@hyperspaceng/neural-ai 0.65.1 → 0.66.2

package/dist/providers/google.js

@@ -224,12 +224,12 @@ export const streamSimpleGoogle = (model, context, options) => {
     }
     const effort = clampReasoning(options.reasoning);
     const googleModel = model;
-    if (isGemini3ProModel(googleModel) || isGemini3FlashModel(googleModel)) {
+    if (isGemini3ProModel(googleModel) || isGemini3FlashModel(googleModel) || isGemma4Model(googleModel)) {
         return streamGoogle(model, context, {
             ...base,
             thinking: {
                 enabled: true,
-                level: getGemini3ThinkingLevel(effort, googleModel),
+                level: getThinkingLevel(effort, googleModel),
             },
         });
     }
@@ -306,6 +306,9 @@ function buildParams(model, context, options = {}) {
     };
     return params;
 }
+function isGemma4Model(model) {
+    return /gemma-?4/.test(model.id.toLowerCase());
+}
 function isGemini3ProModel(model) {
     return /gemini-3(?:\.\d+)?-pro/.test(model.id.toLowerCase());
 }
@@ -322,10 +325,13 @@ function getDisabledThinkingConfig(model) {
     if (isGemini3FlashModel(model)) {
         return { thinkingLevel: "MINIMAL" };
     }
+    if (isGemma4Model(model)) {
+        return { thinkingLevel: "MINIMAL" };
+    }
     // Gemini 2.x supports disabling via thinkingBudget = 0.
     return { thinkingBudget: 0 };
 }
-function getGemini3ThinkingLevel(effort, model) {
+function getThinkingLevel(effort, model) {
     if (isGemini3ProModel(model)) {
         switch (effort) {
             case "minimal":
@@ -336,6 +342,16 @@ function getGemini3ThinkingLevel(effort, model) {
                 return "HIGH";
         }
     }
+    if (isGemma4Model(model)) {
+        switch (effort) {
+            case "minimal":
+            case "low":
+                return "MINIMAL";
+            case "medium":
+            case "high":
+                return "HIGH";
+        }
+    }
     switch (effort) {
         case "minimal":
             return "MINIMAL";
@@ -360,6 +376,15 @@ function getGoogleBudget(model, effort, customBudgets) {
         };
         return budgets[effort];
     }
+    if (model.id.includes("2.5-flash-lite")) {
+        const budgets = {
+            minimal: 512,
+            low: 2048,
+            medium: 8192,
+            high: 24576,
+        };
+        return budgets[effort];
+    }
     if (model.id.includes("2.5-flash")) {
         const budgets = {
             minimal: 128,

package/dist/providers/google.js.map

@@ -1 +1 @@
-{"version":3,"file":"google.js","sourceRoot":"","sources":["../../src/providers/google.ts"],"names":[],"mappings":"AAAA,OAAO,EAGN,WAAW,GAEX,MAAM,eAAe,CAAC;AACvB,OAAO,EAAE,YAAY,EAAE,MAAM,oBAAoB,CAAC;AAClD,OAAO,EAAE,aAAa,EAAE,MAAM,cAAc,CAAC;AAe7C,OAAO,EAAE,2BAA2B,EAAE,MAAM,0BAA0B,CAAC;AACvE,OAAO,EAAE,kBAAkB,EAAE,MAAM,8BAA8B,CAAC;AAElE,OAAO,EACN,eAAe,EACf,YAAY,EACZ,cAAc,EACd,aAAa,EACb,aAAa,EACb,sBAAsB,GACtB,MAAM,oBAAoB,CAAC;AAC5B,OAAO,EAAE,gBAAgB,EAAE,cAAc,EAAE,MAAM,qBAAqB,CAAC;AAWvE,8CAA8C;AAC9C,IAAI,eAAe,GAAG,CAAC,CAAC;AAExB,MAAM,CAAC,MAAM,YAAY,GAA0D,CAClF,KAAoC,EACpC,OAAgB,EAChB,OAAuB,EACO,EAAE,CAAC;IACjC,MAAM,MAAM,GAAG,IAAI,2BAA2B,EAAE,CAAC;IAEjD,CAAC,KAAK,IAAI,EAAE,CAAC;QACZ,MAAM,MAAM,GAAqB;YAChC,IAAI,EAAE,WAAW;YACjB,OAAO,EAAE,EAAE;YACX,GAAG,EAAE,sBAA6B;YAClC,QAAQ,EAAE,KAAK,CAAC,QAAQ;YACxB,KAAK,EAAE,KAAK,CAAC,EAAE;YACf,KAAK,EAAE;gBACN,KAAK,EAAE,CAAC;gBACR,MAAM,EAAE,CAAC;gBACT,SAAS,EAAE,CAAC;gBACZ,UAAU,EAAE,CAAC;gBACb,WAAW,EAAE,CAAC;gBACd,IAAI,EAAE,EAAE,KAAK,EAAE,CAAC,EAAE,MAAM,EAAE,CAAC,EAAE,SAAS,EAAE,CAAC,EAAE,UAAU,EAAE,CAAC,EAAE,KAAK,EAAE,CAAC,EAAE;aACpE;YACD,UAAU,EAAE,MAAM;YAClB,SAAS,EAAE,IAAI,CAAC,GAAG,EAAE;SACrB,CAAC;QAEF,IAAI,CAAC;YACJ,MAAM,MAAM,GAAG,OAAO,EAAE,MAAM,IAAI,YAAY,CAAC,KAAK,CAAC,QAAQ,CAAC,IAAI,EAAE,CAAC;YACrE,MAAM,MAAM,GAAG,YAAY,CAAC,KAAK,EAAE,MAAM,EAAE,OAAO,EAAE,OAAO,CAAC,CAAC;YAC7D,IAAI,MAAM,GAAG,WAAW,CAAC,KAAK,EAAE,OAAO,EAAE,OAAO,CAAC,CAAC;YAClD,MAAM,UAAU,GAAG,MAAM,OAAO,EAAE,SAAS,EAAE,CAAC,MAAM,EAAE,KAAK,CAAC,CAAC;YAC7D,IAAI,UAAU,KAAK,SAAS,EAAE,CAAC;gBAC9B,MAAM,GAAG,UAAuC,CAAC;YAClD,CAAC;YACD,MAAM,YAAY,GAAG,MAAM,MAAM,CAAC,MAAM,CAAC,qBAAqB,CAAC,MAAM,CAAC,CAAC;YAEvE,MAAM,CAAC,IAAI,CAAC,EAAE,IAAI,EAAE,OAAO,EAAE,OAAO,EAAE,MAAM,EAAE,CAAC,CAAC;YAChD,IAAI,YAAY,GAAyC,IAAI,CAAC;YAC9D,MAAM,MAAM,GAAG,MAAM,CAAC,OAAO,CAAC;YAC9B,MAAM,UAAU,GAAG,GAAG,EAAE,CAAC,MAAM,CAAC,MAAM,GAAG,CAAC,CAAC;YAC3C,IAAI,KAAK,EAAE,MAAM,KAAK,IAAI,YAAY,EAAE,CAAC;gBACxC,qFAAqF;gBACrF,gFAAgF;gBAChF,MAAM,CAAC,UAAU,KAAK,KAAK,CAAC,UAAU,CAAC;gBACvC,MAAM,SAAS,GAAG,KAAK,CAAC,UAAU,EAAE,CAAC,CAAC,CAAC,CAAC;gBACxC,IAAI,SAAS,EAAE,OAAO,EAAE,KAAK,EAAE,CAAC;oBAC/B,KAAK,MAAM,IAAI,IAAI,SAAS,CAAC,OAAO,CAAC,KAAK,EAAE,CAAC;wBAC5C,IAAI,IAAI,CAAC,IAAI,KAAK,SAAS,EAAE,CAAC;4BAC7B,MAAM,UAAU,GAAG,cAAc,CAAC,IAAI,CAAC,CAAC;4BACxC,IACC,CAAC,YAAY;gCACb,CAAC,UAAU,IAAI,YAAY,CAAC,IAAI,KAAK,UAAU,CAAC;gCAChD,CAAC,CAAC,UAAU,IAAI,YAAY,CAAC,IAAI,KAAK,MAAM,CAAC,EAC5C,CAAC;gCACF,IAAI,YAAY,EAAE,CAAC;oCAClB,IAAI,YAAY,CAAC,IAAI,KAAK,MAAM,EAAE,CAAC;wCAClC,MAAM,CAAC,IAAI,CAAC;4CACX,IAAI,EAAE,UAAU;4CAChB,YAAY,EAAE,MAAM,CAAC,MAAM,GAAG,CAAC;4CAC/B,OAAO,EAAE,YAAY,CAAC,IAAI;4CAC1B,OAAO,EAAE,MAAM;yCACf,CAAC,CAAC;oCACJ,CAAC;yCAAM,CAAC;wCACP,MAAM,CAAC,IAAI,CAAC;4CACX,IAAI,EAAE,cAAc;4CACpB,YAAY,EAAE,UAAU,EAAE;4CAC1B,OAAO,EAAE,YAAY,CAAC,QAAQ;4CAC9B,OAAO,EAAE,MAAM;yCACf,CAAC,CAAC;oCACJ,CAAC;gCACF,CAAC;gCACD,IAAI,UAAU,EAAE,CAAC;oCAChB,YAAY,GAAG,EAAE,IAAI,EAAE,UAAU,EAAE,QAAQ,EAAE,EAAE,EAAE,iBAAiB,EAAE,SAAS,EAAE,CAAC;oCAChF,MAAM,CAAC,OAAO,CAAC,IAAI,CAAC,YAAY,CAAC,CAAC;oCAClC,MAAM,CAAC,IAAI,CAAC,EAAE,IAAI,EAAE,gBAAgB,EAAE,YAAY,EAAE,UAAU,EAAE,EAAE,OAAO,EAAE,MAAM,EAAE,CAAC,CAAC;gCACtF,CAAC;qCAAM,CAAC;oCACP,YAAY,GAAG,EAAE,IAAI,EAAE,MAAM,EAAE,IAAI,EAAE,EAAE,EAAE,CAAC;oCAC1C,MAAM,CAAC,OAAO,CAAC,IAAI,CAAC,YAAY,CAAC,CAAC;oCAClC,MAAM,CAAC,IAAI,CAAC,EAAE,IAAI,EAAE,YAAY,EAAE,YAAY,EAAE,UAAU,EAAE,EAAE,OAAO,EAAE,MAAM,EAAE,CAAC,CAAC;gCAClF,CAAC;4BACF,CAAC;4BACD,IAAI,YAAY,CAAC,IAAI,KAAK,UAAU,EAAE,CAAC;gCACtC,YAAY,CAAC,QAAQ,IAAI,IAAI,CAAC,IAAI,CAAC;gCACnC,YAAY,CAAC,iBAAiB,GAAG,sBAAsB,CACtD,YAAY,CAAC,iBAAiB,EAC9B,IAAI,CAAC,gBAAgB,CACrB,CAAC;gCACF,MAAM,CAAC,IAAI,CAAC;oCACX,IAAI,EAAE,gBAAgB;oCACtB,YAAY,EAAE,UAAU,EAAE;oCAC1B,KAAK,EAAE,IAAI,CAAC,IAAI;oCAChB,OAAO,EAAE,MAAM;iCACf,CAAC,CAAC;4BACJ,CAAC;iCAAM,CAAC;gCACP,YAAY,CAAC,IAAI,IAAI,IAAI,CAAC,IAAI,CAAC;gCAC/B,YAAY,CAAC,aAAa,GAAG,sBAAsB,CAClD,YAAY,CAAC,aAAa,EAC1B,IAAI,CAAC,gBAAgB,CACrB,CAAC;gCACF,MAAM,CAAC,IAAI,CAAC;oCACX,IAAI,EAAE,YAAY;oCAClB,YAAY,EAAE,UAAU,EAAE;oCAC1B,KAAK,EAAE,IAAI,CAAC,IAAI;oCAChB,OAAO,EAAE,MAAM;iCACf,CAAC,CAAC;4BACJ,CAAC;wBACF,CAAC;wBAED,IAAI,IAAI,CAAC,YAAY,EAAE,CAAC;4BACvB,IAAI,YAAY,EAAE,CAAC;gCAClB,IAAI,YAAY,CAAC,IAAI,KAAK,MAAM,EAAE,CAAC;oCAClC,MAAM,CAAC,IAAI,CAAC;wCACX,IAAI,EAAE,UAAU;wCAChB,YAAY,EAAE,UAAU,EAAE;wCAC1B,OAAO,EAAE,YAAY,CAAC,IAAI;wCAC1B,OAAO,EAAE,MAAM;qCACf,CAAC,CAAC;gCACJ,CAAC;qCAAM,CAAC;oCACP,MAAM,CAAC,IAAI,CAAC;wCACX,IAAI,EAAE,cAAc;wCACpB,YAAY,EAAE,UAAU,EAAE;wCAC1B,OAAO,EAAE,YAAY,CAAC,QAAQ;wCAC9B,OAAO,EAAE,MAAM;qCACf,CAAC,CAAC;gCACJ,CAAC;gCACD,YAAY,GAAG,IAAI,CAAC;4BACrB,CAAC;4BAED,4DAA4D;4BAC5D,MAAM,UAAU,GAAG,IAAI,CAAC,YAAY,CAAC,EAAE,CAAC;4BACxC,MAAM,UAAU,GACf,CAAC,UAAU,IAAI,MAAM,CAAC,OAAO,CAAC,IAAI,CAAC,CAAC,CAAC,EAAE,EAAE,CAAC,CAAC,CAAC,IAAI,KAAK,UAAU,IAAI,CAAC,CAAC,EAAE,KAAK,UAAU,CAAC,CAAC;4BACzF,MAAM,UAAU,GAAG,UAAU;gCAC5B,CAAC,CAAC,GAAG,IAAI,CAAC,YAAY,CAAC,IAAI,IAAI,IAAI,CAAC,GAAG,EAAE,IAAI,EAAE,eAAe,EAAE;gCAChE,CAAC,CAAC,UAAU,CAAC;4BAEd,MAAM,QAAQ,GAAa;gCAC1B,IAAI,EAAE,UAAU;gCAChB,EAAE,EAAE,UAAU;gCACd,IAAI,EAAE,IAAI,CAAC,YAAY,CAAC,IAAI,IAAI,EAAE;gCAClC,SAAS,EAAG,IAAI,CAAC,YAAY,CAAC,IAA4B,IAAI,EAAE;gCAChE,GAAG,CAAC,IAAI,CAAC,gBAAgB,IAAI,EAAE,gBAAgB,EAAE,IAAI,CAAC,gBAAgB,EAAE,CAAC;6BACzE,CAAC;4BAEF,MAAM,CAAC,OAAO,CAAC,IAAI,CAAC,QAAQ,CAAC,CAAC;4BAC9B,MAAM,CAAC,IAAI,CAAC,EAAE,IAAI,EAAE,gBAAgB,EAAE,YAAY,EAAE,UAAU,EAAE,EAAE,OAAO,EAAE,MAAM,EAAE,CAAC,CAAC;4BACrF,MAAM,CAAC,IAAI,CAAC;gCACX,IAAI,EAAE,gBAAgB;gCACtB,YAAY,EAAE,UAAU,EAAE;gCAC1B,KAAK,EAAE,IAAI,CAAC,SAAS,CAAC,QAAQ,CAAC,SAAS,CAAC;gCACzC,OAAO,EAAE,MAAM;6BACf,CAAC,CAAC;4BACH,MAAM,CAAC,IAAI,CAAC,EAAE,IAAI,EAAE,cAAc,EAAE,YAAY,EAAE,UAAU,EAAE,EAAE,QAAQ,EAAE,OAAO,EAAE,MAAM,EAAE,CAAC,CAAC;wBAC9F,CAAC;oBACF,CAAC;gBACF,CAAC;gBAED,IAAI,SAAS,EAAE,YAAY,EAAE,CAAC;oBAC7B,MAAM,CAAC,UAAU,GAAG,aAAa,CAAC,SAAS,CAAC,YAAY,CAAC,CAAC;oBAC1D,IAAI,MAAM,CAAC,OAAO,CAAC,IAAI,CAAC,CAAC,CAAC,EAAE,EAAE,CAAC,CAAC,CAAC,IAAI,KAAK,UAAU,CAAC,EAAE,CAAC;wBACvD,MAAM,CAAC,UAAU,GAAG,SAAS,CAAC;oBAC/B,CAAC;gBACF,CAAC;gBAED,IAAI,KAAK,CAAC,aAAa,EAAE,CAAC;oBACzB,MAAM,CAAC,KAAK,GAAG;wBACd,KAAK,EACJ,CAAC,KAAK,CAAC,aAAa,CAAC,gBAAgB,IAAI,CAAC,CAAC,GAAG,CAAC,KAAK,CAAC,aAAa,CAAC,uBAAuB,IAAI,CAAC,CAAC;wBACjG,MAAM,EACL,CAAC,KAAK,CAAC,aAAa,CAAC,oBAAoB,IAAI,CAAC,CAAC,GAAG,CAAC,KAAK,CAAC,aAAa,CAAC,kBAAkB,IAAI,CAAC,CAAC;wBAChG,SAAS,EAAE,KAAK,CAAC,aAAa,CAAC,uBAAuB,IAAI,CAAC;wBAC3D,UAAU,EAAE,CAAC;wBACb,WAAW,EAAE,KAAK,CAAC,aAAa,CAAC,eAAe,IAAI,CAAC;wBACrD,IAAI,EAAE;4BACL,KAAK,EAAE,CAAC;4BACR,MAAM,EAAE,CAAC;4BACT,SAAS,EAAE,CAAC;4BACZ,UAAU,EAAE,CAAC;4BACb,KAAK,EAAE,CAAC;yBACR;qBACD,CAAC;oBACF,aAAa,CAAC,KAAK,EAAE,MAAM,CAAC,KAAK,CAAC,CAAC;gBACpC,CAAC;YACF,CAAC;YAED,IAAI,YAAY,EAAE,CAAC;gBAClB,IAAI,YAAY,CAAC,IAAI,KAAK,MAAM,EAAE,CAAC;oBAClC,MAAM,CAAC,IAAI,CAAC;wBACX,IAAI,EAAE,UAAU;wBAChB,YAAY,EAAE,UAAU,EAAE;wBAC1B,OAAO,EAAE,YAAY,CAAC,IAAI;wBAC1B,OAAO,EAAE,MAAM;qBACf,CAAC,CAAC;gBACJ,CAAC;qBAAM,CAAC;oBACP,MAAM,CAAC,IAAI,CAAC;wBACX,IAAI,EAAE,cAAc;wBACpB,YAAY,EAAE,UAAU,EAAE;wBAC1B,OAAO,EAAE,YAAY,CAAC,QAAQ;wBAC9B,OAAO,EAAE,MAAM;qBACf,CAAC,CAAC;gBACJ,CAAC;YACF,CAAC;YAED,IAAI,OAAO,EAAE,MAAM,EAAE,OAAO,EAAE,CAAC;gBAC9B,MAAM,IAAI,KAAK,CAAC,qBAAqB,CAAC,CAAC;YACxC,CAAC;YAED,IAAI,MAAM,CAAC,UAAU,KAAK,SAAS,IAAI,MAAM,CAAC,UAAU,KAAK,OAAO,EAAE,CAAC;gBACtE,MAAM,IAAI,KAAK,CAAC,2BAA2B,CAAC,CAAC;YAC9C,CAAC;YAED,MAAM,CAAC,IAAI,CAAC,EAAE,IAAI,EAAE,MAAM,EAAE,MAAM,EAAE,MAAM,CAAC,UAAU,EAAE,OAAO,EAAE,MAAM,EAAE,CAAC,CAAC;YAC1E,MAAM,CAAC,GAAG,EAAE,CAAC;QACd,CAAC;QAAC,OAAO,KAAK,EAAE,CAAC;YAChB,uDAAuD;YACvD,KAAK,MAAM,KAAK,IAAI,MAAM,CAAC,OAAO,EAAE,CAAC;gBACpC,IAAI,OAAO,IAAI,KAAK,EAAE,CAAC;oBACtB,OAAQ,KAA4B,CAAC,KAAK,CAAC;gBAC5C,CAAC;YACF,CAAC;YACD,MAAM,CAAC,UAAU,GAAG,OAAO,EAAE,MAAM,EAAE,OAAO,CAAC,CAAC,CAAC,SAAS,CAAC,CAAC,CAAC,OAAO,CAAC;YACnE,MAAM,CAAC,YAAY,GAAG,KAAK,YAAY,KAAK,CAAC,CAAC,CAAC,KAAK,CAAC,OAAO,CAAC,CAAC,CAAC,IAAI,CAAC,SAAS,CAAC,KAAK,CAAC,CAAC;YACrF,MAAM,CAAC,IAAI,CAAC,EAAE,IAAI,EAAE,OAAO,EAAE,MAAM,EAAE,MAAM,CAAC,UAAU,EAAE,KAAK,EAAE,MAAM,EAAE,CAAC,CAAC;YACzE,MAAM,CAAC,GAAG,EAAE,CAAC;QACd,CAAC;IAAA,CACD,CAAC,EAAE,CAAC;IAEL,OAAO,MAAM,CAAC;AAAA,CACd,CAAC;AAEF,MAAM,CAAC,MAAM,kBAAkB,GAAgE,CAC9F,KAAoC,EACpC,OAAgB,EAChB,OAA6B,EACC,EAAE,CAAC;IACjC,MAAM,MAAM,GAAG,OAAO,EAAE,MAAM,IAAI,YAAY,CAAC,KAAK,CAAC,QAAQ,CAAC,CAAC;IAC/D,IAAI,CAAC,MAAM,EAAE,CAAC;QACb,MAAM,IAAI,KAAK,CAAC,4BAA4B,KAAK,CAAC,QAAQ,EAAE,CAAC,CAAC;IAC/D,CAAC;IAED,MAAM,IAAI,GAAG,gBAAgB,CAAC,KAAK,EAAE,OAAO,EAAE,MAAM,CAAC,CAAC;IACtD,IAAI,CAAC,OAAO,EAAE,SAAS,EAAE,CAAC;QACzB,OAAO,YAAY,CAAC,KAAK,EAAE,OAAO,EAAE,EAAE,GAAG,IAAI,EAAE,QAAQ,EAAE,EAAE,OAAO,EAAE,KAAK,EAAE,EAA0B,CAAC,CAAC;IACxG,CAAC;IAED,MAAM,MAAM,GAAG,cAAc,CAAC,OAAO,CAAC,SAAS,CAAE,CAAC;IAClD,MAAM,WAAW,GAAG,KAAsC,CAAC;IAE3D,IAAI,iBAAiB,CAAC,WAAW,CAAC,IAAI,mBAAmB,CAAC,WAAW,CAAC,EAAE,CAAC;QACxE,OAAO,YAAY,CAAC,KAAK,EAAE,OAAO,EAAE;YACnC,GAAG,IAAI;YACP,QAAQ,EAAE;gBACT,OAAO,EAAE,IAAI;gBACb,KAAK,EAAE,uBAAuB,CAAC,MAAM,EAAE,WAAW,CAAC;aACnD;SACuB,CAAC,CAAC;IAC5B,CAAC;IAED,OAAO,YAAY,CAAC,KAAK,EAAE,OAAO,EAAE;QACnC,GAAG,IAAI;QACP,QAAQ,EAAE;YACT,OAAO,EAAE,IAAI;YACb,YAAY,EAAE,eAAe,CAAC,WAAW,EAAE,MAAM,EAAE,OAAO,CAAC,eAAe,CAAC;SAC3E;KACuB,CAAC,CAAC;AAAA,CAC3B,CAAC;AAEF,SAAS,YAAY,CACpB,KAAoC,EACpC,MAAe,EACf,cAAuC,EACzB;IACd,MAAM,WAAW,GAAgF,EAAE,CAAC;IACpG,IAAI,KAAK,CAAC,OAAO,EAAE,CAAC;QACnB,WAAW,CAAC,OAAO,GAAG,KAAK,CAAC,OAAO,CAAC;QACpC,WAAW,CAAC,UAAU,GAAG,EAAE,CAAC,CAAC,sDAAsD;IACpF,CAAC;IACD,IAAI,KAAK,CAAC,OAAO,IAAI,cAAc,EAAE,CAAC;QACrC,WAAW,CAAC,OAAO,GAAG,EAAE,GAAG,KAAK,CAAC,OAAO,EAAE,GAAG,cAAc,EAAE,CAAC;IAC/D,CAAC;IAED,OAAO,IAAI,WAAW,CAAC;QACtB,MAAM;QACN,WAAW,EAAE,MAAM,CAAC,IAAI,CAAC,WAAW,CAAC,CAAC,MAAM,GAAG,CAAC,CAAC,CAAC,CAAC,WAAW,CAAC,CAAC,CAAC,SAAS;KAC1E,CAAC,CAAC;AAAA,CACH;AAED,SAAS,WAAW,CACnB,KAAoC,EACpC,OAAgB,EAChB,OAAO,GAAkB,EAAE,EACC;IAC5B,MAAM,QAAQ,GAAG,eAAe,CAAC,KAAK,EAAE,OAAO,CAAC,CAAC;IAEjD,MAAM,gBAAgB,GAA0B,EAAE,CAAC;IACnD,IAAI,OAAO,CAAC,WAAW,KAAK,SAAS,EAAE,CAAC;QACvC,gBAAgB,CAAC,WAAW,GAAG,OAAO,CAAC,WAAW,CAAC;IACpD,CAAC;IACD,IAAI,OAAO,CAAC,SAAS,KAAK,SAAS,EAAE,CAAC;QACrC,gBAAgB,CAAC,eAAe,GAAG,OAAO,CAAC,SAAS,CAAC;IACtD,CAAC;IAED,MAAM,MAAM,GAA0B;QACrC,GAAG,CAAC,MAAM,CAAC,IAAI,CAAC,gBAAgB,CAAC,CAAC,MAAM,GAAG,CAAC,IAAI,gBAAgB,CAAC;QACjE,GAAG,CAAC,OAAO,CAAC,YAAY,IAAI,EAAE,iBAAiB,EAAE,kBAAkB,CAAC,OAAO,CAAC,YAAY,CAAC,EAAE,CAAC;QAC5F,GAAG,CAAC,OAAO,CAAC,KAAK,IAAI,OAAO,CAAC,KAAK,CAAC,MAAM,GAAG,CAAC,IAAI,EAAE,KAAK,EAAE,YAAY,CAAC,OAAO,CAAC,KAAK,CAAC,EAAE,CAAC;KACxF,CAAC;IAEF,IAAI,OAAO,CAAC,KAAK,IAAI,OAAO,CAAC,KAAK,CAAC,MAAM,GAAG,CAAC,IAAI,OAAO,CAAC,UAAU,EAAE,CAAC;QACrE,MAAM,CAAC,UAAU,GAAG;YACnB,qBAAqB,EAAE;gBACtB,IAAI,EAAE,aAAa,CAAC,OAAO,CAAC,UAAU,CAAC;aACvC;SACD,CAAC;IACH,CAAC;SAAM,CAAC;QACP,MAAM,CAAC,UAAU,GAAG,SAAS,CAAC;IAC/B,CAAC;IAED,IAAI,OAAO,CAAC,QAAQ,EAAE,OAAO,IAAI,KAAK,CAAC,SAAS,EAAE,CAAC;QAClD,MAAM,cAAc,GAAmB,EAAE,eAAe,EAAE,IAAI,EAAE,CAAC;QACjE,IAAI,OAAO,CAAC,QAAQ,CAAC,KAAK,KAAK,SAAS,EAAE,CAAC;YAC1C,uFAAuF;YACvF,cAAc,CAAC,aAAa,GAAG,OAAO,CAAC,QAAQ,CAAC,KAAY,CAAC;QAC9D,CAAC;aAAM,IAAI,OAAO,CAAC,QAAQ,CAAC,YAAY,KAAK,SAAS,EAAE,CAAC;YACxD,cAAc,CAAC,cAAc,GAAG,OAAO,CAAC,QAAQ,CAAC,YAAY,CAAC;QAC/D,CAAC;QACD,MAAM,CAAC,cAAc,GAAG,cAAc,CAAC;IACxC,CAAC;SAAM,IAAI,KAAK,CAAC,SAAS,IAAI,OAAO,CAAC,QAAQ,IAAI,CAAC,OAAO,CAAC,QAAQ,CAAC,OAAO,EAAE,CAAC;QAC7E,MAAM,CAAC,cAAc,GAAG,yBAAyB,CAAC,KAAK,CAAC,CAAC;IAC1D,CAAC;IAED,IAAI,OAAO,CAAC,MAAM,EAAE,CAAC;QACpB,IAAI,OAAO,CAAC,MAAM,CAAC,OAAO,EAAE,CAAC;YAC5B,MAAM,IAAI,KAAK,CAAC,iBAAiB,CAAC,CAAC;QACpC,CAAC;QACD,MAAM,CAAC,WAAW,GAAG,OAAO,CAAC,MAAM,CAAC;IACrC,CAAC;IAED,MAAM,MAAM,GAA8B;QACzC,KAAK,EAAE,KAAK,CAAC,EAAE;QACf,QAAQ;QACR,MAAM;KACN,CAAC;IAEF,OAAO,MAAM,CAAC;AAAA,CACd;AAID,SAAS,iBAAiB,CAAC,KAAoC,EAAW;IACzE,OAAO,wBAAwB,CAAC,IAAI,CAAC,KAAK,CAAC,EAAE,CAAC,WAAW,EAAE,CAAC,CAAC;AAAA,CAC7D;AAED,SAAS,mBAAmB,CAAC,KAAoC,EAAW;IAC3E,OAAO,0BAA0B,CAAC,IAAI,CAAC,KAAK,CAAC,EAAE,CAAC,WAAW,EAAE,CAAC,CAAC;AAAA,CAC/D;AAED,SAAS,yBAAyB,CAAC,KAAoC,EAAkB;IACxF,uFAAuF;IACvF,yFAAyF;IACzF,oFAAoF;IACpF,IAAI,iBAAiB,CAAC,KAAK,CAAC,EAAE,CAAC;QAC9B,OAAO,EAAE,aAAa,EAAE,KAAY,EAAE,CAAC;IACxC,CAAC;IACD,IAAI,mBAAmB,CAAC,KAAK,CAAC,EAAE,CAAC;QAChC,OAAO,EAAE,aAAa,EAAE,SAAgB,EAAE,CAAC;IAC5C,CAAC;IAED,wDAAwD;IACxD,OAAO,EAAE,cAAc,EAAE,CAAC,EAAE,CAAC;AAAA,CAC7B;AAED,SAAS,uBAAuB,CAC/B,MAA4B,EAC5B,KAAoC,EACd;IACtB,IAAI,iBAAiB,CAAC,KAAK,CAAC,EAAE,CAAC;QAC9B,QAAQ,MAAM,EAAE,CAAC;YAChB,KAAK,SAAS,CAAC;YACf,KAAK,KAAK;gBACT,OAAO,KAAK,CAAC;YACd,KAAK,QAAQ,CAAC;YACd,KAAK,MAAM;gBACV,OAAO,MAAM,CAAC;QAChB,CAAC;IACF,CAAC;IACD,QAAQ,MAAM,EAAE,CAAC;QAChB,KAAK,SAAS;YACb,OAAO,SAAS,CAAC;QAClB,KAAK,KAAK;YACT,OAAO,KAAK,CAAC;QACd,KAAK,QAAQ;YACZ,OAAO,QAAQ,CAAC;QACjB,KAAK,MAAM;YACV,OAAO,MAAM,CAAC;IAChB,CAAC;AAAA,CACD;AAED,SAAS,eAAe,CACvB,KAAoC,EACpC,MAA4B,EAC5B,aAA+B,EACtB;IACT,IAAI,aAAa,EAAE,CAAC,MAAM,CAAC,KAAK,SAAS,EAAE,CAAC;QAC3C,OAAO,aAAa,CAAC,MAAM,CAAE,CAAC;IAC/B,CAAC;IAED,IAAI,KAAK,CAAC,EAAE,CAAC,QAAQ,CAAC,SAAS,CAAC,EAAE,CAAC;QAClC,MAAM,OAAO,GAAyC;YACrD,OAAO,EAAE,GAAG;YACZ,GAAG,EAAE,IAAI;YACT,MAAM,EAAE,IAAI;YACZ,IAAI,EAAE,KAAK;SACX,CAAC;QACF,OAAO,OAAO,CAAC,MAAM,CAAC,CAAC;IACxB,CAAC;IAED,IAAI,KAAK,CAAC,EAAE,CAAC,QAAQ,CAAC,WAAW,CAAC,EAAE,CAAC;QACpC,MAAM,OAAO,GAAyC;YACrD,OAAO,EAAE,GAAG;YACZ,GAAG,EAAE,IAAI;YACT,MAAM,EAAE,IAAI;YACZ,IAAI,EAAE,KAAK;SACX,CAAC;QACF,OAAO,OAAO,CAAC,MAAM,CAAC,CAAC;IACxB,CAAC;IAED,OAAO,CAAC,CAAC,CAAC;AAAA,CACV","sourcesContent":["import {\n\ttype GenerateContentConfig,\n\ttype GenerateContentParameters,\n\tGoogleGenAI,\n\ttype ThinkingConfig,\n} from \"@google/genai\";\nimport { getEnvApiKey } from \"../env-api-keys.js\";\nimport { calculateCost } from \"../models.js\";\nimport type {\n\tApi,\n\tAssistantMessage,\n\tContext,\n\tModel,\n\tSimpleStreamOptions,\n\tStreamFunction,\n\tStreamOptions,\n\tTextContent,\n\tThinkingBudgets,\n\tThinkingContent,\n\tThinkingLevel,\n\tToolCall,\n} from \"../types.js\";\nimport { AssistantMessageEventStream } from \"../utils/event-stream.js\";\nimport { sanitizeSurrogates } from \"../utils/sanitize-unicode.js\";\nimport type { GoogleThinkingLevel } from \"./google-gemini-cli.js\";\nimport {\n\tconvertMessages,\n\tconvertTools,\n\tisThinkingPart,\n\tmapStopReason,\n\tmapToolChoice,\n\tretainThoughtSignature,\n} from \"./google-shared.js\";\nimport { buildBaseOptions, clampReasoning } from \"./simple-options.js\";\n\nexport interface GoogleOptions extends StreamOptions {\n\ttoolChoice?: \"auto\" | \"none\" | \"any\";\n\tthinking?: {\n\t\tenabled: boolean;\n\t\tbudgetTokens?: number; // -1 for dynamic, 0 to disable\n\t\tlevel?: GoogleThinkingLevel;\n\t};\n}\n\n// Counter for generating unique tool call IDs\nlet toolCallCounter = 0;\n\nexport const streamGoogle: StreamFunction<\"google-generative-ai\", GoogleOptions> = (\n\tmodel: Model<\"google-generative-ai\">,\n\tcontext: Context,\n\toptions?: GoogleOptions,\n): AssistantMessageEventStream => {\n\tconst stream = new AssistantMessageEventStream();\n\n\t(async () => {\n\t\tconst output: AssistantMessage = {\n\t\t\trole: \"assistant\",\n\t\t\tcontent: [],\n\t\t\tapi: \"google-generative-ai\" as Api,\n\t\t\tprovider: model.provider,\n\t\t\tmodel: model.id,\n\t\t\tusage: {\n\t\t\t\tinput: 0,\n\t\t\t\toutput: 0,\n\t\t\t\tcacheRead: 0,\n\t\t\t\tcacheWrite: 0,\n\t\t\t\ttotalTokens: 0,\n\t\t\t\tcost: { input: 0, output: 0, cacheRead: 0, cacheWrite: 0, total: 0 },\n\t\t\t},\n\t\t\tstopReason: \"stop\",\n\t\t\ttimestamp: Date.now(),\n\t\t};\n\n\t\ttry {\n\t\t\tconst apiKey = options?.apiKey || getEnvApiKey(model.provider) || \"\";\n\t\t\tconst client = createClient(model, apiKey, options?.headers);\n\t\t\tlet params = buildParams(model, context, options);\n\t\t\tconst nextParams = await options?.onPayload?.