement streaming responses."
	)

Minor, separately: metrics_workflow.py:17 retypes "local-metrics-model" instead of importing MODEL_NAME from :8 here.