Propagate agent runner model usage context

docs/agent-runner-pluginization/PROTOCOL_V1.md

@@ -475,8 +475,11 @@ Host 必须校验 `state.updated` 的 scope、key、value 大小和 JSON 可序
 ```python
 # Model
 await api.invoke_llm(llm_model_uuid, messages, funcs=None, extra_args=None)
+await api.invoke_llm_with_usage(llm_model_uuid, messages, funcs=None, extra_args=None)
 async for chunk in api.invoke_llm_stream(llm_model_uuid, messages, funcs=None, extra_args=None):
     ...
+async for event in api.invoke_llm_stream_events(llm_model_uuid, messages, funcs=None, extra_args=None):
+    ...
 await api.invoke_rerank(rerank_model_id, query, documents, top_k=None)
 
 # Tool
@@ -519,6 +522,16 @@ await api.get_langbot_version()
 `llm_model_uuid`，wire payload 字段也是 `llm_model_uuid`。该值对 runner
 仍是 opaque identifier，不应解析其内部格式。
 
+`invoke_llm()` 和 `invoke_llm_stream()` 保持兼容：前者返回 `Message`，后者只
+yield `MessageChunk`。需要 provider 真实 token 计量的 runner 应使用
+`invoke_llm_with_usage()` 或 `invoke_llm_stream_events()`。Host response 可在
+原有 `{message: ...}` / `{chunk: ...}` 外额外携带可选 `usage` 字段；streaming
+场景允许在所有 chunk 之后追加一个 usage-only event。`usage` 至少保留
+OpenAI-compatible 的 `prompt_tokens`、`completion_tokens`、`total_tokens`，
+若 provider 返回 `prompt_tokens_details` / `completion_tokens_details` 或
+cache token counters，Host / SDK 不应丢弃这些字段。没有 usage 的 provider
+必须继续返回成功响应，SDK 将 usage 置为 `None`。
+
 `get_prompt()` 返回当前 query-backed run 的 Host effective prompt messages：
 `list[Message]` 的 JSON 形式。该能力只在 `ctx.context.available_apis.prompt_get`
 为 true 时可用；没有 query 缓存、prompt 已过期或非 query entry run 时 Host

src/langbot/pkg/agent/runner/context_builder.py

@@ -179,6 +179,52 @@ class AgentRunContextBuilder:
     def __init__(self, ap: app.Application):
         self.ap = ap
 
+    @staticmethod
+    def _positive_int(value: typing.Any) -> int | None:
+        if isinstance(value, bool):
+            return None
+        if isinstance(value, int) and value > 0:
+            return value
+        if isinstance(value, str) and value.isdigit():
+            parsed_value = int(value)
+            if parsed_value > 0:
+                return parsed_value
+        return None
+
+    @staticmethod
+    def _is_llm_model_resource(model_resource: ModelResource) -> bool:
+        operations = model_resource.get('operations')
+        if isinstance(operations, list) and operations:
+            return bool({'invoke', 'stream'} & {str(operation) for operation in operations})
+        return model_resource.get('model_type') != 'rerank'
+
+    async def _build_model_context_window_tokens(self, resources: AgentResources) -> int | None:
+        model_mgr = getattr(self.ap, 'model_mgr', None)
+        if model_mgr is None:
+            return None
+
+        for model_resource in resources.get('models', []):
+            if not self._is_llm_model_resource(model_resource):
+                continue
+
+            model_uuid = model_resource.get('model_id')
+            if not isinstance(model_uuid, str) or not model_uuid:
+                continue
+
+            try:
+                model = await model_mgr.get_model_by_uuid(model_uuid)
+            except Exception as exc:
+                logger = getattr(self.ap, 'logger', None)
+                if logger is not None:
+                    logger.debug(f'Failed to resolve model context window for {model_uuid}: {exc}')
+                continue
+
+            model_entity = getattr(model, 'model_entity', None)
+            context_length = self._positive_int(getattr(model_entity, 'context_length', None))
+            return context_length
+
+        return None
+
     async def build_context_from_event(
         self,
         event: AgentEventEnvelope,
@@ -270,6 +316,8 @@ class AgentRunContextBuilder:
         persistent_state_store = get_persistent_state_store(self.ap.persistence_mgr.get_db_engine())
         state: AgentRunState = await persistent_state_store.build_snapshot_from_event(event, binding, descriptor)
 
+        model_context_window_tokens = await self._build_model_context_window_tokens(resources)
+
         # Build runtime context
         runtime: AgentRuntimeContext = {
             'langbot_version': self.ap.ver_mgr.get_current_version(),
@@ -279,10 +327,7 @@ class AgentRunContextBuilder:
                 'bot_id': event.bot_id,
                 'workspace_id': event.workspace_id,
                 'streaming_supported': event.delivery.supports_streaming,
-                'model_context_window_tokens': None,
+                'model_context_window_tokens': model_context_window_tokens,
-                # TODO(model-info): populate model_context_window_tokens after
-                # LiteLLM/model metadata lands. Runners fall back to their
-                # ctx.config until Host can provide the real window.
