Skip to content

7.4 billion tokens per month. 34 free LLM providers. 635 free model endpoints. All behind one /v1 endpoint, plus any custom OpenAI-compatible endpoint. Smart routing, automatic failover, encrypted keys. Personal experimentation only.

TypeScript

fix: bound the empty-completion exemption, guarantee a failover hop, and surface client aborts (#751, #752) (#755)

Tashfeen44af06be7b60
diff --git a/client/src/i18n/locales/am.json b/client/src/i18n/locales/am.json
index 95dd564a..d922fce0 100644
--- a/client/src/i18n/locales/am.json
+++ b/client/src/i18n/locales/am.json
@@ -550,6 +550,7 @@
"requestedModelHint": "የተጠየቀ {model}፣ በመውደቅ አገልግሏል",
"attempts": "ሙከራዎች",
"filterAll": "ሁሉም",
+ "filterCanceled": "የተሰረዙ",
"providerBreakdown": "በእያንዳንዱ አቅራቢ መከፋፈል",
"tokensPerSec": "ቶክ / ዎች",
"allProviders": "ሁሉም አቅራቢዎች",
diff --git a/client/src/i18n/locales/ar.json b/client/src/i18n/locales/ar.json
index 6e02e86b..a612c2ca 100644
--- a/client/src/i18n/locales/ar.json
+++ b/client/src/i18n/locales/ar.json
@@ -550,6 +550,7 @@
"requestedModelHint": "تم طلب {model}، ويتم تقديمه بطريقة احتياطية",
"attempts": "محاولات",
"filterAll": "الكل",
+ "filterCanceled": "ملغاة",
"providerBreakdown": "تفصيل لكل مزود",
"tokensPerSec": "توك/س",
"allProviders": "جميع مقدمي الخدمة",
diff --git a/client/src/i18n/locales/az.json b/client/src/i18n/locales/az.json
index 9e05dfa3..c08183fe 100644
--- a/client/src/i18n/locales/az.json
+++ b/client/src/i18n/locales/az.json
@@ -550,6 +550,7 @@
"requestedModelHint": "{model} tələb olundu, ehtiyat qüvvə ilə təqdim edildi",
"attempts": "Cəhdlər",
"filterAll": "Hamı",
+ "filterCanceled": "Ləğv edilmiş",
"providerBreakdown": "Təminatçı üzrə bölgü",
"tokensPerSec": "Tok/s",
"allProviders": "Bütün təminatçılar",
diff --git a/client/src/i18n/locales/bg.json b/client/src/i18n/locales/bg.json
index 7cffee38..b630b884 100644
--- a/client/src/i18n/locales/bg.json
+++ b/client/src/i18n/locales/bg.json
@@ -550,6 +550,7 @@
"requestedModelHint": "Поискано {model}, обслужено като резерв",
"attempts": "Опити",
"filterAll": "Всички",
+ "filterCanceled": "Отменени",
"providerBreakdown": "Разпределение на всеки доставчик",
"tokensPerSec": "Ток/с",
"allProviders": "Всички доставчици",
diff --git a/client/src/i18n/locales/bn.json b/client/src/i18n/locales/bn.json
index 9be315fb..b3c1bc0a 100644
--- a/client/src/i18n/locales/bn.json
+++ b/client/src/i18n/locales/bn.json
@@ -550,6 +550,7 @@
"requestedModelHint": "অনুরোধ করা হয়েছে {model}, ফলব্যাক দ্বারা পরিবেশিত৷",
"attempts": "প্রচেষ্টা",
"filterAll": "সব",
+ "filterCanceled": "বাতিল",
"providerBreakdown": "প্রতি-প্রদানকারী ব্রেকডাউন",
"tokensPerSec": "টোক/সে",
"allProviders": "সমস্ত প্রদানকারী",
diff --git a/client/src/i18n/locales/cs.json b/client/src/i18n/locales/cs.json
index 20550ac0..94e47908 100644
--- a/client/src/i18n/locales/cs.json
+++ b/client/src/i18n/locales/cs.json
@@ -550,6 +550,7 @@
"requestedModelHint": "Požadovaný {model}, sloužen záložní variantou",
"attempts": "Pokusy",
"filterAll": "Všichni",
+ "filterCanceled": "Zrušené",
"providerBreakdown": "Rozdělení podle poskytovatele",
"tokensPerSec": "Tok/s",
"allProviders": "Všichni poskytovatelé",
diff --git a/client/src/i18n/locales/da.json b/client/src/i18n/locales/da.json
index 79bbe9e9..624e83ce 100644
--- a/client/src/i18n/locales/da.json
+++ b/client/src/i18n/locales/da.json
@@ -550,6 +550,7 @@
"requestedModelHint": "Anmodede {model}, betjent via fallback",
"attempts": "Forsøg",
"filterAll": "Alle",
+ "filterCanceled": "Annullerede",
"providerBreakdown": "Opdeling pr. udbyder",
"tokensPerSec": "Tok/s",
"allProviders": "Alle udbydere",
diff --git a/client/src/i18n/locales/de.json b/client/src/i18n/locales/de.json
index 66e84468..c9db7ad6 100644
--- a/client/src/i18n/locales/de.json
+++ b/client/src/i18n/locales/de.json
@@ -550,6 +550,7 @@
"requestedModelHint": "Angeforderter {model}, bereitgestellt durch Fallback",
"attempts": "Versuche",
"filterAll": "Alle",
+ "filterCanceled": "Abgebrochen",
"providerBreakdown": "Aufschlüsselung nach Anbieter",
"tokensPerSec": "Tok/s",
"allProviders": "Alle Anbieter",
diff --git a/client/src/i18n/locales/el.json b/client/src/i18n/locales/el.json
index 252b3fcd..0eecf1b3 100644
--- a/client/src/i18n/locales/el.json
+++ b/client/src/i18n/locales/el.json
@@ -550,6 +550,7 @@
"requestedModelHint": "Ζητήθηκε {model}, εξυπηρετείται από εναλλακτική",
"attempts": "Προσπάθειες",
"filterAll": "Όλα",
+ "filterCanceled": "Ακυρωμένες",
"providerBreakdown": "Ανάλυση ανά πάροχο",
"tokensPerSec": "Τοκ/α",
"allProviders": "Όλοι οι πάροχοι",
diff --git a/client/src/i18n/locales/en.json b/client/src/i18n/locales/en.json
index 29d31c73..f58f9be0 100644
--- a/client/src/i18n/locales/en.json
+++ b/client/src/i18n/locales/en.json
@@ -567,6 +567,7 @@
"requestedModelHint": "Requested {model}, served by fallback",
"attempts": "Attempts",
"filterAll": "All",
+ "filterCanceled": "Canceled",
"providerBreakdown": "Per-provider breakdown",
"tokensPerSec": "Tok/s",
"allProviders": "All providers",
diff --git a/client/src/i18n/locales/es.json b/client/src/i18n/locales/es.json
index 92ec7ffe..117ac765 100644
--- a/client/src/i18n/locales/es.json
+++ b/client/src/i18n/locales/es.json
@@ -550,6 +550,7 @@
"requestedModelHint": "Se solicitó {model}, servido por respaldo",
"attempts": "Intentos",
"filterAll": "Todas",
+ "filterCanceled": "Canceladas",
"providerBreakdown": "Desglose por proveedor",
"tokensPerSec": "Tok/s",
"allProviders": "Todos los proveedores",
diff --git a/client/src/i18n/locales/fa.json b/client/src/i18n/locales/fa.json
index 45160162..532a4776 100644
--- a/client/src/i18n/locales/fa.json
+++ b/client/src/i18n/locales/fa.json
@@ -550,6 +550,7 @@
"requestedModelHint": "درخواست {model}، ارائه شده توسط بازگشت",
"attempts": "تلاش ها",
"filterAll": "همه",
+ "filterCanceled": "لغوشده",
"providerBreakdown": "تفکیک هر ارائه دهنده",
"tokensPerSec": "Tok/s",
"allProviders": "همه ارائه دهندگان",
diff --git a/client/src/i18n/locales/fi.json b/client/src/i18n/locales/fi.json
index f03cd26a..97fe1ae3 100644
--- a/client/src/i18n/locales/fi.json
+++ b/client/src/i18n/locales/fi.json
@@ -550,6 +550,7 @@
"requestedModelHint": "Pyydetty {model}, toimitettu varasuunnitelmana",
"attempts": "Yritykset",
"filterAll": "Kaikki",
+ "filterCanceled": "Peruutetut",
"providerBreakdown": "Palveluntarjoajakohtainen erittely",
"tokensPerSec": "Tok/s",
"allProviders": "Kaikki palveluntarjoajat",
diff --git a/client/src/i18n/locales/fr.json b/client/src/i18n/locales/fr.json
index d4d175bc..58494dd2 100644
--- a/client/src/i18n/locales/fr.json
+++ b/client/src/i18n/locales/fr.json
@@ -550,6 +550,7 @@
"requestedModelHint": "{model} demandé, servi par repli",
"attempts": "Tentatives",
"filterAll": "Toutes",
+ "filterCanceled": "Annulées",
"providerBreakdown": "Détail par fournisseur",
"tokensPerSec": "Tok/s",
"allProviders": "Tous les fournisseurs",
diff --git a/client/src/i18n/locales/gu.json b/client/src/i18n/locales/gu.json
index 3ea2c112..f386a555 100644
--- a/client/src/i18n/locales/gu.json
+++ b/client/src/i18n/locales/gu.json
@@ -550,6 +550,7 @@
"requestedModelHint": "{model} વિનંતી કરી, ફોલબેક દ્વારા સેવા આપવામાં આવી",
"attempts": "પ્રયાસો",
"filterAll": "બધા",
+ "filterCanceled": "રદ થયેલ",
"providerBreakdown": "પ્રતિ-પ્રદાતા બ્રેકડાઉન",
"tokensPerSec": "ટોક/સે",
"allProviders": "બધા પ્રદાતાઓ",
diff --git a/client/src/i18n/locales/ha.json b/client/src/i18n/locales/ha.json
index 68c81548..66ca20cd 100644
--- a/client/src/i18n/locales/ha.json
+++ b/client/src/i18n/locales/ha.json
@@ -550,6 +550,7 @@
"requestedModelHint": "Neman {model}, wanda aka yi amfani da shi ta koma baya",
"attempts": "Ƙoƙari",
"filterAll": "Duka",
+ "filterCanceled": "An soke",
"providerBreakdown": "Rushewar kowane mai bayarwa",
"tokensPerSec": "Tok/s",
"allProviders": "Duk masu samarwa",
diff --git a/client/src/i18n/locales/he.json b/client/src/i18n/locales/he.json
index a6e7feb1..bf0f133e 100644
--- a/client/src/i18n/locales/he.json