(params, model);\n\t\t\tif (nextParams !== undefined) {\n\t\t\t\tparams = nextParams as GenerateContentParameters;\n\t\t\t}\n\t\t\tconst googleStream = await client.models.generateContentStream(params);\n\n\t\t\tstream.push({ type: \"start\", partial: output });\n\t\t\tlet currentBlock: TextContent | ThinkingContent | null = null;\n\t\t\tconst blocks = output.content;\n\t\t\tconst blockIndex = () => blocks.length - 1;\n\t\t\tfor await (const chunk of googleStream) {\n\t\t\t\t// @google/genai documents GenerateContentResponse.responseId as an output-only field\n\t\t\t\t// used to identify each response. Keep the first non-empty one from the stream.\n\t\t\t\toutput.responseId ||= chunk.responseId;\n\t\t\t\tconst candidate = chunk.candidates?.[0];\n\t\t\t\tif (candidate?.content?.parts) {\n\t\t\t\t\tfor (const part of candidate.content.parts) {\n\t\t\t\t\t\tif (part.text !== undefined) {\n\t\t\t\t\t\t\tconst isThinking = isThinkingPart(part);\n\t\t\t\t\t\t\tif (\n\t\t\t\t\t\t\t\t!currentBlock ||\n\t\t\t\t\t\t\t\t(isThinking && currentBlock.type !== \"thinking\") ||\n\t\t\t\t\t\t\t\t(!isThinking && currentBlock.type !== \"text\")\n\t\t\t\t\t\t\t) {\n\t\t\t\t\t\t\t\tif (currentBlock) {\n\t\t\t\t\t\t\t\t\tif (currentBlock.type === \"text\") {\n\t\t\t\t\t\t\t\t\t\tstream.push({\n\t\t\t\t\t\t\t\t\t\t\ttype: \"text_end\",\n\t\t\t\t\t\t\t\t\t\t\tcontentIndex: blocks.length - 1,\n\t\t\t\t\t\t\t\t\t\t\tcontent: currentBlock.text,\n\t\t\t\t\t\t\t\t\t\t\tpartial: output,\n\t\t\t\t\t\t\t\t\t\t});\n\t\t\t\t\t\t\t\t\t} else {\n\t\t\t\t\t\t\t\t\t\tstream.push({\n\t\t\t\t\t\t\t\t\t\t\ttype: \"thinking_end\",\n\t\t\t\t\t\t\t\t\t\t\tcontentIndex: blockIndex(),\n\t\t\t\t\t\t\t\t\t\t\tcontent: currentBlock.thinking,\n\t\t\t\t\t\t\t\t\t\t\tpartial: output,\n\t\t\t\t\t\t\t\t\t\t});\n\t\t\t\t\t\t\t\t\t}\n\t\t\t\t\t\t\t\t}\n\t\t\t\t\t\t\t\tif (isThinking) {\n\t\t\t\t\t\t\t\t\tcurrentBlock = { type: \"thinking\", thinking: \"\", thinkingSignature: undefined };\n\t\t\t\t\t\t\t\t\toutput.content.push(currentBlock);\n\t\t\t\t\t\t\t\t\tstream.push({ type: \"thinking_start\", contentIndex: blockIndex(), partial: output });\n\t\t\t\t\t\t\t\t} else {\n\t\t\t\t\t\t\t\t\tcurrentBlock = { type: \"text\", text: \"\" };\n\t\t\t\t\t\t\t\t\toutput.content.push(currentBlock);\n\t\t\t\t\t\t\t\t\tstream.push({ type: \"text_start\", contentIndex: blockIndex(), partial: output });\n\t\t\t\t\t\t\t\t}\n\t\t\t\t\t\t\t}\n\t\t\t\t\t\t\tif (currentBlock.type === \"thinking\") {\n\t\t\t\t\t\t\t\tcurrentBlock.thinking += part.text;\n\t\t\t\t\t\t\t\tcurrentBlock.thinkingSignature = retainThoughtSignature(\n\t\t\t\t\t\t\t\t\tcurrentBlock.thinkingSignature,\n\t\t\t\t\t\t\t\t\tpart.thoughtSignature,\n\t\t\t\t\t\t\t\t);\n\t\t\t\t\t\t\t\tstream.push({\n\t\t\t\t\t\t\t\t\ttype: \"thinking_delta\",\n\t\t\t\t\t\t\t\t\tcontentIndex: blockIndex(),\n\t\t\t\t\t\t\t\t\tdelta: part.text,\n\t\t\t\t\t\t\t\t\tpartial: output,\n\t\t\t\t\t\t\t\t});\n\t\t\t\t\t\t\t} else {\n\t\t\t\t\t\t\t\tcurrentBlock.text += part.text;\n\t\t\t\t\t\t\t\tcurrentBlock.textSignature = retainThoughtSignature(\n\t\t\t\t\t\t\t\t\tcurrentBlock.textSignature,\n\t\t\t\t\t\t\t\t\tpart.thoughtSignature,\n\t\t\t\t\t\t\t\t);\n\t\t\t\t\t\t\t\tstream.push({\n\t\t\t\t\t\t\t\t\ttype: \"text_delta\",\n\t\t\t\t\t\t\t\t\tcontentIndex: blockIndex(),\n\t\t\t\t\t\t\t\t\tdelta: part.text,\n\t\t\t\t\t\t\t\t\tpartial: output,\n\t\t\t\t\t\t\t\t});\n\t\t\t\t\t\t\t}\n\t\t\t\t\t\t}\n\n\t\t\t\t\t\tif (part.functionCall) {\n\t\t\t\t\t\t\tif (currentBlock) {\n\t\t\t\t\t\t\t\tif (currentBlock.type === \"text\") {\n\t\t\t\t\t\t\t\t\tstream.push({\n\t\t\t\t\t\t\t\t\t\ttype: \"text_end\",\n\t\t\t\t\t\t\t\t\t\tcontentIndex: blockIndex(),\n\t\t\t\t\t\t\t\t\t\tcontent: currentBlock.text,\n\t\t\t\t\t\t\t\t\t\tpartial: output,\n\t\t\t\t\t\t\t\t\t});\n\t\t\t\t\t\t\t\t} else {\n\t\t\t\t\t\t\t\t\tstream.push({\n\t\t\t\t\t\t\t\t\t\ttype: \"thinking_end\",\n\t\t\t\t\t\t\t\t\t\tcontentIndex: blockIndex(),\n\t\t\t\t\t\t\t\t\t\tcontent: currentBlock.thinking,\n\t\t\t\t\t\t\t\t\t\tpartial: output,\n\t\t\t\t\t\t\t\t\t});\n\t\t\t\t\t\t\t\t}\n\t\t\t\t\t\t\t\tcurrentBlock = null;\n\t\t\t\t\t\t\t}\n\n\t\t\t\t\t\t\t// Generate unique ID if not provided or if it's a duplicate\n\t\t\t\t\t\t\tconst providedId = part.functionCall.id;\n\t\t\t\t\t\t\tconst needsNewId =\n\t\t\t\t\t\t\t\t!providedId || output.content.some((b) => b.type === \"toolCall\" && b.id === providedId);\n\t\t\t\t\t\t\tconst toolCallId = needsNewId\n\t\t\t\t\t\t\t\t? `${part.functionCall.name}_${Date.now()}_${++toolCallCounter}`\n\t\t\t\t\t\t\t\t: providedId;\n\n\t\t\t\t\t\t\tconst toolCall: ToolCall = {\n\t\t\t\t\t\t\t\ttype: \"toolCall\",\n\t\t\t\t\t\t\t\tid: toolCallId,\n\t\t\t\t\t\t\t\tname: part.functionCall.name || \"\",\n\t\t\t\t\t\t\t\targuments: (part.functionCall.args as Record<string, any>) ?? {},\n\t\t\t\t\t\t\t\t...(part.thoughtSignature && { thoughtSignature: part.thoughtSignature }),\n\t\t\t\t\t\t\t};\n\n\t\t\t\t\t\t\toutput.content.push(toolCall);\n\t\t\t\t\t\t\tstream.push({ type: \"toolcall_start\", contentIndex: blockIndex(), partial: output });\n\t\t\t\t\t\t\tstream.push({\n\t\t\t\t\t\t\t\ttype: \"toolcall_delta\",\n\t\t\t\t\t\t\t\tcontentIndex: blockIndex(),\n\t\t\t\t\t\t\t\tdelta: JSON.stringify(toolCall.arguments),\n\t\t\t\t\t\t\t\tpartial: output,\n\t\t\t\t\t\t\t});\n\t\t\t\t\t\t\tstream.push({ type: \"toolcall_end\", contentIndex: blockIndex(), toolCall, partial: output });\n\t\t\t\t\t\t}\n\t\t\t\t\t}\n\t\t\t\t}\n\n\t\t\t\tif (candidate?.finishReason) {\n\t\t\t\t\toutput.stopReason = mapStopReason(candidate.finishReason);\n\t\t\t\t\tif (output.content.some((b) => b.type === \"toolCall\")) {\n\t\t\t\t\t\toutput.stopReason = \"toolUse\";\n\t\t\t\t\t}\n\t\t\t\t}\n\n\t\t\t\tif (chunk.usageMetadata) {\n\t\t\t\t\toutput.usage = {\n\t\t\t\t\t\tinput:\n\t\t\t\t\t\t\t(chunk.usageMetadata.promptTokenCount || 0) - (chunk.usageMetadata.cachedContentTokenCount || 0),\n\t\t\t\t\t\toutput:\n\t\t\t\t\t\t\t(chunk.usageMetadata.candidatesTokenCount || 0) + (chunk.usageMetadata.thoughtsTokenCount || 0),\n\t\t\t\t\t\tcacheRead: chunk.usageMetadata.cachedContentTokenCount || 0,\n\t\t\t\t\t\tcacheWrite: 0,\n\t\t\t\t\t\ttotalTokens: chunk.usageMetadata.totalTokenCount || 0,\n\t\t\t\t\t\tcost: {\n\t\t\t\t\t\t\tinput: 0,\n\t\t\t\t\t\t\toutput: 0,\n\t\t\t\t\t\t\tcacheRead: 0,\n\t\t\t\t\t\t\tcacheWrite: 0,\n\t\t\t\t\t\t\ttotal: 0,\n\t\t\t\t\t\t},\n\t\t\t\t\t};\n\t\t\t\t\tcalculateCost(model, output.usage);\n\t\t\t\t}\n\t\t\t}\n\n\t\t\tif (currentBlock) {\n\t\t\t\tif (currentBlock.type === \"text\") {\n\t\t\t\t\tstream.push({\n\t\t\t\t\t\ttype: \"text_end\",\n\t\t\t\t\t\tcontentIndex: blockIndex(),\n\t\t\t\t\t\tcontent: currentBlock.text,\n\t\t\t\t\t\tpartial: output,\n\t\t\t\t\t});\n\t\t\t\t} else {\n\t\t\t\t\tstream.push({\n\t\t\t\t\t\ttype: \"thinking_end\",\n\t\t\t\t\t\tcontentIndex: blockIndex(),\n\t\t\t\t\t\tcontent: currentBlock.thinking,\n\t\t\t\t\t\tpartial: output,\n\t\t\t\t\t});\n\t\t\t\t}\n\t\t\t}\n\n\t\t\tif (options?.signal?.aborted) {\n\t\t\t\tthrow new Error(\"Request was aborted\");\n\t\t\t}\n\n\t\t\tif (output.stopReason === \"aborted\" || output.stopReason === \"error\") {\n\t\t\t\tthrow new Error(\"An unknown error occurred\");\n\t\t\t}\n\n\t\t\tstream.push({ type: \"done\", reason: output.stopReason, message: output });\n\t\t\tstream.end();\n\t\t} catch (error) {\n\t\t\t// Remove internal index property used during streaming\n\t\t\tfor (const block of output.content) {\n\t\t\t\tif (\"index\" in block) {\n\t\t\t\t\tdelete (block as { index?: number }).index;\n\t\t\t\t}\n\t\t\t}\n\t\t\toutput.stopReason = options?.signal?.aborted ? \"aborted\" : \"error\";\n\t\t\toutput.errorMessage = error instanceof Error ? error.message : JSON.stringify(error);\n\t\t\tstream.push({ type: \"error\", reason: output.stopReason, error: output });\n\t\t\tstream.end();\n\t\t}\n\t})();\n\n\treturn stream;\n};\n\nexport const streamSimpleGoogle: StreamFunction<\"google-generative-ai\", SimpleStreamOptions> = (\n\tmodel: Model<\"google-generative-ai\">,\n\tcontext: Context,\n\toptions?: SimpleStreamOptions,\n): AssistantMessageEventStream => {\n\tconst apiKey = options?.apiKey || getEnvApiKey(model.provider);\n\tif (!apiKey) {\n\t\tthrow new Error(`No API key for provider: ${model.provider}`);\n\t}\n\n\tconst base = buildBaseOptions(model, options, apiKey);\n\tif (!options?.reasoning) {\n\t\treturn streamGoogle(model, context, { ...base, thinking: { enabled: false } } satisfies GoogleOptions);\n\t}\n\n\tconst effort = clampReasoning(options.reasoning)!;\n\tconst googleModel = model as Model<\"google-generative-ai\">;\n\n\tif (isGemini3ProModel(googleModel) || isGemini3FlashModel(googleModel)) {\n\t\treturn streamGoogle(model, context, {\n\t\t\t...base,\n\t\t\tthinking: {\n\t\t\t\tenabled: true,\n\t\t\t\tlevel: getGemini3ThinkingLevel(effort, googleModel),\n\t\t\t},\n\t\t} satisfies GoogleOptions);\n\t}\n\n\treturn streamGoogle(model, context, {\n\t\t...base,\n\t\tthinking: {\n\t\t\tenabled: true,\n\t\t\tbudgetTokens: getGoogleBudget(googleModel, effort, options.thinkingBudgets),\n\t\t},\n\t} satisfies GoogleOptions);\n};\n\nfunction createClient(\n\tmodel: Model<\"google-generative-ai\">,\n\tapiKey?: string,\n\toptionsHeaders?: Record<string, string>,\n): GoogleGenAI {\n\tconst httpOptions: { baseUrl?: string; apiVersion?: string; headers?: Record<string, string> } = {};\n\tif (model.baseUrl) {\n\t\thttpOptions.baseUrl = model.baseUrl;\n\t\thttpOptions.apiVersion = \"\"; // baseUrl already includes version path, don't append\n\t}\n\tif (model.headers || optionsHeaders) {\n\t\thttpOptions.headers = { ...model.headers, ...optionsHeaders };\n\t}\n\n\treturn new GoogleGenAI({\n\t\tapiKey,\n\t\thttpOptions: Object.keys(httpOptions).length > 0 ? httpOptions : undefined,\n\t});\n}\n\nfunction buildParams(\n\tmodel: Model<\"google-generative-ai\">,\n\tcontext: Context,\n\toptions: GoogleOptions = {},\n): GenerateContentParameters {\n\tconst contents = convertMessages(model, context);\n\n\tconst generationConfig: GenerateContentConfig = {};\n\tif (options.temperature !== undefined) {\n\t\tgenerationConfig.temperature = options.temperature;\n\t}\n\tif (options.maxTokens !== undefined) {\n\t\tgenerationConfig.maxOutputTokens = options.maxTokens;\n\t}\n\n\tconst config: GenerateContentConfig = {\n\t\t...(Object.keys(generationConfig).length > 0 && generationConfig),\n\t\t...(context.systemPrompt && { systemInstruction: sanitizeSurrogates(context.systemPrompt) }),\n\t\t...(context.tools && context.tools.length > 0 && { tools: convertTools(context.tools) }),\n\t};\n\n\tif (context.tools && context.tools.length > 0 && options.toolChoice) {\n\t\tconfig.toolConfig = {\n\t\t\tfunctionCallingConfig: {\n\t\t\t\tmode: mapToolChoice(options.toolChoice),\n\t\t\t},\n\t\t};\n\t} else {\n\t\tconfig.toolConfig = undefined;\n\t}\n\n\tif (options.thinking?.enabled && model.reasoning) {\n\t\tconst thinkingConfig: ThinkingConfig = { includeThoughts: true };\n\t\tif (options.thinking.level !== undefined) {\n\t\t\t// Cast to any since our GoogleThinkingLevel mirrors Google's ThinkingLevel enum values\n\t\t\tthinkingConfig.thinkingLevel = options.thinking.level as any;\n\t\t} else if (options.thinking.budgetTokens !== undefined) {\n\t\t\tthinkingConfig.thinkingBudget = options.thinking.budgetTokens;\n\t\t}\n\t\tconfig.thinkingConfig = thinkingConfig;\n\t} else if (model.reasoning && options.thinking && !options.thinking.enabled) {\n\t\tconfig.thinkingConfig = getDisabledThinkingConfig(model);\n\t}\n\n\tif (options.signal) {\n\t\tif (options.signal.aborted) {\n\t\t\tthrow new Error(\"Request aborted\");\n\t\t}\n\t\tconfig.abortSignal = options.signal;\n\t}\n\n\tconst params: GenerateContentParameters = {\n\t\tmodel: model.id,\n\t\tcontents,\n\t\tconfig,\n\t};\n\n\treturn params;\n}\n\ntype ClampedThinkingLevel = Exclude<ThinkingLevel, \"xhigh\">;\n\nfunction isGemini3ProModel(model: Model<\"google-generative-ai\">): boolean {\n\treturn /gemini-3(?:\\.\\d+)?-pro/.test(model.id.toLowerCase());\n}\n\nfunction isGemini3FlashModel(model: Model<\"google-generative-ai\">): boolean {\n\treturn /gemini-3(?:\\.\\d+)?-flash/.test(model.id.toLowerCase());\n}\n\nfunction getDisabledThinkingConfig(model: Model<\"google-generative-ai\">): ThinkingConfig {\n\t// Google docs: Gemini 3.1 Pro cannot disable thinking, and Gemini 3 Flash / Flash-Lite\n\t// do not support full thinking-off either. For Gemini 3 models, use the lowest supported\n\t// thinkingLevel without includeThoughts so hidden thinking remains invisible to pi.\n\tif (isGemini3ProModel(model)) {\n\t\treturn { thinkingLevel: \"LOW\" as any };\n\t}\n\tif (isGemini3FlashModel(model)) {\n\t\treturn { thinkingLevel: \"MINIMAL\" as any };\n\t}\n\n\t// Gemini 2.x supports disabling via thinkingBudget = 0.\n\treturn { thinkingBudget: 0 };\n}\n\nfunction getGemini3ThinkingLevel(\n\teffort: ClampedThinkingLevel,\n\tmodel: Model<\"google-generative-ai\">,\n): GoogleThinkingLevel {\n\tif (isGemini3ProModel(model)) {\n\t\tswitch (effort) {\n\t\t\tcase \"minimal\":\n\t\t\tcase \"low\":\n\t\t\t\treturn \"LOW\";\n\t\t\tcase \"medium\":\n\t\t\tcase \"high\":\n\t\t\t\treturn \"HIGH\";\n\t\t}\n\t}\n\tswitch (effort) {\n\t\tcase \"minimal\":\n\t\t\treturn \"MINIMAL\";\n\t\tcase \"low\":\n\t\t\treturn \"LOW\";\n\t\tcase \"medium\":\n\t\t\treturn \"MEDIUM\";\n\t\tcase \"high\":\n\t\t\treturn \"HIGH\";\n\t}\n}\n\nfunction getGoogleBudget(\n\tmodel: Model<\"google-generative-ai\">,\n\teffort: ClampedThinkingLevel,\n\tcustomBudgets?: ThinkingBudgets,\n): number {\n\tif (customBudgets?.[effort] !== undefined) {\n\t\treturn customBudgets[effort]!;\n\t}\n\n\tif (model.id.includes(\"2.5-pro\")) {\n\t\tconst budgets: Record<ClampedThinkingLevel, number> = {\n\t\t\tminimal: 128,\n\t\t\tlow: 2048,\n\t\t\tmedium: 8192,\n\t\t\thigh: 32768,\n\t\t};\n\t\treturn budgets[effort];\n\t}\n\n\tif (model.id.includes(\"2.5-flash\")) {\n\t\tconst budgets: Record<ClampedThinkingLevel, number> = {\n\t\t\tminimal: 128,\n\t\t\tlow: 2048,\n\t\t\tmedium: 8192,\n\t\t\thigh: 24576,\n\t\t};\n\t\treturn budgets[effort];\n\t}\n\n\treturn -1;\n}\n"]}
