NVIDIA
diff --git a/‎docs/inference/set-up-llama-cpp.mdx‎
Lines changed: 8 additions & 0 deletions b/‎docs/inference/set-up-llama-cpp.mdx‎
Lines changed: 8 additions & 0 deletions
diff --git a/‎src/lib/inference/llama-cpp/index.test.ts‎
Lines changed: 80 additions & 0 deletions b/‎src/lib/inference/llama-cpp/index.test.ts‎
Lines changed: 80 additions & 0 deletions
diff --git a/‎src/lib/inference/llama-cpp/index.ts‎
Lines changed: 16 additions & 7 deletions b/‎src/lib/inference/llama-cpp/index.ts‎
Lines changed: 16 additions & 7 deletions
@@ -37,9 +37,17 @@ The server must satisfy all of these requirements:
 - Return native llama.cpp model metadata without conflicting model entries.
 
 Configure the server with its own supported mechanism before you run NemoClaw.
+Start the server with an explicit served model alias.
+A server that starts without an alias reports its model file path as the model ID.
+NemoClaw rejects a path as a served model alias.
 Enable metrics when available, and use an API-key file instead of placing the key in process arguments when your llama.cpp build supports those options.
 If the server exposes multiple models, identify the exact served alias through `NEMOCLAW_MODEL`.
 
+NemoClaw takes the served model alias from the `/v1/models` entry that it selects.
+It also compares that alias with `model_alias` in `/props` when `/props` returns that field.
+NemoClaw refuses the attachment when the two values differ.
+If `/props` omits `model_alias`, NemoClaw attaches the server using the remaining native evidence.
+
 For interactive onboarding, run:
 
 ```bash
 
@@ -348,6 +348,86 @@ describe("probeLlamaCppAttachment", () => {
     ).toMatchObject({ ok: false, reason: "conflicting-fingerprint" });
   });
 
+  it("attaches a native server whose properties omit the served model alias (#9603)", () => {
+    const responses = nativeResponses();
+    responses[3] = response(
+      200,
+      JSON.stringify({
+        model_path: "/models/model.gguf",
+        total_slots: 2,
+        default_generation_settings: { params: {} },
+      }),
+    );
+
+    expect(
+      probeLlamaCppAttachment("secret-token", { runCurlProbeImpl: scriptedProbe(responses) }),
+    ).toEqual({ ok: true, model: "team/model-alias" });
+  });
+
+  it("rejects properties that report a null served model alias (#9603)", () => {
+    const responses = nativeResponses();
+    responses[3] = response(
+      200,
+      JSON.stringify({
+        model_alias: null,
+        model_path: "/models/model.gguf",
+        total_slots: 2,
+        default_generation_settings: { params: {} },
+      }),
+    );
+
+    expect(
+      probeLlamaCppAttachment("secret-token", { runCurlProbeImpl: scriptedProbe(responses) }),
+    ).toMatchObject({ ok: false, reason: "conflicting-fingerprint" });
+  });
+
+  it("names the health endpoint when the server reports a loading model (#9603)", () => {
+    const responses = nativeResponses();
+    responses[2] = response(200, '{"status":"loading model"}');
+
+    expect(
+      probeLlamaCppAttachment("secret-token", { runCurlProbeImpl: scriptedProbe(responses) }),
+    ).toMatchObject({
+      ok: false,
+      reason: "conflicting-fingerprint",
+      message: expect.stringContaining("health endpoint"),
+    });
+  });
+
+  it("names the properties endpoint when model_alias differs from the served model alias (#9603)", () => {
+    const responses = nativeResponses();
+    responses[3] = response(
+      200,
+      JSON.stringify({
+        model_alias: "different/model",
+        model_path: "/models/model.gguf",
+        total_slots: 2,
+        default_generation_settings: { params: {} },
+      }),
+    );
+
+    expect(
+      probeLlamaCppAttachment("secret-token", { runCurlProbeImpl: scriptedProbe(responses) }),
+    ).toMatchObject({
+      ok: false,
+      reason: "conflicting-fingerprint",
+      message: expect.stringContaining("properties endpoint"),
+    });
+  });
+
+  it("names the metrics endpoint when the response has no llama.cpp metrics (#9603)", () => {
+    const responses = nativeResponses();
+    responses[4] = response(200, "# TYPE go_gc_duration_seconds summary\ngo_goroutines 12\n");
+
+    expect(
+      probeLlamaCppAttachment("secret-token", { runCurlProbeImpl: scriptedProbe(responses) }),
+    ).toMatchObject({
+      ok: false,
+      reason: "conflicting-fingerprint",
+      message: expect.stringContaining("metrics endpoint"),
+    });
+  });
+
   it.each([
     "/models/model.gguf",
     "C:\\models\\model.gguf",
 
@@ -122,7 +122,8 @@ function hasHealthyNativeResponse(result: CurlProbeResult): boolean {
 function hasMatchingNativeProps(result: CurlProbeResult, model: string): boolean {
   if (!result.ok) return false;
   const body = parseJsonObject(result.body);
-  if (!body || body.model_alias !== model || typeof body.model_path !== "string") return false;
+  if (!body || typeof body.model_path !== "string") return false;
+  if (body.model_alias !== undefined && body.model_alias !== model) return false;
   if (typeof body.total_slots !== "number" || body.total_slots <= 0) return false;
   const defaults = body.default_generation_settings;
   return (
@@ -348,14 +349,22 @@ export function probeLlamaCppAttachment(
       const boundFailure = boundedProbeFailure(result);
       if (boundFailure) return boundFailure;
     }
-    if (
-      !hasHealthyNativeResponse(health) ||
-      !hasMatchingNativeProps(props, model) ||
-      !hasNativeMetricsResponse(metrics)
-    ) {
+    if (!hasHealthyNativeResponse(health)) {
       return failure(
         "conflicting-fingerprint",
-        "The server returned conflicting or incomplete native llama.cpp evidence.",
+        "The llama.cpp health endpoint did not report the native status ok.",
+      );
+    }
+    if (!hasMatchingNativeProps(props, model)) {
+      return failure(
+        "conflicting-fingerprint",
+        "The llama.cpp properties endpoint did not return complete native evidence for the selected served model.",
+      );
+    }
+    if (!hasNativeMetricsResponse(metrics)) {
+      return failure(
+        "conflicting-fingerprint",
+        "The llama.cpp metrics endpoint returned neither llama.cpp metrics nor the native metrics-not-supported response.",
       );
     }
     return { ok: true, model };