+++ b/client/src/i18n/locales/he.json
@@ -550,6 +550,7 @@
"requestedModelHint": "מבוקש {model}, משרת על ידי גיבוי",
"attempts": "ניסיונות",
"filterAll": "הכל",
+ "filterCanceled": "בוטלו",
"providerBreakdown": "פירוט לכל ספק",
"tokensPerSec": "Tok/s",
"allProviders": "כל הספקים",
diff --git a/client/src/i18n/locales/hi.json b/client/src/i18n/locales/hi.json
index 83ee03fa..5721b29a 100644
--- a/client/src/i18n/locales/hi.json
+++ b/client/src/i18n/locales/hi.json
@@ -550,6 +550,7 @@
"requestedModelHint": "अनुरोधित {model}, फ़ॉलबैक द्वारा सेवा प्रदान की गई",
"attempts": "प्रयास",
"filterAll": "सब",
+ "filterCanceled": "रद्द",
"providerBreakdown": "प्रति-प्रदाता टूटना",
"tokensPerSec": "टोक/एस",
"allProviders": "सभी प्रदाता",
diff --git a/client/src/i18n/locales/hr.json b/client/src/i18n/locales/hr.json
index 1613c420..6350164f 100644
--- a/client/src/i18n/locales/hr.json
+++ b/client/src/i18n/locales/hr.json
@@ -550,6 +550,7 @@
"requestedModelHint": "Zatraženo {model}, posluženo kao rezerva",
"attempts": "Pokušaji",
"filterAll": "Svi",
+ "filterCanceled": "Otkazani",
"providerBreakdown": "Raspodjela po pružatelju usluga",
"tokensPerSec": "Tok/s",
"allProviders": "Svi pružatelji",
diff --git a/client/src/i18n/locales/hu.json b/client/src/i18n/locales/hu.json
index 127acad3..f6fa510f 100644
--- a/client/src/i18n/locales/hu.json
+++ b/client/src/i18n/locales/hu.json
@@ -550,6 +550,7 @@
"requestedModelHint": "Kértem {model}, tartalék kézbe szállítva",
"attempts": "Próbálkozások",
"filterAll": "Mind",
+ "filterCanceled": "Megszakítva",
"providerBreakdown": "Szolgáltatónkénti bontás",
"tokensPerSec": "Tok/s",
"allProviders": "Minden szolgáltató",
diff --git a/client/src/i18n/locales/id.json b/client/src/i18n/locales/id.json
index 9c53c931..d1fa0c34 100644
--- a/client/src/i18n/locales/id.json
+++ b/client/src/i18n/locales/id.json
@@ -550,6 +550,7 @@
"requestedModelHint": "{model} yang diminta, dilayani oleh fallback",
"attempts": "Upaya",
"filterAll": "Semua",
+ "filterCanceled": "Dibatalkan",
"providerBreakdown": "Perincian per penyedia",
"tokensPerSec": "Tok/dtk",
"allProviders": "Semua penyedia",
diff --git a/client/src/i18n/locales/ig.json b/client/src/i18n/locales/ig.json
index e92a53e4..87909d34 100644
--- a/client/src/i18n/locales/ig.json
+++ b/client/src/i18n/locales/ig.json
@@ -550,6 +550,7 @@
"requestedModelHint": "Arịrịọ {model}, nke na-azaghachi azụ azụ",
"attempts": "Mgbalị",
"filterAll": "Ha niile",
+ "filterCanceled": "Ndị a kagburu",
"providerBreakdown": "Mmebi nke onye na-eweta ya",
"tokensPerSec": "Tok/s",
"allProviders": "Ndị na-eweta niile",
diff --git a/client/src/i18n/locales/it.json b/client/src/i18n/locales/it.json
index df2a1f96..f31b095c 100644
--- a/client/src/i18n/locales/it.json
+++ b/client/src/i18n/locales/it.json
@@ -550,6 +550,7 @@
"requestedModelHint": "Richiesto {model}, servito dal fallback",
"attempts": "Tentativi",
"filterAll": "Tutte",
+ "filterCanceled": "Annullate",
"providerBreakdown": "Dettaglio per provider",
"tokensPerSec": "Tok/s",
"allProviders": "Tutti i provider",
diff --git a/client/src/i18n/locales/ja.json b/client/src/i18n/locales/ja.json
index feb5133c..0d507785 100644
--- a/client/src/i18n/locales/ja.json
+++ b/client/src/i18n/locales/ja.json
@@ -550,6 +550,7 @@
"requestedModelHint": "リクエストされた {model}、フォールバックによって提供される",
"attempts": "試み",
"filterAll": "すべて",
+ "filterCanceled": "キャンセル",
"providerBreakdown": "プロバイダーごとの内訳",
"tokensPerSec": "トク/秒",
"allProviders": "すべてのプロバイダー",
diff --git a/client/src/i18n/locales/ka.json b/client/src/i18n/locales/ka.json
index 8e25c4f1..32eac33f 100644
--- a/client/src/i18n/locales/ka.json
+++ b/client/src/i18n/locales/ka.json
@@ -550,6 +550,7 @@
"requestedModelHint": "მოითხოვა {model}, მოწოდებული სარეზერვო საშუალებით",
"attempts": "მცდელობები",
"filterAll": "ყველა",
+ "filterCanceled": "გაუქმებული",
"providerBreakdown": "თითო პროვაიდერის დაშლა",
"tokensPerSec": "ტოკ/ს",
"allProviders": "ყველა პროვაიდერი",
diff --git a/client/src/i18n/locales/km.json b/client/src/i18n/locales/km.json
index b78df3a5..15707c98 100644
--- a/client/src/i18n/locales/km.json
+++ b/client/src/i18n/locales/km.json
@@ -550,6 +550,7 @@
"requestedModelHint": "បានស្នើសុំ {model} បម្រើដោយជម្រើសបម្រុង",
"attempts": "ការព្យាយាម",
"filterAll": "ទាំងអស់",
+ "filterCanceled": "បានបោះបង់",
"providerBreakdown": "ការបែងចែកតាមអ្នកផ្តល់សេវា",
"tokensPerSec": "តុក/វិនាទី",
"allProviders": "អ្នកផ្តល់សេវាទាំងអស់",
diff --git a/client/src/i18n/locales/kn.json b/client/src/i18n/locales/kn.json
index 5ade755e..66772513 100644
--- a/client/src/i18n/locales/kn.json
+++ b/client/src/i18n/locales/kn.json
@@ -550,6 +550,7 @@
"requestedModelHint": "ವಿನಂತಿಸಿದ {model}, ಫಾಲ್‌ಬ್ಯಾಕ್ ಮೂಲಕ ಸೇವೆ ಸಲ್ಲಿಸಲಾಗಿದೆ",
"attempts": "ಪ್ರಯತ್ನಗಳು",
"filterAll": "ಎಲ್ಲಾ",
+ "filterCanceled": "ರದ್ದಾದವು",
"providerBreakdown": "ಪ್ರತಿ-ಒದಗಿಸುವವರ ಸ್ಥಗಿತ",
"tokensPerSec": "ಟೋಕ್/ರು",
"allProviders": "ಎಲ್ಲಾ ಪೂರೈಕೆದಾರರು",
diff --git a/client/src/i18n/locales/ko.json b/client/src/i18n/locales/ko.json
index 06e88bc8..fd14d30c 100644
--- a/client/src/i18n/locales/ko.json
+++ b/client/src/i18n/locales/ko.json
@@ -550,6 +550,7 @@
"requestedModelHint": "대체 서비스로 제공되는 {model} 요청",
"attempts": "시도",
"filterAll": "모두",
+ "filterCanceled": "취소됨",
"providerBreakdown": "공급자별 분석",
"tokensPerSec": "톡/초",
"allProviders": "모든 제공업체",
diff --git a/client/src/i18n/locales/lt.json b/client/src/i18n/locales/lt.json
index d8821b76..1c162c3e 100644
--- a/client/src/i18n/locales/lt.json
+++ b/client/src/i18n/locales/lt.json
@@ -550,6 +550,7 @@
"requestedModelHint": "Prašyta {model}, aptarnauta atsarginiu būdu",
"attempts": "Bandymai",
"filterAll": "Visi",
+ "filterCanceled": "Atšauktos",
"providerBreakdown": "Pasiskirstymas pagal tiekėją",
"tokensPerSec": "Tok/s",
"allProviders": "Visi tiekėjai",
diff --git a/client/src/i18n/locales/ml.json b/client/src/i18n/locales/ml.json
index aef638d2..31e3b6e2 100644
--- a/client/src/i18n/locales/ml.json
+++ b/client/src/i18n/locales/ml.json
@@ -550,6 +550,7 @@
"requestedModelHint": "{model} അഭ്യർത്ഥിച്ചു, ഫാൾബാക്ക് നൽകി",
"attempts": "ശ്രമങ്ങൾ",
"filterAll": "എല്ലാം",
+ "filterCanceled": "റദ്ദാക്കിയവ",
"providerBreakdown": "ഓരോ ദാതാവിനും തകരാർ",
"tokensPerSec": "ടോക്ക്/സെ",
"allProviders": "എല്ലാ ദാതാക്കളും",
diff --git a/client/src/i18n/locales/mr.json b/client/src/i18n/locales/mr.json
index 416c7cfc..b7cee743 100644
--- a/client/src/i18n/locales/mr.json
+++ b/client/src/i18n/locales/mr.json
@@ -550,6 +550,7 @@
"requestedModelHint": "{model} ची विनंती केली, फॉलबॅकद्वारे सेवा दिली",
"attempts": "प्रयत्न",
"filterAll": "सर्व",
+ "filterCanceled": "रद्द",
"providerBreakdown": "प्रति-प्रदाता ब्रेकडाउन",
"tokensPerSec": "टोक/से",
"allProviders": "सर्व प्रदाते",
diff --git a/client/src/i18n/locales/ms.json b/client/src/i18n/locales/ms.json
index 7491a65c..c3c9f81c 100644
--- a/client/src/i18n/locales/ms.json
+++ b/client/src/i18n/locales/ms.json
@@ -550,6 +550,7 @@
"requestedModelHint": "Permintaan {model}, dilayani sebagai sandaran",
"attempts": "Percubaan",
"filterAll": "Semua",
+ "filterCanceled": "Dibatalkan",
"providerBreakdown": "Pecahan setiap penyedia",
"tokensPerSec": "Tok/s",
"allProviders": "Semua penyedia",
diff --git a/client/src/i18n/locales/my.json b/client/src/i18n/locales/my.json
index f5e90ce5..7d81765c 100644
--- a/client/src/i18n/locales/my.json
+++ b/client/src/i18n/locales/my.json
@@ -550,6 +550,7 @@
"requestedModelHint": "တောင်းဆိုထားသော {model}၊ fallback ဖြင့် ဝန်ဆောင်မှု",
"attempts": "ကြိုးစားမှုများ",
"filterAll": "အားလုံး",
+ "filterCanceled": "ပယ်ဖျက်ပြီး",
"providerBreakdown": "ပံ့ပိုးသူအလိုက် ခွဲခြားချက်",
"tokensPerSec": "Tok/s",
"allProviders": "ပံ့ပိုးသူအားလုံး",
diff --git a/client/src/i18n/locales/ne.json b/client/src/i18n/locales/ne.json
index de266cc2..234521e0 100644
--- a/client/src/i18n/locales/ne.json
+++ b/client/src/i18n/locales/ne.json
@@ -550,6 +550,7 @@
"requestedModelHint": "{model} अनुरोध गरियो, फलब्याकद्वारा सेवा गरियो",
"attempts": "प्रयासहरू",
"filterAll": "सबै",
+ "filterCanceled": "रद्द गरिएका",
