[azure-ai-projects] Client trace instrumentation lacks Foundry project identity

[jpalvarezl]

• Package Name: azure-ai-projects
• Package Version: 2.3.0
• Operating System: Windows 11 Enterprise 10.0.26200
• Python Version: 3.13.15

Describe the bug

Client-side invoke_agent spans created through the documented AIProjectInstrumentor and Azure Monitor path do not include microsoft.foundry.project.id. The spans reach the Application Insights resource connected to the Foundry project and include gen_ai.response.id, but they are not discoverable in that agent's Foundry Traces view.

AIProjectClient exposes the project endpoint but no public project ARM resource ID. The endpoint contains the account and project names, but not the subscription ID or resource group needed to construct the full ARM ID. AIProjectInstrumentor.instrument() also has no project identity option and does not enrich its spans with this attribute.

The current workaround is to list project connections and derive the parent project resource ID from Connection.id:

connection = next(iter(project_client.connections.list()), None)
project_arm_id = connection.id.rsplit("/connections/", 1)[0] if connection else None

This pattern also appears in the repository's human-evaluations sample, but it is not documented for client tracing and requires an extra service call plus resource-ID parsing.

To Reproduce

Steps to reproduce the behavior:

1. Create or use a Foundry project with Application Insights connected and a published Prompt Agent.
2. Set FOUNDRY_PROJECT_ENDPOINT, FOUNDRY_AGENT_NAME, and FOUNDRY_AGENT_VERSION.
3. Run:

import os

from azure.ai.projects import AIProjectClient
from azure.ai.projects.telemetry import AIProjectInstrumentor
from azure.identity import DefaultAzureCredential
from azure.monitor.opentelemetry import configure_azure_monitor

os.environ["AZURE_EXPERIMENTAL_ENABLE_GENAI_TRACING"] = "true"
AIProjectInstrumentor().instrument()

with (
    DefaultAzureCredential() as credential,
    AIProjectClient(
        endpoint=os.environ["FOUNDRY_PROJECT_ENDPOINT"],
        credential=credential,
    ) as project_client,
):
    configure_azure_monitor(
        connection_string=project_client.telemetry.get_application_insights_connection_string(),
        enable_live_metrics=False,
    )

    openai_client = project_client.get_openai_client()
    openai_client.responses.create(
        input="Reply with exactly OK.",
        extra_body={
            "agent_reference": {
                "name": os.environ["FOUNDRY_AGENT_NAME"],
                "version": os.environ["FOUNDRY_AGENT_VERSION"],
                "type": "agent_reference",
            }
        },
    )

4. Inspect the resulting invoke_agent dependency in the connected Application Insights resource. It has the agent/response attributes, but not microsoft.foundry.project.id.
5. Open the published agent's Traces view in Foundry. The client-side trace is not listed.
6. Add the full project ARM ID as microsoft.foundry.project.id to the root invoke_agent span and repeat. The client-side trace is listed.

This was also reproduced through Microsoft Agent Framework and isolated with controlled live traces in microsoft/agent-framework#7492.

Expected behavior

The supported azure-ai-projects client-tracing setup should provide a public, documented way to associate client spans with the current Foundry project so they appear in the project's Foundry Traces view.

Possible SDK-level solutions include:

• Expose the project's ARM resource ID directly from AIProjectClient.
• Add a telemetry operation that returns project identity together with the Application Insights connection information.
• Allow AIProjectInstrumentor.instrument(project_arm_id=...) and use it to enrich generated invoke_agent spans.
• Automatically resolve and cache project identity inside the instrumentor.

Consumers should not need to parse a child connection resource ID to obtain project identity.

Screenshots

The original Agent Framework report contains Foundry and Application Insights screenshots demonstrating the missing client trace: microsoft/agent-framework#7492.

Additional context

• AIProjectInstrumentor.instrument() currently accepts enable_content_recording, enable_trace_context_propagation, and enable_baggage_propagation, but no project identity.
• The official human-evaluations sample derives project_resource_id from Connection.id: https://github.com/Azure/azure-sdk-for-python/blob/main/sdk/ai/azure-ai-projects/samples/evaluations/sample_human_evaluations.py
• Foundry-hosted agentserver code receives FOUNDRY_PROJECT_ARM_ID from the platform and emits microsoft.foundry.project.id, but external client applications do not receive that environment variable automatically.
• This report is specifically about Foundry trace-list discovery. HTTPX/W3C trace-context propagation is tracked separately by configure_azure_monitor() does not auto-instrument httpx — breaks distributed tracing for OpenAI SDK / Azure AI Agent Framework (MAF) #46286 and Add httpx support #47953.

[github-actions]

Thanks for the feedback! We are routing this to the appropriate team for follow-up. cc retry-recv (@bobogogo1990) Darren Cohen (@dargilco) Glenn Harper (@glharper) Howie Leung (@howieleung) kingernupur Nikolay Rovinskiy (@nick863) Travis Angevine (@trangevi) Travis Wilson (@trrwilson).

[dargilco]

M-Hietala to triage