feat(ai): add strictResponsesPairing for Azure OpenAI Responses API

Split OpenAICompat into OpenAICompletionsCompat and OpenAIResponsesCompat
for type-safe API-specific compat settings. Added strictResponsesPairing
option to suppress orphaned reasoning/tool calls on incomplete turns,
fixing 400 errors on Azure's Responses API which requires strict pairing.

Closes #768

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

@@ -15,7 +15,7 @@ import type {
 	Context,
 	Message,
 	Model,
-	OpenAICompat,
+	OpenAICompletionsCompat,
 	StopReason,
 	StreamFunction,
 	StreamOptions,
@@ -452,7 +452,7 @@ function maybeAddOpenRouterAnthropicCacheControl(
 function convertMessages(
 	model: Model<"openai-completions">,
 	context: Context,
-	compat: Required<OpenAICompat>,
+	compat: Required<OpenAICompletionsCompat>,
 ): ChatCompletionMessageParam[] {
 	const params: ChatCompletionMessageParam[] = [];
 
@@ -681,9 +681,9 @@ function mapStopReason(reason: ChatCompletionChunk.Choice["finish_reason"]): Sto
 /**
  * Detect compatibility settings from provider and baseUrl for known providers.
  * Provider takes precedence over URL-based detection since it's explicitly configured.
- * Returns a fully resolved OpenAICompat object with all fields set.
+ * Returns a fully resolved OpenAICompletionsCompat object with all fields set.
  */
-function detectCompat(model: Model<"openai-completions">): Required<OpenAICompat> {
+function detectCompat(model: Model<"openai-completions">): Required<OpenAICompletionsCompat> {
 	const provider = model.provider;
 	const baseUrl = model.baseUrl;
 
@@ -725,7 +725,7 @@ function detectCompat(model: Model<"openai-completions">): Required<OpenAICompat
  * Get resolved compatibility settings for a model.
  * Uses explicit model.compat if provided, otherwise auto-detects from provider/URL.
  */
-function getCompat(model: Model<"openai-completions">): Required<OpenAICompat> {
+function getCompat(model: Model<"openai-completions">): Required<OpenAICompletionsCompat> {
 	const detected = detectCompat(model);
 	if (!model.compat) return detected;
 

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

@@ -461,10 +461,22 @@ function convertMessages(model: Model<"openai-responses">, context: Context): Re
 			}
 		} else if (msg.role === "assistant") {
 			const output: ResponseInput = [];
+			const strictResponsesPairing = model.compat?.strictResponsesPairing ?? false;
+			let isIncomplete = false;
+			let shouldReplayReasoning = msg.stopReason !== "error";
+			let allowToolCalls = msg.stopReason !== "error";
+			if (strictResponsesPairing) {
+				isIncomplete = msg.stopReason === "error" || msg.stopReason === "aborted";
+				const hasPairedContent = msg.content.some(
+					(b) => b.type === "toolCall" || (b.type === "text" && (b as TextContent).text.trim().length > 0),
+				);
+				shouldReplayReasoning = !isIncomplete && hasPairedContent;
+				allowToolCalls = !isIncomplete;
+			}
 
 			for (const block of msg.content) {
 				// Do not submit thinking blocks if the completion had an error (i.e. abort)
-				if (block.type === "thinking" && msg.stopReason !== "error") {
+				if (block.type === "thinking" && shouldReplayReasoning) {
 					if (block.thinkingSignature) {
 						const reasoningItem = JSON.parse(block.thinkingSignature);
 						output.push(reasoningItem);
@@ -475,6 +487,11 @@ function convertMessages(model: Model<"openai-responses">, context: Context): Re
 					let msgId = textBlock.textSignature;
 					if (!msgId) {
 						msgId = `msg_${msgIndex}`;
+					}
+					// For incomplete turns, never replay the original message id (if any).
+					// Generate a stable synthetic id so strict pairing providers do not expect a paired reasoning item.
+					if (strictResponsesPairing && isIncomplete) {
+						msgId = `msg_${msgIndex}_${shortHash(textBlock.text)}`;
 					} else if (msgId.length > 64) {
 						msgId = `msg_${shortHash(msgId)}`;
 					}
@@ -486,7 +503,7 @@ function convertMessages(model: Model<"openai-responses">, context: Context): Re
 						id: msgId,
 					} satisfies ResponseOutputMessage);
 					// Do not submit toolcall blocks if the completion had an error (i.e. abort)
-				} else if (block.type === "toolCall" && msg.stopReason !== "error") {
+				} else if (block.type === "toolCall" && allowToolCalls) {
 					const toolCall = block as ToolCall;
 					output.push({
 						type: "function_call",

packages/ai/src/types.ts

@@ -204,10 +204,10 @@ export type AssistantMessageEvent =
 	| { type: "error"; reason: Extract<StopReason, "aborted" | "error">; error: AssistantMessage };
 
 /**
- * Compatibility settings for openai-completions API.
+ * Compatibility settings for OpenAI-compatible completions APIs.
  * Use this to override URL-based auto-detection for custom providers.
  */
-export interface OpenAICompat {
+export interface OpenAICompletionsCompat {
 	/** Whether the provider supports the `store` field. Default: auto-detected from URL. */
 	supportsStore?: boolean;
 	/** Whether the provider supports the `developer` role (vs `system`). Default: auto-detected from URL. */
@@ -230,6 +230,12 @@ export interface OpenAICompat {
 	thinkingFormat?: "openai" | "zai";
 }
 
+/** Compatibility settings for OpenAI Responses APIs. */
+export interface OpenAIResponsesCompat {
+	/** Whether OpenAI Responses history replay requires strict reasoning/message pairing (for providers like Azure). */
+	strictResponsesPairing?: boolean;
+}
+
 // Model interface for the unified model system
 export interface Model<TApi extends Api> {
 	id: string;
@@ -248,6 +254,10 @@ export interface Model<TApi extends Api> {
 	contextWindow: number;
 	maxTokens: number;
 	headers?: Record<string, string>;
-	/** Compatibility overrides for openai-completions API. If not set, auto-detected from baseUrl. */
-	compat?: TApi extends "openai-completions" ? OpenAICompat : never;
+	/** Compatibility overrides for OpenAI-compatible APIs. If not set, auto-detected from baseUrl. */
+	compat?: TApi extends "openai-completions"
+		? OpenAICompletionsCompat
+		: TApi extends "openai-responses"
+			? OpenAIResponsesCompat
+			: never;
 }