"providerBreakdown": "प्रत्येक प्रदायक क्रमाङ्कन",
"tokensPerSec": "टक/एस",
"allProviders": "सबै प्रदायकहरू",
diff --git a/client/src/i18n/locales/nl.json b/client/src/i18n/locales/nl.json
index e8026978..ca054618 100644
--- a/client/src/i18n/locales/nl.json
+++ b/client/src/i18n/locales/nl.json
@@ -550,6 +550,7 @@
"requestedModelHint": "Gevraagd {model}, bediend via reserveplan",
"attempts": "Pogingen",
"filterAll": "Allemaal",
+ "filterCanceled": "Geannuleerd",
"providerBreakdown": "Uitsplitsing per aanbieder",
"tokensPerSec": "Tok/s",
"allProviders": "Alle aanbieders",
diff --git a/client/src/i18n/locales/no.json b/client/src/i18n/locales/no.json
index 50dda12c..73794364 100644
--- a/client/src/i18n/locales/no.json
+++ b/client/src/i18n/locales/no.json
@@ -550,6 +550,7 @@
"requestedModelHint": "Bedt om {model}, betjent som reserveplan",
"attempts": "Forsøk",
"filterAll": "Alle",
+ "filterCanceled": "Avbrutt",
"providerBreakdown": "Fordeling per leverandør",
"tokensPerSec": "Tok/s",
"allProviders": "Alle tilbydere",
diff --git a/client/src/i18n/locales/or.json b/client/src/i18n/locales/or.json
index a48a0e61..cd748b2c 100644
--- a/client/src/i18n/locales/or.json
+++ b/client/src/i18n/locales/or.json
@@ -550,6 +550,7 @@
"requestedModelHint": "{model} ଅନୁରୋଧ କରାଯାଇଛି, ଫଲବ୍ୟାକ୍ ଦ୍ୱାରା ପରିବେଷଣ କରାଯାଇଛି",
"attempts": "ଉଦ୍ୟମ",
"filterAll": "ସମସ୍ତ",
+ "filterCanceled": "ବାତିଲ",
"providerBreakdown": "ପ୍ରତି-ପ୍ରଦାନକାରୀ ବ୍ରେକଡାଉନ୍",
"tokensPerSec": "ଟୋକ୍ / ସେକେଣ୍ଡ",
"allProviders": "ସମସ୍ତ ପ୍ରଦାନକାରୀ",
diff --git a/client/src/i18n/locales/pa.json b/client/src/i18n/locales/pa.json
index 7d81a677..9d99f47a 100644
--- a/client/src/i18n/locales/pa.json
+++ b/client/src/i18n/locales/pa.json
@@ -550,6 +550,7 @@
"requestedModelHint": "{model} ਦੀ ਬੇਨਤੀ ਕੀਤੀ, ਫਾਲਬੈਕ ਦੁਆਰਾ ਸੇਵਾ ਕੀਤੀ ਗਈ",
"attempts": "ਕੋਸ਼ਿਸ਼ਾਂ",
"filterAll": "ਸਾਰੇ",
+ "filterCanceled": "ਰੱਦ ਕੀਤੇ",
"providerBreakdown": "ਪ੍ਰਤੀ-ਪ੍ਰਦਾਤਾ ਟੁੱਟਣ",
"tokensPerSec": "Tok/s",
"allProviders": "ਸਾਰੇ ਪ੍ਰਦਾਤਾ",
diff --git a/client/src/i18n/locales/pl.json b/client/src/i18n/locales/pl.json
index 1265b8f4..4d6294ff 100644
--- a/client/src/i18n/locales/pl.json
+++ b/client/src/i18n/locales/pl.json
@@ -550,6 +550,7 @@
"requestedModelHint": "Żądano {model}, obsługiwane przez rezerwę",
"attempts": "Próby",
"filterAll": "Wszystko",
+ "filterCanceled": "Anulowane",
"providerBreakdown": "Podział na dostawcę",
"tokensPerSec": "Tok/s",
"allProviders": "Wszyscy dostawcy",
diff --git a/client/src/i18n/locales/pt-BR.json b/client/src/i18n/locales/pt-BR.json
index c2bd6f85..5dd444c2 100644
--- a/client/src/i18n/locales/pt-BR.json
+++ b/client/src/i18n/locales/pt-BR.json
@@ -550,6 +550,7 @@
"requestedModelHint": "Solicitado {model}, servido por fallback",
"attempts": "Tentativas",
"filterAll": "Todas",
+ "filterCanceled": "Canceladas",
"providerBreakdown": "Detalhamento por provedor",
"tokensPerSec": "Tok/s",
"allProviders": "Todos os provedores",
diff --git a/client/src/i18n/locales/pt-PT.json b/client/src/i18n/locales/pt-PT.json
index 328b777f..4ff11348 100644
--- a/client/src/i18n/locales/pt-PT.json
+++ b/client/src/i18n/locales/pt-PT.json
@@ -550,6 +550,7 @@
"requestedModelHint": "Solicitado {model}, servido por recuo",
"attempts": "Tentativas",
"filterAll": "Todos",
+ "filterCanceled": "Canceladas",
"providerBreakdown": "Distribuição por fornecedor",
"tokensPerSec": "Tok/s",
"allProviders": "Todos os prestadores",
diff --git a/client/src/i18n/locales/ro.json b/client/src/i18n/locales/ro.json
index 4e949b2d..95b55d50 100644
--- a/client/src/i18n/locales/ro.json
+++ b/client/src/i18n/locales/ro.json
@@ -550,6 +550,7 @@
"requestedModelHint": "Solicitat {model}, deservit de rezervă",
"attempts": "Încercări",
"filterAll": "Toate",
+ "filterCanceled": "Anulate",
"providerBreakdown": "Defalcare pe furnizor",
"tokensPerSec": "Tok/s",
"allProviders": "Toți furnizorii",
diff --git a/client/src/i18n/locales/ru.json b/client/src/i18n/locales/ru.json
index b75a3adb..b4612584 100644
--- a/client/src/i18n/locales/ru.json
+++ b/client/src/i18n/locales/ru.json
@@ -550,6 +550,7 @@
"requestedModelHint": "Запрошен {model}, обслуживается резервным способом.",
"attempts": "Попытки",
"filterAll": "Все",
+ "filterCanceled": "Отменённые",
"providerBreakdown": "Разбивка по провайдерам",
"tokensPerSec": "Ток/с",
"allProviders": "Все провайдеры",
diff --git a/client/src/i18n/locales/si.json b/client/src/i18n/locales/si.json
index 3650c1d8..8e9c62c4 100644
--- a/client/src/i18n/locales/si.json
+++ b/client/src/i18n/locales/si.json
@@ -550,6 +550,7 @@
"requestedModelHint": "{model} ඉල්ලන ලදී, පසුබැසීම මගින් සේවය කරන ලදී",
"attempts": "උත්සාහයන්",
"filterAll": "සියල්ල",
+ "filterCanceled": "අවලංගු කළ",
"providerBreakdown": "එක්-සැපයුම්කරු බිඳ වැටීම",
"tokensPerSec": "Tok/s",
"allProviders": "සියලුම සපයන්නන්",
diff --git a/client/src/i18n/locales/sk.json b/client/src/i18n/locales/sk.json
index b7a3aa36..e2f2a091 100644
--- a/client/src/i18n/locales/sk.json
+++ b/client/src/i18n/locales/sk.json
@@ -550,6 +550,7 @@
"requestedModelHint": "Požiadaný {model}, obsluhovaný záložnou zálohou",
"attempts": "Pokusy",
"filterAll": "Všetko",
+ "filterCanceled": "Zrušené",
"providerBreakdown": "Rozdelenie podľa poskytovateľa",
"tokensPerSec": "Tok/s",
"allProviders": "Všetci poskytovatelia",
diff --git a/client/src/i18n/locales/sr.json b/client/src/i18n/locales/sr.json
index 238deed0..0bb1f114 100644
--- a/client/src/i18n/locales/sr.json
+++ b/client/src/i18n/locales/sr.json
@@ -550,6 +550,7 @@
"requestedModelHint": "Захтевано {model}, сервирано резервним путем",
"attempts": "Покушаји",
"filterAll": "Све",
+ "filterCanceled": "Отказани",
"providerBreakdown": "Анализа по добављачима",
"tokensPerSec": "Ток/с",
"allProviders": "Сви провајдери",
diff --git a/client/src/i18n/locales/sv.json b/client/src/i18n/locales/sv.json
index f7fb2dae..a89feb06 100644
--- a/client/src/i18n/locales/sv.json
+++ b/client/src/i18n/locales/sv.json
@@ -550,6 +550,7 @@
"requestedModelHint": "Begärt {model}, servat som reservplan",
"attempts": "Försök",
"filterAll": "Alla",
+ "filterCanceled": "Avbrutna",
"providerBreakdown": "Uppdelning per leverantör",
"tokensPerSec": "Tok/s",
"allProviders": "Alla leverantörer",
diff --git a/client/src/i18n/locales/sw.json b/client/src/i18n/locales/sw.json
index 76e48554..b6177671 100644
--- a/client/src/i18n/locales/sw.json
+++ b/client/src/i18n/locales/sw.json
@@ -550,6 +550,7 @@
"requestedModelHint": "Umeomba {model}, ukitolewa kwa njia mbadala",
"attempts": "Majaribio",
"filterAll": "Wote",
+ "filterCanceled": "Zilizoghairiwa",
"providerBreakdown": "Uchanganuzi wa kila mtoaji",
"tokensPerSec": "Tok/s",
"allProviders": "Watoa huduma wote",
diff --git a/client/src/i18n/locales/ta.json b/client/src/i18n/locales/ta.json
index 4de6b6ad..8b08f26a 100644
--- a/client/src/i18n/locales/ta.json
+++ b/client/src/i18n/locales/ta.json
@@ -550,6 +550,7 @@
"requestedModelHint": "{model} கோரப்பட்டது, ஃபால்பேக் மூலம் சேவை வழங்கப்பட்டது",
"attempts": "முயற்சிகள்",
"filterAll": "அனைத்து",
+ "filterCanceled": "ரத்தானவை",
"providerBreakdown": "ஒவ்வொரு வழங்குநருக்கும் முறிவு",
"tokensPerSec": "டோக்/கள்",
"allProviders": "அனைத்து வழங்குநர்கள்",
diff --git a/client/src/i18n/locales/te.json b/client/src/i18n/locales/te.json
index 40671d80..9ca38ff1 100644
--- a/client/src/i18n/locales/te.json
+++ b/client/src/i18n/locales/te.json
@@ -550,6 +550,7 @@
"requestedModelHint": "ఫాల్‌బ్యాక్ ద్వారా అందించబడిన {model} అభ్యర్థించబడింది",
"attempts": "ప్రయత్నాలు",
"filterAll": "అన్నీ",
+ "filterCanceled": "రద్దైనవి",
"providerBreakdown": "ప్రతి ప్రొవైడర్ బ్రేక్‌డౌన్",
"tokensPerSec": "టోక్/లు",
"allProviders": "అందరు ప్రొవైడర్లు",
diff --git a/client/src/i18n/locales/th.json b/client/src/i18n/locales/th.json
index 8e41bd96..72aafb7a 100644
--- a/client/src/i18n/locales/th.json
+++ b/client/src/i18n/locales/th.json
@@ -550,6 +550,7 @@
"requestedModelHint": "ร้องขอ {model} ซึ่งให้บริการโดยทางเลือก",
"attempts": "ความพยายาม",
"filterAll": "ทั้งหมด",
