Python: flaky integration tests blocking merge queue (Foundry Hosting + Foundry embeddings)

[moonbox3]

Two integration tests are flaking in the merge queue and blocking merges into main. Both are already marked @pytest.mark.flaky and have a 3-attempt retry policy, but all three attempts fail back-to-back. They have been temporarily skipped to unblock the queue (see follow-up PR) and need a real fix.

Observed failing on at least PRs #5301 and #5531 — these flakes are systemic, not PR-specific. Example failing run: https://github.com/microsoft/agent-framework/actions/runs/25083550113

Test 1 — test_temperature_and_max_tokens

• Path: packages/foundry_hosting/tests/test_responses_int.py::TestOptions::test_temperature_and_max_tokens
• Job: Python Tests - Foundry Hosting Integration

The test posts a /responses request with max_output_tokens=50 and asserts that the response contains exactly one message-type output:

output_messages = [o for o in body["output"] if o["type"] == "message"]
assert len(output_messages) == 1

It fails with assert 0 == 1 — i.e. the response status is completed (HTTP 200), but the output array contains no type == "message" entries. The captured server log shows output_count=1, so an output item exists; it's just not a message (likely a reasoning item only, with the message truncated by the 50-token cap). Failed all 3 attempts.

Suggested fix: raise max_output_tokens so the model has room to emit a user-visible message after any reasoning output, or relax the assertion to allow any output type while still verifying status/limits were honored.

Test 2 — test_text_embedding_live

• Path: packages/foundry/tests/foundry/test_foundry_embedding_client.py::TestFoundryEmbeddingIntegration::test_text_embedding_live
• Job: Python Integration Tests - Foundry

Calls FoundryEmbeddingClient().get_embeddings(["Hello, world!"]) against the live endpoint. Fails with:

json.decoder.JSONDecodeError: Expecting value: line 1 column 1 (char 0)

The Azure SDK got an empty response body from azure.ai.inference and tried to loads(''). Failed all 3 attempts. This looks like a transient endpoint issue or a missing-credential / 204 response that the client isn't surfacing as a clearer error.

Suggested fix: investigate why the embeddings endpoint returns an empty body in CI (auth? region? service-side flake?); also consider catching the empty-response case in the client and raising a typed error rather than letting json.loads blow up.

Why this also keeps biting unrelated PRs

Both tests are gated by the core path filter in .github/workflows/python-merge-tests.yml:

core:
  - 'python/packages/core/agent_framework/_*.py'
  - 'python/packages/core/agent_framework/_workflows/**'
  - 'python/packages/core/agent_framework/exceptions.py'
  - 'python/packages/core/agent_framework/observability.py'

Any change under _workflows/** triggers every provider's live integration suite via coreChanged == 'true', so a workflows-only PR ends up gated on Foundry / Foundry Hosting / OpenAI / etc. flakes. Worth a follow-up to narrow this filter (e.g. split workflows out of core) once the flakes themselves are fixed.

Action items

• Fix test_temperature_and_max_tokens (foundry hosting)
• Fix test_text_embedding_live (foundry embeddings)
• Re-enable both tests (remove @pytest.mark.skip)
• (Optional follow-up) Narrow the core path filter so workflows-only PRs don't fan out to every provider