chore: import upstream snapshot with attribution

2026-07-13 13:39:25 +08:00
commit db620d33df
5151 changed files with 925932 additions and 0 deletions
@@ -0,0 +1 @@
+# Copyright (c) Microsoft. All rights reserved.
@@ -0,0 +1,13 @@
+# Copyright (c) Microsoft. All rights reserved.
+
+"""Conftest for golden tests — ensures parent test dir is importable."""
+
+import sys
+from pathlib import Path
+
+
+def pytest_configure() -> None:
+    """Ensure parent test directory is on sys.path for helper module imports."""
+    parent_test_dir = str(Path(__file__).resolve().parent.parent)
+    if parent_test_dir not in sys.path:
+        sys.path.insert(0, parent_test_dir)
@@ -0,0 +1,140 @@
+# Copyright (c) Microsoft. All rights reserved.
+
+"""Golden event-stream tests for the basic agentic chat scenario."""
+
+from __future__ import annotations
+
+from typing import Any
+
+from agent_framework import AgentResponseUpdate, Content
+from conftest import StubAgent  # pyrefly: ignore[missing-import] # pyright: ignore[reportMissingImports]
+from event_stream import EventStream  # pyrefly: ignore[missing-import] # pyright: ignore[reportMissingImports]
+
+from agent_framework_ag_ui import AgentFrameworkAgent
+
+
+def _build_agent(updates: list[AgentResponseUpdate], **kwargs: Any) -> AgentFrameworkAgent:
+    stub = StubAgent(updates=updates)
+    return AgentFrameworkAgent(agent=stub, **kwargs)
+
+
+async def _run(agent: AgentFrameworkAgent, payload: dict[str, Any]) -> EventStream:
+    return EventStream([event async for event in agent.run(payload)])
+
+
+BASIC_PAYLOAD: dict[str, Any] = {
+    "thread_id": "thread-chat",
+    "run_id": "run-chat",
+    "messages": [{"role": "user", "content": "Hello"}],
+}
+
+
+def _text_update(text: str) -> AgentResponseUpdate:
+    return AgentResponseUpdate(contents=[Content.from_text(text=text)], role="assistant")
+
+
+def _snapshot_role(msg: Any) -> str:
+    """Extract role string from a snapshot message (Pydantic model or dict)."""
+    role = getattr(msg, "role", None) or (msg.get("role") if isinstance(msg, dict) else None)
+    if role is None:
+        return ""
+    return str(getattr(role, "value", role))
+
+
+def _snapshot_content(msg: Any) -> str:
+    """Extract content string from a snapshot message."""
+    content = getattr(msg, "content", None) or (msg.get("content") if isinstance(msg, dict) else "")
+    return str(content) if content else ""
+
+
+# ── Golden stream tests ──
+
+
+async def test_basic_chat_golden_event_sequence() -> None:
+    """Assert the exact event type sequence for a single text response."""
+    agent = _build_agent([_text_update("Hi there!")])
+    stream = await _run(agent, BASIC_PAYLOAD)
+
+    stream.assert_strict_types(
+        [
+            "RUN_STARTED",
+            "TEXT_MESSAGE_START",
+            "TEXT_MESSAGE_CONTENT",
+            "TEXT_MESSAGE_END",
+            "MESSAGES_SNAPSHOT",
+            "RUN_FINISHED",
+        ]
+    )
+
+
+async def test_basic_chat_bookends() -> None:
+    """RUN_STARTED is first, RUN_FINISHED is last."""
+    agent = _build_agent([_text_update("reply")])
+    stream = await _run(agent, BASIC_PAYLOAD)
+    stream.assert_bookends()
+
+
+async def test_basic_chat_text_messages_balanced() -> None:
+    """Every TEXT_MESSAGE_START has a matching TEXT_MESSAGE_END."""
+    agent = _build_agent([_text_update("reply")])
+    stream = await _run(agent, BASIC_PAYLOAD)
+    stream.assert_text_messages_balanced()
+
+
+async def test_basic_chat_no_errors() -> None:
+    """No RUN_ERROR events in a normal flow."""
+    agent = _build_agent([_text_update("reply")])
+    stream = await _run(agent, BASIC_PAYLOAD)
+    stream.assert_no_run_error()
+
+
+async def test_basic_chat_message_id_consistency() -> None:
+    """All text events reference the same message_id."""
+    agent = _build_agent([_text_update("reply")])
+    stream = await _run(agent, BASIC_PAYLOAD)
+
+    start = stream.first("TEXT_MESSAGE_START")
+    content = stream.first("TEXT_MESSAGE_CONTENT")
+    end = stream.first("TEXT_MESSAGE_END")
+    assert start.message_id == content.message_id == end.message_id
+
+
+async def test_multi_chunk_text_golden_sequence() -> None:
+    """Streaming multiple chunks produces START + multiple CONTENT + END."""
+    agent = _build_agent([_text_update("Hello "), _text_update("world!")])
+    stream = await _run(agent, BASIC_PAYLOAD)
+
+    stream.assert_strict_types(
+        [
+            "RUN_STARTED",
+            "TEXT_MESSAGE_START",
+            "TEXT_MESSAGE_CONTENT",
+            "TEXT_MESSAGE_CONTENT",
+            "TEXT_MESSAGE_END",
+            "MESSAGES_SNAPSHOT",
+            "RUN_FINISHED",
+        ]
+    )
+    stream.assert_text_messages_balanced()
+    stream.assert_message_ids_consistent()
+
+
+async def test_messages_snapshot_contains_assistant_reply() -> None:
+    """MessagesSnapshotEvent includes the assistant's accumulated text."""
+    agent = _build_agent([_text_update("Hello there")])
+    stream = await _run(agent, BASIC_PAYLOAD)
+
+    snapshot = stream.messages_snapshot()
+    assistant_msgs = [m for m in snapshot if _snapshot_role(m) == "assistant"]
+    assert assistant_msgs, "No assistant message in snapshot"
+    assert any("Hello there" in _snapshot_content(m) for m in assistant_msgs)
+
+
+async def test_empty_messages_produces_start_and_finish() -> None:
+    """Empty message list still produces RUN_STARTED and RUN_FINISHED."""
+    agent = _build_agent([_text_update("reply")])
+    payload = {"thread_id": "t1", "run_id": "r1", "messages": []}
+    stream = await _run(agent, payload)
+
+    stream.assert_bookends()
+    assert "TEXT_MESSAGE_START" not in stream.types()
@@ -0,0 +1,236 @@
+# Copyright (c) Microsoft. All rights reserved.
+
+"""Golden event-stream tests for the backend (server-side) tools scenario."""
+
+from __future__ import annotations
+
+from typing import Any
+
+from agent_framework import AgentResponseUpdate, Content
+from conftest import StubAgent  # pyrefly: ignore[missing-import] # pyright: ignore[reportMissingImports]
+from event_stream import EventStream  # pyrefly: ignore[missing-import] # pyright: ignore[reportMissingImports]
+
+from agent_framework_ag_ui import AgentFrameworkAgent
+
+
+def _build_agent(updates: list[AgentResponseUpdate], **kwargs: Any) -> AgentFrameworkAgent:
+    stub = StubAgent(updates=updates)
+    return AgentFrameworkAgent(agent=stub, **kwargs)
+
+
+async def _run(agent: AgentFrameworkAgent, payload: dict[str, Any]) -> EventStream:
+    return EventStream([event async for event in agent.run(payload)])
+
+
+PAYLOAD: dict[str, Any] = {
+    "thread_id": "thread-tools",
+    "run_id": "run-tools",
+    "messages": [{"role": "user", "content": "What's the weather?"}],
+}
+
+
+# ── Golden stream tests ──
+
+
+async def test_tool_call_lifecycle_golden_sequence() -> None:
+    """Assert the full event sequence for a tool call → result → text response."""
+    updates = [
+        # LLM calls the tool
+        AgentResponseUpdate(
+            contents=[Content.from_function_call(name="get_weather", call_id="call-1", arguments='{"city": "SF"}')],
+            role="assistant",
+        ),
+        # Tool result comes back
+        AgentResponseUpdate(
+            contents=[Content.from_function_result(call_id="call-1", result="72°F and sunny")],
+            role="assistant",
+        ),
+        # LLM responds with text
+        AgentResponseUpdate(
+            contents=[Content.from_text(text="It's 72°F and sunny in SF!")],
+            role="assistant",
+        ),
+    ]
+    agent = _build_agent(updates)
+    stream = await _run(agent, PAYLOAD)
+
+    stream.assert_ordered_types(
+        [
+            "RUN_STARTED",
+            "TEXT_MESSAGE_START",  # Synthetic start for tool-only message
+            "TOOL_CALL_START",
+            "TOOL_CALL_ARGS",
+            "TOOL_CALL_END",
+            "TOOL_CALL_RESULT",
+            "TEXT_MESSAGE_END",  # End of synthetic message
+            "TEXT_MESSAGE_START",  # New message for text response
+            "TEXT_MESSAGE_CONTENT",
+            "TEXT_MESSAGE_END",
+            "MESSAGES_SNAPSHOT",
+            "RUN_FINISHED",
+        ]
+    )
+
+
+async def test_tool_calls_balanced() -> None:
+    """Every TOOL_CALL_START has a matching TOOL_CALL_END."""
+    updates = [
+        AgentResponseUpdate(
+            contents=[Content.from_function_call(name="get_weather", call_id="call-1", arguments='{"city": "SF"}')],
+            role="assistant",
+        ),
+        AgentResponseUpdate(
+            contents=[Content.from_function_result(call_id="call-1", result="72°F")],
+            role="assistant",
+        ),
+        AgentResponseUpdate(
+            contents=[Content.from_text(text="It's 72°F!")],
+            role="assistant",
+        ),
+    ]
+    agent = _build_agent(updates)
+    stream = await _run(agent, PAYLOAD)
+
+    stream.assert_tool_calls_balanced()
+
+
+async def test_text_messages_balanced_with_tools() -> None:
+    """Text messages are properly balanced even around tool calls."""
+    updates = [
+        AgentResponseUpdate(
+            contents=[Content.from_function_call(name="get_weather", call_id="call-1", arguments='{"city": "SF"}')],
+            role="assistant",
+        ),
+        AgentResponseUpdate(
+            contents=[Content.from_function_result(call_id="call-1", result="72°F")],
+            role="assistant",
+        ),
+        AgentResponseUpdate(
+            contents=[Content.from_text(text="It's 72°F!")],
+            role="assistant",
+        ),
+    ]
+    agent = _build_agent(updates)
+    stream = await _run(agent, PAYLOAD)
+
+    stream.assert_text_messages_balanced()
+
+
+async def test_tool_call_id_matches_result() -> None:
+    """TOOL_CALL_START and TOOL_CALL_RESULT reference the same tool_call_id."""
+    updates = [
+        AgentResponseUpdate(
+            contents=[Content.from_function_call(name="get_weather", call_id="call-1", arguments="{}")],
+            role="assistant",
+        ),
+        AgentResponseUpdate(
+            contents=[Content.from_function_result(call_id="call-1", result="72°F")],
+            role="assistant",
+        ),
+    ]
+    agent = _build_agent(updates)
+    stream = await _run(agent, PAYLOAD)
+
+    start = stream.first("TOOL_CALL_START")
+    result = stream.first("TOOL_CALL_RESULT")
+    assert start.tool_call_id == result.tool_call_id == "call-1"
+
+
+async def test_tool_result_content_preserved() -> None:
+    """TOOL_CALL_RESULT event carries the tool's result content."""
+    updates = [
+        AgentResponseUpdate(
+            contents=[Content.from_function_call(name="get_weather", call_id="call-1", arguments="{}")],
+            role="assistant",
+        ),
+        AgentResponseUpdate(
+            contents=[Content.from_function_result(call_id="call-1", result="72°F and sunny")],
+            role="assistant",
+        ),
+    ]
+    agent = _build_agent(updates)
+    stream = await _run(agent, PAYLOAD)
+
+    result = stream.first("TOOL_CALL_RESULT")
+    assert result.content == "72°F and sunny"
+
+
+async def test_no_run_error_on_tool_flow() -> None:
+    """Tool call flow doesn't produce RUN_ERROR."""
