CRITICAL - google-adk 2.2.0 breaks Agent Engine/Runtime observability.

[vishal-bulbule]

Describe the Bug:

google-adk 2.2.0 breaks Vertex AI Agent Engine observability. Since v2.2.0, the Agent Engine dashboard (Sessions, Agent invocations, Model calls, Models panels) shows 0 for every deployment, even though spans are ingested correctly.

Root cause: get_gcp_resource() in src/google/adk/telemetry/google_cloud.py:252 emits the cloud resource identifier under the key cloud.resource.id (dots):

if cloud_resource_id is not None:
    resource_attributes["cloud.resource.id"] = cloud_resource_id   # should be "cloud.resource_id"

The OpenTelemetry semantic convention — and the key the dashboard's BigQuery query filters on — is cloud.resource_id (underscore):

>>> from opentelemetry.semconv.resource import ResourceAttributes
>>> ResourceAttributes.CLOUD_RESOURCE_ID
'cloud.resource_id'

The dashboard filters per deployment with ENDS_WITH(JSON_VALUE(resource.attributes,'$."cloud.resource_id"'), <agentEngineUri>). Because ADK writes the non-standard cloud.resource.id, this filter matches nothing → every panel reads 0. The value is correct (a GCP full resource name); only the attribute key is wrong.

Steps to Reproduce:

1. Install google-adk==2.2.0.
2. Reproduce directly in the library (no GCP required):

   python3 - <<'PY'
   import os
   os.environ["GOOGLE_CLOUD_AGENT_ENGINE_ID"] = "1234567890"
   os.environ["GOOGLE_CLOUD_LOCATION"] = "us-central1"
   from google.adk.telemetry.google_cloud import get_gcp_resource
   r = get_gcp_resource("my-project")
   print("cloud.resource_id :", r.attributes.get("cloud.resource_id"))
   print("cloud.resource.id :", r.attributes.get("cloud.resource.id"))
   PY

3. (End-to-end) Deploy any agent to Agent Engine: adk deploy agent_engine --project=PROJECT --region=us-central1 <agent_dir>.
4. Invoke the deployed agent through a session (Playground or stream_query), then open the deployment's Dashboard → Overview / Models in the console.
5. Observe all panels show 0 / "no rows".

Expected Behavior:

get_gcp_resource() sets the OTel-standard cloud.resource_id attribute, so the dashboard's filter matches and the Sessions / Agent invocations / Model calls / Models panels populate.

Observed Behavior:

cloud.resource_id is None; the value is emitted under cloud.resource.id instead, so the dashboard matches 0 rows and every per-deployment panel shows 0.

Output of step 2:

cloud.resource_id : None
cloud.resource.id : //aiplatform.googleapis.com/projects/my-project/locations/us-central1/reasoningEngines/1234567890

Environment Details:

• ADK Library Version (pip show google-adk): 2.2.0 (also present in 2.3.0 and main)
• Desktop OS: macOS
• Python Version (python -V): Python 3.12.12

Model Information:

• Are you using LiteLLM: No
• Which model is being used: gemini-2.5-flash

---

🟡 Optional Information

Regression:

Yes — this is a regression introduced in v2.2.0 (commit ffa057c1, 2026-06-04, "feat: Add --trigger_sources and ADK service options to cli_deploy_agent_engine"). Present in v2.3.0 and main.

Confirmed by A/B deploying the same agent code on the same project:

ADK version	Resource key emitted in _Trace	Dashboard Overview
2.1.0	cloud.resource_id (correct)	Sessions 2, turns/session 1.5, invocations 3 ✅
2.2.0	cloud.resource.id (wrong)	Sessions 0 ❌

Mechanism: in ≤2.1.0 the correct cloud.resource_id was produced by GoogleCloudResourceDetector. The v2.2.0 agent-engine branch in get_gcp_resource() returns early, skips GoogleCloudResourceDetector, and hardcodes the typo'd cloud.resource.id — so it both drops the correct key and adds a wrong one.

Logs:

Span resource keys in the Agent Engine _Trace store, per deployed engine (BigQuery over the _Trace linked dataset). Every 2.2.0/2.3.0 engine has 0 spans with the standard key; the 2.1.0 engine has the correct key:

+---------------------+--------------+------------+   engine (ADK ver)
| engine_id           | standard_key | dotted_key |
+---------------------+--------------+------------+
| 2361379891288473600 |           16 |          0 |   2.1.0  -> dashboard works
| 4220240637485645824 |            0 |         22 |   2.2.0  -> dashboard 0
+---------------------+--------------+------------+

Raw resource.attributes of a 2.2.0 span (note the dotted key):

{
  "cloud.platform": "gcp.agent_engine",
  "cloud.resource.id": "//aiplatform.googleapis.com/projects/.../locations/us-central1/reasoningEngines/...",
  "service.name": "...",
  "telemetry.sdk.name": "opentelemetry",
  "telemetry.sdk.version": "1.41.1"
}

Screenshots / Video: Click to Open

Additional Context:

Suggested fix — use the OTel constant so the key can't drift from the standard:

from opentelemetry.semconv.resource import ResourceAttributes
resource_attributes[ResourceAttributes.CLOUD_RESOURCE_ID] = cloud_resource_id  # "cloud.resource_id"

(Minimal equivalent: resource_attributes["cloud.resource_id"] = cloud_resource_id.)

The existing tests/unittests/telemetry/test_google_cloud.py::test_get_gcp_resource only asserts gcp.project_id, so no test pins the wrong key; a case asserting cloud.resource_id is set when the agent-engine env vars are present would lock in the fix.

Minimal Reproduction Code:

import os
os.environ["GOOGLE_CLOUD_AGENT_ENGINE_ID"] = "1234567890"
os.environ["GOOGLE_CLOUD_LOCATION"] = "us-central1"

from google.adk.telemetry.google_cloud import get_gcp_resource
r = get_gcp_resource(project_id="my-project")

print("cloud.resource_id :", r.attributes.get("cloud.resource_id"))  # None  <-- dashboard expects this
print("cloud.resource.id :", r.attributes.get("cloud.resource.id"))  # value (wrong key)

How often has this issue occurred?:

• Always (100%)