+ "filterCanceled": "ยกเลิกแล้ว",
"providerBreakdown": "รายละเอียดต่อผู้ให้บริการ",
"tokensPerSec": "ต๊อก/ส",
"allProviders": "ผู้ให้บริการทั้งหมด",
diff --git a/client/src/i18n/locales/tl.json b/client/src/i18n/locales/tl.json
index 6ff5f02b..b7802cce 100644
--- a/client/src/i18n/locales/tl.json
+++ b/client/src/i18n/locales/tl.json
@@ -550,6 +550,7 @@
"requestedModelHint": "Hiniling na {model}, pinagsilbihan bilang fallback",
"attempts": "Mga Pagtatangka",
"filterAll": "Lahat",
+ "filterCanceled": "Kanselado",
"providerBreakdown": "Pagkakahati-hati ng bawat provider",
"tokensPerSec": "Tok/s",
"allProviders": "Lahat ng provider",
diff --git a/client/src/i18n/locales/tr.json b/client/src/i18n/locales/tr.json
index db791843..c0b6c08b 100644
--- a/client/src/i18n/locales/tr.json
+++ b/client/src/i18n/locales/tr.json
@@ -550,6 +550,7 @@
"requestedModelHint": "İstenen {model}, yedek tarafından sunuldu",
"attempts": "denemeler",
"filterAll": "Hepsi",
+ "filterCanceled": "İptal edilenler",
"providerBreakdown": "Sağlayıcı başına döküm",
"tokensPerSec": "Tok/s",
"allProviders": "Tüm sağlayıcılar",
diff --git a/client/src/i18n/locales/uk.json b/client/src/i18n/locales/uk.json
index e8038a4c..5eef604a 100644
--- a/client/src/i18n/locales/uk.json
+++ b/client/src/i18n/locales/uk.json
@@ -550,6 +550,7 @@
"requestedModelHint": "Запитано {model}, обслуговується резервним способом",
"attempts": "Спроби",
"filterAll": "все",
+ "filterCanceled": "Скасовані",
"providerBreakdown": "Розподіл за постачальником",
"tokensPerSec": "Ток/с",
"allProviders": "Всі провайдери",
diff --git a/client/src/i18n/locales/ur.json b/client/src/i18n/locales/ur.json
index 3a06858c..449e4fa8 100644
--- a/client/src/i18n/locales/ur.json
+++ b/client/src/i18n/locales/ur.json
@@ -550,6 +550,7 @@
"requestedModelHint": "{model} کی درخواست کی گئی، فال بیک کے ذریعے پیش کی گئی۔",
"attempts": "کوششیں",
"filterAll": "تمام",
+ "filterCanceled": "منسوخ شدہ",
"providerBreakdown": "فی فراہم کنندہ کی خرابی",
"tokensPerSec": "Tok/s",
"allProviders": "تمام فراہم کنندگان",
diff --git a/client/src/i18n/locales/uz.json b/client/src/i18n/locales/uz.json
index a2d156a7..5f932a2b 100644
--- a/client/src/i18n/locales/uz.json
+++ b/client/src/i18n/locales/uz.json
@@ -550,6 +550,7 @@
"requestedModelHint": "Soʻralgan {model}, zaxira orqali xizmat koʻrsatadi",
"attempts": "Urinishlar",
"filterAll": "Hammasi",
+ "filterCanceled": "Bekor qilingan",
"providerBreakdown": "Har bir provayder bo'yicha taqsimot",
"tokensPerSec": "Tok/s",
"allProviders": "Barcha provayderlar",
diff --git a/client/src/i18n/locales/vi.json b/client/src/i18n/locales/vi.json
index 59329cd9..f3e17ed6 100644
--- a/client/src/i18n/locales/vi.json
+++ b/client/src/i18n/locales/vi.json
@@ -550,6 +550,7 @@
"requestedModelHint": "{model} được yêu cầu, được cung cấp bởi dự phòng",
"attempts": "Nỗ lực",
"filterAll": "Tất cả",
+ "filterCanceled": "Đã hủy",
"providerBreakdown": "Phân tích theo nhà cung cấp",
"tokensPerSec": "Tok/giây",
"allProviders": "Tất cả các nhà cung cấp",
diff --git a/client/src/i18n/locales/yo.json b/client/src/i18n/locales/yo.json
index 94a199da..ad6a3088 100644
--- a/client/src/i18n/locales/yo.json
+++ b/client/src/i18n/locales/yo.json
@@ -550,6 +550,7 @@
"requestedModelHint": "Ti beere {model}, yoo wa nipasẹ fallback",
"attempts": "Awọn igbiyanju",
"filterAll": "Gbogbo",
+ "filterCanceled": "Ti fagile",
"providerBreakdown": "Pipin-olupese",
"tokensPerSec": "Tok/s",
"allProviders": "Gbogbo olupese",
diff --git a/client/src/i18n/locales/zh-CN.json b/client/src/i18n/locales/zh-CN.json
index f128c6c7..ca1c37b2 100644
--- a/client/src/i18n/locales/zh-CN.json
+++ b/client/src/i18n/locales/zh-CN.json
@@ -550,6 +550,7 @@
"requestedModelHint": "请求了 {model},由回退模型提供",
"attempts": "尝试次数",
"filterAll": "全部",
+ "filterCanceled": "已取消",
"providerBreakdown": "按提供方细分",
"tokensPerSec": "词元/秒",
"allProviders": "全部提供方",
diff --git a/client/src/i18n/locales/zh-TW.json b/client/src/i18n/locales/zh-TW.json
index 379a55e5..441561ed 100644
--- a/client/src/i18n/locales/zh-TW.json
+++ b/client/src/i18n/locales/zh-TW.json
@@ -550,6 +550,7 @@
"requestedModelHint": "請求了 {model},由備援模型提供",
"attempts": "嘗試次數",
"filterAll": "全部",
+ "filterCanceled": "已取消",
"providerBreakdown": "按提供者細分",
"tokensPerSec": "Token/秒",
"allProviders": "全部提供者",
diff --git a/client/src/pages/AnalyticsPage.tsx b/client/src/pages/AnalyticsPage.tsx
index 8f4ef133..c1d0e959 100644
--- a/client/src/pages/AnalyticsPage.tsx
+++ b/client/src/pages/AnalyticsPage.tsx
@@ -187,7 +187,15 @@ interface RequestDetail extends Omit<RecentCallRow, 'attemptCount'> {
attempts: RequestAttempt[]
}
-type StatusFilter = 'all' | 'success' | 'error'
+type StatusFilter = 'all' | 'success' | 'error' | 'canceled'
+
+// 'canceled' (#752 — the client hung up mid-request) is neither success nor
+// error: neutral amber, not destructive red.
+function statusTextClass(status: string): string {
+ if (status === 'success') return 'text-muted-foreground'
+ if (status === 'canceled') return 'text-amber-600 dark:text-amber-400'
+ return 'text-destructive'
+}
// First product token of the UA ("python-requests/2.32.3", "curl/8.6.0", …)
// is enough to tell callers apart in a narrow cell; full string on hover.
@@ -310,7 +318,7 @@ function RequestDetailDialog({ requestId, onClose }: { requestId: number | null;
<DetailField
label={t('common.status')}
value={
- <span className={detail.status === 'success' ? '' : 'text-destructive'}>{detail.status}</span>
+ <span className={detail.status === 'success' ? '' : statusTextClass(detail.status)}>{detail.status}</span>
}
/>
<DetailField
@@ -362,7 +370,8 @@ function RequestDetailDialog({ requestId, onClose }: { requestId: number | null;
<span className="font-medium">{a.platform}</span>
<span className="text-muted-foreground truncate" title={a.modelId}>{a.modelId}</span>
<Badge variant="outline">{t('analytics.keyOrdinal', { n: a.keyOrdinal })}</Badge>
- <Badge variant={a.outcome === 'ok' || a.outcome === 'committed' ? 'secondary' : 'destructive'}>
+ {/* client_abort is the caller's doing, not a hop failure. */}
+ <Badge variant={a.outcome === 'ok' || a.outcome === 'committed' ? 'secondary' : a.outcome === 'client_abort' ? 'outline' : 'destructive'}>
{a.outcome}
</Badge>
<span
@@ -765,6 +774,7 @@ export default function AnalyticsPage() {
{ value: 'all', label: t('analytics.filterAll') },
{ value: 'success', label: t('common.success') },
{ value: 'error', label: t('analytics.errors') },
+ { value: 'canceled', label: t('analytics.filterCanceled') },
]}
ariaLabel={t('common.status')}
/>
@@ -827,7 +837,7 @@ export default function AnalyticsPage() {
{r.requestedModel && r.requestedModel !== r.modelId ? ' *' : ''}
</TableCell>
<TableCell className="text-xs text-muted-foreground">{r.platform}</TableCell>
- <TableCell className={`text-xs ${r.status === 'success' ? 'text-muted-foreground' : 'text-destructive'}`} title={r.error ?? undefined}>
+ <TableCell className={`text-xs ${statusTextClass(r.status)}`} title={r.error ?? undefined}>
{r.status}
</TableCell>
{/* >1 = the request burned failover hops; that is the
diff --git a/server/src/__tests__/lib/fallback-loop-client-abort.test.ts b/server/src/__tests__/lib/fallback-loop-client-abort.test.ts
index f98f3b7f..0fe59405 100644
--- a/server/src/__tests__/lib/fallback-loop-client-abort.test.ts
+++ b/server/src/__tests__/lib/fallback-loop-client-abort.test.ts
@@ -11,7 +11,9 @@ import type { RouteResult } from '../../services/router.js';
* the in-flight attempt with the marked error from newClientAbortError. That
* throw must stop the ladder immediately, release the in-flight lease via the
* existing finally, and leave NO provider-failure trace — no cooldown, no
- * skip-key entry, no logFailure row, no exhaustion render.
+ * skip-key entry, no logFailure row, no exhaustion render. What it DOES leave
+ * (#752) is a 'canceled' requests row, so the dashboard shows the request
+ * happened; stats/scoring queries exclude that status everywhere.
