add option to skip provider tool call validation

packages/agent/src/agent.ts

@@ -221,6 +221,7 @@ export class Agent {
 			tools: this._state.tools,
 			model,
 			reasoning,
+			validateToolCallsAtProvider: false,
 			getQueuedMessages: async <T>() => {
 				// Return queued messages based on queue mode
 				if (this.queueMode === "one-at-a-time") {

packages/agent/src/transports/AppTransport.ts

@@ -77,6 +77,7 @@ function streamSimpleProxy(
 						temperature: options.temperature,
 						maxTokens: options.maxTokens,
 						reasoning: options.reasoning,
+						validateToolCallsAtProvider: options.validateToolCallsAtProvider,
 						// Don't send apiKey or signal - those are added server-side
 					},
 				}),
@@ -365,6 +366,7 @@ export class AppTransport implements AgentTransport {
 			model: cfg.model,
 			reasoning: cfg.reasoning,
 			getQueuedMessages: cfg.getQueuedMessages,
+			validateToolCallsAtProvider: cfg.validateToolCallsAtProvider ?? false,
 		};
 
 		// Yield events from the upstream agentLoop iterator

packages/agent/src/transports/ProviderTransport.ts

@@ -65,6 +65,7 @@ export class ProviderTransport implements AgentTransport {
 			reasoning: cfg.reasoning,
 			apiKey,
 			getQueuedMessages: cfg.getQueuedMessages,
+			validateToolCallsAtProvider: cfg.validateToolCallsAtProvider ?? false,
 		};
 
 		// Yield events from agentLoop

packages/agent/src/transports/types.ts

@@ -9,6 +9,7 @@ export interface AgentRunConfig {
 	model: Model<any>;
 	reasoning?: "low" | "medium" | "high";
 	getQueuedMessages?: <T>() => Promise<QueuedMessage<T>[]>;
+	validateToolCallsAtProvider?: boolean;
 }
 
 /**

packages/ai/CHANGELOG.md

@@ -2,6 +2,10 @@
 
 ## [Unreleased]
 
+### Added
+
+- Added `validateToolCallsAtProvider` option to streaming and agent APIs to optionally skip provider-level tool-call validation (default on), allowing agent loops to surface schema errors as toolResult messages and retry.
+
 ## [0.13.0] - 2025-12-06
 
 ### Breaking Changes

packages/ai/README.md

@@ -270,6 +270,20 @@ for await (const event of s) {
 - Full validation only occurs at `toolcall_end` when arguments are complete
 - The Google provider does not support function call streaming. Instead, you will receive a single `toolcall_delta` event with the full arguments.
 
+### Provider tool-call validation
+
+By default, providers validate streamed tool calls against your tool schema and abort the stream on validation errors. Set `validateToolCallsAtProvider: false` on `stream`, `streamSimple`, `complete`, `completeSimple`, or `AgentLoopConfig` to skip provider-level validation and let downstream code (for example, `agentLoop` via `executeToolCalls` → `validateToolArguments`) surface schema errors as `toolResult` messages. This enables the model to retry after receiving a validation error.
+
+```typescript
+await streamSimple(model, context, {
+  apiKey: 'your-key',
+  validateToolCallsAtProvider: false
+});
+```
+
+- `true` (default): Provider validates tool calls and emits an error if arguments do not match the schema
+- `false`: Provider emits tool calls even when arguments are invalid; callers must validate and handle errors themselves
+
 ### Complete Event Reference
 
 All streaming events emitted during assistant message generation:

packages/ai/src/providers/anthropic.ts

@@ -92,6 +92,7 @@ export const streamAnthropic: StreamFunction<"anthropic-messages"> = (
 	options?: AnthropicOptions,
 ): AssistantMessageEventStream => {
 	const stream = new AssistantMessageEventStream();
+	const shouldValidateToolCalls = options?.validateToolCallsAtProvider !== false;
 
 	(async () => {
 		const output: AssistantMessage = {
@@ -233,7 +234,7 @@ export const streamAnthropic: StreamFunction<"anthropic-messages"> = (
 							block.arguments = parseStreamingJson(block.partialJson);
 