             },
         }
 

src/langbot/pkg/plugin/handler.py

@@ -21,6 +21,7 @@ import langbot_plugin.api.entities.builtin.resource.tool as resource_tool
 
 from ..entity.persistence import plugin as persistence_plugin
 from ..entity.persistence import bstorage as persistence_bstorage
+from ..provider.modelmgr import requester as model_requester
 
 from ..core import app
 from ..utils import constants
@@ -43,6 +44,18 @@ def _make_rag_error_response(error: Exception, error_type: str, **extra_context)
     return handler.ActionResponse.error(message=message)
 
 
+def _pop_query_llm_usage(query: Any) -> dict[str, Any] | None:
+    """Read provider usage stashed on a query by RuntimeProvider."""
+    if query is None or not getattr(query, 'variables', None):
+        return None
+    usage = query.variables.pop(model_requester.LLM_USAGE_QUERY_VARIABLE, None)
+    if usage is None:
+        return None
+    if isinstance(usage, dict):
+        return dict(usage)
+    return None
+
+
 def _i18n_to_dict(value: Any) -> dict[str, Any]:
     """Convert SDK i18n values to plain dictionaries."""
     if value is None:
@@ -802,10 +815,20 @@ class RuntimeConnectionHandler(handler.Handler):
                 remove_think=remove_think,
             )
 
+            usage = None
+            if isinstance(result, tuple):
+                result, usage = result
+            if usage is None:
+                usage = _pop_query_llm_usage(query)
+
+            response_data = {
+                'message': result.model_dump(),
+            }
+            if usage is not None:
+                response_data['usage'] = usage
+
             return handler.ActionResponse.success(
-                data={
+                data=response_data,
-                    'message': result.model_dump(),
-                },
             )
 
         @self.action(PluginToRuntimeAction.INVOKE_LLM_STREAM)
@@ -867,6 +890,13 @@ class RuntimeConnectionHandler(handler.Handler):
                         'chunk': chunk.model_dump(),
                     },
                 )
+            usage = _pop_query_llm_usage(query)
+            if usage is not None:
+                yield handler.ActionResponse.success(
+                    data={
+                        'usage': usage,
+                    },
+                )
 
         @self.action(PluginToRuntimeAction.CALL_TOOL)
         async def call_tool(data: dict[str, Any]) -> handler.ActionResponse:

src/langbot/pkg/provider/modelmgr/requester.py

@@ -12,6 +12,19 @@ import langbot_plugin.api.entities.builtin.pipeline.query as pipeline_query
 import langbot_plugin.api.entities.builtin.provider.message as provider_message
 
 
+LLM_USAGE_QUERY_VARIABLE = '_llm_usage'
+STREAM_USAGE_QUERY_VARIABLE = '_stream_usage'
+
+
+def _store_llm_usage(query: pipeline_query.Query | None, usage_info: dict | None) -> None:
+    """Store the latest provider usage on the query for upstream action handlers."""