*/
let keySeq = 900;
@@ -123,6 +125,26 @@ describe('fallback loop on client abort', () => {
expect(state.skipModels.size).toBe(0);
});
+ it("writes a 'canceled' requests row so the abort is visible in the dashboard (#752)", async () => {
+ const state = newFallbackState();
+ const { route } = leasedRoute();
+ await runFallbackLoop(hooks(state, {
+ route: () => route,
+ dispatch: async () => { throw newClientAbortError(); },
+ }));
+
+ const row = getDb().prepare(
+ 'SELECT id, platform, model_id, key_id, status, error FROM requests ORDER BY id DESC LIMIT 1',
+ ).get() as any;
+ expect(row).toMatchObject({ platform: 'fake', model_id: 'fake-model', status: 'canceled' });
+ expect(row.error).toContain('client disconnected');
+ expect(row.error).toContain('upstream request canceled');
+ // The attempt trace persists too, keyed to the canceled row — pure aborts
+ // used to leave no parent row and therefore no ladder either.
+ const attempts = getDb().prepare('SELECT outcome FROM request_attempts WHERE request_id = ?').all(row.id) as any[];
+ expect(attempts.map(a => a.outcome)).toEqual(['client_abort']);
+ });
+
it('clientGone still stops the loop before the next attempt after an ordinary failure', async () => {
// The pre-existing boundary: a NON-abort provider failure that lands after
// the disconnect is still recorded normally, but the loop must not start
diff --git a/server/src/__tests__/lib/fallback-loop.test.ts b/server/src/__tests__/lib/fallback-loop.test.ts
index bea71005..dff16745 100644
--- a/server/src/__tests__/lib/fallback-loop.test.ts
+++ b/server/src/__tests__/lib/fallback-loop.test.ts
@@ -7,7 +7,11 @@ import { describe, it, expect, beforeAll, beforeEach, vi } from 'vitest';
// contract enforcement.
const { mockCheckKeyHealth } = vi.hoisted(() => ({ mockCheckKeyHealth: vi.fn() }));
-vi.mock('../../services/health.js', () => ({ checkKeyHealth: mockCheckKeyHealth }));
+vi.mock('../../services/health.js', () => ({
+ checkKeyHealth: mockCheckKeyHealth,
+ // recordUpstreamSuccess touches this on the streak-reset test path.
+ markKeyHealthyFromRequest: vi.fn(),
+}));
import { initDb, getDb } from '../../db/index.js';
import {
@@ -15,6 +19,8 @@ import {
newFallbackState,
cooldownForError,
recordRetryableFailure,
+ recordUpstreamSuccess,
+ resetEmptyCompletionStreaks,
exhaustedRetryError,
formatAttemptTrail,
classifyAttemptError,
@@ -22,6 +28,7 @@ import {
getFallbackTimeBudgetMs,
DEFAULT_FALLBACK_TIME_BUDGET_MS,
AUTH_FAILURE_COOLDOWN_MS,
+ EMPTY_COMPLETION_STREAK_LIMIT,
type AttemptRecord,
type FallbackHooks,
} from '../../lib/fallback-loop.js';
@@ -68,6 +75,7 @@ beforeEach(() => {
mockCheckKeyHealth.mockReset();
mockCheckKeyHealth.mockResolvedValue('invalid');
getDb().prepare('DELETE FROM rate_limit_cooldowns').run();
+ resetEmptyCompletionStreaks();
});
describe('isKeyAuthError (401 = key-fatal, rotate instead of 502)', () => {
@@ -189,6 +197,80 @@ describe('recordRetryableFailure skipBench exemption (reasoning truncation)', ()
});
});
+describe('empty-completion streak lifts the skipBench exemption (#751)', () => {
+ const emptyLengthErr = (route: RouteResult) =>
+ Object.assign(new Error(`empty completion from ${route.displayName}`), { skipBench: true });
+ const cooldownFor = (route: RouteResult) =>
+ getDb().prepare('SELECT 1 FROM rate_limit_cooldowns WHERE platform = ? AND key_id = ?').get('fake', route.keyId);
+
+ it('benches + penalizes from the Nth consecutive empty completion on the same model+key', () => {
+ const route = fakeRoute();
+ for (let i = 1; i < EMPTY_COMPLETION_STREAK_LIMIT; i++) {
+ expect(recordRetryableFailure(route, emptyLengthErr(route), newFallbackState())).toBe(true);
+ expect(cooldownFor(route)).toBeUndefined();
+ }
+ // Streak limit reached: the exemption lifts and the normal path runs.
+ expect(recordRetryableFailure(route, emptyLengthErr(route), newFallbackState())).toBe(false);
+ expect(cooldownFor(route)).toBeDefined();
+ expect(getAllPenalties().some(p => p.modelDbId === route.modelDbId)).toBe(true);
+ });
+
+ it('a success on the model+key resets the streak', () => {
+ const route = fakeRoute();
+ for (let i = 1; i < EMPTY_COMPLETION_STREAK_LIMIT; i++) {
+ recordRetryableFailure(route, emptyLengthErr(route), newFallbackState());
+ }
+ recordUpstreamSuccess(route, 0);
+ // The full pre-limit run is available again.
+ for (let i = 1; i < EMPTY_COMPLETION_STREAK_LIMIT; i++) {
+ expect(recordRetryableFailure(route, emptyLengthErr(route), newFallbackState())).toBe(true);
+ }
+ expect(cooldownFor(route)).toBeUndefined();
+ });
+
+ it('a normally-penalized failure resets the streak too', () => {
+ const route = fakeRoute();
+ for (let i = 1; i < EMPTY_COMPLETION_STREAK_LIMIT; i++) {
+ recordRetryableFailure(route, emptyLengthErr(route), newFallbackState());
+ }
+ // A plain 429 takes the cooldown ladder — and breaks the streak.
+ recordRetryableFailure(route, Object.assign(new Error('429 Too Many Requests'), { status: 429 }), newFallbackState());
+ getDb().prepare('DELETE FROM rate_limit_cooldowns').run();
+ for (let i = 1; i < EMPTY_COMPLETION_STREAK_LIMIT; i++) {
+ expect(recordRetryableFailure(route, emptyLengthErr(route), newFallbackState())).toBe(true);
+ }
+ expect(cooldownFor(route)).toBeUndefined();
+ });
+
+ it('format violations (skipModelForRequest) stay exempt and never accrue', () => {
+ const route = fakeRoute();
+ const formatErr = () => Object.assign(
+ new Error(`${route.displayName} ignored response_format (returned non-JSON despite json_schema)`),
+ { skipBench: true, skipModelForRequest: true },
+ );
+ for (let i = 0; i < EMPTY_COMPLETION_STREAK_LIMIT * 2; i++) {
+ expect(recordRetryableFailure(route, formatErr(), newFallbackState())).toBe(true);
+ }
+ // …and they did not pre-charge the empty-completion streak either.
+ expect(recordRetryableFailure(route, emptyLengthErr(route), newFallbackState())).toBe(true);
+ expect(cooldownFor(route)).toBeUndefined();
+ });
+
+ it('an over-the-limit empty completion counts toward the breaker', async () => {
+ const route = fakeRoute();
+ const onExhausted = vi.fn();
+ const dispatch = vi.fn(async () => { throw emptyLengthErr(route); });
+
+ await runFallbackLoop(hooksSkeleton({ maxRetries: 20, breakerLimit: 2, route: () => route, dispatch, onExhausted }));
+
+ // The first LIMIT-1 failures are exempt (invisible to the breaker); the
+ // next two count and trip the 2-limit breaker.
+ expect(dispatch).toHaveBeenCalledTimes(EMPTY_COMPLETION_STREAK_LIMIT - 1 + 2);
+ expect(onExhausted).toHaveBeenCalledTimes(1);
+ expect(onExhausted.mock.calls[0][0].status).toBe(503);
+ });
+});
+
describe('exhaustedRetryError attempt trail + auth exhaustion', () => {
it('all-auth attempts produce a distinct 502 provider_error body, not a rate-limit 429', () => {
const body = exhaustedRetryError(new Error('Groq API error 401: Invalid API Key'), 20, { attempts: authRecord(3) });
@@ -275,7 +357,7 @@ describe('runFallbackLoop: auth rotation (401 is key-fatal, not request-fatal)',
});
describe('runFallbackLoop: wall-clock retry budget', () => {
- it('stops starting new attempts once the budget is spent and reports timedOut', async () => {
+ it('always allows one failover hop, then stops and reports timedOut (#751)', async () => {
const dispatch = vi.fn(async () => {
await new Promise(r => setTimeout(r, 25));
throw Object.assign(new Error('429 Too Many Requests'), { status: 429 });
@@ -288,12 +370,15 @@ describe('runFallbackLoop: wall-clock retry budget', () => {
onExhausted,
}));
- expect(dispatch).toHaveBeenCalledTimes(1); // first attempt always runs, no second start
+ // Attempt 0 alone consumed the budget, but attempt 1 still starts — a
+ // slow-failing model must never make failover structurally impossible.
+ // Attempt 2 is where the spent budget stops the chain.
+ expect(dispatch).toHaveBeenCalledTimes(2);
expect(onExhausted).toHaveBeenCalledTimes(1);
const [body, info] = onExhausted.mock.calls[0];
expect(info.timedOut).toBe(true);
expect(body.message).toContain('retry time budget');
- expect(info.attempts).toHaveLength(1);
+ expect(info.attempts).toHaveLength(2);
});
it('budget 0 disables the check', async () => {
diff --git a/server/src/__tests__/routes/analytics-requests.test.ts b/server/src/__tests__/routes/analytics-requests.test.ts
index d47fc1cd..cd676196 100644
--- a/server/src/__tests__/routes/analytics-requests.test.ts
+++ b/server/src/__tests__/routes/analytics-requests.test.ts
@@ -101,6 +101,18 @@ describe('GET /api/analytics/requests', () => {
expect(successes.body.rows.map((r: any) => r.status)).toEqual(['success']);
});
+ it("filters by the 'canceled' status (#752)", async () => {
+ insertCall(recentUtcTimestamp(8).sql, '10.0.0.1', 'ua', 'success');
+ insertCall(recentUtcTimestamp(9).sql, '10.0.0.1', 'ua', 'canceled');
+
+ const canceled = await request(app, '/api/analytics/requests?range=7d&status=canceled');
+ expect(canceled.status).toBe(200);
+ expect(canceled.body.total).toBe(1);
+ expect(canceled.body.rows.map((r: any) => r.status)).toEqual(['canceled']);
+ // ...and canceled rows still show up unfiltered.
