@aria-framework/ai 0.8.0 → 0.8.1

package/health.js

@@ -200,8 +200,12 @@ function createHealthChecker(opts = {}) {
     // generation speed — separating those would need time-to-first-token, which no adapter
     // reports — but it is what the caller actually experienced, which is the number worth planning
     // against. Only completions count; a health probe generates no tokens and would drag it to 0.
-    if (ok && info.tokens > 0 && info.ms > 0) {
-      e.samples.push({ tokens: info.tokens, ms: info.ms });
+    // Guarded on TOKENS, not on duration. Requiring ms > 0 here silently dropped any call that
+    // completed inside a millisecond — which a fast local model or a stub genuinely does, so the
+    // sample count became timing-dependent. The division is where zero actually matters, so that
+    // is where it is handled.
+    if (ok && info.tokens > 0) {
+      e.samples.push({ tokens: info.tokens, ms: Math.max(0, Number(info.ms) || 0) });
       if (e.samples.length > WINDOW) e.samples.shift();
     }
     if (ok) {
@@ -303,11 +307,14 @@ function createHealthChecker(opts = {}) {
         modelPresent: e.modelPresent,
         // NULL until something has actually generated tokens. Reporting 0 tok/s for a provider
         // nobody has used yet reads as "it is slow" rather than "we do not know".
-        tokensPerSec: e.samples.length
-          ? Math.round(
-            e.samples.reduce((n, x) => n + x.tokens, 0)
-            / (e.samples.reduce((n, x) => n + x.ms, 0) / 1000))
-          : null,
+        // Null rather than Infinity when every sample was too fast to measure: "we cannot tell"
+        // is the honest answer, and a card showing Infinity tok/s is worse than showing nothing.
+        tokensPerSec: (() => {
+          if (!e.samples.length) return null;
+          const ms = e.samples.reduce((n, x) => n + x.ms, 0);
+          if (ms <= 0) return null;
+          return Math.round(e.samples.reduce((n, x) => n + x.tokens, 0) / (ms / 1000));
+        })(),
         samples: e.samples.length,
         cooldownRemainingMs: down ? e.downUntil - now() : 0
       };

package/package.json

@@ -1,7 +1,7 @@
 {
   "name": "@aria-framework/ai",
   "description": "Aria App Framework — AI module. A dependency-injected model seam (createAiClient) over several providers (LM Studio / OpenAI-compatible / Anthropic), with a fact-preservation guard, generic Polish and Generate writing engines, and a browser polish widget. Prompts and config stay in the consuming app.",
-  "version": "0.8.0",
+  "version": "0.8.1",
   "license": "UNLICENSED",
   "private": false,
   "publishConfig": {