@sammorrowdrums/mcpi-ai 0.70.9 → 0.85.0

package/dist/{providers → api}/google-shared.js

@@ -1,8 +1,10 @@
 /**
- * Shared utilities for Google Generative AI and Google Cloud Code Assist providers.
+ * Shared utilities for Google Generative AI and Google Vertex providers.
  */
 import { FinishReason, FunctionCallingConfigMode } from "@google/genai";
+import { retryProviderRequest } from "../utils/provider-retry.js";
 import { sanitizeSurrogates } from "../utils/sanitize-unicode.js";
+import { getJsonSchemaToolParameters, resolveJsonSchemaStrictSampling } from "./constrained-sampling.js";
 import { transformMessages } from "./transform-messages.js";
 /**
  * Determines whether a streamed Gemini `Part` should be treated as "thinking".
@@ -38,10 +40,6 @@ export function retainThoughtSignature(existing, incoming) {
 }
 // Thought signatures must be base64 for Google APIs (TYPE_BYTES).
 const base64SignaturePattern = /^[A-Za-z0-9+/]+={0,2}$/;
-// Sentinel value that tells the Gemini API to skip thought signature validation.
-// Used for unsigned function call parts (e.g. replayed from providers without thought signatures).
-// See: https://ai.google.dev/gemini-api/docs/thought-signatures
-const SKIP_THOUGHT_SIGNATURE = "skip_thought_signature_validator";
 function isValidThoughtSignature(signature) {
     if (!signature)
         return false;
@@ -59,7 +57,10 @@ function resolveThoughtSignature(isSameProviderAndModel, signature) {
  * Models via Google APIs that require explicit tool call IDs in function calls/responses.
  */
 export function requiresToolCallId(modelId) {
-    return modelId.startsWith("claude-") || modelId.startsWith("gpt-oss-");
+    const geminiMajorVersion = getGeminiMajorVersion(modelId);
+    return (modelId.startsWith("claude-") ||
+        modelId.startsWith("gpt-oss-") ||
+        (geminiMajorVersion !== undefined && geminiMajorVersion >= 3));
 }
 function getGeminiMajorVersion(modelId) {
     const match = modelId.toLowerCase().match(/^gemini(?:-live)?-(\d+)/);
@@ -121,23 +122,27 @@ export function convertMessages(model, context) {
             const isSameProviderAndModel = msg.provider === model.provider && msg.model === model.id;
             for (const block of msg.content) {
                 if (block.type === "text") {
-                    // Skip empty text blocks - they can cause issues with some models (e.g. Claude via Antigravity)
-                    if (!block.text || block.text.trim() === "")
-                        continue;
                     const thoughtSignature = resolveThoughtSignature(isSameProviderAndModel, block.textSignature);
+                    // Skip empty text blocks — unless they carry a thought signature. Gemini can attach
+                    // the signature to a part whose visible text is empty and requires it echoed back;
+                    // dropping it breaks the reasoning chain and the model intermittently ends mid-task
+                    // turns with a thought-only STOP (empty completion, no tool call).
+                    if ((!block.text || block.text.trim() === "") && !thoughtSignature)
+                        continue;
                     parts.push({
                         text: sanitizeSurrogates(block.text),
                         ...(thoughtSignature && { thoughtSignature }),
                     });
                 }
                 else if (block.type === "thinking") {
-                    // Skip empty thinking blocks
-                    if (!block.thinking || block.thinking.trim() === "")
-                        continue;
                     // Only keep as thinking block if same provider AND same model
                     // Otherwise convert to plain text (no tags to avoid model mimicking them)
                     if (isSameProviderAndModel) {
                         const thoughtSignature = resolveThoughtSignature(isSameProviderAndModel, block.thinkingSignature);
+                        // Same rule as text blocks: an empty thinking block is dropped only when it
+                        // carries no signature (mirrors the anthropic converter's handling).
+                        if ((!block.thinking || block.thinking.trim() === "") && !thoughtSignature)
+                            continue;
                         parts.push({
                             thought: true,
                             text: sanitizeSurrogates(block.thinking),
@@ -145,6 +150,9 @@ export function convertMessages(model, context) {
                         });
                     }
                     else {
+                        // Cross-provider/model: the signature is unusable, empty blocks stay dropped.
+                        if (!block.thinking || block.thinking.trim() === "")
+                            continue;
                         parts.push({
                             text: sanitizeSurrogates(block.thinking),
                         });
@@ -152,18 +160,13 @@ export function convertMessages(model, context) {
                 }
                 else if (block.type === "toolCall") {
                     const thoughtSignature = resolveThoughtSignature(isSameProviderAndModel, block.thoughtSignature);
-                    // Gemini 3 requires thoughtSignature on all function calls when thinking mode is enabled.
-                    // Use the skip_thought_signature_validator sentinel for unsigned function calls
-                    // (e.g. replayed from providers without thought signatures like Claude via Antigravity).
-                    const isGemini3 = model.id.toLowerCase().includes("gemini-3");
-                    const effectiveSignature = thoughtSignature || (isGemini3 ? SKIP_THOUGHT_SIGNATURE : undefined);
                     const part = {
                         functionCall: {
                             name: block.name,
                             args: block.arguments ?? {},
                             ...(requiresToolCallId(model.id) ? { id: block.id } : {}),
                         },
-                        ...(effectiveSignature && { thoughtSignature: effectiveSignature }),
+                        ...(thoughtSignature && { thoughtSignature }),
                     };
                     parts.push(part);
                 }
@@ -186,7 +189,7 @@ export function convertMessages(model, context) {
             const hasImages = imageContent.length > 0;
             // Gemini 3+ models support multimodal function responses with images nested inside
             // functionResponse.parts. Claude and other non-Gemini models behind Cloud Code Assist /
-            // Antigravity also accept this shape. Gemini < 3 still needs a separate user image turn.
+            // Gemini < 3 still needs a separate user image turn.
             const modelSupportsMultimodalFunctionResponse = supportsMultimodalFunctionResponse(model.id);
             // Use "output" key for success, "error" key for errors as per SDK documentation
             const responseValue = hasText ? sanitizeSurrogates(textResult) : hasImages ? "(see attached image)" : "";
@@ -261,24 +264,31 @@ function sanitizeForOpenApi(schema) {
  * field instead (OpenAPI 3.03 Schema). This is needed for Cloud Code Assist with Claude
  * models, where the API translates `parameters` into Anthropic's `input_schema`.
  */
-export function convertTools(tools, useParameters = false) {
+export function convertTools(tools, useParameters = false, supportsStrictMode = true) {
     if (tools.length === 0)
         return undefined;
     return [
         {
-            functionDeclarations: tools.map((tool) => ({
-                name: tool.name,
-                description: tool.description,
-                ...(useParameters
-                    ? { parameters: sanitizeForOpenApi(tool.parameters) }
-                    : { parametersJsonSchema: tool.parameters }),
-            })),
+            functionDeclarations: tools.map((tool) => {
+                const strict = resolveJsonSchemaStrictSampling(tool, supportsStrictMode);
+                const parameters = getJsonSchemaToolParameters(tool, strict);
+                return {
+                    name: tool.name,
+                    description: tool.description,
+                    ...(useParameters
+                        ? { parameters: sanitizeForOpenApi(parameters) }
+                        : { parametersJsonSchema: parameters }),
+                };
+            }),
         },
     ];
 }
-/**
- * Map tool choice string to Gemini FunctionCallingConfigMode.
- */
+/** Gemini 3+ enforces required function parameters in validated tool-calling modes. */
+export function supportsGoogleStrictToolSampling(modelId) {
+    const majorVersion = getGeminiMajorVersion(modelId);
+    return majorVersion !== undefined && majorVersion >= 3;
+}
+/** Map tool choice string to Gemini FunctionCallingConfigMode. */
 export function mapToolChoice(choice) {
     switch (choice) {
         case "auto":
@@ -291,6 +301,16 @@ export function mapToolChoice(choice) {
             return FunctionCallingConfigMode.AUTO;
     }
 }
+export function resolveGoogleFunctionCallingMode(tools, toolChoice, supportsStrictMode) {
+    const useStrictMode = tools.some((tool) => resolveJsonSchemaStrictSampling(tool, supportsStrictMode) === true);
+    if (toolChoice === "none" || toolChoice === "any") {
+        return mapToolChoice(toolChoice);
+    }
+    if (useStrictMode) {
+        return FunctionCallingConfigMode.VALIDATED;
+    }
+    return toolChoice ? mapToolChoice(toolChoice) : undefined;
+}
 /**
  * Map Gemini FinishReason to our StopReason.
