chore: merge master into dev/4.11.x

tests/integration/api/test_plugins_security.py

@@ -106,7 +106,12 @@ async def plugin_security_api(plugin_module):
     application.plugin_connector.require_workspace_context = AsyncMock()
     application.plugin_connector.list_plugins = AsyncMock(return_value=[raw_plugin])
     application.plugin_connector.get_plugin_info = AsyncMock(return_value=raw_plugin)
-    application.plugin_connector.get_debug_info = AsyncMock(return_value={'plugin_debug_key': 'runtime-debug-secret'})
+    application.plugin_connector.get_debug_info = AsyncMock(
+        return_value={
+            'plugin_debug_key': 'runtime-debug-secret',
+            'expires_at': '2026-08-04T12:00:00Z',
+        }
+    )
     application.plugin_connector.get_plugin_logs = AsyncMock(return_value=['private runtime line'])
     application.plugin_connector.set_plugin_config = AsyncMock()
 
@@ -230,10 +235,22 @@ async def test_debug_key_requires_resource_manage_permission(plugin_security_api
     assert operator_denied.status_code == 403
     assert allowed.status_code == 200
     assert (await allowed.get_json())['data'] == {
-        'debug_url': 'http://localhost:5401',
+        'debug_url': 'ws://localhost:5401/plugin/debug/ws',
         'plugin_debug_key': 'runtime-debug-secret',
+        'expires_at': '2026-08-04T12:00:00Z',
     }
-    application.plugin_connector.get_debug_info.assert_awaited_once_with()
+    application.plugin_connector.get_debug_info.assert_awaited_once()
+
+
+@pytest.mark.asyncio
+async def test_debug_info_uses_websocket_endpoint_for_legacy_config(plugin_security_api):
+    application, client, _ = plugin_security_api
+    application.instance_config.data['plugin'].pop('display_plugin_debug_url')
+
+    response = await client.get('/api/v1/plugins/debug-info', headers=_headers('manager-token'))
+
+    assert response.status_code == 200
+    assert (await response.get_json())['data']['debug_url'] == 'ws://localhost:5401/plugin/debug/ws'
 
 
 @pytest.mark.asyncio

tests/integration/api/test_user_space_oauth.py

@@ -2,6 +2,7 @@
 
 from __future__ import annotations
 
+import datetime
 from types import SimpleNamespace
 from unittest.mock import AsyncMock, Mock
 from urllib.parse import parse_qs, urlsplit
@@ -14,6 +15,7 @@ from langbot.pkg.api.http.controller.groups.user import UserRouterGroup
 
 pytestmark = pytest.mark.integration
 WORKSPACE_UUID = '11111111-1111-4111-8111-111111111111'
+WORKSPACE_CREATED_AT = datetime.datetime(2026, 1, 2, 3, 4, 5, tzinfo=datetime.UTC)
 
 
 @pytest.fixture
@@ -58,6 +60,12 @@ async def space_oauth_api():
         return_value={'account_uuid': 'account-a', 'workspace_uuid': WORKSPACE_UUID}
     )
     application.workspace_collaboration_service.resolve_account_workspace = AsyncMock(return_value=access)
+    application.workspace_service.get_execution_binding = AsyncMock(
+        return_value=SimpleNamespace(
+            workspace_uuid=WORKSPACE_UUID,
+            workspace_created_at=WORKSPACE_CREATED_AT,
+        )
+    )
     application.space_service.get_oauth_authorize_url = Mock(
         side_effect=lambda redirect_uri, state: f'https://space.example/authorize?state={state}'
     )
@@ -157,34 +165,51 @@ async def test_bind_state_is_account_bound_and_requires_authentication(space_oau
 
 
 @pytest.mark.asyncio
-async def test_redirect_origin_and_callback_path_are_restricted(space_oauth_api):
+async def test_redirect_allows_any_http_or_https_origin(space_oauth_api):
     _, client = space_oauth_api
 
-    wrong_origin = await client.get(
-        '/api/v1/user/space/authorize-url',
-        query_string={'redirect_uri': 'https://evil.example/auth/space/callback'},
-        headers={'Origin': 'http://localhost'},
-    )
-    wrong_path = await client.get(
-        '/api/v1/user/space/authorize-url',
-        query_string={'redirect_uri': 'http://localhost/arbitrary'},
-        headers={'Origin': 'http://localhost'},
-    )
-    forged_origin = await client.get(
-        '/api/v1/user/space/authorize-url',
-        query_string={'redirect_uri': 'https://evil.example/auth/space/callback'},
-        headers={'Origin': 'https://evil.example'},
-    )
-    forged_host = await client.get(
-        '/api/v1/user/space/authorize-url',
-        query_string={'redirect_uri': 'https://evil.example/auth/space/callback'},
-        headers={'Host': 'evil.example'},
-    )
+    responses = [
+        await client.get(
+            '/api/v1/user/space/authorize-url',
+            query_string={'redirect_uri': redirect_uri},
+            headers={'Origin': 'https://irrelevant.example'},
+        )
+        for redirect_uri in (
+            'https://langbot.example/auth/space/callback',
+            'https://gateway.example:8443/auth/space/callback',
+            'https://192.0.2.10/auth/space/callback',
+            'http://localhost:5300/auth/space/callback',
+            'http://127.0.0.1:5300/auth/space/callback',
+            'http://[::1]:5300/auth/space/callback',
+            'http://langbot.example/auth/space/callback',
+            'http://192.0.2.10:5300/auth/space/callback',
+        )
+    ]
 
-    assert (await wrong_origin.get_json())['code'] == 1
-    assert (await wrong_path.get_json())['code'] == 1
-    assert (await forged_origin.get_json())['code'] == 1
-    assert (await forged_host.get_json())['code'] == 1
+    assert all(response.status_code == 200 for response in responses)
+    payloads = [await response.get_json() for response in responses]
+    assert all(payload['code'] == 0 for payload in payloads)
+
+
+@pytest.mark.asyncio
+async def test_redirect_rejects_invalid_callback_shape(space_oauth_api):
+    _, client = space_oauth_api
+
+    responses = [
+        await client.get(
+            '/api/v1/user/space/authorize-url',
+            query_string={'redirect_uri': redirect_uri},
+        )
+        for redirect_uri in (
+            'https://langbot.example/arbitrary',
+            'https://langbot.example/auth/space/callback?next=https://evil.example',
+            'https://user@langbot.example/auth/space/callback',
+            'https://langbot.example/auth/space/callback#fragment',
+        )
+    ]
+
+    payloads = [await response.get_json() for response in responses]
+    assert all(payload['code'] == 1 for payload in payloads)
 
 
 @pytest.mark.asyncio
@@ -234,7 +259,11 @@ async def test_login_callback_requires_and_consumes_server_state(space_oauth_api
     assert response.status_code == 200
     assert (await response.get_json())['data']['token'] == 'space-login-token'
     application.user_service.consume_space_oauth_state_details.assert_awaited_once_with('opaque-login-state', 'login')
-    application.space_service.exchange_oauth_code.assert_awaited_once_with('oauth-code')
+    application.space_service.exchange_oauth_code.assert_awaited_once_with(
+        'oauth-code',
+        [WORKSPACE_UUID],
+        {WORKSPACE_UUID: int(WORKSPACE_CREATED_AT.timestamp())},
+    )
 
 
 @pytest.mark.asyncio

tests/integration/persistence/test_membership_source_migration.py

@@ -0,0 +1,70 @@
+from __future__ import annotations
+
+import pytest
+import sqlalchemy as sa
+from sqlalchemy.ext.asyncio import create_async_engine
+
+from langbot.pkg.persistence.alembic_runner import run_alembic_stamp, run_alembic_upgrade
+
+
+@pytest.mark.asyncio
+async def test_membership_source_migration_backfills_existing_rows_as_local_and_enforces_constraint(tmp_path):
+    engine = create_async_engine(f'sqlite+aiosqlite:///{tmp_path / "membership-source.db"}')
+    try:
+        async with engine.begin() as connection:
+            await connection.execute(
+                sa.text(
+                    """
+                    CREATE TABLE workspace_memberships (
+                        uuid VARCHAR(36) PRIMARY KEY,
+                        workspace_uuid VARCHAR(36) NOT NULL,
+                        account_uuid VARCHAR(36) NOT NULL,
+                        role VARCHAR(32) NOT NULL,
+                        status VARCHAR(32) NOT NULL,
+                        projection_revision BIGINT NOT NULL DEFAULT 0,
+                        created_at DATETIME NOT NULL DEFAULT CURRENT_TIMESTAMP,
+                        updated_at DATETIME NOT NULL DEFAULT CURRENT_TIMESTAMP
+                    )
+                    """
+                )
+            )
+            await connection.execute(
+                sa.text(
+                    """
+                    INSERT INTO workspace_memberships
+                        (uuid, workspace_uuid, account_uuid, role, status, projection_revision)
+                    VALUES
+                        ('00000000-0000-4000-8000-000000000001', 'workspace', 'local-account',
+                         'viewer', 'active', 0),
+                        ('00000000-0000-4000-8000-000000000002', 'workspace', 'cloud-account',
+                         'viewer', 'active', 0)
+                    """
+                )
+            )
+
+        await run_alembic_stamp(engine, '0019_single_workspace_owner')
+        await run_alembic_upgrade(engine, 'head')
+
+        async with engine.connect() as connection:
+            rows = (
+                await connection.execute(sa.text('SELECT uuid, source FROM workspace_memberships ORDER BY uuid'))
+            ).all()
+            columns = await connection.run_sync(
+                lambda sync_connection: {
+                    column['name']: column
+                    for column in sa.inspect(sync_connection).get_columns('workspace_memberships')
+                }
+            )
+        assert rows == [
+            ('00000000-0000-4000-8000-000000000001', 'local'),
+            ('00000000-0000-4000-8000-000000000002', 'local'),
+        ]
+        assert columns['source']['nullable'] is False
+
+        with pytest.raises(sa.exc.IntegrityError):
+            async with engine.begin() as connection:
+                await connection.execute(
+                    sa.text("UPDATE workspace_memberships SET source = 'guessed-from-user-source'")
+                )
+    finally:
+        await engine.dispose()

tests/integration/persistence/test_migrations.py

@@ -9,8 +9,11 @@ Run: uv run pytest tests/integration/persistence/test_migrations.py -q
 
 from __future__ import annotations
 
+import json
+
 import pytest
 import sqlalchemy
+from sqlalchemy import text
 from sqlalchemy.ext.asyncio import create_async_engine
 
 from langbot.pkg.entity.persistence.base import Base
@@ -105,7 +108,7 @@ class TestSQLiteMigrationUpgrade:
         await run_alembic_upgrade(sqlite_engine, 'head')
 
         assert await get_alembic_current(sqlite_engine) == _get_script_head()
-        assert _get_script_head() == '0020_merge_agent_cloud_heads'
+        assert _get_script_head() == '0022_merge_agent_reasoning_heads'
 
     @pytest.mark.asyncio
     async def test_upgrade_from_development_workspace_head_to_merged_head(self, sqlite_engine):
@@ -117,7 +120,18 @@ class TestSQLiteMigrationUpgrade:
         await run_alembic_upgrade(sqlite_engine, 'head')
 
         assert await get_alembic_current(sqlite_engine) == _get_script_head()
-        assert _get_script_head() == '0020_merge_agent_cloud_heads'
+        assert _get_script_head() == '0022_merge_agent_reasoning_heads'
+
+    @pytest.mark.asyncio
+    async def test_upgrade_from_reasoning_config_head_to_merged_head(self, sqlite_engine):
+        """A database that already ran the feature migration must remain upgradable."""
+        async with sqlite_engine.begin() as conn:
+            await conn.run_sync(Base.metadata.create_all)
+
+        await run_alembic_stamp(sqlite_engine, '0018_llm_reasoning_config')
+        await run_alembic_upgrade(sqlite_engine, 'head')
+
+        assert await get_alembic_current(sqlite_engine) == '0022_merge_agent_reasoning_heads'
 
     @pytest.mark.asyncio
     async def test_upgrade_from_baseline_to_head(self, sqlite_engine):
@@ -214,6 +228,66 @@ class TestSQLiteMigrationUpgrade:
         await run_alembic_upgrade(sqlite_engine, 'head')
         assert await get_alembic_current(sqlite_engine) == _get_script_head()
 
+    @pytest.mark.asyncio
+    async def test_reasoning_config_migrates_existing_models(self, sqlite_engine):
+        """Upgrade from 0017 backfills reasoning config and keeps a database default."""
+        async with sqlite_engine.begin() as conn:
+            await conn.execute(
+                text(
+                    """
+                    CREATE TABLE llm_models (
+                        uuid VARCHAR(255) PRIMARY KEY,
+                        name VARCHAR(255) NOT NULL,
+                        provider_uuid VARCHAR(255) NOT NULL,
+                        abilities JSON NOT NULL,
+                        context_length INTEGER,
+                        extra_args JSON NOT NULL,
+                        prefered_ranking INTEGER NOT NULL DEFAULT 0
+                    )
+                    """
+                )
+            )
+            await conn.execute(
+                text(
+                    """
+                    INSERT INTO llm_models (
+                        uuid, name, provider_uuid, abilities, extra_args, prefered_ranking
+                    ) VALUES (
+                        'existing-model', 'Existing Model', 'provider', '[]', '{}', 0
+                    )
+                    """
+                )
+            )
+
+        await run_alembic_stamp(sqlite_engine, '0017_oss_workspace_identity')
+        await run_alembic_upgrade(sqlite_engine, 'head')
+
+        async with sqlite_engine.begin() as conn:
+            columns = await conn.run_sync(lambda sync_conn: sqlalchemy.inspect(sync_conn).get_columns('llm_models'))
+            reasoning_column = next(column for column in columns if column['name'] == 'reasoning_config')
+            assert reasoning_column['nullable'] is False
+
+            existing_value = (
+                await conn.execute(text("SELECT reasoning_config FROM llm_models WHERE uuid = 'existing-model'"))
+            ).scalar_one()
+            assert json.loads(existing_value) == {'level': 'provider_default'}
+
+            await conn.execute(
+                text(
+                    """
+                    INSERT INTO llm_models (
+                        uuid, name, provider_uuid, abilities, extra_args, prefered_ranking
+                    ) VALUES (
+                        'new-model', 'New Model', 'provider', '[]', '{}', 0
+                    )
+                    """
+                )
+            )
+            new_value = (
+                await conn.execute(text("SELECT reasoning_config FROM llm_models WHERE uuid = 'new-model'"))
+            ).scalar_one()
+            assert json.loads(new_value) == {'level': 'provider_default'}
+
 
 class TestSQLiteMigrationFreshDatabase:
     """Tests for fresh database workflow."""