+    if query is None or not usage_info:
+        return
+    if query.variables is None:
+        query.variables = {}
+    query.variables[LLM_USAGE_QUERY_VARIABLE] = dict(usage_info)
+
+
 class RuntimeProvider:
     """运行时模型提供商"""
 
@@ -67,6 +80,7 @@ class RuntimeProvider:
             if isinstance(result, tuple):
                 msg, usage_info = result
                 if usage_info:
+                    _store_llm_usage(query, usage_info)
                     input_tokens = usage_info.get('prompt_tokens', 0)
                     output_tokens = usage_info.get('completion_tokens', 0)
                 return msg
@@ -146,11 +160,12 @@ class RuntimeProvider:
             if query:
                 if query.variables is None:
                     query.variables = {}
-                if '_stream_usage' in query.variables:
+                if STREAM_USAGE_QUERY_VARIABLE in query.variables:
-                    usage_info = query.variables['_stream_usage']
+                    usage_info = query.variables[STREAM_USAGE_QUERY_VARIABLE]
+                    _store_llm_usage(query, usage_info)
                     input_tokens = usage_info.get('prompt_tokens', 0)
                     output_tokens = usage_info.get('completion_tokens', 0)
-                    del query.variables['_stream_usage']
+                    del query.variables[STREAM_USAGE_QUERY_VARIABLE]
         except Exception as e:
             status = 'error'
             error_message = str(e)

src/langbot/pkg/provider/modelmgr/requesters/litellmchat.py

@@ -250,32 +250,81 @@ class LiteLLMRequester(requester.ProviderAPIRequester):
         - dict with the same keys
         - missing ``total_tokens`` (derived from prompt + completion)
         - ``None`` / partially-populated usage (defaults to 0)
+        - provider-specific token details, including cache token counters
         """
-        if usage is None:
+        def _plain_value(value: typing.Any) -> typing.Any:
-            return {'prompt_tokens': 0, 'completion_tokens': 0, 'total_tokens': 0}
+            if value is None:
+                return None
+            if isinstance(value, dict):
+                return {k: _plain_value(v) for k, v in value.items() if v is not None}
+            if isinstance(value, (list, tuple)):
+                return [_plain_value(v) for v in value]
 
-        def _get(key: str) -> typing.Any:
+            model_dump = getattr(value, 'model_dump', None)
-            if isinstance(usage, dict):
+            if callable(model_dump):
-                return usage.get(key)
+                try:
-            return getattr(usage, key, None)
+                    dumped = model_dump()
+                    if isinstance(dumped, dict):
+                        return _plain_value(dumped)
+                except Exception:
+                    pass
 
-        prompt_tokens = _get('prompt_tokens') or 0
+            return value
-        completion_tokens = _get('completion_tokens') or 0
+
-        total_tokens = _get('total_tokens') or 0
+        def _usage_dict(value: typing.Any) -> dict[str, typing.Any]:
+            if value is None:
+                return {}
+            plain = _plain_value(value)
+            if isinstance(plain, dict):
+                return plain
+
+            def _is_mock_attr(attr: typing.Any) -> bool:
+                return type(attr).__module__.startswith('unittest.mock')
+
+            data: dict[str, typing.Any] = {}
+            for key in (
+                'prompt_tokens',
+                'completion_tokens',
+                'total_tokens',
+                'prompt_tokens_details',
+                'completion_tokens_details',
+                'cache_creation_input_tokens',
+                'cache_read_input_tokens',
+                'input_token_details',
+                'output_token_details',
+            ):
+                attr_value = getattr(value, key, None)
+                if attr_value is not None and not _is_mock_attr(attr_value):
+                    data[key] = _plain_value(attr_value)
+            return data
+
+        def _to_int(value: typing.Any) -> int:
+            try:
+                return int(value or 0)
+            except (TypeError, ValueError):
+                return 0
+
+        normalized = _usage_dict(usage)
+
+        prompt_tokens = _to_int(normalized.get('prompt_tokens'))
+        completion_tokens = _to_int(normalized.get('completion_tokens'))
+        total_tokens = _to_int(normalized.get('total_tokens'))
 
         # Some providers omit total_tokens in streaming usage; derive it.
         if not total_tokens:
             total_tokens = prompt_tokens + completion_tokens
 
-        return {
+        normalized['prompt_tokens'] = prompt_tokens
-            'prompt_tokens': int(prompt_tokens),
+        normalized['completion_tokens'] = completion_tokens
-            'completion_tokens': int(completion_tokens),
+        normalized['total_tokens'] = total_tokens
-            'total_tokens': int(total_tokens),
+        return normalized
-        }
 
-    def _extract_usage(self, response) -> dict:
+    def _extract_usage(self, response) -> dict | None:
         """Extract usage info from a non-streaming LiteLLM response."""