  */
@@ -335,4 +355,30 @@ export function mapStopReasonString(reason) {
             return "error";
     }
 }
+/**
+ * Run a Google GenAI SDK request with the shared provider retry policy
+ * (408/409/429/5xx with backoff, honoring retry-after), mirroring how the
+ * Anthropic and OpenAI adapters wrap their initial request in
+ * retryProviderRequest. The SDK's ApiError has a `status` property but no
+ * `headers` property, and retryProviderRequest only retries errors that carry
+ * both, so normalize the error by adding the missing `headers` before
+ * rethrowing.
+ */
+export function retryGoogleRequest(request, options) {
+    return retryProviderRequest(async () => {
+        try {
+            return await request();
+        }
+        catch (error) {
+            if (error instanceof Error && "status" in error && !("headers" in error)) {
+                error.headers = undefined;
+            }
+            throw error;
+        }
+    }, {
+        maxRetries: options?.maxRetries,
+        maxRetryDelayMs: options?.maxRetryDelayMs,
+        signal: options?.signal,
+    });
+}
 //# sourceMappingURL=google-shared.js.map

package/dist/api/google-shared.js.map

@@ -0,0 +1 @@
+{"version":3,"file":"google-shared.js","sourceRoot":"","sources":["../../src/api/google-shared.ts"],"names":[],"mappings":"AAAA;;GAEG;AAEH,OAAO,EAAgB,YAAY,EAAE,yBAAyB,EAAa,MAAM,eAAe,CAAC;AAEjG,OAAO,EAAE,oBAAoB,EAAE,MAAM,4BAA4B,CAAC;AAClE,OAAO,EAAE,kBAAkB,EAAE,MAAM,8BAA8B,CAAC;AAClE,OAAO,EAAE,2BAA2B,EAAE,+BAA+B,EAAE,MAAM,2BAA2B,CAAC;AACzG,OAAO,EAAE,iBAAiB,EAAE,MAAM,yBAAyB,CAAC;AAU5D;;;;;;;;;;;;;;GAcG;AACH,MAAM,UAAU,cAAc,CAAC,IAAgD,EAAW;IACzF,OAAO,IAAI,CAAC,OAAO,KAAK,IAAI,CAAC;AAAA,CAC7B;AAED;;;;;;;;GAQG;AACH,MAAM,UAAU,sBAAsB,CAAC,QAA4B,EAAE,QAA4B,EAAsB;IACtH,IAAI,OAAO,QAAQ,KAAK,QAAQ,IAAI,QAAQ,CAAC,MAAM,GAAG,CAAC;QAAE,OAAO,QAAQ,CAAC;IACzE,OAAO,QAAQ,CAAC;AAAA,CAChB;AAED,kEAAkE;AAClE,MAAM,sBAAsB,GAAG,wBAAwB,CAAC;AAExD,SAAS,uBAAuB,CAAC,SAA6B,EAAW;IACxE,IAAI,CAAC,SAAS;QAAE,OAAO,KAAK,CAAC;IAC7B,IAAI,SAAS,CAAC,MAAM,GAAG,CAAC,KAAK,CAAC;QAAE,OAAO,KAAK,CAAC;IAC7C,OAAO,sBAAsB,CAAC,IAAI,CAAC,SAAS,CAAC,CAAC;AAAA,CAC9C;AAED;;GAEG;AACH,SAAS,uBAAuB,CAAC,sBAA+B,EAAE,SAA6B,EAAsB;IACpH,OAAO,sBAAsB,IAAI,uBAAuB,CAAC,SAAS,CAAC,CAAC,CAAC,CAAC,SAAS,CAAC,CAAC,CAAC,SAAS,CAAC;AAAA,CAC5F;AAED;;GAEG;AACH,MAAM,UAAU,kBAAkB,CAAC,OAAe,EAAW;IAC5D,MAAM,kBAAkB,GAAG,qBAAqB,CAAC,OAAO,CAAC,CAAC;IAC1D,OAAO,CACN,OAAO,CAAC,UAAU,CAAC,SAAS,CAAC;QAC7B,OAAO,CAAC,UAAU,CAAC,UAAU,CAAC;QAC9B,CAAC,kBAAkB,KAAK,SAAS,IAAI,kBAAkB,IAAI,CAAC,CAAC,CAC7D,CAAC;AAAA,CACF;AAED,SAAS,qBAAqB,CAAC,OAAe,EAAsB;IACnE,MAAM,KAAK,GAAG,OAAO,CAAC,WAAW,EAAE,CAAC,KAAK,CAAC,yBAAyB,CAAC,CAAC;IACrE,IAAI,CAAC,KAAK;QAAE,OAAO,SAAS,CAAC;IAC7B,OAAO,MAAM,CAAC,QAAQ,CAAC,KAAK,CAAC,CAAC,CAAC,EAAE,EAAE,CAAC,CAAC;AAAA,CACrC;AAED,SAAS,kCAAkC,CAAC,OAAe,EAAW;IACrE,MAAM,kBAAkB,GAAG,qBAAqB,CAAC,OAAO,CAAC,CAAC;IAC1D,IAAI,kBAAkB,KAAK,SAAS,EAAE,CAAC;QACtC,OAAO,kBAAkB,IAAI,CAAC,CAAC;IAChC,CAAC;IACD,OAAO,IAAI,CAAC;AAAA,CACZ;AAED;;GAEG;AACH,MAAM,UAAU,eAAe,CAA0B,KAAe,EAAE,OAAgB,EAAa;IACtG,MAAM,QAAQ,GAAc,EAAE,CAAC;IAC/B,MAAM,mBAAmB,GAAG,CAAC,EAAU,EAAU,EAAE,CAAC;QACnD,IAAI,CAAC,kBAAkB,CAAC,KAAK,CAAC,EAAE,CAAC;YAAE,OAAO,EAAE,CAAC;QAC7C,OAAO,EAAE,CAAC,OAAO,CAAC,iBAAiB,EAAE,GAAG,CAAC,CAAC,KAAK,CAAC,CAAC,EAAE,EAAE,CAAC,CAAC;IAAA,CACvD,CAAC;IAEF,MAAM,mBAAmB,GAAG,iBAAiB,CAAC,OAAO,CAAC,QAAQ,EAAE,KAAK,EAAE,mBAAmB,CAAC,CAAC;IAE5F,KAAK,MAAM,GAAG,IAAI,mBAAmB,EAAE,CAAC;QACvC,IAAI,GAAG,CAAC,IAAI,KAAK,MAAM,EAAE,CAAC;YACzB,IAAI,OAAO,GAAG,CAAC,OAAO,KAAK,QAAQ,EAAE,CAAC;gBACrC,QAAQ,CAAC,IAAI,CAAC;oBACb,IAAI,EAAE,MAAM;oBACZ,KAAK,EAAE,CAAC,EAAE,IAAI,EAAE,kBAAkB,CAAC,GAAG,CAAC,OAAO,CAAC,EAAE,CAAC;iBAClD,CAAC,CAAC;YACJ,CAAC;iBAAM,CAAC;gBACP,MAAM,KAAK,GAAW,GAAG,CAAC,OAAO,CAAC,GAAG,CAAC,CAAC,IAAI,EAAE,EAAE,CAAC;oBAC/C,IAAI,IAAI,CAAC,IAAI,KAAK,MAAM,EAAE,CAAC;wBAC1B,OAAO,EAAE,IAAI,EAAE,kBAAkB,CAAC,IAAI,CAAC,IAAI,CAAC,EAAE,CAAC;oBAChD,CAAC;yBAAM,CAAC;wBACP,OAAO;4BACN,UAAU,EAAE;gCACX,QAAQ,EAAE,IAAI,CAAC,QAAQ;gCACvB,IAAI,EAAE,IAAI,CAAC,IAAI;6BACf;yBACD,CAAC;oBACH,CAAC;gBAAA,CACD,CAAC,CAAC;gBACH,IAAI,KAAK,CAAC,MAAM,KAAK,CAAC;oBAAE,SAAS;gBACjC,QAAQ,CAAC,IAAI,CAAC;oBACb,IAAI,EAAE,MAAM;oBACZ,KAAK;iBACL,CAAC,CAAC;YACJ,CAAC;QACF,CAAC;aAAM,IAAI,GAAG,CAAC,IAAI,KAAK,WAAW,EAAE,CAAC;YACrC,MAAM,KAAK,GAAW,EAAE,CAAC;YACzB,oFAAoF;YACpF,MAAM,sBAAsB,GAAG,GAAG,CAAC,QAAQ,KAAK,KAAK,CAAC,QAAQ,IAAI,GAAG,CAAC,KAAK,KAAK,KAAK,CAAC,EAAE,CAAC;YAEzF,KAAK,MAAM,KAAK,IAAI,GAAG,CAAC,OAAO,EAAE,CAAC;gBACjC,IAAI,KAAK,CAAC,IAAI,KAAK,MAAM,EAAE,CAAC;oBAC3B,MAAM,gBAAgB,GAAG,uBAAuB,CAAC,sBAAsB,EAAE,KAAK,CAAC,aAAa,CAAC,CAAC;oBAC9F,sFAAoF;oBACpF,mFAAmF;oBACnF,oFAAoF;oBACpF,mEAAmE;oBACnE,IAAI,CAAC,CAAC,KAAK,CAAC,IAAI,IAAI,KAAK,CAAC,IAAI,CAAC,IAAI,EAAE,KAAK,EAAE,CAAC,IAAI,CAAC,gBAAgB;wBAAE,SAAS;oBAC7E,KAAK,CAAC,IAAI,CAAC;wBACV,IAAI,EAAE,kBAAkB,CAAC,KAAK,CAAC,IAAI,CAAC;wBACpC,GAAG,CAAC,gBAAgB,IAAI,EAAE,gBAAgB,EAAE,CAAC;qBAC7C,CAAC,CAAC;gBACJ,CAAC;qBAAM,IAAI,KAAK,CAAC,IAAI,KAAK,UAAU,EAAE,CAAC;oBACtC,8DAA8D;oBAC9D,0EAA0E;oBAC1E,IAAI,sBAAsB,EAAE,CAAC;wBAC5B,MAAM,gBAAgB,GAAG,uBAAuB,CAAC,sBAAsB,EAAE,KAAK,CAAC,iBAAiB,CAAC,CAAC;wBAClG,4EAA4E;wBAC5E,qEAAqE;wBACrE,IAAI,CAAC,CAAC,KAAK,CAAC,QAAQ,IAAI,KAAK,CAAC,QAAQ,CAAC,IAAI,EAAE,KAAK,EAAE,CAAC,IAAI,CAAC,gBAAgB;4BAAE,SAAS;wBACrF,KAAK,CAAC,IAAI,CAAC;4BACV,OAAO,EAAE,IAAI;4BACb,IAAI,EAAE,kBAAkB,CAAC,KAAK,CAAC,QAAQ,CAAC;4BACxC,GAAG,CAAC,gBAAgB,IAAI,EAAE,gBAAgB,EAAE,CAAC;yBAC7C,CAAC,CAAC;oBACJ,CAAC;yBAAM,CAAC;wBACP,8EAA8E;wBAC9E,IAAI,CAAC,KAAK,CAAC,QAAQ,IAAI,KAAK,CAAC,QAAQ,CAAC,IAAI,EAAE,KAAK,EAAE;4BAAE,SAAS;wBAC9D,KAAK,CAAC,IAAI,CAAC;4BACV,IAAI,EAAE,kBAAkB,CAAC,KAAK,CAAC,QAAQ,CAAC;yBACxC,CAAC,CAAC;oBACJ,CAAC;gBACF,CAAC;qBAAM,IAAI,KAAK,CAAC,IAAI,KAAK,UAAU,EAAE,CAAC;oBACtC,MAAM,gBAAgB,GAAG,uBAAuB,CAAC,sBAAsB,EAAE,KAAK,CAAC,gBAAgB,CAAC,CAAC;oBACjG,MAAM,IAAI,GAAS;wBAClB,YAAY,EAAE;4BACb,IAAI,EAAE,KAAK,CAAC,IAAI;4BAChB,IAAI,EAAE,KAAK,CAAC,SAAS,IAAI,EAAE;4BAC3B,GAAG,CAAC,kBAAkB,CAAC,KAAK,CAAC,EAAE,CAAC,CAAC,CAAC,CAAC,EAAE,EAAE,EAAE,KAAK,CAAC,EAAE,EAAE,CAAC,CAAC,CAAC,EAAE,CAAC;yBACzD;wBACD,GAAG,CAAC,gBAAgB,IAAI,EAAE,gBAAgB,EAAE,CAAC;qBAC7C,CAAC;oBACF,KAAK,CAAC,IAAI,CAAC,IAAI,CAAC,CAAC;gBAClB,CAAC;YACF,CAAC;YAED,IAAI,KAAK,CAAC,MAAM,KAAK,CAAC;gBAAE,SAAS;YACjC,QAAQ,CAAC,IAAI,CAAC;gBACb,IAAI,EAAE,OAAO;gBACb,KAAK;aACL,CAAC,CAAC;QACJ,CAAC;aAAM,IAAI,GAAG,CAAC,IAAI,KAAK,YAAY,EAAE,CAAC;YACtC,iCAAiC;YACjC,MAAM,WAAW,GAAG,GAAG,CAAC,OAAO,CAAC,MAAM,CAAC,CAAC,CAAC,EAAoB,EAAE,CAAC,CAAC,CAAC,IAAI,KAAK,MAAM,CAAC,CAAC;YACnF,MAAM,UAAU,GAAG,WAAW,CAAC,GAAG,CAAC,CAAC,CAAC,EAAE,EAAE,CAAC,CAAC,CAAC,IAAI,CAAC,CAAC,IAAI,CAAC,IAAI,CAAC,CAAC;YAC7D,MAAM,YAAY,GAAG,KAAK,CAAC,KAAK,CAAC,QAAQ,CAAC,OAAO,CAAC;gBACjD,CAAC,CAAC,GAAG,CAAC,OAAO,CAAC,MAAM,CAAC,CAAC,CAAC,EAAqB,EAAE,CAAC,CAAC,CAAC,IAAI,KAAK,OAAO,CAAC;gBAClE,CAAC,CAAC,EAAE,CAAC;YAEN,MAAM,OAAO,GAAG,UAAU,CAAC,MAAM,GAAG,CAAC,CAAC;YACtC,MAAM,SAAS,GAAG,YAAY,CAAC,MAAM,GAAG,CAAC,CAAC;YAE1C,mFAAmF;YACnF,wFAAwF;YACxF,qDAAqD;YACrD,MAAM,uCAAuC,GAAG,kCAAkC,CAAC,KAAK,CAAC,EAAE,CAAC,CAAC;YAE7F,gFAAgF;YAChF,MAAM,aAAa,GAAG,OAAO,CAAC,CAAC,CAAC,kBAAkB,CAAC,UAAU,CAAC,CAAC,CAAC,CAAC,SAAS,CAAC,CAAC,CAAC,sBAAsB,CAAC,CAAC,CAAC,EAAE,CAAC;YAEzG,MAAM,UAAU,GAAW,YAAY,CAAC,GAAG,CAAC,CAAC,UAAU,EAAE,EAAE,CAAC,CAAC;gBAC5D,UAAU,EAAE;oBACX,QAAQ,EAAE,UAAU,CAAC,QAAQ;oBAC7B,IAAI,EAAE,UAAU,CAAC,IAAI;iBACrB;aACD,CAAC,CAAC,CAAC;YAEJ,MAAM,SAAS,GAAG,kBAAkB,CAAC,KAAK,CAAC,EAAE,CAAC,CAAC;YAC/C,MAAM,oBAAoB,GAAS;gBAClC,gBAAgB,EAAE;oBACjB,IAAI,EAAE,GAAG,CAAC,QAAQ;oBAClB,QAAQ,EAAE,GAAG,CAAC,OAAO,CAAC,CAAC,CAAC,EAAE,KAAK,EAAE,aAAa,EAAE,CAAC,CAAC,CAAC,EAAE,MAAM,EAAE,aAAa,EAAE;oBAC5E,GAAG,CAAC,SAAS,IAAI,uCAAuC,IAAI,EAAE,KAAK,EAAE,UAAU,EAAE,CAAC;oBAClF,GAAG,CAAC,SAAS,CAAC,CAAC,CAAC,EAAE,EAAE,EAAE,GAAG,CAAC,UAAU,EAAE,CAAC,CAAC,CAAC,EAAE,CAAC;iBAC5C;aACD,CAAC;YAEF,qFAAqF;YACrF,sFAAsF;YACtF,MAAM,WAAW,GAAG,QAAQ,CAAC,QAAQ,CAAC,MAAM,GAAG,CAAC,CAAC,CAAC;YAClD,IAAI,WAAW,EAAE,IAAI,KAAK,MAAM,IAAI,WAAW,CAAC,KAAK,EAAE,IAAI,CAAC,CAAC,CAAC,EAAE,EAAE,CAAC,CAAC,CAAC,gBAAgB,CAAC,EAAE,CAAC;gBACxF,WAAW,CAAC,KAAK,CAAC,IAAI,CAAC,oBAAoB,CAAC,CAAC;YAC9C,CAAC;iBAAM,CAAC;gBACP,QAAQ,CAAC,IAAI,CAAC;oBACb,IAAI,EAAE,MAAM;oBACZ,KAAK,EAAE,CAAC,oBAAoB,CAAC;iBAC7B,CAAC,CAAC;YACJ,CAAC;YAED,wDAAwD;YACxD,IAAI,SAAS,IAAI,CAAC,uCAAuC,EAAE,CAAC;gBAC3D,QAAQ,CAAC,IAAI,CAAC;oBACb,IAAI,EAAE,MAAM;oBACZ,KAAK,EAAE,CAAC,EAAE,IAAI,EAAE,oBAAoB,EAAE,EAAE,GAAG,UAAU,CAAC;iBACtD,CAAC,CAAC;YACJ,CAAC;QACF,CAAC;IACF,CAAC;IAED,OAAO,QAAQ,CAAC;AAAA,CAChB;AAED,MAAM,6BAA6B,GAAG,IAAI,GAAG,CAAC;IAC7C,SAAS;IACT,KAAK;IACL,SAAS;IACT,gBAAgB;IAChB,aAAa;IACb,UAAU;IACV,OAAO;IACP,aAAa,EAAE,wCAAwC;CACvD,CAAC,CAAC;AAEH;;GAEG;AACH,SAAS,kBAAkB,CAAC,MAAe,EAAW;IACrD,IAAI,OAAO,MAAM,KAAK,QAAQ,IAAI,MAAM,KAAK,IAAI,IAAI,KAAK,CAAC,OAAO,CAAC,MAAM,CAAC,EAAE,CAAC;QAC5E,OAAO,MAAM,CAAC;IACf,CAAC;IAED,MAAM,MAAM,GAA4B,EAAE,CAAC;IAC3C,KAAK,MAAM,CAAC,GAAG,EAAE,KAAK,CAAC,IAAI,MAAM,CAAC,OAAO,CAAC,MAAM,CAAC,EAAE,CAAC;QACnD,IAAI,6BAA6B,CAAC,GAAG,CAAC,GAAG,CAAC;YAAE,SAAS;QACrD,MAAM,CAAC,GAAG,CAAC,GAAG,kBAAkB,CAAC,KAAK,CAAC,CAAC;IACzC,CAAC;IACD,OAAO,MAAM,CAAC;AAAA,CACd;AAED;;;;;;;GAOG;AACH,MAAM,UAAU,YAAY,CAC3B,KAAa,EACb,aAAa,GAAG,KAAK,EACrB,kBAAkB,GAAG,IAAI,EAC2C;IACpE,IAAI,KAAK,CAAC,MAAM,KAAK,CAAC;QAAE,OAAO,SAAS,CAAC;IACzC,OAAO;QACN;YACC,oBAAoB,EAAE,KAAK,CAAC,GAAG,CAAC,CAAC,IAAI,EAAE,EAAE,CAAC;gBACzC,MAAM,MAAM,GAAG,+BAA+B,CAAC,IAAI,EAAE,kBAAkB,CAAC,CAAC;gBACzE,MAAM,UAAU,GAAG,2BAA2B,CAAC,IAAI,EAAE,MAAM,CAAC,CAAC;gBAC7D,OAAO;oBACN,IAAI,EAAE,IAAI,CAAC,IAAI;oBACf,WAAW,EAAE,IAAI,CAAC,WAAW;oBAC7B,GAAG,CAAC,aAAa;wBAChB,CAAC,CAAC,EAAE,UAAU,EAAE,kBAAkB,CAAC,UAAqB,CAAC,EAAE;wBAC3D,CAAC,CAAC,EAAE,oBAAoB,EAAE,UAAU,EAAE,CAAC;iBACxC,CAAC;YAAA,CACF,CAAC;SACF;KACD,CAAC;AAAA,CACF;AAED,uFAAuF;AACvF,MAAM,UAAU,gCAAgC,CAAC,OAAe,EAAW;IAC1E,MAAM,YAAY,GAAG,qBAAqB,CAAC,OAAO,CAAC,CAAC;IACpD,OAAO,YAAY,KAAK,SAAS,IAAI,YAAY,IAAI,CAAC,CAAC;AAAA,CACvD;AAED,kEAAkE;AAClE,MAAM,UAAU,aAAa,CAAC,MAAc,EAA6B;IACxE,QAAQ,MAAM,EAAE,CAAC;QAChB,KAAK,MAAM;YACV,OAAO,yBAAyB,CAAC,IAAI,CAAC;QACvC,KAAK,MAAM;YACV,OAAO,yBAAyB,CAAC,IAAI,CAAC;QACvC,KAAK,KAAK;YACT,OAAO,yBAAyB,CAAC,GAAG,CAAC;QACtC;YACC,OAAO,yBAAyB,CAAC,IAAI,CAAC;IACxC,CAAC;AAAA,CACD;AAED,MAAM,UAAU,gCAAgC,CAC/C,KAAa,EACb,UAA8B,EAC9B,kBAA2B,EACa;IACxC,MAAM,aAAa,GAAG,KAAK,CAAC,IAAI,CAAC,CAAC,IAAI,EAAE,EAAE,CAAC,+BAA+B,CAAC,IAAI,EAAE,kBAAkB,CAAC,KAAK,IAAI,CAAC,CAAC;IAC/G,IAAI,UAAU,KAAK,MAAM,IAAI,UAAU,KAAK,KAAK,EAAE,CAAC;QACnD,OAAO,aAAa,CAAC,UAAU,CAAC,CAAC;IAClC,CAAC;IACD,IAAI,aAAa,EAAE,CAAC;QACnB,OAAO,yBAAyB,CAAC,SAAS,CAAC;IAC5C,CAAC;IACD,OAAO,UAAU,CAAC,CAAC,CAAC,aAAa,CAAC,UAAU,CAAC,CAAC,CAAC,CAAC,SAAS,CAAC;AAAA,CAC1D;AAED;;GAEG;AACH,MAAM,UAAU,aAAa,CAAC,MAAoB,EAAc;IAC/D,QAAQ,MAAM,EAAE,CAAC;QAChB,KAAK,YAAY,CAAC,IAAI;YACrB,OAAO,MAAM,CAAC;QACf,KAAK,YAAY,CAAC,UAAU;YAC3B,OAAO,QAAQ,CAAC;QACjB,KAAK,YAAY,CAAC,SAAS,CAAC;QAC5B,KAAK,YAAY,CAAC,kBAAkB,CAAC;QACrC,KAAK,YAAY,CAAC,IAAI,CAAC;QACvB,KAAK,YAAY,CAAC,MAAM,CAAC;QACzB,KAAK,YAAY,CAAC,YAAY,CAAC;QAC/B,KAAK,YAAY,CAAC,wBAAwB,CAAC;QAC3C,KAAK,YAAY,CAAC,gBAAgB,CAAC;QACnC,KAAK,YAAY,CAAC,WAAW,CAAC;QAC9B,KAAK,YAAY,CAAC,UAAU,CAAC;QAC7B,KAAK,YAAY,CAAC,yBAAyB,CAAC;QAC5C,KAAK,YAAY,CAAC,KAAK,CAAC;QACxB,KAAK,YAAY,CAAC,QAAQ,CAAC;QAC3B,KAAK,YAAY,CAAC,uBAAuB,CAAC;QAC1C,KAAK,YAAY,CAAC,oBAAoB,CAAC;QACvC,KAAK,YAAY,CAAC,QAAQ;YACzB,OAAO,OAAO,CAAC;QAChB,SAAS,CAAC;YACT,MAAM,WAAW,GAAU,MAAM,CAAC;YAClC,MAAM,IAAI,KAAK,CAAC,0BAA0B,WAAW,EAAE,CAAC,CAAC;QAC1D,CAAC;IACF,CAAC;AAAA,CACD;AAED;;GAEG;AACH,MAAM,UAAU,mBAAmB,CAAC,MAAc,EAAc;IAC/D,QAAQ,MAAM,EAAE,CAAC;QAChB,KAAK,MAAM;YACV,OAAO,MAAM,CAAC;QACf,KAAK,YAAY;YAChB,OAAO,QAAQ,CAAC;QACjB;YACC,OAAO,OAAO,CAAC;IACjB,CAAC;AAAA,CACD;AAED;;;;;;;;GAQG;AACH,MAAM,UAAU,kBAAkB,CACjC,OAAyB,EACzB,OAA0E,EAC7D;IACb,OAAO,oBAAoB,CAC1B,KAAK,IAAI,EAAE,CAAC;QACX,IAAI,CAAC;YACJ,OAAO,MAAM,OAAO,EAAE,CAAC;QACxB,CAAC;QAAC,OAAO,KAAK,EAAE,CAAC;YAChB,IAAI,KAAK,YAAY,KAAK,IAAI,QAAQ,IAAI,KAAK,IAAI,CAAC,CAAC,SAAS,IAAI,KAAK,CAAC,EAAE,CAAC;gBACzE,KAA+B,CAAC,OAAO,GAAG,SAAS,CAAC;YACtD,CAAC;YACD,MAAM,KAAK,CAAC;QACb,CAAC;IAAA,CACD,EACD;QACC,UAAU,EAAE,OAAO,EAAE,UAAU;QAC/B,eAAe,EAAE,OAAO,EAAE,eAAe;QACzC,MAAM,EAAE,OAAO,EAAE,MAAM;KACvB,CACD,CAAC;AAAA,CACF","sourcesContent":["/**\n * Shared utilities for Google Generative AI and Google Vertex providers.