Python: Fix FoundryAgent inheriting OPENAI_CHAT_MODEL for agent-reference requests

[Thota Sai Karthik (karthik-0306)]

Motivation & Context

RawOpenAIChatClient.__init__ resolves model from OPENAI_CHAT_MODEL/OPENAI_MODEL environment variables even when a pre-configured async_client is passed. RawFoundryAgentChatClient inherits this, causing self.model to get silently populated from an unrelated environment variable.

Agent.__init__ then copies client.model into default_options["model"]. During request preparation, _prepare_options's strip condition (options.get("model") is None) never fired because the value was already a non-None string — so an unrelated model was sent alongside the Foundry agent_reference, which the backend rejected because it must match the persisted agent's model.

Description & Review Guide

• What are the major changes?

  1. RawFoundryAgentChatClient.__init__ now explicitly sets self.model = "" after super().__init__() — Foundry agent models are resolved server-side via the persisted agent definition, not from local env vars.
  2. Updated _prepare_options's strip condition from options.get("model") is None to not options.get("model") defensively to catch empty-string models while preserving explicit caller overrides (e.g. model="gpt-5.4").
  3. Added regression tests in packages/foundry/tests/foundry/test_foundry_agent.py.

• What is the impact of these changes?

  • Prevents FoundryAgent and RawFoundryAgentChatClient from inheriting unrelated OPENAI_CHAT_MODEL, OPENAI_MODEL, or AZURE_OPENAI_* environment variables.
  • Ensures clean payloads for agent-reference requests.

• What do you want reviewers to focus on?

  • Verify that self.model = "" on RawFoundryAgentChatClient cleanly isolates client construction from OPENAI_* env vars.
  • Confirm that explicit caller model overrides (e.g., via default_options={"model": "gpt-5.4"}) survive unstripped.

Related Issue

Fixes #7272

Contribution Checklist

• The code builds clean without any errors or warnings
• All unit tests pass, and I have added new tests where possible
• The PR follows the Contribution Guidelines
• This PR is linked to an issue and there is no other open PR for this issue (see Related Issue above).
• This is not a breaking change. If it is a breaking change, add the breaking change label (or add "[BREAKING]" to the title prefix, before or after any language prefix) — a workflow keeps the label and title prefix in sync automatically.

[Copilot]

Pull request overview

Fixes a Foundry agent-reference bug where RawFoundryAgentChatClient could inherit an unrelated model from OPENAI_*/AZURE_OPENAI_* environment variables and then send that model alongside agent_reference, causing Foundry to reject the request.

Changes:

• Force RawFoundryAgentChatClient.model to "" after RawOpenAIChatClient initialization to prevent env-var model leakage.
• Broaden model-stripping logic for agent-reference requests to treat empty-string models as “unset”.
• Add regression tests ensuring env-var models don’t pollute client.model, Agent.default_options, or outbound payloads, while explicit model overrides still work.

Reviewed changes

Copilot reviewed 2 out of 2 changed files in this pull request and generated no comments.

File	Description
python/packages/foundry/agent_framework_foundry/_agent.py	Clears inherited model and updates stripping logic for agent-reference requests.
python/packages/foundry/tests/foundry/test_foundry_agent.py	Adds regression coverage for env-var model leakage and explicit model overrides.

Comments suppressed due to low confidence (1)

python/packages/foundry/agent_framework_foundry/_agent.py:386

• When self.model is forced to "" (to avoid env-var leakage), Agent.__init__ will still copy that empty string into default_options["model"] (it adds the key whenever the value is not None). For continuation calls where conversation_id is a standard resp_*/conv_* (so conversation_id is None is false), should_strip_model becomes false and the client can end up sending model="" alongside the request. That likely produces an invalid/undesired payload; it would be safer to always drop a falsy run_options["model"] (""/None) regardless of conversation_id, while still preserving non-empty explicit models for resp_* continuity.

        if not self.allow_preview:
            extra_body.setdefault("agent_reference", _build_agent_reference(self.agent_name, self.agent_version))
            should_strip_model = _uses_foundry_agent_session(conversation_id) or (
                conversation_id is None and not options.get("model")  # falsy catches None and "" (#7272)
            )
            if should_strip_model:
                run_options.pop("model", None)
        if extra_body:

[Eduard van Valkenburg (eavanvalkenburg)]

please use the defined PR template Thota Sai Karthik (@karthik-0306)

[Thota Sai Karthik (karthik-0306)]

please use the defined PR template Thota Sai Karthik (@karthik-0306)

Eduard van Valkenburg (@eavanvalkenburg) thanks for the note — updated the description to follow the template.

[github-actions]

Python Test Coverage Report •

File	Stmts	Miss	Cover	Missing
packages/foundry/agent_framework_foundry
_agent.py	262	40	84%	119, 122, 244–245, 249–251, 256–259, 359, 467–468, 492, 495, 506–518, 522–523, 721–722, 725, 751, 761, 783, 871, 876, 880
TOTAL	46057	4474	90%

Python Unit Test Overview

Tests	Skipped	Failures	Errors	Time
9540	34 💤	0 ❌	0 🔥	1m 58s ⏱️