Strip reasoning parts from message history (OpenAI fix)

src/services/aiService.ts

@@ -19,6 +19,7 @@ import {
   validateAnthropicCompliance,
   addInterruptedSentinel,
   filterEmptyAssistantMessages,
+  stripReasoningForOpenAI,
 } from "@/utils/messages/modelMessageTransform";
 import { applyCacheControl } from "@/utils/ai/cacheStrategy";
 import type { HistoryService } from "./historyService";
@@ -278,11 +279,23 @@ export class AIService extends EventEmitter {
       // Dump original messages for debugging
       log.debug_obj(`${workspaceId}/1_original_messages.json`, messages);
 
+      // Extract provider name from modelString (e.g., "anthropic:claude-opus-4-1" -> "anthropic")
+      const [providerName] = modelString.split(":");
+
       // Filter out assistant messages with only reasoning (no text/tools)
-      const filteredMessages = filterEmptyAssistantMessages(messages);
+      let filteredMessages = filterEmptyAssistantMessages(messages);
       log.debug(`Filtered ${messages.length - filteredMessages.length} empty assistant messages`);
       log.debug_obj(`${workspaceId}/1a_filtered_messages.json`, filteredMessages);
 
+      // OpenAI-specific: Strip reasoning parts from history
+      // OpenAI manages reasoning via previousResponseId; sending Anthropic-style reasoning
+      // parts creates orphaned reasoning items that cause API errors
+      if (providerName === "openai") {
+        filteredMessages = stripReasoningForOpenAI(filteredMessages);
+        log.debug("Stripped reasoning parts for OpenAI");
+        log.debug_obj(`${workspaceId}/1b_openai_stripped.json`, filteredMessages);
+      }
+
       // Add [INTERRUPTED] sentinel to partial messages (for model context)
       const messagesWithSentinel = addInterruptedSentinel(filteredMessages);
 
@@ -293,9 +306,6 @@ export class AIService extends EventEmitter {
 
       log.debug_obj(`${workspaceId}/2_model_messages.json`, modelMessages);
 
-      // Extract provider name from modelString (e.g., "anthropic:claude-opus-4-1" -> "anthropic")
-      const [providerName] = modelString.split(":");
-
       // Apply ModelMessage transforms based on provider requirements
       const transformedMessages = transformModelMessages(modelMessages, providerName);
 

src/utils/messages/modelMessageTransform.ts

@@ -10,6 +10,10 @@ import type { CmuxMessage } from "@/types/message";
  * Filter out assistant messages that only contain reasoning parts (no text or tool parts).
  * These messages are invalid for the API and provide no value to the model.
  * This happens when a message is interrupted during thinking before producing any text.
+ *
+ * Note: This function filters out reasoning-only messages but does NOT strip reasoning
+ * parts from messages that have other content. Reasoning parts are handled differently
+ * per provider (see stripReasoningForOpenAI).
  */
 export function filterEmptyAssistantMessages(messages: CmuxMessage[]): CmuxMessage[] {
   return messages.filter((msg) => {
@@ -27,6 +31,36 @@ export function filterEmptyAssistantMessages(messages: CmuxMessage[]): CmuxMessa
   });
 }
 
+/**
+ * Strip reasoning parts from messages for OpenAI.
+ *
+ * OpenAI's Responses API uses encrypted reasoning items (with IDs like rs_*) that are
+ * managed automatically via previous_response_id. When reasoning parts from history
+ * (which are Anthropic-style text-based reasoning) are sent to OpenAI, they create
+ * orphaned reasoning items that cause "reasoning without following item" errors.
+ *
+ * Anthropic's reasoning (text-based) is different and SHOULD be sent back via sendReasoning.
+ *
+ * @param messages - Messages that may contain reasoning parts
+ * @returns Messages with reasoning parts stripped (for OpenAI only)
+ */
+export function stripReasoningForOpenAI(messages: CmuxMessage[]): CmuxMessage[] {
+  return messages.map((msg) => {
+    // Only process assistant messages
+    if (msg.role !== "assistant") {
+      return msg;
+    }
+
+    // Strip reasoning parts - OpenAI manages reasoning via previousResponseId
+    const filteredParts = msg.parts.filter((part) => part.type !== "reasoning");
+
+    return {
+      ...msg,
+      parts: filteredParts,
+    };
+  });
+}
+
 /**
  * Add [INTERRUPTED] sentinel to partial messages by inserting a user message.
  * This helps the model understand that a message was interrupted and incomplete.