\n */\n\nimport { type Content, FinishReason, FunctionCallingConfigMode, type Part } from \"@google/genai\";\nimport type { Context, ImageContent, Model, StopReason, StreamOptions, TextContent, Tool } from \"../types.ts\";\nimport { retryProviderRequest } from \"../utils/provider-retry.ts\";\nimport { sanitizeSurrogates } from \"../utils/sanitize-unicode.ts\";\nimport { getJsonSchemaToolParameters, resolveJsonSchemaStrictSampling } from \"./constrained-sampling.ts\";\nimport { transformMessages } from \"./transform-messages.ts\";\n\ntype GoogleApiType = \"google-generative-ai\" | \"google-vertex\";\n\n/**\n * Thinking level for Gemini 3 models.\n * Mirrors Google's ThinkingLevel enum values.\n */\nexport type GoogleThinkingLevel = \"THINKING_LEVEL_UNSPECIFIED\" | \"MINIMAL\" | \"LOW\" | \"MEDIUM\" | \"HIGH\";\n\n/**\n * Determines whether a streamed Gemini `Part` should be treated as \"thinking\".\n *\n * Protocol note (Gemini / Vertex AI thought signatures):\n * - `thought: true` is the definitive marker for thinking content (thought summaries).\n * - `thoughtSignature` is an encrypted representation of the model's internal thought process\n *   used to preserve reasoning context across multi-turn interactions.\n * - `thoughtSignature` can appear on ANY part type (text, functionCall, etc.) - it does NOT\n *   indicate the part itself is thinking content.\n * - For non-functionCall responses, the signature appears on the last part for context replay.\n * - When persisting/replaying model outputs, signature-bearing parts must be preserved as-is;\n *   do not merge/move signatures across parts.\n *\n * See: https://ai.google.dev/gemini-api/docs/thought-signatures\n */\nexport function isThinkingPart(part: Pick<Part, \"thought\" | \"thoughtSignature\">): boolean {\n\treturn part.thought === true;\n}\n\n/**\n * Retain thought signatures during streaming.\n *\n * Some backends only send `thoughtSignature` on the first delta for a given part/block; later deltas may omit it.\n * This helper preserves the last non-empty signature for the current block.\n *\n * Note: this does NOT merge or move signatures across distinct response parts. It only prevents\n * a signature from being overwritten with `undefined` within the same streamed block.\n */\nexport function retainThoughtSignature(existing: string | undefined, incoming: string | undefined): string | undefined {\n\tif (typeof incoming === \"string\" && incoming.length > 0) return incoming;\n\treturn existing;\n}\n\n// Thought signatures must be base64 for Google APIs (TYPE_BYTES).\nconst base64SignaturePattern = /^[A-Za-z0-9+/]+={0,2}$/;\n\nfunction isValidThoughtSignature(signature: string | undefined): boolean {\n\tif (!signature) return false;\n\tif (signature.length % 4 !== 0) return false;\n\treturn base64SignaturePattern.test(signature);\n}\n\n/**\n * Only keep signatures from the same provider/model and with valid base64.\n */\nfunction resolveThoughtSignature(isSameProviderAndModel: boolean, signature: string | undefined): string | undefined {\n\treturn isSameProviderAndModel && isValidThoughtSignature(signature) ? signature : undefined;\n}\n\n/**\n * Models via Google APIs that require explicit tool call IDs in function calls/responses.\n */\nexport function requiresToolCallId(modelId: string): boolean {\n\tconst geminiMajorVersion = getGeminiMajorVersion(modelId);\n\treturn (\n\t\tmodelId.startsWith(\"claude-\") ||\n\t\tmodelId.startsWith(\"gpt-oss-\") ||\n\t\t(geminiMajorVersion !== undefined && geminiMajorVersion >= 3)\n\t);\n}\n\nfunction getGeminiMajorVersion(modelId: string): number | undefined {\n\tconst match = modelId.toLowerCase().match(/^gemini(?:-live)?-(\\d+)/);\n\tif (!match) return undefined;\n\treturn Number.parseInt(match[1], 10);\n}\n\nfunction supportsMultimodalFunctionResponse(modelId: string): boolean {\n\tconst geminiMajorVersion = getGeminiMajorVersion(modelId);\n\tif (geminiMajorVersion !== undefined) {\n\t\treturn geminiMajorVersion >= 3;\n\t}\n\treturn true;\n}\n\n/**\n * Convert internal messages to Gemini Content[] format.\n */\nexport function convertMessages<T extends GoogleApiType>(model: Model<T>, context: Context): Content[] {\n\tconst contents: Content[] = [];\n\tconst normalizeToolCallId = (id: string): string => {\n\t\tif (!requiresToolCallId(model.id)) return id;\n\t\treturn id.replace(/[^a-zA-Z0-9_-]/g, \"_\").slice(0, 64);\n\t};\n\n\tconst transformedMessages = transformMessages(context.messages, model, normalizeToolCallId);\n\n\tfor (const msg of transformedMessages) {\n\t\tif (msg.role === \"user\") {\n\t\t\tif (typeof msg.content === \"string\") {\n\t\t\t\tcontents.push({\n\t\t\t\t\trole: \"user\",\n\t\t\t\t\tparts: [{ text: sanitizeSurrogates(msg.content) }],\n\t\t\t\t});\n\t\t\t} else {\n\t\t\t\tconst parts: Part[] = msg.content.map((item) => {\n\t\t\t\t\tif (item.type === \"text\") {\n\t\t\t\t\t\treturn { text: sanitizeSurrogates(item.text) };\n\t\t\t\t\t} else {\n\t\t\t\t\t\treturn {\n\t\t\t\t\t\t\tinlineData: {\n\t\t\t\t\t\t\t\tmimeType: item.mimeType,\n\t\t\t\t\t\t\t\tdata: item.data,\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\tif (parts.length === 0) continue;\n\t\t\t\tcontents.push({\n\t\t\t\t\trole: \"user\",\n\t\t\t\t\tparts,\n\t\t\t\t});\n\t\t\t}\n\t\t} else if (msg.role === \"assistant\") {\n\t\t\tconst parts: Part[] = [];\n\t\t\t// Check if message is from same provider and model - only then keep thinking blocks\n\t\t\tconst isSameProviderAndModel = msg.provider === model.provider && msg.model === model.id;\n\n\t\t\tfor (const block of msg.content) {\n\t\t\t\tif (block.type === \"text\") {\n\t\t\t\t\tconst thoughtSignature = resolveThoughtSignature(isSameProviderAndModel, block.textSignature);\n\t\t\t\t\t// Skip empty text blocks — unless they carry a thought signature. Gemini can attach\n\t\t\t\t\t// the signature to a part whose visible text is empty and requires it echoed back;\n\t\t\t\t\t// dropping it breaks the reasoning chain and the model intermittently ends mid-task\n\t\t\t\t\t// turns with a thought-only STOP (empty completion, no tool call).\n\t\t\t\t\tif ((!block.text || block.text.trim() === \"\") && !thoughtSignature) continue;\n\t\t\t\t\tparts.push({\n\t\t\t\t\t\ttext: sanitizeSurrogates(block.text),\n\t\t\t\t\t\t...(thoughtSignature && { thoughtSignature }),\n\t\t\t\t\t});\n\t\t\t\t} else if (block.type === \"thinking\") {\n\t\t\t\t\t// Only keep as thinking block if same provider AND same model\n\t\t\t\t\t// Otherwise convert to plain text (no tags to avoid model mimicking them)\n\t\t\t\t\tif (isSameProviderAndModel) {\n\t\t\t\t\t\tconst thoughtSignature = resolveThoughtSignature(isSameProviderAndModel, block.thinkingSignature);\n\t\t\t\t\t\t// Same rule as text blocks: an empty thinking block is dropped only when it\n\t\t\t\t\t\t// carries no signature (mirrors the anthropic converter's handling).\n\t\t\t\t\t\tif ((!block.thinking || block.thinking.trim() === \"\") && !thoughtSignature) continue;\n\t\t\t\t\t\tparts.push({\n\t\t\t\t\t\t\tthought: true,\n\t\t\t\t\t\t\ttext: sanitizeSurrogates(block.thinking),\n\t\t\t\t\t\t\t...(thoughtSignature && { thoughtSignature }),\n\t\t\t\t\t\t});\n\t\t\t\t\t} else {\n\t\t\t\t\t\t// Cross-provider/model: the signature is unusable, empty blocks stay dropped.\n\t\t\t\t\t\tif (!block.thinking || block.thinking.trim() === \"\") continue;\n\t\t\t\t\t\tparts.push({\n\t\t\t\t\t\t\ttext: sanitizeSurrogates(block.thinking),\n\t\t\t\t\t\t});\n\t\t\t\t\t}\n\t\t\t\t} else if (block.type === \"toolCall\") {\n\t\t\t\t\tconst thoughtSignature = resolveThoughtSignature(isSameProviderAndModel, block.thoughtSignature);\n\t\t\t\t\tconst part: Part = {\n\t\t\t\t\t\tfunctionCall: {\n\t\t\t\t\t\t\tname: block.name,\n\t\t\t\t\t\t\targs: block.arguments ?? {},\n\t\t\t\t\t\t\t...(requiresToolCallId(model.id) ? { id: block.id } : {}),\n\t\t\t\t\t\t},\n\t\t\t\t\t\t...(thoughtSignature && { thoughtSignature }),\n\t\t\t\t\t};\n\t\t\t\t\tparts.push(part);\n\t\t\t\t}\n\t\t\t}\n\n\t\t\tif (parts.length === 0) continue;\n\t\t\tcontents.push({\n\t\t\t\trole: \"model\",\n\t\t\t\tparts,\n\t\t\t});\n\t\t} else if (msg.role === \"toolResult\") {\n\t\t\t// Extract text and image content\n\t\t\tconst textContent = msg.content.filter((c): c is TextContent => c.type === \"text\");\n\t\t\tconst textResult = textContent.map((c) => c.text).join(\"\\n\");\n\t\t\tconst imageContent = model.input.includes(\"image\")\n\t\t\t\t? msg.content.filter((c): c is ImageContent => c.type === \"image\")\n\t\t\t\t: [];\n\n\t\t\tconst hasText = textResult.length > 0;\n\t\t\tconst hasImages = imageContent.length > 0;\n\n\t\t\t// Gemini 3+ models support multimodal function responses with images nested inside\n\t\t\t// functionResponse.parts. Claude and other non-Gemini models behind Cloud Code Assist /\n\t\t\t// Gemini < 3 still needs a separate user image turn.\n\t\t\tconst modelSupportsMultimodalFunctionResponse = supportsMultimodalFunctionResponse(model.id);\n\n\t\t\t// Use \"output\" key for success, \"error\" key for errors as per SDK documentation\n\t\t\tconst responseValue = hasText ? sanitizeSurrogates(textResult) : hasImages ? \"(see attached image)\" : \"\";\n\n\t\t\tconst imageParts: Part[] = imageContent.map((imageBlock) => ({\n\t\t\t\tinlineData: {\n\t\t\t\t\tmimeType: imageBlock.mimeType,\n\t\t\t\t\tdata: imageBlock.data,\n\t\t\t\t},\n\t\t\t}));\n\n\t\t\tconst includeId = requiresToolCallId(model.id);\n\t\t\tconst functionResponsePart: Part = {\n\t\t\t\tfunctionResponse: {\n\t\t\t\t\tname: msg.toolName,\n\t\t\t\t\tresponse: msg.isError ? { error: responseValue } : { output: responseValue },\n\t\t\t\t\t...(hasImages && modelSupportsMultimodalFunctionResponse && { parts: imageParts }),\n\t\t\t\t\t...(includeId ? { id: msg.toolCallId } : {}),\n\t\t\t\t},\n\t\t\t};\n\n\t\t\t// Cloud Code Assist API requires all function responses to be in a single user turn.\n\t\t\t// Check if the last content is already a user turn with function responses and merge.\n\t\t\tconst lastContent = contents[contents.length - 1];\n\t\t\tif (lastContent?.role === \"user\" && lastContent.parts?.some((p) => p.functionResponse)) {\n\t\t\t\tlastContent.parts.push(functionResponsePart);\n\t\t\t} else {\n\t\t\t\tcontents.push({\n\t\t\t\t\trole: \"user\",\n\t\t\t\t\tparts: [functionResponsePart],\n\t\t\t\t});\n\t\t\t}\n\n\t\t\t// For Gemini < 3, add images in a separate user message\n\t\t\tif (hasImages && !modelSupportsMultimodalFunctionResponse) {\n\t\t\t\tcontents.push({\n\t\t\t\t\trole: \"user\",\n\t\t\t\t\tparts: [{ text: \"Tool result image:\" }, ...imageParts],\n\t\t\t\t});\n\t\t\t}\n\t\t}\n\t}\n\n\treturn contents;\n}\n\nconst JSON_SCHEMA_META_DECLARATIONS = new Set([\n\t\"$schema\",\n\t\"$id\",\n\t\"$anchor\",\n\t\"$dynamicAnchor\",\n\t\"$vocabulary\",\n\t\"$comment\",\n\t\"$defs\",\n\t\"definitions\", // pre-draft-2019-09 equivalent of $defs\n]);\n\n/**\n * Strip meta-declarations from a schema obj\n */\nfunction sanitizeForOpenApi(schema: unknown): unknown {\n\tif (typeof schema !