+{"version":3,"file":"google.js","sourceRoot":"","sources":["../../src/providers/google.ts"],"names":[],"mappings":"AAAA,OAAO,EAGN,WAAW,GAEX,MAAM,eAAe,CAAC;AACvB,OAAO,EAAE,YAAY,EAAE,MAAM,oBAAoB,CAAC;AAClD,OAAO,EAAE,aAAa,EAAE,MAAM,cAAc,CAAC;AAe7C,OAAO,EAAE,2BAA2B,EAAE,MAAM,0BAA0B,CAAC;AACvE,OAAO,EAAE,kBAAkB,EAAE,MAAM,8BAA8B,CAAC;AAElE,OAAO,EACN,eAAe,EACf,YAAY,EACZ,cAAc,EACd,aAAa,EACb,aAAa,EACb,sBAAsB,GACtB,MAAM,oBAAoB,CAAC;AAC5B,OAAO,EAAE,gBAAgB,EAAE,cAAc,EAAE,MAAM,qBAAqB,CAAC;AAWvE,8CAA8C;AAC9C,IAAI,eAAe,GAAG,CAAC,CAAC;AAExB,MAAM,CAAC,MAAM,YAAY,GAA0D,CAClF,KAAoC,EACpC,OAAgB,EAChB,OAAuB,EACO,EAAE,CAAC;IACjC,MAAM,MAAM,GAAG,IAAI,2BAA2B,EAAE,CAAC;IAEjD,CAAC,KAAK,IAAI,EAAE,CAAC;QACZ,MAAM,MAAM,GAAqB;YAChC,IAAI,EAAE,WAAW;YACjB,OAAO,EAAE,EAAE;YACX,GAAG,EAAE,sBAA6B;YAClC,QAAQ,EAAE,KAAK,CAAC,QAAQ;YACxB,KAAK,EAAE,KAAK,CAAC,EAAE;YACf,KAAK,EAAE;gBACN,KAAK,EAAE,CAAC;gBACR,MAAM,EAAE,CAAC;gBACT,SAAS,EAAE,CAAC;gBACZ,UAAU,EAAE,CAAC;gBACb,WAAW,EAAE,CAAC;gBACd,IAAI,EAAE,EAAE,KAAK,EAAE,CAAC,EAAE,MAAM,EAAE,CAAC,EAAE,SAAS,EAAE,CAAC,EAAE,UAAU,EAAE,CAAC,EAAE,KAAK,EAAE,CAAC,EAAE;aACpE;YACD,UAAU,EAAE,MAAM;YAClB,SAAS,EAAE,IAAI,CAAC,GAAG,EAAE;SACrB,CAAC;QAEF,IAAI,CAAC;YACJ,MAAM,MAAM,GAAG,OAAO,EAAE,MAAM,IAAI,YAAY,CAAC,KAAK,CAAC,QAAQ,CAAC,IAAI,EAAE,CAAC;YACrE,MAAM,MAAM,GAAG,YAAY,CAAC,KAAK,EAAE,MAAM,EAAE,OAAO,EAAE,OAAO,CAAC,CAAC;YAC7D,IAAI,MAAM,GAAG,WAAW,CAAC,KAAK,EAAE,OAAO,EAAE,OAAO,CAAC,CAAC;YAClD,MAAM,UAAU,GAAG,MAAM,OAAO,EAAE,SAAS,EAAE,CAAC,MAAM,EAAE,KAAK,CAAC,CAAC;YAC7D,IAAI,UAAU,KAAK,SAAS,EAAE,CAAC;gBAC9B,MAAM,GAAG,UAAuC,CAAC;YAClD,CAAC;YACD,MAAM,YAAY,GAAG,MAAM,MAAM,CAAC,MAAM,CAAC,qBAAqB,CAAC,MAAM,CAAC,CAAC;YAEvE,MAAM,CAAC,IAAI,CAAC,EAAE,IAAI,EAAE,OAAO,EAAE,OAAO,EAAE,MAAM,EAAE,CAAC,CAAC;YAChD,IAAI,YAAY,GAAyC,IAAI,CAAC;YAC9D,MAAM,MAAM,GAAG,MAAM,CAAC,OAAO,CAAC;YAC9B,MAAM,UAAU,GAAG,GAAG,EAAE,CAAC,MAAM,CAAC,MAAM,GAAG,CAAC,CAAC;YAC3C,IAAI,KAAK,EAAE,MAAM,KAAK,IAAI,YAAY,EAAE,CAAC;gBACxC,qFAAqF;gBACrF,gFAAgF;gBAChF,MAAM,CAAC,UAAU,KAAK,KAAK,CAAC,UAAU,CAAC;gBACvC,MAAM,SAAS,GAAG,KAAK,CAAC,UAAU,EAAE,CAAC,CAAC,CAAC,CAAC;gBACxC,IAAI,SAAS,EAAE,OAAO,EAAE,KAAK,EAAE,CAAC;oBAC/B,KAAK,MAAM,IAAI,IAAI,SAAS,CAAC,OAAO,CAAC,KAAK,EAAE,CAAC;wBAC5C,IAAI,IAAI,CAAC,IAAI,KAAK,SAAS,EAAE,CAAC;4BAC7B,MAAM,UAAU,GAAG,cAAc,CAAC,IAAI,CAAC,CAAC;4BACxC,IACC,CAAC,YAAY;gCACb,CAAC,UAAU,IAAI,YAAY,CAAC,IAAI,KAAK,UAAU,CAAC;gCAChD,CAAC,CAAC,UAAU,IAAI,YAAY,CAAC,IAAI,KAAK,MAAM,CAAC,EAC5C,CAAC;gCACF,IAAI,YAAY,EAAE,CAAC;oCAClB,IAAI,YAAY,CAAC,IAAI,KAAK,MAAM,EAAE,CAAC;wCAClC,MAAM,CAAC,IAAI,CAAC;4CACX,IAAI,EAAE,UAAU;4CAChB,YAAY,EAAE,MAAM,CAAC,MAAM,GAAG,CAAC;4CAC/B,OAAO,EAAE,YAAY,CAAC,IAAI;4CAC1B,OAAO,EAAE,MAAM;yCACf,CAAC,CAAC;oCACJ,CAAC;yCAAM,CAAC;wCACP,MAAM,CAAC,IAAI,CAAC;4CACX,IAAI,EAAE,cAAc;4CACpB,YAAY,EAAE,UAAU,EAAE;4CAC1B,OAAO,EAAE,YAAY,CAAC,QAAQ;4CAC9B,OAAO,EAAE,MAAM;yCACf,CAAC,CAAC;oCACJ,CAAC;gCACF,CAAC;gCACD,IAAI,UAAU,EAAE,CAAC;oCAChB,YAAY,GAAG,EAAE,IAAI,EAAE,UAAU,EAAE,QAAQ,EAAE,EAAE,EAAE,iBAAiB,EAAE,SAAS,EAAE,CAAC;oCAChF,MAAM,CAAC,OAAO,CAAC,IAAI,CAAC,YAAY,CAAC,CAAC;oCAClC,MAAM,CAAC,IAAI,CAAC,EAAE,IAAI,EAAE,gBAAgB,EAAE,YAAY,EAAE,UAAU,EAAE,EAAE,OAAO,EAAE,MAAM,EAAE,CAAC,CAAC;gCACtF,CAAC;qCAAM,CAAC;oCACP,YAAY,GAAG,EAAE,IAAI,EAAE,MAAM,EAAE,IAAI,EAAE,EAAE,EAAE,CAAC;oCAC1C,MAAM,CAAC,OAAO,CAAC,IAAI,CAAC,YAAY,CAAC,CAAC;oCAClC,MAAM,CAAC,IAAI,CAAC,EAAE,IAAI,EAAE,YAAY,EAAE,YAAY,EAAE,UAAU,EAAE,EAAE,OAAO,EAAE,MAAM,EAAE,CAAC,CAAC;gCAClF,CAAC;4BACF,CAAC;4BACD,IAAI,YAAY,CAAC,IAAI,KAAK,UAAU,EAAE,CAAC;gCACtC,YAAY,CAAC,QAAQ,IAAI,IAAI,CAAC,IAAI,CAAC;gCACnC,YAAY,CAAC,iBAAiB,GAAG,sBAAsB,CACtD,YAAY,CAAC,iBAAiB,EAC9B,IAAI,CAAC,gBAAgB,CACrB,CAAC;gCACF,MAAM,CAAC,IAAI,CAAC;oCACX,IAAI,EAAE,gBAAgB;oCACtB,YAAY,EAAE,UAAU,EAAE;oCAC1B,KAAK,EAAE,IAAI,CAAC,IAAI;oCAChB,OAAO,EAAE,MAAM;iCACf,CAAC,CAAC;4BACJ,CAAC;iCAAM,CAAC;gCACP,YAAY,CAAC,IAAI,IAAI,IAAI,CAAC,IAAI,CAAC;gCAC/B,YAAY,CAAC,aAAa,GAAG,sBAAsB,CAClD,YAAY,CAAC,aAAa,EAC1B,IAAI,CAAC,gBAAgB,CACrB,CAAC;gCACF,MAAM,CAAC,IAAI,CAAC;oCACX,IAAI,EAAE,YAAY;oCAClB,YAAY,EAAE,UAAU,EAAE;oCAC1B,KAAK,EAAE,IAAI,CAAC,IAAI;oCAChB,OAAO,EAAE,MAAM;iCACf,CAAC,CAAC;4BACJ,CAAC;wBACF,CAAC;wBAED,IAAI,IAAI,CAAC,YAAY,EAAE,CAAC;4BACvB,IAAI,YAAY,EAAE,CAAC;gCAClB,IAAI,YAAY,CAAC,IAAI,KAAK,MAAM,EAAE,CAAC;oCAClC,MAAM,CAAC,IAAI,CAAC;wCACX,IAAI,EAAE,UAAU;wCAChB,YAAY,EAAE,UAAU,EAAE;wCAC1B,OAAO,EAAE,YAAY,CAAC,IAAI;wCAC1B,OAAO,EAAE,MAAM;qCACf,CAAC,CAAC;gCACJ,CAAC;qCAAM,CAAC;oCACP,MAAM,CAAC,IAAI,CAAC;wCACX,IAAI,EAAE,cAAc;wCACpB,YAAY,EAAE,UAAU,EAAE;wCAC1B,OAAO,EAAE,YAAY,CAAC,QAAQ;wCAC9B,OAAO,EAAE,MAAM;qCACf,CAAC,CAAC;gCACJ,CAAC;gCACD,YAAY,GAAG,IAAI,CAAC;4BACrB,CAAC;4BAED,4DAA4D;4BAC5D,MAAM,UAAU,GAAG,IAAI,CAAC,YAAY,CAAC,EAAE,CAAC;4BACxC,MAAM,UAAU,GACf,CAAC,UAAU,IAAI,MAAM,CAAC,OAAO,CAAC,IAAI,CAAC,CAAC,CAAC,EAAE,EAAE,CAAC,CAAC,CAAC,IAAI,KAAK,UAAU,IAAI,CAAC,CAAC,EAAE,KAAK,UAAU,CAAC,CAAC;4BACzF,MAAM,UAAU,GAAG,UAAU;gCAC5B,CAAC,CAAC,GAAG,IAAI,CAAC,YAAY,CAAC,IAAI,IAAI,IAAI,CAAC,GAAG,EAAE,IAAI,EAAE,eAAe,EAAE;gCAChE,CAAC,CAAC,UAAU,CAAC;4BAEd,MAAM,QAAQ,GAAa;gCAC1B,IAAI,EAAE,UAAU;gCAChB,EAAE,EAAE,UAAU;gCACd,IAAI,EAAE,IAAI,CAAC,YAAY,CAAC,IAAI,IAAI,EAAE;gCAClC,SAAS,EAAG,IAAI,CAAC,YAAY,CAAC,IAA4B,IAAI,EAAE;gCAChE,GAAG,CAAC,IAAI,CAAC,gBAAgB,IAAI,EAAE,gBAAgB,EAAE,IAAI,CAAC,gBAAgB,EAAE,CAAC;6BACzE,CAAC;4BAEF,MAAM,CAAC,OAAO,CAAC,IAAI,CAAC,QAAQ,CAAC,CAAC;4BAC9B,MAAM,CAAC,IAAI,CAAC,EAAE,IAAI,EAAE,gBAAgB,EAAE,YAAY,EAAE,UAAU,EAAE,EAAE,OAAO,EAAE,MAAM,EAAE,CAAC,CAAC;4BACrF,MAAM,CAAC,IAAI,CAAC;gCACX,IAAI,EAAE,gBAAgB;gCACtB,YAAY,EAAE,UAAU,EAAE;gCAC1B,KAAK,EAAE,IAAI,CAAC,SAAS,CAAC,QAAQ,CAAC,SAAS,CAAC;gCACzC,OAAO,EAAE,MAAM;6BACf,CAAC,CAAC;4BACH,MAAM,CAAC,IAAI,CAAC,EAAE,IAAI,EAAE,cAAc,EAAE,YAAY,EAAE,UAAU,EAAE,EAAE,QAAQ,EAAE,OAAO,EAAE,MAAM,EAAE,CAAC,CAAC;wBAC9F,CAAC;oBACF,CAAC;gBACF,CAAC;gBAED,IAAI,SAAS,EAAE,YAAY,EAAE,CAAC;oBAC7B,MAAM,CAAC,UAAU,GAAG,aAAa,CAAC,SAAS,CAAC,YAAY,CAAC,CAAC;oBAC1D,IAAI,MAAM,CAAC,OAAO,CAAC,IAAI,CAAC,CAAC,CAAC,EAAE,EAAE,CAAC,CAAC,CAAC,IAAI,KAAK,UAAU,CAAC,EAAE,CAAC;wBACvD,MAAM,CAAC,UAAU,GAAG,SAAS,CAAC;oBAC/B,CAAC;gBACF,CAAC;gBAED,IAAI,KAAK,CAAC,aAAa,EAAE,CAAC;oBACzB,MAAM,CAAC,KAAK,GAAG;wBACd,KAAK,EACJ,CAAC,KAAK,CAAC,aAAa,CAAC,gBAAgB,IAAI,CAAC,CAAC,GAAG,CAAC,KAAK,CAAC,aAAa,CAAC,uBAAuB,IAAI,CAAC,CAAC;wBACjG,MAAM,EACL,CAAC,KAAK,CAAC,aAAa,CAAC,oBAAoB,IAAI,CAAC,CAAC,GAAG,CAAC,KAAK,CAAC,aAAa,CAAC,kBAAkB,IAAI,CAAC,CAAC;wBAChG,SAAS,EAAE,KAAK,CAAC,aAAa,CAAC,uBAAuB,IAAI,CAAC;wBAC3D,UAAU,EAAE,CAAC;wBACb,WAAW,EAAE,KAAK,CAAC,aAAa,CAAC,eAAe,IAAI,CAAC;wBACrD,IAAI,EAAE;4BACL,KAAK,EAAE,CAAC;4BACR,MAAM,EAAE,CAAC;4BACT,SAAS,EAAE,CAAC;4BACZ,UAAU,EAAE,CAAC;4BACb,KAAK,EAAE,CAAC;yBACR;qBACD,CAAC;oBACF,aAAa,CAAC,KAAK,EAAE,MAAM,CAAC,KAAK,CAAC,CAAC;gBACpC,CAAC;YACF,CAAC;YAED,IAAI,YAAY,EAAE,CAAC;gBAClB,IAAI,YAAY,CAAC,IAAI,KAAK,MAAM,EAAE,CAAC;oBAClC,MAAM,CAAC,IAAI,CAAC;wBACX,IAAI,EAAE,UAAU;wBAChB,YAAY,EAAE,UAAU,EAAE;wBAC1B,OAAO,EAAE,YAAY,CAAC,IAAI;wBAC1B,OAAO,EAAE,MAAM;qBACf,CAAC,CAAC;gBACJ,CAAC;qBAAM,CAAC;oBACP,MAAM,CAAC,IAAI,CAAC;wBACX,IAAI,EAAE,cAAc;wBACpB,YAAY,EAAE,UAAU,EAAE;wBAC1B,OAAO,EAAE,YAAY,CAAC,QAAQ;wBAC9B,OAAO,EAAE,MAAM;qBACf,CAAC,CAAC;gBACJ,CAAC;YACF,CAAC;YAED,IAAI,OAAO,EAAE,MAAM,EAAE,OAAO,EAAE,CAAC;gBAC9B,MAAM,IAAI,KAAK,CAAC,qBAAqB,CAAC,CAAC;YACxC,CAAC;YAED,IAAI,MAAM,CAAC,UAAU,KAAK,SAAS,IAAI,MAAM,CAAC,UAAU,KAAK,OAAO,EAAE,CAAC;gBACtE,MAAM,IAAI,KAAK,CAAC,2BAA2B,CAAC,CAAC;YAC9C,CAAC;YAED,MAAM,CAAC,IAAI,CAAC,EAAE,IAAI,EAAE,MAAM,EAAE,MAAM,EAAE,MAAM,CAAC,UAAU,EAAE,OAAO,EAAE,MAAM,EAAE,CAAC,CAAC;YAC1E,MAAM,CAAC,GAAG,EAAE,CAAC;QACd,CAAC;QAAC,OAAO,KAAK,EAAE,CAAC;YAChB,uDAAuD;YACvD,KAAK,MAAM,KAAK,IAAI,MAAM,CAAC,OAAO,EAAE,CAAC;gBACpC,IAAI,OAAO,IAAI,KAAK,EAAE,CAAC;oBACtB,OAAQ,KAA4B,CAAC,KAAK,CAAC;gBAC5C,CAAC;YACF,CAAC;YACD,MAAM,CAAC,UAAU,GAAG,OAAO,EAAE,MAAM,EAAE,OAAO,CAAC,CAAC,CAAC,SAAS,CAAC,CAAC,CAAC,OAAO,CAAC;YACnE,MAAM,CAAC,YAAY,GAAG,KAAK,YAAY,KAAK,CAAC,CAAC,CAAC,KAAK,CAAC,OAAO,CAAC,CAAC,CAAC,IAAI,CAAC,SAAS,CAAC,KAAK,CAAC,CAAC;YACrF,MAAM,CAAC,IAAI,CAAC,EAAE,IAAI,EAAE,OAAO,EAAE,MAAM,EAAE,MAAM,CAAC,UAAU,EAAE,KAAK,EAAE,MAAM,EAAE,CAAC,CAAC;YACzE,MAAM,CAAC,GAAG,EAAE,CAAC;QACd,CAAC;IAAA,CACD,CAAC,EAAE,CAAC;IAEL,OAAO,MAAM,CAAC;AAAA,CACd,CAAC;AAEF,MAAM,CAAC,MAAM,kBAAkB,GAAgE,CAC9F,KAAoC,EACpC,OAAgB,EAChB,OAA6B,EACC,EAAE,CAAC;IACjC,MAAM,MAAM,GAAG,OAAO,EAAE,MAAM,IAAI,YAAY,CAAC,KAAK,CAAC,QAAQ,CAAC,CAAC;IAC/D,IAAI,CAAC,MAAM,EAAE,CAAC;QACb,MAAM,IAAI,KAAK,CAAC,4BAA4B,KAAK,CAAC,QAAQ,EAAE,CAAC,CAAC;IAC/D,CAAC;IAED,MAAM,IAAI,GAAG,gBAAgB,CAAC,KAAK,EAAE,OAAO,EAAE,MAAM,CAAC,CAAC;IACtD,IAAI,CAAC,OAAO,EAAE,SAAS,EAAE,CAAC;QACzB,OAAO,YAAY,CAAC,KAAK,EAAE,OAAO,EAAE,EAAE,GAAG,IAAI,EAAE,QAAQ,EAAE,EAAE,OAAO,EAAE,KAAK,EAAE,EAA0B,CAAC,CAAC;IACxG,CAAC;IAED,MAAM,MAAM,GAAG,cAAc,CAAC,OAAO,CAAC,SAAS,CAAE,CAAC;IAClD,MAAM,WAAW,GAAG,KAAsC,CAAC;IAE3D,IAAI,iBAAiB,CAAC,WAAW,CAAC,IAAI,mBAAmB,CAAC,WAAW,CAAC,IAAI,aAAa,CAAC,WAAW,CAAC,EAAE,CAAC;QACtG,OAAO,YAAY,CAAC,KAAK,EAAE,OAAO,EAAE;YACnC,GAAG,IAAI;YACP,QAAQ,EAAE;gBACT,OAAO,EAAE,IAAI;gBACb,KAAK,EAAE,gBAAgB,CAAC,MAAM,EAAE,WAAW,CAAC;aAC5C;SACuB,CAAC,CAAC;IAC5B,CAAC;IAED,OAAO,YAAY,CAAC,KAAK,EAAE,OAAO,EAAE;QACnC,GAAG,IAAI;QACP,QAAQ,EAAE;YACT,OAAO,EAAE,IAAI;YACb,YAAY,EAAE,eAAe,CAAC,WAAW,EAAE,MAAM,EAAE,OAAO,CAAC,eAAe,CAAC;SAC3E;KACuB,CAAC,CAAC;AAAA,CAC3B,CAAC;AAEF,SAAS,YAAY,CACpB,KAAoC,EACpC,MAAe,EACf,cAAuC,EACzB;IACd,MAAM,WAAW,GAAgF,EAAE,CAAC;IACpG,IAAI,KAAK,CAAC,OAAO,EAAE,CAAC;QACnB,WAAW,CAAC,OAAO,GAAG,KAAK,CAAC,OAAO,CAAC;QACpC,WAAW,CAAC,UAAU,GAAG,EAAE,CAAC,CAAC,sDAAsD;IACpF,CAAC;IACD,IAAI,KAAK,CAAC,OAAO,IAAI,cAAc,EAAE,CAAC;QACrC,WAAW,CAAC,OAAO,GAAG,EAAE,GAAG,KAAK,CAAC,OAAO,EAAE,GAAG,cAAc,EAAE,CAAC;IAC/D,CAAC;IAED,OAAO,IAAI,WAAW,CAAC;QACtB,MAAM;QACN,WAAW,EAAE,MAAM,CAAC,IAAI,CAAC,WAAW,CAAC,CAAC,MAAM,GAAG,CAAC,CAAC,CAAC,CAAC,WAAW,CAAC,CAAC,CAAC,SAAS;KAC1E,CAAC,CAAC;AAAA,CACH;AAED,SAAS,WAAW,CACnB,KAAoC,EACpC,OAAgB,EAChB,OAAO,GAAkB,EAAE,EACC;IAC5B,MAAM,QAAQ,GAAG,eAAe,CAAC,KAAK,EAAE,OAAO,CAAC,CAAC;IAEjD,MAAM,gBAAgB,GAA0B,EAAE,CAAC;IACnD,IAAI,OAAO,CAAC,WAAW,KAAK,SAAS,EAAE,CAAC;QACvC,gBAAgB,CAAC,WAAW,GAAG,OAAO,CAAC,WAAW,CAAC;IACpD,CAAC;IACD,IAAI,OAAO,CAAC,SAAS,KAAK,SAAS,EAAE,CAAC;QACrC,gBAAgB,CAAC,eAAe,GAAG,OAAO,CAAC,SAAS,CAAC;IACtD,CAAC;IAED,MAAM,MAAM,GAA0B;QACrC,GAAG,CAAC,MAAM,CAAC,IAAI,CAAC,gBAAgB,CAAC,CAAC,MAAM,GAAG,CAAC,IAAI,gBAAgB,CAAC;QACjE,GAAG,CAAC,OAAO,CAAC,YAAY,IAAI,EAAE,iBAAiB,EAAE,kBAAkB,CAAC,OAAO,CAAC,YAAY,CAAC,EAAE,CAAC;QAC5F,GAAG,CAAC,OAAO,CAAC,KAAK,IAAI,OAAO,CAAC,KAAK,CAAC,MAAM,GAAG,CAAC,IAAI,EAAE,KAAK,EAAE,YAAY,CAAC,OAAO,CAAC,KAAK,CAAC,EAAE,CAAC;KACxF,CAAC;IAEF,IAAI,OAAO,CAAC,KAAK,IAAI,OAAO,CAAC,KAAK,CAAC,MAAM,GAAG,CAAC,IAAI,OAAO,CAAC,UAAU,EAAE,CAAC;QACrE,MAAM,CAAC,UAAU,GAAG;YACnB,qBAAqB,EAAE;gBACtB,IAAI,EAAE,aAAa,CAAC,OAAO,CAAC,UAAU,CAAC;aACvC;SACD,CAAC;IACH,CAAC;SAAM,CAAC;QACP,MAAM,CAAC,UAAU,GAAG,SAAS,CAAC;IAC/B,CAAC;IAED,IAAI,OAAO,CAAC,QAAQ,EAAE,OAAO,IAAI,KAAK,CAAC,SAAS,EAAE,CAAC;QAClD,MAAM,cAAc,GAAmB,EAAE,eAAe,EAAE,IAAI,EAAE,CAAC;QACjE,IAAI,OAAO,CAAC,QAAQ,CAAC,KAAK,KAAK,SAAS,EAAE,CAAC;YAC1C,uFAAuF;YACvF,cAAc,CAAC,aAAa,GAAG,OAAO,CAAC,QAAQ,CAAC,KAAY,CAAC;QAC9D,CAAC;aAAM,IAAI,OAAO,CAAC,QAAQ,CAAC,YAAY,KAAK,SAAS,EAAE,CAAC;YACxD,cAAc,CAAC,cAAc,GAAG,OAAO,CAAC,QAAQ,CAAC,YAAY,CAAC;QAC/D,CAAC;QACD,MAAM,CAAC,cAAc,GAAG,cAAc,CAAC;IACxC,CAAC;SAAM,IAAI,KAAK,CAAC,SAAS,IAAI,OAAO,CAAC,QAAQ,IAAI,CAAC,OAAO,CAAC,QAAQ,CAAC,OAAO,EAAE,CAAC;QAC7E,MAAM,CAAC,cAAc,GAAG,yBAAyB,CAAC,KAAK,CAAC,CAAC;IAC1D,CAAC;IAED,IAAI,OAAO,CAAC,MAAM,EAAE,CAAC;QACpB,IAAI,OAAO,CAAC,MAAM,CAAC,OAAO,EAAE,CAAC;YAC5B,MAAM,IAAI,KAAK,CAAC,iBAAiB,CAAC,CAAC;QACpC,CAAC;QACD,MAAM,CAAC,WAAW,GAAG,OAAO,CAAC,MAAM,CAAC;IACrC,CAAC;IAED,MAAM,MAAM,GAA8B;QACzC,KAAK,EAAE,KAAK,CAAC,EAAE;QACf,QAAQ;QACR,MAAM;KACN,CAAC;IAEF,OAAO,MAAM,CAAC;AAAA,CACd;AAID,SAAS,aAAa,CAAC,KAAoC,EAAW;IACrE,OAAO,UAAU,CAAC,IAAI,CAAC,KAAK,CAAC,EAAE,CAAC,WAAW,EAAE,CAAC,CAAC;AAAA,CAC/C;AAED,SAAS,iBAAiB,CAAC,KAAoC,EAAW;IACzE,OAAO,wBAAwB,CAAC,IAAI,CAAC,KAAK,CAAC,EAAE,CAAC,WAAW,EAAE,CAAC,CAAC;AAAA,CAC7D;AAED,SAAS,mBAAmB,CAAC,KAAoC,EAAW;IAC3E,OAAO,0BAA0B,CAAC,IAAI,CAAC,KAAK,CAAC,EAAE,CAAC,WAAW,EAAE,CAAC,CAAC;AAAA,CAC/D;AAED,SAAS,yBAAyB,CAAC,KAAoC,EAAkB;IACxF,uFAAuF;IACvF,yFAAyF;IACzF,oFAAoF;IACpF,IAAI,iBAAiB,CAAC,KAAK,CAAC,EAAE,CAAC;QAC9B,OAAO,EAAE,aAAa,EAAE,KAAY,EAAE,CAAC;IACxC,CAAC;IACD,IAAI,mBAAmB,CAAC,KAAK,CAAC,EAAE,CAAC;QAChC,OAAO,EAAE,aAAa,EAAE,SAAgB,EAAE,CAAC;IAC5C,CAAC;IACD,IAAI,aAAa,CAAC,KAAK,CAAC,EAAE,CAAC;QAC1B,OAAO,EAAE,aAAa,EAAE,SAAgB,EAAE,CAAC;IAC5C,CAAC;IAED,wDAAwD;IACxD,OAAO,EAAE,cAAc,EAAE,CAAC,EAAE,CAAC;AAAA,CAC7B;AAED,SAAS,gBAAgB,CAAC,MAA4B,EAAE,KAAoC,EAAuB;IAClH,IAAI,iBAAiB,CAAC,KAAK,CAAC,EAAE,CAAC;QAC9B,QAAQ,MAAM,EAAE,CAAC;YAChB,KAAK,SAAS,CAAC;YACf,KAAK,KAAK;gBACT,OAAO,KAAK,CAAC;YACd,KAAK,QAAQ,CAAC;YACd,KAAK,MAAM;gBACV,OAAO,MAAM,CAAC;QAChB,CAAC;IACF,CAAC;IACD,IAAI,aAAa,CAAC,KAAK,CAAC,EAAE,CAAC;QAC1B,QAAQ,MAAM,EAAE,CAAC;YAChB,KAAK,SAAS,CAAC;YACf,KAAK,KAAK;gBACT,OAAO,SAAS,CAAC;YAClB,KAAK,QAAQ,CAAC;YACd,KAAK,MAAM;gBACV,OAAO,MAAM,CAAC;QAChB,CAAC;IACF,CAAC;IACD,QAAQ,MAAM,EAAE,CAAC;QAChB,KAAK,SAAS;YACb,OAAO,SAAS,CAAC;QAClB,KAAK,KAAK;YACT,OAAO,KAAK,CAAC;QACd,KAAK,QAAQ;YACZ,OAAO,QAAQ,CAAC;QACjB,KAAK,MAAM;YACV,OAAO,MAAM,CAAC;IAChB,CAAC;AAAA,CACD;AAED,SAAS,eAAe,CACvB,KAAoC,EACpC,MAA4B,EAC5B,aAA+B,EACtB;IACT,IAAI,aAAa,EAAE,CAAC,MAAM,CAAC,KAAK,SAAS,EAAE,CAAC;QAC3C,OAAO,aAAa,CAAC,MAAM,CAAE,CAAC;IAC/B,CAAC;IAED,IAAI,KAAK,CAAC,EAAE,CAAC,QAAQ,CAAC,SAAS,CAAC,EAAE,CAAC;QAClC,MAAM,OAAO,GAAyC;YACrD,OAAO,EAAE,GAAG;YACZ,GAAG,EAAE,IAAI;YACT,MAAM,EAAE,IAAI;YACZ,IAAI,EAAE,KAAK;SACX,CAAC;QACF,OAAO,OAAO,CAAC,MAAM,CAAC,CAAC;IACxB,CAAC;IAED,IAAI,KAAK,CAAC,EAAE,CAAC,QAAQ,CAAC,gBAAgB,CAAC,EAAE,CAAC;QACzC,MAAM,OAAO,GAAyC;YACrD,OAAO,EAAE,GAAG;YACZ,GAAG,EAAE,IAAI;YACT,MAAM,EAAE,IAAI;YACZ,IAAI,EAAE,KAAK;SACX,CAAC;QACF,OAAO,OAAO,CAAC,MAAM,CAAC,CAAC;IACxB,CAAC;IAED,IAAI,KAAK,CAAC,EAAE,CAAC,QAAQ,CAAC,WAAW,CAAC,EAAE,CAAC;QACpC,MAAM,OAAO,GAAyC;YACrD,OAAO,EAAE,GAAG;YACZ,GAAG,EAAE,IAAI;YACT,MAAM,EAAE,IAAI;YACZ,IAAI,EAAE,KAAK;SACX,CAAC;QACF,OAAO,OAAO,CAAC,MAAM,CAAC,CAAC;IACxB,CAAC;IAED,OAAO,CAAC,CAAC,CAAC;AAAA,CACV","sourcesContent":["import {\n\ttype GenerateContentConfig,\n\ttype GenerateContentParameters,\n\tGoogleGenAI,\n\ttype ThinkingConfig,\n} from \"@google/genai\";\nimport { getEnvApiKey } from \"../env-api-keys.js\";\nimport { calculateCost } from \"../models.js\";\nimport type {\n\tApi,\n\tAssistantMessage,\n\tContext,\n\tModel,\n\tSimpleStreamOptions,\n\tStreamFunction,\n\tStreamOptions,\n\tTextContent,\n\tThinkingBudgets,\n\tThinkingContent,\n\tThinkingLevel,\n\tToolCall,\n} from \"../types.js\";\nimport { AssistantMessageEventStream } from \"../utils/event-stream.js\";\nimport { sanitizeSurrogates } from \"../utils/sanitize-unicode.js\";\nimport type { GoogleThinkingLevel } from \"./google-gemini-cli.js\";\nimport {\n\tconvertMessages,\n\tconvertTools,\n\tisThinkingPart,\n\tmapStopReason,\n\tmapToolChoice,\n\tretainThoughtSignature,\n} from \"./google-shared.js\";\nimport { buildBaseOptions, clampReasoning } from \"./simple-options.js\";\n\nexport interface GoogleOptions extends StreamOptions {\n\ttoolChoice?: \"auto\" | \"none\" | \"any\";\n\tthinking?: {\n\t\tenabled: boolean;\n\t\tbudgetTokens?: number; // -1 for dynamic, 0 to disable\n\t\tlevel?: GoogleThinkingLevel;\n\t};\n}\n\n// Counter for generating unique tool call IDs\nlet toolCallCounter = 0;\n\nexport const streamGoogle: StreamFunction<\"google-generative-ai\", GoogleOptions> = (\n\tmodel: Model<\"google-generative-ai\">,\n\tcontext: Context,\n\toptions?: GoogleOptions,\n): AssistantMessageEventStream => {\n\tconst stream = new AssistantMessageEventStream();\n\n\t(async () => {\n\t\tconst output: AssistantMessage = {\n\t\t\trole: \"assistant\",\n\t\t\tcontent: [],\n\t\t\tapi: \"google-generative-ai\" as Api,\n\t\t\tprovider: model.provider,\n\t\t\tmodel: model.id,\n\t\t\tusage: {\n\t\t\t\tinput: 0,\n\t\t\t\toutput: 0,\n\t\t\t\tcacheRead: 0,\n\t\t\t\tcacheWrite: 0,\n\t\t\t\ttotalTokens: 0,\n\t\t\t\tcost: { input: 0, output: 0, cacheRead: 0, cacheWrite: 0, total: 0 },\n\t\t\t},\n\t\t\tstopReason: \"stop\",\n\t\t\ttimestamp: Date.now(),\n\t\t};\n\n\t\ttry {\n\t\t\tconst apiKey = options?.apiKey || getEnvApiKey(model.provider) || \"\";\n\t\t\tconst client = createClient(model, apiKey, options?.headers);\n\t\t\tlet params = buildParams(model, context, options);\n\t\t\tconst nextParams = await options?.onPayload?.(params, model);\n\t\t\tif (nextParams !== undefined) {\n\t\t\t\tparams = nextParams as GenerateContentParameters;\n\t\t\t}\n\t\t\tconst googleStream = await client.models.generateContentStream(params);\n\n\t\t\tstream.push({ type: \"start\", partial: output });\n\t\t\tlet currentBlock: TextContent | ThinkingContent | null = null;\n\t\t\tconst blocks = output.content;\n\t\t\tconst blockIndex = () => blocks.length - 1;\n\t\t\tfor await (const chunk of googleStream) {\n\t\t\t\t// @google/genai documents GenerateContentResponse.responseId as an output-only field\n\t\t\t\t// used to identify each response. Keep the first non-empty one from the stream.\n\t\t\t\toutput.responseId ||= chunk.responseId;\n\t\t\t\tconst candidate = chunk.candidates?.[0];\n\t\t\t\tif (candidate?.content?.parts) {\n\t\t\t\t\tfor (const part of candidate.content.parts) {\n\t\t\t\t\t\tif (part.text !== undefined) {\n\t\t\t\t\t\t\tconst isThinking = isThinkingPart(part);\n\t\t\t\t\t\t\tif (\n\t\t\t\t\t\t\t\t!currentBlock ||\n\t\t\t\t\t\t\t\t(isThinking && currentBlock.type !== \"thinking\") ||\n\t\t\t\t\t\t\t\t(!isThinking && currentBlock.type !