tests/integration/persistence/test_pgvector_postgres.py

@@ -85,6 +85,32 @@ async def clean_database(postgres_engine: AsyncEngine):
     await clean()
 
 
+async def test_upgrade_adds_3072_dimension_index_and_constraint(
+    postgres_engine: AsyncEngine,
+    clean_database,
+) -> None:
+    async with postgres_engine.begin() as conn:
+        await conn.execute(text('CREATE EXTENSION IF NOT EXISTS vector'))
+        await conn.run_sync(Base.metadata.create_all)
+    await run_alembic_stamp(postgres_engine, '0010_scope_resources')
+    await run_alembic_upgrade(postgres_engine, 'head')
+
+    async with postgres_engine.connect() as conn:
+        constraint = await conn.scalar(
+            text(
+                'SELECT pg_get_constraintdef(oid) FROM pg_constraint '
+                "WHERE conrelid = 'langbot_vectors'::regclass "
+                "AND conname = 'ck_langbot_vectors_embedding_dimension_enabled'"
+            )
+        )
+        assert '3072' in constraint
+        index_definition = await conn.scalar(
+            text("SELECT indexdef FROM pg_indexes WHERE indexname = 'ix_langbot_vectors_hnsw_cosine_3072'")
+        )
+        assert 'halfvec(3072)' in index_definition
+        assert 'halfvec_cosine_ops' in index_definition
+
+
 async def test_legacy_upgrade_temporarily_suspends_and_restores_source_rls_for_unprivileged_owner(
     postgres_url: str,
     postgres_engine: AsyncEngine,

tests/integration/persistence/test_release_migration_postgres.py

@@ -92,7 +92,7 @@ def _application(postgres_url: str, *, runtime_role: str = 'langbot_runtime_not_
                     'use': 'pgvector',
                     'pgvector': {
                         'use_business_database': True,
-                        'allowed_dimensions': [384, 512, 768, 1024, 1536],
+                        'allowed_dimensions': [384, 512, 768, 1024, 1536, 3072],
                     },
                 },
             }

tests/unit_tests/api/service/test_model_service.py

@@ -17,12 +17,14 @@ import pytest
 from unittest.mock import AsyncMock, Mock
 from types import SimpleNamespace
 
+from langbot.pkg.api.http.context import ExecutionContext
 from langbot.pkg.api.http.service.model import (
     LLMModelsService,
     EmbeddingModelsService,
     RerankModelsService,
     _parse_provider_api_keys,
     _runtime_model_data,
+    _serialize_llm_model,
     _validate_provider_supports,
 )
 from langbot.pkg.api.http.service import model as model_service_module
@@ -64,15 +66,19 @@ def _create_mock_llm_model(
     abilities: list = None,
     context_length: int | None = None,
     extra_args: dict = None,
+    reasoning_config: dict = None,
 ) -> Mock:
     """Helper to create mock LLMModel entity."""
     model = Mock(spec=LLMModel)
+    model.workspace_uuid = WORKSPACE_UUID
     model.uuid = model_uuid
     model.name = name
     model.provider_uuid = provider_uuid
     model.abilities = abilities or []
     model.context_length = context_length
     model.extra_args = extra_args or {}
+    model.reasoning_config = reasoning_config or {'level': 'provider_default'}
+    model.prefered_ranking = 0
     return model
 
 
@@ -156,6 +162,26 @@ def _create_runtime_model_mgr() -> SimpleNamespace:
     return manager
 
 
+def _create_reasoning_runtime_provider(capabilities: dict) -> SimpleNamespace:
+    execution_context = ExecutionContext(
+        instance_uuid='instance-test',
+        workspace_uuid=WORKSPACE_UUID,
+        placement_generation=1,
+    )
+    return SimpleNamespace(
+        execution_context=execution_context,
+        provider_entity=ModelProvider(
+            workspace_uuid=WORKSPACE_UUID,
+            uuid='provider-uuid',
+            name='Reasoning Provider',
+            requester='openai',
+            base_url='https://api.openai.com',
+            api_keys=[],
+        ),
+        requester=SimpleNamespace(get_reasoning_capabilities=Mock(return_value=capabilities)),
+    )
+
+
 class TestParseProviderApiKeys:
     """Tests for _parse_provider_api_keys helper function."""
 
@@ -209,6 +235,42 @@ class TestRuntimeModelData:
         assert result['extra_args'] == {'temp': 0.7}
 
 
+class TestSerializeLLMModel:
+    def test_includes_runtime_reasoning_capabilities(self):
+        model = _create_mock_llm_model(
+            abilities=['reasoning'],
+            reasoning_config={'level': 'high'},
+        )
+        capabilities = {
+            'supported': True,
+            'levels': ['provider_default', 'low', 'high'],
+            'source': 'litellm',
+        }
+        runtime_model = SimpleNamespace(
+            model_entity=model,
+            provider=SimpleNamespace(
+                requester=SimpleNamespace(get_reasoning_capabilities=Mock(return_value=capabilities))
+            ),
+        )
+        ap = SimpleNamespace(
+            persistence_mgr=SimpleNamespace(
+                serialize_model=Mock(
+                    return_value={
+                        'uuid': model.uuid,
+                        'name': model.name,
+                        'reasoning_config': {'level': 'high'},
+                    }
+                )
+            ),
+            model_mgr=SimpleNamespace(llm_model_dict={('workspace', model.uuid): runtime_model}),
+        )
+
+        serialized = _serialize_llm_model(ap, model)
+
+        assert serialized['reasoning_config'] == {'level': 'high'}
+        assert serialized['reasoning_capabilities'] == capabilities
+
+
 class TestLLMModelsServiceGetLLMModels:
     """Tests for LLMModelsService.get_llm_models method."""
 
@@ -580,6 +642,66 @@ class TestLLMModelsServiceCreateLLMModel:
         ap.provider_service.find_or_create_provider.assert_called_once()
         assert result_uuid is not None
 
+    async def test_create_llm_model_validates_explicit_reasoning_level(self):
+        ap = SimpleNamespace()
+        ap.persistence_mgr = SimpleNamespace(execute_async=AsyncMock(return_value=_create_mock_result([])))
+        runtime_provider = _create_reasoning_runtime_provider(
+            {
+                'supported': True,
+                'levels': ['provider_default', 'low', 'high'],
+                'source': 'litellm',
+            }
+        )
+        ap.model_mgr = _create_runtime_model_mgr()
+        ap.model_mgr.provider_dict = {'provider-uuid': runtime_provider}
+
+        service = LLMModelsService(ap)
+        await service.create_llm_model(
+            WORKSPACE_UUID,
+            {
+                'uuid': 'reasoning-model',
+                'name': 'Reasoning Model',
+                'provider_uuid': 'provider-uuid',
+                'abilities': ['reasoning'],
+                'reasoning_config': {'level': 'high'},
+                'extra_args': {},
+            },
+            preserve_uuid=True,
+            auto_set_to_default_pipeline=False,
+        )
+
+        runtime_entity = ap.model_mgr.load_llm_model_with_provider.await_args.args[1]
+        assert runtime_entity.reasoning_config == {'level': 'high'}
+
+    async def test_create_llm_model_rejects_unsupported_reasoning_before_insert(self):
+        ap = SimpleNamespace()
+        ap.persistence_mgr = SimpleNamespace(execute_async=AsyncMock())
+        runtime_provider = _create_reasoning_runtime_provider(
+            {
+                'supported': True,
+                'levels': ['provider_default'],
+                'source': 'manual',
+            }
+        )
+        ap.model_mgr = _create_runtime_model_mgr()
+        ap.model_mgr.provider_dict = {'provider-uuid': runtime_provider}
+
+        service = LLMModelsService(ap)
+        with pytest.raises(ValueError, match='Available levels: provider_default'):
+            await service.create_llm_model(
+                WORKSPACE_UUID,
+                {
+                    'name': 'Unknown Reasoning Model',
+                    'provider_uuid': 'provider-uuid',
+                    'abilities': ['reasoning'],
+                    'reasoning_config': {'level': 'high'},
+                    'extra_args': {},
+                },
+                auto_set_to_default_pipeline=False,
+            )
+
+        ap.persistence_mgr.execute_async.assert_not_awaited()
+
 
 class TestLLMModelsServiceUpdateLLMModel:
     """Tests for LLMModelsService.update_llm_model method."""
@@ -595,7 +717,10 @@ class TestLLMModelsServiceUpdateLLMModel:
         ap.model_mgr.remove_llm_model = AsyncMock()
         ap.model_mgr.load_llm_model_with_provider = AsyncMock(return_value=Mock())
 
-        ap.persistence_mgr.execute_async = AsyncMock()
+        existing_model = _create_mock_llm_model()
+        ap.persistence_mgr.execute_async = AsyncMock(
+            side_effect=[_create_mock_result(first_item=existing_model), _create_mock_result()]
+        )
 
         service = LLMModelsService(ap)
         service.get_llm_model = AsyncMock(return_value=_existing_llm_data())
@@ -623,7 +748,8 @@ class TestLLMModelsServiceUpdateLLMModel:
         ap.model_mgr.provider_dict = {}  # Empty
         ap.model_mgr.remove_llm_model = AsyncMock()
 
-        ap.persistence_mgr.execute_async = AsyncMock()
+        existing_model = _create_mock_llm_model()
+        ap.persistence_mgr.execute_async = AsyncMock(return_value=_create_mock_result(first_item=existing_model))
 
         service = LLMModelsService(ap)
         service.get_llm_model = AsyncMock(return_value=_existing_llm_data('nonexistent-provider'))

tests/unit_tests/api/service/test_space_service.py

@@ -25,6 +25,7 @@ import time
 
 from langbot.pkg.api.http.service.space import SpaceService
 from langbot.pkg.entity.persistence.user import User
+from langbot.pkg.utils import constants
 
 
 pytestmark = pytest.mark.asyncio
@@ -573,10 +574,20 @@ class TestSpaceServiceExchangeOAuthCode:
             mock_session_obj.post.return_value.__aexit__ = AsyncMock(return_value=None)
 
             # Execute
-            result = await service.exchange_oauth_code('auth_code')
+            result = await service.exchange_oauth_code(
+                'auth_code',
+                ['workspace-1'],
+                {'workspace-1': 1_700_000_000},
+            )
 
         # Verify
         assert result['access_token'] == 'new_access_token'
+        assert mock_session_obj.post.call_args.kwargs['json'] == {
+            'code': 'auth_code',
+            'instance_id': constants.instance_id,
+            'workspace_uuids': ['workspace-1'],
+            'workspace_created_ats': {'workspace-1': 1_700_000_000},
+        }
 
     async def test_exchange_oauth_code_api_error(self):
         """Raises ValueError on API error."""

tests/unit_tests/api/service/test_user_service.py

@@ -377,7 +377,7 @@ class TestUserServiceAuthenticate:
         service = UserService(ap)
 
         # Execute & Verify
-        with pytest.raises(ValueError, match='请使用 Space 账户登录'):
+        with pytest.raises(ValueError, match='请使用 LangBot 账号登录'):
             await service.authenticate('space@example.com', 'password')
 
 
@@ -726,7 +726,7 @@ class TestUserServiceCreateOrUpdateSpaceUser:
         )
         service = UserService(ap)
 
-        with pytest.raises(ControlPlaneDirectoryRequiredError, match='Space account'):
+        with pytest.raises(ControlPlaneDirectoryRequiredError, match='LangBot Account'):
             await service.register_invited_account('invite-token', 'member@example.com', 'password')
 
     async def test_create_or_update_new_space_user_first_init(self):

tests/unit_tests/api/test_plugin_runtime_route_fence.py

@@ -124,24 +124,37 @@ async def test_background_plugin_operation_refences_captured_generation(plugin_r
 
 
 @pytest.mark.asyncio
-async def test_background_plugin_operation_revalidates_inside_short_tenant_uow(plugin_router_cls):
+async def test_background_plugin_operation_revalidates_and_runs_inside_tenant_uow(plugin_router_cls):
     scopes = []
+    active_scope = None
+
+    transaction_active = False
 
     @asynccontextmanager
-    async def tenant_uow(workspace_uuid):
+    async def tenant_scope(workspace_uuid):
+        nonlocal active_scope
         scopes.append(workspace_uuid)
-        yield
+        active_scope = workspace_uuid
+        try:
+            yield
+        finally:
+            active_scope = None
 
     connector = SimpleNamespace(
         require_workspace_context=AsyncMock(side_effect=lambda context: context),
     )
-    operation = AsyncMock(return_value='done')
+
+    async def operation():
+        assert active_scope == CONTEXT.workspace_uuid
+        assert transaction_active is False
+        return 'done'
+
     router = object.__new__(plugin_router_cls)
     router.ap = SimpleNamespace(
         plugin_connector=connector,
         persistence_mgr=SimpleNamespace(
             mode=SimpleNamespace(value='cloud_runtime'),
-            tenant_uow=tenant_uow,
+            tenant_scope=tenant_scope,
         ),
     )
 
@@ -150,4 +163,3 @@ async def test_background_plugin_operation_revalidates_inside_short_tenant_uow(p
     assert result == 'done'
     assert scopes == [CONTEXT.workspace_uuid]
     connector.require_workspace_context.assert_awaited_once_with(CONTEXT)
-    operation.assert_awaited_once()

tests/unit_tests/box/test_box_connector.py

@@ -24,7 +24,7 @@ from langbot.pkg.box.connector import BoxRuntimeConnector
 _CONTROL_TOKEN = 'box-control-token-that-is-longer-than-32-bytes'
 
 
-def make_app(logger: Mock, runtime_endpoint: str = ''):
+def make_app(logger: Mock, runtime_endpoint: str = '', *, cloud: bool = False):
     return SimpleNamespace(
         logger=logger,
         workspace_service=SimpleNamespace(instance_uuid='instance-a'),
@@ -42,6 +42,7 @@ def make_app(logger: Mock, runtime_endpoint: str = ''):
                 }
             }
         ),
+        deployment=SimpleNamespace(mode='cloud' if cloud else 'oss'),
     )
 
 
@@ -306,10 +307,27 @@ def test_box_runtime_connector_rejects_relay_context_from_other_instance(
         )
 
 
-def test_external_box_runtime_fails_closed_without_control_token(monkeypatch: pytest.MonkeyPatch):
+def test_external_box_runtime_control_headers_are_tokenless_when_secret_is_unset(
+    monkeypatch: pytest.MonkeyPatch,
+):
     monkeypatch.delenv(BOX_CONTROL_TOKEN_ENV, raising=False)
     connector = BoxRuntimeConnector(make_app(Mock(), runtime_endpoint='http://box-runtime:5410'))
 