== \"object\" || schema === null || Array.isArray(schema)) {\n\t\treturn schema;\n\t}\n\n\tconst result: Record<string, unknown> = {};\n\tfor (const [key, value] of Object.entries(schema)) {\n\t\tif (JSON_SCHEMA_META_DECLARATIONS.has(key)) continue;\n\t\tresult[key] = sanitizeForOpenApi(value);\n\t}\n\treturn result;\n}\n\n/**\n * Convert tools to Gemini function declarations format.\n *\n * By default uses `parametersJsonSchema` which supports full JSON Schema (including\n * anyOf, oneOf, const, etc.). Set `useParameters` to true to use the legacy `parameters`\n * field instead (OpenAPI 3.03 Schema). This is needed for Cloud Code Assist with Claude\n * models, where the API translates `parameters` into Anthropic's `input_schema`.\n */\nexport function convertTools(\n\ttools: Tool[],\n\tuseParameters = false,\n\tsupportsStrictMode = true,\n): { functionDeclarations: Record<string, unknown>[] }[] | undefined {\n\tif (tools.length === 0) return undefined;\n\treturn [\n\t\t{\n\t\t\tfunctionDeclarations: tools.map((tool) => {\n\t\t\t\tconst strict = resolveJsonSchemaStrictSampling(tool, supportsStrictMode);\n\t\t\t\tconst parameters = getJsonSchemaToolParameters(tool, strict);\n\t\t\t\treturn {\n\t\t\t\t\tname: tool.name,\n\t\t\t\t\tdescription: tool.description,\n\t\t\t\t\t...(useParameters\n\t\t\t\t\t\t? { parameters: sanitizeForOpenApi(parameters as unknown) }\n\t\t\t\t\t\t: { parametersJsonSchema: parameters }),\n\t\t\t\t};\n\t\t\t}),\n\t\t},\n\t];\n}\n\n/** Gemini 3+ enforces required function parameters in validated tool-calling modes. */\nexport function supportsGoogleStrictToolSampling(modelId: string): boolean {\n\tconst majorVersion = getGeminiMajorVersion(modelId);\n\treturn majorVersion !== undefined && majorVersion >= 3;\n}\n\n/** Map tool choice string to Gemini FunctionCallingConfigMode. */\nexport function mapToolChoice(choice: string): FunctionCallingConfigMode {\n\tswitch (choice) {\n\t\tcase \"auto\":\n\t\t\treturn FunctionCallingConfigMode.AUTO;\n\t\tcase \"none\":\n\t\t\treturn FunctionCallingConfigMode.NONE;\n\t\tcase \"any\":\n\t\t\treturn FunctionCallingConfigMode.ANY;\n\t\tdefault:\n\t\t\treturn FunctionCallingConfigMode.AUTO;\n\t}\n}\n\nexport function resolveGoogleFunctionCallingMode(\n\ttools: Tool[],\n\ttoolChoice: string | undefined,\n\tsupportsStrictMode: boolean,\n): FunctionCallingConfigMode | undefined {\n\tconst useStrictMode = tools.some((tool) => resolveJsonSchemaStrictSampling(tool, supportsStrictMode) === true);\n\tif (toolChoice === \"none\" || toolChoice === \"any\") {\n\t\treturn mapToolChoice(toolChoice);\n\t}\n\tif (useStrictMode) {\n\t\treturn FunctionCallingConfigMode.VALIDATED;\n\t}\n\treturn toolChoice ? mapToolChoice(toolChoice) : undefined;\n}\n\n/**\n * Map Gemini FinishReason to our StopReason.\n */\nexport function mapStopReason(reason: FinishReason): StopReason {\n\tswitch (reason) {\n\t\tcase FinishReason.STOP:\n\t\t\treturn \"stop\";\n\t\tcase FinishReason.MAX_TOKENS:\n\t\t\treturn \"length\";\n\t\tcase FinishReason.BLOCKLIST:\n\t\tcase FinishReason.PROHIBITED_CONTENT:\n\t\tcase FinishReason.SPII:\n\t\tcase FinishReason.SAFETY:\n\t\tcase FinishReason.IMAGE_SAFETY:\n\t\tcase FinishReason.IMAGE_PROHIBITED_CONTENT:\n\t\tcase FinishReason.IMAGE_RECITATION:\n\t\tcase FinishReason.IMAGE_OTHER:\n\t\tcase FinishReason.RECITATION:\n\t\tcase FinishReason.FINISH_REASON_UNSPECIFIED:\n\t\tcase FinishReason.OTHER:\n\t\tcase FinishReason.LANGUAGE:\n\t\tcase FinishReason.MALFORMED_FUNCTION_CALL:\n\t\tcase FinishReason.UNEXPECTED_TOOL_CALL:\n\t\tcase FinishReason.NO_IMAGE:\n\t\t\treturn \"error\";\n\t\tdefault: {\n\t\t\tconst _exhaustive: never = reason;\n\t\t\tthrow new Error(`Unhandled stop reason: ${_exhaustive}`);\n\t\t}\n\t}\n}\n\n/**\n * Map string finish reason to our StopReason (for raw API responses).\n */\nexport function mapStopReasonString(reason: string): StopReason {\n\tswitch (reason) {\n\t\tcase \"STOP\":\n\t\t\treturn \"stop\";\n\t\tcase \"MAX_TOKENS\":\n\t\t\treturn \"length\";\n\t\tdefault:\n\t\t\treturn \"error\";\n\t}\n}\n\n/**\n * Run a Google GenAI SDK request with the shared provider retry policy\n * (408/409/429/5xx with backoff, honoring retry-after), mirroring how the\n * Anthropic and OpenAI adapters wrap their initial request in\n * retryProviderRequest. The SDK's ApiError has a `status` property but no\n * `headers` property, and retryProviderRequest only retries errors that carry\n * both, so normalize the error by adding the missing `headers` before\n * rethrowing.\n */\nexport function retryGoogleRequest<T>(\n\trequest: () => Promise<T>,\n\toptions?: Pick<StreamOptions, \"maxRetries\" | \"maxRetryDelayMs\" | \"signal\">,\n): Promise<T> {\n\treturn retryProviderRequest(\n\t\tasync () => {\n\t\t\ttry {\n\t\t\t\treturn await request();\n\t\t\t} catch (error) {\n\t\t\t\tif (error instanceof Error && \"status\" in error && !(\"headers\" in error)) {\n\t\t\t\t\t(error as { headers?: Headers }).headers = undefined;\n\t\t\t\t}\n\t\t\t\tthrow error;\n\t\t\t}\n\t\t},\n\t\t{\n\t\t\tmaxRetries: options?.maxRetries,\n\t\t\tmaxRetryDelayMs: options?.maxRetryDelayMs,\n\t\t\tsignal: options?.signal,\n\t\t},\n\t);\n}\n"]}

package/dist/api/google-vertex.d.ts

@@ -0,0 +1,15 @@
+import type { SimpleStreamOptions, StreamFunction, StreamOptions } from "../types.ts";
+import type { GoogleThinkingLevel } from "./google-shared.ts";
+export interface GoogleVertexOptions extends StreamOptions {