== \"text\")\n\t\t\t\t\t\t\t) {\n\t\t\t\t\t\t\t\tif (currentBlock) {\n\t\t\t\t\t\t\t\t\tif (currentBlock.type === \"text\") {\n\t\t\t\t\t\t\t\t\t\tstream.push({\n\t\t\t\t\t\t\t\t\t\t\ttype: \"text_end\",\n\t\t\t\t\t\t\t\t\t\t\tcontentIndex: blocks.length - 1,\n\t\t\t\t\t\t\t\t\t\t\tcontent: currentBlock.text,\n\t\t\t\t\t\t\t\t\t\t\tpartial: output,\n\t\t\t\t\t\t\t\t\t\t});\n\t\t\t\t\t\t\t\t\t} else {\n\t\t\t\t\t\t\t\t\t\tstream.push({\n\t\t\t\t\t\t\t\t\t\t\ttype: \"thinking_end\",\n\t\t\t\t\t\t\t\t\t\t\tcontentIndex: blockIndex(),\n\t\t\t\t\t\t\t\t\t\t\tcontent: currentBlock.thinking,\n\t\t\t\t\t\t\t\t\t\t\tpartial: output,\n\t\t\t\t\t\t\t\t\t\t});\n\t\t\t\t\t\t\t\t\t}\n\t\t\t\t\t\t\t\t}\n\t\t\t\t\t\t\t\tif (isThinking) {\n\t\t\t\t\t\t\t\t\tcurrentBlock = { type: \"thinking\", thinking: \"\", thinkingSignature: undefined };\n\t\t\t\t\t\t\t\t\toutput.content.push(currentBlock);\n\t\t\t\t\t\t\t\t\tstream.push({ type: \"thinking_start\", contentIndex: blockIndex(), partial: output });\n\t\t\t\t\t\t\t\t} else {\n\t\t\t\t\t\t\t\t\tcurrentBlock = { type: \"text\", text: \"\" };\n\t\t\t\t\t\t\t\t\toutput.content.push(currentBlock);\n\t\t\t\t\t\t\t\t\tstream.push({ type: \"text_start\", contentIndex: blockIndex(), partial: output });\n\t\t\t\t\t\t\t\t}\n\t\t\t\t\t\t\t}\n\t\t\t\t\t\t\tif (currentBlock.type === \"thinking\") {\n\t\t\t\t\t\t\t\tcurrentBlock.thinking += part.text;\n\t\t\t\t\t\t\t\tcurrentBlock.thinkingSignature = retainThoughtSignature(\n\t\t\t\t\t\t\t\t\tcurrentBlock.thinkingSignature,\n\t\t\t\t\t\t\t\t\tpart.thoughtSignature,\n\t\t\t\t\t\t\t\t);\n\t\t\t\t\t\t\t\tstream.push({\n\t\t\t\t\t\t\t\t\ttype: \"thinking_delta\",\n\t\t\t\t\t\t\t\t\tcontentIndex: blockIndex(),\n\t\t\t\t\t\t\t\t\tdelta: part.text,\n\t\t\t\t\t\t\t\t\tpartial: output,\n\t\t\t\t\t\t\t\t});\n\t\t\t\t\t\t\t} else {\n\t\t\t\t\t\t\t\tcurrentBlock.text += part.text;\n\t\t\t\t\t\t\t\tcurrentBlock.textSignature = retainThoughtSignature(\n\t\t\t\t\t\t\t\t\tcurrentBlock.textSignature,\n\t\t\t\t\t\t\t\t\tpart.thoughtSignature,\n\t\t\t\t\t\t\t\t);\n\t\t\t\t\t\t\t\tstream.push({\n\t\t\t\t\t\t\t\t\ttype: \"text_delta\",\n\t\t\t\t\t\t\t\t\tcontentIndex: blockIndex(),\n\t\t\t\t\t\t\t\t\tdelta: part.text,\n\t\t\t\t\t\t\t\t\tpartial: output,\n\t\t\t\t\t\t\t\t});\n\t\t\t\t\t\t\t}\n\t\t\t\t\t\t}\n\n\t\t\t\t\t\tif (part.functionCall) {\n\t\t\t\t\t\t\tif (currentBlock) {\n\t\t\t\t\t\t\t\tif (currentBlock.type === \"text\") {\n\t\t\t\t\t\t\t\t\tstream.push({\n\t\t\t\t\t\t\t\t\t\ttype: \"text_end\",\n\t\t\t\t\t\t\t\t\t\tcontentIndex: blockIndex(),\n\t\t\t\t\t\t\t\t\t\tcontent: currentBlock.text,\n\t\t\t\t\t\t\t\t\t\tpartial: output,\n\t\t\t\t\t\t\t\t\t});\n\t\t\t\t\t\t\t\t} else {\n\t\t\t\t\t\t\t\t\tstream.push({\n\t\t\t\t\t\t\t\t\t\ttype: \"thinking_end\",\n\t\t\t\t\t\t\t\t\t\tcontentIndex: blockIndex(),\n\t\t\t\t\t\t\t\t\t\tcontent: currentBlock.thinking,\n\t\t\t\t\t\t\t\t\t\tpartial: output,\n\t\t\t\t\t\t\t\t\t});\n\t\t\t\t\t\t\t\t}\n\t\t\t\t\t\t\t\tcurrentBlock = null;\n\t\t\t\t\t\t\t}\n\n\t\t\t\t\t\t\t// Generate unique ID if not provided or if it's a duplicate\n\t\t\t\t\t\t\tconst providedId = part.functionCall.id;\n\t\t\t\t\t\t\tconst needsNewId =\n\t\t\t\t\t\t\t\t!providedId || output.content.some((b) => b.type === \"toolCall\" && b.id === providedId);\n\t\t\t\t\t\t\tconst toolCallId = needsNewId\n\t\t\t\t\t\t\t\t? `${part.functionCall.name}_${Date.now()}_${++toolCallCounter}`\n\t\t\t\t\t\t\t\t: providedId;\n\n\t\t\t\t\t\t\tconst toolCall: ToolCall = {\n\t\t\t\t\t\t\t\ttype: \"toolCall\",\n\t\t\t\t\t\t\t\tid: toolCallId,\n\t\t\t\t\t\t\t\tname: part.functionCall.name || \"\",\n\t\t\t\t\t\t\t\targuments: (part.functionCall.args as Record<string, any>) ?? {},\n\t\t\t\t\t\t\t\t...(part.thoughtSignature && { thoughtSignature: part.thoughtSignature }),\n\t\t\t\t\t\t\t};\n\n\t\t\t\t\t\t\toutput.content.push(toolCall);\n\t\t\t\t\t\t\tstream.push({ type: \"toolcall_start\", contentIndex: blockIndex(), partial: output });\n\t\t\t\t\t\t\tstream.push({\n\t\t\t\t\t\t\t\ttype: \"toolcall_delta\",\n\t\t\t\t\t\t\t\tcontentIndex: blockIndex(),\n\t\t\t\t\t\t\t\tdelta: JSON.stringify(toolCall.arguments),\n\t\t\t\t\t\t\t\tpartial: output,\n\t\t\t\t\t\t\t});\n\t\t\t\t\t\t\tstream.push({ type: \"toolcall_end\", contentIndex: blockIndex(), toolCall, partial: output });\n\t\t\t\t\t\t}\n\t\t\t\t\t}\n\t\t\t\t}\n\n\t\t\t\tif (candidate?.finishReason) {\n\t\t\t\t\toutput.stopReason = mapStopReason(candidate.finishReason);\n\t\t\t\t\tif (output.content.some((b) => b.type === \"toolCall\")) {\n\t\t\t\t\t\toutput.stopReason = \"toolUse\";\n\t\t\t\t\t}\n\t\t\t\t}\n\n\t\t\t\tif (chunk.usageMetadata) {\n\t\t\t\t\toutput.usage = {\n\t\t\t\t\t\tinput:\n\t\t\t\t\t\t\t(chunk.usageMetadata.promptTokenCount || 0) - (chunk.usageMetadata.cachedContentTokenCount || 0),\n\t\t\t\t\t\toutput:\n\t\t\t\t\t\t\t(chunk.usageMetadata.candidatesTokenCount || 0) + (chunk.usageMetadata.thoughtsTokenCount || 0),\n\t\t\t\t\t\tcacheRead: chunk.usageMetadata.cachedContentTokenCount || 0,\n\t\t\t\t\t\tcacheWrite: 0,\n\t\t\t\t\t\ttotalTokens: chunk.usageMetadata.totalTokenCount || 0,\n\t\t\t\t\t\tcost: {\n\t\t\t\t\t\t\tinput: 0,\n\t\t\t\t\t\t\toutput: 0,\n\t\t\t\t\t\t\tcacheRead: 0,\n\t\t\t\t\t\t\tcacheWrite: 0,\n\t\t\t\t\t\t\ttotal: 0,\n\t\t\t\t\t\t},\n\t\t\t\t\t};\n\t\t\t\t\tcalculateCost(model, output.usage);\n\t\t\t\t}\n\t\t\t}\n\n\t\t\tif (currentBlock) {\n\t\t\t\tif (currentBlock.type === \"text\") {\n\t\t\t\t\tstream.push({\n\t\t\t\t\t\ttype: \"text_end\",\n\t\t\t\t\t\tcontentIndex: blockIndex(),\n\t\t\t\t\t\tcontent: currentBlock.text,\n\t\t\t\t\t\tpartial: output,\n\t\t\t\t\t});\n\t\t\t\t} else {\n\t\t\t\t\tstream.push({\n\t\t\t\t\t\ttype: \"thinking_end\",\n\t\t\t\t\t\tcontentIndex: blockIndex(),\n\t\t\t\t\t\tcontent: currentBlock.thinking,\n\t\t\t\t\t\tpartial: output,\n\t\t\t\t\t});\n\t\t\t\t}\n\t\t\t}\n\n\t\t\tif (options?.signal?.aborted) {\n\t\t\t\tthrow new Error(\"Request was aborted\");\n\t\t\t}\n\n\t\t\tif (output.stopReason === \"aborted\" || output.stopReason === \"error\") {\n\t\t\t\tthrow new Error(\"An unknown error occurred\");\n\t\t\t}\n\n\t\t\tstream.push({ type: \"done\", reason: output.stopReason, message: output });\n\t\t\tstream.end();\n\t\t} catch (error) {\n\t\t\t// Remove internal index property used during streaming\n\t\t\tfor (const block of output.content) {\n\t\t\t\tif (\"index\" in block) {\n\t\t\t\t\tdelete (block as { index?: number }).index;\n\t\t\t\t}\n\t\t\t}\n\t\t\toutput.stopReason = options?.signal?.aborted ? \"aborted\" : \"error\";\n\t\t\toutput.errorMessage = error instanceof Error ? error.message : JSON.stringify(error);\n\t\t\tstream.push({ type: \"error\", reason: output.stopReason, error: output });\n\t\t\tstream.end();\n\t\t}\n\t})();\n\n\treturn stream;\n};\n\nexport const streamSimpleGoogle: StreamFunction<\"google-generative-ai\", SimpleStreamOptions> = (\n\tmodel: Model<\"google-generative-ai\">,\n\tcontext: Context,\n\toptions?: SimpleStreamOptions,\n): AssistantMessageEventStream => {\n\tconst apiKey = options?.apiKey || getEnvApiKey(model.provider);\n\tif (!apiKey) {\n\t\tthrow new Error(`No API key for provider: ${model.provider}`);\n\t}\n\n\tconst base = buildBaseOptions(model, options, apiKey);\n\tif (!options?.reasoning) {\n\t\treturn streamGoogle(model, context, { ...base, thinking: { enabled: false } } satisfies GoogleOptions);\n\t}\n\n\tconst effort = clampReasoning(options.reasoning)!;\n\tconst googleModel = model as Model<\"google-generative-ai\">;\n\n\tif (isGemini3ProModel(googleModel) || isGemini3FlashModel(googleModel) || isGemma4Model(googleModel)) {\n\t\treturn streamGoogle(model, context, {\n\t\t\t...base,\n\t\t\tthinking: {\n\t\t\t\tenabled: true,\n\t\t\t\tlevel: getThinkingLevel(effort, googleModel),\n\t\t\t},\n\t\t} satisfies GoogleOptions);\n\t}\n\n\treturn streamGoogle(model, context, {\n\t\t...base,\n\t\tthinking: {\n\t\t\tenabled: true,\n\t\t\tbudgetTokens: getGoogleBudget(googleModel, effort, options.thinkingBudgets),\n\t\t},\n\t} satisfies GoogleOptions);\n};\n\nfunction createClient(\n\tmodel: Model<\"google-generative-ai\">,\n\tapiKey?: string,\n\toptionsHeaders?: Record<string, string>,\n): GoogleGenAI {\n\tconst httpOptions: { baseUrl?: string; apiVersion?: string; headers?: Record<string, string> } = {};\n\tif (model.baseUrl) {\n\t\thttpOptions.baseUrl = model.baseUrl;\n\t\thttpOptions.apiVersion = \"\"; // baseUrl already includes version path, don't append\n\t}\n\tif (model.headers || optionsHeaders) {\n\t\thttpOptions.headers = { ...model.headers, ...optionsHeaders };\n\t}\n\n\treturn new GoogleGenAI({\n\t\tapiKey,\n\t\thttpOptions: Object.keys(httpOptions).length > 0 ? httpOptions : undefined,\n\t});\n}\n\nfunction buildParams(\n\tmodel: Model<\"google-generative-ai\">,\n\tcontext: Context,\n\toptions: GoogleOptions = {},\n): GenerateContentParameters {\n\tconst contents = convertMessages(model, context);\n\n\tconst generationConfig: GenerateContentConfig = {};\n\tif (options.temperature !== undefined) {\n\t\tgenerationConfig.temperature = options.temperature;\n\t}\n\tif (options.maxTokens !== undefined) {\n\t\tgenerationConfig.maxOutputTokens = options.maxTokens;\n\t}\n\n\tconst config: GenerateContentConfig = {\n\t\t...(Object.keys(generationConfig).length > 0 && generationConfig),\n\t\t...(context.systemPrompt && { systemInstruction: sanitizeSurrogates(context.systemPrompt) }),\n\t\t...(context.tools && context.tools.length > 0 && { tools: convertTools(context.tools) }),\n\t};\n\n\tif (context.tools && context.tools.length > 0 && options.toolChoice) {\n\t\tconfig.toolConfig = {\n\t\t\tfunctionCallingConfig: {\n\t\t\t\tmode: mapToolChoice(options.toolChoice),\n\t\t\t},\n\t\t};\n\t} else {\n\t\tconfig.toolConfig = undefined;\n\t}\n\n\tif (options.thinking?.enabled && model.reasoning) {\n\t\tconst thinkingConfig: ThinkingConfig = { includeThoughts: true };\n\t\tif (options.thinking.level !== undefined) {\n\t\t\t// Cast to any since our GoogleThinkingLevel mirrors Google's ThinkingLevel enum values\n\t\t\tthinkingConfig.thinkingLevel = options.thinking.level as any;\n\t\t} else if (options.thinking.budgetTokens !== undefined) {\n\t\t\tthinkingConfig.thinkingBudget = options.thinking.budgetTokens;\n\t\t}\n\t\tconfig.thinkingConfig = thinkingConfig;\n\t} else if (model.reasoning && options.thinking && !options.thinking.enabled) {\n\t\tconfig.thinkingConfig = getDisabledThinkingConfig(model);\n\t}\n\n\tif (options.signal) {\n\t\tif (options.signal.aborted) {\n\t\t\tthrow new Error(\"Request aborted\");\n\t\t}\n\t\tconfig.abortSignal = options.signal;\n\t}\n\n\tconst params: GenerateContentParameters = {\n\t\tmodel: model.id,\n\t\tcontents,\n\t\tconfig,\n\t};\n\n\treturn params;\n}\n\ntype ClampedThinkingLevel = Exclude<ThinkingLevel, \"xhigh\">;\n\nfunction isGemma4Model(model: Model<\"google-generative-ai\">): boolean {\n\treturn /gemma-?4/.test(model.id.toLowerCase());\n}\n\nfunction isGemini3ProModel(model: Model<\"google-generative-ai\">): boolean {\n\treturn /gemini-3(?:\\.\\d+)?-pro/.test(model.id.toLowerCase());\n}\n\nfunction isGemini3FlashModel(model: Model<\"google-generative-ai\">): boolean {\n\treturn /gemini-3(?:\\.\\d+)?-flash/.test(model.id.toLowerCase());\n}\n\nfunction getDisabledThinkingConfig(model: Model<\"google-generative-ai\">): ThinkingConfig {\n\t// Google docs: Gemini 3.1 Pro cannot disable thinking, and Gemini 3 Flash / Flash-Lite\n\t// do not support full thinking-off either. For Gemini 3 models, use the lowest supported\n\t// thinkingLevel without includeThoughts so hidden thinking remains invisible to pi.\n\tif (isGemini3ProModel(model)) {\n\t\treturn { thinkingLevel: \"LOW\" as any };\n\t}\n\tif (isGemini3FlashModel(model)) {\n\t\treturn { thinkingLevel: \"MINIMAL\" as any };\n\t}\n\tif (isGemma4Model(model)) {\n\t\treturn { thinkingLevel: \"MINIMAL\" as any };\n\t}\n\n\t// Gemini 2.x supports disabling via thinkingBudget = 0.\n\treturn { thinkingBudget: 0 };\n}\n\nfunction getThinkingLevel(effort: ClampedThinkingLevel, model: Model<\"google-generative-ai\">): GoogleThinkingLevel {\n\tif (isGemini3ProModel(model)) {\n\t\tswitch (effort) {\n\t\t\tcase \"minimal\":\n\t\t\tcase \"low\":\n\t\t\t\treturn \"LOW\";\n\t\t\tcase \"medium\":\n\t\t\tcase \"high\":\n\t\t\t\treturn \"HIGH\";\n\t\t}\n\t}\n\tif (isGemma4Model(model)) {\n\t\tswitch (effort) {\n\t\t\tcase \"minimal\":\n\t\t\tcase \"low\":\n\t\t\t\treturn \"MINIMAL\";\n\t\t\tcase \"medium\":\n\t\t\tcase \"high\":\n\t\t\t\treturn \"HIGH\";\n\t\t}\n\t}\n\tswitch (effort) {\n\t\tcase \"minimal\":\n\t\t\treturn \"MINIMAL\";\n\t\tcase \"low\":\n\t\t\treturn \"LOW\";\n\t\tcase \"medium\":\n\t\t\treturn \"MEDIUM\";\n\t\tcase \"high\":\n\t\t\treturn \"HIGH\";\n\t}\n}\n\nfunction getGoogleBudget(\n\tmodel: Model<\"google-generative-ai\">,\n\teffort: ClampedThinkingLevel,\n\tcustomBudgets?: ThinkingBudgets,\n): number {\n\tif (customBudgets?.[effort] !== undefined) {\n\t\treturn customBudgets[effort]!;\n\t}\n\n\tif (model.id.includes(\"2.5-pro\")) {\n\t\tconst budgets: Record<ClampedThinkingLevel, number> = {\n\t\t\tminimal: 128,\n\t\t\tlow: 2048,\n\t\t\tmedium: 8192,\n\t\t\thigh: 32768,\n\t\t};\n\t\treturn budgets[effort];\n\t}\n\n\tif (model.id.includes(\"2.5-flash-lite\")) {\n\t\tconst budgets: Record<ClampedThinkingLevel, number> = {\n\t\t\tminimal: 512,\n\t\t\tlow: 2048,\n\t\t\tmedium: 8192,\n\t\t\thigh: 24576,\n\t\t};\n\t\treturn budgets[effort];\n\t}\n\n\tif (model.id.includes(\"2.5-flash\")) {\n\t\tconst budgets: Record<ClampedThinkingLevel, number> = {\n\t\t\tminimal: 128,\n\t\t\tlow: 2048,\n\t\t\tmedium: 8192,\n\t\t\thigh: 24576,\n\t\t};\n\t\treturn budgets[effort];\n\t}\n\n\treturn -1;\n}\n"]}

package/dist/providers/openai-completions.d.ts.map

@@ -1 +1 @@
-{"version":3,"file":"openai-completions.d.ts","sourceRoot":"","sources":["../../src/providers/openai-completions.ts"],"names":[],"mappings":"AACA,OAAO,KAAK,EAMX,0BAA0B,EAE1B,MAAM,sCAAsC,CAAC;AAG9C,OAAO,KAAK,EAEX,OAAO,EAEP,KAAK,EACL,uBAAuB,EACvB,mBAAmB,EAEnB,cAAc,EACd,aAAa,EAMb,MAAM,aAAa,CAAC;AA2BrB,MAAM,WAAW,wBAAyB,SAAQ,aAAa;IAC9D,UAAU,CAAC,EAAE,MAAM,GAAG,MAAM,GAAG,UAAU,GAAG;QAAE,IAAI,EAAE,UAAU,CAAC;QAAC,QAAQ,EAAE;YAAE,IAAI,EAAE,MAAM,CAAA;SAAE,CAAA;KAAE,CAAC;IAC7F,eAAe,CAAC,EAAE,SAAS,GAAG,KAAK,GAAG,QAAQ,GAAG,MAAM,GAAG,OAAO,CAAC;CAClE;AAED,eAAO,MAAM,uBAAuB,EAAE,cAAc,CAAC,oBAAoB,EAAE,wBAAwB,CAoPlG,CAAC;AAEF,eAAO,MAAM,6BAA6B,EAAE,cAAc,CAAC,oBAAoB,EAAE,mBAAmB,CAmBnG,CAAC;AAqKF,wBAAgB,eAAe,CAC9B,KAAK,EAAE,KAAK,CAAC,oBAAoB,CAAC,EAClC,OAAO,EAAE,OAAO,EAChB,MAAM,EAAE,QAAQ,CAAC,uBAAuB,CAAC,GACvC,0BAA0B,EAAE,CA6N9B","sourcesContent":["import OpenAI from \"openai\";\nimport type {\n\tChatCompletionAssistantMessageParam,\n\tChatCompletionChunk,\n\tChatCompletionContentPart,\n\tChatCompletionContentPartImage,\n\tChatCompletionContentPartText,\n\tChatCompletionMessageParam,\n\tChatCompletionToolMessageParam,\n} from \"openai/resources/chat/completions.js\";\nimport { getEnvApiKey } from \"../env-api-keys.js\";\nimport { calculateCost, supportsXhigh } from \"../models.js\";\nimport type {\n\tAssistantMessage,\n\tContext,\n\tMessage,\n\tModel,\n\tOpenAICompletionsCompat,\n\tSimpleStreamOptions,\n\tStopReason,\n\tStreamFunction,\n\tStreamOptions,\n\tTextContent,\n\tThinkingContent,\n\tTool,\n\tToolCall,\n\tToolResultMessage,\n} from \"../types.js\";\nimport { AssistantMessageEventStream } from \"../utils/event-stream.js\";\nimport { parseStreamingJson } from \"../utils/json-parse.js\";\nimport { sanitizeSurrogates } from \"../utils/sanitize-unicode.js\";\nimport { buildCopilotDynamicHeaders, hasCopilotVisionInput } from \"./github-copilot-headers.js\";\nimport { buildBaseOptions, clampReasoning } from \"./simple-options.js\";\nimport { transformMessages } from \"./transform-messages.js\";\n\n/**\n * Check if conversation messages contain tool calls or tool results.\n * This is needed because Anthropic (via proxy) requires the tools param\n * to be present when messages include tool_calls or tool role messages.\n */\nfunction hasToolHistory(messages: Message[]): boolean {\n\tfor (const msg of messages) {\n\t\tif (msg.role === \"toolResult\") {\n\t\t\treturn true;\n\t\t}\n\t\tif (msg.role === \"assistant\") {\n\t\t\tif (msg.content.some((block) => block.type === \"toolCall\")) {\n\t\t\t\treturn true;\n\t\t\t}\n\t\t}\n\t}\n\treturn false;\n}\n\nexport interface OpenAICompletionsOptions extends StreamOptions {\n\ttoolChoice?: \"auto\" | \"none\" | \"required\" | { type: \"function\"; function: { name: string } };\n\treasoningEffort?: \"minimal\" | \"low\" | \"medium\" | \"high\" | \"xhigh\";\n}\n\nexport const streamOpenAICompletions: StreamFunction<\"openai-completions\", OpenAICompletionsOptions> = (\n\tmodel: Model<\"openai-completions\">,\n\tcontext: Context,\n\toptions?: OpenAICompletionsOptions,\n): AssistantMessageEventStream => {\n\tconst stream = new AssistantMessageEventStream();\n\n\t(async () => {\n\t\tconst output: AssistantMessage = {\n\t\t\trole: \"assistant\",\n\t\t\tcontent: [],\n\t\t\tapi: model.api,\n\t\t\tprovider: model.provider,\n\t\t\tmodel: model.id,\n\t\t\tusage: {\n\t\t\t\tinput: 0,\n\t\t\t\toutput: 0,\n\t\t\t\tcacheRead: 0,\n\t\t\t\tcacheWrite: 0,\n\t\t\t\ttotalTokens: 0,\n\t\t\t\tcost: { input: 0, output: 0, cacheRead: 0, cacheWrite: 0, total: 0 },\n\t\t\t},\n\t\t\tstopReason: \"stop\",\n\t\t\ttimestamp: Date.now(),\n\t\t};\n\n\t\ttry {\n\t\t\tconst apiKey = options?.apiKey || getEnvApiKey(model.provider) || \"\";\n\t\t\tconst client = createClient(model, context, apiKey, options?.headers);\n\t\t\tlet params = buildParams(model, context, options);\n\t\t\tconst nextParams = await options?.onPayload?.(params, model);\n\t\t\tif (nextParams !