+    assert connector.get_control_headers() == {BOX_INSTANCE_HEADER: 'instance-a'}
+
+
+def test_cloud_box_runtime_rejects_missing_control_secret(monkeypatch: pytest.MonkeyPatch):
+    monkeypatch.delenv(BOX_CONTROL_TOKEN_ENV, raising=False)
+    connector = BoxRuntimeConnector(make_app(Mock(), runtime_endpoint='http://box-runtime:5410', cloud=True))
+
+    with pytest.raises(BoxRuntimeUnavailableError, match=BOX_CONTROL_TOKEN_ENV):
+        connector.get_control_headers()
+
+
+def test_external_box_runtime_rejects_invalid_configured_control_token(monkeypatch: pytest.MonkeyPatch):
+    monkeypatch.setenv(BOX_CONTROL_TOKEN_ENV, 'too-short')
+    connector = BoxRuntimeConnector(make_app(Mock(), runtime_endpoint='http://box-runtime:5410'))
+
     with pytest.raises(BoxRuntimeUnavailableError, match=BOX_CONTROL_TOKEN_ENV):
         connector.get_control_headers()
 

tests/unit_tests/cloud/test_bootstrap.py

@@ -108,7 +108,7 @@ def _cloud_config() -> dict:
             'use': 'pgvector',
             'pgvector': {
                 'use_business_database': True,
-                'allowed_dimensions': [384, 768, 1536],
+                'allowed_dimensions': [384, 768, 1536, 3072],
             },
         },
         'mcp': {'stdio': {'enabled': False}},
@@ -216,7 +216,6 @@ async def test_cloud_directory_capacity_contract_is_fail_closed(directory_config
     [
         ({'use_business_database': False, 'allowed_dimensions': [1536]}, 'use_business_database=true'),
         ({'use_business_database': True, 'allowed_dimensions': []}, 'allowed_dimensions'),
-        ({'use_business_database': True, 'allowed_dimensions': [3072]}, 'allowed_dimensions'),
         ({'use_business_database': True, 'allowed_dimensions': [True]}, 'allowed_dimensions'),
     ],
 )

tests/unit_tests/cloud/test_directory_projection.py

@@ -181,6 +181,39 @@ def _delta(
     )
 
 
+async def test_directory_delta_requests_model_catalog_sync_after_commit(projection_context):
+    application, _session_factory = projection_context
+    request_sync = Mock()
+    application.cloud_model_catalog_service = SimpleNamespace(request_sync=request_sync)
+    event = DirectoryEvent(
+        cursor=2,
+        uuid='20000000-0000-4000-8000-000000000002',
+        aggregate_uuid=WORKSPACE_UUID,
+        event_type='directory.changed',
+        revision=2,
+        payload={'workspace_uuid': WORKSPACE_UUID, 'directory_revision': 2},
+        created_at=datetime.datetime(2026, 7, 24, 12, 30, tzinfo=datetime.UTC),
+    )
+    batch = DirectoryEventBatch(
+        instance_uuid=INSTANCE_UUID,
+        after_cursor=1,
+        cursor=2,
+        high_water_cursor=2,
+        events=[event],
+    )
+    service = DirectoryProjectionService(
+        application,
+        _Provider([_snapshot(1)], [batch], [_delta(workspaces=[_workspace(revision=2)])]),
+        INSTANCE_UUID,
+    )
+    await service.initialize()
+    request_sync.reset_mock()
+
+    await service.sync_once()
+
+    request_sync.assert_called_once_with()
+
+
 async def test_initial_snapshot_projects_core_owned_rows(projection_context):
     application, session_factory = projection_context
     reconcile_execution_projection = Mock()
@@ -1023,7 +1056,7 @@ async def test_snapshot_for_another_instance_is_rejected(projection_context):
         await service.initialize()
 
 
-async def test_core_owned_membership_survives_directory_updates_and_omission(projection_context):
+async def test_directory_revision_zero_membership_is_adopted(projection_context):
     application, session_factory = projection_context
     service = DirectoryProjectionService(application, _Provider([_snapshot(1)]), INSTANCE_UUID)
     await service.initialize()
@@ -1034,29 +1067,125 @@ async def test_core_owned_membership_survives_directory_updates_and_omission(pro
             membership.role = 'viewer'
             membership.status = 'active'
             membership.projection_revision = 0
-            session.add(
-                WorkspaceMembership(
-                    uuid=SECOND_MEMBERSHIP_UUID,
-                    workspace_uuid=WORKSPACE_UUID,
-                    account_uuid='20000000-0000-0000-0000-000000000099',
-                    role='viewer',
-                    status='active',
-                    joined_at=membership.joined_at,
-                    projection_revision=0,
-                )
-            )
 
     projected_member = _member(revision=2).model_copy(update={'role': 'owner', 'membership_status': 'removed'})
     projected_workspace = _workspace(revision=2).model_copy(update={'members': (projected_member,)})
     await service.apply_snapshot(_snapshot(2, workspaces=[projected_workspace]))
 
     async with session_factory() as session:
-        memberships = {
-            membership.uuid: membership
-            for membership in (await session.scalars(sqlalchemy.select(WorkspaceMembership))).all()
-        }
-    assert memberships[MEMBERSHIP_UUID].role == 'viewer'
-    assert memberships[MEMBERSHIP_UUID].status == 'active'
-    assert memberships[MEMBERSHIP_UUID].projection_revision == 0
-    assert memberships[SECOND_MEMBERSHIP_UUID].status == 'active'
-    assert memberships[SECOND_MEMBERSHIP_UUID].projection_revision == 0
+        membership = await session.scalar(sqlalchemy.select(WorkspaceMembership))
+    assert membership.source == 'cloud_projection'
+    assert membership.role == 'owner'
+    assert membership.status == 'removed'
+    assert membership.projection_revision == 2
+
+
+async def test_directory_revision_zero_membership_omitted_from_snapshot_is_removed(projection_context):
+    application, session_factory = projection_context
+    service = DirectoryProjectionService(application, _Provider([_snapshot(1)]), INSTANCE_UUID)
+    await service.initialize()
+
+    historical_account_uuid = '20000000-0000-0000-0000-000000000099'
+    async with session_factory() as session:
+        async with session.begin():
+            membership = await session.scalar(sqlalchemy.select(WorkspaceMembership))
+            session.add(
+                User(
+                    uuid=historical_account_uuid,
+                    user='Historical Space Member',
+                    normalized_email='historical@example.com',
+                    password='',
+                    status='active',
+                    source='cloud_projection',
+                    projection_revision=1,
+                    account_type='space',
+                    space_account_uuid=historical_account_uuid,
+                )
+            )
+            session.add(
+                WorkspaceMembership(
+                    uuid=SECOND_MEMBERSHIP_UUID,
+                    workspace_uuid=WORKSPACE_UUID,
+                    account_uuid=historical_account_uuid,
+                    role='viewer',
+                    status='active',
+                    source='cloud_projection',
+                    joined_at=membership.joined_at,
+                    projection_revision=0,
+                )
+            )
+
+    await service.apply_snapshot(_snapshot(2))
+
+    async with session_factory() as session:
+        historical = await session.get(WorkspaceMembership, SECOND_MEMBERSHIP_UUID)
+    assert historical.status == 'removed'
+    assert historical.projection_revision == 2
+
+
+async def test_cloud_account_core_invitation_membership_survives_directory_omission(projection_context):
+    application, session_factory = projection_context
+    service = DirectoryProjectionService(application, _Provider([_snapshot(1)]), INSTANCE_UUID)
+    await service.initialize()
+
+    invited_account_uuid = '20000000-0000-0000-0000-000000000098'
+    async with session_factory() as session:
+        async with session.begin():
+            projected_membership = await session.scalar(sqlalchemy.select(WorkspaceMembership))
+            session.add(
+                User(
+                    uuid=invited_account_uuid,
+                    user='Invited Cloud Account',
+                    normalized_email='invited-cloud@example.com',
+                    password='',
+                    status='active',
+                    source='cloud_projection',
+                    projection_revision=1,
+                    account_type='space',
+                    space_account_uuid=invited_account_uuid,
+                )
+            )
+            session.add(
+                WorkspaceMembership(
+                    uuid=SECOND_MEMBERSHIP_UUID,
+                    workspace_uuid=WORKSPACE_UUID,
+                    account_uuid=invited_account_uuid,
+                    role='viewer',
+                    status='active',
+                    source='local',
+                    joined_at=projected_membership.joined_at,
+                    projection_revision=0,
+                )
+            )
+
+    await service.apply_snapshot(_snapshot(2))
+
+    async with session_factory() as session:
+        membership = await session.get(WorkspaceMembership, SECOND_MEMBERSHIP_UUID)
+    assert membership.source == 'local'
+    assert membership.status == 'active'
+    assert membership.projection_revision == 0
+
+
+async def test_directory_does_not_adopt_local_membership_with_different_uuid_for_same_cloud_account(projection_context):
+    application, session_factory = projection_context
+    service = DirectoryProjectionService(application, _Provider([_snapshot(1)]), INSTANCE_UUID)
+    await service.initialize()
+
+    async with session_factory() as session:
+        async with session.begin():
+            membership = await session.scalar(sqlalchemy.select(WorkspaceMembership))
+            membership.uuid = SECOND_MEMBERSHIP_UUID
+            membership.source = 'local'
+            membership.projection_revision = 0
+
+    projected_member = _member(revision=2).model_copy(update={'role': 'owner', 'membership_status': 'removed'})
+    projected_workspace = _workspace(revision=2).model_copy(update={'members': (projected_member,)})
+    await service.apply_snapshot(_snapshot(2, workspaces=[projected_workspace]))
+
+    async with session_factory() as session:
+        membership = await session.get(WorkspaceMembership, SECOND_MEMBERSHIP_UUID)
+    assert membership.source == 'local'
+    assert membership.role == 'developer'
+    assert membership.status == 'active'
+    assert membership.projection_revision == 0

tests/unit_tests/cloud/test_model_catalog.py

@@ -273,6 +273,94 @@ async def test_snapshot_must_cover_every_active_workspace() -> None:
         await service.sync_once()
 
 
+async def test_periodic_sync_discovers_workspace_created_after_startup_cache_release(tmp_path) -> None:
+    engine = create_async_engine(f'sqlite+aiosqlite:///{tmp_path / "model-catalog-new-workspace.db"}')
+    manager = PersistenceManager(object(), mode=PersistenceMode.CLOUD_RUNTIME)
+    manager.db = SimpleNamespace(get_engine=lambda: engine)
+    startup_bindings = [
+        SimpleNamespace(instance_uuid=INSTANCE_UUID, workspace_uuid=WORKSPACE_A, placement_generation=1)
+    ]
+    live_bindings = [
+        *startup_bindings,
+        SimpleNamespace(instance_uuid=INSTANCE_UUID, workspace_uuid=WORKSPACE_B, placement_generation=1),
+    ]
+
+    class _WorkspaceService:
+        startup_released = False
+
+        async def list_active_execution_bindings(self):
+            return list(live_bindings if self.startup_released else startup_bindings)
+
+        def release_startup_execution_bindings(self):
+            self.startup_released = True
+
+    workspace_service = _WorkspaceService()
+    app = SimpleNamespace(
+        persistence_mgr=manager,
+        workspace_service=workspace_service,
+        model_mgr=SimpleNamespace(load_models_from_db=_AsyncCounter()),
+        logger=logging.getLogger(__name__),
+    )
+    service = CloudModelCatalogSyncService(app, _CatalogProvider(_snapshot()), INSTANCE_UUID)
+
+    try:
+        async with engine.begin() as connection:
+            await connection.run_sync(Base.metadata.create_all)
+            await connection.execute(
+                sqlalchemy.insert(Workspace),
+                [
+                    {
+                        'uuid': WORKSPACE_A,
+                        'instance_uuid': INSTANCE_UUID,
+                        'name': 'A',
+                        'slug': 'a',
+                        'source': 'cloud_projection',
+                    },
+                    {
+                        'uuid': WORKSPACE_B,
+                        'instance_uuid': INSTANCE_UUID,
+                        'name': 'B',
+                        'slug': 'b',
+                        'source': 'cloud_projection',
+                    },
+                ],
+            )
+
+        await service.initialize()
+        workspace_service.release_startup_execution_bindings()
+        await service.sync_once()
+
+        async with engine.connect() as connection:
+            provider_b = await connection.scalar(
+                sqlalchemy.select(ModelProvider).where(ModelProvider.uuid == system_provider_uuid(WORKSPACE_B))
+            )
+        assert provider_b is not None
+    finally:
+        await engine.dispose()
+
+
+async def test_catalog_run_wakes_immediately_when_directory_changes() -> None:
+    sync_started = asyncio.Event()
+
+    class _WakeService(CloudModelCatalogSyncService):
+        async def sync_once(self, *, reload_runtime: bool = True):
+            del reload_runtime
+            sync_started.set()
+            return {'workspaces': 0, 'created': 0, 'updated': 0, 'deleted': 0}
+
+    app = SimpleNamespace(logger=logging.getLogger(__name__))
+    service = _WakeService(app, _CatalogProvider(_snapshot()), INSTANCE_UUID, sync_interval_seconds=3600)
+    task = asyncio.create_task(service.run())
+    try:
+        await asyncio.sleep(0)
+        service.request_sync()
+        await asyncio.wait_for(sync_started.wait(), timeout=0.2)
+    finally:
+        task.cancel()
+        with pytest.raises(asyncio.CancelledError):
+            await task
+
+
 async def _async_value(value):
     return value
 

tests/unit_tests/core/test_app_shutdown.py

@@ -87,7 +87,10 @@ async def test_runtime_resource_stats_are_aggregate_and_constant_time() -> None:
     app.platform_mgr = SimpleNamespace(_bots_by_key={})
     app.pipeline_mgr = SimpleNamespace(_pipelines_by_key={})
     app.rag_mgr = SimpleNamespace(knowledge_bases={})
-    app.plugin_connector = SimpleNamespace(_known_desired_states={'installation': object()})
+    app.plugin_connector = SimpleNamespace(
+        _known_desired_states={'installation': object()},
+        _runtime_available=lambda: True,
+    )
     app.persistence_mgr = SimpleNamespace(
         get_resource_stats=lambda: {
             'configured_capacity': 20,
@@ -140,3 +143,4 @@ async def test_runtime_resource_stats_are_aggregate_and_constant_time() -> None:
     }
     assert stats['models']['providers'] == 1
     assert stats['runtimes']['plugin_installations'] == 1
+    assert stats['runtimes']['plugin_runtime_connected'] is True

tests/unit_tests/core/test_load_config.py

@@ -319,6 +319,7 @@ class TestApplyEnvOverridesToConfig:
         load_config = get_load_config_module()
         cfg = {
             'plugin': {
+                'connect_timeout_seconds': 30.0,
                 'worker': {
                     'max_cpus': 1.0,
                     'max_memory_mb': 512,
@@ -329,11 +330,12 @@ class TestApplyEnvOverridesToConfig:
                     'restart_failure_threshold': 8,
                     'restart_failure_window_seconds': 30.0,
                     'restart_circuit_open_seconds': 60.0,
-                }
+                },
             },
             'mcp': {'stdio': {'enabled': True}},
         }
         env = {
+            'PLUGIN__CONNECT_TIMEOUT_SECONDS': '180',
             'PLUGIN__WORKER__MAX_CPUS': '2.5',
             'PLUGIN__WORKER__MAX_MEMORY_MB': '1024',
             'PLUGIN__WORKER__MAX_PIDS': '64',
@@ -349,6 +351,7 @@ class TestApplyEnvOverridesToConfig:
         with patch.dict(os.environ, env, clear=True):
             result = load_config._apply_env_overrides_to_config(cfg)
 