-        return self._normalize_usage(getattr(response, 'usage', None))
+        usage = getattr(response, 'usage', None)
+        if usage is None:
+            return None
+        return self._normalize_usage(usage)
 
     @staticmethod
     def _as_dict(value: typing.Any) -> dict:
@@ -474,7 +523,7 @@ class LiteLLMRequester(requester.ProviderAPIRequester):
                     if query is not None:
                         if query.variables is None:
                             query.variables = {}
-                        query.variables['_stream_usage'] = usage_info
+                        query.variables[requester.STREAM_USAGE_QUERY_VARIABLE] = usage_info
 
                 if not hasattr(chunk, 'choices') or not chunk.choices:
                     continue

tests/unit_tests/agent/test_context_validation.py

@@ -2,6 +2,7 @@
 from __future__ import annotations
 
 import pytest
+from types import SimpleNamespace
 from unittest.mock import MagicMock, AsyncMock, patch
 
 # SDK imports for validation
@@ -174,6 +175,87 @@ class TestContextValidation:
         # Verify input
         assert validated.input.text == "Hello world"
 
+    @pytest.mark.asyncio
+    async def test_build_context_from_event_populates_model_context_window(self):
+        """Runtime metadata should expose the selected LLM model context window."""
+        mock_app = self._make_mock_app()
+        mock_app.model_mgr = MagicMock()
+        mock_app.model_mgr.get_model_by_uuid = AsyncMock(
+            return_value=SimpleNamespace(
+                model_entity=SimpleNamespace(context_length=128000),
+            )
+        )
+        builder = AgentRunContextBuilder(mock_app)
+
+        event = self._make_event_envelope()
+        binding = self._make_binding()
+        resources = self._make_resources()
+        resources['models'] = [
+            {
+                'model_id': 'rerank-model',
+                'model_type': 'rerank',
+                'provider': 'test-provider',
+                'operations': ['rerank'],
+            },
+            {
+                'model_id': 'llm-model',
+                'model_type': 'llm',
+                'provider': 'test-provider',
+                'operations': ['invoke', 'stream'],
+            },
+        ]
+        descriptor = self._make_descriptor()
+
+        with patch('langbot.pkg.agent.runner.context_builder.get_persistent_state_store') as mock_get_store:
+            mock_store = AsyncMock()
+            mock_store.build_snapshot_from_event = AsyncMock(return_value={
+                'conversation': {},
+                'actor': {},
+                'subject': {},
+                'runner': {},
+            })
+            mock_get_store.return_value = mock_store
+
+            context_dict = await builder.build_context_from_event(
+                event=event,
+                binding=binding,
+                descriptor=descriptor,
+                resources=resources,
+            )
+
+        assert context_dict['runtime']['metadata']['model_context_window_tokens'] == 128000
+        mock_app.model_mgr.get_model_by_uuid.assert_awaited_once_with('llm-model')
+
+    @pytest.mark.asyncio
+    async def test_model_context_window_uses_primary_llm_only(self):
+        """Fallback model windows should not replace missing primary model metadata."""
+        mock_app = self._make_mock_app()
+        mock_app.model_mgr = MagicMock()
+        mock_app.model_mgr.get_model_by_uuid = AsyncMock(
+            return_value=SimpleNamespace(
+                model_entity=SimpleNamespace(context_length=None),
+            )
+        )
+        builder = AgentRunContextBuilder(mock_app)
+        resources = self._make_resources()
+        resources['models'] = [
+            {
+                'model_id': 'primary-model',
+                'model_type': 'llm',
+                'provider': 'test-provider',
+                'operations': ['invoke', 'stream'],
+            },
+            {
+                'model_id': 'fallback-model',
+                'model_type': 'llm',
+                'provider': 'test-provider',
+                'operations': ['invoke', 'stream'],
+            },
+        ]
+
+        assert await builder._build_model_context_window_tokens(resources) is None
+        mock_app.model_mgr.get_model_by_uuid.assert_awaited_once_with('primary-model')
+
     @pytest.mark.asyncio
     async def test_build_context_preserves_subject_data_for_non_message_events(self):
         """Non-message EBA events keep subject.data instead of relying on message text."""