== undefined) {\n\t\t\t\tparams = nextParams as OpenAI.Chat.Completions.ChatCompletionCreateParamsStreaming;\n\t\t\t}\n\t\t\tconst openaiStream = await client.chat.completions.create(params, { signal: options?.signal });\n\t\t\tstream.push({ type: \"start\", partial: output });\n\n\t\t\tlet currentBlock: TextContent | ThinkingContent | (ToolCall & { partialArgs?: string }) | null = null;\n\t\t\tconst blocks = output.content;\n\t\t\tconst blockIndex = () => blocks.length - 1;\n\t\t\tconst finishCurrentBlock = (block?: typeof currentBlock) => {\n\t\t\t\tif (block) {\n\t\t\t\t\tif (block.type === \"text\") {\n\t\t\t\t\t\tstream.push({\n\t\t\t\t\t\t\ttype: \"text_end\",\n\t\t\t\t\t\t\tcontentIndex: blockIndex(),\n\t\t\t\t\t\t\tcontent: block.text,\n\t\t\t\t\t\t\tpartial: output,\n\t\t\t\t\t\t});\n\t\t\t\t\t} else if (block.type === \"thinking\") {\n\t\t\t\t\t\tstream.push({\n\t\t\t\t\t\t\ttype: \"thinking_end\",\n\t\t\t\t\t\t\tcontentIndex: blockIndex(),\n\t\t\t\t\t\t\tcontent: block.thinking,\n\t\t\t\t\t\t\tpartial: output,\n\t\t\t\t\t\t});\n\t\t\t\t\t} else if (block.type === \"toolCall\") {\n\t\t\t\t\t\tblock.arguments = parseStreamingJson(block.partialArgs);\n\t\t\t\t\t\tdelete block.partialArgs;\n\t\t\t\t\t\tstream.push({\n\t\t\t\t\t\t\ttype: \"toolcall_end\",\n\t\t\t\t\t\t\tcontentIndex: blockIndex(),\n\t\t\t\t\t\t\ttoolCall: block,\n\t\t\t\t\t\t\tpartial: output,\n\t\t\t\t\t\t});\n\t\t\t\t\t}\n\t\t\t\t}\n\t\t\t};\n\n\t\t\tfor await (const chunk of openaiStream) {\n\t\t\t\tif (!chunk || typeof chunk !== \"object\") continue;\n\n\t\t\t\t// OpenAI documents ChatCompletionChunk.id as the unique chat completion identifier,\n\t\t\t\t// and each chunk in a streamed completion carries the same id.\n\t\t\t\toutput.responseId ||= chunk.id;\n\t\t\t\tif (chunk.usage) {\n\t\t\t\t\toutput.usage = parseChunkUsage(chunk.usage, model);\n\t\t\t\t}\n\n\t\t\t\tconst choice = Array.isArray(chunk.choices) ? chunk.choices[0] : undefined;\n\t\t\t\tif (!choice) continue;\n\n\t\t\t\t// Fallback: some providers (e.g., Moonshot) return usage\n\t\t\t\t// in choice.usage instead of the standard chunk.usage\n\t\t\t\tif (!chunk.usage && (choice as any).usage) {\n\t\t\t\t\toutput.usage = parseChunkUsage((choice as any).usage, model);\n\t\t\t\t}\n\n\t\t\t\tif (choice.finish_reason) {\n\t\t\t\t\tconst finishReasonResult = mapStopReason(choice.finish_reason);\n\t\t\t\t\toutput.stopReason = finishReasonResult.stopReason;\n\t\t\t\t\tif (finishReasonResult.errorMessage) {\n\t\t\t\t\t\toutput.errorMessage = finishReasonResult.errorMessage;\n\t\t\t\t\t}\n\t\t\t\t}\n\n\t\t\t\tif (choice.delta) {\n\t\t\t\t\tif (\n\t\t\t\t\t\tchoice.delta.content !== null &&\n\t\t\t\t\t\tchoice.delta.content !== undefined &&\n\t\t\t\t\t\tchoice.delta.content.length > 0\n\t\t\t\t\t) {\n\t\t\t\t\t\tif (!currentBlock || currentBlock.type !== \"text\") {\n\t\t\t\t\t\t\tfinishCurrentBlock(currentBlock);\n\t\t\t\t\t\t\tcurrentBlock = { type: \"text\", text: \"\" };\n\t\t\t\t\t\t\toutput.content.push(currentBlock);\n\t\t\t\t\t\t\tstream.push({ type: \"text_start\", contentIndex: blockIndex(), partial: output });\n\t\t\t\t\t\t}\n\n\t\t\t\t\t\tif (currentBlock.type === \"text\") {\n\t\t\t\t\t\t\tcurrentBlock.text += choice.delta.content;\n\t\t\t\t\t\t\tstream.push({\n\t\t\t\t\t\t\t\ttype: \"text_delta\",\n\t\t\t\t\t\t\t\tcontentIndex: blockIndex(),\n\t\t\t\t\t\t\t\tdelta: choice.delta.content,\n\t\t\t\t\t\t\t\tpartial: output,\n\t\t\t\t\t\t\t});\n\t\t\t\t\t\t}\n\t\t\t\t\t}\n\n\t\t\t\t\t// Some endpoints return reasoning in reasoning_content (llama.cpp),\n\t\t\t\t\t// or reasoning (other openai compatible endpoints)\n\t\t\t\t\t// Use the first non-empty reasoning field to avoid duplication\n\t\t\t\t\t// (e.g., chutes.ai returns both reasoning_content and reasoning with same content)\n\t\t\t\t\tconst reasoningFields = [\"reasoning_content\", \"reasoning\", \"reasoning_text\"];\n\t\t\t\t\tlet foundReasoningField: string | null = null;\n\t\t\t\t\tfor (const field of reasoningFields) {\n\t\t\t\t\t\tif (\n\t\t\t\t\t\t\t(choice.delta as any)[field] !== null &&\n\t\t\t\t\t\t\t(choice.delta as any)[field] !== undefined &&\n\t\t\t\t\t\t\t(choice.delta as any)[field].length > 0\n\t\t\t\t\t\t) {\n\t\t\t\t\t\t\tif (!foundReasoningField) {\n\t\t\t\t\t\t\t\tfoundReasoningField = field;\n\t\t\t\t\t\t\t\tbreak;\n\t\t\t\t\t\t\t}\n\t\t\t\t\t\t}\n\t\t\t\t\t}\n\n\t\t\t\t\tif (foundReasoningField) {\n\t\t\t\t\t\tif (!currentBlock || currentBlock.type !== \"thinking\") {\n\t\t\t\t\t\t\tfinishCurrentBlock(currentBlock);\n\t\t\t\t\t\t\tcurrentBlock = {\n\t\t\t\t\t\t\t\ttype: \"thinking\",\n\t\t\t\t\t\t\t\tthinking: \"\",\n\t\t\t\t\t\t\t\tthinkingSignature: foundReasoningField,\n\t\t\t\t\t\t\t};\n\t\t\t\t\t\t\toutput.content.push(currentBlock);\n\t\t\t\t\t\t\tstream.push({ type: \"thinking_start\", contentIndex: blockIndex(), partial: output });\n\t\t\t\t\t\t}\n\n\t\t\t\t\t\tif (currentBlock.type === \"thinking\") {\n\t\t\t\t\t\t\tconst delta = (choice.delta as any)[foundReasoningField];\n\t\t\t\t\t\t\tcurrentBlock.thinking += delta;\n\t\t\t\t\t\t\tstream.push({\n\t\t\t\t\t\t\t\ttype: \"thinking_delta\",\n\t\t\t\t\t\t\t\tcontentIndex: blockIndex(),\n\t\t\t\t\t\t\t\tdelta,\n\t\t\t\t\t\t\t\tpartial: output,\n\t\t\t\t\t\t\t});\n\t\t\t\t\t\t}\n\t\t\t\t\t}\n\n\t\t\t\t\tif (choice?.delta?.tool_calls) {\n\t\t\t\t\t\tfor (const toolCall of choice.delta.tool_calls) {\n\t\t\t\t\t\t\tif (\n\t\t\t\t\t\t\t\t!currentBlock ||\n\t\t\t\t\t\t\t\tcurrentBlock.type !== \"toolCall\" ||\n\t\t\t\t\t\t\t\t(toolCall.id && currentBlock.id !== toolCall.id)\n\t\t\t\t\t\t\t) {\n\t\t\t\t\t\t\t\tfinishCurrentBlock(currentBlock);\n\t\t\t\t\t\t\t\tcurrentBlock = {\n\t\t\t\t\t\t\t\t\ttype: \"toolCall\",\n\t\t\t\t\t\t\t\t\tid: toolCall.id || \"\",\n\t\t\t\t\t\t\t\t\tname: toolCall.function?.name || \"\",\n\t\t\t\t\t\t\t\t\targuments: {},\n\t\t\t\t\t\t\t\t\tpartialArgs: \"\",\n\t\t\t\t\t\t\t\t};\n\t\t\t\t\t\t\t\toutput.content.push(currentBlock);\n\t\t\t\t\t\t\t\tstream.push({ type: \"toolcall_start\", contentIndex: blockIndex(), partial: output });\n\t\t\t\t\t\t\t}\n\n\t\t\t\t\t\t\tif (currentBlock.type === \"toolCall\") {\n\t\t\t\t\t\t\t\tif (toolCall.id) currentBlock.id = toolCall.id;\n\t\t\t\t\t\t\t\tif (toolCall.function?.name) currentBlock.name = toolCall.function.name;\n\t\t\t\t\t\t\t\tlet delta = \"\";\n\t\t\t\t\t\t\t\tif (toolCall.function?.arguments) {\n\t\t\t\t\t\t\t\t\tdelta = toolCall.function.arguments;\n\t\t\t\t\t\t\t\t\tcurrentBlock.partialArgs += toolCall.function.arguments;\n\t\t\t\t\t\t\t\t\tcurrentBlock.arguments = parseStreamingJson(currentBlock.partialArgs);\n\t\t\t\t\t\t\t\t}\n\t\t\t\t\t\t\t\tstream.push({\n\t\t\t\t\t\t\t\t\ttype: \"toolcall_delta\",\n\t\t\t\t\t\t\t\t\tcontentIndex: blockIndex(),\n\t\t\t\t\t\t\t\t\tdelta,\n\t\t\t\t\t\t\t\t\tpartial: output,\n\t\t\t\t\t\t\t\t});\n\t\t\t\t\t\t\t}\n\t\t\t\t\t\t}\n\t\t\t\t\t}\n\n\t\t\t\t\tconst reasoningDetails = (choice.delta as any).reasoning_details;\n\t\t\t\t\tif (reasoningDetails && Array.isArray(reasoningDetails)) {\n\t\t\t\t\t\tfor (const detail of reasoningDetails) {\n\t\t\t\t\t\t\tif (detail.type === \"reasoning.encrypted\" && detail.id && detail.data) {\n\t\t\t\t\t\t\t\tconst matchingToolCall = output.content.find(\n\t\t\t\t\t\t\t\t\t(b) => b.type === \"toolCall\" && b.id === detail.id,\n\t\t\t\t\t\t\t\t) as ToolCall | undefined;\n\t\t\t\t\t\t\t\tif (matchingToolCall) {\n\t\t\t\t\t\t\t\t\tmatchingToolCall.thoughtSignature = JSON.stringify(detail);\n\t\t\t\t\t\t\t\t}\n\t\t\t\t\t\t\t}\n\t\t\t\t\t\t}\n\t\t\t\t\t}\n\t\t\t\t}\n\t\t\t}\n\n\t\t\tfinishCurrentBlock(currentBlock);\n\t\t\tif (options?.signal?.aborted) {\n\t\t\t\tthrow new Error(\"Request was aborted\");\n\t\t\t}\n\n\t\t\tif (output.stopReason === \"aborted\") {\n\t\t\t\tthrow new Error(\"Request was aborted\");\n\t\t\t}\n\t\t\tif (output.stopReason === \"error\") {\n\t\t\t\tthrow new Error(output.errorMessage || \"Provider returned an error stop reason\");\n\t\t\t}\n\n\t\t\tstream.push({ type: \"done\", reason: output.stopReason, message: output });\n\t\t\tstream.end();\n\t\t} catch (error) {\n\t\t\tfor (const block of output.content) delete (block as any).index;\n\t\t\toutput.stopReason = options?.signal?.aborted ? \"aborted\" : \"error\";\n\t\t\toutput.errorMessage = error instanceof Error ? error.message : JSON.stringify(error);\n\t\t\t// Some providers via OpenRouter give additional information in this field.\n\t\t\tconst rawMetadata = (error as any)?.error?.metadata?.raw;\n\t\t\tif (rawMetadata) output.errorMessage += `\\n${rawMetadata}`;\n\t\t\tstream.push({ type: \"error\", reason: output.stopReason, error: output });\n\t\t\tstream.end();\n\t\t}\n\t})();\n\n\treturn stream;\n};\n\nexport const streamSimpleOpenAICompletions: StreamFunction<\"openai-completions\", SimpleStreamOptions> = (\n\tmodel: Model<\"openai-completions\">,\n\tcontext: Context,\n\toptions?: SimpleStreamOptions,\n): AssistantMessageEventStream => {\n\tconst apiKey = options?.apiKey || getEnvApiKey(model.provider);\n\tif (!apiKey) {\n\t\tthrow new Error(`No API key for provider: ${model.provider}`);\n\t}\n\n\tconst base = buildBaseOptions(model, options, apiKey);\n\tconst reasoningEffort = supportsXhigh(model) ? options?.reasoning : clampReasoning(options?.reasoning);\n\tconst toolChoice = (options as OpenAICompletionsOptions | undefined)?.toolChoice;\n\n\treturn streamOpenAICompletions(model, context, {\n\t\t...base,\n\t\treasoningEffort,\n\t\ttoolChoice,\n\t} satisfies OpenAICompletionsOptions);\n};\n\nfunction createClient(\n\tmodel: Model<\"openai-completions\">,\n\tcontext: Context,\n\tapiKey?: string,\n\toptionsHeaders?: Record<string, string>,\n) {\n\tif (!apiKey) {\n\t\tif (!process.env.OPENAI_API_KEY) {\n\t\t\tthrow new Error(\n\t\t\t\t\"OpenAI API key is required. Set OPENAI_API_KEY environment variable or pass it as an argument.\",\n\t\t\t);\n\t\t}\n\t\tapiKey = process.env.OPENAI_API_KEY;\n\t}\n\n\tconst headers = { ...model.headers };\n\tif (model.provider === \"github-copilot\") {\n\t\tconst hasImages = hasCopilotVisionInput(context.messages);\n\t\tconst copilotHeaders = buildCopilotDynamicHeaders({\n\t\t\tmessages: context.messages,\n\t\t\thasImages,\n\t\t});\n\t\tObject.assign(headers, copilotHeaders);\n\t}\n\n\t// Merge options headers last so they can override defaults\n\tif (optionsHeaders) {\n\t\tObject.assign(headers, optionsHeaders);\n\t}\n\n\treturn new OpenAI({\n\t\tapiKey,\n\t\tbaseURL: model.baseUrl,\n\t\tdangerouslyAllowBrowser: true,\n\t\tdefaultHeaders: headers,\n\t});\n}\n\nfunction buildParams(model: Model<\"openai-completions\">, context: Context, options?: OpenAICompletionsOptions) {\n\tconst compat = getCompat(model);\n\tconst messages = convertMessages(model, context, compat);\n\tmaybeAddOpenRouterAnthropicCacheControl(model, messages);\n\n\tconst params: OpenAI.Chat.Completions.ChatCompletionCreateParamsStreaming = {\n\t\tmodel: model.id,\n\t\tmessages,\n\t\tstream: true,\n\t};\n\n\tif (compat.supportsUsageInStreaming !== false) {\n\t\t(params as any).stream_options = { include_usage: true };\n\t}\n\n\tif (compat.supportsStore) {\n\t\tparams.store = false;\n\t}\n\n\tif (options?.maxTokens) {\n\t\tif (compat.maxTokensField === \"max_tokens\") {\n\t\t\t(params as any).max_tokens = options.maxTokens;\n\t\t} else {\n\t\t\tparams.max_completion_tokens = options.maxTokens;\n\t\t}\n\t}\n\n\tif (options?.temperature !== undefined) {\n\t\tparams.temperature = options.temperature;\n\t}\n\n\tif (context.tools) {\n\t\tparams.tools = convertTools(context.tools, compat);\n\t\tif (compat.zaiToolStream) {\n\t\t\t(params as any).tool_stream = true;\n\t\t}\n\t} else if (hasToolHistory(context.messages)) {\n\t\t// Anthropic (via LiteLLM/proxy) requires tools param when conversation has tool_calls/tool_results\n\t\tparams.tools = [];\n\t}\n\n\tif (options?.toolChoice) {\n\t\tparams.tool_choice = options.toolChoice;\n\t}\n\n\tif (compat.thinkingFormat === \"zai\" && model.reasoning) {\n\t\t(params as any).enable_thinking = !!options?.reasoningEffort;\n\t} else if (compat.thinkingFormat === \"qwen\" && model.reasoning) {\n\t\t(params as any).enable_thinking = !!options?.reasoningEffort;\n\t} else if (compat.thinkingFormat === \"qwen-chat-template\" && model.reasoning) {\n\t\t(params as any).chat_template_kwargs = { enable_thinking: !!options?.reasoningEffort };\n\t} else if (compat.thinkingFormat === \"openrouter\" && model.reasoning) {\n\t\t// OpenRouter normalizes reasoning across providers via a nested reasoning object.\n\t\tconst openRouterParams = params as typeof params & { reasoning?: { effort?: string } };\n\t\tif (options?.reasoningEffort) {\n\t\t\topenRouterParams.reasoning = {\n\t\t\t\teffort: mapReasoningEffort(options.reasoningEffort, compat.reasoningEffortMap),\n\t\t\t};\n\t\t} else {\n\t\t\topenRouterParams.reasoning = { effort: \"none\" };\n\t\t}\n\t} else if (options?.reasoningEffort && model.reasoning && compat.supportsReasoningEffort) {\n\t\t// OpenAI-style reasoning_effort\n\t\t(params as any).reasoning_effort = mapReasoningEffort(options.reasoningEffort, compat.reasoningEffortMap);\n\t}\n\n\t// OpenRouter provider routing preferences\n\tif (model.baseUrl.includes(\"openrouter.ai\") && model.compat?.openRouterRouting) {\n\t\t(params as any).provider = model.compat.openRouterRouting;\n\t}\n\n\t// Vercel AI Gateway provider routing preferences\n\tif (model.baseUrl.includes(\"ai-gateway.vercel.sh\") && model.compat?.vercelGatewayRouting) {\n\t\tconst routing = model.compat.vercelGatewayRouting;\n\t\tif (routing.only || routing.order) {\n\t\t\tconst gatewayOptions: Record<string, string[]> = {};\n\t\t\tif (routing.only) gatewayOptions.only = routing.only;\n\t\t\tif (routing.order) gatewayOptions.order = routing.order;\n\t\t\t(params as any).providerOptions = { gateway: gatewayOptions };\n\t\t}\n\t}\n\n\treturn params;\n}\n\nfunction mapReasoningEffort(\n\teffort: NonNullable<OpenAICompletionsOptions[\"reasoningEffort\"]>,\n\treasoningEffortMap: Partial<Record<NonNullable<OpenAICompletionsOptions[\"reasoningEffort\"]>, string>>,\n): string {\n\treturn reasoningEffortMap[effort] ?? effort;\n}\n\nfunction maybeAddOpenRouterAnthropicCacheControl(\n\tmodel: Model<\"openai-completions\">,\n\tmessages: ChatCompletionMessageParam[],\n): void {\n\tif (model.provider !== \"openrouter\" || !model.id.startsWith(\"anthropic/\")) return;\n\n\t// Anthropic-style caching requires cache_control on a text part. Add a breakpoint\n\t// on the last user/assistant message (walking backwards until we find text content).\n\tfor (let i = messages.length - 1; i >= 0; i--) {\n\t\tconst msg = messages[i];\n\t\tif (msg.role !== \"user\" && msg.role !== \"assistant\") continue;\n\n\t\tconst content = msg.content;\n\t\tif (typeof content === \"string\") {\n\t\t\tmsg.content = [\n\t\t\t\tObject.assign({ type: \"text\" as const, text: content }, { cache_control: { type: \"ephemeral\" } }),\n\t\t\t];\n\t\t\treturn;\n\t\t}\n\n\t\tif (!Array.isArray(content)) continue;\n\n\t\t// Find last text part and add cache_control\n\t\tfor (let j = content.length - 1; j >= 0; j--) {\n\t\t\tconst part = content[j];\n\t\t\tif (part?.type === \"text\") {\n\t\t\t\tObject.assign(part, { cache_control: { type: \"ephemeral\" } });\n\t\t\t\treturn;\n\t\t\t}\n\t\t}\n\t}\n}\n\nexport function convertMessages(\n\tmodel: Model<\"openai-completions\">,\n\tcontext: Context,\n\tcompat: Required<OpenAICompletionsCompat>,\n): ChatCompletionMessageParam[] {\n\tconst params: ChatCompletionMessageParam[] = [];\n\n\tconst normalizeToolCallId = (id: string): string => {\n\t\t// Handle pipe-separated IDs from OpenAI Responses API\n\t\t// Format: {call_id}|{id} where {id} can be 400+ chars with special chars (+, /, =)\n\t\t// These come from providers like github-copilot, openai-codex, opencode\n\t\t// Extract just the call_id part and normalize it\n\t\tif (id.includes(\"|\")) {\n\t\t\tconst [callId] = id.split(\"|\");\n\t\t\t// Sanitize to allowed chars and truncate to 40 chars (OpenAI limit)\n\t\t\treturn callId.replace(/[^a-zA-Z0-9_-]/g, \"_\").slice(0, 40);\n\t\t}\n\n\t\tif (model.provider === \"openai\") return id.length > 40 ? id.slice(0, 40) : id;\n\t\treturn id;\n\t};\n\n\tconst transformedMessages = transformMessages(context.messages, model, (id) => normalizeToolCallId(id));\n\n\tif (context.systemPrompt) {\n\t\tconst useDeveloperRole = model.reasoning && compat.supportsDeveloperRole;\n\t\tconst role = useDeveloperRole ? \"developer\" : \"system\";\n\t\tparams.push({ role: role, content: sanitizeSurrogates(context.systemPrompt) });\n\t}\n\n\tlet lastRole: string | null = null;\n\n\tfor (let i = 0; i < transformedMessages.length; i++) {\n\t\tconst msg = transformedMessages[i];\n\t\t// Some providers don't allow user messages directly after tool results\n\t\t// Insert a synthetic assistant message to bridge the gap\n\t\tif (compat.requiresAssistantAfterToolResult && lastRole === \"toolResult\" && msg.role === \"user\") {\n\t\t\tparams.push({\n\t\t\t\trole: \"assistant\",\n\t\t\t\tcontent: \"I have processed the tool results.