 							// Validate tool arguments if tool definition is available
-							if (context.tools) {
+							if (shouldValidateToolCalls && context.tools) {
 								const tool = context.tools.find((t) => t.name === block.name);
 								if (tool) {
 									block.arguments = validateToolArguments(tool, block);

packages/ai/src/providers/google.ts

@@ -43,6 +43,7 @@ export const streamGoogle: StreamFunction<"google-generative-ai"> = (
 	options?: GoogleOptions,
 ): AssistantMessageEventStream => {
 	const stream = new AssistantMessageEventStream();
+	const shouldValidateToolCalls = options?.validateToolCallsAtProvider !== false;
 
 	(async () => {
 		const output: AssistantMessage = {
@@ -167,7 +168,7 @@ export const streamGoogle: StreamFunction<"google-generative-ai"> = (
 							};
 
 							// Validate tool arguments if tool definition is available
-							if (context.tools) {
+							if (shouldValidateToolCalls && context.tools) {
 								const tool = context.tools.find((t) => t.name === toolCall.name);
 								if (tool) {
 									toolCall.arguments = validateToolArguments(tool, toolCall);

packages/ai/src/providers/openai-completions.ts

@@ -37,6 +37,7 @@ export const streamOpenAICompletions: StreamFunction<"openai-completions"> = (
 	options?: OpenAICompletionsOptions,
 ): AssistantMessageEventStream => {
 	const stream = new AssistantMessageEventStream();
+	const shouldValidateToolCalls = options?.validateToolCallsAtProvider !== false;
 
 	(async () => {
 		const output: AssistantMessage = {
@@ -86,7 +87,7 @@ export const streamOpenAICompletions: StreamFunction<"openai-completions"> = (
 						block.arguments = JSON.parse(block.partialArgs || "{}");
 
 						// Validate tool arguments if tool definition is available
-						if (context.tools) {
+						if (shouldValidateToolCalls && context.tools) {
 							const tool = context.tools.find((t) => t.name === block.name);
 							if (tool) {
 								block.arguments = validateToolArguments(tool, block);

packages/ai/src/providers/openai-responses.ts

@@ -45,6 +45,7 @@ export const streamOpenAIResponses: StreamFunction<"openai-responses"> = (
 	options?: OpenAIResponsesOptions,
 ): AssistantMessageEventStream => {
 	const stream = new AssistantMessageEventStream();
+	const shouldValidateToolCalls = options?.validateToolCallsAtProvider !== false;
 
 	// Start async processing
 	(async () => {
@@ -240,7 +241,7 @@ export const streamOpenAIResponses: StreamFunction<"openai-responses"> = (
 						};
 
 						// Validate tool arguments if tool definition is available
-						if (context.tools) {
+						if (shouldValidateToolCalls && context.tools) {
 							const tool = context.tools.find((t) => t.name === toolCall.name);
 							if (tool) {
 								toolCall.arguments = validateToolArguments(tool, toolCall);

packages/ai/src/stream.ts

@@ -120,6 +120,7 @@ function mapOptionsForApi<TApi extends Api>(
 		maxTokens: options?.maxTokens || Math.min(model.maxTokens, 32000),
 		signal: options?.signal,
 		apiKey: apiKey || options?.apiKey,
+		validateToolCallsAtProvider: options?.validateToolCallsAtProvider ?? true,
 	};
 
 	switch (model.api) {

packages/ai/src/types.ts

@@ -37,6 +37,12 @@ export interface StreamOptions {
 	maxTokens?: number;
 	signal?: AbortSignal;
 	apiKey?: string;
+	/**
+	 * Controls whether providers validate streamed tool calls against the tool schema.
+	 * Defaults to true. Set to false to skip provider-level validation and allow
+	 * downstream consumers (e.g., agentLoop) to handle validation failures.
+	 */
+	validateToolCallsAtProvider?: boolean;
 }
 
 // Unified options with reasoning passed to streamSimple() and completeSimple()