+    updates = [
+        AgentResponseUpdate(
+            contents=[Content.from_function_call(name="get_weather", call_id="call-1", arguments="{}")],
+            role="assistant",
+        ),
+        AgentResponseUpdate(
+            contents=[Content.from_function_result(call_id="call-1", result="72°F")],
+            role="assistant",
+        ),
+    ]
+    agent = _build_agent(updates)
+    stream = await _run(agent, PAYLOAD)
+
+    stream.assert_no_run_error()
+    stream.assert_bookends()
+
+
+async def test_multiple_sequential_tool_calls() -> None:
+    """Multiple sequential tool calls each produce balanced START/END pairs."""
+    updates = [
+        AgentResponseUpdate(
+            contents=[Content.from_function_call(name="tool_a", call_id="call-a", arguments="{}")],
+            role="assistant",
+        ),
+        AgentResponseUpdate(
+            contents=[Content.from_function_result(call_id="call-a", result="result-a")],
+            role="assistant",
+        ),
+        AgentResponseUpdate(
+            contents=[Content.from_function_call(name="tool_b", call_id="call-b", arguments="{}")],
+            role="assistant",
+        ),
+        AgentResponseUpdate(
+            contents=[Content.from_function_result(call_id="call-b", result="result-b")],
+            role="assistant",
+        ),
+        AgentResponseUpdate(
+            contents=[Content.from_text(text="Done!")],
+            role="assistant",
+        ),
+    ]
+    agent = _build_agent(updates)
+    stream = await _run(agent, PAYLOAD)
+
+    stream.assert_tool_calls_balanced()
+    stream.assert_text_messages_balanced()
+    stream.assert_bookends()
+
+    # Both tool calls should appear
+    starts = stream.get("TOOL_CALL_START")
+    assert len(starts) == 2
+    assert {s.tool_call_name for s in starts} == {"tool_a", "tool_b"}
+
+
+async def test_messages_snapshot_includes_tool_calls() -> None:
+    """MessagesSnapshotEvent includes tool call and result messages."""
+    updates = [
+        AgentResponseUpdate(
+            contents=[Content.from_function_call(name="get_weather", call_id="call-1", arguments='{"city":"SF"}')],
+            role="assistant",
+        ),
+        AgentResponseUpdate(
+            contents=[Content.from_function_result(call_id="call-1", result="72°F")],
+            role="assistant",
+        ),
+        AgentResponseUpdate(
+            contents=[Content.from_text(text="It's warm!")],
+            role="assistant",
+        ),
+    ]
+    agent = _build_agent(updates)
+    stream = await _run(agent, PAYLOAD)
+
+    stream.assert_has_type("MESSAGES_SNAPSHOT")
+    snapshot = stream.messages_snapshot()
+    # Should have: user message, assistant with tool_calls, tool result, assistant text
+    assert len(snapshot) >= 3
@@ -0,0 +1,364 @@
+# Copyright (c) Microsoft. All rights reserved.
+
+"""Golden event-stream tests for the deterministic tool-driven state scenario.
+
+Covers issue https://github.com/microsoft/agent-framework/issues/3167 — a tool
+returning :func:`agent_framework_ag_ui.state_update` must push a deterministic
+``StateSnapshotEvent`` derived from its actual return value, orthogonal to the
+optimistic ``predict_state_config`` path. These golden tests pin the user-visible
+event stream so additive changes cannot silently regress it.
+"""
+
+from __future__ import annotations
+
+from typing import Any
+
+from agent_framework import AgentResponseUpdate, Content
+from conftest import StubAgent  # pyrefly: ignore[missing-import] # pyright: ignore[reportMissingImports]
+from event_stream import EventStream  # pyrefly: ignore[missing-import] # pyright: ignore[reportMissingImports]
+
+from agent_framework_ag_ui import AgentFrameworkAgent, state_update
+
+STATE_SCHEMA = {
+    "weather": {"type": "object", "description": "Last fetched weather"},
+}
+
+
+def _build_agent(updates: list[AgentResponseUpdate], **kwargs: Any) -> AgentFrameworkAgent:
+    stub = StubAgent(updates=updates)
+    kwargs.setdefault("state_schema", STATE_SCHEMA)
+    return AgentFrameworkAgent(agent=stub, **kwargs)
+
+
+async def _run(agent: AgentFrameworkAgent, payload: dict[str, Any]) -> EventStream:
+    return EventStream([event async for event in agent.run(payload)])
+
+
+PAYLOAD: dict[str, Any] = {
+    "thread_id": "thread-det-state",
+    "run_id": "run-det-state",
+    "messages": [{"role": "user", "content": "What's the weather in SF?"}],
+    "state": {"weather": {}},
+}
+
+
+def _tool_call(call_id: str, name: str, arguments: str) -> AgentResponseUpdate:
+    return AgentResponseUpdate(
+        contents=[Content.from_function_call(name=name, call_id=call_id, arguments=arguments)],
+        role="assistant",
+    )
+
+
+def _tool_result_with_state(call_id: str, text: str, state: dict[str, Any]) -> AgentResponseUpdate:
+    """Build a function_result update whose inner item carries a state marker.
+
+    This mirrors what the core framework produces when a real ``@tool`` returns
+    :func:`state_update`: ``parse_result`` keeps the ``Content`` as-is, and
+    ``Content.from_function_result`` preserves its ``additional_properties``
+    inside ``items``.
+    """
+    return AgentResponseUpdate(
+        contents=[
+            Content.from_function_result(
+                call_id=call_id,
+                result=[state_update(text=text, state=state)],
+            )
+        ],
+        role="assistant",
+    )
+
+
+def _tool_result_with_display(call_id: str, text: str, tool_result: Any, **kwargs: Any) -> AgentResponseUpdate:
+    """Build a function_result update carrying an optional UI display marker."""
+    return AgentResponseUpdate(
+        contents=[
+            Content.from_function_result(
+                call_id=call_id,
+                result=[state_update(text=text, tool_result=tool_result, **kwargs)],
+            )
+        ],
+        role="assistant",
+    )
+
+
+# ── Golden stream tests ──
+
+
+async def test_deterministic_state_emits_snapshot_after_tool_result() -> None:
+    """The happy path: STATE_SNAPSHOT follows TOOL_CALL_RESULT in order."""
+    updates = [
+        _tool_call("call-1", "get_weather", '{"city": "SF"}'),
+        _tool_result_with_state(
+            "call-1",
+            text="Weather in SF: 14°C foggy",
+            state={"weather": {"city": "SF", "temp": 14, "conditions": "foggy"}},
+        ),
+        AgentResponseUpdate(
+            contents=[Content.from_text(text="It's 14°C and foggy in SF.")],
+            role="assistant",
+        ),
+    ]
+    agent = _build_agent(updates)
+    stream = await _run(agent, PAYLOAD)
+
+    stream.assert_bookends()
+    stream.assert_no_run_error()
+    stream.assert_tool_calls_balanced()
+    stream.assert_text_messages_balanced()
+
+    # Ordered subsequence: the deterministic STATE_SNAPSHOT must follow the
+    # TOOL_CALL_RESULT. This is the central contract for #3167.
+    stream.assert_ordered_types(
+        [
+            "RUN_STARTED",
+            "TOOL_CALL_START",
+            "TOOL_CALL_ARGS",
+            "TOOL_CALL_END",
+            "TOOL_CALL_RESULT",
+            "STATE_SNAPSHOT",
+            "RUN_FINISHED",
+        ]
+    )
+
+    # The final STATE_SNAPSHOT must carry the tool-driven state.
+    snapshot = stream.snapshot()
+    assert snapshot["weather"] == {"city": "SF", "temp": 14, "conditions": "foggy"}
+
+
+async def test_deterministic_state_does_not_fire_for_plain_tool_result() -> None:
+    """Regression guard: tools returning plain strings must NOT emit a new STATE_SNAPSHOT.
+
+    The initial STATE_SNAPSHOT fires once from the schema + initial payload
+    state. A plain (non-state_update) tool result must not add another one.
+    """
+    updates = [
+        _tool_call("call-1", "get_weather", '{"city": "SF"}'),
+        AgentResponseUpdate(
+            contents=[Content.from_function_result(call_id="call-1", result="14°C foggy")],
+            role="assistant",
+        ),
+        AgentResponseUpdate(
+            contents=[Content.from_text(text="It's 14°C and foggy.")],
+            role="assistant",
+        ),
+    ]
+    agent = _build_agent(updates)
+    stream = await _run(agent, PAYLOAD)
+
+    stream.assert_bookends()
+    stream.assert_no_run_error()
+
+    snapshots = stream.get("STATE_SNAPSHOT")
+    # Only the initial snapshot (from state_schema + payload state) should exist.
+    # No deterministic snapshot should have been added by the plain tool result.
+    assert len(snapshots) == 1, (
+        f"Expected exactly 1 STATE_SNAPSHOT (initial only) for plain tool result; "
+        f"got {len(snapshots)}. Snapshots: {[s.snapshot for s in snapshots]}"
+    )
+
+
+async def test_deterministic_state_merges_into_initial_state() -> None:
+    """The tool-driven snapshot must merge into, not replace, pre-existing state keys."""
+    payload = dict(PAYLOAD)
+    payload["state"] = {"weather": {}, "user_preferences": {"unit": "C"}}
+
+    updates = [
+        _tool_call("call-1", "get_weather", '{"city": "SF"}'),
+        _tool_result_with_state(
+            "call-1",
+            text="Weather: 14°C",
+            state={"weather": {"city": "SF", "temp": 14}},
+        ),
+    ]
+    agent = _build_agent(updates, state_schema={**STATE_SCHEMA, "user_preferences": {"type": "object"}})
+    stream = await _run(agent, payload)
+
+    stream.assert_bookends()
+    stream.assert_no_run_error()
+
+    final_snapshot = stream.snapshot()
+    assert final_snapshot["weather"] == {"city": "SF", "temp": 14}
+    assert final_snapshot["user_preferences"] == {"unit": "C"}, (
+        "Pre-existing state keys must survive the deterministic merge"
+    )
+
+
+async def test_deterministic_state_llm_visible_text_is_clean() -> None:
+    """The LLM-visible TOOL_CALL_RESULT content must not leak the state marker key."""
+    updates = [
+        _tool_call("call-1", "get_weather", '{"city": "SF"}'),
+        _tool_result_with_state(
+            "call-1",
+            text="Weather in SF: 14°C foggy",
+            state={"weather": {"city": "SF", "temp": 14}},
+        ),
+    ]
+    agent = _build_agent(updates)
+    stream = await _run(agent, PAYLOAD)
+
+    result = stream.first("TOOL_CALL_RESULT")
+    assert result.content == "Weather in SF: 14°C foggy"
+    # The marker key must never appear in the content sent back to the LLM.
+    assert "__ag_ui_tool_result_state__" not in result.content
+    assert "weather" not in result.content  # not as a raw state dump
+
+
+async def test_deterministic_state_multiple_tools_merge_in_order() -> None:
+    """Two state-updating tools in one run merge in order; later wins on key collisions."""
+    updates = [
+        _tool_call("call-a", "get_weather", '{"city": "SF"}'),
+        _tool_result_with_state(
+            "call-a",
+            text="First result",
+            state={"weather": {"city": "SF", "temp": 14}, "source": "primary"},
+        ),
+        _tool_call("call-b", "get_weather_refined", '{"city": "SF"}'),
+        _tool_result_with_state(
+            "call-b",
+            text="Refined result",
+            state={"source": "refined"},
+        ),
+        AgentResponseUpdate(
+            contents=[Content.from_text(text="Here you go.")],
+            role="assistant",
+        ),
+    ]
+    agent = _build_agent(
+        updates,
+        state_schema={**STATE_SCHEMA, "source": {"type": "string"}},
+    )
+    stream = await _run(agent, PAYLOAD)
+
+    stream.assert_bookends()
+    stream.assert_tool_calls_balanced()
+    stream.assert_no_run_error()
+
+    # Two tool-driven snapshots emitted (one per tool) plus the initial snapshot.
