fix(runner): keep installation context within each stream resume

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

@@ -48,12 +48,15 @@ class RunnerInvoker:
                 context=context,
             )
 
-            while True:
-                try:
-                    result_dict = await self._next_with_deadline(gen, descriptor, context)
-                except StopAsyncIteration:
-                    break
-                yield result_dict
+            try:
+                while True:
+                    try:
+                        result_dict = await self._next_with_deadline(gen, descriptor, context)
+                    except StopAsyncIteration:
+                        break
+                    yield result_dict
+            finally:
+                await self._close_generator(gen, descriptor)
 
         except asyncio.TimeoutError as e:
             raise RunnerExecutionError(
@@ -128,4 +131,4 @@ class RunnerInvoker:
         try:
             await gen.aclose()
         except Exception as e:
-            self.ap.logger.warning(f'Failed to close timed-out runner {descriptor.id}: {e}')
+            self.ap.logger.warning(f'Failed to close runner {descriptor.id}: {e}')

src/langbot/pkg/plugin/connector.py

@@ -2237,12 +2237,12 @@ class PluginRuntimeConnector(ManagedRuntimeConnector):
             include_plugins=bound_plugins,
         )
         runtime_handler = self._runtime_handler()
-        with runtime_handler.installation_scope(binding):
-            gen = runtime_handler.execute_command(
-                command_ctx.model_dump(serialize_as_any=True),
-                include_plugins=bound_plugins,
-            )
-            async for ret in gen:
+        gen = runtime_handler.execute_command(
+            command_ctx.model_dump(serialize_as_any=True),
+            include_plugins=bound_plugins,
+        )
+        async with contextlib.aclosing(self._installation_scoped_stream(runtime_handler, binding, gen)) as scoped:
+            async for ret in scoped:
                 yield command_context.CommandReturn.model_validate(ret)
 
     # Runner methods
@@ -2308,15 +2308,31 @@ class PluginRuntimeConnector(ManagedRuntimeConnector):
             require_enabled=True,
         )
         runtime_handler = self._runtime_handler()
-        with runtime_handler.installation_scope(binding):
-            async for ret in runtime_handler.run_runner(
-                plugin_author,
-                plugin_name,
-                runner_name,
-                context,
-            ):
+        gen = runtime_handler.run_runner(plugin_author, plugin_name, runner_name, context)
+        async with contextlib.aclosing(self._installation_scoped_stream(runtime_handler, binding, gen)) as scoped:
+            async for ret in scoped:
                 yield ret
 
+    @staticmethod
+    async def _installation_scoped_stream(runtime_handler, binding, gen):
+        """Keep ContextVar tokens inside a single resume, never across yields.
+
+        Consumers may use a different Task for each anext (e.g. wait_for).
+        Reset the installation before exposing a result to the consumer, and
+        re-enter the same immutable binding for transport cleanup.
+        """
+        try:
+            while True:
+                with runtime_handler.installation_scope(binding):
+                    try:
+                        result = await anext(gen)
+                    except StopAsyncIteration:
+                        return
+                yield result
+        finally:
+            with runtime_handler.installation_scope(binding):
+                await gen.aclose()
+
     async def retrieve_knowledge(
         self,
         plugin_author: str,

src/langbot/pkg/plugin/handler.py

@@ -2933,8 +2933,9 @@ class RuntimeConnectionHandler(handler.Handler):
             timeout=timeout,
         )
 
-        async for ret in gen:
-            yield ret
+        async with contextlib.aclosing(gen):
+            async for ret in gen:
+                yield ret
 
     def _get_runner_action_timeout(self, context: dict[str, Any]) -> float:
         """Use the run deadline as the transport idle timeout when available."""
@@ -3143,8 +3144,9 @@ class RuntimeConnectionHandler(handler.Handler):
             timeout=180,
         )
 
-        async for ret in gen:
-            yield ret
+        async with contextlib.aclosing(gen):
+            async for ret in gen:
+                yield ret
 
     async def retrieve_knowledge(
         self,