+ expect((await request(app, '/api/analytics/requests?range=7d')).body.total).toBe(2);
+ });
+
it('filters by platform, and combines with the status filter', async () => {
insertCall(recentUtcTimestamp(8).sql, '10.0.0.1', 'ua', 'success', 'groq');
insertCall(recentUtcTimestamp(9).sql, '10.0.0.1', 'ua', 'error', 'groq');
diff --git a/server/src/__tests__/routes/analytics.test.ts b/server/src/__tests__/routes/analytics.test.ts
index c5ab711e..b668b412 100644
--- a/server/src/__tests__/routes/analytics.test.ts
+++ b/server/src/__tests__/routes/analytics.test.ts
@@ -169,6 +169,27 @@ describe('Analytics API', () => {
expect(body.lifetimeTotalRequests).toBe(3);
});
+ it("counts 'canceled' rows in totals but in no success/error rate (#752)", async () => {
+ vi.useRealTimers();
+ logRequest('groq', 'llama-3.3-70b-versatile', 0, 'success', 100, 50, 12, null);
+ logRequest('groq', 'llama-3.3-70b-versatile', 0, 'error', 30, 0, 9, 'boom');
+ logRequest('groq', 'llama-3.3-70b-versatile', 0, 'canceled', 0, 0, 800,
+ 'client disconnected after 0.8s; upstream request canceled');
+
+ const summary = await request(app, '/api/analytics/summary?range=24h');
+ expect(summary.status).toBe(200);
+ expect(summary.body.totalRequests).toBe(3); // the canceled request happened
+ expect(summary.body.successRate).toBe(50); // 1 of 2 decided — not 1 of 3
+
+ const byModel = await request(app, '/api/analytics/by-model?range=24h');
+ expect(byModel.body[0].requests).toBe(3);
+ expect(byModel.body[0].successRate).toBe(50);
+
+ const byPlatform = await request(app, '/api/analytics/by-platform?range=24h');
+ expect(byPlatform.body[0].successRate).toBe(50);
+ expect(byPlatform.body[0].errorCount).toBe(1); // canceled is not an error
+ });
+
it('prices savings at the served model paid-equivalent rate', async () => {
// groq/llama-3.3-70b-versatile is mapped at $0.10/M in, $0.32/M out
// (db/model-pricing.ts): 10M in + 5M out → 1.00 + 1.60 = $2.60
diff --git a/server/src/__tests__/routes/fallback-hardening.test.ts b/server/src/__tests__/routes/fallback-hardening.test.ts
index 112fcf7a..19015958 100644
--- a/server/src/__tests__/routes/fallback-hardening.test.ts
+++ b/server/src/__tests__/routes/fallback-hardening.test.ts
@@ -29,7 +29,7 @@ const { createApp } = await import('../../app.js');
const { initDb, getDb, getUnifiedApiKey } = await import('../../db/index.js');
const { encrypt } = await import('../../lib/crypto.js');
const { setRoutingStrategy, getAllPenalties } = await import('../../services/router.js');
-const { msUntilNextUtcMidnight } = await import('../../lib/fallback-loop.js');
+const { msUntilNextUtcMidnight, resetEmptyCompletionStreaks, EMPTY_COMPLETION_STREAK_LIMIT } = await import('../../lib/fallback-loop.js');
async function post(app: Express, path: string, body: any, key: string) {
const server = app.listen(0);
@@ -97,6 +97,7 @@ describe('fallback hardening (items 3, 4, 5, 6)', () => {
db.prepare('DELETE FROM rate_limit_cooldowns').run();
db.prepare('DELETE FROM rate_limit_usage').run();
db.prepare('DELETE FROM requests').run();
+ resetEmptyCompletionStreaks();
});
afterEach(() => {
@@ -181,7 +182,30 @@ describe('fallback hardening (items 3, 4, 5, 6)', () => {
expect(cooldownRowFor(firstModel)).toBeUndefined();
});
- it('item 5: the wall-clock budget stops retries and returns the rich exhaustion error', async () => {
+ it('item 4 bound (#751): consecutive empty-length completions bench the model+key at the streak limit', async () => {
+ // One empty-length turn is reasoning truncation; the Nth in a row on the
+ // same model+key is a broken model. Each request here fails over to a
+ // healthy sibling, but the streak accrues on the FIRST model.
+ let firstModel = '';
+ for (let i = 1; i <= EMPTY_COMPLETION_STREAK_LIMIT; i++) {
+ chatCompletion.mockReset();
+ chatCompletion
+ .mockResolvedValueOnce(emptyWithFinish('length'))
+ .mockResolvedValueOnce(GOOD_RESULT);
+ const { status } = await post(app, '/v1/chat/completions', {
+ messages: [{ role: 'user', content: `empty streak request ${i}` }],
+ }, key);
+ expect(status).toBe(200); // every request still failed over and served
+ firstModel = chatCompletion.mock.calls[0][2] as string;
+ if (i < EMPTY_COMPLETION_STREAK_LIMIT) {
+ expect(cooldownRowFor(firstModel)).toBeUndefined(); // still exempt
+ }
+ }
+ // The limit-th consecutive empty completion finally benches it.
+ expect(cooldownRowFor(firstModel)).toBeDefined();
+ });
+
+ it('item 5: the wall-clock budget allows one failover hop, then stops with the rich exhaustion error', async () => {
process.env.FALLBACK_TIME_BUDGET_MS = '1';
chatCompletion.mockImplementation(async () => {
await new Promise(r => setTimeout(r, 15)); // ensure the budget is spent after attempt 0
@@ -193,10 +217,12 @@ describe('fallback hardening (items 3, 4, 5, 6)', () => {
}, key);
expect(status).toBe(429);
- expect(chatCompletion).toHaveBeenCalledTimes(1); // the first attempt always runs; no second start
+ // The first attempt always runs, and so does the first RETRY even on a
+ // spent budget (#751) — attempt 2 is where the chain stops.
+ expect(chatCompletion).toHaveBeenCalledTimes(2);
expect(body.error.message).toContain('retry time budget');
expect(body.error.message).toContain('Attempt trail:');
- expect(headers.get('x-fallback-attempts')).toBe('1');
+ expect(headers.get('x-fallback-attempts')).toBe('2');
});
it('item 6: /v1/chat/completions logs estimated tokens when the provider omits usage', async () => {
diff --git a/server/src/__tests__/services/router-speed-scoring.test.ts b/server/src/__tests__/services/router-speed-scoring.test.ts
index 37a835b7..a3a14d4b 100644
--- a/server/src/__tests__/services/router-speed-scoring.test.ts
+++ b/server/src/__tests__/services/router-speed-scoring.test.ts
@@ -142,6 +142,26 @@ describe('speed scoring: timeouts count against speed (#619)', () => {
const fast = scores.find(s => s.modelId === 'fast')!;
expect(refusing.reliability).toBeLessThan(fast.reliability);
});
+
+ it("a 'canceled' row (#752 — client hung up) affects neither speed nor reliability", () => {
+ addModel({ platform: 'google', modelId: 'steady', name: 'Steady', priority: 1 });
+ addModel({ platform: 'groq', modelId: 'hungup', name: 'HungUp', priority: 2 });
+ for (const [platform, modelId] of [['google', 'steady'], ['groq', 'hungup']] as const) {
+ addRequests(platform, modelId, { count: 40, status: 'success', outTokens: 100, latencyMs: 1000, ttfbMs: 200 });
+ }
+ // Long-latency canceled rows would read as timeouts/failures if counted.
+ addRequests('groq', 'hungup', {
+ count: 20, status: 'canceled', outTokens: 0, latencyMs: 120_000,
+ error: 'client disconnected after 120.0s; upstream request canceled',
+ });
+
+ refreshStatsCache(getDb(), true);
+ expect(speedOf('hungup')).toBeCloseTo(speedOf('steady'), 10);
+ const { scores } = getRoutingScores();
+ const hungup = scores.find(s => s.modelId === 'hungup')!;
+ const steady = scores.find(s => s.modelId === 'steady')!;
+ expect(hungup.reliability).toBeCloseTo(steady.reliability, 10);
+ });
});
describe('observed speed_rank writeback (#619)', () => {
diff --git a/server/src/lib/fallback-loop.ts b/server/src/lib/fallback-loop.ts
index e5b4200c..e775df5c 100644
--- a/server/src/lib/fallback-loop.ts
+++ b/server/src/lib/fallback-loop.ts
@@ -49,7 +49,7 @@ import { noteModelRetirementSignal } from '../services/model-retirement.js';
import { getSetting } from '../db/index.js';
import { newBreaker, recordBreakerFailure } from './guardrails.js';
import { getRequestTrace, newRequestTrace, runWithRequestTrace, type AttemptOutcome, type RequestTrace } from './attempt-trace.js';
-import { persistRequestAttempts } from './request-log.js';
+import { logRequest, persistRequestAttempts } from './request-log.js';
// Every surface caps failover hops at the same number.
export const FALLBACK_MAX_RETRIES = 20;
@@ -57,9 +57,13 @@ export const FALLBACK_MAX_RETRIES = 20;
// ── Wall-clock retry budget ──────────────────────────────────────────────────
// Serial failover has no time bound of its own: the observed worst case was a
// 38.8s TTFB over 11 attempts, and the theoretical worst is maxRetries x the
-// per-attempt HTTP timeout. The budget is checked before STARTING each retry
-// (the first attempt always runs), so one slow attempt is never aborted
-// mid-flight — it just becomes the last one. 0 disables the budget entirely.
+// per-attempt HTTP timeout. The budget is checked before STARTING each retry,
+// so one slow attempt is never aborted mid-flight — it just becomes the last
+// one. The first attempt always runs, and so does the FIRST retry: when
+// attempt 0 alone consumes the whole budget (a slow-failing model), refusing
+// attempt 1 would make failover structurally impossible for exactly the
+// requests that need it (#751). The budget stops attempts >= 2 only.
+// 0 disables the budget entirely.
// Precedence mirrors the response cache: the settings-table value wins when
// present (runtime-tunable), then the env var, then the default.
// TODO(fallback-v2): AbortController hedging so a stalled attempt can be
@@ -153,6 +157,40 @@ export function cooldownDecisionForError(route: RouteResult, err: any): Cooldown
);
}
+// ── Empty-completion streak (issue #751) ─────────────────────────────────────
+// The skipBench exemption below assumes an empty 'length' completion is
+// REQUEST behavior (this turn's max_tokens spent on hidden reasoning). A
+// model+key that produces them consecutively is broken, not unlucky — with an
+// unconditional exemption it stayed penalty-free forever while failing every
+// request routed to it. At the streak limit the exemption lifts and the
+// failure takes the normal cooldown/penalty/limit-learning path (and counts
+// toward the breaker). Format violations (skipModelForRequest) never accrue:
+// they are model behavior for THIS request's response_format, not key health.
+export const EMPTY_COMPLETION_STREAK_LIMIT = 3;
+const emptyCompletionStreaks = new Map<string, number>(); // "platform:modelId:keyId"
+
+export function resetEmptyCompletionStreaks(): void {
+ emptyCompletionStreaks.clear();
+}
+
+// Advance (or break) the streak for this failure and report whether the
+// skipBench exemption still holds. Called exactly once per retryable failure,
+// from recordRetryableFailure; the loop reuses its returned decision for the
+// breaker so the two consumers can never disagree.
