fix(openai-responses): emit reasoning blocks from responses stream

packages/core/src/providers/openai/parseResponsesStream.reasoning.test.ts

@@ -0,0 +1,196 @@
+import { describe, it, expect } from 'vitest';
+import { parseResponsesStream } from './parseResponsesStream.js';
+
+function createSSEStream(chunks: string[]): ReadableStream<Uint8Array> {
+  const encoder = new TextEncoder();
+  let index = 0;
+
+  return new ReadableStream<Uint8Array>({
+    async pull(controller) {
+      if (index < chunks.length) {
+        const chunk = chunks[index++];
+        controller.enqueue(encoder.encode(chunk));
+      } else {
+        controller.close();
+      }
+    },
+  });
+}
+
+describe('parseResponsesStream - Reasoning/Thinking Support', () => {
+  it('should parse reasoning-only stream with delta and done events', async () => {
+    const chunks = [
+      'data: {"type":"response.reasoning_text.delta","sequence_number":1,"delta":"Let me think about this..."}\n\n',
+      'data: {"type":"response.reasoning_text.delta","sequence_number":2,"delta":" The user wants to know..."}\n\n',
+      'data: {"type":"response.reasoning_text.done","sequence_number":3}\n\n',
+    ];
+
+    const stream = createSSEStream(chunks);
+    const messages = [];
+
+    for await (const message of parseResponsesStream(stream)) {
+      messages.push(message);
+    }
+
+    // Should have one thinking block message
+    const thinkingMessage = messages.find((m) =>
+      m.blocks.some((block) => block.type === 'thinking'),
+    );
+    expect(thinkingMessage).toBeDefined();
+    expect(thinkingMessage?.speaker).toBe('ai');
+    expect(thinkingMessage?.blocks).toHaveLength(1);
+    expect(thinkingMessage?.blocks[0]).toEqual({
+      type: 'thinking',
+      thought: 'Let me think about this... The user wants to know...',
+      sourceField: 'reasoning_content',
+    });
+  });
+
+  it('should handle interleaved reasoning, text, and tool calls', async () => {
+    const chunks = [
+      // Reasoning starts
+      'data: {"type":"response.reasoning_text.delta","sequence_number":1,"delta":"I need to search for this information."}\n\n',
+      'data: {"type":"response.reasoning_text.done","sequence_number":2}\n\n',
+      // Text content
+      'data: {"type":"response.output_text.delta","delta":"Let me search for that..."}\n\n',
+      // Tool call
+      'data: {"type":"response.output_item.added","sequence_number":4,"output_index":1,"item":{"id":"fc_search","type":"function_call","status":"in_progress","arguments":"","call_id":"call_search","name":"search"}}\n\n',
+      'data: {"type":"response.function_call_arguments.delta","sequence_number":5,"item_id":"fc_search","output_index":1,"delta":"{\\"query\\":\\"test\\"}"}\n\n',
+      'data: {"type":"response.output_item.done","sequence_number":6,"output_index":1,"item":{"id":"fc_search","type":"function_call","status":"completed","arguments":"{\\"query\\":\\"test\\"}","call_id":"call_search","name":"search"}}\n\n',
+    ];
+
+    const stream = createSSEStream(chunks);
+    const messages = [];
+
+    for await (const message of parseResponsesStream(stream)) {
+      messages.push(message);
+    }
+
+    // Should have reasoning, text, and tool call messages
+    expect(
+      messages.some((m) => m.blocks.some((block) => block.type === 'thinking')),
+    ).toBe(true);
+    expect(
+      messages.some((m) =>
+        m.blocks.some(
+          (block) =>
+            block.type === 'text' &&
+            (block as { type: 'text'; text: string }).text ===
+              'Let me search for that...',
+        ),
+      ),
+    ).toBe(true);
+    expect(
+      messages.some((m) =>
+        m.blocks.some((block) => block.type === 'tool_call'),
+      ),
+    ).toBe(true);
+  });
+
+  it('should not yield thinking block for empty/whitespace-only reasoning', async () => {
+    const chunks = [
+      'data: {"type":"response.reasoning_text.delta","sequence_number":1,"delta":"   "}\n\n',
+      'data: {"type":"response.reasoning_text.delta","sequence_number":2,"delta":"\\n\\t"}\n\n',
+      'data: {"type":"response.reasoning_text.done","sequence_number":3}\n\n',
+      'data: {"type":"response.output_text.delta","delta":"Hello!"}\n\n',
+    ];
+
+    const stream = createSSEStream(chunks);
+    const messages = [];
+
+    for await (const message of parseResponsesStream(stream)) {
+      messages.push(message);
+    }
+
+    // Should not have thinking block, only text
+    expect(
+      messages.some((m) => m.blocks.some((block) => block.type === 'thinking')),
+    ).toBe(false);
+    expect(
+      messages.some((m) => m.blocks.some((block) => block.type === 'text')),
+    ).toBe(true);
+  });
+
+  it('should accumulate multiple reasoning deltas into single block', async () => {
+    const chunks = [
+      'data: {"type":"response.reasoning_text.delta","sequence_number":1,"delta":"First chunk."}\n\n',
+      'data: {"type":"response.reasoning_text.delta","sequence_number":2,"delta":" Second chunk."}\n\n',
+      'data: {"type":"response.reasoning_text.delta","sequence_number":3,"delta":" Third chunk."}\n\n',
+      'data: {"type":"response.reasoning_text.done","sequence_number":4}\n\n',
+    ];
+
+    const stream = createSSEStream(chunks);
+    const messages = [];
+
+    for await (const message of parseResponsesStream(stream)) {
+      messages.push(message);
+    }
+
+    const thinkingMessages = messages.filter((m) =>
+      m.blocks.some((block) => block.type === 'thinking'),
+    );
+    // Should have exactly one thinking message with all deltas accumulated
+    expect(thinkingMessages).toHaveLength(1);
+    const thinkingBlock = thinkingMessages[0]?.blocks[0] as {
+      type: 'thinking';
+      thought: string;
+    };
+    expect(thinkingBlock.thought).toBe(
+      'First chunk. Second chunk. Third chunk.',
+    );
+  });
+
+  it('should handle reasoning with usage metadata', async () => {
+    const chunks = [
+      'data: {"type":"response.reasoning_text.delta","sequence_number":1,"delta":"Thinking deeply..."}\n\n',
+      'data: {"type":"response.reasoning_text.done","sequence_number":2}\n\n',
+      'data: {"type":"response.output_text.delta","delta":"Here is my answer."}\n\n',
+      'data: {"type":"response.completed","sequence_number":4,"response":{"id":"resp_123","object":"response","model":"gpt-5.2","status":"completed","usage":{"input_tokens":100,"output_tokens":50,"total_tokens":150}}}\n\n',
+    ];
+
+    const stream = createSSEStream(chunks);
+    const messages = [];
+
+    for await (const message of parseResponsesStream(stream)) {
+      messages.push(message);
+    }
+
+    // Should have thinking block, text, and usage
+    expect(
+      messages.some((m) => m.blocks.some((block) => block.type === 'thinking')),
+    ).toBe(true);
+    expect(
+      messages.some((m) => m.blocks.some((block) => block.type === 'text')),
+    ).toBe(true);
+    const usageMessage = messages.find((m) => m.metadata?.usage);
+    expect(usageMessage?.metadata?.usage).toEqual({
+      promptTokens: 100,
+      completionTokens: 50,
+      totalTokens: 150,
+      cachedTokens: 0,
+    });
+  });
+
+  it('should yield reasoning before response.completed', async () => {
+    const chunks = [
+      'data: {"type":"response.reasoning_text.delta","sequence_number":1,"delta":"Reasoning content"}\n\n',
+      'data: {"type":"response.completed","sequence_number":2,"response":{"id":"resp_123","object":"response","model":"gpt-5.2","status":"completed","usage":{"input_tokens":10,"output_tokens":5,"total_tokens":15}}}\n\n',
+    ];
+
+    const stream = createSSEStream(chunks);
+    const messages = [];
+
+    for await (const message of parseResponsesStream(stream)) {
+      messages.push(message);
+    }
+
+    // Thinking should come before usage
+    const thinkingIndex = messages.findIndex((m) =>
+      m.blocks.some((block) => block.type === 'thinking'),
+    );
+    const usageIndex = messages.findIndex((m) => m.metadata?.usage);
+    expect(thinkingIndex).toBeGreaterThanOrEqual(0);
+    expect(usageIndex).toBeGreaterThanOrEqual(0);
+    expect(thinkingIndex).toBeLessThan(usageIndex);
+  });
+});

