fix(llama-cpp): report cleanup recovery to initiating request

Punchcard-Session: frost-brook-timber-mx

extensions/llama-cpp/src/inference-provider.test.ts

@@ -768,11 +768,14 @@ describe("llama.cpp inference provider", () => {
       rejectCleanup = reject;
     });
     mocks.contextDispose.mockImplementationOnce(async () => await cleanup);
-    await createTestStream({ prompt: "three" });
+    const failedSwitch = await createTestStream({ prompt: "three" });
     await vi.waitFor(() => expect(mocks.contextDispose).toHaveBeenCalledTimes(2));
     const unavailable = await createTestStream({ selectedModel: otherModel, prompt: "four" });
     const disposing = inferenceRuntime.dispose();
     rejectCleanup(new Error("context cleanup failed"));
+    await expect(failedSwitch.result()).resolves.toMatchObject({
+      errorMessage: expect.stringContaining("openclaw gateway restart"),
+    });
     await expect(unavailable.result()).resolves.toMatchObject({
       errorMessage: expect.stringContaining("openclaw gateway restart"),
     });
@@ -784,7 +787,10 @@ describe("llama.cpp inference provider", () => {
   it("records cleanup failure during partial model initialization", async () => {
     mocks.model.createContext.mockRejectedValueOnce(new Error("context creation failed"));
     mocks.modelDispose.mockRejectedValueOnce(new Error("model cleanup failed"));
-    await collectTestEvents();
+    const failedInitialization = await createTestStream();
+    await expect(failedInitialization.result()).resolves.toMatchObject({
+      errorMessage: expect.stringContaining("openclaw gateway restart"),
+    });
     await expect(inferenceRuntime.dispose()).rejects.toThrow("model cleanup failed");
     expectDisposeCalls(0, 1, 0);
   });

extensions/llama-cpp/src/inference-provider.ts

@@ -86,6 +86,12 @@ function buildMessage(params: {
   };
 }
 
+function runtimeUnavailableErrorMessage(state: LlamaCppInferenceRuntimeState): string {
+  return state.cleanupFailure
+    ? "llama.cpp runtime stopped after cleanup failed. Run `openclaw gateway restart` to recover."
+    : "llama.cpp runtime is stopping";
+}
+
 function runtimeUnavailableMessage(
   state: LlamaCppInferenceRuntimeState,
   model: Parameters<StreamFn>[0],
@@ -94,9 +100,7 @@ function runtimeUnavailableMessage(
     model,
     content: [],
     stopReason: "error",
-    errorMessage: state.cleanupFailure
-      ? "llama.cpp runtime stopped after cleanup failed. Run `openclaw gateway restart` to recover."
-      : "llama.cpp runtime is stopping",
+    errorMessage: runtimeUnavailableErrorMessage(state),
   });
 }
 
@@ -682,7 +686,11 @@ function createLlamaCppStreamFnForRuntime(
       } catch (error) {
         const aborted = generationAborted || options?.signal?.aborted === true;
         const reason = aborted ? "aborted" : "error";
-        const errorMessage = aborted ? "Request was aborted" : formatLlamaCppSetupError(error);
+        const errorMessage = aborted
+          ? "Request was aborted"
+          : state.cleanupFailure
+            ? runtimeUnavailableErrorMessage(state)
+            : formatLlamaCppSetupError(error);
         stream.push({
           type: "error",
           reason,