+    snapshots = stream.get("STATE_SNAPSHOT")
+    assert len(snapshots) >= 2, f"Expected at least 2 STATE_SNAPSHOTs; got {len(snapshots)}"
+
+    final = stream.snapshot()
+    assert final["weather"] == {"city": "SF", "temp": 14}
+    # Later tool must override earlier tool on the shared key.
+    assert final["source"] == "refined"
+
+
+async def test_deterministic_state_coexists_with_predict_state_config() -> None:
+    """Predictive state and deterministic state must coexist without clobbering each other."""
+    predict_config = {
+        "draft": {
+            "tool": "write_draft",
+            "tool_argument": "body",
+        }
+    }
+    updates = [
+        # Predictive tool: its argument "body" populates state.draft optimistically.
+        _tool_call("call-1", "write_draft", '{"body": "Hello world"}'),
+        # Then a deterministic tool result landing a different key.
+        _tool_result_with_state(
+            "call-1",
+            text="Draft saved",
+            state={"weather": {"city": "SF", "temp": 14}},
+        ),
+    ]
+    agent = _build_agent(
+        updates,
+        state_schema={**STATE_SCHEMA, "draft": {"type": "string"}},
+        predict_state_config=predict_config,
+        require_confirmation=False,
+    )
+    payload = dict(PAYLOAD)
+    payload["state"] = {"weather": {}, "draft": ""}
+    stream = await _run(agent, payload)
+
+    stream.assert_bookends()
+    stream.assert_no_run_error()
+    stream.assert_tool_calls_balanced()
+
+    # The final observed state must contain both the deterministic and predictive contributions.
+    final = stream.snapshot()
+    assert final["weather"] == {"city": "SF", "temp": 14}, f"Deterministic state missing from final snapshot: {final}"
+
+
+async def test_tool_result_display_payload_reaches_ui_event_only() -> None:
+    """Rich display payload overrides TOOL_CALL_RESULT without leaking marker keys."""
+    updates = [
+        _tool_call("call-1", "get_weather", '{"city": "SF"}'),
+        _tool_result_with_display(
+            "call-1",
+            text="Weather in SF: 14°C foggy",
+            tool_result={"city": "SF", "temp": 14, "conditions": "foggy"},
+        ),
+    ]
+    agent = _build_agent(updates)
+    stream = await _run(agent, PAYLOAD)
+
+    stream.assert_bookends()
+    stream.assert_no_run_error()
+    stream.assert_tool_calls_balanced()
+
+    result = stream.first("TOOL_CALL_RESULT")
+    assert result.content == '{"city": "SF", "temp": 14, "conditions": "foggy"}'
+    assert "__ag_ui_tool_result_display__" not in result.content
+    assert "__ag_ui_tool_result_state__" not in result.content
+
+
+async def test_tool_result_display_falls_back_to_text_when_unset() -> None:
+    """Without a display marker, the UI event keeps the existing text content."""
+    updates = [
+        _tool_call("call-1", "get_weather", '{"city": "SF"}'),
+        _tool_result_with_state(
+            "call-1",
+            text="Weather in SF: 14°C foggy",
+            state={"weather": {"city": "SF", "temp": 14}},
+        ),
+    ]
+    agent = _build_agent(updates)
+    stream = await _run(agent, PAYLOAD)
+
+    stream.assert_bookends()
+    stream.assert_no_run_error()
+    stream.assert_tool_calls_balanced()
+
+    result = stream.first("TOOL_CALL_RESULT")
+    assert result.content == "Weather in SF: 14°C foggy"
+    assert "__ag_ui_tool_result_display__" not in result.content
+    assert "__ag_ui_tool_result_state__" not in result.content
+
+
+async def test_tool_result_display_coexists_with_state_snapshot() -> None:
+    """Display and durable state markers produce one deterministic state snapshot."""
+    updates = [
+        _tool_call("call-1", "get_weather", '{"city": "SF"}'),
+        _tool_result_with_display(
+            "call-1",
+            text="Weather in SF: 14°C foggy",
+            tool_result={"city": "SF", "temp": 14, "conditions": "foggy"},
+            state={"weather": {"city": "SF", "temp": 14, "conditions": "foggy"}},
+        ),
+    ]
+    agent = _build_agent(updates)
+    stream = await _run(agent, PAYLOAD)
+
+    stream.assert_bookends()
+    stream.assert_no_run_error()
+    stream.assert_tool_calls_balanced()
+    stream.assert_ordered_types(["TOOL_CALL_RESULT", "STATE_SNAPSHOT", "RUN_FINISHED"])
+
+    result = stream.first("TOOL_CALL_RESULT")
+    assert result.content == '{"city": "SF", "temp": 14, "conditions": "foggy"}'
+
+    result_idx = stream.events.index(result)
+    deterministic_snapshots = [
+        event
+        for event in stream.events[result_idx + 1 :]
+        if getattr(getattr(event, "type", None), "value", getattr(event, "type", None)) == "STATE_SNAPSHOT"
+    ]
+    assert len(deterministic_snapshots) == 1
+    assert deterministic_snapshots[0].snapshot["weather"] == {
+        "city": "SF",
+        "temp": 14,
+        "conditions": "foggy",
+    }
+    assert "__ag_ui_tool_result_display__" not in str(deterministic_snapshots[0].snapshot)
+    assert "__ag_ui_tool_result_state__" not in str(deterministic_snapshots[0].snapshot)
@@ -0,0 +1,90 @@
+# Copyright (c) Microsoft. All rights reserved.
+
+"""Golden event-stream tests for the generative UI (workflow-as-agent) scenario."""
+
+from __future__ import annotations
+
+from typing import Any
+
+from agent_framework import WorkflowBuilder, WorkflowContext, executor
+from event_stream import EventStream  # pyrefly: ignore[missing-import] # pyright: ignore[reportMissingImports]
+
+from agent_framework_ag_ui import AgentFrameworkWorkflow
+
+
+async def _run(wrapper: AgentFrameworkWorkflow, payload: dict[str, Any]) -> EventStream:
+    return EventStream([event async for event in wrapper.run(payload)])
+
+
+PAYLOAD: dict[str, Any] = {
+    "thread_id": "thread-gen-ui-agent",
+    "run_id": "run-gen-ui-agent",
+    "messages": [{"role": "user", "content": "Generate a UI"}],
+}
+
+
+# ── Golden stream tests ──
+
+
+async def test_workflow_agent_golden_sequence() -> None:
+    """Workflow-as-agent: emits step events and text content."""
+
+    @executor(id="generator")
+    async def generator(message: Any, ctx: WorkflowContext[Any, str]) -> None:
+        await ctx.yield_output("Here is your generated UI content!")
+
+    workflow = WorkflowBuilder(start_executor=generator).build()
+    wrapper = AgentFrameworkWorkflow(workflow=workflow)
+    stream = await _run(wrapper, PAYLOAD)
+
+    stream.assert_bookends()
+    stream.assert_no_run_error()
+    stream.assert_text_messages_balanced()
+
+    # Should have step events for the executor
+    stream.assert_has_type("STEP_STARTED")
+    stream.assert_has_type("STEP_FINISHED")
+
+    # Should have text message content
+    stream.assert_has_type("TEXT_MESSAGE_CONTENT")
+
+
+async def test_workflow_agent_step_names_match() -> None:
+    """Step started/finished events reference the executor name."""
+
+    @executor(id="my_executor")
+    async def my_executor(message: Any, ctx: WorkflowContext[Any, str]) -> None:
+        await ctx.yield_output("Done!")
+
+    workflow = WorkflowBuilder(start_executor=my_executor).build()
+    wrapper = AgentFrameworkWorkflow(workflow=workflow)
+    stream = await _run(wrapper, PAYLOAD)
+
+    started = [e for e in stream.get("STEP_STARTED") if getattr(e, "step_name", "") == "my_executor"]
+    finished = [e for e in stream.get("STEP_FINISHED") if getattr(e, "step_name", "") == "my_executor"]
+    assert started, "Expected STEP_STARTED for 'my_executor'"
+    assert finished, "Expected STEP_FINISHED for 'my_executor'"
+
+
+async def test_workflow_agent_ordered_events() -> None:
+    """Workflow events follow expected ordering: RUN_STARTED → STEP_STARTED → content → STEP_FINISHED → RUN_FINISHED."""
+
+    @executor(id="my_step")
+    async def my_step(message: Any, ctx: WorkflowContext[Any, str]) -> None:
+        await ctx.yield_output("Generated content")
+
+    workflow = WorkflowBuilder(start_executor=my_step).build()
+    wrapper = AgentFrameworkWorkflow(workflow=workflow)
+    stream = await _run(wrapper, PAYLOAD)
+
+    stream.assert_ordered_types(
+        [
+            "RUN_STARTED",
+            "STEP_STARTED",
+            "TEXT_MESSAGE_START",
+            "TEXT_MESSAGE_CONTENT",
+            "STEP_FINISHED",
+            "TEXT_MESSAGE_END",
+            "RUN_FINISHED",
+        ]
+    )
@@ -0,0 +1,135 @@
+# Copyright (c) Microsoft. All rights reserved.
+
+"""Golden event-stream tests for the client-side (declaration-only) tools scenario."""
+
+from __future__ import annotations
+
+from typing import Any
+
+from agent_framework import AgentResponseUpdate, Content
+from conftest import StubAgent  # pyrefly: ignore[missing-import] # pyright: ignore[reportMissingImports]
+from event_stream import EventStream  # pyrefly: ignore[missing-import] # pyright: ignore[reportMissingImports]
+
+from agent_framework_ag_ui import AgentFrameworkAgent
+
+
+def _build_agent(updates: list[AgentResponseUpdate], **kwargs: Any) -> AgentFrameworkAgent:
+    stub = StubAgent(updates=updates)
+    return AgentFrameworkAgent(agent=stub, **kwargs)
+
+
+async def _run(agent: AgentFrameworkAgent, payload: dict[str, Any]) -> EventStream:
+    return EventStream([event async for event in agent.run(payload)])
+
+
+PAYLOAD: dict[str, Any] = {
+    "thread_id": "thread-gen-ui-tool",
+    "run_id": "run-gen-ui-tool",
+    "messages": [{"role": "user", "content": "Show me a chart"}],
+    "tools": [
+        {
+            "type": "function",
+            "function": {
+                "name": "render_chart",
+                "description": "Render a chart in the UI",
+                "parameters": {
+                    "type": "object",
+                    "properties": {"data": {"type": "array"}},
+                },
+            },
+        }
+    ],
+}
+
+
+# ── Golden stream tests ──
+
+
+async def test_declaration_only_tool_golden_sequence() -> None:
+    """Declaration-only tool: TOOL_CALL_START/ARGS emitted, TOOL_CALL_END at stream end."""
+    # The LLM calls a client-side tool (no server-side execution)
+    updates = [
+        AgentResponseUpdate(
+            contents=[
+                Content.from_function_call(
+                    name="render_chart",
+                    call_id="call-chart",
+                    arguments='{"data": [1, 2, 3]}',
+                )
+            ],
+            role="assistant",
+        ),
+    ]
+    agent = _build_agent(updates)
+    stream = await _run(agent, PAYLOAD)
+
+    stream.assert_bookends()
+    stream.assert_no_run_error()
+
+    # Tool call start and args should be present
+    stream.assert_has_type("TOOL_CALL_START")
+    stream.assert_has_type("TOOL_CALL_ARGS")
+
+    # TOOL_CALL_END should be emitted (via get_pending_without_end)
+    stream.assert_has_type("TOOL_CALL_END")
+    stream.assert_tool_calls_balanced()
+
+
+async def test_declaration_only_tool_no_tool_call_result() -> None:
+    """Declaration-only tools should NOT produce TOOL_CALL_RESULT events."""
