Python: Fix core observability unsafe serialization of function-call arguments containing dataclass/framework objects

[Evan Mattson (moonbox3)]

Motivation and Context

Core observability in agent_framework/observability.py passed raw function_call arguments directly into json.dumps, which raises TypeError when those arguments contain dataclass or framework objects (e.g. HandoffAgentUserRequest) emitted by workflow request_info / handoff flows. This affected any Foundry-hosted or handoff scenario with sensitive telemetry enabled.

Fixes#5733

Description

The root cause was that _to_otel_part's function_call branch kept content.arguments as-is, with no safe-serialization step, mirroring the already-fixed AG-UI path. The fix lifts a make_json_safe() helper into agent_framework/_serialization.py — which recursively converts dataclasses, Pydantic models, datetimes, and __dict__-bearing objects to JSON-serializable primitives — and calls it on content.arguments before the value is included in the OTel span attribute. Tests covering dataclass payloads and nested non-primitive objects in function_call arguments were added to test_observability.py to prevent regression.

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

[Copilot]

Pull request overview

This PR fixes a runtime failure in Python core observability where OpenTelemetry span attributes could not be serialized when function_call arguments contained dataclasses or framework objects (common in workflow request_info / handoff flows), by introducing a safe-serialization step and adding regression tests.

Changes:

• Add make_json_safe() helper in core serialization utilities and apply it to function_call arguments in _to_otel_part.
• Add tests ensuring dataclass and nested-object tool arguments are serializable and _capture_messages doesn’t raise.
• Minor formatting/import cleanups in Foundry hosting and A2A tests, plus a lockfile metadata normalization.

Reviewed changes

Copilot reviewed 6 out of 7 changed files in this pull request and generated 2 comments.

Show a summary per file

File	Description
python/packages/core/agent_framework/observability.py	Serializes function_call arguments via make_json_safe before emitting OTel message parts.
python/packages/core/agent_framework/_serialization.py	Introduces make_json_safe() recursive conversion helper for JSON-serializable telemetry payloads.
python/packages/core/tests/core/test_observability.py	Adds regression tests for dataclass/nested objects in tool-call arguments and span capture.
python/packages/foundry_hosting/agent_framework_foundry_hosting/_responses.py	Consolidates imports (minor cleanup).
python/packages/foundry_hosting/tests/test_responses.py	Adds spacing to satisfy formatting/lint expectations around region boundaries.
python/packages/a2a/tests/test_a2a_agent.py	Formats a test function signature to a single line.
python/uv.lock	Normalizes dependency specifier ordering for github-copilot-sdk.

[Evan Mattson (moonbox3)]

Python Test Coverage Report •

File	Stmts	Miss	Cover	Missing
packages/core/agent_framework
_serialization.py	150	15	90%	517, 534, 544, 612, 637, 646–649, 651–654, 658, 661
packages/core/agent_framework/_workflows
_agent.py	363	64	82%	68, 76–82, 118–119, 212, 273, 286, 353, 364, 366, 426, 432, 449–450, 457, 459, 465, 531–532, 541, 580, 615, 690, 720, 737, 765, 776–779, 785, 791, 795–796, 799–805, 809–810, 818, 879, 886, 892–893, 904, 936, 943, 964, 973, 977, 979–981, 988
_functional.py	491	31	93%	398–399, 465, 488, 497–499, 502, 532, 713–714, 721, 969, 1001, 1004–1005, 1013–1015, 1090–1091, 1097, 1480–1481, 1483, 1490–1493, 1514, 1539
TOTAL	37429	4354	88%

Python Unit Test Overview

Tests	Skipped	Failures	Errors	Time
7458	34 💤	0 ❌	0 🔥	1m 58s ⏱️

[Evan Mattson (moonbox3)]

Automated Code Review

Reviewers: 2 | Confidence: 92% | Result: All clear

Reviewed: Security Reliability, Design Approach

---

Automated review by moonbox3's agents

[github-actions]

Automated Code Review

Reviewers: 4 | Confidence: 83% | Result: All clear

Reviewed: Correctness, Security Reliability, Test Coverage, Design Approach

---

Automated review by moonbox3's agents

[Evan Mattson (moonbox3)]

Automated Code Review

Reviewers: 4 | Confidence: 86%

✓ Correctness

The PR correctly fixes two bugs: (1) hasattr checks replaced with callable(getattr(..., None)) to skip non-callable attributes named model_dump/to_dict/dict, and (2) dict keys are now stringified so json.dumps never raises TypeError on non-string keys. The try/except TypeError wrappers provide a safe fallback chain. The __dict__ branch retains bare key references but vars(obj) always yields string keys, so no issue there. No correctness problems found.

✓ Security Reliability

The PR correctly fixes two bugs: (1) using callable(getattr(...)) to guard against non-callable attributes named model_dump/to_dict/dict, and (2) converting dict keys to strings so json.dumps never raises TypeError on non-string keys. Both fixes are well-motivated and tested. One reliability concern: the str(key) conversion at line 656 can silently clobber data when a dict contains both an integer key and its string equivalent (e.g., {42: 'a', '42': 'b'}), since iteration order determines which value survives with no warning. This is a silent data-loss scenario not covered by the new tests. The TypeError suppression around serialization method calls is reasonable in scope since TypeError on a zero-argument call is a narrow signal, but it could mask bugs in model_dump/to_dict implementations that raise TypeError internally — the fallback to __dict__ may then silently serialize incorrect/partial state.

✓ Test Coverage

The fix correctly guards against non-callable model_dump/to_dict/dict attributes and converts dict keys to strings. Two test coverage gaps: (1) the assertion in test_make_json_safe_non_callable_method_attribute is trivially true (json.dumps never returns None), masking whether the correct fallback value is returned; (2) the new try/except TypeError path — where a callable model_dump() raises TypeError and falls through — is never exercised by a test.

✓ Design Approach

I did not find a design-approach issue in this bounded pass. The change stays focused on making observability serialization tolerate the two concrete edge cases covered by the diff: non-callable serializer-like attributes and non-string dict keys that would otherwise break JSON encoding.

---

Automated review by moonbox3's agents