fix(translator): stop emitting empty <think> markers into OpenAI content

open-sse/transformer/responsesTransformer.js

@@ -333,6 +333,9 @@ export function createResponsesApiTransformStream(logger = null) {
 
           // Regular text content
           if (content) {
+            // The answer starts, so thinking is over. Upstreams that send reasoning via
+            // reasoning_content never emit "</think>", so close it here rather than at finish.
+            closeReasoning(controller);
             if (!state.msgItemAdded[idx]) {
               state.msgItemAdded[idx] = true;
               const msgId = `msg_${state.responseId}_${idx}`;
@@ -372,6 +375,7 @@ export function createResponsesApiTransformStream(logger = null) {
 
         // Handle tool_calls
         if (delta.tool_calls) {
+          closeReasoning(controller);
           closeMessage(controller, idx);
 
           for (const tc of delta.tool_calls) {

open-sse/translator/response/claude-to-openai.js

@@ -59,10 +59,6 @@ export function claudeToOpenAIResponse(chunk, state) {
       }
       if (block?.type === CLAUDE_BLOCK.TEXT) {
         state.textBlockStarted = true;
-      } else if (block?.type === CLAUDE_BLOCK.THINKING) {
-        state.inThinkingBlock = true;
-        state.currentBlockIndex = chunk.index;
-        results.push(createChunk(state, { content: "<think>" }));
       } else if (block?.type === CLAUDE_BLOCK.TOOL_USE) {
         const toolCallIndex = state.toolCallIndex++;
         // Restore original tool name from mapping (Claude OAuth)
@@ -89,6 +85,8 @@ export function claudeToOpenAIResponse(chunk, state) {
       if (delta?.type === "text_delta" && delta.text) {
         results.push(createChunk(state, { content: delta.text }));
       } else if (delta?.type === "thinking_delta" && delta.thinking) {
+        // Thinking travels only in reasoning_content. No "<think>" markers in
+        // content: OpenAI-format clients render them as literal text.
         results.push(createChunk(state, reasoningDelta(delta.thinking)));
       } else if (delta?.type === "input_json_delta" && delta.partial_json) {
         const toolCall = state.toolCalls.get(chunk.index);
@@ -112,10 +110,6 @@ export function claudeToOpenAIResponse(chunk, state) {
         state.serverToolBlockIndex = -1;
         break;
       }
-      if (state.inThinkingBlock && chunk.index === state.currentBlockIndex) {
-        results.push(createChunk(state, { content: "</think>" }));
-        state.inThinkingBlock = false;
-      }
       state.textBlockStarted = false;
       state.thinkingBlockStarted = false;
       break;

open-sse/translator/response/openai-responses.js

@@ -129,12 +129,16 @@ export function openaiToOpenAIResponsesResponse(chunk, state) {
     }
 
     if (content) {
+      // The answer starts, so thinking is over. Upstreams that send reasoning via
+      // reasoning_content never emit "</think>", so close it here rather than at finish.
+      closeReasoning(state, emit);
       emitTextContent(state, emit, idx, content);
     }
   }
 
   // Handle tool_calls (empty array is truthy; require a real call)
   if (delta.tool_calls && delta.tool_calls.length) {
+    closeReasoning(state, emit);
     closeMessage(state, emit, idx);
     for (const tc of delta.tool_calls) {
       emitToolCall(state, emit, tc);