+        assert result['plugin']['connect_timeout_seconds'] == 180.0
         assert result['plugin']['worker'] == {
             'max_cpus': 2.5,
             'max_memory_mb': 1024,
@@ -393,6 +396,14 @@ class TestApplyEnvOverridesToConfig:
         assert isinstance(result['plugin']['worker']['max_memory_mb'], int)
         assert result['mcp']['stdio']['enabled'] is False
 
+    def test_runtime_policy_defaults_add_typed_plugin_connect_timeout(self):
+        load_config = get_load_config_module()
+
+        completed = load_config._complete_runtime_policy_defaults({'plugin': {'enable': True}})
+
+        assert completed['plugin']['connect_timeout_seconds'] == 180.0
+        assert isinstance(completed['plugin']['connect_timeout_seconds'], float)
+
     def test_webhook_prefix_override(self):
         """Test overriding webhook_prefix via environment variable."""
         load_config = get_load_config_module()

tests/unit_tests/persistence/test_tenant_uow.py

@@ -7,7 +7,7 @@ from types import SimpleNamespace
 
 import pytest
 import sqlalchemy as sa
-from pgvector.sqlalchemy import Vector
+from pgvector.sqlalchemy import HALFVEC, Vector
 from sqlalchemy.dialects.postgresql import insert as postgresql_insert
 from sqlalchemy.dialects.sqlite import insert as sqlite_insert
 from sqlalchemy.ext.asyncio import create_async_engine
@@ -967,6 +967,7 @@ async def test_scoped_session_rejects_raw_or_unapproved_sql(
         ),
         sa.select(sa.column('embedding').op('<=>')(sa.literal([0.1]))),
         sa.select(sa.cast(sa.column('embedding'), Vector(384))),
+        sa.select(sa.cast(sa.column('embedding'), HALFVEC(3072))),
         sa.insert(sa.table('rows', sa.column('id'))).values(id=1),
         _multi_value_statement(value=1),
         _on_conflict_statement(update_value=sa.func.coalesce(sa.literal(1), sa.literal(0))),

tests/unit_tests/pipeline/test_controller_tenancy.py

@@ -29,6 +29,57 @@ def _prepare_scheduler(mock_app):
     return query_pool, session
 
 
+@pytest.mark.asyncio
+async def test_consumer_schedules_query_after_running_transition(
+    mock_app,
+    sample_query,
+):
+    query_pool = MagicMock()
+    query_pool.queries = [sample_query]
+    query_pool.__aenter__ = AsyncMock(return_value=query_pool)
+    query_pool.__aexit__ = AsyncMock(return_value=None)
+    query_pool.remove_query = AsyncMock(return_value=True)
+    wait_for_query = asyncio.Event()
+    query_pool.condition = SimpleNamespace(
+        wait=AsyncMock(side_effect=wait_for_query.wait),
+        notify_all=Mock(),
+    )
+    query_pool.mark_query_running_locked = Mock(side_effect=query_pool.queries.remove)
+    mock_app.query_pool = query_pool
+
+    session = SimpleNamespace(_semaphore=asyncio.Semaphore(1))
+    mock_app.sess_mgr.get_session = AsyncMock(return_value=session)
+    runtime_pipeline = SimpleNamespace(run=AsyncMock())
+    mock_app.pipeline_mgr = SimpleNamespace(get_pipeline_by_uuid=AsyncMock(return_value=runtime_pipeline))
+
+    task_created = asyncio.Event()
+    process_tasks = []
+
+    def create_process_task(coro, **_kwargs):
+        process_tasks.append(asyncio.create_task(coro))
+        task_created.set()
+
+    mock_app.task_mgr.create_task = Mock(side_effect=create_process_task)
+    controller = Controller(mock_app)
+    initial_slots = controller.semaphore._value
+    consumer_task = asyncio.create_task(controller.consumer())
+
+    try:
+        await asyncio.wait_for(task_created.wait(), timeout=2)
+    finally:
+        consumer_task.cancel()
+        with pytest.raises(asyncio.CancelledError):
+            await consumer_task
+        await asyncio.gather(*process_tasks)
+
+    query_pool.mark_query_running_locked.assert_called_once_with(sample_query)
+    runtime_pipeline.run.assert_awaited_once_with(sample_query)
+    query_pool.remove_query.assert_awaited_once_with(sample_query)
+    assert query_pool.queries == []
+    assert session._semaphore._value == 1
+    assert controller.semaphore._value == initial_slots
+
+
 @pytest.mark.asyncio
 async def test_controller_drops_stale_query_before_pipeline_lookup(
     mock_app,

tests/unit_tests/platform/test_websocket_adapter_attachments.py

@@ -2,8 +2,8 @@
 
 The web debug client uploads Image / Voice / File components carrying a storage
 key in ``path``. This helper resolves each to a base64 data URI (so multimodal
-LLM input and the Box sandbox inbox have usable bytes) while retaining the key
-for browser history. Covers mimetype selection per type and fail-closed error
+LLM input and the Box sandbox inbox have usable bytes), then deletes the
+consumed upload. Covers mimetype selection per type and fail-closed error
 handling.
 """
 
@@ -52,7 +52,7 @@ def _make_adapter(load_return=b'hello', load_side_effect=None):
 
 
 @pytest.mark.asyncio
-async def test_image_jpeg_mimetype_and_retained_storage_key():
+async def test_image_jpeg_mimetype_and_consumed_storage_key():
     adapter, storage_mgr, _ = _make_adapter(load_return=b'\xff\xd8\xff')
     path = f'{_UPLOAD_PREFIX}photo.jpg'
     chain = [{'type': 'Image', 'path': path}]
@@ -61,8 +61,12 @@ async def test_image_jpeg_mimetype_and_retained_storage_key():
 
     expected_b64 = base64.b64encode(b'\xff\xd8\xff').decode('utf-8')
     assert chain[0]['base64'] == f'data:image/jpeg;base64,{expected_b64}'
-    assert chain[0]['path'] == path
-    storage_mgr.delete_scoped_object_key.assert_not_awaited()
+    assert chain[0]['path'] == ''
+    storage_mgr.delete_scoped_object_key.assert_awaited_once_with(
+        _CONTEXT,
+        path,
+        expected_owner_type='upload_image',
+    )
 
 
 def test_history_retains_storage_key_without_large_base64_payload():

tests/unit_tests/platform/test_websocket_session_isolation.py

@@ -7,8 +7,8 @@ from unittest.mock import AsyncMock, Mock
 
 import pytest
 
-import langbot_plugin.api.entities.builtin.platform.events as platform_events
 import langbot_plugin.api.entities.builtin.platform.entities as platform_entities
+import langbot_plugin.api.entities.builtin.platform.events as platform_events
 import langbot_plugin.api.entities.builtin.platform.message as platform_message
 import langbot_plugin.api.entities.builtin.provider.message as provider_message
 from langbot.pkg.platform.sources import websocket_adapter as websocket_adapter_module
@@ -347,9 +347,9 @@ async def test_stable_session_launcher_resolves_to_active_connection(monkeypatch
 
 
 @pytest.mark.asyncio
-async def test_dashboard_reply_uses_event_pipeline_after_connection_closes(monkeypatch):
+async def test_dashboard_reply_survives_connection_replacement(monkeypatch):
     manager = WebSocketConnectionManager()
-    connection = await manager.add_connection(
+    original = await manager.add_connection(
         websocket=Mock(),
         scope=SCOPE_A,
         pipeline_uuid='pipeline-1',
@@ -357,22 +357,36 @@ async def test_dashboard_reply_uses_event_pipeline_after_connection_closes(monke
     )
     monkeypatch.setattr(websocket_adapter_module, 'ws_connection_manager', manager)
 
-    app = Mock()
-    app.platform_mgr.websocket_proxy_bot.bot_entity = Mock(spec=[])
-    adapter = WebSocketAdapter.model_construct(ap=app, logger=AsyncMock())
-    message_source = platform_events.FriendMessage(
-        sender=platform_entities.Friend(
-            id=f'websocket_{connection.connection_id}',
-            nickname='User',
-            remark='User',
-        ),
-        message_chain=platform_message.MessageChain([platform_message.Plain(text='hello')]),
-        time=1,
-    )
-    object.__setattr__(message_source, '_langbot_pipeline_uuid', 'pipeline-1')
-    await manager.remove_connection(connection.connection_id)
+    adapter = WebSocketAdapter.model_construct(ap=Mock(), logger=_adapter_logger())
+    adapter.websocket_person_session = WebSocketSession(id='person')
+    adapter.websocket_group_session = WebSocketSession(id='group')
+    received = []
 
-    assert await adapter._get_message_context(message_source) == ('pipeline-1', None)
+    async def listener(event, _callback_adapter):
+        received.append(event)
+
+    adapter.listeners = {platform_events.FriendMessage: listener}
+    await adapter.handle_websocket_message(
+        original,
+        {'message': [{'type': 'Plain', 'text': 'hello'}], 'stream': False},
+    )
+    await asyncio.sleep(0)
+    await manager.remove_connection(original.connection_id)
+    replacement = await manager.add_connection(
+        websocket=Mock(),
+        scope=SCOPE_A,
+        pipeline_uuid='pipeline-1',
+        session_type='person',
+    )
+
+    await adapter.reply_message(
+        received[0],
+        platform_message.MessageChain([platform_message.Plain(text='done')]),
+    )
+
+    response = await replacement.send_queue.get()
+    assert response['type'] == 'response'
+    assert response['data']['content'] == 'done'
 
 
 @pytest.mark.asyncio
@@ -468,7 +482,7 @@ async def test_attachment_key_must_belong_to_connection_upload_scope():
     await adapter._process_image_components(connection, message_chain)
 
     assert message_chain[0]['base64'].startswith('data:image/png;base64,')
-    assert message_chain[0]['path'] == 'v1/current/upload_image/key.png'
+    assert message_chain[0]['path'] == ''
     storage_mgr.scoped_prefix.assert_called_once_with(
         connection.execution_context,
         owner_type='upload_image',
@@ -482,7 +496,11 @@ async def test_attachment_key_must_belong_to_connection_upload_scope():
         'v1/current/upload_image/key.png',
         expected_owner_type='upload_image',
     )
-    storage_mgr.delete_scoped_object_key.assert_not_awaited()
+    storage_mgr.delete_scoped_object_key.assert_awaited_once_with(
+        connection.execution_context,
+        'v1/current/upload_image/key.png',
+        expected_owner_type='upload_image',
+    )
 
     with pytest.raises(ValueError, match='does not belong'):
         await adapter._process_image_components(

tests/unit_tests/plugin/test_connector_methods.py

@@ -93,14 +93,10 @@ class TestRunAgent:
     @pytest.mark.asyncio
     async def test_revalidates_trusted_execution_context(self):
         connector = create_mock_connector()
-        connector._current_execution_context = AsyncMock(
-            return_value=TEST_EXECUTION_CONTEXT
-        )
+        connector._current_execution_context = AsyncMock(return_value=TEST_EXECUTION_CONTEXT)
 
         class RuntimeHandler:
-            installation_scope = Mock(
-                side_effect=lambda _binding: nullcontext()
-            )
+            installation_scope = Mock(side_effect=lambda _binding: nullcontext())
 
             async def run_agent(self, *_args):
                 yield {'type': 'run.completed'}
@@ -113,16 +109,12 @@ class TestRunAgent:
         )
 
         assert results == [{'type': 'run.completed'}]
-        connector.require_workspace_context.assert_awaited_once_with(
-            TEST_EXECUTION_CONTEXT
-        )
+        connector.require_workspace_context.assert_awaited_once_with(TEST_EXECUTION_CONTEXT)
 
     @pytest.mark.asyncio
     async def test_rejects_payload_workspace_mismatch(self):
         connector = create_mock_connector()
-        connector._current_execution_context = AsyncMock(
-            return_value=TEST_EXECUTION_CONTEXT
-        )
+        connector._current_execution_context = AsyncMock(return_value=TEST_EXECUTION_CONTEXT)
         configure_handler(connector, AsyncMock())
 
         with pytest.raises(WorkspaceNotFoundError, match='Plugin resource not found'):
@@ -670,8 +662,13 @@ class TestDisabledPluginEarlyReturns:
         mock_app.instance_config.data = {'plugin': {'enable': False}}
 
         connector = connector_module.PluginRuntimeConnector(mock_app, mock_disconnect)
+        execution_context = connector_module.ExecutionContext(
+            instance_uuid='instance-a',
+            workspace_uuid='workspace-a',
+            placement_generation=1,
+        )
 
-        result = await connector.get_debug_info()
+        result = await connector.get_debug_info(execution_context)
 
         assert result == {}
 

tests/unit_tests/plugin/test_connector_ping.py

@@ -15,7 +15,7 @@ from langbot_plugin.runtime.security import (
 )
 
 
-def make_connector() -> PluginRuntimeConnector:
+def make_connector(*, cloud: bool = False) -> PluginRuntimeConnector:
     app = SimpleNamespace(
         logger=Mock(),
         instance_config=SimpleNamespace(
@@ -34,6 +34,7 @@ def make_connector() -> PluginRuntimeConnector:
                 'space': {'url': ''},
             }
         ),
+        deployment=SimpleNamespace(mode='cloud' if cloud else 'oss'),
     )
     return PluginRuntimeConnector(app, AsyncMock())
 