\",\n\t\t\t});\n\t\t}\n\n\t\tif (msg.role === \"user\") {\n\t\t\tif (typeof msg.content === \"string\") {\n\t\t\t\tparams.push({\n\t\t\t\t\trole: \"user\",\n\t\t\t\t\tcontent: sanitizeSurrogates(msg.content),\n\t\t\t\t});\n\t\t\t} else {\n\t\t\t\tconst content: ChatCompletionContentPart[] = msg.content.map((item): ChatCompletionContentPart => {\n\t\t\t\t\tif (item.type === \"text\") {\n\t\t\t\t\t\treturn {\n\t\t\t\t\t\t\ttype: \"text\",\n\t\t\t\t\t\t\ttext: sanitizeSurrogates(item.text),\n\t\t\t\t\t\t} satisfies ChatCompletionContentPartText;\n\t\t\t\t\t} else {\n\t\t\t\t\t\treturn {\n\t\t\t\t\t\t\ttype: \"image_url\",\n\t\t\t\t\t\t\timage_url: {\n\t\t\t\t\t\t\t\turl: `data:${item.mimeType};base64,${item.data}`,\n\t\t\t\t\t\t\t},\n\t\t\t\t\t\t} satisfies ChatCompletionContentPartImage;\n\t\t\t\t\t}\n\t\t\t\t});\n\t\t\t\tconst filteredContent = !model.input.includes(\"image\")\n\t\t\t\t\t? content.filter((c) => c.type !== \"image_url\")\n\t\t\t\t\t: content;\n\t\t\t\tif (filteredContent.length === 0) continue;\n\t\t\t\tparams.push({\n\t\t\t\t\trole: \"user\",\n\t\t\t\t\tcontent: filteredContent,\n\t\t\t\t});\n\t\t\t}\n\t\t} else if (msg.role === \"assistant\") {\n\t\t\t// Some providers don't accept null content, use empty string instead\n\t\t\tconst assistantMsg: ChatCompletionAssistantMessageParam = {\n\t\t\t\trole: \"assistant\",\n\t\t\t\tcontent: compat.requiresAssistantAfterToolResult ? \"\" : null,\n\t\t\t};\n\n\t\t\tconst textBlocks = msg.content.filter((b) => b.type === \"text\") as TextContent[];\n\t\t\t// Filter out empty text blocks to avoid API validation errors\n\t\t\tconst nonEmptyTextBlocks = textBlocks.filter((b) => b.text && b.text.trim().length > 0);\n\t\t\tif (nonEmptyTextBlocks.length > 0) {\n\t\t\t\t// Always send assistant content as a plain string (OpenAI Chat Completions\n\t\t\t\t// API standard format). Sending as an array of {type:\"text\", text:\"...\"}\n\t\t\t\t// objects is non-standard and causes some models (e.g. DeepSeek V3.2 via\n\t\t\t\t// NVIDIA NIM) to mirror the content-block structure literally in their\n\t\t\t\t// output, producing recursive nesting like [{'type':'text','text':'[{...}]'}].\n\t\t\t\tassistantMsg.content = nonEmptyTextBlocks.map((b) => sanitizeSurrogates(b.text)).join(\"\");\n\t\t\t}\n\n\t\t\t// Handle thinking blocks\n\t\t\tconst thinkingBlocks = msg.content.filter((b) => b.type === \"thinking\") as ThinkingContent[];\n\t\t\t// Filter out empty thinking blocks to avoid API validation errors\n\t\t\tconst nonEmptyThinkingBlocks = thinkingBlocks.filter((b) => b.thinking && b.thinking.trim().length > 0);\n\t\t\tif (nonEmptyThinkingBlocks.length > 0) {\n\t\t\t\tif (compat.requiresThinkingAsText) {\n\t\t\t\t\t// Convert thinking blocks to plain text (no tags to avoid model mimicking them)\n\t\t\t\t\tconst thinkingText = nonEmptyThinkingBlocks.map((b) => b.thinking).join(\"\\n\\n\");\n\t\t\t\t\tconst textContent = assistantMsg.content as Array<{ type: \"text\"; text: string }> | null;\n\t\t\t\t\tif (textContent) {\n\t\t\t\t\t\ttextContent.unshift({ type: \"text\", text: thinkingText });\n\t\t\t\t\t} else {\n\t\t\t\t\t\tassistantMsg.content = [{ type: \"text\", text: thinkingText }];\n\t\t\t\t\t}\n\t\t\t\t} else {\n\t\t\t\t\t// Use the signature from the first thinking block if available (for llama.cpp server + gpt-oss)\n\t\t\t\t\tconst signature = nonEmptyThinkingBlocks[0].thinkingSignature;\n\t\t\t\t\tif (signature && signature.length > 0) {\n\t\t\t\t\t\t(assistantMsg as any)[signature] = nonEmptyThinkingBlocks.map((b) => b.thinking).join(\"\\n\");\n\t\t\t\t\t}\n\t\t\t\t}\n\t\t\t}\n\n\t\t\tconst toolCalls = msg.content.filter((b) => b.type === \"toolCall\") as ToolCall[];\n\t\t\tif (toolCalls.length > 0) {\n\t\t\t\tassistantMsg.tool_calls = toolCalls.map((tc) => ({\n\t\t\t\t\tid: tc.id,\n\t\t\t\t\ttype: \"function\" as const,\n\t\t\t\t\tfunction: {\n\t\t\t\t\t\tname: tc.name,\n\t\t\t\t\t\targuments: JSON.stringify(tc.arguments),\n\t\t\t\t\t},\n\t\t\t\t}));\n\t\t\t\tconst reasoningDetails = toolCalls\n\t\t\t\t\t.filter((tc) => tc.thoughtSignature)\n\t\t\t\t\t.map((tc) => {\n\t\t\t\t\t\ttry {\n\t\t\t\t\t\t\treturn JSON.parse(tc.thoughtSignature!);\n\t\t\t\t\t\t} catch {\n\t\t\t\t\t\t\treturn null;\n\t\t\t\t\t\t}\n\t\t\t\t\t})\n\t\t\t\t\t.filter(Boolean);\n\t\t\t\tif (reasoningDetails.length > 0) {\n\t\t\t\t\t(assistantMsg as any).reasoning_details = reasoningDetails;\n\t\t\t\t}\n\t\t\t}\n\t\t\t// Skip assistant messages that have no content and no tool calls.\n\t\t\t// Some providers require \"either content or tool_calls, but not none\".\n\t\t\t// Other providers also don't accept empty assistant messages.\n\t\t\t// This handles aborted assistant responses that got no content.\n\t\t\tconst content = assistantMsg.content;\n\t\t\tconst hasContent =\n\t\t\t\tcontent !== null &&\n\t\t\t\tcontent !== undefined &&\n\t\t\t\t(typeof content === \"string\" ? content.length > 0 : content.length > 0);\n\t\t\tif (!hasContent && !assistantMsg.tool_calls) {\n\t\t\t\tcontinue;\n\t\t\t}\n\t\t\tparams.push(assistantMsg);\n\t\t} else if (msg.role === \"toolResult\") {\n\t\t\tconst imageBlocks: Array<{ type: \"image_url\"; image_url: { url: string } }> = [];\n\t\t\tlet j = i;\n\n\t\t\tfor (; j < transformedMessages.length && transformedMessages[j].role === \"toolResult\"; j++) {\n\t\t\t\tconst toolMsg = transformedMessages[j] as ToolResultMessage;\n\n\t\t\t\t// Extract text and image content\n\t\t\t\tconst textResult = toolMsg.content\n\t\t\t\t\t.filter((c) => c.type === \"text\")\n\t\t\t\t\t.map((c) => (c as any).text)\n\t\t\t\t\t.join(\"\\n\");\n\t\t\t\tconst hasImages = toolMsg.content.some((c) => c.type === \"image\");\n\n\t\t\t\t// Always send tool result with text (or placeholder if only images)\n\t\t\t\tconst hasText = textResult.length > 0;\n\t\t\t\t// Some providers require the 'name' field in tool results\n\t\t\t\tconst toolResultMsg: ChatCompletionToolMessageParam = {\n\t\t\t\t\trole: \"tool\",\n\t\t\t\t\tcontent: sanitizeSurrogates(hasText ? textResult : \"(see attached image)\"),\n\t\t\t\t\ttool_call_id: toolMsg.toolCallId,\n\t\t\t\t};\n\t\t\t\tif (compat.requiresToolResultName && toolMsg.toolName) {\n\t\t\t\t\t(toolResultMsg as any).name = toolMsg.toolName;\n\t\t\t\t}\n\t\t\t\tparams.push(toolResultMsg);\n\n\t\t\t\tif (hasImages && model.input.includes(\"image\")) {\n\t\t\t\t\tfor (const block of toolMsg.content) {\n\t\t\t\t\t\tif (block.type === \"image\") {\n\t\t\t\t\t\t\timageBlocks.push({\n\t\t\t\t\t\t\t\ttype: \"image_url\",\n\t\t\t\t\t\t\t\timage_url: {\n\t\t\t\t\t\t\t\t\turl: `data:${(block as any).mimeType};base64,${(block as any).data}`,\n\t\t\t\t\t\t\t\t},\n\t\t\t\t\t\t\t});\n\t\t\t\t\t\t}\n\t\t\t\t\t}\n\t\t\t\t}\n\t\t\t}\n\n\t\t\ti = j - 1;\n\n\t\t\tif (imageBlocks.length > 0) {\n\t\t\t\tif (compat.requiresAssistantAfterToolResult) {\n\t\t\t\t\tparams.push({\n\t\t\t\t\t\trole: \"assistant\",\n\t\t\t\t\t\tcontent: \"I have processed the tool results.\",\n\t\t\t\t\t});\n\t\t\t\t}\n\n\t\t\t\tparams.push({\n\t\t\t\t\trole: \"user\",\n\t\t\t\t\tcontent: [\n\t\t\t\t\t\t{\n\t\t\t\t\t\t\ttype: \"text\",\n\t\t\t\t\t\t\ttext: \"Attached image(s) from tool result:\",\n\t\t\t\t\t\t},\n\t\t\t\t\t\t...imageBlocks,\n\t\t\t\t\t],\n\t\t\t\t});\n\t\t\t\tlastRole = \"user\";\n\t\t\t} else {\n\t\t\t\tlastRole = \"toolResult\";\n\t\t\t}\n\t\t\tcontinue;\n\t\t}\n\n\t\tlastRole = msg.role;\n\t}\n\n\treturn params;\n}\n\nfunction convertTools(\n\ttools: Tool[],\n\tcompat: Required<OpenAICompletionsCompat>,\n): OpenAI.Chat.Completions.ChatCompletionTool[] {\n\treturn tools.map((tool) => ({\n\t\ttype: \"function\",\n\t\tfunction: {\n\t\t\tname: tool.name,\n\t\t\tdescription: tool.description,\n\t\t\tparameters: tool.parameters as any, // TypeBox already generates JSON Schema\n\t\t\t// Only include strict if provider supports it. Some reject unknown fields.\n\t\t\t...(compat.supportsStrictMode !== false && { strict: false }),\n\t\t},\n\t}));\n}\n\nfunction parseChunkUsage(\n\trawUsage: {\n\t\tprompt_tokens?: number;\n\t\tcompletion_tokens?: number;\n\t\tprompt_tokens_details?: { cached_tokens?: number };\n\t\tcompletion_tokens_details?: { reasoning_tokens?: number };\n\t},\n\tmodel: Model<\"openai-completions\">,\n): AssistantMessage[\"usage\"] {\n\tconst cachedTokens = rawUsage.prompt_tokens_details?.cached_tokens || 0;\n\tconst reasoningTokens = rawUsage.completion_tokens_details?.reasoning_tokens || 0;\n\t// OpenAI includes cached tokens in prompt_tokens, so subtract to get non-cached input\n\tconst input = (rawUsage.prompt_tokens || 0) - cachedTokens;\n\t// Compute totalTokens ourselves since we add reasoning_tokens to output\n\t// and some providers (e.g., Groq) don't include them in total_tokens\n\tconst outputTokens = (rawUsage.completion_tokens || 0) + reasoningTokens;\n\tconst usage: AssistantMessage[\"usage\"] = {\n\t\tinput,\n\t\toutput: outputTokens,\n\t\tcacheRead: cachedTokens,\n\t\tcacheWrite: 0,\n\t\ttotalTokens: input + outputTokens + cachedTokens,\n\t\tcost: { input: 0, output: 0, cacheRead: 0, cacheWrite: 0, total: 0 },\n\t};\n\tcalculateCost(model, usage);\n\treturn usage;\n}\n\nfunction mapStopReason(reason: ChatCompletionChunk.Choice[\"finish_reason\"] | string): {\n\tstopReason: StopReason;\n\terrorMessage?: string;\n} {\n\tif (reason === null) return { stopReason: \"stop\" };\n\tswitch (reason) {\n\t\tcase \"stop\":\n\t\tcase \"end\":\n\t\t\treturn { stopReason: \"stop\" };\n\t\tcase \"length\":\n\t\t\treturn { stopReason: \"length\" };\n\t\tcase \"function_call\":\n\t\tcase \"tool_calls\":\n\t\t\treturn { stopReason: \"toolUse\" };\n\t\tcase \"content_filter\":\n\t\t\treturn { stopReason: \"error\", errorMessage: \"Provider finish_reason: content_filter\" };\n\t\tcase \"network_error\":\n\t\t\treturn { stopReason: \"error\", errorMessage: \"Provider finish_reason: network_error\" };\n\t\tdefault:\n\t\t\treturn {\n\t\t\t\tstopReason: \"error\",\n\t\t\t\terrorMessage: `Provider finish_reason: ${reason}`,\n\t\t\t};\n\t}\n}\n\n/**\n * Detect compatibility settings from provider and baseUrl for known providers.\n * Provider takes precedence over URL-based detection since it's explicitly configured.\n * Returns a fully resolved OpenAICompletionsCompat object with all fields set.\n */\nfunction detectCompat(model: Model<\"openai-completions\">): Required<OpenAICompletionsCompat> {\n\tconst provider = model.provider;\n\tconst baseUrl = model.baseUrl;\n\n\tconst isZai = provider === \"zai\" || baseUrl.includes(\"api.z.ai\");\n\n\tconst isNonStandard =\n\t\tprovider === \"cerebras\" ||\n\t\tbaseUrl.includes(\"cerebras.ai\") ||\n\t\tprovider === \"xai\" ||\n\t\tbaseUrl.includes(\"api.x.ai\") ||\n\t\tbaseUrl.includes(\"chutes.ai\") ||\n\t\tbaseUrl.includes(\"deepseek.com\") ||\n\t\tisZai ||\n\t\tprovider === \"opencode\" ||\n\t\tbaseUrl.includes(\"opencode.ai\");\n\n\tconst useMaxTokens = baseUrl.includes(\"chutes.ai\");\n\n\tconst isGrok = provider === \"xai\" || baseUrl.includes(\"api.x.ai\");\n\tconst isGroq = provider === \"groq\" || baseUrl.includes(\"groq.com\");\n\n\tconst reasoningEffortMap =\n\t\tisGroq && model.id === \"qwen/qwen3-32b\"\n\t\t\t? {\n\t\t\t\t\tminimal: \"default\",\n\t\t\t\t\tlow: \"default\",\n\t\t\t\t\tmedium: \"default\",\n\t\t\t\t\thigh: \"default\",\n\t\t\t\t\txhigh: \"default\",\n\t\t\t\t}\n\t\t\t: {};\n\treturn {\n\t\tsupportsStore: !isNonStandard,\n\t\tsupportsDeveloperRole: !isNonStandard,\n\t\tsupportsReasoningEffort: !isGrok && !isZai,\n\t\treasoningEffortMap,\n\t\tsupportsUsageInStreaming: true,\n\t\tmaxTokensField: useMaxTokens ? \"max_tokens\" : \"max_completion_tokens\",\n\t\trequiresToolResultName: false,\n\t\trequiresAssistantAfterToolResult: false,\n\t\trequiresThinkingAsText: false,\n\t\tthinkingFormat: isZai\n\t\t\t? \"zai\"\n\t\t\t: provider === \"openrouter\" || baseUrl.includes(\"openrouter.ai\")\n\t\t\t\t? \"openrouter\"\n\t\t\t\t: \"openai\",\n\t\topenRouterRouting: {},\n\t\tvercelGatewayRouting: {},\n\t\tzaiToolStream: false,\n\t\tsupportsStrictMode: true,\n\t};\n}\n\n/**\n * Get resolved compatibility settings for a model.\n * Uses explicit model.compat if provided, otherwise auto-detects from provider/URL.\n */\nfunction getCompat(model: Model<\"openai-completions\">): Required<OpenAICompletionsCompat> {\n\tconst detected = detectCompat(model);\n\tif (!model.compat) return detected;\n\n\treturn {\n\t\tsupportsStore: model.compat.supportsStore ?? detected.supportsStore,\n\t\tsupportsDeveloperRole: model.compat.supportsDeveloperRole ?? detected.supportsDeveloperRole,\n\t\tsupportsReasoningEffort: model.compat.supportsReasoningEffort ?? detected.supportsReasoningEffort,\n\t\treasoningEffortMap: model.compat.reasoningEffortMap ?? detected.reasoningEffortMap,\n\t\tsupportsUsageInStreaming: model.compat.supportsUsageInStreaming ?? detected.supportsUsageInStreaming,\n\t\tmaxTokensField: model.compat.maxTokensField ?? detected.maxTokensField,\n\t\trequiresToolResultName: model.compat.requiresToolResultName ?? detected.requiresToolResultName,\n\t\trequiresAssistantAfterToolResult:\n\t\t\tmodel.compat.requiresAssistantAfterToolResult ?? detected.requiresAssistantAfterToolResult,\n\t\trequiresThinkingAsText: model.compat.requiresThinkingAsText ?? detected.requiresThinkingAsText,\n\t\tthinkingFormat: model.compat.thinkingFormat ?? detected.thinkingFormat,\n\t\topenRouterRouting: model.compat.openRouterRouting ?? {},\n\t\tvercelGatewayRouting: model.compat.vercelGatewayRouting ?? detected.vercelGatewayRouting,\n\t\tzaiToolStream: model.compat.zaiToolStream ?? detected.zaiToolStream,\n\t\tsupportsStrictMode: model.compat.supportsStrictMode ?? detected.supportsStrictMode,\n\t};\n}\n"]}
+{"version":3,"file":"openai-completions.d.ts","sourceRoot":"","sources":["../../src/providers/openai-completions.ts"],"names":[],"mappings":"AACA,OAAO,KAAK,EAMX,0BAA0B,EAE1B,MAAM,sCAAsC,CAAC;AAG9C,OAAO,KAAK,EAEX,OAAO,EAEP,KAAK,EACL,uBAAuB,EACvB,mBAAmB,EAEnB,cAAc,EACd,aAAa,EAMb,MAAM,aAAa,CAAC;AA2BrB,MAAM,WAAW,wBAAyB,SAAQ,aAAa;IAC9D,UAAU,CAAC,EAAE,MAAM,GAAG,MAAM,GAAG,UAAU,GAAG;QAAE,IAAI,EAAE,UAAU,CAAC;QAAC,QAAQ,EAAE;YAAE,IAAI,EAAE,MAAM,CAAA;SAAE,CAAA;KAAE,CAAC;IAC7F,eAAe,CAAC,EAAE,SAAS,GAAG,KAAK,GAAG,QAAQ,GAAG,MAAM,GAAG,OAAO,CAAC;CAClE;AAED,eAAO,MAAM,uBAAuB,EAAE,cAAc,CAAC,oBAAoB,EAAE,wBAAwB,CAoPlG,CAAC;AAEF,eAAO,MAAM,6BAA6B,EAAE,cAAc,CAAC,oBAAoB,EAAE,mBAAmB,CAmBnG,CAAC;AAqKF,wBAAgB,eAAe,CAC9B,KAAK,EAAE,KAAK,CAAC,oBAAoB,CAAC,EAClC,OAAO,EAAE,OAAO,EAChB,MAAM,EAAE,QAAQ,CAAC,uBAAuB,CAAC,GACvC,0BAA0B,EAAE,CA6N9B","sourcesContent":["import OpenAI from \"openai\";\nimport type {\n\tChatCompletionAssistantMessageParam,\n\tChatCompletionChunk,\n\tChatCompletionContentPart,\n\tChatCompletionContentPartImage,\n\tChatCompletionContentPartText,\n\tChatCompletionMessageParam,\n\tChatCompletionToolMessageParam,\n} from \"openai/resources/chat/completions.js\";\nimport { getEnvApiKey } from \"../env-api-keys.js\";\nimport { calculateCost, supportsXhigh } from \"../models.js\";\nimport type {\n\tAssistantMessage,\n\tContext,\n\tMessage,\n\tModel,\n\tOpenAICompletionsCompat,\n\tSimpleStreamOptions,\n\tStopReason,\n\tStreamFunction,\n\tStreamOptions,\n\tTextContent,\n\tThinkingContent,\n\tTool,\n\tToolCall,\n\tToolResultMessage,\n} from \"../types.js\";\nimport { AssistantMessageEventStream } from \"../utils/event-stream.js\";\nimport { parseStreamingJson } from \"../utils/json-parse.js\";\nimport { sanitizeSurrogates } from \"../utils/sanitize-unicode.js\";\nimport { buildCopilotDynamicHeaders, hasCopilotVisionInput } from \"./github-copilot-headers.js\";\nimport { buildBaseOptions, clampReasoning } from \"./simple-options.js\";\nimport { transformMessages } from \"./transform-messages.js\";\n\n/**\n * Check if conversation messages contain tool calls or tool results.\n * This is needed because Anthropic (via proxy) requires the tools param\n * to be present when messages include tool_calls or tool role messages.\n */\nfunction hasToolHistory(messages: Message[]): boolean {\n\tfor (const msg of messages) {\n\t\tif (msg.role === \"toolResult\") {\n\t\t\treturn true;\n\t\t}\n\t\tif (msg.role === \"assistant\") {\n\t\t\tif (msg.content.some((block) => block.type === \"toolCall\")) {\n\t\t\t\treturn true;\n\t\t\t}\n\t\t}\n\t}\n\treturn false;\n}\n\nexport interface OpenAICompletionsOptions extends StreamOptions {\n\ttoolChoice?: \"auto\" | \"none\" | \"required\" | { type: \"function\"; function: { name: string } };\n\treasoningEffort?: \"minimal\" | \"low\" | \"medium\" | \"high\" | \"xhigh\";\n}\n\nexport const streamOpenAICompletions: StreamFunction<\"openai-completions\", OpenAICompletionsOptions> = (\n\tmodel: Model<\"openai-completions\">,\n\tcontext: Context,\n\toptions?: OpenAICompletionsOptions,\n): AssistantMessageEventStream => {\n\tconst stream = new AssistantMessageEventStream();\n\n\t(async () => {\n\t\tconst output: AssistantMessage = {\n\t\t\trole: \"assistant\",\n\t\t\tcontent: [],\n\t\t\tapi: model.api,\n\t\t\tprovider: model.provider,\n\t\t\tmodel: model.id,\n\t\t\tusage: {\n\t\t\t\tinput: 0,\n\t\t\t\toutput: 0,\n\t\t\t\tcacheRead: 0,\n\t\t\t\tcacheWrite: 0,\n\t\t\t\ttotalTokens: 0,\n\t\t\t\tcost: { input: 0, output: 0, cacheRead: 0, cacheWrite: 0, total: 0 },\n\t\t\t},\n\t\t\tstopReason: \"stop\",\n\t\t\ttimestamp: Date.now(),\n\t\t};\n\n\t\ttry {\n\t\t\tconst apiKey = options?.apiKey || getEnvApiKey(model.provider) || \"\";\n\t\t\tconst client = createClient(model, context, apiKey, options?.headers);\n\t\t\tlet params = buildParams(model, context, options);\n\t\t\tconst nextParams = await options?.onPayload?.