tests/unit_tests/plugin/test_handler_actions.py

@@ -388,6 +388,7 @@ class TestAgentRunProxyActions:
     def query(remove_think=True):
         return SimpleNamespace(
             pipeline_config={'output': {'misc': {'remove-think': remove_think}}},
+            variables={},
             prompt=SimpleNamespace(
                 messages=[provider_message.Message(role='system', content='effective prompt')]
             ),
@@ -488,6 +489,60 @@ class TestAgentRunProxyActions:
         assert kwargs['remove_think'] is True
         assert [tool.name for tool in kwargs['funcs']] == ['search']
 
+    @pytest.mark.asyncio
+    async def test_invoke_llm_returns_provider_usage(self, app):
+        """INVOKE_LLM includes optional provider usage in the action response."""
+        from langbot.pkg.agent.runner.session_registry import get_session_registry
+        from langbot.pkg.provider.modelmgr import requester as model_requester
+
+        usage = {
+            'prompt_tokens': 11,
+            'completion_tokens': 7,
+            'total_tokens': 18,
+            'prompt_tokens_details': {'cached_tokens': 3},
+        }
+
+        class UsageProvider:
+            async def invoke_llm(self, **kwargs):
+                kwargs['query'].variables[model_requester.LLM_USAGE_QUERY_VARIABLE] = usage
+                return provider_message.Message(role='assistant', content='ok')
+
+        run_id = 'run_proxy_invoke_llm_usage'
+        query = self.query()
+        app.query_pool.cached_queries[905] = query
+
+        registry = get_session_registry()
+        await registry.unregister(run_id)
+        await registry.register(
+            run_id=run_id,
+            runner_id='plugin:test/runner/default',
+            query_id=905,
+            plugin_identity='test/runner',
+            resources=make_agent_resources(models=[{'model_id': 'llm_usage_001'}]),
+        )
+
+        model = SimpleNamespace(
+            model_entity=SimpleNamespace(abilities=[], extra_args={}),
+            provider=UsageProvider(),
+        )
+        app.model_mgr.get_model_by_uuid.return_value = model
+        runtime_handler = make_handler(app)
+
+        try:
+            response = await runtime_handler.actions[PluginToRuntimeAction.INVOKE_LLM.value]({
+                'run_id': run_id,
+                'caller_plugin_identity': 'test/runner',
+                'llm_model_uuid': 'llm_usage_001',
+                'messages': [{'role': 'user', 'content': 'hello'}],
+            })
+        finally:
+            await registry.unregister(run_id)
+
+        assert response.code == 0
+        assert response.data['message']['content'] == 'ok'
+        assert response.data['usage'] == usage
+        assert model_requester.LLM_USAGE_QUERY_VARIABLE not in query.variables
+
     @pytest.mark.asyncio
     async def test_invoke_llm_stream_restores_query_and_options(self, app):
         """INVOKE_LLM_STREAM applies the same host context as non-streaming calls."""
@@ -598,6 +653,63 @@ class TestAgentRunProxyActions:
         assert [response.code for response in responses] == [0, 0]
         assert [response.data['chunk']['content'] for response in responses] == ['ok', ' done']
 
+    @pytest.mark.asyncio
+    async def test_invoke_llm_stream_returns_provider_usage_event(self, app):
+        """INVOKE_LLM_STREAM emits a final usage-only action response when available."""