@@ -142,6 +143,49 @@ async def test_stdio_runtime_connection_does_not_capture_unconsumed_stderr(
     await connector.aclose()
 
 
+@pytest.mark.asyncio
+async def test_invalid_connect_timeout_is_rejected_before_transport_startup(
+    monkeypatch: pytest.MonkeyPatch,
+):
+    connector = make_connector()
+    connector.ap.instance_config.data['plugin']['connect_timeout_seconds'] = 0
+    stdio_controller = Mock()
+    websocket_controller = Mock()
+    create_task = Mock()
+    get_platform = Mock(return_value='linux')
+    use_websocket = Mock(return_value=False)
+    connector._start_runtime_subprocess = AsyncMock()
+    monkeypatch.setattr(connector_module.constants, 'instance_id', 'instance-a')
+    monkeypatch.setattr(connector_module.asyncio, 'create_task', create_task)
+    monkeypatch.setattr(connector_module.platform, 'get_platform', get_platform)
+    monkeypatch.setattr(
+        connector_module.platform,
+        'use_websocket_to_connect_plugin_runtime',
+        use_websocket,
+    )
+    monkeypatch.setattr(
+        connector_module.stdio_client_controller,
+        'StdioClientController',
+        stdio_controller,
+    )
+    monkeypatch.setattr(
+        connector_module.ws_client_controller,
+        'WebSocketClientController',
+        websocket_controller,
+    )
+
+    with pytest.raises(ValueError, match='plugin.connect_timeout_seconds'):
+        await connector.initialize()
+
+    get_platform.assert_not_called()
+    use_websocket.assert_not_called()
+    stdio_controller.assert_not_called()
+    websocket_controller.assert_not_called()
+    connector._start_runtime_subprocess.assert_not_awaited()
+    create_task.assert_not_called()
+    assert connector._transport_task is None
+
+
 @pytest.mark.asyncio
 async def test_runtime_disconnect_notifies_once_and_clears_handler(
     monkeypatch: pytest.MonkeyPatch,
@@ -292,10 +336,17 @@ def test_closed_deployment_selects_instance_scoped_shared_profile():
     assert connector.runtime_profile == 'shared'
 
 
-def test_external_runtime_control_headers_require_strong_secret(monkeypatch):
+def test_external_runtime_control_headers_are_empty_when_secret_is_unset(monkeypatch):
     monkeypatch.delenv(PLUGIN_RUNTIME_CONTROL_TOKEN_ENV, raising=False)
     connector = make_connector()
 
+    assert connector._control_headers(allow_generate=False) == {}
+
+
+def test_cloud_runtime_rejects_missing_control_secret(monkeypatch):
+    monkeypatch.delenv(PLUGIN_RUNTIME_CONTROL_TOKEN_ENV, raising=False)
+    connector = make_connector(cloud=True)
+
     with pytest.raises(PluginRuntimeNotConnectedError, match=PLUGIN_RUNTIME_CONTROL_TOKEN_ENV):
         connector._control_headers(allow_generate=False)
 

tests/unit_tests/plugin/test_connector_reconcile.py

@@ -153,6 +153,31 @@ async def test_empty_projected_workspaces_do_not_retain_installation_sets():
     connector.handler.reconcile_plugin_installations.assert_awaited_once_with(())
 
 
+@pytest.mark.asyncio
+async def test_shared_reconcile_logs_workspace_installation_counts_and_elapsed_time():
+    binding_a = execution_binding('workspace-a')
+    binding_b = execution_binding('workspace-b')
+    setting_a = plugin_setting('01', 'a' * 64)
+    setting_b = plugin_setting('02', 'b' * 64)
+    connector = shared_connector(
+        [[binding_a, binding_b]],
+        {'workspace-a': [setting_a], 'workspace-b': [setting_b]},
+    )
+    connector.handler = runtime_handler()
+    await connector._prepare_connected_runtime()
+
+    matching_calls = [
+        call
+        for call in connector.ap.logger.info.call_args_list
+        if call.args
+        and call.args[0]
+        == 'Shared plugin runtime reconcile completed: workspaces=%d desired_installations=%d elapsed_seconds=%.3f'
+    ]
+    assert len(matching_calls) == 1
+    assert matching_calls[0].args[1:3] == (2, 2)
+    assert matching_calls[0].args[3] >= 0
+
+
 @pytest.mark.asyncio
 async def test_fresh_shared_runtime_cache_replays_persisted_local_package():
     package = b'local-lbpkg-bytes'

tests/unit_tests/plugin/test_connector_static.py

@@ -6,9 +6,10 @@ Tests cover:
 
 from __future__ import annotations
 
-import pytest
 from importlib import import_module
 
+import pytest
+
 
 def get_connector_module():
     """Lazy import to avoid circular import issues."""
@@ -60,3 +61,28 @@ def test_runtime_id_is_stable_across_core_restarts(monkeypatch):
     monkeypatch.setattr(connector.constants, 'instance_id', 'instance-a')
 
     assert connector.PluginRuntimeConnector._build_runtime_id() == 'instance-a:plugin-runtime'
+
+
+def test_runtime_connect_timeout_defaults_to_three_minutes():
+    connector = get_connector_module()
+    assert connector.PluginRuntimeConnector._runtime_connect_timeout({}) == 180.0
+
+
+def test_runtime_connect_timeout_reads_typed_plugin_config():
+    connector = get_connector_module()
+    assert connector.PluginRuntimeConnector._runtime_connect_timeout({'connect_timeout_seconds': 45.5}) == 45.5
+
+
+@pytest.mark.parametrize('value', [True, False, None, 0, -1, float('nan'), float('inf'), '180', object()])
+def test_runtime_connect_timeout_rejects_invalid_values(value):
+    connector = get_connector_module()
+    with pytest.raises(ValueError, match='plugin.connect_timeout_seconds'):
+        connector.PluginRuntimeConnector._runtime_connect_timeout({'connect_timeout_seconds': value})
+
+
+def test_runtime_connect_timeout_error_displays_actual_seconds():
+    connector = get_connector_module()
+
+    assert connector.PluginRuntimeConnector._runtime_connect_timeout_error(45.5) == (
+        'Plugin runtime did not become ready within 45.5 seconds'
+    )

tests/unit_tests/plugin/test_handler.py

@@ -9,8 +9,8 @@ from types import SimpleNamespace
 from unittest.mock import AsyncMock, MagicMock, Mock
 import pytest
 
-from langbot_plugin.entities.io.actions.enums import PluginToRuntimeAction
-from langbot_plugin.entities.io.context import ActionContext, InstallationBinding
+from langbot_plugin.entities.io.actions.enums import LangBotToRuntimeAction, PluginToRuntimeAction
+from langbot_plugin.entities.io.context import ActionContext, InstallationBinding, PluginInstallationDesiredState
 
 
 def make_handler(app):
@@ -67,6 +67,20 @@ def make_handler(app):
     return runtime_handler
 
 
+@pytest.mark.asyncio
+async def test_reconcile_plugin_installations_allows_cloud_cold_start_to_finish():
+    app = SimpleNamespace()
+    runtime_handler = make_handler(app)
+    runtime_handler.call_action = AsyncMock(return_value={})
+    binding = next(iter(runtime_handler._installation_bindings.values()))[0]
+    desired = PluginInstallationDesiredState(binding=binding, enabled=True)
+
+    await runtime_handler.reconcile_plugin_installations((desired,))
+
+    assert runtime_handler.call_action.await_args.args[0] == LangBotToRuntimeAction.RECONCILE_PLUGIN_INSTALLATIONS
+    assert runtime_handler.call_action.await_args.kwargs['timeout'] == 300
+
+
 class TestHandlerQueryVariables:
     """Tests for handler query variable logic."""
 

tests/unit_tests/plugin/test_handler_tenancy.py

@@ -396,9 +396,7 @@ async def test_legacy_oss_knowledge_file_reply_uses_complete_installation_bindin
         get_file_stream=AsyncMock(return_value=b'knowledge-file'),
     )
     runtime_handler.send_file = AsyncMock(return_value='knowledge-file-key')
-    legacy_context = workspace_context().for_installation(
-        installation_context.installation_uuid
-    )
+    legacy_context = workspace_context().for_installation(installation_context.installation_uuid)
 
     response = await invoke_with_context(
         runtime_handler,
@@ -505,3 +503,23 @@ async def test_host_to_runtime_action_carries_trusted_connector_context():
         'runtime_id': 'runtime-a',
     }
     assert request.get('context') is None
+
+
+@pytest.mark.asyncio
+async def test_get_debug_info_converts_execution_context_to_sdk_action_context():
+    runtime_handler, _app, _installation_context = make_handler()
+    runtime_handler.call_action = AsyncMock(return_value={'plugin_debug_key': 'debug-key'})
+    execution_context = ExecutionContext(
+        instance_uuid='instance-a',
+        workspace_uuid='workspace-a',
+        placement_generation=7,
+    )
+
+    result = await runtime_handler.get_debug_info(execution_context)
+
+    assert result == {'plugin_debug_key': 'debug-key'}
+    assert runtime_handler.call_action.await_args.kwargs['action_context'] == ActionContext(
+        instance_uuid='instance-a',
+        workspace_uuid='workspace-a',
+        placement_generation=7,
+    )

tests/unit_tests/provider/test_litellmchat.py

@@ -1305,6 +1305,7 @@ class TestScanModels:
         )
         requester._supports_function_calling = Mock(side_effect=lambda model_id: model_id == 'gpt-4o')
         requester._supports_vision = Mock(side_effect=lambda model_id: model_id == 'gpt-4o')
+        requester._supports_reasoning = Mock(side_effect=lambda model_id: model_id == 'o3')
         requester._safe_context_length = Mock(side_effect=lambda model_id: 128000 if model_id == 'gpt-4o' else None)
 
         mock_response = Mock()
@@ -1312,6 +1313,7 @@ class TestScanModels:
             return_value={
                 'data': [
                     {'id': 'gpt-4o'},
+                    {'id': 'o3'},
                     {'id': 'text-embedding-3-small'},
                     {'id': 'bge-reranker-v2'},
                 ]
@@ -1328,6 +1330,7 @@ class TestScanModels:
         by_id = {model['id']: model for model in result['models']}
         assert by_id['gpt-4o']['abilities'] == ['func_call', 'vision']
         assert by_id['gpt-4o']['context_length'] == 128000
+        assert by_id['o3']['abilities'] == ['reasoning']
         assert by_id['text-embedding-3-small']['type'] == 'embedding'
         assert by_id['bge-reranker-v2']['type'] == 'rerank'
 
@@ -1375,8 +1378,8 @@ class TestScanModels:
         )
 
         with patch.object(litellmchat.litellm, 'get_model_info') as mock_get_model_info:
-            mock_get_model_info.side_effect = (
-                lambda model: {'max_input_tokens': 131072} if model == 'moonshot/moonshot-v1-128k' else {}
+            mock_get_model_info.side_effect = lambda model: (
+                {'max_input_tokens': 131072} if model == 'moonshot/moonshot-v1-128k' else {}
             )
 
             assert requester._safe_context_length('moonshot-v1-128k') == 131072
@@ -1405,8 +1408,8 @@ class TestScanModels:
         )
 
         with patch.object(litellmchat.litellm, 'supports_function_calling') as mock_supports_function_calling:
-            mock_supports_function_calling.side_effect = (
-                lambda model, custom_llm_provider=None: model == 'moonshot/kimi-k2.6' and custom_llm_provider is None
+            mock_supports_function_calling.side_effect = lambda model, custom_llm_provider=None: (
+                model == 'moonshot/kimi-k2.6' and custom_llm_provider is None
             )
 