+function consumeSkipBenchExemption(route: RouteResult, err: any): boolean {
+ const key = `${route.platform}:${route.modelId}:${route.keyId}`;
+ if (err?.skipBench !== true) {
+ // A normally-penalized failure breaks the streak: the cooldown ladder is
+ // already handling whatever is wrong with this model+key.
+ emptyCompletionStreaks.delete(key);
+ return false;
+ }
+ if (err?.skipModelForRequest === true) return true;
+ const streak = (emptyCompletionStreaks.get(key) ?? 0) + 1;
+ emptyCompletionStreaks.set(key, streak);
+ return streak < EMPTY_COMPLETION_STREAK_LIMIT;
+}
+
/**
* Apply the full per-key failure bookkeeping shared by every surface after a
* retryable failure:
@@ -172,11 +210,17 @@ export function cooldownDecisionForError(route: RouteResult, err: any): Cooldown
* finish_reason 'length') still fails over — the key is skipped for THIS
* request — but is NOT a provider-health signal, so no cooldown, no model
* penalty, and no limit-learning are recorded. Benching those was costing
- * healthy models a 90s cooldown + a scorer penalty per truncated turn.
+ * healthy models a 90s cooldown + a scorer penalty per truncated turn. The
+ * exemption is streak-bounded (#751): from the EMPTY_COMPLETION_STREAK_LIMITth
+ * consecutive empty completion on the same model+key it stops applying, until
+ * a success (or a normally-penalized failure) resets the streak.
+ *
+ * Returns whether the skipBench exemption held for this failure, so the loop
+ * can keep the breaker in lockstep with the bench decision.
*
* Callers add the just-failed key to skipKeys via this function (do not pre-add).
*/
-export function recordRetryableFailure(route: RouteResult, err: any, state: FallbackState): void {
+export function recordRetryableFailure(route: RouteResult, err: any, state: FallbackState): boolean {
// `skipModelForRequest: true` = the failure is MODEL behavior, not key
// state (ignored response_format, JSON truncated at max_tokens): a sibling
// key would reproduce it exactly, so rule out the whole model for this
@@ -194,7 +238,7 @@ export function recordRetryableFailure(route: RouteResult, err: any, state: Fall
// sibling keys counts as the single observation it is.
noteModelRetirementSignal(route, err, getRequestTrace());
state.skipKeys.add(`${route.platform}:${route.modelId}:${route.keyId}`);
- if (err?.skipBench === true) return;
+ if (consumeSkipBenchExemption(route, err)) return true;
const decision = cooldownDecisionForError(route, err);
setCooldown(route.platform, route.modelId, route.keyId, decision.durationMs, decision.source);
// Model-level penalty only when no sibling key can still serve (#454).
@@ -202,6 +246,7 @@ export function recordRetryableFailure(route: RouteResult, err: any, state: Fall
recordRateLimitHit(route.modelDbId);
}
learnLimitFromError(route.modelDbId, err);
+ return false;
}
// ── Upstream 401 handling (key-fatal, not request-fatal) ─────────────────────
@@ -252,6 +297,9 @@ export function recordUpstreamSuccess(route: RouteResult, rateLimitTokens: numbe
recordRequest(route.platform, route.modelId, route.keyId);
recordTokens(route.platform, route.modelId, route.keyId, rateLimitTokens);
recordSuccess(route.modelDbId);
+ // A served request proves the model+key can complete: the empty-completion
+ // streak (#751) starts over.
+ emptyCompletionStreaks.delete(`${route.platform}:${route.modelId}:${route.keyId}`);
// A served request is the strongest possible evidence the key works, so clear
// any stale 'error' status left by an earlier transport blip instead of waiting
// for the next health pass to make the key routable again.
@@ -452,7 +500,7 @@ export function exhaustedRetryError(lastError: any, maxRetries?: number, ctx?: E
const attempts = ctx?.attempts ?? [];
const trail = attempts.length > 0 ? ` Attempt trail: ${formatAttemptTrail(attempts)}.` : '';
const budgetNote = ctx?.timedOut
- ? ` (stopped early: retry time budget ${Math.round((ctx.budgetMs ?? 0) / 1000)}s exceeded — the attempt in flight is never aborted mid-flight, the budget only stops STARTING further retries; raise FALLBACK_TIME_BUDGET_MS or the fallback_time_budget_ms setting to allow a longer failover chain)`
+ ? ` (stopped early: retry time budget ${Math.round((ctx.budgetMs ?? 0) / 1000)}s exceeded — the attempt in flight is never aborted mid-flight and one failover hop is always allowed, the budget only stops STARTING further retries; raise FALLBACK_TIME_BUDGET_MS or the fallback_time_budget_ms setting to allow a longer failover chain)`
: '';
const everyAttempt = (cls: AttemptErrorClass | ReadonlySet<AttemptErrorClass>): boolean =>
attempts.length > 0 && attempts.every(a => (cls instanceof Set ? cls.has(a.errorClass) : a.errorClass === cls));
@@ -748,7 +796,8 @@ export async function runFallbackLoop(hooks: FallbackHooks): Promise<void> {
// report back the row ids it writes without any surface changing. The flush
// runs after the loop returns — i.e. after the response is finished — so it
// never sits on the client's latency path, and it is a no-op when no
- // `requests` row was written (pure client aborts).
+ // `requests` row was written (a pure abort writes its own 'canceled' row,
+ // so its trace persists too).
const trace = newRequestTrace();
try {
await runWithRequestTrace(trace, () => runFallbackLoopAttempts(hooks, trace));
@@ -807,9 +856,12 @@ async function runFallbackLoopAttempts(hooks: FallbackHooks, trace: RequestTrace
}
// Wall-clock budget: refuse to START another retry once spent. The first
- // attempt always runs; a slow attempt is never aborted mid-flight (that is
- // the TODO(fallback-v2) hedging work), it just becomes the last one.
- if (attempt > 0 && budgetMs > 0 && Date.now() - startedAt >= budgetMs) {
+ // attempt always runs, and so does the first RETRY — when attempt 0 alone
+ // consumed the budget, refusing attempt 1 would make failover impossible
+ // for exactly the slow-failing models that need it (#751). A slow attempt
+ // is never aborted mid-flight (that is the TODO(fallback-v2) hedging
+ // work), it just becomes the last one.
+ if (attempt > 1 && budgetMs > 0 && Date.now() - startedAt >= budgetMs) {
hooks.onExhausted(
exhaustedRetryError(lastError, maxRetries, { attempts, timedOut: true, budgetMs }),
{ attempts, timedOut: true },
@@ -867,10 +919,15 @@ async function runFallbackLoopAttempts(hooks: FallbackHooks, trace: RequestTrace
// waiting, and there is no socket to render anything to. The finally
// below still frees the in-flight lease.
if (isClientAbortError(err)) {
+ const elapsedMs = Date.now() - startedAt;
console.log(`[FallbackLoop] client disconnected mid-attempt on ${route.platform}/${route.modelId} — upstream canceled, stopping without benching`);
- // Trace only: persisted iff an earlier failure already wrote a
- // `requests` row (the requests table records nothing for a pure abort,
- // and the trace stays consistent with that).
+ // Visibility row (#752): a pure abort used to leave NOTHING in the
+ // requests table, so the dashboard showed no trace of the request.
+ // Status 'canceled' is neither success nor failure — every stats/
+ // scoring query excludes it — and this row is also the parent the
+ // attempt-trace batch below keys to.
+ logRequest(route.platform, route.modelId, route.keyId, 'canceled', 0, 0, elapsedMs,
+ `client disconnected after ${(elapsedMs / 1000).toFixed(1)}s; upstream request canceled`);
traceAttempt('client_abort');
return;
}
@@ -886,17 +943,19 @@ async function runFallbackLoopAttempts(hooks: FallbackHooks, trace: RequestTrace
continue;
}
if (isRetryableError(err)) {
- recordRetryableFailure(route, err, hooks.state);
+ const exempt = recordRetryableFailure(route, err, hooks.state);
const errorClass = classifyAttemptError(err);
attempts.push({ platform: route.platform, modelId: route.modelId, keyOrdinal: keyOrdinal(route), errorClass });
traceAttempt(errorClass, err);
lastError = err;
- // skipBench failures (format ignored, hidden-reasoning truncation) are
- // model behavior, not provider health — recordRetryableFailure already
- // skips the cooldown/penalty for them, and they must not count toward
- // the "pool looks unhealthy" breaker either: three prose answers to a
- // json_schema request say nothing about whether candidate four is up.
- if (err?.skipBench !== true && stopIfBreakerTripped()) return;
+ // A still-exempt skipBench failure (format ignored, hidden-reasoning
+ // truncation under the streak limit) is model behavior, not provider
+ // health — recordRetryableFailure skipped the cooldown/penalty for it,
+ // and it must not count toward the "pool looks unhealthy" breaker
+ // either: three prose answers to a json_schema request say nothing
+ // about whether candidate four is up. Once the empty-completion streak
+ // lifts the exemption (#751), the failure counts everywhere.
+ if (!exempt && stopIfBreakerTripped()) return;
continue;
}
traceAttempt(classifyAttemptError(err), err);
diff --git a/server/src/lib/request-log.ts b/server/src/lib/request-log.ts
index cf8e44d2..e14af39c 100644
--- a/server/src/lib/request-log.ts
+++ b/server/src/lib/request-log.ts
@@ -34,6 +34,11 @@ function setSettingIfMissing(db: LogTx, key: string, value: string): void {
// is logged identically. Lives in a neutral lib module to avoid an import cycle
// between the fusion service and the proxy route that both call it.
//
+// Status is 'success', 'error', or 'canceled' (#752 — the client hung up
+// mid-attempt). A canceled request counts toward request totals — it happened —
+// but toward NEITHER success nor error: rates and scoring must read
+// success/(success+error), never success/total.
+//
// In addition to the raw row, we update two durable aggregates so analytics
// totals survive the raw-row prune (REQUEST_ANALYTICS_MAX_ROWS):
// - request_hourly: per-hour bucket counts and tokens (max window = 30d).
@@ -113,9 +118,10 @@ export function logRequest(
// mid-stream error row, or the last per-attempt failure row). Called once per
// request by the fallback loop AFTER the response is finished, so the write is
// off the client's latency path. Zero-failure single-attempt successes write
-// exactly one 'ok' row; a trace with no parent row (e.g. a client abort before
-// any attempt was logged) writes nothing — consistent with the `requests`
-// table, which records nothing for those either.
