[Core][Frontend] Unified Full-duplex Framework (#7413)

Signed-off-by: chickeyton <ngton2014@gmail.com>
Co-authored-by: Claude Opus 5 (1M context) <noreply@anthropic.com>

tests/entrypoints/openai/test_duplex_session_attachment.py

@@ -133,7 +133,7 @@ async def test_registry_failed_send_is_accepted_only_when_journaled(mocker, jour
         assert accepted.call_args_list == expected_calls
         raise RuntimeError("transport lost")
 
-    await registry.create("sid-send-failure", incarnation=0, send=failing_send, close=mocker.AsyncMock())
+    await registry.create("sid-send-failure", send=failing_send, close=mocker.AsyncMock())
 
     with pytest.raises(RuntimeError, match="transport lost"):
         await registry.send_event(
@@ -152,7 +152,7 @@ async def test_registry_detached_event_is_accepted_only_when_journaled(mocker, j
     registry = DuplexSessionAttachmentRegistry(replay_ttl_s=60.0, replay_max_bytes_per_session=4096)
     send = mocker.AsyncMock()
     accepted = mocker.Mock()
-    created = await registry.create("sid-detached-acceptance", incarnation=0, send=send, close=mocker.AsyncMock())
+    created = await registry.create("sid-detached-acceptance", send=send, close=mocker.AsyncMock())
     await registry.detach("sid-detached-acceptance", attachment_generation=created.attachment_generation)
 
     entry = await registry.send_event(
@@ -175,7 +175,7 @@ async def test_registry_unjournaled_event_is_accepted_after_successful_send(mock
     async def send(payload):
         accepted.assert_not_called()
 
-    await registry.create("sid-live-acceptance", incarnation=0, send=send, close=mocker.AsyncMock())
+    await registry.create("sid-live-acceptance", send=send, close=mocker.AsyncMock())
 
     entry = await registry.send_event(
         "sid-live-acceptance",
@@ -193,7 +193,7 @@ async def test_registry_overflow_does_not_accept_or_send_event(mocker) -> None:
     registry = DuplexSessionAttachmentRegistry(replay_ttl_s=60.0, replay_max_bytes_per_session=100)
     accepted = mocker.Mock()
     send = mocker.AsyncMock()
-    await registry.create("sid-overflow-acceptance", incarnation=0, send=send, close=mocker.AsyncMock())
+    await registry.create("sid-overflow-acceptance", send=send, close=mocker.AsyncMock())
 
     with pytest.raises(DuplexJournalOverflowError, match="byte limit"):
         await registry.send_event(
@@ -231,13 +231,12 @@ async def test_registry_resume_rotates_token_replays_and_atomically_replaces_att
     async def close_b(reason):
         closes_b.append(reason)
 
-    created = await registry.create("sid", incarnation=2, send=send_a, close=close_a)
+    created = await registry.create("sid", send=send_a, close=close_a)
     await registry.send_event("sid", {"type": "event-1"})
     await registry.send_event("sid", {"type": "event-2"})
 
     resumed = await registry.resume(
         "sid",
-        incarnation=2,
         resume_token=created.resume_token.plaintext,
         last_received_server_event_seq=1,
         send=send_b,
@@ -260,6 +259,34 @@ async def test_registry_resume_rotates_token_replays_and_atomically_replaces_att
     assert closes_a == [] and sends_b == [] and closes_b == []
 
 
+@pytest.mark.asyncio
+async def test_registry_detach_without_a_generation_targets_the_current_attachment_once() -> None:
+    registry = DuplexSessionAttachmentRegistry(replay_ttl_s=60.0, replay_max_bytes_per_session=4096)
+    sends: list[dict] = []
+
+    async def send(payload):
+        sends.append(payload)
+
+    async def close(reason):
+        return None
+
+    await registry.create("sid-current", send=send, close=close)
+
+    # The outbound pump outlives any one connection and only knows the session:
+    # ``None`` means "whichever socket is attached right now".
+    assert await registry.detach("sid-current", attachment_generation=None) is True
+    # A second report of the same dead socket is not a new detach, so it cannot
+    # restart the engine's disconnect grace window.
+    assert await registry.detach("sid-current", attachment_generation=None) is False
+    assert await registry.detach("sid-current", attachment_generation=1) is False
+    assert await registry.is_current_attachment("sid-current", 1) is False
+
+    # A detached session still journals for a later resume.
+    await registry.send_event("sid-current", {"type": "event-1"})
+    assert sends == []
+    assert await registry.detach("missing-session", attachment_generation=None) is False
+
+
 @pytest.mark.asyncio
 async def test_registry_resume_sends_activation_then_replay_before_new_live_events() -> None:
     registry = DuplexSessionAttachmentRegistry(
@@ -282,13 +309,12 @@ async def test_registry_resume_sends_activation_then_replay_before_new_live_even
             activation_started.set()
             await release_activation.wait()
 