+    updates = [
+        AgentResponseUpdate(
+            contents=[
+                Content.from_function_call(
+                    name="render_chart",
+                    call_id="call-chart",
+                    arguments='{"data": [1, 2, 3]}',
+                )
+            ],
+            role="assistant",
+        ),
+    ]
+    agent = _build_agent(updates)
+    stream = await _run(agent, PAYLOAD)
+
+    assert "TOOL_CALL_RESULT" not in stream.types(), "Declaration-only tools should not have TOOL_CALL_RESULT"
+
+
+async def test_declaration_only_tool_text_messages_balanced() -> None:
+    """Text messages remain balanced even with declaration-only tools."""
+    updates = [
+        AgentResponseUpdate(
+            contents=[
+                Content.from_function_call(
+                    name="render_chart",
+                    call_id="call-chart",
+                    arguments='{"data": [1, 2, 3]}',
+                )
+            ],
+            role="assistant",
+        ),
+    ]
+    agent = _build_agent(updates)
+    stream = await _run(agent, PAYLOAD)
+
+    stream.assert_text_messages_balanced()
+
+
+async def test_declaration_only_tool_messages_snapshot() -> None:
+    """MessagesSnapshotEvent includes the tool call for declaration-only tools."""
+    updates = [
+        AgentResponseUpdate(
+            contents=[
+                Content.from_function_call(
+                    name="render_chart",
+                    call_id="call-chart",
+                    arguments='{"data": [1, 2, 3]}',
+                )
+            ],
+            role="assistant",
+        ),
+    ]
+    agent = _build_agent(updates)
+    stream = await _run(agent, PAYLOAD)
+
+    stream.assert_has_type("MESSAGES_SNAPSHOT")
@@ -0,0 +1,194 @@
+# Copyright (c) Microsoft. All rights reserved.
+
+"""Golden event-stream tests for the HITL (human-in-the-loop) approval scenario."""
+
+from __future__ import annotations
+
+import json
+from typing import Any
+
+from agent_framework import AgentResponseUpdate, Content
+from conftest import StubAgent  # pyrefly: ignore[missing-import] # pyright: ignore[reportMissingImports]
+from event_stream import EventStream  # pyrefly: ignore[missing-import] # pyright: ignore[reportMissingImports]
+
+from agent_framework_ag_ui import AgentFrameworkAgent
+
+PREDICT_CONFIG = {
+    "tasks": {
+        "tool": "generate_task_steps",
+        "tool_argument": "steps",
+    }
+}
+
+STATE_SCHEMA = {
+    "tasks": {"type": "array", "items": {"type": "object"}},
+}
+
+
+def _build_agent(updates: list[AgentResponseUpdate], **kwargs: Any) -> AgentFrameworkAgent:
+    stub = StubAgent(updates=updates)
+    return AgentFrameworkAgent(
+        agent=stub,
+        state_schema=STATE_SCHEMA,
+        predict_state_config=PREDICT_CONFIG,
+        require_confirmation=True,
+        **kwargs,
+    )
+
+
+async def _run(agent: AgentFrameworkAgent, payload: dict[str, Any]) -> EventStream:
+    return EventStream([event async for event in agent.run(payload)])
+
+
+STEPS = [
+    {"description": "Step 1: Plan", "status": "enabled"},
+    {"description": "Step 2: Execute", "status": "enabled"},
+]
+
+
+PAYLOAD: dict[str, Any] = {
+    "thread_id": "thread-hitl",
+    "run_id": "run-hitl",
+    "messages": [{"role": "user", "content": "Plan my tasks"}],
+    "state": {"tasks": []},
+}
+
+
+# ── Turn 1: Tool call → confirm_changes → interrupt ──
+
+
+async def test_hitl_turn1_golden_sequence() -> None:
+    """Turn 1 emits tool call, confirm_changes, and finishes with interrupt."""
+    updates = [
+        AgentResponseUpdate(
+            contents=[
+                Content.from_function_call(
+                    name="generate_task_steps",
+                    call_id="call-steps",
+                    arguments=json.dumps({"steps": STEPS}),
+                )
+            ],
+            role="assistant",
+        ),
+    ]
+    agent = _build_agent(updates)
+    stream = await _run(agent, PAYLOAD)
+
+    # Should have: tool call start/args/end for the primary tool,
+    # then TOOL_CALL_END, STATE_SNAPSHOT, confirm_changes cycle
+    stream.assert_bookends()
+    stream.assert_no_run_error()
+
+    # confirm_changes tool call should be present
+    tool_starts = stream.get("TOOL_CALL_START")
+    tool_names = [getattr(s, "tool_call_name", None) for s in tool_starts]
+    assert "generate_task_steps" in tool_names
+    assert "confirm_changes" in tool_names
+
+    # RUN_FINISHED should have interrupt metadata
+    interrupt = stream.run_finished_interrupts()
+    assert len(interrupt) > 0
+
+
+async def test_hitl_turn1_tool_calls_balanced() -> None:
+    """All tool calls in turn 1 (primary + confirm_changes) are balanced."""
+    updates = [
+        AgentResponseUpdate(
+            contents=[
+                Content.from_function_call(
+                    name="generate_task_steps",
+                    call_id="call-steps",
+                    arguments=json.dumps({"steps": STEPS}),
+                )
+            ],
+            role="assistant",
+        ),
+    ]
+    agent = _build_agent(updates)
+    stream = await _run(agent, PAYLOAD)
+
+    stream.assert_tool_calls_balanced()
+
+
+async def test_hitl_turn1_text_messages_balanced() -> None:
+    """Text messages are balanced even in the approval flow."""
+    updates = [
+        AgentResponseUpdate(
+            contents=[
+                Content.from_function_call(
+                    name="generate_task_steps",
+                    call_id="call-steps",
+                    arguments=json.dumps({"steps": STEPS}),
+                )
+            ],
+            role="assistant",
+        ),
+    ]
+    agent = _build_agent(updates)
+    stream = await _run(agent, PAYLOAD)
+
+    stream.assert_text_messages_balanced()
+
+
+# ── Turn 2: Resume with approval → confirmation message → no interrupt ──
+
+
+async def test_hitl_turn2_resume_with_approval() -> None:
+    """Resuming with confirm_changes result emits confirmation text and finishes cleanly."""
+    # Turn 2: user sends confirm_changes result as resume
+    # The agent wrapper sees a confirm_changes response and emits a confirmation message
+    confirm_result = json.dumps(
+        {
+            "accepted": True,
+            "steps": STEPS,
+        }
+    )
+
+    # Build payload with resume containing the approval
+    # For confirm_changes, the messages should include the tool result
+    payload: dict[str, Any] = {
+        "thread_id": "thread-hitl",
+        "run_id": "run-hitl-2",
+        "messages": [
+            {"role": "user", "content": "Plan my tasks"},
+            {
+                "role": "assistant",
+                "tool_calls": [
+                    {
+                        "id": "confirm-id-1",
+                        "type": "function",
+                        "function": {"name": "confirm_changes", "arguments": json.dumps({"steps": STEPS})},
+                    }
+                ],
+            },
+            {
+                "role": "tool",
+                "toolCallId": "confirm-id-1",
+                "content": confirm_result,
+            },
+        ],
+        "state": {"tasks": []},
+    }
+
+    # In turn 2, the agent sees the confirm_changes result and emits a confirmation text
+    updates = [
+        AgentResponseUpdate(
+            contents=[Content.from_text(text="Tasks confirmed!")],
+            role="assistant",
+        ),
+    ]
+    agent = _build_agent(updates)
+    stream = await _run(agent, payload)
+
+    stream.assert_bookends()
+    stream.assert_text_messages_balanced()
+    stream.assert_no_run_error()
+
+    # Should have text message content (the confirmation message)
+    text_events = stream.get("TEXT_MESSAGE_CONTENT")
+    assert text_events, "Expected confirmation text message"
+
+    # RUN_FINISHED should NOT have interrupt (approval completed)
+    finished = stream.last("RUN_FINISHED")
+    dumped = finished.model_dump(by_alias=True, exclude_none=True)
+    assert "outcome" not in dumped, f"Expected no interrupt after approval, got {dumped.get('outcome')}"
@@ -0,0 +1,130 @@
+# Copyright (c) Microsoft. All rights reserved.
+
+"""Golden event-stream tests for the predictive state scenario."""
+
+from __future__ import annotations
+
+from typing import Any
+
+from agent_framework import AgentResponseUpdate, Content
+from conftest import StubAgent  # pyrefly: ignore[missing-import] # pyright: ignore[reportMissingImports]
+from event_stream import EventStream  # pyrefly: ignore[missing-import] # pyright: ignore[reportMissingImports]
+
+from agent_framework_ag_ui import AgentFrameworkAgent
+
+PREDICT_CONFIG = {
+    "document": {
+        "tool": "update_document",
+        "tool_argument": "content",
+    }
+}
+
+STATE_SCHEMA = {
+    "document": {"type": "string"},
+}
+
+
+def _build_agent(updates: list[AgentResponseUpdate], **kwargs: Any) -> AgentFrameworkAgent:
+    stub = StubAgent(updates=updates)
+    return AgentFrameworkAgent(
+        agent=stub,
+        state_schema=STATE_SCHEMA,
+        predict_state_config=PREDICT_CONFIG,
+        require_confirmation=False,
+        **kwargs,
+    )
+
+
+async def _run(agent: AgentFrameworkAgent, payload: dict[str, Any]) -> EventStream:
+    return EventStream([event async for event in agent.run(payload)])
+
+
+PAYLOAD: dict[str, Any] = {
+    "thread_id": "thread-predict",
+    "run_id": "run-predict",
+    "messages": [{"role": "user", "content": "Write a document"}],
+    "state": {"document": ""},
+}
+
+
+# ── Golden stream tests ──
+
+
+async def test_predictive_state_emits_deltas_during_tool_args() -> None:
+    """STATE_DELTA events are emitted as tool arguments stream in."""
+    updates = [
+        AgentResponseUpdate(
+            contents=[Content.from_function_call(name="update_document", call_id="call-1", arguments="")],
+            role="assistant",
+        ),
+        AgentResponseUpdate(
+            contents=[
+                Content.from_function_call(name="update_document", call_id="call-1", arguments='{"content": "Hello')
+            ],
+            role="assistant",
+        ),
+        AgentResponseUpdate(
+            contents=[Content.from_function_call(name="update_document", call_id="call-1", arguments=' world"}')],
+            role="assistant",
+        ),
+    ]
+    agent = _build_agent(updates)
+    stream = await _run(agent, PAYLOAD)
+
+    stream.assert_bookends()
+    stream.assert_no_run_error()
+
+    # PredictState custom event should be present
+    custom_events = stream.get("CUSTOM")
+    predict_events = [e for e in custom_events if getattr(e, "name", None) == "PredictState"]
+    assert predict_events, "Expected PredictState custom event"
+
+    # STATE_DELTA events should be emitted during tool arg streaming
+    assert "STATE_DELTA" in stream.types(), "Expected STATE_DELTA events during predictive streaming"
+
+
+async def test_predictive_state_snapshot_after_tool_end() -> None:
+    """STATE_SNAPSHOT is emitted when a predictive tool completes (no confirmation)."""
+    updates = [
+        AgentResponseUpdate(
+            contents=[
+                Content.from_function_call(
+                    name="update_document", call_id="call-1", arguments='{"content": "Final text"}'
+                )
+            ],
+            role="assistant",
+        ),
+    ]
+    agent = _build_agent(updates)
+    stream = await _run(agent, PAYLOAD)
+
+    stream.assert_bookends()
+
+    # Should have initial state snapshot + updated snapshot after tool completion
+    snapshots = stream.get("STATE_SNAPSHOT")
+    assert len(snapshots) >= 1, "Expected at least one STATE_SNAPSHOT"
+
+
+async def test_predictive_state_ordered_events() -> None:
+    """Event ordering: RUN_STARTED → PredictState → STATE_SNAPSHOT → TOOL_CALL_* → STATE_SNAPSHOT → RUN_FINISHED."""
+    updates = [
+        AgentResponseUpdate(
+            contents=[
+                Content.from_function_call(name="update_document", call_id="call-1", arguments='{"content": "doc"}')
+            ],
+            role="assistant",
+        ),
+    ]
+    agent = _build_agent(updates)
+    stream = await _run(agent, PAYLOAD)
+
+    stream.assert_ordered_types(
+        [
+            "RUN_STARTED",
+            "CUSTOM",  # PredictState
+            "STATE_SNAPSHOT",  # Initial state
+            "TOOL_CALL_START",
+            "TOOL_CALL_ARGS",
+            "RUN_FINISHED",
+        ]
+    )
@@ -0,0 +1,110 @@
+# Copyright (c) Microsoft. All rights reserved.
