Python: [Bug]: `oauth_consent_request` produces error within same session

[cecheta]

Description

Following the sample https://github.com/microsoft/agent-framework/tree/main/python/samples/04-hosting/foundry-hosted-agents/responses/04_foundry_toolbox, using the ghmcpoauth auth, I found that after accepting the oauth_consent_request by clicking the consent link, then sending a follow-up request using the previous response ID (as explained here), all subsequent requests failed with the following:

"error": {
  "code": "server_error",
  "message": "MCP server failed to initialize: Cancelled via cancel scope 7f3360511e50"
},

If you try again with a new conversation, then it works.

Code Sample

Error Messages / Stack Traces

"error": {
  "code": "server_error",
  "message": "MCP server failed to initialize: Cancelled via cancel scope 7f3360511e50"
},

Package Versions

agent-framework-foundry-hosting: 1.0.0a260709

Python Version

Python 3.13.14

Additional Context

No response

[github-actions]

🤖 Automated triage reproduction notes (agent-authored — trust but verify)

Agent analysis

Repro: FoundryToolbox.close() in packages/foundry_hosting/agent_framework_foundry_hosting/_toolbox.py (lines 188-190) permanently nulls out _httpx_client, which carries the _ToolboxAuth bearer-token handler. After a consent-required McpError from load_tools() in _mcp.py:1373 propagates to MCPTool.aenter (line 2650) which calls self.close(), all subsequent get_mcp_client() calls pass http_client=None to streamable_http_client, losing authentication. Minimal repro: construct FoundryToolbox with a credential, call close(), then observe _httpx_client is None and get_mcp_client() produces an unauthenticated transport.

• Failing test: python/packages/foundry_hosting/tests/test_repro_7166.py
• Files examined: python/packages/foundry_hosting/agent_framework_foundry_hosting/_toolbox.py, python/packages/foundry_hosting/agent_framework_foundry_hosting/_responses.py, python/packages/core/agent_framework/_mcp.py, python/packages/core/agent_framework/_agents.py, python/packages/core/agent_framework/exceptions.py, python/packages/foundry_hosting/tests/test_responses.py, python/packages/foundry_hosting/tests/test_toolbox.py
• Tests run: python/packages/foundry_hosting/tests/test_repro_7166.py
• Reported version: 1.0.0a260709
• Current version: 1.0.0a260709

[hxaxd]

I traced this to the FoundryToolbox lifecycle. The consent-required tool-loading response causes the toolbox to close its authenticated HTTP client, and the next connection on the same toolbox object falls back to a client without the Foundry bearer-token and call-id auth handler.

I'd like to submit a fix that retains the client configuration and recreates the same authenticated client when the toolbox reconnects, with a regression test covering close followed by an authenticated reconnect. Would that approach work for you?

[pamelafox]

Confirmed that this also affects a Python Foundry hosted agent using a Work IQ RemoteA2A toolbox connection.

Observed with:

• agent-framework-core==1.11.0
• agent-framework-foundry==1.10.1
• agent-framework-foundry-hosting==1.0.0a260709
• Python 3.13

Sequence from the hosted session logs:

1. The toolbox MCP initialization and tool-list requests returned 200/204.
2. Work IQ returned CONSENT_REQUIRED, which the host surfaced as oauth_consent_request using the compatibility handling described in .NET: Python: [Bug]: ResponsesHostServer does not surface Work IQ CONSENT_REQUIRED from a2a_preview source #7227.
3. The user completed browser consent successfully.
4. Resending in the same hosted session/conversation caused the toolbox MCP initialization request to return 401 Unauthorized, followed by MCP server failed to initialize: Cancelled via cancel scope ....
5. Starting a fresh agent session and conversation after consent succeeded: MCP initialization and tool discovery returned 200, and work-iq-connection___SendMessage completed successfully.

This matches the triage root cause here and confirms the impact is not limited to ghmcpoauth; it also affects Work IQ delegated OAuth. #5594 tracks automatic turn resumption, but until that exists, the documented manual retry needs the authenticated FoundryToolbox reconnect fixed by this issue.

This comment was written by GitHub Copilot from reproduced hosted-agent logs and verified package source inspection.