+    toolChoice?: "auto" | "none" | "any";
+    thinking?: {
+        enabled: boolean;
+        budgetTokens?: number;
+        level?: GoogleThinkingLevel;
+    };
+    project?: string;
+    location?: string;
+}
+export declare const stream: StreamFunction<"google-vertex", GoogleVertexOptions>;
+export declare const streamSimple: StreamFunction<"google-vertex", SimpleStreamOptions>;
+//# sourceMappingURL=google-vertex.d.ts.map

package/dist/api/google-vertex.d.ts.map

@@ -0,0 +1 @@
+{"version":3,"file":"google-vertex.d.ts","sourceRoot":"","sources":["../../src/api/google-vertex.ts"],"names":[],"mappings":"AAUA,OAAO,KAAK,EAQX,mBAAmB,EACnB,cAAc,EACd,aAAa,EAKb,MAAM,aAAa,CAAC;AAMrB,OAAO,KAAK,EAAE,mBAAmB,EAAE,MAAM,oBAAoB,CAAC;AAa9D,MAAM,WAAW,mBAAoB,SAAQ,aAAa;IACzD,UAAU,CAAC,EAAE,MAAM,GAAG,MAAM,GAAG,KAAK,CAAC;IACrC,QAAQ,CAAC,EAAE;QACV,OAAO,EAAE,OAAO,CAAC;QACjB,YAAY,CAAC,EAAE,MAAM,CAAC;QACtB,KAAK,CAAC,EAAE,mBAAmB,CAAC;KAC5B,CAAC;IACF,OAAO,CAAC,EAAE,MAAM,CAAC;IACjB,QAAQ,CAAC,EAAE,MAAM,CAAC;CAClB;AAgBD,eAAO,MAAM,MAAM,EAAE,cAAc,CAAC,eAAe,EAAE,mBAAmB,CAiPvE,CAAC;AAEF,eAAO,MAAM,YAAY,EAAE,cAAc,CAAC,eAAe,EAAE,mBAAmB,CAkC7E,CAAC","sourcesContent":["import {\n\ttype GenerateContentConfig,\n\ttype GenerateContentParameters,\n\tGoogleGenAI,\n\ttype HttpOptions,\n\tResourceScope,\n\ttype ThinkingConfig,\n\tThinkingLevel,\n} from \"@google/genai\";\nimport { calculateCost, clampThinkingLevel } from \"../models.ts\";\nimport type {\n\tApi,\n\tAssistantMessage,\n\tContext,\n\tModel,\n\tThinkingLevel as PiThinkingLevel,\n\tProviderEnv,\n\tProviderHeaders,\n\tSimpleStreamOptions,\n\tStreamFunction,\n\tStreamOptions,\n\tTextContent,\n\tThinkingBudgets,\n\tThinkingContent,\n\tToolCall,\n} from \"../types.ts\";\nimport { formatProviderError, normalizeProviderError } from \"../utils/error-body.ts\";\nimport { AssistantMessageEventStream } from \"../utils/event-stream.ts\";\nimport { providerHeadersToRecord } from \"../utils/headers.ts\";\nimport { getProviderEnvValue } from \"../utils/provider-env.ts\";\nimport { sanitizeSurrogates } from \"../utils/sanitize-unicode.ts\";\nimport type { GoogleThinkingLevel } from \"./google-shared.ts\";\nimport {\n\tconvertMessages,\n\tconvertTools,\n\tisThinkingPart,\n\tmapStopReason,\n\tresolveGoogleFunctionCallingMode,\n\tretainThoughtSignature,\n\tretryGoogleRequest,\n\tsupportsGoogleStrictToolSampling,\n} from \"./google-shared.ts\";\nimport { buildBaseOptions } from \"./simple-options.ts\";\n\nexport interface GoogleVertexOptions 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\tproject?: string;\n\tlocation?: string;\n}\n\nconst API_VERSION = \"v1\";\nconst GCP_VERTEX_CREDENTIALS_MARKER = \"gcp-vertex-credentials\";\n\nconst THINKING_LEVEL_MAP: Record<GoogleThinkingLevel, ThinkingLevel> = {\n\tTHINKING_LEVEL_UNSPECIFIED: ThinkingLevel.THINKING_LEVEL_UNSPECIFIED,\n\tMINIMAL: ThinkingLevel.MINIMAL,\n\tLOW: ThinkingLevel.LOW,\n\tMEDIUM: ThinkingLevel.MEDIUM,\n\tHIGH: ThinkingLevel.HIGH,\n};\n\n// Counter for generating unique tool call IDs\nlet toolCallCounter = 0;\n\nexport const stream: StreamFunction<\"google-vertex\", GoogleVertexOptions> = (\n\tmodel: Model<\"google-vertex\">,\n\tcontext: Context,\n\toptions?: GoogleVertexOptions,\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-vertex\" 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: \"pending\",\n\t\t\ttimestamp: Date.now(),\n\t\t};\n\n\t\ttry {\n\t\t\tif (options?.fetch && options.fetch !== globalThis.fetch) {\n\t\t\t\tthrow new Error(\"Custom fetch is not supported by the Google Vertex adapter\");\n\t\t\t}\n\t\t\tconst apiKey = resolveApiKey(options);\n\t\t\t// Create the client using either a Vertex API key, if provided, or ADC with project and location\n\t\t\tconst client = apiKey\n\t\t\t\t? createClientWithApiKey(model, apiKey, options?.headers)\n\t\t\t\t: createClient(model, resolveProject(options), resolveLocation(options), options?.headers, options?.env);\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 retryGoogleRequest(() => client.models.generateContentStream(params), options);\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// Vertex uses the same @google/genai GenerateContentResponse type as Gemini.\n\t\t\t\t// responseId is documented there as an output-only identifier for each response.\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\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.rawStopReason = 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\") && output.stopReason === \"stop\") {\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\treasoning: chunk.usageMetadata.thoughtsTokenCount || 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 === \"pending\") {\n\t\t\t\tthrow new Error(\"Google Vertex stream ended without a finish reason\");\n\t\t\t}\n\t\t\tif (output.stopReason === \"aborted\" || output.stopReason === \"error\") {\n\t\t\t\tconst errorMessage = output.rawStopReason\n\t\t\t\t\t? `Provider stopped with: ${output.rawStopReason}`\n\t\t\t\t\t: \"An unknown error occurred\";\n\t\t\t\tthrow new Error(errorMessage);\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 = formatProviderError(normalizeProviderError(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 streamSimple: StreamFunction<\"google-vertex\", SimpleStreamOptions> = (\n\tmodel: Model<\"google-vertex\">,\n\tcontext: Context,\n\toptions?: SimpleStreamOptions,\n): AssistantMessageEventStream => {\n\tconst base = buildBaseOptions(model, context, options, undefined);\n\tif (!options?.reasoning) {\n\t\treturn stream(model, context, {\n\t\t\t...base,\n\t\t\tthinking: { enabled: false },\n\t\t} satisfies GoogleVertexOptions);\n\t}\n\n\tconst