             assert requester._supports_function_calling('kimi-k2.6') is True

tests/unit_tests/provider/test_reasoning_control.py

@@ -0,0 +1,884 @@
+from __future__ import annotations
+
+from types import SimpleNamespace
+from unittest.mock import AsyncMock
+
+import pytest
+
+import langbot_plugin.api.entities.builtin.provider.message as provider_message
+from langbot.pkg.api.http.context import ExecutionContext
+from langbot.pkg.entity.persistence import model as persistence_model
+from langbot.pkg.provider.modelmgr import errors, reasoning, requester
+from langbot.pkg.provider.modelmgr.requesters import litellmchat
+from langbot.pkg.provider.modelmgr.requesters.litellmchat import LiteLLMRequester
+
+
+def _runtime_model(
+    request: LiteLLMRequester,
+    level: str = 'provider_default',
+    name: str = 'reasoning-model',
+    abilities: list[str] | None = None,
+    requester_name: str | None = None,
+) -> requester.RuntimeLLMModel:
+    execution_context = ExecutionContext(
+        instance_uuid='instance-test',
+        workspace_uuid='workspace-test',
+        placement_generation=1,
+    )
+    entity = persistence_model.LLMModel(
+        workspace_uuid='workspace-test',
+        uuid='reasoning-model',
+        name=name,
+        provider_uuid='provider-test',
+        abilities=abilities if abilities is not None else ['reasoning'],
+        reasoning_config={'level': level},
+        extra_args={},
+    )
+    provider = SimpleNamespace(
+        execution_context=execution_context,
+        provider_entity=persistence_model.ModelProvider(
+            workspace_uuid='workspace-test',
+            uuid='provider-test',
+            name='provider',
+            requester=requester_name or request.requester_cfg.get('requester_name') or 'custom-requester',
+            base_url='https://example.com',
+            api_keys=[],
+        ),
+        requester=request,
+        token_mgr=SimpleNamespace(),
+    )
+    return requester.RuntimeLLMModel(execution_context, entity, provider)
+
+
+def _requester(provider: str = '', requester_name: str = '') -> LiteLLMRequester:
+    return LiteLLMRequester(
+        SimpleNamespace(),
+        {
+            'custom_llm_provider': provider,
+            'requester_name': requester_name,
+        },
+    )
+
+
+def test_reasoning_config_normalization_and_conflicts():
+    assert reasoning.normalize_reasoning_config(None) == {'level': 'provider_default'}
+    assert reasoning.normalize_reasoning_config({}) == {'level': 'provider_default'}
+    assert reasoning.validate_reasoning_config(
+        {'level': 'high'},
+        ['reasoning'],
+        {},
+    ) == {'level': 'high'}
+
+    with pytest.raises(ValueError, match='Unsupported reasoning level'):
+        reasoning.normalize_reasoning_config({'level': 'turbo'})
+    with pytest.raises(ValueError, match='reasoning ability'):
+        reasoning.validate_reasoning_config({'level': 'low'}, [], {})
+    with pytest.raises(ValueError, match='extra_body.thinking_budget'):
+        reasoning.validate_reasoning_config(
+            {'level': 'low'},
+            ['reasoning'],
+            {'extra_body': {'thinking_budget': 1024}},
+        )
+    assert reasoning.find_reasoning_arg_conflicts(
+        {
+            'enable_thinking': True,
+            'extra_body': {'reasoning_effort': 'high'},
+        }
+    ) == ['enable_thinking', 'extra_body.reasoning_effort']
+
+
+def test_manual_reasoning_model_without_known_protocol_stays_conservative(monkeypatch):
+    request = _requester()
+    monkeypatch.setattr(request, '_supports_reasoning', lambda _: False)
+
+    capabilities = request.get_reasoning_capabilities(_runtime_model(request))
+
+    assert capabilities == {
+        'supported': True,
+        'levels': ['provider_default'],
+        'source': 'manual',
+    }
+
+
+def test_openai_protocol_does_not_mark_unknown_models_as_reasoning(monkeypatch):
+    request = _requester('openai', 'openai-chat-completions')
+    monkeypatch.setattr(request, '_supports_reasoning', lambda _: False)
+
+    capabilities = request.get_reasoning_capabilities(
+        _runtime_model(request, name='future-reasoning-model', abilities=[])
+    )
+
+    assert capabilities == {
+        'supported': False,
+        'levels': ['provider_default'],
+        'source': 'unknown',
+    }
+
+
+def test_unknown_unmarked_model_without_provider_stays_safe(monkeypatch):
+    request = _requester()
+    monkeypatch.setattr(request, '_supports_reasoning', lambda _: False)
+
+    capabilities = request.get_reasoning_capabilities(_runtime_model(request, name='unknown-model', abilities=[]))
+
+    assert capabilities == {
+        'supported': False,
+        'levels': ['provider_default'],
+        'source': 'unknown',
+    }
+
+
+def test_mimo_exposes_off_on_without_fake_effort_levels(monkeypatch):
+    request = _requester('openai', 'mimo-chat-completions')
+    monkeypatch.setattr(request, '_supports_reasoning', lambda _: False)
+
+    capabilities = request.get_reasoning_capabilities(_runtime_model(request, name='mimo-v2.5', abilities=[]))
+
+    assert capabilities == {
+        'supported': True,
+        'levels': ['provider_default', 'disabled', 'enabled'],
+        'source': 'provider',
+    }
+
+
+def test_openai_reasoning_levels_follow_litellm_metadata(monkeypatch):
+    request = _requester('openai', 'openai-chat-completions')
+    monkeypatch.setattr(request, '_supports_reasoning', lambda _: True)
+    monkeypatch.setattr(
+        request,
+        '_safe_model_info',
+        lambda _: {
+            'supports_none_reasoning_effort': True,
+            'supports_minimal_reasoning_effort': False,
+            'supports_low_reasoning_effort': True,
+            'supports_xhigh_reasoning_effort': True,
+        },
+    )
+
+    capabilities = request.get_reasoning_capabilities(_runtime_model(request, name='gpt-5'))
+
+    assert capabilities['source'] == 'litellm'
+    assert capabilities['levels'] == [
+        'provider_default',
+        'disabled',
+        'low',
+        'medium',
+        'high',
+        'xhigh',
+    ]
+
+
+def test_anthropic_adaptive_and_always_on_profiles(monkeypatch):
+    request = _requester('anthropic', 'anthropic-messages')
+    monkeypatch.setattr(request, '_supports_reasoning', lambda _: False)
+
+    adaptive = request.get_reasoning_capabilities(_runtime_model(request, name='claude-sonnet-4-6', abilities=[]))
+    assert adaptive['levels'] == [
+        'provider_default',
+        'disabled',
+        'low',
+        'medium',
+        'high',
+        'xhigh',
+        'max',
+    ]
+
+    always_on = request.get_reasoning_capabilities(_runtime_model(request, name='claude-fable-5', abilities=[]))
+    assert 'disabled' not in always_on['levels']
+
+    legacy = request.get_reasoning_capabilities(_runtime_model(request, name='claude-3-5-sonnet', abilities=[]))
+    assert legacy['levels'] == ['provider_default', 'low', 'medium', 'high']
+
+
+def test_deepseek_profiles_match_model_generation(monkeypatch):
+    request = _requester('deepseek', 'deepseek-chat-completions')
+    monkeypatch.setattr(request, '_supports_reasoning', lambda _: False)
+
+    assert request.get_reasoning_capabilities(_runtime_model(request, name='deepseek-v4-flash', abilities=[]))[
+        'levels'
+    ] == ['provider_default', 'disabled', 'low', 'high', 'xhigh', 'max']
+    assert request.get_reasoning_capabilities(_runtime_model(request, name='deepseek-chat', abilities=[]))[
+        'levels'
+    ] == ['provider_default', 'disabled', 'enabled']
+    assert request.get_reasoning_capabilities(_runtime_model(request, name='deepseek-r1', abilities=[]))['levels'] == [
+        'provider_default'
+    ]
+
+
+@pytest.mark.parametrize(
+    ('model_name', 'expected_levels'),
+    [
+        ('kimi-k3', ['provider_default', 'low', 'high', 'max']),
+        ('kimi-k2.7-code', ['provider_default']),
+        ('kimi-k2.6', ['provider_default', 'disabled', 'enabled']),
+        ('kimi-k2.5', ['provider_default', 'disabled', 'enabled']),
+    ],
+)
+def test_kimi_profiles(model_name, expected_levels, monkeypatch):
+    request = _requester('openai', 'moonshot-chat-completions')
+    monkeypatch.setattr(request, '_supports_reasoning', lambda _: False)
+
+    capabilities = request.get_reasoning_capabilities(_runtime_model(request, name=model_name, abilities=[]))
+
+    assert capabilities['levels'] == expected_levels
+
+
+def test_qwen_mixed_and_dedicated_thinking_profiles(monkeypatch):
+    request = _requester('openai', 'bailian-chat-completions')
+    monkeypatch.setattr(request, '_supports_reasoning', lambda _: False)
+
+    mixed = request.get_reasoning_capabilities(_runtime_model(request, name='qwen-plus', abilities=[]))
+    dedicated = request.get_reasoning_capabilities(
+        _runtime_model(request, name='qwen3-235b-a22b-thinking-2507', abilities=[])
+    )
+
+    assert mixed['levels'] == ['provider_default', 'disabled', 'enabled']
+    assert dedicated['levels'] == ['provider_default', 'low', 'medium', 'high']
+
+
+def test_qwen3_exposes_budget_based_reasoning_levels(monkeypatch):
+    request = _requester('openai', 'bailian-chat-completions')
+    monkeypatch.setattr(request, '_supports_reasoning', lambda _: False)
+
+    mixed = request.get_reasoning_capabilities(_runtime_model(request, name='qwen3.8-max', abilities=[]))
+    dedicated = request.get_reasoning_capabilities(_runtime_model(request, name='qwen3.7-max-preview', abilities=[]))
+
+    assert mixed['levels'] == ['provider_default', 'disabled', 'low', 'medium', 'high']
+    assert mixed['legacy_levels'] == ['enabled']
+    assert dedicated['levels'] == ['provider_default', 'low', 'medium', 'high']
+
+
+def test_qwen3_legacy_enabled_config_remains_supported(monkeypatch):
+    request = _requester('openai', 'bailian-chat-completions')
+    monkeypatch.setattr(request, '_supports_reasoning', lambda _: False)
+
+    assert request._build_reasoning_args(_runtime_model(request, 'enabled', name='qwen3.8-max')) == {
+        'extra_body': {'enable_thinking': True}
+    }
+
+
+@pytest.mark.parametrize(
+    ('level', 'budget'),
+    [('low', 1024), ('medium', 4096), ('high', 8192)],
+)
+def test_qwen3_reasoning_levels_translate_to_thinking_budget(level, budget, monkeypatch):
+    request = _requester('openai', 'bailian-chat-completions')
+    monkeypatch.setattr(request, '_supports_reasoning', lambda _: False)
+
+    assert request._build_reasoning_args(_runtime_model(request, level, name='qwen3.8-max')) == {
+        'extra_body': {
+            'enable_thinking': True,
+            'thinking_budget': budget,
+        }
+    }
+
+
+@pytest.mark.parametrize('model_name', ['qwen3.7-max-preview', 'qwen3.7-max-2026-05-17'])
+def test_qwen_dedicated_thinking_release_models_are_not_toggleable(model_name, monkeypatch):
+    request = _requester('openai', 'bailian-chat-completions')
+    monkeypatch.setattr(request, '_supports_reasoning', lambda _: False)
+
+    capabilities = request.get_reasoning_capabilities(_runtime_model(request, name=model_name, abilities=[]))
+
+    assert capabilities['levels'] == ['provider_default', 'low', 'medium', 'high']
+
+
+@pytest.mark.parametrize(
+    ('model_name', 'expected_levels'),
+    [
+        ('kimi-k2.6', ['provider_default', 'disabled', 'enabled']),
+        ('kimi-k2.5', ['provider_default', 'disabled', 'enabled']),
+        ('kimi-k2.7-code', ['provider_default']),
+        ('kimi-k2-thinking', ['provider_default']),
+    ],
+)
+def test_bailian_kimi_profiles_use_kimi_model_rules(model_name, expected_levels, monkeypatch):
+    request = _requester('openai', 'bailian-chat-completions')
+    monkeypatch.setattr(request, '_supports_reasoning', lambda _: False)
+
+    capabilities = request.get_reasoning_capabilities(_runtime_model(request, name=model_name, abilities=[]))
+
+    assert capabilities['levels'] == expected_levels
+
+
+def test_bailian_kimi_uses_thinking_protocol_instead_of_qwen_protocol(monkeypatch):
+    request = _requester('openai', 'bailian-chat-completions')
+    monkeypatch.setattr(request, '_supports_reasoning', lambda _: False)
+
+    assert request._build_reasoning_args(_runtime_model(request, 'disabled', name='kimi-k2.6')) == {
+        'extra_body': {'thinking': {'type': 'disabled'}}
+    }
+
+
+def test_doubao_exposes_documented_effort_range(monkeypatch):
+    request = _requester('openai', 'doubao-chat-completions')
+    monkeypatch.setattr(request, '_supports_reasoning', lambda _: False)
+
+    capabilities = request.get_reasoning_capabilities(
+        _runtime_model(request, name='doubao-seed-2-1-pro-260628', abilities=[])
+    )
+
+    assert capabilities['levels'] == ['provider_default', 'disabled', 'low', 'medium', 'high']
+
+
+@pytest.mark.parametrize(
+    ('model_name', 'expected_levels'),
+    [
+        ('gpt-5', ['provider_default', 'low', 'medium', 'high']),
+        (
+            'claude-sonnet-4-6',
+            ['provider_default', 'disabled', 'low', 'medium', 'high', 'xhigh', 'max'],
+        ),
+        ('deepseek-v4-flash', ['provider_default', 'disabled', 'low', 'high', 'xhigh', 'max']),
+        ('kimi-k2.6', ['provider_default', 'disabled', 'enabled']),
+        ('qwen-plus', ['provider_default', 'disabled', 'enabled']),
+        ('doubao-seed-2-1-pro-260628', ['provider_default', 'disabled', 'low', 'medium', 'high']),
+        ('mimo-v2.5', ['provider_default', 'disabled', 'enabled']),
+    ],
+)
+def test_new_api_infers_upstream_protocol_from_model_name(model_name, expected_levels, monkeypatch):
+    request = _requester('openai', 'new-api-chat-completions')
+    monkeypatch.setattr(request, '_supports_reasoning', lambda _: False)
+    monkeypatch.setattr(request, '_safe_model_info', lambda _: {})
+
+    capabilities = request.get_reasoning_capabilities(_runtime_model(request, name=model_name, abilities=[]))
+
+    assert capabilities['levels'] == expected_levels
+
+
+@pytest.mark.parametrize(
+    ('provider', 'requester_name', 'model_name'),
+    [
+        ('openai', 'openai-chat-completions', 'gpt-5'),
+        ('anthropic', 'anthropic-messages', 'claude-sonnet-4-6'),
+        ('deepseek', 'deepseek-chat-completions', 'deepseek-v4-flash'),
+        ('openai', 'mimo-chat-completions', 'mimo-v2.5'),
+        ('openai', 'moonshot-chat-completions', 'kimi-k2.6'),
+        ('openai', 'bailian-chat-completions', 'qwen-plus'),
+        ('openai', 'doubao-chat-completions', 'doubao-seed-2-1-pro-260628'),
+        ('openai', 'new-api-chat-completions', 'deepseek-v4-flash'),
+    ],
+)
+def test_scanned_known_models_gain_reasoning_ability(provider, requester_name, model_name, monkeypatch):
+    request = _requester(provider, requester_name)
+    monkeypatch.setattr(request, '_supports_reasoning', lambda _: False)
+    monkeypatch.setattr(request, '_supports_function_calling', lambda _: False)
+    monkeypatch.setattr(request, '_supports_vision', lambda _: False)
+    monkeypatch.setattr(request, '_safe_context_length', lambda _: None)
+
+    scanned = request._enrich_scanned_model(model_name)
+
+    assert scanned['abilities'] == ['reasoning']
+
+
+def test_new_api_unknown_alias_stays_conservative(monkeypatch):
+    request = _requester('openai', 'new-api-chat-completions')
+    monkeypatch.setattr(request, '_supports_reasoning', lambda _: False)
+
+    capabilities = request.get_reasoning_capabilities(
+        _runtime_model(request, name='company-internal-alias', abilities=[])
+    )
+
+    assert capabilities == {
+        'supported': False,
+        'levels': ['provider_default'],
+        'source': 'unknown',
+    }
+
+
+def test_reasoning_argument_translation(monkeypatch):
+    openai_request = _requester('openai', 'openai-chat-completions')
+    monkeypatch.setattr(openai_request, '_supports_reasoning', lambda _: True)
+    monkeypatch.setattr(openai_request, '_safe_model_info', lambda _: {'supports_none_reasoning_effort': True})
+    assert openai_request._build_reasoning_args(_runtime_model(openai_request, 'disabled', name='gpt-5')) == {
+        'reasoning_effort': 'none'
+    }
+
+    anthropic_request = _requester('anthropic', 'anthropic-messages')
+    assert anthropic_request._build_reasoning_args(
+        _runtime_model(anthropic_request, 'disabled', name='claude-sonnet-4-6')
+    ) == {'thinking': {'type': 'disabled'}}
+
+    deepseek_request = _requester('deepseek', 'deepseek-chat-completions')
+    assert deepseek_request._build_reasoning_args(
+        _runtime_model(deepseek_request, 'high', name='deepseek-v4-flash')
+    ) == {
+        'extra_body': {
+            'thinking': {'type': 'enabled'},
+            'reasoning_effort': 'high',
+        }
+    }
+
+    kimi_request = _requester('openai', 'moonshot-chat-completions')
+    assert kimi_request._build_reasoning_args(_runtime_model(kimi_request, 'enabled', name='kimi-k2.6')) == {
+        'extra_body': {'thinking': {'type': 'enabled'}}
+    }
+    assert kimi_request._build_reasoning_args(_runtime_model(kimi_request, 'high', name='kimi-k3')) == {
+        'reasoning_effort': 'high'
+    }
+
+    qwen_request = _requester('openai', 'bailian-chat-completions')
+    assert qwen_request._build_reasoning_args(_runtime_model(qwen_request, 'disabled', name='qwen-plus')) == {
+        'extra_body': {'enable_thinking': False}
+    }
+
+    doubao_request = _requester('openai', 'doubao-chat-completions')
+    assert doubao_request._build_reasoning_args(