+
+"""Golden event-stream tests for the shared state (structured output) scenario."""
+
+from __future__ import annotations
+
+from typing import Any
+
+from agent_framework import AgentResponseUpdate, Content
+from conftest import StubAgent  # pyrefly: ignore[missing-import] # pyright: ignore[reportMissingImports]
+from event_stream import EventStream  # pyrefly: ignore[missing-import] # pyright: ignore[reportMissingImports]
+from pydantic import BaseModel
+
+from agent_framework_ag_ui import AgentFrameworkAgent
+
+
+class RecipeState(BaseModel):
+    recipe_title: str = ""
+    ingredients: list[str] = []
+    message: str = ""
+
+
+def _build_agent(updates: list[AgentResponseUpdate], **kwargs: Any) -> AgentFrameworkAgent:
+    stub = StubAgent(
+        updates=updates,
+        default_options={"tools": None, "response_format": RecipeState},
+    )
+    return AgentFrameworkAgent(
+        agent=stub,
+        state_schema={
+            "recipe_title": {"type": "string"},
+            "ingredients": {"type": "array", "items": {"type": "string"}},
+        },
+        **kwargs,
+    )
+
+
+async def _run(agent: AgentFrameworkAgent, payload: dict[str, Any]) -> EventStream:
+    return EventStream([event async for event in agent.run(payload)])
+
+
+PAYLOAD: dict[str, Any] = {
+    "thread_id": "thread-state",
+    "run_id": "run-state",
+    "messages": [{"role": "user", "content": "Give me a pasta recipe"}],
+    "state": {"recipe_title": "", "ingredients": []},
+}
+
+
+# ── Golden stream tests ──
+
+
+async def test_shared_state_emits_state_snapshot() -> None:
+    """Structured output agent emits STATE_SNAPSHOT with parsed model fields."""
+    # The structured output agent gets a response that the framework parses as RecipeState
+    updates = [
+        AgentResponseUpdate(
+            contents=[
+                Content.from_text(
+                    text='{"recipe_title": "Pasta Carbonara", "ingredients": ["pasta", "eggs", "cheese"], "message": "Here is your recipe!"}'
+                )
+            ],
+            role="assistant",
+        ),
+    ]
+    agent = _build_agent(updates)
+    stream = await _run(agent, PAYLOAD)
+
+    stream.assert_bookends()
+    stream.assert_no_run_error()
+
+    # Should have STATE_SNAPSHOT with the initial state at minimum
+    stream.assert_has_type("STATE_SNAPSHOT")
+
+
+async def test_shared_state_initial_snapshot_on_first_update() -> None:
+    """When state_schema and state are provided, initial STATE_SNAPSHOT is emitted after RUN_STARTED."""
+    updates = [
+        AgentResponseUpdate(
+            contents=[Content.from_text(text='{"recipe_title": "Test", "ingredients": [], "message": "hi"}')],
+            role="assistant",
+        ),
+    ]
+    agent = _build_agent(updates)
+    stream = await _run(agent, PAYLOAD)
+
+    # RUN_STARTED should be followed by STATE_SNAPSHOT (initial state)
+    stream.assert_ordered_types(["RUN_STARTED", "STATE_SNAPSHOT"])
+
+
+async def test_shared_state_text_emitted_from_message_field() -> None:
+    """Structured output's 'message' field is emitted as text message events."""
+    updates = [
+        AgentResponseUpdate(
+            contents=[
+                Content.from_text(
+                    text='{"recipe_title": "Pasta", "ingredients": ["pasta"], "message": "Enjoy your pasta!"}'
+                )
+            ],
+            role="assistant",
+        ),
+    ]
+    agent = _build_agent(updates)
+    stream = await _run(agent, PAYLOAD)
+
+    # Text should be emitted from the message field
+    text_contents = stream.get("TEXT_MESSAGE_CONTENT")
+    if text_contents:
+        combined = "".join(getattr(e, "delta", "") for e in text_contents)
+        assert "Enjoy your pasta!" in combined
@@ -0,0 +1,207 @@
+# Copyright (c) Microsoft. All rights reserved.
+
+"""Golden event-stream tests for the workflow HITL (subgraphs) scenario.
+
+Extends the existing test_subgraphs_example_agent.py with EventStream assertions
+on full event ordering, balancing, and interrupt structure.
+"""
+
+from __future__ import annotations
+
+import json
+from typing import Any
+
+from event_stream import EventStream  # pyrefly: ignore[missing-import] # pyright: ignore[reportMissingImports]
+
+from agent_framework_ag_ui_examples.agents.subgraphs_agent import subgraphs_agent
+
+
+async def _run(agent: Any, payload: dict[str, Any]) -> EventStream:
+    return EventStream([event async for event in agent.run(payload)])
+
+
+# ── Turn 1: Initial request → flight interrupt ──
+
+
+async def test_subgraphs_turn1_golden_bookends() -> None:
+    """Turn 1 starts with RUN_STARTED and ends with RUN_FINISHED."""
+    agent = subgraphs_agent()
+    stream = await _run(
+        agent,
+        {
+            "thread_id": "thread-sub-golden-1",
+            "run_id": "run-1",
+            "messages": [{"role": "user", "content": "Plan a trip to San Francisco"}],
+        },
+    )
+    stream.assert_bookends()
+
+
+async def test_subgraphs_turn1_no_errors() -> None:
+    """Turn 1 completes without errors."""
+    agent = subgraphs_agent()
+    stream = await _run(
+        agent,
+        {
+            "thread_id": "thread-sub-golden-2",
+            "run_id": "run-1",
+            "messages": [{"role": "user", "content": "Plan a trip"}],
+        },
+    )
+    stream.assert_no_run_error()
+
+
+async def test_subgraphs_turn1_has_step_events() -> None:
+    """Turn 1 emits STEP_STARTED and STEP_FINISHED for workflow executors."""
+    agent = subgraphs_agent()
+    stream = await _run(
+        agent,
+        {
+            "thread_id": "thread-sub-golden-3",
+            "run_id": "run-1",
+            "messages": [{"role": "user", "content": "Plan a trip"}],
+        },
+    )
+    stream.assert_has_type("STEP_STARTED")
+    stream.assert_has_type("STEP_FINISHED")
+
+
+async def test_subgraphs_turn1_interrupt_structure() -> None:
+    """Turn 1 RUN_FINISHED carries flight interrupt with correct structure."""
+    agent = subgraphs_agent()
+    stream = await _run(
+        agent,
+        {
+            "thread_id": "thread-sub-golden-4",
+            "run_id": "run-1",
+            "messages": [{"role": "user", "content": "Plan a trip to SF"}],
+        },
+    )
+
+    interrupt = stream.run_finished_interrupts()
+    assert len(interrupt) > 0
+    interrupt_value = stream.interrupt_metadata_value(interrupt[0])
+    assert interrupt_value["agent"] == "flights"
+    assert len(interrupt_value["options"]) == 2
+
+
+async def test_subgraphs_turn1_text_messages_balanced() -> None:
+    """All text messages in turn 1 are properly balanced."""
+    agent = subgraphs_agent()
+    stream = await _run(
+        agent,
+        {
+            "thread_id": "thread-sub-golden-5",
+            "run_id": "run-1",
+            "messages": [{"role": "user", "content": "Plan a trip"}],
+        },
+    )
+    stream.assert_text_messages_balanced()
+
+
+async def test_subgraphs_turn1_ordered_flow() -> None:
+    """Turn 1 event ordering: RUN_STARTED → STATE_SNAPSHOT → STEP_* → TOOL_CALL_* → RUN_FINISHED."""
+    agent = subgraphs_agent()
+    stream = await _run(
+        agent,
+        {
+            "thread_id": "thread-sub-golden-6",
+            "run_id": "run-1",
+            "messages": [{"role": "user", "content": "Plan a trip"}],
+        },
+    )
+    stream.assert_ordered_types(
+        [
+            "RUN_STARTED",
+            "STATE_SNAPSHOT",
+            "STEP_STARTED",
+            "RUN_FINISHED",
+        ]
+    )
+
+
+# ── Multi-turn: Flight selection → hotel interrupt → completion ──
+
+
+async def test_subgraphs_full_flow_event_ordering() -> None:
+    """Complete 3-turn flow maintains proper event ordering throughout."""
+    agent = subgraphs_agent()
+    thread_id = "thread-sub-golden-full"
+
+    # Turn 1
+    stream1 = await _run(
+        agent,
+        {
+            "thread_id": thread_id,
+            "run_id": "run-1",
+            "messages": [{"role": "user", "content": "Plan a trip to SF from Amsterdam"}],
+        },
+    )
+    stream1.assert_bookends()
+    stream1.assert_no_run_error()
+
+    # Extract flight interrupt
+    interrupt1 = stream1.run_finished_interrupts()[0]
+
+    # Turn 2: Select flight
+    stream2 = await _run(
+        agent,
+        {
+            "thread_id": thread_id,
+            "run_id": "run-2",
+            "resume": {
+                "interrupts": [
+                    {
+                        "id": interrupt1["id"],
+                        "value": json.dumps(
+                            {
+                                "airline": "United",
+                                "departure": "Amsterdam (AMS)",
+                                "arrival": "San Francisco (SFO)",
+                                "price": "$720",
+                                "duration": "12h 15m",
+                            }
+                        ),
+                    }
+                ]
+            },
+        },
+    )
+    stream2.assert_bookends()
+    stream2.assert_no_run_error()
+
+    # Should now have hotel interrupt
+    interrupt2 = stream2.run_finished_interrupts()
+    assert stream2.interrupt_metadata_value(interrupt2[0])["agent"] == "hotels"
+
+    # Turn 3: Select hotel
+    stream3 = await _run(
+        agent,
+        {
+            "thread_id": thread_id,
+            "run_id": "run-3",
+            "resume": {
+                "interrupts": [
+                    {
+                        "id": interrupt2[0]["id"],
+                        "value": json.dumps(
+                            {
+                                "name": "The Ritz-Carlton",
+                                "location": "Nob Hill",
+                                "price_per_night": "$550/night",
+                                "rating": "4.8 stars",
+                            }
+                        ),
+                    }
+                ]
+            },
+        },
+    )
+    stream3.assert_bookends()
+    stream3.assert_no_run_error()
+    stream3.assert_text_messages_balanced()
+
+    # Final turn should not have interrupt
+    finished3 = stream3.last("RUN_FINISHED")
+    final_dump = finished3.model_dump(by_alias=True, exclude_none=True)
+    assert "outcome" not in final_dump, f"Expected no interrupt after completion, got {final_dump.get('outcome')}"
@@ -0,0 +1,945 @@
+# Copyright (c) Microsoft. All rights reserved.
+
+"""Comprehensive golden event-stream tests for AgentFrameworkWorkflow.
+
+Covers the full matrix of workflow-specific AG-UI patterns:
+- request_info → TOOL_CALL lifecycle and balancing
+- Executor step events and activity snapshots
+- Text output, dict output, BaseEvent passthrough, AgentResponse output
+- Text deduplication across workflow outputs
+- Workflow error handling → RUN_ERROR
+- Multi-turn interrupt/resume round-trips
+- Empty turns with pending requests
+- Custom workflow events
+- Text message draining on request_info and executor boundaries
+"""
+
+import json
+from typing import Any, cast
+
+from ag_ui.core import EventType, StateSnapshotEvent
+from agent_framework import (
+    AgentResponse,
+    Content,
+    Executor,
+    Message,
+    WorkflowBuilder,
+    WorkflowContext,
+    WorkflowEvent,
+    executor,
+    handler,
+    response_handler,
+)
+from event_stream import EventStream  # pyrefly: ignore[missing-import] # pyright: ignore[reportMissingImports]
+
+from agent_framework_ag_ui import AgentFrameworkWorkflow
+
+
+async def _run(wrapper: AgentFrameworkWorkflow, payload: dict[str, Any]) -> EventStream:
+    return EventStream([event async for event in wrapper.run(payload)])
+
+
+def _payload(
+    msg: str = "go",
+    *,
+    thread_id: str = "thread-wf",
+    run_id: str = "run-wf",
+    **extra: Any,
+) -> dict[str, Any]:
+    return {"thread_id": thread_id, "run_id": run_id, "messages": [{"role": "user", "content": msg}], **extra}
+
+
+# ──────────────────────────────────────────────────────────────────────
+# 1. Basic workflow text output
+# ──────────────────────────────────────────────────────────────────────
+
+
+async def test_workflow_text_output_golden_sequence() -> None:
+    """Simple text output: RUN_STARTED → STEP_STARTED → TEXT_* → STEP_FINISHED → TEXT_MESSAGE_END → RUN_FINISHED."""