(params, model);\n\t\t\tif (nextParams !== undefined) {\n\t\t\t\tparams = nextParams as OpenAI.Chat.Completions.ChatCompletionCreateParamsStreaming;\n\t\t\t}\n\t\t\tconst openaiStream = await client.chat.completions.create(params, { signal: options?.signal });\n\t\t\tstream.push({ type: \"start\", partial: output });\n\n\t\t\tlet currentBlock: TextContent | ThinkingContent | (ToolCall & { partialArgs?: string }) | null = null;\n\t\t\tconst blocks = output.content;\n\t\t\tconst blockIndex = () => blocks.length - 1;\n\t\t\tconst finishCurrentBlock = (block?: typeof currentBlock) => {\n\t\t\t\tif (block) {\n\t\t\t\t\tif (block.type === \"text\") {\n\t\t\t\t\t\tstream.push({\n\t\t\t\t\t\t\ttype: \"text_end\",\n\t\t\t\t\t\t\tcontentIndex: blockIndex(),\n\t\t\t\t\t\t\tcontent: block.text,\n\t\t\t\t\t\t\tpartial: output,\n\t\t\t\t\t\t});\n\t\t\t\t\t} else if (block.type === \"thinking\") {\n\t\t\t\t\t\tstream.push({\n\t\t\t\t\t\t\ttype: \"thinking_end\",\n\t\t\t\t\t\t\tcontentIndex: blockIndex(),\n\t\t\t\t\t\t\tcontent: block.thinking,\n\t\t\t\t\t\t\tpartial: output,\n\t\t\t\t\t\t});\n\t\t\t\t\t} else if (block.type === \"toolCall\") {\n\t\t\t\t\t\tblock.arguments = parseStreamingJson(block.partialArgs);\n\t\t\t\t\t\tdelete block.partialArgs;\n\t\t\t\t\t\tstream.push({\n\t\t\t\t\t\t\ttype: \"toolcall_end\",\n\t\t\t\t\t\t\tcontentIndex: blockIndex(),\n\t\t\t\t\t\t\ttoolCall: block,\n\t\t\t\t\t\t\tpartial: output,\n\t\t\t\t\t\t});\n\t\t\t\t\t}\n\t\t\t\t}\n\t\t\t};\n\n\t\t\tfor await (const chunk of openaiStream) {\n\t\t\t\tif (!chunk || typeof chunk !== \"object\") continue;\n\n\t\t\t\t// OpenAI documents ChatCompletionChunk.id as the unique chat completion identifier,\n\t\t\t\t// and each chunk in a streamed completion carries the same id.\n\t\t\t\toutput.responseId ||= chunk.id;\n\t\t\t\tif (chunk.usage) {\n\t\t\t\t\toutput.usage = parseChunkUsage(chunk.usage, model);\n\t\t\t\t}\n\n\t\t\t\tconst choice = Array.isArray(chunk.choices) ? chunk.choices[0] : undefined;\n\t\t\t\tif (!choice) continue;\n\n\t\t\t\t// Fallback: some providers (e.g., Moonshot) return usage\n\t\t\t\t// in choice.usage instead of the standard chunk.usage\n\t\t\t\tif (!chunk.usage && (choice as any).usage) {\n\t\t\t\t\toutput.usage = parseChunkUsage((choice as any).usage, model);\n\t\t\t\t}\n\n\t\t\t\tif (choice.finish_reason) {\n\t\t\t\t\tconst finishReasonResult = mapStopReason(choice.finish_reason);\n\t\t\t\t\toutput.stopReason = finishReasonResult.stopReason;\n\t\t\t\t\tif (finishReasonResult.errorMessage) {\n\t\t\t\t\t\toutput.errorMessage = finishReasonResult.errorMessage;\n\t\t\t\t\t}\n\t\t\t\t}\n\n\t\t\t\tif (choice.delta) {\n\t\t\t\t\tif (\n\t\t\t\t\t\tchoice.delta.content !== null &&\n\t\t\t\t\t\tchoice.delta.content !== undefined &&\n\t\t\t\t\t\tchoice.delta.content.length > 0\n\t\t\t\t\t) {\n\t\t\t\t\t\tif (!currentBlock || currentBlock.type !== \"text\") {\n\t\t\t\t\t\t\tfinishCurrentBlock(currentBlock);\n\t\t\t\t\t\t\tcurrentBlock = { type: \"text\", text: \"\" };\n\t\t\t\t\t\t\toutput.content.push(currentBlock);\n\t\t\t\t\t\t\tstream.push({ type: \"text_start\", contentIndex: blockIndex(), partial: output });\n\t\t\t\t\t\t}\n\n\t\t\t\t\t\tif (currentBlock.type === \"text\") {\n\t\t\t\t\t\t\tcurrentBlock.text += choice.delta.content;\n\t\t\t\t\t\t\tstream.push({\n\t\t\t\t\t\t\t\ttype: \"text_delta\",\n\t\t\t\t\t\t\t\tcontentIndex: blockIndex(),\n\t\t\t\t\t\t\t\tdelta: choice.delta.content,\n\t\t\t\t\t\t\t\tpartial: output,\n\t\t\t\t\t\t\t});\n\t\t\t\t\t\t}\n\t\t\t\t\t}\n\n\t\t\t\t\t// Some endpoints return reasoning in reasoning_content (llama.cpp),\n\t\t\t\t\t// or reasoning (other openai compatible endpoints)\n\t\t\t\t\t// Use the first non-empty reasoning field to avoid duplication\n\t\t\t\t\t// (e.g., chutes.ai returns both reasoning_content and reasoning with same content)\n\t\t\t\t\tconst reasoningFields = [\"reasoning_content\", \"reasoning\", \"reasoning_text\"];\n\t\t\t\t\tlet foundReasoningField: string | null = null;\n\t\t\t\t\tfor (const field of reasoningFields) {\n\t\t\t\t\t\tif (\n\t\t\t\t\t\t\t(choice.delta as any)[field] !== null &&\n\t\t\t\t\t\t\t(choice.delta as any)[field] !== undefined &&\n\t\t\t\t\t\t\t(choice.delta as any)[field].length > 0\n\t\t\t\t\t\t) {\n\t\t\t\t\t\t\tif (!foundReasoningField) {\n\t\t\t\t\t\t\t\tfoundReasoningField = field;\n\t\t\t\t\t\t\t\tbreak;\n\t\t\t\t\t\t\t}\n\t\t\t\t\t\t}\n\t\t\t\t\t}\n\n\t\t\t\t\tif (foundReasoningField) {\n\t\t\t\t\t\tif (!currentBlock || currentBlock.type !== \"thinking\") {\n\t\t\t\t\t\t\tfinishCurrentBlock(currentBlock);\n\t\t\t\t\t\t\tcurrentBlock = {\n\t\t\t\t\t\t\t\ttype: \"thinking\",\n\t\t\t\t\t\t\t\tthinking: \"\",\n\t\t\t\t\t\t\t\tthinkingSignature: foundReasoningField,\n\t\t\t\t\t\t\t};\n\t\t\t\t\t\t\toutput.content.push(currentBlock);\n\t\t\t\t\t\t\tstream.push({ type: \"thinking_start\", contentIndex: blockIndex(), partial: output });\n\t\t\t\t\t\t}\n\n\t\t\t\t\t\tif (currentBlock.type === \"thinking\") {\n\t\t\t\t\t\t\tconst delta = (choice.delta as any)[foundReasoningField];\n\t\t\t\t\t\t\tcurrentBlock.thinking += delta;\n\t\t\t\t\t\t\tstream.push({\n\t\t\t\t\t\t\t\ttype: \"thinking_delta\",\n\t\t\t\t\t\t\t\tcontentIndex: blockIndex(),\n\t\t\t\t\t\t\t\tdelta,\n\t\t\t\t\t\t\t\tpartial: output,\n\t\t\t\t\t\t\t});\n\t\t\t\t\t\t}\n\t\t\t\t\t}\n\n\t\t\t\t\tif (choice?.delta?.tool_calls) {\n\t\t\t\t\t\tfor (const toolCall of choice.delta.tool_calls) {\n\t\t\t\t\t\t\tif (\n\t\t\t\t\t\t\t\t!currentBlock ||\n\t\t\t\t\t\t\t\tcurrentBlock.type !== \"toolCall\" ||\n\t\t\t\t\t\t\t\t(toolCall.id && currentBlock.id !== toolCall.id)\n\t\t\t\t\t\t\t) {\n\t\t\t\t\t\t\t\tfinishCurrentBlock(currentBlock);\n\t\t\t\t\t\t\t\tcurrentBlock = {\n\t\t\t\t\t\t\t\t\ttype: \"toolCall\",\n\t\t\t\t\t\t\t\t\tid: toolCall.id || \"\",\n\t\t\t\t\t\t\t\t\tname: toolCall.function?.name || \"\",\n\t\t\t\t\t\t\t\t\targuments: {},\n\t\t\t\t\t\t\t\t\tpartialArgs: \"\",\n\t\t\t\t\t\t\t\t};\n\t\t\t\t\t\t\t\toutput.content.push(currentBlock);\n\t\t\t\t\t\t\t\tstream.push({ type: \"toolcall_start\", contentIndex: blockIndex(), partial: output });\n\t\t\t\t\t\t\t}\n\n\t\t\t\t\t\t\tif (currentBlock.type === \"toolCall\") {\n\t\t\t\t\t\t\t\tif (toolCall.id) currentBlock.id = toolCall.id;\n\t\t\t\t\t\t\t\tif (toolCall.function?.name) currentBlock.name = toolCall.function.name;\n\t\t\t\t\t\t\t\tlet delta = \"\";\n\t\t\t\t\t\t\t\tif (toolCall.function?.arguments) {\n\t\t\t\t\t\t\t\t\tdelta = toolCall.function.arguments;\n\t\t\t\t\t\t\t\t\tcurrentBlock.partialArgs += toolCall.function.arguments;\n\t\t\t\t\t\t\t\t\tcurrentBlock.arguments = parseStreamingJson(currentBlock.partialArgs);\n\t\t\t\t\t\t\t\t}\n\t\t\t\t\t\t\t\tstream.push({\n\t\t\t\t\t\t\t\t\ttype: \"toolcall_delta\",\n\t\t\t\t\t\t\t\t\tcontentIndex: blockIndex(),\n\t\t\t\t\t\t\t\t\tdelta,\n\t\t\t\t\t\t\t\t\tpartial: output,\n\t\t\t\t\t\t\t\t});\n\t\t\t\t\t\t\t}\n\t\t\t\t\t\t}\n\t\t\t\t\t}\n\n\t\t\t\t\tconst reasoningDetails = (choice.delta as any).reasoning_details;\n\t\t\t\t\tif (reasoningDetails && Array.isArray(reasoningDetails)) {\n\t\t\t\t\t\tfor (const detail of reasoningDetails) {\n\t\t\t\t\t\t\tif (detail.type === \"reasoning.encrypted\" && detail.id && detail.data) {\n\t\t\t\t\t\t\t\tconst matchingToolCall = output.content.find(\n\t\t\t\t\t\t\t\t\t(b) => b.type === \"toolCall\" && b.id === detail.id,\n\t\t\t\t\t\t\t\t) as ToolCall | undefined;\n\t\t\t\t\t\t\t\tif (matchingToolCall) {\n\t\t\t\t\t\t\t\t\tmatchingToolCall.thoughtSignature = JSON.stringify(detail);\n\t\t\t\t\t\t\t\t}\n\t\t\t\t\t\t\t}\n\t\t\t\t\t\t}\n\t\t\t\t\t}\n\t\t\t\t}\n\t\t\t}\n\n\t\t\tfinishCurrentBlock(currentBlock);\n\t\t\tif (options?.signal?.aborted) {\n\t\t\t\tthrow new Error(\"Request was aborted\");\n\t\t\t}\n\n\t\t\tif (output.stopReason === \"aborted\") {\n\t\t\t\tthrow new Error(\"Request was aborted\");\n\t\t\t}\n\t\t\tif (output.stopReason === \"error\") {\n\t\t\t\tthrow new Error(output.errorMessage || \"Provider returned an error stop reason\");\n\t\t\t}\n\n\t\t\tstream.push({ type: \"done\", reason: output.stopReason, message: output });\n\t\t\tstream.end();\n\t\t} catch (error) {\n\t\t\tfor (const block of output.content) delete (block as any).index;\n\t\t\toutput.stopReason = options?.signal?.aborted ? \"aborted\" : \"error\";\n\t\t\toutput.errorMessage = error instanceof Error ? error.message : JSON.stringify(error);\n\t\t\t// Some providers via OpenRouter give additional information in this field.\n\t\t\tconst rawMetadata = (error as any)?.error?.metadata?.raw;\n\t\t\tif (rawMetadata) output.errorMessage += `\\n${rawMetadata}`;\n\t\t\tstream.push({ type: \"error\", reason: output.stopReason, error: output });\n\t\t\tstream.end();\n\t\t}\n\t})();\n\n\treturn stream;\n};\n\nexport const streamSimpleOpenAICompletions: StreamFunction<\"openai-completions\", SimpleStreamOptions> = (\n\tmodel: Model<\"openai-completions\">,\n\tcontext: Context,\n\toptions?: SimpleStreamOptions,\n): AssistantMessageEventStream => {\n\tconst apiKey = options?.apiKey || getEnvApiKey(model.provider);\n\tif (!apiKey) {\n\t\tthrow new Error(`No API key for provider: ${model.provider}`);\n\t}\n\n\tconst base = buildBaseOptions(model, options, apiKey);\n\tconst reasoningEffort = supportsXhigh(model) ? options?.reasoning : clampReasoning(options?.reasoning);\n\tconst toolChoice = (options as OpenAICompletionsOptions | undefined)?.toolChoice;\n\n\treturn streamOpenAICompletions(model, context, {\n\t\t...base,\n\t\treasoningEffort,\n\t\ttoolChoice,\n\t} satisfies OpenAICompletionsOptions);\n};\n\nfunction createClient(\n\tmodel: Model<\"openai-completions\">,\n\tcontext: Context,\n\tapiKey?: string,\n\toptionsHeaders?: Record<string, string>,\n) {\n\tif (!apiKey) {\n\t\tif (!process.env.OPENAI_API_KEY) {\n\t\t\tthrow new Error(\n\t\t\t\t\"OpenAI API key is required. Set OPENAI_API_KEY environment variable or pass it as an argument.\",\n\t\t\t);\n\t\t}\n\t\tapiKey = process.env.OPENAI_API_KEY;\n\t}\n\n\tconst headers = { ...model.headers };\n\tif (model.provider === \"github-copilot\") {\n\t\tconst hasImages = hasCopilotVisionInput(context.messages);\n\t\tconst copilotHeaders = buildCopilotDynamicHeaders({\n\t\t\tmessages: context.messages,\n\t\t\thasImages,\n\t\t});\n\t\tObject.assign(headers, copilotHeaders);\n\t}\n\n\t// Merge options headers last so they can override defaults\n\tif (optionsHeaders) {\n\t\tObject.assign(headers, optionsHeaders);\n\t}\n\n\treturn new OpenAI({\n\t\tapiKey,\n\t\tbaseURL: model.baseUrl,\n\t\tdangerouslyAllowBrowser: true,\n\t\tdefaultHeaders: headers,\n\t});\n}\n\nfunction buildParams(model: Model<\"openai-completions\">, context: Context, options?: OpenAICompletionsOptions) {\n\tconst compat = getCompat(model);\n\tconst messages = convertMessages(model, context, compat);\n\tmaybeAddOpenRouterAnthropicCacheControl(model, messages);\n\n\tconst params: OpenAI.Chat.Completions.ChatCompletionCreateParamsStreaming = {\n\t\tmodel: model.id,\n\t\tmessages,\n\t\tstream: true,\n\t};\n\n\tif (compat.supportsUsageInStreaming !== false) {\n\t\t(params as any).stream_options = { include_usage: true };\n\t}\n\n\tif (compat.supportsStore) {\n\t\tparams.store = false;\n\t}\n\n\tif (options?.maxTokens) {\n\t\tif (compat.maxTokensField === \"max_tokens\") {\n\t\t\t(params as any).max_tokens = options.maxTokens;\n\t\t} else {\n\t\t\tparams.max_completion_tokens = options.maxTokens;\n\t\t}\n\t}\n\n\tif (options?.temperature !== undefined) {\n\t\tparams.temperature = options.temperature;\n\t}\n\n\tif (context.tools) {\n\t\tparams.tools = convertTools(context.tools, compat);\n\t\tif (compat.zaiToolStream) {\n\t\t\t(params as any).tool_stream = true;\n\t\t}\n\t} else if (hasToolHistory(context.messages)) {\n\t\t// Anthropic (via LiteLLM/proxy) requires tools param when conversation has tool_calls/tool_results\n\t\tparams.tools = [];\n\t}\n\n\tif (options?.toolChoice) {\n\t\tparams.tool_choice = options.toolChoice;\n\t}\n\n\tif (compat.thinkingFormat === \"zai\" && model.reasoning) {\n\t\t(params as any).enable_thinking = !!options?.reasoningEffort;\n\t} else if (compat.thinkingFormat === \"qwen\" && model.reasoning) {\n\t\t(params as any).enable_thinking = !!options?.reasoningEffort;\n\t} else if (compat.thinkingFormat === \"qwen-chat-template\" && model.reasoning) {\n\t\t(params as any).chat_template_kwargs = { enable_thinking: !!options?.reasoningEffort };\n\t} else if (compat.thinkingFormat === \"openrouter\" && model.reasoning) {\n\t\t// OpenRouter normalizes reasoning across providers via a nested reasoning object.\n\t\tconst openRouterParams = params as typeof params & { reasoning?: { effort?: string } };\n\t\tif (options?.reasoningEffort) {\n\t\t\topenRouterParams.reasoning = {\n\t\t\t\teffort: mapReasoningEffort(options.reasoningEffort, compat.reasoningEffortMap),\n\t\t\t};\n\t\t} else {\n\t\t\topenRouterParams.reasoning = { effort: \"none\" };\n\t\t}\n\t} else if (options?.reasoningEffort && model.reasoning && compat.supportsReasoningEffort) {\n\t\t// OpenAI-style reasoning_effort\n\t\t(params as any).reasoning_effort = mapReasoningEffort(options.reasoningEffort, compat.reasoningEffortMap);\n\t}\n\n\t// OpenRouter provider routing preferences\n\tif (model.baseUrl.includes(\"openrouter.ai\") && model.compat?.openRouterRouting) {\n\t\t(params as any).provider = model.compat.openRouterRouting;\n\t}\n\n\t// Vercel AI Gateway provider routing preferences\n\tif (model.baseUrl.includes(\"ai-gateway.vercel.sh\") && model.compat?.vercelGatewayRouting) {\n\t\tconst routing = model.compat.vercelGatewayRouting;\n\t\tif (routing.only || routing.order) {\n\t\t\tconst gatewayOptions: Record<string, string[]> = {};\n\t\t\tif (routing.only) gatewayOptions.only = routing.only;\n\t\t\tif (routing.order) gatewayOptions.order = routing.order;\n\t\t\t(params as any).providerOptions = { gateway: gatewayOptions };\n\t\t}\n\t}\n\n\treturn params;\n}\n\nfunction mapReasoningEffort(\n\teffort: NonNullable<OpenAICompletionsOptions[\"reasoningEffort\"]>,\n\treasoningEffortMap: Partial<Record<NonNullable<OpenAICompletionsOptions[\"reasoningEffort\"]>, string>>,\n): string {\n\treturn reasoningEffortMap[effort] ?? effort;\n}\n\nfunction maybeAddOpenRouterAnthropicCacheControl(\n\tmodel: Model<\"openai-completions\">,\n\tmessages: ChatCompletionMessageParam[],\n): void {\n\tif (model.provider !== \"openrouter\" || !model.id.startsWith(\"anthropic/\")) return;\n\n\t// Anthropic-style caching requires cache_control on a text part. Add a breakpoint\n\t// on the last user/assistant message (walking backwards until we find text content).\n\tfor (let i = messages.length - 1; i >= 0; i--) {\n\t\tconst msg = messages[i];\n\t\tif (msg.role !== \"user\" && msg.role !== \"assistant\") continue;\n\n\t\tconst content = msg.content;\n\t\tif (typeof content === \"string\") {\n\t\t\tmsg.content = [\n\t\t\t\tObject.assign({ type: \"text\" as const, text: content }, { cache_control: { type: \"ephemeral\" } }),\n\t\t\t];\n\t\t\treturn;\n\t\t}\n\n\t\tif (!Array.isArray(content)) continue;\n\n\t\t// Find last text part and add cache_control\n\t\tfor (let j = content.length - 1; j >= 0; j--) {\n\t\t\tconst part = content[j];\n\t\t\tif (part?.type === \"text\") {\n\t\t\t\tObject.assign(part, { cache_control: { type: \"ephemeral\" } });\n\t\t\t\treturn;\n\t\t\t}\n\t\t}\n\t}\n}\n\nexport function convertMessages(\n\tmodel: Model<\"openai-completions\">,\n\tcontext: Context,\n\tcompat: Required<OpenAICompletionsCompat>,\n): ChatCompletionMessageParam[] {\n\tconst params: ChatCompletionMessageParam[] = [];\n\n\tconst normalizeToolCallId = (id: string): string => {\n\t\t// Handle pipe-separated IDs from OpenAI Responses API\n\t\t// Format: {call_id}|{id} where {id} can be 400+ chars with special chars (+, /, =)\n\t\t// These come from providers like github-copilot, openai-codex, opencode\n\t\t// Extract just the call_id part and normalize it\n\t\tif (id.includes(\"|\")) {\n\t\t\tconst [callId] = id.split(\"|\");\n\t\t\t// Sanitize to allowed chars and truncate to 40 chars (OpenAI limit)\n\t\t\treturn callId.replace(/[^a-zA-Z0-9_-]/g, \"_\").slice(0, 40);\n\t\t}\n\n\t\tif (model.provider === \"openai\") return id.length > 40 ? id.slice(0, 40) : id;\n\t\treturn id;\n\t};\n\n\tconst transformedMessages = transformMessages(context.messages, model, (id) => normalizeToolCallId(id));\n\n\tif (context.systemPrompt) {\n\t\tconst useDeveloperRole = model.reasoning && compat.supportsDeveloperRole;\n\t\tconst role = useDeveloperRole ? \"developer\" : \"system\";\n\t\tparams.push({ role: role, content: sanitizeSurrogates(context.systemPrompt) });\n\t}\n\n\tlet lastRole: string | null = null;\n\n\tfor (let i = 0; i < transformedMessages.length; i++) {\n\t\tconst msg = transformedMessages[i];\n\t\t// Some providers don't allow user messages directly after tool results\n\t\t// Insert a synthetic assistant message to bridge the gap\n\t\tif (compat.requiresAssistantAfterToolResult && lastRole === \"toolResult\" && msg.role === \"user\") {\n\t\t\tparams.push({\n\t\t\t\trole: \"assistant\",\n\t\t\t\tcontent: \"I have processed the tool results.