+// exactly one 'ok' row. A trace with no parent row writes nothing — since the
+// fallback loop logs a 'canceled' row for pure client aborts (#752), that is
+// now only the loop-top stop paths, whose failed attempts each wrote their own
+// row already.
export function persistRequestAttempts(trace: RequestTrace): void {
if (trace.records.length === 0 || trace.lastRequestRowId == null) return;
try {
diff --git a/server/src/routes/analytics.ts b/server/src/routes/analytics.ts
index 61fa5de8..178f9a23 100644
--- a/server/src/routes/analytics.ts
+++ b/server/src/routes/analytics.ts
@@ -42,6 +42,7 @@ function readAggregateSince(since: string) {
SELECT
COALESCE(SUM(total_requests), 0) as total_requests,
COALESCE(SUM(success_count), 0) as success_count,
+ COALESCE(SUM(error_count), 0) as error_count,
COALESCE(SUM(input_tokens), 0) as total_input_tokens,
COALESCE(SUM(output_tokens), 0) as total_output_tokens,
MIN(hour) as first_request_at
@@ -50,6 +51,7 @@ function readAggregateSince(since: string) {
`).get(aggregateSince) as {
total_requests: number;
success_count: number;
+ error_count: number;
total_input_tokens: number;
total_output_tokens: number;
first_request_at: string | null;
@@ -80,7 +82,11 @@ analyticsRouter.get('/summary', (req: Request, res: Response) => {
// is the source of truth for headline numbers.
const aggregate = readAggregateSince(since);
const totalRequests = aggregate.total_requests ?? 0;
- const successRate = totalRequests > 0 ? (aggregate.success_count / totalRequests) * 100 : 0;
+ // Success rate over success+error only: a 'canceled' request (#752 — client
+ // hung up) still counts in the totals but is neither a success nor a
+ // failure, so it must not dilute the rate.
+ const decidedRequests = (aggregate.success_count ?? 0) + (aggregate.error_count ?? 0);
+ const successRate = decidedRequests > 0 ? (aggregate.success_count / decidedRequests) * 100 : 0;
// Avg latency is only meaningful at the raw row level; the hourly bucket
// doesn't preserve it. Fall back to a 0/null when no recent raw rows exist.
@@ -201,7 +207,8 @@ analyticsRouter.get('/by-model', (req: Request, res: Response) => {
r.model_id,
m.display_name,
COUNT(*) as requests,
- SUM(CASE WHEN r.status = 'success' THEN 1 ELSE 0 END) * 100.0 / COUNT(*) as success_rate,
+ -- Rate over success+error only: 'canceled' (#752) is neither.
+ SUM(CASE WHEN r.status = 'success' THEN 1 ELSE 0 END) * 100.0 / NULLIF(SUM(CASE WHEN r.status <> 'canceled' THEN 1 ELSE 0 END), 0) as success_rate,
AVG(r.latency_ms) as avg_latency_ms,
SUM(r.input_tokens) as total_input_tokens,
SUM(r.output_tokens) as total_output_tokens,
@@ -222,7 +229,8 @@ analyticsRouter.get('/by-model', (req: Request, res: Response) => {
modelId: r.model_id,
displayName: r.display_name ?? r.model_id,
requests: r.requests,
- successRate: Math.round(r.success_rate * 10) / 10,
+ // success_rate is NULL when every row in the group was canceled.
+ successRate: Math.round((r.success_rate ?? 0) * 10) / 10,
avgLatencyMs: Math.round(r.avg_latency_ms),
totalInputTokens: r.total_input_tokens ?? 0,
totalOutputTokens: r.total_output_tokens ?? 0,
@@ -243,7 +251,7 @@ analyticsRouter.get('/by-platform', (req: Request, res: Response) => {
platform,
COUNT(*) as requests,
COUNT(latency_ms) as latency_count,
- SUM(CASE WHEN status = 'success' THEN 1 ELSE 0 END) * 100.0 / COUNT(*) as success_rate,
+ SUM(CASE WHEN status = 'success' THEN 1 ELSE 0 END) * 100.0 / NULLIF(SUM(CASE WHEN status <> 'canceled' THEN 1 ELSE 0 END), 0) as success_rate,
AVG(latency_ms) as avg_latency_ms,
AVG(ttfb_ms) as avg_ttfb_ms,
SUM(CASE WHEN status = 'error' THEN 1 ELSE 0 END) as error_count,
@@ -279,7 +287,7 @@ analyticsRouter.get('/by-platform', (req: Request, res: Response) => {
return {
platform: r.platform,
requests: r.requests,
- successRate: Math.round(r.success_rate * 10) / 10,
+ successRate: Math.round((r.success_rate ?? 0) * 10) / 10,
avgLatencyMs: Math.round(r.avg_latency_ms),
p95LatencyMs: p95Row ? Math.round(p95Row.latency_ms) : null,
avgTtfbMs: r.avg_ttfb_ms != null ? Math.round(r.avg_ttfb_ms) : null,
@@ -300,7 +308,7 @@ analyticsRouter.get('/by-client', (req: Request, res: Response) => {
SELECT
COALESCE(client_agent, 'unknown') AS client_agent,
COUNT(*) AS requests,
- SUM(CASE WHEN status = 'success' THEN 1 ELSE 0 END) * 100.0 / COUNT(*) AS success_rate,
+ SUM(CASE WHEN status = 'success' THEN 1 ELSE 0 END) * 100.0 / NULLIF(SUM(CASE WHEN status <> 'canceled' THEN 1 ELSE 0 END), 0) AS success_rate,
AVG(latency_ms) AS avg_latency_ms,
SUM(input_tokens) AS total_input_tokens,
SUM(output_tokens) AS total_output_tokens,
@@ -336,7 +344,7 @@ analyticsRouter.get('/by-key', (req: Request, res: Response) => {
k.label as label,
k.platform as platform,
COUNT(*) as requests,
- SUM(CASE WHEN r.status = 'success' THEN 1 ELSE 0 END) * 100.0 / COUNT(*) as success_rate,
+ SUM(CASE WHEN r.status = 'success' THEN 1 ELSE 0 END) * 100.0 / NULLIF(SUM(CASE WHEN r.status <> 'canceled' THEN 1 ELSE 0 END), 0) as success_rate,
AVG(r.latency_ms) as avg_latency_ms,
SUM(r.input_tokens) as total_input_tokens,
SUM(r.output_tokens) as total_output_tokens
@@ -355,7 +363,7 @@ analyticsRouter.get('/by-key', (req: Request, res: Response) => {
label: r.label ?? null,
platform: r.platform ?? null,
requests: r.requests,
- successRate: Math.round(r.success_rate * 10) / 10,
+ successRate: Math.round((r.success_rate ?? 0) * 10) / 10,
avgLatencyMs: Math.round(r.avg_latency_ms),
totalInputTokens: r.total_input_tokens ?? 0,
totalOutputTokens: r.total_output_tokens ?? 0,
@@ -500,8 +508,8 @@ analyticsRouter.get('/requests', (req: Request, res: Response) => {
// Optional filters. Both are validated (whitelist / shape) and applied as
// bound parameters; absent filters keep the default behavior identical.
const status = req.query.status as string | undefined;
- if (status !== undefined && status !== 'success' && status !== 'error') {
- res.status(400).json({ error: "invalid status filter (expected 'success' or 'error')" });
+ if (status !== undefined && status !== 'success' && status !== 'error' && status !== 'canceled') {
+ res.status(400).json({ error: "invalid status filter (expected 'success', 'error' or 'canceled')" });
return;
}
// Platform ids are short slugs ('groq', 'pt-custom_1'); anything else is a
diff --git a/server/src/routes/mcp.ts b/server/src/routes/mcp.ts
index 4803a80e..1a638dfd 100644
--- a/server/src/routes/mcp.ts
+++ b/server/src/routes/mcp.ts
@@ -137,10 +137,11 @@ function usageSummary(args: Record<string, unknown>): unknown {
const totals = db.prepare(`
SELECT COALESCE(SUM(total_requests), 0) AS requests,
COALESCE(SUM(success_count), 0) AS successes,
+ COALESCE(SUM(error_count), 0) AS errors,
COALESCE(SUM(input_tokens), 0) AS input_tokens,
COALESCE(SUM(output_tokens), 0) AS output_tokens
FROM request_hourly WHERE hour >= ?
- `).get(since.slice(0, 13) + ':00:00') as { requests: number; successes: number; input_tokens: number; output_tokens: number };
+ `).get(since.slice(0, 13) + ':00:00') as { requests: number; successes: number; errors: number; input_tokens: number; output_tokens: number };
const topModels = db.prepare(`
SELECT platform, model_id, COUNT(*) AS requests,
SUM(CASE WHEN status = 'success' THEN 1 ELSE 0 END) AS successes
@@ -150,7 +151,8 @@ function usageSummary(args: Record<string, unknown>): unknown {
return {
range,
requests: totals.requests,
- success_rate: totals.requests > 0 ? Math.round((totals.successes / totals.requests) * 1000) / 10 : null,
+ // Over success+error only: 'canceled' rows (#752) are neither.
+ success_rate: totals.successes + totals.errors > 0 ? Math.round((totals.successes / (totals.successes + totals.errors)) * 1000) / 10 : null,
input_tokens: totals.input_tokens,
output_tokens: totals.output_tokens,
top_models: topModels,
diff --git a/server/src/services/router.ts b/server/src/services/router.ts
index 7ff1c911..a6cea647 100644
--- a/server/src/services/router.ts
+++ b/server/src/services/router.ts
@@ -529,12 +529,14 @@ function decayWeight(ageDays: number): number {
return Math.pow(0.5, Math.max(0, ageDays) / HALF_LIFE_DAYS);
}
-// SQL predicate for "this row is a timed-out request" (#619). `requests.status`
-// only ever holds 'success' or 'error' — a timeout is an error row whose text
-// carries one of the shared timeout markers (lib/error-classify.ts), which is
-// also what the failover attempt trail classifies on. The markers are
-// hard-coded lowercase identifiers from our own source, never user input, so
-// interpolating them into the LIKE list is safe.
+// SQL predicate for "this row is a timed-out request" (#619). A timeout is an
+// error row whose text carries one of the shared timeout markers
+// (lib/error-classify.ts), which is also what the failover attempt trail
+// classifies on. 'canceled' rows (#752 — client hung up) never reach this
+// predicate: the stats query below filters them out entirely, because a
+// vanished client says nothing about the model's reliability or speed. The
+// markers are hard-coded lowercase identifiers from our own source, never
+// user input, so interpolating them into the LIKE list is safe.
const IS_TIMEOUT_SQL = `(status != 'success' AND (${
TIMEOUT_ERROR_MARKERS.map(m => `LOWER(COALESCE(error, '')) LIKE '%${m}%'`).join(' OR ')
}))`;
@@ -570,7 +572,7 @@ export function refreshStatsCache(db: Db, force = false): void {
SUM(CASE WHEN ${IS_TIMEOUT_SQL} THEN 1 ELSE 0 END) AS timeouts,
SUM(CASE WHEN ${IS_TIMEOUT_SQL} THEN MIN(MAX(latency_ms, 0), ${TIMEOUT_LATENCY_CAP_MS}) ELSE 0 END) AS timeout_lat
FROM requests
- WHERE created_at >= ?
+ WHERE created_at >= ? AND status <> 'canceled'
GROUP BY platform, model_id, key_id, age_days
`).all(since) as Array<{
platform: string; model_id: string; key_id: number | null; age_days: number; total: number; successes: number;