Python: Hosted Responses sample loses invoke_agent token telemetry with latest hosting stack

[singankit]

Describe the bug

The basic Agent Framework Responses hosted-agent sample does not emit the Agent Framework invoke_agent span or token usage attributes when deployed with the latest unpinned agent-framework-foundry-hosting dependency stack.

Sample:
https://github.com/microsoft-foundry/foundry-samples/tree/main/samples/python/hosted-agents/agent-framework/responses/01-basic

After pinning agent-framework-foundry-hosting==1.0.0b260730, the Agent Framework invoke_agent span and token attributes are emitted again.

Reproduction

1. Deploy the sample to Microsoft Foundry using the Responses protocol and a gpt-5 deployment.
2. Because Foundry invokes the hosted endpoint with store=true, enable resilient tasks before host startup:

from azure.ai.agentserver.core.tasks import set_resilient_tasks_enabled

set_resilient_tasks_enabled(True)

3. Invoke the deployed agent normally.
4. Inspect the correlated trace in Application Insights.

Failing dependency resolution

The unpinned sample resolved to:

• azure-ai-agentserver-core==2.1.0b2
• azure-ai-agentserver-responses==2.1.0b1
• latest available agent-framework-foundry-hosting

Observed behavior:

• The request succeeds after resilient tasks are enabled.
• Only the platform invoke_agent span is present.
• The Agent Framework invoke_agent span is absent.
• No gen_ai.usage.input_tokens or gen_ai.usage.output_tokens attributes are available.
• Runtime log contains:

WARNING agent_framework_foundry_hosting._responses: Content type 'usage' is not supported yet. This is usually safe to ignore.

Working dependency resolution

Change only the hosting requirement:

agent-framework-foundry-hosting==1.0.0b260730

This transitively resolves azure-ai-agentserver-responses==1.0.0b9 while retaining azure-ai-agentserver-core==2.1.0b2.

Observed behavior for the same sample and request:

• The Agent Framework invoke_agent span is emitted.
• gen_ai.usage.input_tokens, gen_ai.usage.output_tokens, and gen_ai.usage.reasoning.output_tokens are populated.
• Example observed values: input 29, output 160, reasoning 128.

The Content type 'usage' is not supported yet warning is still logged in the working configuration, so the warning alone does not explain the missing span. This appears to be a compatibility/regression issue involving the newer hosting/AgentServer Responses stack.

Expected behavior

The current unpinned hosted-agent stack should emit the Agent Framework invoke_agent span with model token usage attributes for successful Responses invocations, matching the behavior of the older compatible stack.

Additional context

Without set_resilient_tasks_enabled(True), store=true requests fail earlier with:

Resilient task subsystem missing in hosted environment

That execution issue is resolved by enabling resilient tasks; this report concerns the remaining missing Agent Framework span and token telemetry after successful execution.

[github-actions]

👋 Hi Ankit Singhal (@singankit) — thanks for the report!

Your reported dependency resolution and Content type 'usage' is not supported yet warning are reproducible on the current stack. However, a deterministic resilient store=true request completed and exported the Agent Framework invoke_agent span with input, output, and reasoning token attributes. This suggests the missing telemetry may depend on your deployed Foundry/Application Insights export or correlation boundary; the exact resolved hosting version and exported span hierarchy with trace, span, and parent IDs would help isolate it.

If this still reproduces after trying the above, please reopen and we'll take another look. 🙏