-    created = await registry.create("sid-order", incarnation=0, send=old_send, close=close)
+    created = await registry.create("sid-order", send=old_send, close=close)
     await registry.detach("sid-order", attachment_generation=1)
     await registry.send_event("sid-order", {"type": "replayed"})
     resume_task = asyncio.create_task(
         registry.resume(
             "sid-order",
-            incarnation=0,
             resume_token=created.resume_token.plaintext,
             last_received_server_event_seq=0,
             send=new_send,
@@ -322,13 +348,12 @@ async def test_registry_resume_delivery_failure_keeps_one_shot_old_token_recover
     async def close(reason):
         del reason
 
-    created = await registry.create("sid-recovery", incarnation=0, send=send, close=close)
+    created = await registry.create("sid-recovery", send=send, close=close)
     await registry.detach("sid-recovery", attachment_generation=1)
 
     with pytest.raises(RuntimeError, match="transport lost"):
         await registry.resume(
             "sid-recovery",
-            incarnation=0,
             resume_token=created.resume_token.plaintext,
             last_received_server_event_seq=0,
             send=failing_send,
@@ -342,7 +367,6 @@ async def test_registry_resume_delivery_failure_keeps_one_shot_old_token_recover
 
     recovered = await registry.resume(
         "sid-recovery",
-        incarnation=0,
         resume_token=created.resume_token.plaintext,
         last_received_server_event_seq=0,
         send=send,
@@ -353,7 +377,6 @@ async def test_registry_resume_delivery_failure_keeps_one_shot_old_token_recover
     with pytest.raises(InvalidResumeTokenError):
         await registry.resume(
             "sid-recovery",
-            incarnation=0,
             resume_token=created.resume_token.plaintext,
             last_received_server_event_seq=0,
             send=send,
@@ -374,13 +397,12 @@ async def test_registry_concurrent_resume_allows_exactly_one_rotated_token_winne
     async def close(reason):
         del reason
 
-    created = await registry.create("sid-race", incarnation=0, send=send, close=close)
+    created = await registry.create("sid-race", send=send, close=close)
 
     async def attempt():
         try:
             return await registry.resume(
                 "sid-race",
-                incarnation=0,
                 resume_token=created.resume_token.plaintext,
                 last_received_server_event_seq=0,
                 send=send,
@@ -396,7 +418,7 @@ async def test_registry_concurrent_resume_allows_exactly_one_rotated_token_winne
 
 
 @pytest.mark.asyncio
-async def test_registry_keeps_sessions_journals_and_incarnations_isolated() -> None:
+async def test_registry_keeps_sessions_journals_and_tokens_isolated() -> None:
     registry = DuplexSessionAttachmentRegistry(
         replay_ttl_s=60.0,
         replay_max_bytes_per_session=4096,
@@ -408,8 +430,8 @@ async def test_registry_keeps_sessions_journals_and_incarnations_isolated() -> N
     async def close(reason):
         del reason
 
-    created_a = await registry.create("sid-a", incarnation=1, send=send, close=close)
-    created_b = await registry.create("sid-b", incarnation=4, send=send, close=close)
+    created_a = await registry.create("sid-a", send=send, close=close)
+    created_b = await registry.create("sid-b", send=send, close=close)
     await registry.detach("sid-a", attachment_generation=1)
     await registry.detach("sid-b", attachment_generation=1)
     await registry.send_event("sid-a", {"type": "a-only"})
@@ -417,7 +439,6 @@ async def test_registry_keeps_sessions_journals_and_incarnations_isolated() -> N
 