+        from langbot.pkg.agent.runner.session_registry import get_session_registry
+        from langbot.pkg.provider.modelmgr import requester as model_requester
+
+        usage = {
+            'prompt_tokens': 9,
+            'completion_tokens': 4,
+            'total_tokens': 13,
+            'prompt_tokens_details': {'cached_tokens': 2},
+        }
+
+        class StreamProvider:
+            async def invoke_llm_stream(self, **kwargs):
+                yield provider_message.MessageChunk(role='assistant', content='ok')
+                kwargs['query'].variables[model_requester.LLM_USAGE_QUERY_VARIABLE] = usage
+
+        run_id = 'run_proxy_invoke_llm_stream_usage'
+        query = self.query()
+        app.query_pool.cached_queries[906] = query
+
+        registry = get_session_registry()
+        await registry.unregister(run_id)
+        await registry.register(
+            run_id=run_id,
+            runner_id='plugin:test/runner/default',
+            query_id=906,
+            plugin_identity='test/runner',
+            resources=make_agent_resources(models=[{'model_id': 'llm_stream_usage_001'}]),
+        )
+
+        model = SimpleNamespace(
+            model_entity=SimpleNamespace(abilities=[], extra_args={}),
+            provider=StreamProvider(),
+        )
+        app.model_mgr.get_model_by_uuid.return_value = model
+        runtime_handler = make_handler(app)
+
+        responses = []
+        try:
+            stream = runtime_handler.actions[PluginToRuntimeAction.INVOKE_LLM_STREAM.value]({
+                'run_id': run_id,
+                'caller_plugin_identity': 'test/runner',
+                'llm_model_uuid': 'llm_stream_usage_001',
+                'messages': [{'role': 'user', 'content': 'hello'}],
+            })
+            async for response in stream:
+                responses.append(response)
+        finally:
+            await registry.unregister(run_id)
+
+        assert [response.code for response in responses] == [0, 0]
+        assert responses[0].data['chunk']['content'] == 'ok'
+        assert responses[1].data == {'usage': usage}
+        assert model_requester.LLM_USAGE_QUERY_VARIABLE not in query.variables
+
     @pytest.mark.asyncio
     async def test_call_tool_passes_current_query(self, app):
         """CALL_TOOL passes the current Query back into tool execution."""

tests/unit_tests/provider/test_litellmchat.py

@@ -115,6 +115,15 @@ class TestExtractUsage:
         assert result['prompt_tokens'] == 0
         assert result['completion_tokens'] == 0
 
+    def test_extract_usage_without_provider_usage(self):
+        """Missing provider usage is not treated as authoritative zero usage."""
+        requester = litellmchat.LiteLLMRequester(ap=Mock(), config={})
+
+        response = Mock()
+        response.usage = None
+
+        assert requester._extract_usage(response) is None
+
 
 class TestNormalizeUsage:
     """Test _normalize_usage helper covering real-world usage shapes"""
@@ -131,6 +140,22 @@ class TestNormalizeUsage:
         )
         assert result == {'prompt_tokens': 12, 'completion_tokens': 8, 'total_tokens': 20}
 
+    def test_preserves_token_details(self):
+        """Provider token details such as cache counters are preserved."""
+        result = litellmchat.LiteLLMRequester._normalize_usage(
+            {
+                'prompt_tokens': 12,
+                'completion_tokens': 8,
+                'total_tokens': 20,
+                'prompt_tokens_details': {'cached_tokens': 7},
+                'completion_tokens_details': {'reasoning_tokens': 3},
+            }
+        )
+
+        assert result['prompt_tokens'] == 12
+        assert result['prompt_tokens_details'] == {'cached_tokens': 7}
+        assert result['completion_tokens_details'] == {'reasoning_tokens': 3}
+
     def test_missing_total_is_derived(self):
         """When total_tokens is absent/zero it is derived from prompt + completion"""
         usage = Mock()

tests/unit_tests/provider/test_requester_base.py

@@ -299,6 +299,59 @@ async def test_runtime_provider_invoke_llm_delegates(runtime_provider, runtime_l
     assert result.role == 'assistant'
 
 
+@pytest.mark.asyncio
+async def test_runtime_provider_invoke_llm_stashes_usage(runtime_provider, runtime_llm_model):
+    """RuntimeProvider preserves requester usage for upstream action handlers."""