clampedReasoning = clampThinkingLevel(model, options.reasoning);\n\tconst effort = (clampedReasoning === \"off\" ? \"high\" : clampedReasoning) as ClampedThinkingLevel;\n\tconst geminiModel = model as unknown as Model<\"google-generative-ai\">;\n\n\tif (isGemini3ProModel(geminiModel) || isGemini3FlashModel(geminiModel)) {\n\t\treturn stream(model, context, {\n\t\t\t...base,\n\t\t\tthinking: {\n\t\t\t\tenabled: true,\n\t\t\t\tlevel: getGemini3ThinkingLevel(effort, geminiModel),\n\t\t\t},\n\t\t} satisfies GoogleVertexOptions);\n\t}\n\n\treturn stream(model, context, {\n\t\t...base,\n\t\tthinking: {\n\t\t\tenabled: true,\n\t\t\tbudgetTokens: getGoogleBudget(geminiModel, effort, options.thinkingBudgets),\n\t\t},\n\t} satisfies GoogleVertexOptions);\n};\n\nfunction createClient(\n\tmodel: Model<\"google-vertex\">,\n\tproject: string,\n\tlocation: string,\n\toptionsHeaders?: ProviderHeaders,\n\tenv?: ProviderEnv,\n): GoogleGenAI {\n\tconst googleAuthOptions = buildGoogleAuthOptions(env);\n\treturn new GoogleGenAI({\n\t\tvertexai: true,\n\t\tproject,\n\t\tlocation,\n\t\tapiVersion: API_VERSION,\n\t\t...(googleAuthOptions ? { googleAuthOptions } : {}),\n\t\thttpOptions: buildHttpOptions(model, optionsHeaders),\n\t});\n}\n\nfunction createClientWithApiKey(\n\tmodel: Model<\"google-vertex\">,\n\tapiKey: string,\n\toptionsHeaders?: ProviderHeaders,\n): GoogleGenAI {\n\treturn new GoogleGenAI({\n\t\tvertexai: true,\n\t\tapiKey,\n\t\tapiVersion: API_VERSION,\n\t\thttpOptions: buildHttpOptions(model, optionsHeaders),\n\t});\n}\n\nfunction buildHttpOptions(model: Model<\"google-vertex\">, optionsHeaders?: ProviderHeaders): HttpOptions | undefined {\n\tconst httpOptions: HttpOptions = {};\n\tconst baseUrl = resolveCustomBaseUrl(model.baseUrl);\n\tif (baseUrl) {\n\t\thttpOptions.baseUrl = baseUrl;\n\t\thttpOptions.baseUrlResourceScope = ResourceScope.COLLECTION;\n\t\tif (baseUrlIncludesApiVersion(baseUrl)) {\n\t\t\thttpOptions.apiVersion = \"\";\n\t\t}\n\t}\n\n\tconst headers = providerHeadersToRecord({ ...model.headers, ...optionsHeaders });\n\tif (headers) {\n\t\thttpOptions.headers = headers;\n\t}\n\n\treturn Object.keys(httpOptions).length > 0 ? httpOptions : undefined;\n}\n\nfunction resolveCustomBaseUrl(baseUrl: string): string | undefined {\n\tconst trimmed = baseUrl.trim();\n\tif (!trimmed || trimmed.includes(\"{location}\")) {\n\t\treturn undefined;\n\t}\n\treturn trimmed;\n}\n\nfunction baseUrlIncludesApiVersion(baseUrl: string): boolean {\n\ttry {\n\t\tconst url = new URL(baseUrl);\n\t\treturn url.pathname.split(\"/\").some((part) => /^v\\d+(?:beta\\d*)?$/.test(part));\n\t} catch {\n\t\treturn /(?:^|\\/)v\\d+(?:beta\\d*)?(?:\\/|$)/.test(baseUrl);\n\t}\n}\n\nfunction buildGoogleAuthOptions(env?: ProviderEnv): { keyFilename: string } | undefined {\n\tconst keyFilename = getProviderEnvValue(\"GOOGLE_APPLICATION_CREDENTIALS\", env);\n\treturn keyFilename ? { keyFilename } : undefined;\n}\n\nfunction resolveApiKey(options?: GoogleVertexOptions): string | undefined {\n\tconst apiKey = options?.apiKey?.trim();\n\tif (!apiKey || apiKey === GCP_VERTEX_CREDENTIALS_MARKER || isPlaceholderApiKey(apiKey)) {\n\t\treturn undefined;\n\t}\n\treturn apiKey;\n}\n\nfunction isPlaceholderApiKey(apiKey: string): boolean {\n\treturn /^<[^>]+>$/.test(apiKey);\n}\n\nfunction resolveProject(options?: GoogleVertexOptions): string {\n\tconst project =\n\t\toptions?.project ||\n\t\tgetProviderEnvValue(\"GOOGLE_CLOUD_PROJECT\", options?.env) ||\n\t\tgetProviderEnvValue(\"GCLOUD_PROJECT\", options?.env);\n\tif (!project) {\n\t\tthrow new Error(\n\t\t\t\"Vertex AI requires a project ID. Set GOOGLE_CLOUD_PROJECT/GCLOUD_PROJECT or pass project in options.\",\n\t\t);\n\t}\n\treturn project;\n}\n\nfunction resolveLocation(options?: GoogleVertexOptions): string {\n\tconst location = options?.location || getProviderEnvValue(\"GOOGLE_CLOUD_LOCATION\", options?.env);\n\tif (!location) {\n\t\tthrow new Error(\"Vertex AI requires a location. Set GOOGLE_CLOUD_LOCATION or pass location in options.\");\n\t}\n\treturn location;\n}\n\nfunction buildParams(\n\tmodel: Model<\"google-vertex\">,\n\tcontext: Context,\n\toptions: GoogleVertexOptions = {},\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 supportsStrictMode = supportsGoogleStrictToolSampling(model.id);\n\tconst functionCallingMode = context.tools?.length\n\t\t? resolveGoogleFunctionCallingMode(context.tools, options.toolChoice, supportsStrictMode)\n\t\t: undefined;\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 &&\n\t\t\tcontext.tools.length > 0 && {\n\t\t\t\ttools: convertTools(context.tools, false, supportsStrictMode),\n\t\t\t}),\n\t\t...(functionCallingMode !== undefined && {\n\t\t\ttoolConfig: { functionCallingConfig: { mode: functionCallingMode } },\n\t\t}),\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\tthinkingConfig.thinkingLevel = THINKING_LEVEL_MAP[options.thinking.level];\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<PiThinkingLevel, \"xhigh\" | \"max\">;\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\tconst id = model.id.toLowerCase();\n\treturn /gemini-3(?:\\.\\d+)?-flash/.test(id) || id === \"gemini-flash-latest\" || id === \"gemini-flash-lite-latest\";\n}\n\nfunction getDisabledThinkingConfig(model: Model<\"google-vertex\">): 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 mcpi.\n\tconst geminiModel = model as unknown as Model<\"google-generative-ai\">;\n\tif (isGemini3ProModel(geminiModel)) {\n\t\treturn { thinkingLevel: ThinkingLevel.LOW };\n\t}\n\tif (isGemini3FlashModel(geminiModel)) {\n\t\treturn { thinkingLevel: ThinkingLevel.MINIMAL };\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"]}