+        _runtime_model(doubao_request, 'high', name='doubao-seed-2-1-pro-260628')
+    ) == {'reasoning_effort': 'high'}
+
+    mimo_request = _requester('openai', 'mimo-chat-completions')
+    assert mimo_request._build_reasoning_args(_runtime_model(mimo_request, 'disabled', name='mimo-v2.5')) == {
+        'extra_body': {'thinking': {'type': 'disabled'}}
+    }
+
+
+def test_pipeline_reasoning_override_takes_precedence(monkeypatch):
+    request = _requester('openai', 'openai-chat-completions')
+    monkeypatch.setattr(request, '_supports_reasoning', lambda _: True)
+    monkeypatch.setattr(request, '_safe_model_info', lambda _: {})
+    model = _runtime_model(request, 'high', name='gpt-5')
+
+    model.reasoning_config_override = {'level': 'provider_default'}
+    assert request._build_reasoning_args(model) == {}
+
+    model.reasoning_config_override = {'level': 'low'}
+    assert request._build_reasoning_args(model) == {'reasoning_effort': 'low'}
+
+
+def test_always_on_reasoning_models_do_not_offer_disabled(monkeypatch):
+    deepseek_request = _requester('deepseek', 'deepseek-chat-completions')
+    monkeypatch.setattr(deepseek_request, '_supports_reasoning', lambda _: True)
+    monkeypatch.setattr(deepseek_request, '_safe_model_info', lambda _: {})
+    deepseek_capabilities = deepseek_request.get_reasoning_capabilities(
+        _runtime_model(deepseek_request, name='deepseek-r1')
+    )
+    assert deepseek_capabilities['levels'] == ['provider_default']
+
+    gemini_request = _requester('gemini')
+    monkeypatch.setattr(gemini_request, '_supports_reasoning', lambda _: True)
+    monkeypatch.setattr(
+        gemini_request,
+        '_safe_model_info',
+        lambda _: {'supports_none_reasoning_effort': True},
+    )
+    gemini_capabilities = gemini_request.get_reasoning_capabilities(_runtime_model(gemini_request, name='gemini-3-pro'))
+    assert 'disabled' not in gemini_capabilities['levels']
+    with pytest.raises(errors.RequesterError, match='not supported'):
+        gemini_request._build_reasoning_args(_runtime_model(gemini_request, 'disabled', name='gemini-3-pro'))
+
+
+def test_non_target_provider_capabilities_remain_supported(monkeypatch):
+    ollama_request = _requester('ollama', 'ollama')
+    monkeypatch.setattr(ollama_request, '_supports_reasoning', lambda _: False)
+    monkeypatch.setattr(ollama_request, '_safe_model_info', lambda _: {})
+
+    toggle_capabilities = ollama_request.get_reasoning_capabilities(_runtime_model(ollama_request, name='qwen3'))
+    assert toggle_capabilities['levels'] == [
+        'provider_default',
+        'disabled',
+        'enabled',
+    ]
+    assert ollama_request._build_reasoning_args(_runtime_model(ollama_request, 'enabled', name='qwen3')) == {
+        'reasoning_effort': 'low'
+    }
+
+    effort_capabilities = ollama_request.get_reasoning_capabilities(_runtime_model(ollama_request, name='gpt-oss:20b'))
+    assert effort_capabilities['levels'] == [
+        'provider_default',
+        'disabled',
+        'low',
+        'medium',
+        'high',
+    ]
+    assert ollama_request._build_reasoning_args(_runtime_model(ollama_request, 'high', name='gpt-oss:20b')) == {
+        'reasoning_effort': 'high'
+    }
+
+    volcengine_request = _requester('volcengine', 'volcark-chat-completions')
+    monkeypatch.setattr(volcengine_request, '_supports_reasoning', lambda _: False)
+    monkeypatch.setattr(volcengine_request, '_safe_model_info', lambda _: {})
+    assert volcengine_request._build_reasoning_args(
+        _runtime_model(volcengine_request, 'disabled', name='doubao-seed')
+    ) == {'extra_body': {'thinking': {'type': 'disabled'}}}
+
+
+def test_explicit_unsupported_level_raises(monkeypatch):
+    request = _requester()
+    monkeypatch.setattr(request, '_supports_reasoning', lambda _: False)
+    monkeypatch.setattr(request, '_safe_model_info', lambda _: {})
+
+    with pytest.raises(errors.RequesterError, match='Available levels: provider_default'):
+        request._build_reasoning_args(_runtime_model(request, 'high', abilities=[]))
+
+
+def test_provider_inference_rejects_levels_outside_conservative_profile(monkeypatch):
+    request = _requester('openai', 'openai-chat-completions')
+    monkeypatch.setattr(request, '_supports_reasoning', lambda _: False)
+    monkeypatch.setattr(request, '_safe_model_info', lambda _: {})
+
+    with pytest.raises(errors.RequesterError, match='Available levels: provider_default, low, medium, high'):
+        request._build_reasoning_args(_runtime_model(request, 'xhigh', name='gpt-5', abilities=[]))
+
+
+@pytest.mark.asyncio
+async def test_completion_args_reject_reasoning_extra_arg_conflicts(monkeypatch):
+    request = _requester('openai', 'openai-chat-completions')
+    monkeypatch.setattr(request, '_supports_reasoning', lambda _: True)
+    monkeypatch.setattr(request, '_safe_model_info', lambda _: {})
+    model = _runtime_model(request, 'high', name='gpt-5')
+    model.model_entity.extra_args = {'reasoning_effort': 'low'}
+    model.provider.token_mgr.get_token = lambda: 'test-token'
+
+    with pytest.raises(errors.RequesterError, match='conflicts with advanced parameters'):
+        await request._build_completion_args(model, [])
+
+
+@pytest.mark.asyncio
+async def test_openai_compatible_reasoning_effort_is_explicitly_allowed(monkeypatch):
+    request = _requester('openai', 'moonshot-chat-completions')
+    model = _runtime_model(request, 'high', name='kimi-k3')
+    model.model_entity.extra_args = {'allowed_openai_params': ['custom_extension']}
+    model.provider.token_mgr.get_token = lambda: 'test-token'
+
+    args = await request._build_completion_args(model, [])
+
+    assert args['reasoning_effort'] == 'high'
+    assert args['allowed_openai_params'] == ['custom_extension', 'reasoning_effort']
+
+
+@pytest.mark.asyncio
+async def test_provider_default_does_not_allow_or_send_reasoning_effort():
+    request = _requester('openai', 'new-api-chat-completions')
+    model = _runtime_model(request, 'provider_default', name='deepseek-v4-flash')
+    model.provider.token_mgr.get_token = lambda: 'test-token'
+
+    args = await request._build_completion_args(model, [])
+
+    assert 'reasoning_effort' not in args
+    assert 'allowed_openai_params' not in args
+
+
+@pytest.mark.asyncio
+async def test_deepseek_disabled_thinking_is_merged_into_extra_body(monkeypatch):
+    request = _requester('deepseek', 'deepseek-chat-completions')
+    monkeypatch.setattr(request, '_supports_reasoning', lambda _: False)
+    model = _runtime_model(request, 'disabled', name='deepseek-chat')
+    model.model_entity.extra_args = {'extra_body': {'custom_extension': True}}
+    model.provider.token_mgr.get_token = lambda: 'test-token'
+
+    args = await request._build_completion_args(model, [])
+
+    assert args['extra_body'] == {
+        'custom_extension': True,
+        'thinking': {'type': 'disabled'},
+    }
+
+
+@pytest.mark.asyncio
+async def test_openai_compatible_reasoning_history_is_promoted_for_tool_continuity():
+    request = _requester('openai', 'mimo-chat-completions')
+    model = _runtime_model(request, 'enabled', name='mimo-v2.5')
+    model.provider.token_mgr.get_token = lambda: 'test-token'
+    history = [
+        provider_message.Message(
+            role='assistant',
+            content='<think>\nprior reasoning\n</think>\nanswer',
+            provider_specific_fields={'reasoning_content': 'prior reasoning'},
+        )
+    ]
+
+    args = await request._build_completion_args(model, history)
+
+    assert args['messages'][0]['reasoning_content'] == 'prior reasoning'
+    assert args['messages'][0]['content'] == 'answer'
+    assert 'provider_specific_fields' not in args['messages'][0]
+
+
+@pytest.mark.asyncio
+async def test_disabling_reasoning_removes_previous_reasoning_context():
+    request = _requester('openai', 'mimo-chat-completions')
+    model = _runtime_model(request, 'disabled', name='mimo-v2.5')
+    model.provider.token_mgr.get_token = lambda: 'test-token'
+    history = [
+        provider_message.Message(
+            role='assistant',
+            content='answer',
+            provider_specific_fields={'reasoning_content': 'prior reasoning'},
+        )
+    ]
+
+    args = await request._build_completion_args(model, history)
+
+    assert 'reasoning_content' not in args['messages'][0]
+    assert 'provider_specific_fields' not in args['messages'][0]
+
+
+@pytest.mark.asyncio
+async def test_anthropic_history_promotes_thinking_blocks_instead_of_reasoning_content():
+    request = _requester('anthropic', 'anthropic-messages')
+    model = _runtime_model(request, 'high', name='claude-sonnet-4-6')
+    model.provider.token_mgr.get_token = lambda: 'test-token'
+    thinking_blocks = [{'type': 'thinking', 'thinking': 'prior reasoning', 'signature': 'sig'}]
+    history = [
+        provider_message.Message(
+            role='assistant',
+            content='',
+            provider_specific_fields={
+                'reasoning_content': 'prior reasoning',
+                'thinking_blocks': thinking_blocks,
+            },
+        )
+    ]
+
+    args = await request._build_completion_args(model, history)
+
+    assert args['messages'][0]['thinking_blocks'] == thinking_blocks
+    assert 'reasoning_content' not in args['messages'][0]
+    assert 'provider_specific_fields' not in args['messages'][0]
+
+
+@pytest.mark.asyncio
+async def test_non_stream_anthropic_thinking_blocks_are_preserved(monkeypatch):
+    request = _requester('anthropic', 'anthropic-messages')
+    request._build_completion_args = AsyncMock(return_value={})
+    thinking_blocks = [{'type': 'thinking', 'thinking': 'private reasoning', 'signature': 'sig'}]
+    response = SimpleNamespace(
+        choices=[
+            SimpleNamespace(
+                message=_Dumpable(
+                    {
+                        'role': 'assistant',
+                        'content': 'answer',
+                        'thinking_blocks': thinking_blocks,
+                    }
+                )
+            )
+        ],
+        usage=None,
+    )
+    monkeypatch.setattr(litellmchat, 'acompletion', AsyncMock(return_value=response))
+
+    message, _ = await request.invoke_llm(None, _runtime_model(request, 'high', name='claude-sonnet-4-6'), [])
+
+    assert message.content == '<think>\nprivate reasoning\n</think>\nanswer'
+    assert message.provider_specific_fields == {'thinking_blocks': thinking_blocks}
+
+
+class _Dumpable:
+    def __init__(self, data: dict):
+        self.data = data
+
+    def model_dump(self) -> dict:
+        return dict(self.data)
+
+
+@pytest.mark.asyncio
+async def test_non_stream_reasoning_content_is_preserved(monkeypatch):
+    request = _requester('deepseek')
+    request._build_completion_args = AsyncMock(return_value={})
+    response = SimpleNamespace(
+        choices=[
+            SimpleNamespace(
+                message=_Dumpable(
+                    {
+                        'role': 'assistant',
+                        'content': 'answer',
+                        'reasoning_content': 'private reasoning',
+                    }
+                )
+            )
+        ],
+        usage=None,
+    )
+    monkeypatch.setattr(litellmchat, 'acompletion', AsyncMock(return_value=response))
+
+    message, _ = await request.invoke_llm(None, _runtime_model(request), [], remove_think=True)
+
+    assert message.content == 'answer'
+    assert message.provider_specific_fields == {'reasoning_content': 'private reasoning'}
+
+
+@pytest.mark.asyncio
+async def test_stream_reasoning_round_trip_with_hidden_display(monkeypatch):
+    request = _requester('deepseek')
+    request._build_completion_args = AsyncMock(return_value={})
+
+    async def chunks():
+        yield SimpleNamespace(
+            choices=[
+                SimpleNamespace(
+                    delta=_Dumpable({'role': 'assistant', 'reasoning_content': 'private '}),
+                    finish_reason=None,
+                )
+            ],
+            usage=None,
+        )
+        yield SimpleNamespace(
+            choices=[
+                SimpleNamespace(
+                    delta=_Dumpable({'content': 'answer'}),
+                    finish_reason='stop',
+                )
+            ],
+            usage=None,
+        )
+
+    monkeypatch.setattr(litellmchat, 'acompletion', AsyncMock(return_value=chunks()))
+    emitted = [
+        chunk
+        async for chunk in request.invoke_llm_stream(
+            None,
+            _runtime_model(request),
+            [],
+            remove_think=True,
+        )
+    ]
+
+    assert ''.join(chunk.content or '' for chunk in emitted) == 'answer'
+    assert (
+        ''.join(
+            chunk.provider_specific_fields.get('reasoning_content', '')
+            for chunk in emitted
+            if chunk.provider_specific_fields
+        )
+        == 'private '
+    )
+
+
+@pytest.mark.asyncio
+async def test_stream_reasoning_content_is_wrapped_for_display(monkeypatch):
+    request = _requester('deepseek')
+    request._build_completion_args = AsyncMock(return_value={})
+
+    async def chunks():
+        yield SimpleNamespace(
+            choices=[
+                SimpleNamespace(
+                    delta=_Dumpable({'role': 'assistant', 'reasoning_content': 'private '}),
+                    finish_reason=None,
+                )
+            ],
+            usage=None,
+        )
+        yield SimpleNamespace(
+            choices=[
+                SimpleNamespace(
+                    delta=_Dumpable({'content': 'answer'}),
+                    finish_reason='stop',
+                )
+            ],
+            usage=None,
+        )
+
+    monkeypatch.setattr(litellmchat, 'acompletion', AsyncMock(return_value=chunks()))
+    emitted = [
+        chunk
+        async for chunk in request.invoke_llm_stream(
+            None,
+            _runtime_model(request),
+            [],
+            remove_think=False,
+        )
+    ]
+
+    assert ''.join(chunk.content or '' for chunk in emitted) == '<think>\nprivate \n</think>\nanswer'
+    assert (
+        ''.join(
+            chunk.provider_specific_fields.get('reasoning_content', '')
+            for chunk in emitted
+            if chunk.provider_specific_fields
+        )
+        == 'private '
+    )
+
+
+@pytest.mark.asyncio
+async def test_stream_anthropic_thinking_blocks_are_preserved(monkeypatch):
+    request = _requester('anthropic', 'anthropic-messages')
+    request._build_completion_args = AsyncMock(return_value={})
+    thinking_blocks = [{'type': 'thinking', 'thinking': 'private ', 'signature': 'sig'}]
+
+    async def chunks():
+        yield SimpleNamespace(
+            choices=[
+                SimpleNamespace(
+                    delta=_Dumpable({'role': 'assistant', 'thinking_blocks': thinking_blocks}),
+                    finish_reason=None,
+                )
+            ],
+            usage=None,
+        )
+        yield SimpleNamespace(
+            choices=[
+                SimpleNamespace(
+                    delta=_Dumpable({'content': 'answer'}),
+                    finish_reason='stop',
+                )
+            ],
+            usage=None,
+        )
+
+    monkeypatch.setattr(litellmchat, 'acompletion', AsyncMock(return_value=chunks()))
+    emitted = [
+        chunk
+        async for chunk in request.invoke_llm_stream(
+            None,
+            _runtime_model(request, 'high', name='claude-sonnet-4-6'),
+            [],
+            remove_think=False,
+        )
+    ]
+
+    assert ''.join(chunk.content or '' for chunk in emitted) == '<think>\nprivate \n</think>\nanswer'
+    assert (
+        next(
+            chunk.provider_specific_fields['thinking_blocks']
+            for chunk in emitted
+            if chunk.provider_specific_fields and 'thinking_blocks' in chunk.provider_specific_fields
+        )
+        == thinking_blocks
+    )
+
+
+@pytest.mark.asyncio
+async def test_hidden_thinking_does_not_drop_same_delta_tool_call(monkeypatch):
+    request = _requester('openai', 'openai-chat-completions')
+    request._build_completion_args = AsyncMock(return_value={})
+
+    async def chunks():
+        yield SimpleNamespace(
+            choices=[
+                SimpleNamespace(
+                    delta=_Dumpable(
+                        {
+                            'content': '<think>hidden</think>',
+                            'tool_calls': [
+                                {
+                                    'index': 0,
+                                    'id': 'call_1',
+                                    'type': 'function',
+                                    'function': {'name': 'lookup', 'arguments': '{}'},
+                                }
+                            ],
+                        }
+                    ),
+                    finish_reason='tool_calls',
+                )
+            ],
+            usage=None,
+        )
+
+    monkeypatch.setattr(litellmchat, 'acompletion', AsyncMock(return_value=chunks()))
+    collected = [
+        chunk
+        async for chunk in request.invoke_llm_stream(
+            None,
+            _runtime_model(request, 'provider_default'),
+            [],
+            remove_think=True,
+        )
+    ]
+
+    assert len(collected) == 1
+    assert collected[0].tool_calls[0].id == 'call_1'