packages/core/src/providers/openai/parseResponsesStream.ts

@@ -1,4 +1,7 @@
 /**
+ * Parses OpenAI Responses API server-sent events (SSE) and yields IContent messages.
+ * Handles text output, tool calls, reasoning/thinking content, and usage metadata.
+ *
  * @plan PLAN-20250120-DEBUGLOGGING.P15
  * @requirement REQ-INT-001.1
  */
@@ -52,6 +55,7 @@ export async function* parseResponsesStream(
   const decoder = new TextDecoder();
   let buffer = '';
   const functionCalls = new Map<string, FunctionCallState>();
+  let reasoningText = '';
 
   try {
     while (true) {
@@ -88,6 +92,30 @@ export async function* parseResponsesStream(
                 }
                 break;
 
+              case 'response.reasoning_text.delta':
+                // Reasoning content chunk
+                if (event.delta) {
+                  reasoningText += event.delta;
+                }
+                break;
+
+              case 'response.reasoning_text.done':
+                // Reasoning completed - yield accumulated reasoning
+                if (reasoningText.trim()) {
+                  yield {
+                    speaker: 'ai',
+                    blocks: [
+                      {
+                        type: 'thinking',
+                        thought: reasoningText,
+                        sourceField: 'reasoning_content',
+                      },
+                    ],
+                  };
+                }
+                reasoningText = '';
+                break;
+
               case 'response.output_item.added':
                 // New function call started
                 if (event.item?.type === 'function_call' && event.item.id) {
@@ -158,6 +186,21 @@ export async function* parseResponsesStream(
 
               case 'response.completed':
               case 'response.done':
+                // Yield any remaining reasoning before usage data
+                if (reasoningText.trim()) {
+                  yield {
+                    speaker: 'ai',
+                    blocks: [
+                      {
+                        type: 'thinking',
+                        thought: reasoningText,
+                        sourceField: 'reasoning_content',
+                      },
+                    ],
+                  };
+                  reasoningText = '';
+                }
+
                 // Usage data - handle both response.completed (OpenAI) and response.done (Codex)
                 if (event.response?.usage) {
                   yield {