+
+    @executor(id="greeter")
+    async def greeter(message: Any, ctx: WorkflowContext[Any, str]) -> None:
+        await ctx.yield_output("Hello from workflow!")
+
+    workflow = WorkflowBuilder(start_executor=greeter).build()
+    wrapper = AgentFrameworkWorkflow(workflow=workflow)
+    stream = await _run(wrapper, _payload())
+
+    stream.assert_bookends()
+    stream.assert_no_run_error()
+    stream.assert_text_messages_balanced()
+    stream.assert_has_type("TEXT_MESSAGE_START")
+    stream.assert_has_type("TEXT_MESSAGE_CONTENT")
+    stream.assert_has_type("TEXT_MESSAGE_END")
+
+    # Verify actual content
+    deltas = [e.delta for e in stream.get("TEXT_MESSAGE_CONTENT")]
+    assert "Hello from workflow!" in deltas
+
+
+async def test_workflow_text_output_message_id_consistency() -> None:
+    """All text events for a single output share the same message_id."""
+
+    @executor(id="echo")
+    async def echo(message: Any, ctx: WorkflowContext[Any, str]) -> None:
+        await ctx.yield_output("echo reply")
+
+    workflow = WorkflowBuilder(start_executor=echo).build()
+    wrapper = AgentFrameworkWorkflow(workflow=workflow)
+    stream = await _run(wrapper, _payload())
+
+    stream.assert_message_ids_consistent()
+
+
+# ──────────────────────────────────────────────────────────────────────
+# 2. Executor step events and activity snapshots
+# ──────────────────────────────────────────────────────────────────────
+
+
+async def test_workflow_executor_lifecycle_events() -> None:
+    """Executor invocation produces STEP_STARTED, ACTIVITY_SNAPSHOT, STEP_FINISHED."""
+
+    @executor(id="worker")
+    async def worker(message: Any, ctx: WorkflowContext[Any, str]) -> None:
+        await ctx.yield_output("done")
+
+    workflow = WorkflowBuilder(start_executor=worker).build()
+    wrapper = AgentFrameworkWorkflow(workflow=workflow)
+    stream = await _run(wrapper, _payload())
+
+    # Step events with executor ID
+    started = [e for e in stream.get("STEP_STARTED") if getattr(e, "step_name", "") == "worker"]
+    finished = [e for e in stream.get("STEP_FINISHED") if getattr(e, "step_name", "") == "worker"]
+    assert started, "Expected STEP_STARTED for 'worker'"
+    assert finished, "Expected STEP_FINISHED for 'worker'"
+
+    # Activity snapshots
+    activities = stream.get("ACTIVITY_SNAPSHOT")
+    assert activities, "Expected ACTIVITY_SNAPSHOT events"
+    # Check one of them has executor payload
+    executor_activities = [a for a in activities if getattr(a, "activity_type", None) == "executor"]
+    assert executor_activities, "Expected executor-type activity snapshots"
+
+
+async def test_workflow_executor_step_ordering() -> None:
+    """STEP_STARTED comes before content, STEP_FINISHED comes after."""
+
+    @executor(id="orderer")
+    async def orderer(message: Any, ctx: WorkflowContext[Any, str]) -> None:
+        await ctx.yield_output("ordered output")
+
+    workflow = WorkflowBuilder(start_executor=orderer).build()
+    wrapper = AgentFrameworkWorkflow(workflow=workflow)
+    stream = await _run(wrapper, _payload())
+
+    stream.assert_ordered_types(
+        [
+            "RUN_STARTED",
+            "STEP_STARTED",
+            "TEXT_MESSAGE_START",
+            "TEXT_MESSAGE_CONTENT",
+            "STEP_FINISHED",
+            "RUN_FINISHED",
+        ]
+    )
+
+
+# ──────────────────────────────────────────────────────────────────────
+# 3. Dict output → CUSTOM workflow_output
+# ──────────────────────────────────────────────────────────────────────
+
+
+async def test_workflow_dict_output_maps_to_custom_event() -> None:
+    """Non-chat dict output is emitted as CUSTOM workflow_output event."""
+
+    @executor(id="structured")
+    async def structured(message: Any, ctx: WorkflowContext[Any, dict[str, int]]) -> None:
+        await ctx.yield_output({"count": 42, "status": 1})
+
+    workflow = WorkflowBuilder(start_executor=structured).build()
+    wrapper = AgentFrameworkWorkflow(workflow=workflow)
+    stream = await _run(wrapper, _payload())
+
+    stream.assert_bookends()
+    stream.assert_no_run_error()
+
+    customs = [e for e in stream.get("CUSTOM") if getattr(e, "name", None) == "workflow_output"]
+    assert len(customs) == 1
+    assert customs[0].value == {"count": 42, "status": 1}
+
+    # Should NOT have TEXT_MESSAGE events for dict output
+    assert "TEXT_MESSAGE_CONTENT" not in stream.types()
+
+
+# ──────────────────────────────────────────────────────────────────────
+# 4. BaseEvent passthrough
+# ──────────────────────────────────────────────────────────────────────
+
+
+async def test_workflow_base_event_passthrough() -> None:
+    """AG-UI BaseEvent outputs are yielded directly, not wrapped."""
+
+    @executor(id="stateful")
+    async def stateful(message: Any, ctx: WorkflowContext[Any, StateSnapshotEvent]) -> None:
+        await ctx.yield_output(StateSnapshotEvent(type=EventType.STATE_SNAPSHOT, snapshot={"active_agent": "flights"}))
+
+    workflow = WorkflowBuilder(start_executor=stateful).build()
+    wrapper = AgentFrameworkWorkflow(workflow=workflow)
+    stream = await _run(wrapper, _payload())
+
+    stream.assert_bookends()
+    snapshots = stream.get("STATE_SNAPSHOT")
+    assert len(snapshots) == 1
+    assert snapshots[0].snapshot["active_agent"] == "flights"
+
+
+# ──────────────────────────────────────────────────────────────────────
+# 5. AgentResponse output (conversation payload)
+# ──────────────────────────────────────────────────────────────────────
+
+
+async def test_workflow_agent_response_output_extracts_latest_assistant() -> None:
+    """AgentResponse output uses only the latest assistant message, not full history."""
+
+    @executor(id="responder")
+    async def responder(message: Any, ctx: WorkflowContext[Any, AgentResponse]) -> None:
+        response = AgentResponse(
+            messages=[
+                Message(role="user", contents=[Content.from_text("My order is damaged")]),
+                Message(role="assistant", contents=[Content.from_text("I'll process your replacement.")]),
+            ]
+        )
+        await ctx.yield_output(response)
+
+    workflow = WorkflowBuilder(start_executor=responder).build()
+    wrapper = AgentFrameworkWorkflow(workflow=workflow)
+    stream = await _run(wrapper, _payload())
+
+    stream.assert_bookends()
+    stream.assert_text_messages_balanced()
+
+    deltas = [e.delta for e in stream.get("TEXT_MESSAGE_CONTENT")]
+    assert deltas == ["I'll process your replacement."]
+
+
+# ──────────────────────────────────────────────────────────────────────
+# 6. Custom workflow events
+# ──────────────────────────────────────────────────────────────────────
+
+
+class ProgressEvent(WorkflowEvent):
+    """Custom workflow event for testing CUSTOM event mapping."""
+
+    def __init__(self, progress: int) -> None:
+        super().__init__(cast(Any, "custom_progress"), data={"progress": progress})
+
+
+async def test_workflow_custom_events() -> None:
+    """Custom workflow events are mapped to CUSTOM AG-UI events."""
+
+    @executor(id="progress_tracker")
+    async def progress_tracker(message: Any, ctx: WorkflowContext[Any, str]) -> None:
+        await ctx.add_event(ProgressEvent(25))
+        await ctx.yield_output("In progress...")
+        await ctx.add_event(ProgressEvent(100))
+
+    workflow = WorkflowBuilder(start_executor=progress_tracker).build()
+    wrapper = AgentFrameworkWorkflow(workflow=workflow)
+    stream = await _run(wrapper, _payload())
+
+    stream.assert_bookends()
+    stream.assert_no_run_error()
+
+    progress_events = [e for e in stream.get("CUSTOM") if getattr(e, "name", None) == "custom_progress"]
+    assert len(progress_events) == 2
+    assert progress_events[0].value == {"progress": 25}
+    assert progress_events[1].value == {"progress": 100}
+
+
+# ──────────────────────────────────────────────────────────────────────
+# 7. request_info → TOOL_CALL lifecycle
+# ──────────────────────────────────────────────────────────────────────
+
+
+async def test_workflow_request_info_tool_call_lifecycle() -> None:
+    """request_info emits TOOL_CALL_START/ARGS/END cycle plus CUSTOM request_info."""
+
+    @executor(id="requester")
+    async def requester(message: Any, ctx: WorkflowContext) -> None:
+        await ctx.request_info("Need approval", str, request_id="req-1")
+
+    workflow = WorkflowBuilder(start_executor=requester).build()
+    wrapper = AgentFrameworkWorkflow(workflow=workflow)
+    stream = await _run(wrapper, _payload())
+
+    stream.assert_bookends()
+    stream.assert_no_run_error()
+
+    # Tool call lifecycle
+    stream.assert_ordered_types(
+        [
+            "RUN_STARTED",
+            "TOOL_CALL_START",
+            "TOOL_CALL_ARGS",
+            "TOOL_CALL_END",
+            "CUSTOM",  # request_info
+            "RUN_FINISHED",
+        ]
+    )
+
+    # Verify tool call details
+    start = stream.first("TOOL_CALL_START")
+    assert start.tool_call_id == "req-1"
+    assert start.tool_call_name == "request_info"
+
+    # TOOL_CALL_ARGS should contain the request payload
+    args = stream.first("TOOL_CALL_ARGS")
+    assert args.tool_call_id == "req-1"
+    parsed_args = json.loads(args.delta)
+    assert parsed_args["request_id"] == "req-1"
+
+    # Tool calls should be balanced
+    stream.assert_tool_calls_balanced()
+
+
+async def test_workflow_request_info_interrupt_in_run_finished() -> None:
+    """request_info populates RUN_FINISHED.outcome.interrupts with the request metadata."""
+
+    @executor(id="requester")
+    async def requester(message: Any, ctx: WorkflowContext) -> None:
+        await ctx.request_info(
+            {"message": "Choose a flight", "options": [{"airline": "KLM"}], "agent": "flights"},
+            dict,
+            request_id="flights-choice",
+        )
+
+    workflow = WorkflowBuilder(start_executor=requester).build()
+    wrapper = AgentFrameworkWorkflow(workflow=workflow)
+    stream = await _run(wrapper, _payload())
+
+    interrupt = stream.run_finished_interrupts()
+    assert len(interrupt) == 1
+    assert interrupt[0]["id"] == "flights-choice"
+    assert stream.interrupt_metadata_value(interrupt[0])["agent"] == "flights"
+
+
+async def test_workflow_request_info_emits_interrupt_card_event() -> None:
+    """request_info with dict data emits a WorkflowInterruptEvent custom event."""