tests/unit_tests/provider/test_requester_base.py

@@ -510,6 +510,7 @@ def test_runtime_llm_model_initialization(runtime_llm_model, fake_persistence_da
     assert model.model_entity.abilities == model_entity.abilities
     assert model.model_entity.extra_args == model_entity.extra_args
     assert model.provider is not None
+    assert model.reasoning_config_override is None
 
 
 def test_runtime_llm_model_provider_ref(runtime_llm_model):

tests/unit_tests/telemetry/test_heartbeat.py

@@ -3,6 +3,7 @@
 from __future__ import annotations
 
 import json
+from datetime import datetime, timezone
 from types import SimpleNamespace
 
 import pytest
@@ -14,6 +15,12 @@ def get_heartbeat_module():
     return import_module('langbot.pkg.telemetry.heartbeat')
 
 
+def test_workspace_created_timestamp_treats_naive_database_values_as_utc():
+    heartbeat = get_heartbeat_module()
+    created_at = datetime(2026, 8, 4, 0, 0, 0)
+    assert heartbeat._workspace_created_timestamp(created_at) == 1785801600
+
+
 def make_app():
     ap = Mock()
     ap.instance_config = Mock()
@@ -57,15 +64,17 @@ def make_app():
 
 class TestBuildHeartbeatPayload:
     @pytest.mark.asyncio
-    async def test_payload_shape(self):
+    async def test_payload_shape(self, monkeypatch):
         heartbeat = get_heartbeat_module()
+        monkeypatch.setattr(heartbeat.constants, 'instance_id', 'instance-test')
         ap = make_app()
         payload = await heartbeat.build_heartbeat_payload(ap, workspace_uuid='workspace-a')
 
         assert payload['event_type'] == 'instance_heartbeat'
         assert payload['query_id'] == ''
         assert payload['workspace_uuid'] == 'workspace-a'
-        assert 'instance_id' not in payload
+        assert payload['instance_id']
+        assert payload['workspace_create_ts'] == 0
         assert 'instance_create_ts' in payload
         assert 'timestamp' in payload
         f = payload['features']
@@ -100,8 +109,9 @@ class TestBuildHeartbeatPayload:
         assert payload['features']['pipeline_count'] == -1
 
     @pytest.mark.asyncio
-    async def test_cloud_counts_loaded_registries_without_tenant_sql(self):
+    async def test_cloud_counts_loaded_registries_without_tenant_sql(self, monkeypatch):
         heartbeat = get_heartbeat_module()
+        monkeypatch.setattr(heartbeat.constants, 'instance_id', 'instance-test')
         ap = make_app()
         ap.persistence_mgr.mode = SimpleNamespace(value='cloud_runtime')
         ap.persistence_mgr.execute_async = AsyncMock(
@@ -140,7 +150,11 @@ class TestBuildHeartbeatPayload:
         ap.workspace_service.list_active_execution_bindings = AsyncMock(
             return_value=[
                 SimpleNamespace(workspace_uuid='workspace-a', placement_generation=7),
-                SimpleNamespace(workspace_uuid='workspace-b', placement_generation=9),
+                SimpleNamespace(
+                    workspace_uuid='workspace-b',
+                    placement_generation=9,
+                    workspace_created_at=datetime(2026, 8, 4, tzinfo=timezone.utc),
+                ),
             ],
         )
         ap.platform_mgr._bots_by_key[('instance-a', 'workspace-b', 'bot-b')] = SimpleNamespace(
@@ -150,7 +164,9 @@ class TestBuildHeartbeatPayload:
         payloads = await heartbeat.build_heartbeat_payloads(ap)
 
         assert [payload['workspace_uuid'] for payload in payloads] == ['workspace-a', 'workspace-b']
-        assert all('instance_id' not in payload for payload in payloads)
+        assert all(payload['instance_id'] for payload in payloads)
+        assert payloads[0]['workspace_create_ts'] == 0
+        assert payloads[1]['workspace_create_ts'] == 1785801600
         by_workspace = {payload['workspace_uuid']: payload['features'] for payload in payloads}
         assert by_workspace['workspace-a']['pipeline_count'] == 2
         assert by_workspace['workspace-a']['mcp_server_count'] == 3

tests/unit_tests/telemetry/test_telemetry.py

@@ -596,6 +596,36 @@ class TestTelemetryManagedRuntimeAuthentication:
         assert captured['headers'] == {'X-LangBot-Telemetry-Token': 'managed-runtime-secret'}
 
 
+class TestAuthenticatedWorkspaceReporter:
+    @pytest.mark.asyncio
+    async def test_workspace_owner_access_token_is_sent_as_bearer(self):
+        telemetry = get_telemetry_module()
+        mock_app = Mock()
+        mock_app.logger = Mock()
+        mock_app.user_service = Mock()
+        mock_app.user_service.get_workspace_owner = AsyncMock(
+            return_value=Mock(user='owner@example.com', space_access_token='expired-token')
+        )
+        mock_app.space_service = Mock()
+        mock_app.space_service.get_valid_access_token = AsyncMock(return_value='refreshed-workspace-owner-token')
+        manager = telemetry.TelemetryManager(mock_app)
+        manager.telemetry_config = {'url': 'https://example.com'}
+
+        response = Mock(status_code=200, text='')
+        response.json = Mock(return_value={'code': 0})
+        mock_client = Mock()
+        mock_client.post = Mock(return_value=response)
+
+        with patch.object(httpx, 'AsyncClient', return_value=mock_client):
+            await manager.send({'query_id': 'q-1', 'workspace_uuid': 'workspace-1'})
+
+        mock_app.user_service.get_workspace_owner.assert_awaited_once_with('workspace-1')
+        mock_app.space_service.get_valid_access_token.assert_awaited_once_with('owner@example.com')
+        assert mock_client.post.call_args.kwargs['headers'] == {
+            'Authorization': 'Bearer refreshed-workspace-owner-token'
+        }
+
+
 class TestStartSendTask:
     """Tests for start_send_task() method."""
 

tests/unit_tests/telemetry/test_workspace_identity.py

@@ -7,25 +7,28 @@ from types import SimpleNamespace
 def test_standard_oss_instance_id_aligns_to_embedded_uuid():
     from langbot.pkg.workspace.identity import workspace_uuid_from_instance_id
 
-    instance_uuid = "a711d9e4-0953-443f-a0e9-7dd50193a79f"
+    instance_uuid = 'a711d9e4-0953-443f-a0e9-7dd50193a79f'
 
     assert workspace_uuid_from_instance_id(instance_uuid) == instance_uuid
-    assert workspace_uuid_from_instance_id(f"instance_{instance_uuid}") == instance_uuid
+    assert workspace_uuid_from_instance_id(f'instance_{instance_uuid}') == instance_uuid
 
 
 def test_custom_legacy_instance_id_maps_to_stable_valid_uuid():
     from langbot.pkg.workspace.identity import workspace_uuid_from_instance_id
 
-    first = workspace_uuid_from_instance_id("instance_migration_test")
-    second = workspace_uuid_from_instance_id("instance_migration_test")
+    first = workspace_uuid_from_instance_id('instance_migration_test')
+    second = workspace_uuid_from_instance_id('instance_migration_test')
 
     assert first == second
     assert str(uuid.UUID(first)) == first
 
 
-def test_query_telemetry_identity_uses_execution_workspace_only():
+def test_query_telemetry_identity_reports_instance_and_workspace():
     from langbot.pkg.telemetry.identity import workspace_identity
 
-    identity = workspace_identity(SimpleNamespace(workspace_uuid="workspace-a", instance_uuid="instance-a"))
+    identity = workspace_identity(SimpleNamespace(workspace_uuid='workspace-a', instance_uuid='instance-a'))
 
-    assert identity == {"workspace_uuid": "workspace-a"}
+    assert identity == {
+        'instance_id': 'instance-a',
+        'workspace_uuid': 'workspace-a',
+    }

tests/unit_tests/vector/test_mgr.py

@@ -204,7 +204,7 @@ class TestVectorDBManagerInitialization:
                 mock_app,
                 connection_string='postgresql://user:pass@host:5432/langbot',
                 use_business_database=False,
-                allowed_dimensions=[384, 512, 768, 1024, 1536],
+                allowed_dimensions=[384, 512, 768, 1024, 1536, 3072],
             )
 
     @pytest.mark.asyncio
@@ -241,7 +241,7 @@ class TestVectorDBManagerInitialization:
                 user='admin',
                 password='secret',
                 use_business_database=False,
-                allowed_dimensions=[384, 512, 768, 1024, 1536],
+                allowed_dimensions=[384, 512, 768, 1024, 1536, 3072],
             )
 
     @pytest.mark.asyncio
@@ -269,7 +269,7 @@ class TestVectorDBManagerInitialization:
                 user='postgres',
                 password='postgres',
                 use_business_database=False,
-                allowed_dimensions=[384, 512, 768, 1024, 1536],
+                allowed_dimensions=[384, 512, 768, 1024, 1536, 3072],
             )
 
     @pytest.mark.asyncio

tests/unit_tests/workspace/test_workspace_collaboration.py

@@ -99,6 +99,7 @@ async def test_invitation_secret_is_hashed_and_acceptance_is_one_time(collaborat
     membership = await service.accept_invitation(created.token, account.uuid)
     assert membership.workspace_uuid == workspace.uuid
     assert membership.role == 'developer'
+    assert membership.source == 'local'
 
     with pytest.raises(InvitationUsedError):
         await service.accept_invitation(created.token, account.uuid)

tests/unit_tests/workspace/test_workspace_service.py

@@ -153,6 +153,33 @@ async def test_initial_owner_cannot_be_claimed_by_another_account(workspace_test
         ).all()
         assert len(owners) == 1
         assert owners[0].account_uuid == first_account_uuid
+        assert owners[0].source == 'local'
+
+
+async def test_claim_initial_owner_reclassifies_existing_membership_as_local(workspace_test_context):
+    service, session_factory = workspace_test_context
+
+    async with session_factory() as session:
+        async with session.begin():
+            account_uuid = await _insert_account(session, 'reclaimed@example.com')
+            workspace = await service.ensure_singleton_workspace(session=session)
+            session.add(
+                WorkspaceMembership(
+                    uuid='44444444-4444-4444-8444-444444444444',
+                    workspace_uuid=workspace.uuid,
+                    account_uuid=account_uuid,
+                    role='viewer',
+                    status='removed',
+                    source='cloud_projection',
+                    projection_revision=4,
+                )
+            )
+
+    membership = await service.claim_initial_owner(account_uuid)
+
+    assert membership.role == 'owner'
+    assert membership.status == 'active'
+    assert membership.source == 'local'
 
 
 async def test_execution_binding_returns_persisted_generation(workspace_test_context):