fix(ai): skip errored/aborted assistant messages in transform-messages

Fixes OpenAI Responses 400 error 'reasoning without following item' by
skipping errored/aborted assistant messages entirely rather than filtering
at the provider level. This covers openai-responses, openai-codex-responses,
and future providers.

Removes strictResponsesPairing compat option (no longer needed).

Closes #838

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

@@ -329,7 +329,7 @@ function convertAssistantMessage(msg: AssistantMessage): unknown[] {
 	const output: unknown[] = [];
 
 	for (const block of msg.content) {
-		if (block.type === "thinking" && msg.stopReason !== "error" && block.thinkingSignature) {
+		if (block.type === "thinking" && block.thinkingSignature) {
 			output.push(JSON.parse(block.thinkingSignature));
 		} else if (block.type === "text") {
 			output.push({
@@ -338,7 +338,7 @@ function convertAssistantMessage(msg: AssistantMessage): unknown[] {
 				content: [{ type: "output_text", text: sanitizeSurrogates(block.text), annotations: [] }],
 				status: "completed",
 			});
-		} else if (block.type === "toolCall" && msg.stopReason !== "error") {
+		} else if (block.type === "toolCall") {
 			const [callId, id] = block.id.split("|");
 			output.push({
 				type: "function_call",

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

@@ -478,22 +478,9 @@ 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" && shouldReplayReasoning) {
+				if (block.type === "thinking") {
 					if (block.thinkingSignature) {
 						const reasoningItem = JSON.parse(block.thinkingSignature);
 						output.push(reasoningItem);
@@ -504,11 +491,6 @@ 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)}`;
 					}
@@ -519,8 +501,7 @@ function convertMessages(model: Model<"openai-responses">, context: Context): Re
 						status: "completed",
 						id: msgId,
 					} satisfies ResponseOutputMessage);
-					// Do not submit toolcall blocks if the completion had an error (i.e. abort)
-				} else if (block.type === "toolCall" && allowToolCalls) {
+				} else if (block.type === "toolCall") {
 					const toolCall = block as ToolCall;
 					output.push({
 						type: "function_call",

packages/ai/src/providers/transform-messages.ts

@@ -118,27 +118,23 @@ export function transformMessages<TApi extends Api>(
 				existingToolResultIds = new Set();
 			}
 
-			// Track tool calls from this assistant message
-			// Don't track tool calls from errored messages - they will be dropped by
-			// provider-specific converters, so we shouldn't create synthetic results for them
+			// Skip errored/aborted assistant messages entirely.
+			// These are incomplete turns that shouldn't be replayed:
+			// - May have partial content (reasoning without message, incomplete tool calls)
+			// - Replaying them can cause API errors (e.g., OpenAI "reasoning without following item")
+			// - The model should retry from the last valid state
 			const assistantMsg = msg as AssistantMessage;
-			const toolCalls =
-				assistantMsg.stopReason === "error"
-					? []
-					: (assistantMsg.content.filter((b) => b.type === "toolCall") as ToolCall[]);
+			if (assistantMsg.stopReason === "error" || assistantMsg.stopReason === "aborted") {
+				continue;
+			}
+
+			// Track tool calls from this assistant message
+			const toolCalls = assistantMsg.content.filter((b) => b.type === "toolCall") as ToolCall[];
 			if (toolCalls.length > 0) {
 				pendingToolCalls = toolCalls;
 				existingToolResultIds = new Set();
 			}
 
-			// Skip empty assistant messages (no content and no tool calls)
-			// This handles error responses (e.g., 429/500) that produced no content
-			// All providers already filter these in convertMessages, but we do it here
-			// centrally to prevent issues with the tool_use -> tool_result chain
-			if (assistantMsg.content.length === 0 && toolCalls.length === 0) {
-				continue;
-			}
-
 			result.push(msg);
 		} else if (msg.role === "toolResult") {
 			existingToolResultIds.add(msg.toolCallId);

packages/ai/src/types.ts

@@ -236,8 +236,7 @@ export interface OpenAICompletionsCompat {
 
 /** 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;
+	// Reserved for future use
 }
 
 // Model interface for the unified model system