     resumed_a = await registry.resume(
         "sid-a",
-        incarnation=1,
         resume_token=created_a.resume_token.plaintext,
         last_received_server_event_seq=0,
         send=send,
@@ -425,7 +446,6 @@ async def test_registry_keeps_sessions_journals_and_incarnations_isolated() -> N
     )
     resumed_b = await registry.resume(
         "sid-b",
-        incarnation=4,
         resume_token=created_b.resume_token.plaintext,
         last_received_server_event_seq=0,
         send=send,
@@ -434,11 +454,10 @@ async def test_registry_keeps_sessions_journals_and_incarnations_isolated() -> N
 
     assert [entry.payload["type"] for entry in resumed_a.replay_entries] == ["a-only"]
     assert [entry.payload["type"] for entry in resumed_b.replay_entries] == ["b-only"]
-    with pytest.raises(ValueError, match="incarnation mismatch"):
+    with pytest.raises(InvalidResumeTokenError):
         await registry.resume(
             "sid-a",
-            incarnation=2,
-            resume_token=resumed_a.resume_token.plaintext,
+            resume_token=resumed_b.resume_token.plaintext,
             last_received_server_event_seq=1,
             send=send,
             close=close,
@@ -458,7 +477,7 @@ async def test_registry_repr_never_contains_plaintext_tokens() -> None:
     async def close(reason):
         del reason
 
-    created = await registry.create("sid-repr", incarnation=0, send=send, close=close)
+    created = await registry.create("sid-repr", send=send, close=close)
     await registry.detach("sid-repr", attachment_generation=1)
     entry = await registry.send_event(
         "sid-repr",
@@ -469,3 +488,67 @@ async def test_registry_repr_never_contains_plaintext_tokens() -> None:
     assert created.resume_token.plaintext not in repr(created)
     assert created.resume_token.plaintext not in repr(entry)
     assert created.resume_token.plaintext not in repr(registry)
+
+
+@pytest.mark.asyncio
+@pytest.mark.parametrize("cancel_on", ["activation", "replay"])
+async def test_registry_resume_cancelled_mid_delivery_rolls_back_like_a_failure(cancel_on: str) -> None:
+    """Cancellation must take the same rollback as an ordinary delivery failure.
+
+    ``CancelledError`` derives from ``BaseException``, so an ``except
+    Exception`` around the activation and replay sends let a cancelled resume
+    keep its attachment as the current generation while the token it rotated
+    past was already gone — the session left attached to an unusable transport
+    with no way back. Both send points are covered because either can be the
+    one that gets cancelled.
+    """
+    registry = DuplexSessionAttachmentRegistry(replay_ttl_s=60.0, replay_max_bytes_per_session=4096)
+
+    async def send(payload):
+        del payload
+
+    async def close(reason):
+        del reason
+
+    created = await registry.create("sid-cancel", send=send, close=close)
+    # One journaled event, so the replay loop has something to send.
+    await registry.send_event("sid-cancel", {"type": "response.output_text.delta", "delta": "a"})
+    await registry.detach("sid-cancel", attachment_generation=1)
+
+    sends = 0
+    cancel_at = 1 if cancel_on == "activation" else 2
+
+    async def cancelling_send(payload):
+        del payload
+        nonlocal sends
+        sends += 1
+        if sends == cancel_at:
+            raise asyncio.CancelledError
+
+    with pytest.raises(asyncio.CancelledError):
+        await registry.resume(
+            "sid-cancel",
+            resume_token=created.resume_token.plaintext,
+            last_received_server_event_seq=0,
+            send=cancelling_send,
+            close=close,
+            activation_payload_factory=lambda token, generation: {
+                "type": "session.resumed",
+                "resume_token": token.plaintext,
+                "attachment_generation": generation,
+            },
+        )
+
+    assert sends == cancel_at, "the send that was cancelled is the one under test"
+    # The cancelled attachment must not remain current...
+    assert not await registry.has_attachment("sid-cancel")
+    # ...and the token the caller presented has to work again, or the session
+    # is unreachable: resume() had already rotated past it.
+    recovered = await registry.resume(
+        "sid-cancel",
+        resume_token=created.resume_token.plaintext,
+        last_received_server_event_seq=0,
+        send=send,
+        close=close,
+    )
+    assert recovered.attachment_generation == 3