fix(cli): status probe accepts reasoning_content for thinking models (#62)

src/cli/src/commands/status.ts

@@ -227,13 +227,21 @@ function defaultFetchServerLlms(mcpdUrl: string, token: string | null): Promise<
 /**
  * POST a tiny "say hi" prompt to /api/v1/llms/<name>/infer and decide if
  * the LLM actually serves inference. Returns ok=true when the response is
- * 200 with a non-empty assistant message; otherwise ok=false with an
- * error string suitable for one-line display.
+ * 200 with non-empty content OR reasoning_content (thinking models often
+ * spend their token budget on the reasoning trace and never emit a
+ * `content` block, but they're clearly alive if reasoning came back).
+ * Otherwise ok=false with an error string suitable for one-line display.
+ *
+ * `max_tokens: 64` gives reasoning models enough headroom to emit
+ * something visible while still capping latency at ~1-2 sec on cheap
+ * models. The exact wording — "Reply with just: hi" — is more terse and
+ * closer to what a thinking model can short-circuit on without burning
+ * its entire budget on reasoning.
  */
 const PROBE_TIMEOUT_MS = 15_000;
 const PROBE_BODY = JSON.stringify({
-  messages: [{ role: 'user', content: "Say exactly the word 'hi' and nothing else." }],
-  max_tokens: 8,
+  messages: [{ role: 'user', content: 'Reply with just: hi' }],
+  max_tokens: 64,
   temperature: 0,
 });
 
@@ -276,20 +284,31 @@ function defaultProbeServerLlm(mcpdUrl: string, name: string, token: string | nu
             return;
           }
           let content = '';
+          let reasoning = '';
           try {
             const parsed = JSON.parse(body) as {
-              choices?: Array<{ message?: { content?: string } }>;
+              choices?: Array<{ message?: { content?: string; reasoning_content?: string } }>;
             };
-            content = parsed.choices?.[0]?.message?.content?.trim() ?? '';
+            const msg = parsed.choices?.[0]?.message;
+            content = msg?.content?.trim() ?? '';
+            reasoning = msg?.reasoning_content?.trim() ?? '';
           } catch {
             resolve({ ok: false, ms, error: 'invalid response body' });
             return;
           }
-          if (content === '') {
-            resolve({ ok: false, ms, error: 'empty content' });
+          if (content !== '') {
+            resolve({ ok: true, ms, say: content.slice(0, 16) });
             return;
           }
-          resolve({ ok: true, ms, say: content.slice(0, 16) });
+          if (reasoning !== '') {
+            // Thinking model burned its budget on the reasoning trace
+            // before emitting `content`. The LLM is alive — flag it as
+            // ok and surface a short reasoning preview so the user can
+            // tell at a glance.
+            resolve({ ok: true, ms, say: `[thinking] ${reasoning.slice(0, 12)}` });
+            return;
+          }
+          resolve({ ok: false, ms, error: 'empty content' });
         });
       });
     } catch {