\",\n\t\t\t});\n\t\t}\n\n\t\tif (msg.role === \"user\") {\n\t\t\tif (typeof msg.content === \"string\") {\n\t\t\t\tparams.push({\n\t\t\t\t\trole: \"user\",\n\t\t\t\t\tcontent: sanitizeSurrogates(msg.content),\n\t\t\t\t});\n\t\t\t} else {\n\t\t\t\tconst content: ChatCompletionContentPart[] = msg.content.map((item): ChatCompletionContentPart => {\n\t\t\t\t\tif (item.type === \"text\") {\n\t\t\t\t\t\treturn {\n\t\t\t\t\t\t\ttype: \"text\",\n\t\t\t\t\t\t\ttext: sanitizeSurrogates(item.text),\n\t\t\t\t\t\t} satisfies ChatCompletionContentPartText;\n\t\t\t\t\t} else {\n\t\t\t\t\t\treturn {\n\t\t\t\t\t\t\ttype: \"image_url\",\n\t\t\t\t\t\t\timage_url: {\n\t\t\t\t\t\t\t\turl: `data:${item.mimeType};base64,${item.data}`,\n\t\t\t\t\t\t\t},\n\t\t\t\t\t\t} satisfies ChatCompletionContentPartImage;\n\t\t\t\t\t}\n\t\t\t\t});\n\t\t\t\tconst filteredContent = !model.input.includes(\"image\")\n\t\t\t\t\t? content.filter((c) => c.type !== \"image_url\")\n\t\t\t\t\t: content;\n\t\t\t\tif (filteredContent.length === 0) continue;\n\t\t\t\tparams.push({\n\t\t\t\t\trole: \"user\",\n\t\t\t\t\tcontent: filteredContent,\n\t\t\t\t});\n\t\t\t}\n\t\t} else if (msg.role === \"assistant\") {\n\t\t\t// Some providers don't accept null content, use empty string instead\n\t\t\tconst assistantMsg: ChatCompletionAssistantMessageParam = {\n\t\t\t\trole: \"assistant\",\n\t\t\t\tcontent: compat.requiresAssistantAfterToolResult ? \"\" : null,\n\t\t\t};\n\n\t\t\tconst textBlocks = msg.content.filter((b) => b.type === \"text\") as TextContent[];\n\t\t\t// Filter out empty text blocks to avoid API validation errors\n\t\t\tconst nonEmptyTextBlocks = textBlocks.filter((b) => b.text && b.text.trim().length > 0);\n\t\t\tif (nonEmptyTextBlocks.length > 0) {\n\t\t\t\t// Always send assistant content as a plain string (OpenAI Chat Completions\n\t\t\t\t// API standard format). Sending as an array of {type:\"text\", text:\"...\"}\n\t\t\t\t// objects is non-standard and causes some models (e.g. DeepSeek V3.2 via\n\t\t\t\t// NVIDIA NIM) to mirror the content-block structure literally in their\n\t\t\t\t// output, producing recursive nesting like [{'type':'text','text':'[{...}]'}].\n\t\t\t\tassistantMsg.content = nonEmptyTextBlocks.map((b) => sanitizeSurrogates(b.text)).join(\"\");\n\t\t\t}\n\n\t\t\t// Handle thinking blocks\n\t\t\tconst thinkingBlocks = msg.content.filter((b) => b.type === \"thinking\") as ThinkingContent[];\n\t\t\t// Filter out empty thinking blocks to avoid API validation errors\n\t\t\tconst nonEmptyThinkingBlocks = thinkingBlocks.filter((b) => b.thinking && b.thinking.trim().length > 0);\n\t\t\tif (nonEmptyThinkingBlocks.length > 0) {\n\t\t\t\tif (compat.requiresThinkingAsText) {\n\t\t\t\t\t// Convert thinking blocks to plain text (no tags to avoid model mimicking them)\n\t\t\t\t\tconst thinkingText = nonEmptyThinkingBlocks.map((b) => b.thinking).join(\"\\n\\n\");\n\t\t\t\t\tconst textContent = assistantMsg.content as Array<{ type: \"text\"; text: string }> | null;\n\t\t\t\t\tif (textContent) {\n\t\t\t\t\t\ttextContent.unshift({ type: \"text\", text: thinkingText });\n\t\t\t\t\t} else {\n\t\t\t\t\t\tassistantMsg.content = [{ type: \"text\", text: thinkingText }];\n\t\t\t\t\t}\n\t\t\t\t} else {\n\t\t\t\t\t// Use the signature from the first thinking block if available (for llama.cpp server + gpt-oss)\n\t\t\t\t\tconst signature = nonEmptyThinkingBlocks[0].thinkingSignature;\n\t\t\t\t\tif (signature && signature.length > 0) {\n\t\t\t\t\t\t(assistantMsg as any)[signature] = nonEmptyThinkingBlocks.map((b) => b.thinking).join(\"\\n\");\n\t\t\t\t\t}\n\t\t\t\t}\n\t\t\t}\n\n\t\t\tconst toolCalls = msg.content.filter((b) => b.type === \"toolCall\") as ToolCall[];\n\t\t\tif (toolCalls.length > 0) {\n\t\t\t\tassistantMsg.tool_calls = toolCalls.map((tc) => ({\n\t\t\t\t\tid: tc.id,\n\t\t\t\t\ttype: \"function\" as const,\n\t\t\t\t\tfunction: {\n\t\t\t\t\t\tname: tc.name,\n\t\t\t\t\t\targuments: JSON.stringify(tc.arguments),\n\t\t\t\t\t},\n\t\t\t\t}));\n\t\t\t\tconst reasoningDetails = toolCalls\n\t\t\t\t\t.filter((tc) => tc.thoughtSignature)\n\t\t\t\t\t.map((tc) => {\n\t\t\t\t\t\ttry {\n\t\t\t\t\t\t\treturn JSON.parse(tc.thoughtSignature!);\n\t\t\t\t\t\t} catch {\n\t\t\t\t\t\t\treturn null;\n\t\t\t\t\t\t}\n\t\t\t\t\t})\n\t\t\t\t\t.filter(Boolean);\n\t\t\t\tif (reasoningDetails.length > 0) {\n\t\t\t\t\t(assistantMsg as any).reasoning_details = reasoningDetails;\n\t\t\t\t}\n\t\t\t}\n\t\t\t// Skip assistant messages that have no content and no tool calls.\n\t\t\t// Some providers require \"either content or tool_calls, but not none\".\n\t\t\t// Other providers also don't accept empty assistant messages.\n\t\t\t// This handles aborted assistant responses that got no content.\n\t\t\tconst content = assistantMsg.content;\n\t\t\tconst hasContent =\n\t\t\t\tcontent !== null &&\n\t\t\t\tcontent !== undefined &&\n\t\t\t\t(typeof content === \"string\" ? content.length > 0 : content.length > 0);\n\t\t\tif (!hasContent && !assistantMsg.tool_calls) {\n\t\t\t\tcontinue;\n\t\t\t}\n\t\t\tparams.push(assistantMsg);\n\t\t} else if (msg.role === \"toolResult\") {\n\t\t\tconst imageBlocks: Array<{ type: \"image_url\"; image_url: { url: string } }> = [];\n\t\t\tlet j = i;\n\n\t\t\tfor (; j < transformedMessages.length && transformedMessages[j].role === \"toolResult\"; j++) {\n\t\t\t\tconst toolMsg = transformedMessages[j] as ToolResultMessage;\n\n\t\t\t\t// Extract text and image content\n\t\t\t\tconst textResult = toolMsg.content\n\t\t\t\t\t.filter((c) => c.type === \"text\")\n\t\t\t\t\t.map((c) => (c as any).text)\n\t\t\t\t\t.join(\"\\n\");\n\t\t\t\tconst hasImages = toolMsg.content.some((c) => c.type === \"image\");\n\n\t\t\t\t// Always send tool result with text (or placeholder if only images)\n\t\t\t\tconst hasText = textResult.length > 0;\n\t\t\t\t// Some providers require the 'name' field in tool results\n\t\t\t\tconst toolResultMsg: ChatCompletionToolMessageParam = {\n\t\t\t\t\trole: \"tool\",\n\t\t\t\t\tcontent: sanitizeSurrogates(hasText ? textResult : \"(see attached image)\"),\n\t\t\t\t\ttool_call_id: toolMsg.toolCallId,\n\t\t\t\t};\n\t\t\t\tif (compat.requiresToolResultName && toolMsg.toolName) {\n\t\t\t\t\t(toolResultMsg as any).name = toolMsg.toolName;\n\t\t\t\t}\n\t\t\t\tparams.push(toolResultMsg);\n\n\t\t\t\tif (hasImages && model.input.includes(\"image\")) {\n\t\t\t\t\tfor (const block of toolMsg.content) {\n\t\t\t\t\t\tif (block.type === \"image\") {\n\t\t\t\t\t\t\timageBlocks.push({\n\t\t\t\t\t\t\t\ttype: \"image_url\",\n\t\t\t\t\t\t\t\timage_url: {\n\t\t\t\t\t\t\t\t\turl: `data:${(block as any).mimeType};base64,${(block as any).data}`,\n\t\t\t\t\t\t\t\t},\n\t\t\t\t\t\t\t});\n\t\t\t\t\t\t}\n\t\t\t\t\t}\n\t\t\t\t}\n\t\t\t}\n\n\t\t\ti = j - 1;\n\n\t\t\tif (imageBlocks.length > 0) {\n\t\t\t\tif (compat.requiresAssistantAfterToolResult) {\n\t\t\t\t\tparams.push({\n\t\t\t\t\t\trole: \"assistant\",\n\t\t\t\t\t\tcontent: \"I have processed the tool results.\",\n\t\t\t\t\t});\n\t\t\t\t}\n\n\t\t\t\tparams.push({\n\t\t\t\t\trole: \"user\",\n\t\t\t\t\tcontent: [\n\t\t\t\t\t\t{\n\t\t\t\t\t\t\ttype: \"text\",\n\t\t\t\t\t\t\ttext: \"Attached image(s) from tool result:\",\n\t\t\t\t\t\t},\n\t\t\t\t\t\t...imageBlocks,\n\t\t\t\t\t],\n\t\t\t\t});\n\t\t\t\tlastRole = \"user\";\n\t\t\t} else {\n\t\t\t\tlastRole = \"toolResult\";\n\t\t\t}\n\t\t\tcontinue;\n\t\t}\n\n\t\tlastRole = msg.role;\n\t}\n\n\treturn params;\n}\n\nfunction convertTools(\n\ttools: Tool[],\n\tcompat: Required<OpenAICompletionsCompat>,\n): OpenAI.Chat.Completions.ChatCompletionTool[] {\n\treturn tools.map((tool) => ({\n\t\ttype: \"function\",\n\t\tfunction: {\n\t\t\tname: tool.name,\n\t\t\tdescription: tool.description,\n\t\t\tparameters: tool.parameters as any, // TypeBox already generates JSON Schema\n\t\t\t// Only include strict if provider supports it. Some reject unknown fields.\n\t\t\t...(compat.supportsStrictMode !== false && { strict: false }),\n\t\t},\n\t}));\n}\n\nfunction parseChunkUsage(\n\trawUsage: {\n\t\tprompt_tokens?: number;\n\t\tcompletion_tokens?: number;\n\t\tprompt_tokens_details?: { cached_tokens?: number; cache_write_tokens?: number };\n\t\tcompletion_tokens_details?: { reasoning_tokens?: number };\n\t},\n\tmodel: Model<\"openai-completions\">,\n): AssistantMessage[\"usage\"] {\n\tconst promptTokens = rawUsage.prompt_tokens || 0;\n\tconst reportedCachedTokens = rawUsage.prompt_tokens_details?.cached_tokens || 0;\n\tconst cacheWriteTokens = rawUsage.prompt_tokens_details?.cache_write_tokens || 0;\n\tconst reasoningTokens = rawUsage.completion_tokens_details?.reasoning_tokens || 0;\n\n\t// Normalize to pi-ai semantics:\n\t// - cacheRead: hits from cache created by previous requests only\n\t// - cacheWrite: tokens written to cache in this request\n\t// Some OpenAI-compatible providers (observed on OpenRouter) report cached_tokens\n\t// as (previous hits + current writes). In that case, remove cacheWrite from cacheRead.\n\tconst cacheReadTokens =\n\t\tcacheWriteTokens > 0 ? Math.max(0, reportedCachedTokens - cacheWriteTokens) : reportedCachedTokens;\n\n\tconst input = Math.max(0, promptTokens - cacheReadTokens - cacheWriteTokens);\n\t// Compute totalTokens ourselves since we add reasoning_tokens to output\n\t// and some providers (e.g., Groq) don't include them in total_tokens\n\tconst outputTokens = (rawUsage.completion_tokens || 0) + reasoningTokens;\n\tconst usage: AssistantMessage[\"usage\"] = {\n\t\tinput,\n\t\toutput: outputTokens,\n\t\tcacheRead: cacheReadTokens,\n\t\tcacheWrite: cacheWriteTokens,\n\t\ttotalTokens: input + outputTokens + cacheReadTokens + cacheWriteTokens,\n\t\tcost: { input: 0, output: 0, cacheRead: 0, cacheWrite: 0, total: 0 },\n\t};\n\tcalculateCost(model, usage);\n\treturn usage;\n}\n\nfunction mapStopReason(reason: ChatCompletionChunk.Choice[\"finish_reason\"] | string): {\n\tstopReason: StopReason;\n\terrorMessage?: string;\n} {\n\tif (reason === null) return { stopReason: \"stop\" };\n\tswitch (reason) {\n\t\tcase \"stop\":\n\t\tcase \"end\":\n\t\t\treturn { stopReason: \"stop\" };\n\t\tcase \"length\":\n\t\t\treturn { stopReason: \"length\" };\n\t\tcase \"function_call\":\n\t\tcase \"tool_calls\":\n\t\t\treturn { stopReason: \"toolUse\" };\n\t\tcase \"content_filter\":\n\t\t\treturn { stopReason: \"error\", errorMessage: \"Provider finish_reason: content_filter\" };\n\t\tcase \"network_error\":\n\t\t\treturn { stopReason: \"error\", errorMessage: \"Provider finish_reason: network_error\" };\n\t\tdefault:\n\t\t\treturn {\n\t\t\t\tstopReason: \"error\",\n\t\t\t\terrorMessage: `Provider finish_reason: ${reason}`,\n\t\t\t};\n\t}\n}\n\n/**\n * Detect compatibility settings from provider and baseUrl for known providers.\n * Provider takes precedence over URL-based detection since it's explicitly configured.\n * Returns a fully resolved OpenAICompletionsCompat object with all fields set.\n */\nfunction detectCompat(model: Model<\"openai-completions\">): Required<OpenAICompletionsCompat> {\n\tconst provider = model.provider;\n\tconst baseUrl = model.baseUrl;\n\n\tconst isZai = provider === \"zai\" || baseUrl.includes(\"api.z.ai\");\n\n\tconst isNonStandard =\n\t\tprovider === \"cerebras\" ||\n\t\tbaseUrl.includes(\"cerebras.ai\") ||\n\t\tprovider === \"xai\" ||\n\t\tbaseUrl.includes(\"api.x.ai\") ||\n\t\tbaseUrl.includes(\"chutes.ai\") ||\n\t\tbaseUrl.includes(\"deepseek.com\") ||\n\t\tisZai ||\n\t\tprovider === \"opencode\" ||\n\t\tbaseUrl.includes(\"opencode.ai\");\n\n\tconst useMaxTokens = baseUrl.includes(\"chutes.ai\");\n\n\tconst isGrok = provider === \"xai\" || baseUrl.includes(\"api.x.ai\");\n\tconst isGroq = provider === \"groq\" || baseUrl.includes(\"groq.com\");\n\n\tconst reasoningEffortMap =\n\t\tisGroq && model.id === \"qwen/qwen3-32b\"\n\t\t\t? {\n\t\t\t\t\tminimal: \"default\",\n\t\t\t\t\tlow: \"default\",\n\t\t\t\t\tmedium: \"default\",\n\t\t\t\t\thigh: \"default\",\n\t\t\t\t\txhigh: \"default\",\n\t\t\t\t}\n\t\t\t: {};\n\treturn {\n\t\tsupportsStore: !isNonStandard,\n\t\tsupportsDeveloperRole: !isNonStandard,\n\t\tsupportsReasoningEffort: !isGrok && !isZai,\n\t\treasoningEffortMap,\n\t\tsupportsUsageInStreaming: true,\n\t\tmaxTokensField: useMaxTokens ? \"max_tokens\" : \"max_completion_tokens\",\n\t\trequiresToolResultName: false,\n\t\trequiresAssistantAfterToolResult: false,\n\t\trequiresThinkingAsText: false,\n\t\tthinkingFormat: isZai\n\t\t\t? \"zai\"\n\t\t\t: provider === \"openrouter\" || baseUrl.includes(\"openrouter.ai\")\n\t\t\t\t? \"openrouter\"\n\t\t\t\t: \"openai\",\n\t\topenRouterRouting: {},\n\t\tvercelGatewayRouting: {},\n\t\tzaiToolStream: false,\n\t\tsupportsStrictMode: true,\n\t};\n}\n\n/**\n * Get resolved compatibility settings for a model.\n * Uses explicit model.compat if provided, otherwise auto-detects from provider/URL.\n */\nfunction getCompat(model: Model<\"openai-completions\">): Required<OpenAICompletionsCompat> {\n\tconst detected = detectCompat(model);\n\tif (!model.compat) return detected;\n\n\treturn {\n\t\tsupportsStore: model.compat.supportsStore ?? detected.supportsStore,\n\t\tsupportsDeveloperRole: model.compat.supportsDeveloperRole ?? detected.supportsDeveloperRole,\n\t\tsupportsReasoningEffort: model.compat.supportsReasoningEffort ?? detected.supportsReasoningEffort,\n\t\treasoningEffortMap: model.compat.reasoningEffortMap ?? detected.reasoningEffortMap,\n\t\tsupportsUsageInStreaming: model.compat.supportsUsageInStreaming ?? detected.supportsUsageInStreaming,\n\t\tmaxTokensField: model.compat.maxTokensField ?? detected.maxTokensField,\n\t\trequiresToolResultName: model.compat.requiresToolResultName ?? detected.requiresToolResultName,\n\t\trequiresAssistantAfterToolResult:\n\t\t\tmodel.compat.requiresAssistantAfterToolResult ?? detected.requiresAssistantAfterToolResult,\n\t\trequiresThinkingAsText: model.compat.requiresThinkingAsText ?? detected.requiresThinkingAsText,\n\t\tthinkingFormat: model.compat.thinkingFormat ?? detected.thinkingFormat,\n\t\topenRouterRouting: model.compat.openRouterRouting ?? {},\n\t\tvercelGatewayRouting: model.compat.vercelGatewayRouting ?? detected.vercelGatewayRouting,\n\t\tzaiToolStream: model.compat.zaiToolStream ?? detected.zaiToolStream,\n\t\tsupportsStrictMode: model.compat.supportsStrictMode ?? detected.supportsStrictMode,\n\t};\n}\n"]}

package/dist/providers/openai-completions.js

@@ -626,19 +626,26 @@ function convertTools(tools, compat) {
     }));
 }
 function parseChunkUsage(rawUsage, model) {
-    const cachedTokens = rawUsage.prompt_tokens_details?.cached_tokens || 0;
+    const promptTokens = rawUsage.prompt_tokens || 0;
+    const reportedCachedTokens = rawUsage.prompt_tokens_details?.cached_tokens || 0;
+    const cacheWriteTokens = rawUsage.prompt_tokens_details?.cache_write_tokens || 0;
     const reasoningTokens = rawUsage.completion_tokens_details?.reasoning_tokens || 0;
-    // OpenAI includes cached tokens in prompt_tokens, so subtract to get non-cached input
-    const input = (rawUsage.prompt_tokens || 0) - cachedTokens;
+    // Normalize to pi-ai semantics:
+    // - cacheRead: hits from cache created by previous requests only
+    // - cacheWrite: tokens written to cache in this request
+    // Some OpenAI-compatible providers (observed on OpenRouter) report cached_tokens
+    // as (previous hits + current writes). In that case, remove cacheWrite from cacheRead.
+    const cacheReadTokens = cacheWriteTokens > 0 ? Math.max(0, reportedCachedTokens - cacheWriteTokens) : reportedCachedTokens;
+    const input = Math.max(0, promptTokens - cacheReadTokens - cacheWriteTokens);
     // Compute totalTokens ourselves since we add reasoning_tokens to output
     // and some providers (e.g., Groq) don't include them in total_tokens
     const outputTokens = (rawUsage.completion_tokens || 0) + reasoningTokens;
     const usage = {
         input,
         output: outputTokens,
-        cacheRead: cachedTokens,
-        cacheWrite: 0,
-        totalTokens: input + outputTokens + cachedTokens,
+        cacheRead: cacheReadTokens,
+        cacheWrite: cacheWriteTokens,
+        totalTokens: input + outputTokens + cacheReadTokens + cacheWriteTokens,
         cost: { input: 0, output: 0, cacheRead: 0, cacheWrite: 0, total: 0 },
     };
     calculateCost(model, usage);