+    provider = runtime_provider
+
+    import langbot_plugin.api.entities.builtin.provider.message as provider_message
+    import langbot_plugin.api.entities.builtin.pipeline.query as pipeline_query
+
+    query = pipeline_query.Query.model_construct(
+        query_id='test-query-usage',
+        launcher_type='person',
+        launcher_id=12345,
+        sender_id=12345,
+        message_chain=None,
+        message_event=None,
+        adapter=None,
+        pipeline_uuid='pipeline-uuid',
+        bot_uuid='bot-uuid',
+        pipeline_config={'ai': {}, 'output': {}, 'trigger': {}},
+        session=None,
+        prompt=None,
+        messages=[],
+        user_message=None,
+        use_funcs=[],
+        use_llm_model_uuid=None,
+        variables={},
+        resp_messages=[],
+        resp_message_chain=None,
+        current_stage_name=None,
+    )
+    usage = {
+        'prompt_tokens': 11,
+        'completion_tokens': 7,
+        'total_tokens': 18,
+        'prompt_tokens_details': {'cached_tokens': 3},
+    }
+    provider.requester.invoke_llm = AsyncMock(
+        return_value=(
+            provider_message.Message(role='assistant', content='ok'),
+            usage,
+        )
+    )
+
+    result = await provider.invoke_llm(
+        query,
+        runtime_llm_model,
+        [provider_message.Message(role='user', content='Hello')],
+    )
+
+    assert result.content == 'ok'
+    assert query.variables[requester.LLM_USAGE_QUERY_VARIABLE] == usage
+
+
 @pytest.mark.asyncio
 async def test_runtime_provider_invoke_llm_stream_yields_chunks(runtime_provider, runtime_llm_model):
     """Test RuntimeProvider.invoke_llm_stream yields chunks from requester."""
@@ -340,6 +393,62 @@ async def test_runtime_provider_invoke_llm_stream_yields_chunks(runtime_provider
     assert chunks[0].role == 'assistant'
 
 
+@pytest.mark.asyncio
+async def test_runtime_provider_invoke_llm_stream_stashes_usage(runtime_provider, runtime_llm_model):
+    """RuntimeProvider transfers captured stream usage to the public query usage key."""
+    provider = runtime_provider
+
+    import langbot_plugin.api.entities.builtin.provider.message as provider_message
+    import langbot_plugin.api.entities.builtin.pipeline.query as pipeline_query
+
+    query = pipeline_query.Query.model_construct(
+        query_id='test-stream-usage',
+        launcher_type='person',
+        launcher_id=12345,
+        sender_id=12345,
+        message_chain=None,
+        message_event=None,
+        adapter=None,
+        pipeline_uuid='pipeline-uuid',
+        bot_uuid='bot-uuid',
+        pipeline_config={'ai': {}, 'output': {}, 'trigger': {}},
+        session=None,
+        prompt=None,
+        messages=[],
+        user_message=None,
+        use_funcs=[],
+        use_llm_model_uuid=None,
+        variables={},
+        resp_messages=[],
+        resp_message_chain=None,
+        current_stage_name=None,
+    )
+    usage = {
+        'prompt_tokens': 13,
+        'completion_tokens': 2,
+        'total_tokens': 15,
+    }
+
+    async def fake_stream(**kwargs):
+        kwargs['query'].variables[requester.STREAM_USAGE_QUERY_VARIABLE] = usage
+        yield provider_message.MessageChunk(role='assistant', content='ok')
+
+    provider.requester.invoke_llm_stream = fake_stream
+
+    chunks = [
+        chunk
+        async for chunk in provider.invoke_llm_stream(
+            query,
+            runtime_llm_model,
+            [provider_message.Message(role='user', content='Hello')],
+        )
+    ]
+
+    assert len(chunks) == 1
+    assert query.variables[requester.LLM_USAGE_QUERY_VARIABLE] == usage
+    assert requester.STREAM_USAGE_QUERY_VARIABLE not in query.variables
+
+
 @pytest.mark.asyncio
 async def test_runtime_provider_invoke_embedding_returns_vectors(runtime_provider, runtime_embedding_model):
     """Test RuntimeProvider.invoke_embedding returns embedding vectors."""