Python: Fix response_format resolution in streaming finalizer

[moonbox3]

Motivation and Context

When response_format is set in default_options (the typical workflow pattern), AgentResponse.value returns None during streaming, even though the text is correctly parsed. The non-streaming path works fine because it reads from the merged chat_options, but the streaming finalizer only checked the per-call options parameter.

Fixes #3970

Description

The streaming path in BaseAgent.run() used partial(self._finalize_response_updates, response_format=options.get("response_format")) to bind the finalizer at stream-creation time. Since options only contains per-call overrides (not default_options), response_format was None whenever it was set via default_options.

The fix replaces the partial with a _finalizer closure that defers the lookup to finalization time, reading response_format from ctx["chat_options"] — the merged options dict that includes both default_options and per-call overrides. This matches how the non-streaming path already resolves response_format.

A new test verifies that AgentResponse.value is correctly parsed when response_format is set in default_options and the workflow is streamed.

Contribution Checklist

• The code builds clean without any errors or warnings
• The PR follows the Contribution Guidelines
• All unit tests pass, and I have added new tests where possible
• Is this a breaking change? If yes, add "[BREAKING]" prefix to the title of the PR.

---

Note: PR autogenerated by moonbox3's agent

[moonbox3]

Automated Code Review

Reviewers: 3 | Confidence: 82%

✓ Correctness

This diff fixes a bug where response_format set in default_options was not picked up during streaming finalization, because the old partial-based finalizer only consulted per-call options. The new _finalizer closure defers the lookup to finalization time and checks ctx["chat_options"] first (which merges default and per-call options). The fix is logically sound and the new test validates the scenario. One minor robustness concern exists around the dict key access pattern.

✓ Security Reliability

This diff fixes response_format resolution during streaming by capturing the merged chat_options from the runtime context instead of only the per-call options. The change is small and mostly a bug fix. The only reliability concern is that ctx["chat_options"] uses bracket access, which will raise a KeyError if the context dict exists but lacks a chat_options key; using .get() would be more defensive. No injection, secrets, or resource-leak issues identified.

✓ Test Coverage

The diff fixes a bug where response_format from default_options was ignored during streaming because the finalizer only read from per-call options. A new test correctly verifies the fix by setting response_format in default_options and asserting the streamed response is parsed into a Pydantic model. Assertions are meaningful (checking result.value.greeting). However, the new _finalizer has a branching fallback (ctx path vs options path vs None) and only the default_options-via-ctx path is tested; the per-call-override and ctx-is-None paths lack dedicated streaming tests.

Suggestions

• In _finalizer, ctx["chat_options"] uses bracket access which will raise KeyError if chat_options is absent from ctx. Consider using ctx.get("chat_options", {}).get("response_format") for defensive access, unless the presence of chat_options in ctx is guaranteed by contract.
• Consider using ctx.get("chat_options", {}).get("response_format") instead of ctx["chat_options"].get("response_format") to avoid a potential KeyError if ctx is present but missing the chat_options key.
• Add a test where response_format is passed both in default_options and as a per-call option to verify that per-call options take precedence in the streaming finalizer.
• Add a test where response_format is passed only as a per-call kwarg (not in default_options) to ensure the fallback branch in _finalizer (when ctx is None or missing chat_options) still works correctly under the new code path.

---

Automated review by maf-dashboard agent

[markwallace-microsoft]

Python Test Coverage Report •

File	Stmts	Miss	Cover	Missing
packages/core/agent_framework
_agents.py	335	43	87%	425, 429, 481, 846, 882, 898, 977–980, 1041–1043, 1164, 1180, 1182, 1195, 1201, 1237, 1239, 1248–1253, 1258, 1260, 1266–1267, 1274, 1276–1277, 1285–1286, 1289–1291, 1299–1300, 1302, 1307, 1309
TOTAL	22178	2762	87%

Python Unit Test Overview

Tests	Skipped	Failures	Errors	Time
4674	247 💤	0 ❌	0 🔥	1m 18s ⏱️

[Copilot]

Pull request overview

This PR fixes a bug where AgentResponse.value is None when streaming a workflow with an agent that has response_format set in default_options. The issue occurred because the streaming path's finalizer was reading response_format from the raw options parameter instead of from the merged chat options that combine default_options with runtime options.

Changes:

• Fixed the streaming finalizer to access response_format from the merged chat_options in ctx_holder
• Added a test to validate that response_format from default_options is correctly parsed when streaming

Reviewed changes

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

File	Description
python/packages/core/agent_framework/_agents.py	Updated the streaming finalizer to read response_format from merged ctx["chat_options"] instead of raw options parameter
python/packages/core/tests/core/test_agents.py	Added test validating that AgentResponse.value is properly parsed when using response_format in default_options with streaming

[eavanvalkenburg]

thanks