+
+    @executor(id="requester")
+    async def requester(message: Any, ctx: WorkflowContext) -> None:
+        await ctx.request_info(
+            {"message": "Pick one", "options": ["A", "B"]},
+            dict,
+            request_id="pick-1",
+        )
+
+    workflow = WorkflowBuilder(start_executor=requester).build()
+    wrapper = AgentFrameworkWorkflow(workflow=workflow)
+    stream = await _run(wrapper, _payload())
+
+    interrupt_cards = [e for e in stream.get("CUSTOM") if getattr(e, "name", None) == "WorkflowInterruptEvent"]
+    assert interrupt_cards, "Expected WorkflowInterruptEvent custom event"
+
+
+# ──────────────────────────────────────────────────────────────────────
+# 8. Text message draining on request_info boundary
+# ──────────────────────────────────────────────────────────────────────
+
+
+async def test_workflow_text_drained_before_request_info() -> None:
+    """Open text message is closed (TEXT_MESSAGE_END) before request_info tool calls begin."""
+
+    @executor(id="text_then_request")
+    async def text_then_request(message: Any, ctx: WorkflowContext) -> None:
+        await ctx.yield_output("Please confirm this action.")  # type: ignore[arg-type]  # pyrefly: ignore[bad-argument-type]  # ty: ignore[invalid-argument-type]
+        await ctx.request_info("Need approval", str, request_id="approval-1")
+
+    workflow = WorkflowBuilder(start_executor=text_then_request).build()
+    wrapper = AgentFrameworkWorkflow(workflow=workflow)
+    stream = await _run(wrapper, _payload())
+
+    stream.assert_text_messages_balanced()
+    stream.assert_tool_calls_balanced()
+
+    # TEXT_MESSAGE_END must appear before TOOL_CALL_START
+    types = stream.types()
+    text_end_idx = types.index("TEXT_MESSAGE_END")
+    tool_start_idx = types.index("TOOL_CALL_START")
+    assert text_end_idx < tool_start_idx, (
+        f"TEXT_MESSAGE_END (idx={text_end_idx}) must come before TOOL_CALL_START (idx={tool_start_idx})"
+    )
+
+
+# ──────────────────────────────────────────────────────────────────────
+# 9. Text deduplication
+# ──────────────────────────────────────────────────────────────────────
+
+
+async def test_workflow_skips_duplicate_text_from_snapshot() -> None:
+    """Duplicate text from AgentResponse snapshot is not re-emitted."""
+
+    @executor(id="deduper")
+    async def deduper(message: Any, ctx: WorkflowContext[Any, Any]) -> None:
+        text = "Order processed successfully."
+        await ctx.yield_output(text)
+        # Snapshot repeats the same text
+        await ctx.yield_output(
+            AgentResponse(
+                messages=[
+                    Message(role="user", contents=[Content.from_text("process order")]),
+                    Message(role="assistant", contents=[Content.from_text(text)]),
+                ]
+            )
+        )
+
+    workflow = WorkflowBuilder(start_executor=deduper).build()
+    wrapper = AgentFrameworkWorkflow(workflow=workflow)
+    stream = await _run(wrapper, _payload())
+
+    stream.assert_text_messages_balanced()
+    deltas = [e.delta for e in stream.get("TEXT_MESSAGE_CONTENT")]
+    # Text should appear only once
+    assert deltas == ["Order processed successfully."]
+
+
+async def test_workflow_skips_consecutive_duplicate_outputs() -> None:
+    """Consecutive identical text outputs are deduplicated."""
+
+    @executor(id="repeater")
+    async def repeater(message: Any, ctx: WorkflowContext[Any, Any]) -> None:
+        text = "Done!"
+        await ctx.yield_output(text)
+        await ctx.yield_output(text)
+
+    workflow = WorkflowBuilder(start_executor=repeater).build()
+    wrapper = AgentFrameworkWorkflow(workflow=workflow)
+    stream = await _run(wrapper, _payload())
+
+    stream.assert_text_messages_balanced()
+    deltas = [e.delta for e in stream.get("TEXT_MESSAGE_CONTENT")]
+    assert deltas == ["Done!"]
+
+
+async def test_workflow_emits_distinct_consecutive_outputs() -> None:
+    """Distinct text outputs are all emitted, not incorrectly deduplicated."""
+
+    @executor(id="multisayer")
+    async def multisayer(message: Any, ctx: WorkflowContext[Any, str]) -> None:
+        await ctx.yield_output("First part. ")
+        await ctx.yield_output("Second part.")
+
+    workflow = WorkflowBuilder(start_executor=multisayer).build()
+    wrapper = AgentFrameworkWorkflow(workflow=workflow)
+    stream = await _run(wrapper, _payload())
+
+    stream.assert_text_messages_balanced()
+    deltas = [e.delta for e in stream.get("TEXT_MESSAGE_CONTENT")]
+    assert deltas == ["First part. ", "Second part."]
+
+
+# ──────────────────────────────────────────────────────────────────────
+# 10. Workflow error handling → RUN_ERROR
+# ──────────────────────────────────────────────────────────────────────
+
+
+async def test_workflow_error_emits_run_error_event() -> None:
+    """Exceptions during workflow streaming produce RUN_ERROR events."""
+
+    class FailingWorkflow:
+        def run(self, **kwargs: Any):
+            async def _stream():
+                raise RuntimeError("workflow exploded")
+                yield  # pragma: no cover
+
+            return _stream()
+
+    wrapper = AgentFrameworkWorkflow(workflow=cast(Any, FailingWorkflow()))
+    stream = await _run(wrapper, _payload())
+
+    # Should still have RUN_STARTED
+    stream.assert_has_type("RUN_STARTED")
+    # Should have RUN_ERROR
+    stream.assert_has_type("RUN_ERROR")
+    error = stream.first("RUN_ERROR")
+    assert "workflow exploded" in error.message
+
+
+async def test_workflow_error_preserves_bookend_structure() -> None:
+    """Even on error, RUN_STARTED is the first event."""
+
+    class FailingWorkflow:
+        def run(self, **kwargs: Any):
+            async def _stream():
+                raise ValueError("bad input")
+                yield  # pragma: no cover
+
+            return _stream()
+
+    wrapper = AgentFrameworkWorkflow(workflow=cast(Any, FailingWorkflow()))
+    stream = await _run(wrapper, _payload())
+
+    types = stream.types()
+    assert types[0] == "RUN_STARTED"
+    assert "RUN_ERROR" in types
+
+
+# ──────────────────────────────────────────────────────────────────────
+# 11. Multi-turn request_info interrupt/resume
+# ──────────────────────────────────────────────────────────────────────
+
+
+async def test_workflow_interrupt_resume_round_trip() -> None:
+    """Turn 1: request_info → interrupt. Turn 2: resume → completion."""
+
+    class RequesterExecutor(Executor):
+        def __init__(self) -> None:
+            super().__init__(id="requester")
+
+        @handler
+        async def start(self, message: Any, ctx: WorkflowContext) -> None:
+            await ctx.request_info("Choose an option", str, request_id="choice-1")
+
+        @response_handler
+        async def handle_choice(self, original: str, response: str, ctx: WorkflowContext) -> None:
+            await ctx.yield_output(f"You chose: {response}")  # type: ignore[arg-type]  # pyrefly: ignore[bad-argument-type]  # ty: ignore[invalid-argument-type]
+
+    workflow = WorkflowBuilder(start_executor=RequesterExecutor()).build()
+    wrapper = AgentFrameworkWorkflow(workflow=workflow)
+
+    # Turn 1
+    stream1 = await _run(wrapper, _payload(thread_id="thread-resume", run_id="run-1"))
+    stream1.assert_bookends()
+    stream1.assert_no_run_error()
+    stream1.assert_tool_calls_balanced()
+
+    interrupt1 = stream1.run_finished_interrupts()
+    assert interrupt1, "Expected interrupt"
+    assert interrupt1[0]["id"] == "choice-1"
+
+    # Turn 2: resume
+    stream2 = await _run(
+        wrapper,
+        {
+            "thread_id": "thread-resume",
+            "run_id": "run-2",
+            "messages": [],
+            "resume": {"interrupts": [{"id": "choice-1", "value": "Option A"}]},
+        },
+    )
+    stream2.assert_has_run_lifecycle()
+    stream2.assert_no_run_error()
+    stream2.assert_text_messages_balanced()
+
+    # Should have the response text
+    deltas = [e.delta for e in stream2.get("TEXT_MESSAGE_CONTENT")]
+    assert any("Option A" in d for d in deltas), f"Expected 'Option A' in deltas: {deltas}"
+
+    # No interrupt after resume
+    finished2 = stream2.last("RUN_FINISHED")
+    dumped2 = finished2.model_dump(by_alias=True, exclude_none=True)
+    assert "outcome" not in dumped2
+
+
+async def test_workflow_forwarded_props_resume() -> None:
+    """forwarded_props.command.resume should resume with canonical resume entries."""
+
+    @executor(id="requester")
+    async def requester(message: Any, ctx: WorkflowContext) -> None:
+        await ctx.request_info({"options": [{"name": "A"}]}, dict, request_id="pick")
+
+    workflow = WorkflowBuilder(start_executor=requester).build()
+    wrapper = AgentFrameworkWorkflow(workflow=workflow)
+
+    # Turn 1
+    await _run(wrapper, _payload(thread_id="thread-fwd", run_id="run-1"))
+
+    # Turn 2 via forwarded_props
+    stream2 = await _run(
+        wrapper,
+        {
+            "thread_id": "thread-fwd",
+            "run_id": "run-2",
+            "messages": [],
+            "forwarded_props": {
+                "command": {"resume": [{"interruptId": "pick", "status": "resolved", "payload": {"name": "A"}}]}
+            },
+        },
+    )
+    stream2.assert_bookends()
+    stream2.assert_no_run_error()
+
+    finished = stream2.last("RUN_FINISHED")
+    assert "outcome" not in finished.model_dump(by_alias=True, exclude_none=True)
+
+
+# ──────────────────────────────────────────────────────────────────────
+# 12. Empty turns with pending requests
+# ──────────────────────────────────────────────────────────────────────
+
+
+async def test_workflow_empty_turn_with_pending_request_emits_run_error() -> None:
+    """An empty turn with a pending request must provide canonical resume entries."""
+
+    @executor(id="requester")
+    async def requester(message: Any, ctx: WorkflowContext) -> None:
+        await ctx.request_info({"prompt": "choose"}, dict, request_id="pick-one")
+
+    workflow = WorkflowBuilder(start_executor=requester).build()
+    wrapper = AgentFrameworkWorkflow(workflow=workflow)
+
+    # Turn 1: trigger the request
+    await _run(wrapper, _payload(thread_id="thread-empty", run_id="run-1"))
+
+    # Turn 2: empty messages, no resume
+    stream2 = await _run(
+        wrapper,
+        {
+            "thread_id": "thread-empty",
+            "run_id": "run-2",
+            "messages": [],
+        },
+    )
+    stream2.assert_has_type("RUN_STARTED")
+    run_error = stream2.last("RUN_ERROR")
+    assert run_error.code == "WORKFLOW_RESUME_REQUIRED"
+
+
+async def test_workflow_empty_turn_no_pending_requests() -> None:
+    """Empty turn with no pending requests produces clean bookends."""
+
+    @executor(id="noop")
+    async def noop(message: Any, ctx: WorkflowContext[Any, str]) -> None:
+        await ctx.yield_output("done")
+
+    workflow = WorkflowBuilder(start_executor=noop).build()
+    wrapper = AgentFrameworkWorkflow(workflow=workflow)
+
+    # Run once to completion
+    await _run(wrapper, _payload(thread_id="thread-empty-clean", run_id="run-1"))
+
+    # Empty turn
+    stream2 = await _run(
+        wrapper,
+        {
+            "thread_id": "thread-empty-clean",
+            "run_id": "run-2",
+            "messages": [],
+        },
+    )
+    stream2.assert_bookends()
+    stream2.assert_no_run_error()
+
+
+# ──────────────────────────────────────────────────────────────────────
+# 13. Usage content as CUSTOM event
+# ──────────────────────────────────────────────────────────────────────
+
+
+async def test_workflow_usage_output_maps_to_custom_event() -> None:
+    """Usage Content outputs are surfaced as custom usage events."""
