Python: Fix AG-UI resume rejecting JSON for typed responses

[NekoPunch (orangeCatDeveloper)]

Motivation & Context

Resuming a Magentic plan-review interrupt over AG-UI fails for the payload an AG-UI client naturally sends — {"review": []}, the canonical "plan approved" answer:

RUN_ERROR  code=WORKFLOW_RESUME_INVALID_RESPONSE
Workflow resume for interruptId '...' does not match expected response type MagenticPlanReviewResponse.

Apps have to hand-build MagenticPlanReviewResponse to work around it. Magentic is only where this surfaced: any request_info whose response_type is a dataclass or model hits the same wall.

Description & Review Guide

• What are the major changes?

_coerce_response_for_request fell back to isinstance(candidate, target_type), which a JSON dict can never satisfy, so AG-UI rejected the resume before it reached core's try_coerce_to_type — the helper the direct-SDK path already uses to build these objects. That helper had a second defect: target_cls(**data) built the outer object and left nested values as raw dicts, and since _send_responses_internal validates the outer type only, MagenticPlanReviewResponse(review=[dict]) passed.

The coercion therefore lives in core, where every entrypoint benefits: it walks the declared annotations recursively, and a field that does not match its annotation fails the whole build rather than yielding a half-typed object. AG-UI keeps only what is protocol-specific — translating the AG-UI message wire shape into Message for fields declared as such, and turning a coercion failure into RUN_ERROR.

• What is the impact of these changes?

Field validation is stricter: Point(**{"x": 1, "y": "two"}) used to construct and pass, and is now rejected. Message shorthand is translated for first-level Message / list[Message] fields only; nested messages need the canonical contents form. The interrupt responseSchema for structured response types is still an opaque description, left for separate work.

Same resume request, before and after:

POST /workflow  {"resume": [{"interruptId": "...", "status": "resolved", "payload": {"review": []}}]}

before: data: {"type": "RUN_ERROR", "code": "WORKFLOW_RESUME_INVALID_RESPONSE", ...}
after:  data: {"type": "TEXT_MESSAGE_CONTENT", "delta": "FINAL"}
        data: {"type": "RUN_FINISHED", ...}

Nested coercion, which the direct-SDK path got wrong silently:

try_coerce_to_type({"review": [{"role": "user", "contents": [...]}]}, MagenticPlanReviewResponse)

before:  MagenticPlanReviewResponse(review=[dict])      # is_instance_of() said True
after:   MagenticPlanReviewResponse(review=[Message])

Related Issue

Fixes #7513

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

Note

Copilot was unable to run its full agentic suite in this review.

Improves runtime typing/coercion for workflow payloads so JSON-shaped responses can be converted into typed Python objects (dataclasses, containers, unions), and wires that into AG-UI response handling.

Changes:

• Enhanced try_coerce_to_type to handle Union, typed containers, dataclass field coercion/validation, Literal, and from_dict / model_validate.
• Added extensive unit tests around dataclass coercion, Literal, union selection, and container coercion.
• Updated AG-UI workflow response coercion to leverage the core coercion utilities and normalize AG-UI message wire formats by field annotation.

Reviewed changes

Copilot reviewed 5 out of 5 changed files in this pull request and generated 5 comments.

Show a summary per file

File	Description
python/packages/core/tests/workflow/test_typing_utils.py	Adds coverage for new coercion behaviors (dataclasses, containers, literals, unions, bool→float guard, from_dict types).
python/packages/core/agent_framework/_workflows/_typing_utils.py	Implements richer coercion logic and dataclass field-level validation/coercion.
python/packages/ag-ui/agent_framework_ag_ui/_workflow_run.py	Switches response coercion to core coercion utilities and adds message-field normalization based on annotations.
python/packages/ag-ui/tests/ag_ui/test_workflow_run.py	Adds AG-UI tests for dataclass/pydantic response coercion and field-annotation-driven message normalization.
python/packages/ag-ui/tests/ag_ui/test_endpoint.py	Adds endpoint-level regression test for resuming dataclass responses from JSON payloads.

---

💡 Add a code-review agent skill for context-aware, tailored reviews. Learn more in the docs.

[github-actions]

Python Test Coverage Report •

File	Stmts	Miss	Cover	Missing
packages/ag-ui/agent_framework_ag_ui
_workflow_run.py	807	91	88%	98, 112, 114, 116, 185–186, 190, 291–294, 338, 349, 354, 385, 421–424, 452, 457, 473, 475, 480–484, 494, 501, 513, 518, 525, 600, 610, 621, 626, 629, 652, 676, 679, 684, 687, 704–706, 711, 713, 719, 723, 725, 742, 748–749, 759, 765, 769–770, 794–795, 833, 841, 910, 930, 946, 961, 1045, 1067–1069, 1104–1118, 1150–1151, 1185–1186, 1261, 1323
packages/core/agent_framework/_workflows
_typing_utils.py	252	32	87%	22, 202, 226, 299, 301, 415, 484, 486–487, 496, 498, 505, 507, 527, 529, 531, 536–543, 546–547, 549–553, 555
TOTAL	47037	4366	90%

Python Unit Test Overview

Tests	Skipped	Failures	Errors	Time
9524	36 💤	0 ❌	0 🔥	2m 28s ⏱️

[Evan Mattson (moonbox3)]

Please also look at the failing CI/CD checks.

[Evan Mattson (moonbox3)]

Please have a look at some conflicts, thanks. NekoPunch (@orangeCatDeveloper)

[NekoPunch (orangeCatDeveloper)]

Please have a look at some conflicts, thanks. NekoPunch (NekoPunch (@orangeCatDeveloper))

rebased

[Evan Mattson (moonbox3)]

/review

[github-actions]

MAF Automated Review — Iteration 1

Result: Findings reported
Scope: full PR (4 commit(s)): 534af1f093e0, 32230375d11b, 5aa5b5a2511c, 7025872aec80
Model: gpt-5.6-sol

Overview

The PR centralizes recursive typed-response coercion in core and keeps AG-UI-specific Message normalization at the protocol boundary, with strong coverage for nested dataclasses, common containers, constructor failures, and invalid resume handling. Two gaps remain: the new dataclass field discovery rejects valid InitVar constructor arguments, and the new frozenset conversion can accept members that violate the declared element type.

Reviewed the supplied pull-request change set across correctness, security/reliability, architecture, and failure behavior.
2 verified findings remained after source verification (2 medium) across 1 file. Details are attached to the affected lines below.

Affected areas: python/packages/core/agent_framework/_workflows/_typing_utils.py

[NekoPunch (orangeCatDeveloper)]

Rebased to resolve conflict