+
+    @executor(id="usage_reporter")
+    async def usage_reporter(message: Any, ctx: WorkflowContext[Any, Content]) -> None:
+        await ctx.yield_output(
+            Content.from_usage({"input_token_count": 100, "output_token_count": 50, "total_token_count": 150})
+        )
+
+    workflow = WorkflowBuilder(start_executor=usage_reporter).build()
+    wrapper = AgentFrameworkWorkflow(workflow=workflow)
+    stream = await _run(wrapper, _payload())
+
+    stream.assert_bookends()
+    stream.assert_no_run_error()
+
+    usage_events = [e for e in stream.get("CUSTOM") if getattr(e, "name", None) == "usage"]
+    assert len(usage_events) == 1
+    assert usage_events[0].value["input_token_count"] == 100
+    assert usage_events[0].value["total_token_count"] == 150
+
+
+# ──────────────────────────────────────────────────────────────────────
+# 14. Approval flow (Content-based request_info)
+# ──────────────────────────────────────────────────────────────────────
+
+
+async def test_workflow_approval_flow_round_trip() -> None:
+    """function_approval_request via request_info, then resume with approval response."""
+
+    class ApprovalExecutor(Executor):
+        def __init__(self) -> None:
+            super().__init__(id="approval_exec")
+
+        @handler
+        async def start(self, message: Any, ctx: WorkflowContext) -> None:
+            function_call = Content.from_function_call(
+                call_id="refund-call",
+                name="submit_refund",
+                arguments={"order_id": "12345", "amount": "$89.99"},
+            )
+            approval_request = Content.from_function_approval_request(id="approval-1", function_call=function_call)
+            await ctx.request_info(approval_request, Content, request_id="approval-1")
+
+        @response_handler
+        async def handle_approval(self, original_request: Content, response: Content, ctx: WorkflowContext) -> None:
+            status = "approved" if bool(response.approved) else "rejected"
+            await ctx.yield_output(f"Refund {status}.")  # type: ignore[arg-type]  # pyrefly: ignore[bad-argument-type]  # ty: ignore[invalid-argument-type]
+
+    workflow = WorkflowBuilder(start_executor=ApprovalExecutor()).build()
+    wrapper = AgentFrameworkWorkflow(workflow=workflow)
+
+    # Turn 1: request approval
+    stream1 = await _run(wrapper, _payload(thread_id="thread-approval", run_id="run-1"))
+    stream1.assert_bookends()
+    stream1.assert_no_run_error()
+
+    interrupt1 = stream1.run_finished_interrupts()
+    assert interrupt1, "Expected approval interrupt"
+    interrupt_value = stream1.interrupt_metadata_value(interrupt1[0])
+
+    # Turn 2: approve
+    stream2 = await _run(
+        wrapper,
+        {
+            "thread_id": "thread-approval",
+            "run_id": "run-2",
+            "messages": [],
+            "resume": {
+                "interrupts": [
+                    {
+                        "id": "approval-1",
+                        "value": {
+                            "type": "function_approval_response",
+                            "approved": True,
+                            "id": interrupt_value.get("id", "approval-1"),
+                            "function_call": interrupt_value.get("function_call"),
+                        },
+                    }
+                ]
+            },
+        },
+    )
+    stream2.assert_has_run_lifecycle()
+    stream2.assert_no_run_error()
+    stream2.assert_text_messages_balanced()
+
+    deltas = [e.delta for e in stream2.get("TEXT_MESSAGE_CONTENT")]
+    assert any("approved" in d for d in deltas)
+
+    # No more interrupt
+    finished2 = stream2.last("RUN_FINISHED")
+    assert "outcome" not in finished2.model_dump(by_alias=True, exclude_none=True)
+
+
+# ──────────────────────────────────────────────────────────────────────
+# 15. Message list request/response coercion
+# ──────────────────────────────────────────────────────────────────────
+
+
+async def test_workflow_message_list_resume() -> None:
+    """Resume with list[Message] payload coerces correctly into workflow response."""
+
+    class MessageRequestExecutor(Executor):
+        def __init__(self) -> None:
+            super().__init__(id="msg_request")
+
+        @handler
+        async def start(self, message: Any, ctx: WorkflowContext) -> None:
+            await ctx.request_info({"prompt": "Need follow-up"}, list[Message], request_id="handoff")
+
+        @response_handler
+        async def handle_input(self, original: dict, response: list[Message], ctx: WorkflowContext) -> None:
+            user_text = response[0].text if response else ""
+            await ctx.yield_output(f"Got: {user_text}")  # type: ignore[arg-type]  # pyrefly: ignore[bad-argument-type]  # ty: ignore[invalid-argument-type]
+
+    workflow = WorkflowBuilder(start_executor=MessageRequestExecutor()).build()
+    wrapper = AgentFrameworkWorkflow(workflow=workflow)
+
+    # Turn 1
+    await _run(wrapper, _payload(thread_id="thread-msg", run_id="run-1"))
+
+    # Turn 2: resume with message list
+    stream2 = await _run(
+        wrapper,
+        {
+            "thread_id": "thread-msg",
+            "run_id": "run-2",
+            "messages": [],
+            "resume": {
+                "interrupts": [
+                    {
+                        "id": "handoff",
+                        "value": [
+                            {"role": "user", "contents": [{"type": "text", "text": "Ship a replacement"}]},
+                        ],
+                    }
+                ]
+            },
+        },
+    )
+    stream2.assert_has_run_lifecycle()
+    stream2.assert_no_run_error()
+    stream2.assert_text_messages_balanced()
+
+    deltas = [e.delta for e in stream2.get("TEXT_MESSAGE_CONTENT")]
+    assert any("replacement" in d for d in deltas)
+
+
+# ──────────────────────────────────────────────────────────────────────
+# 16. Plain text follow-up does NOT infer interrupt response
+# ──────────────────────────────────────────────────────────────────────
+
+
+async def test_workflow_plain_text_does_not_resume_pending_dict_request() -> None:
+    """Plain text user follow-up should fail instead of being coerced into a dict response."""
+
+    @executor(id="requester")
+    async def requester(message: Any, ctx: WorkflowContext) -> None:
+        await ctx.request_info(
+            {"message": "Choose a flight", "options": [{"airline": "KLM"}], "agent": "flights"},
+            dict,
+            request_id="flights-choice",
+        )
+
+    workflow = WorkflowBuilder(start_executor=requester).build()
+    wrapper = AgentFrameworkWorkflow(workflow=workflow)
+
+    # Turn 1
+    await _run(wrapper, _payload(thread_id="thread-nocoerce", run_id="run-1"))
+
+    # Turn 2: plain text follow-up with request_info tool call in history
+    stream2 = await _run(
+        wrapper,
+        {
+            "thread_id": "thread-nocoerce",
+            "run_id": "run-2",
+            "messages": [
+                {
+                    "role": "assistant",
+                    "content": "",
+                    "tool_calls": [
+                        {
+                            "id": "flights-choice",
+                            "type": "function",
+                            "function": {"name": "request_info", "arguments": "{}"},
+                        }
+                    ],
+                },
+                {"role": "user", "content": "I prefer KLM please"},
+            ],
+        },
+    )
+    stream2.assert_has_type("RUN_STARTED")
+    run_error = stream2.last("RUN_ERROR")
+    assert run_error.code == "WORKFLOW_RESUME_REQUIRED"
+
+
+# ──────────────────────────────────────────────────────────────────────
+# 17. Workflow factory (thread-scoped workflows)
+# ──────────────────────────────────────────────────────────────────────
+
+
+async def test_workflow_factory_thread_scoping() -> None:
+    """workflow_factory creates separate workflow instances per thread_id."""
+
+    def make_workflow(thread_id: str):
+        @executor(id="echo")
+        async def echo(message: Any, ctx: WorkflowContext[Any, str]) -> None:
+            await ctx.yield_output(f"Thread: {thread_id}")
+
+        return WorkflowBuilder(start_executor=echo).build()
+
+    wrapper = AgentFrameworkWorkflow(workflow_factory=make_workflow)
+
+    stream_a = await _run(wrapper, _payload(thread_id="thread-a", run_id="run-a"))
+    stream_b = await _run(wrapper, _payload(thread_id="thread-b", run_id="run-b"))
+
+    stream_a.assert_bookends()
+    stream_b.assert_bookends()
+
+    deltas_a = [e.delta for e in stream_a.get("TEXT_MESSAGE_CONTENT")]
+    deltas_b = [e.delta for e in stream_b.get("TEXT_MESSAGE_CONTENT")]
+    assert any("thread-a" in d for d in deltas_a)
+    assert any("thread-b" in d for d in deltas_b)
+
+
+# ──────────────────────────────────────────────────────────────────────
+# 18. Multiple request_info calls in sequence
+# ──────────────────────────────────────────────────────────────────────
+
+
+async def test_workflow_sequential_request_info_interrupts() -> None:
+    """Two chained executors each requesting info: first triggers interrupt, resume, then second triggers interrupt.
+
+    This mirrors the subgraphs_agent pattern where separate executors handle sequential interactions.
+    """
+
+    class NameRequester(Executor):
+        def __init__(self) -> None:
+            super().__init__(id="name_requester")
+
+        @handler
+        async def start(self, message: Any, ctx: WorkflowContext[str]) -> None:
+            await ctx.request_info("What's your name?", str, request_id="name-req")
+
+        @response_handler
+        async def handle_name(self, original: str, response: str, ctx: WorkflowContext[str]) -> None:
+            await ctx.send_message(response)
+
+    class DestRequester(Executor):
+        def __init__(self) -> None:
+            super().__init__(id="dest_requester")
+
+        @handler
+        async def start(self, message: str, ctx: WorkflowContext[str]) -> None:
+            self._name = message
+            await ctx.request_info("Where to?", str, request_id="dest-req")
+
+        @response_handler
+        async def handle_dest(self, original: str, response: str, ctx: WorkflowContext[str]) -> None:
+            await ctx.yield_output(f"Booking for {self._name} to {response}")  # type: ignore[arg-type]  # pyrefly: ignore[bad-argument-type]  # ty: ignore[invalid-argument-type]
+
+    name_requester = NameRequester()
+    dest_requester = DestRequester()
+    workflow = WorkflowBuilder(start_executor=name_requester).add_chain([name_requester, dest_requester]).build()
+    wrapper = AgentFrameworkWorkflow(workflow=workflow)
+
+    # Turn 1
+    stream1 = await _run(wrapper, _payload(thread_id="thread-seq", run_id="run-1"))
+    stream1.assert_bookends()
+    stream1.assert_tool_calls_balanced()
+    interrupt1 = stream1.run_finished_interrupts()
+    assert interrupt1[0]["id"] == "name-req"
+
+    # Turn 2: answer name → triggers second executor's request_info
+    stream2 = await _run(
+        wrapper,
+        {
+            "thread_id": "thread-seq",
+            "run_id": "run-2",
+            "messages": [],
+            "resume": {"interrupts": [{"id": "name-req", "value": "Alice"}]},
+        },
+    )
+    stream2.assert_has_run_lifecycle()
+    stream2.assert_tool_calls_balanced()
+    interrupt2 = stream2.run_finished_interrupts()
+    assert interrupt2[0]["id"] == "dest-req"
+
+    # Turn 3: answer destination → completion
+    stream3 = await _run(
+        wrapper,
+        {
+            "thread_id": "thread-seq",
+            "run_id": "run-3",
+            "messages": [],
+            "resume": {"interrupts": [{"id": "dest-req", "value": "Paris"}]},
+        },
+    )
+    stream3.assert_has_run_lifecycle()
+    stream3.assert_no_run_error()
+    stream3.assert_text_messages_balanced()
+
+    deltas = [e.delta for e in stream3.get("TEXT_MESSAGE_CONTENT")]
+    assert any("Alice" in d and "Paris" in d for d in deltas)
+    assert "outcome" not in stream3.last("RUN_FINISHED").model_dump(by_alias=True, exclude_none=True)
				`@@ -0,0 +1 @@`
				`# Copyright (c) Microsoft. All rights reserved.`