fix: separate reasoning blocks with newline on markdown boundaries (#569)

src/core/display-pipeline.test.ts

@@ -109,6 +109,68 @@ describe('createDisplayPipeline', () => {
     }
   });
 
+  it('inserts newline before reasoning chunk starting with bold header', async () => {
+    const events = await collect([
+      { type: 'reasoning', content: 'First section ends here.', runId: 'run-1' },
+      { type: 'reasoning', content: '**Second Section**\nMore text.', runId: 'run-1' },
+      { type: 'assistant', content: 'reply', runId: 'run-1' },
+      { type: 'result', success: true, result: 'reply', runIds: ['run-1'] },
+    ]);
+
+    const reasoning = events.find(e => e.type === 'reasoning');
+    expect(reasoning).toBeDefined();
+    if (reasoning?.type === 'reasoning') {
+      expect(reasoning.content).toBe('First section ends here.\n**Second Section**\nMore text.');
+    }
+  });
+
+  it('inserts newline before reasoning chunk starting with markdown heading', async () => {
+    const events = await collect([
+      { type: 'reasoning', content: 'End of thought.', runId: 'run-1' },
+      { type: 'reasoning', content: '## Next Topic\nDetails here.', runId: 'run-1' },
+      { type: 'assistant', content: 'reply', runId: 'run-1' },
+      { type: 'result', success: true, result: 'reply', runIds: ['run-1'] },
+    ]);
+
+    const reasoning = events.find(e => e.type === 'reasoning');
+    expect(reasoning).toBeDefined();
+    if (reasoning?.type === 'reasoning') {
+      expect(reasoning.content).toBe('End of thought.\n## Next Topic\nDetails here.');
+    }
+  });
+
+  it('does not insert separator for token-level streaming chunks', async () => {
+    const events = await collect([
+      { type: 'reasoning', content: "I'm", runId: 'run-1' },
+      { type: 'reasoning', content: ' thinking', runId: 'run-1' },
+      { type: 'reasoning', content: ' about this', runId: 'run-1' },
+      { type: 'assistant', content: 'reply', runId: 'run-1' },
+      { type: 'result', success: true, result: 'reply', runIds: ['run-1'] },
+    ]);
+
+    const reasoning = events.find(e => e.type === 'reasoning');
+    expect(reasoning).toBeDefined();
+    if (reasoning?.type === 'reasoning') {
+      expect(reasoning.content).toBe("I'm thinking about this");
+    }
+  });
+
+  it('skips separator when buffer already ends with newline', async () => {
+    const events = await collect([
+      { type: 'reasoning', content: 'First block.\n', runId: 'run-1' },
+      { type: 'reasoning', content: '**Second block**', runId: 'run-1' },
+      { type: 'assistant', content: 'reply', runId: 'run-1' },
+      { type: 'result', success: true, result: 'reply', runIds: ['run-1'] },
+    ]);
+
+    const reasoning = events.find(e => e.type === 'reasoning');
+    expect(reasoning).toBeDefined();
+    if (reasoning?.type === 'reasoning') {
+      // No double newline -- buffer already ended with \n
+      expect(reasoning.content).toBe('First block.\n**Second block**');
+    }
+  });
+
   it('prefers streamed text over result field on divergence', async () => {
     const events = await collect([
       { type: 'assistant', content: 'streamed reply', runId: 'run-1' },

src/core/display-pipeline.ts

@@ -237,7 +237,17 @@ export async function* createDisplayPipeline(
     // ── Dispatch by type ──
     switch (msg.type) {
       case 'reasoning': {
-        reasoningBuffer += msg.content || '';
+        const chunk = msg.content || '';
+        // When a new chunk starts with a markdown block indicator (bold header,
+        // heading, list item), insert a newline to prevent it running into the
+        // previous text. This separates complete reasoning blocks (common with
+        // OpenAI models that emit whole sections) without affecting token-level
+        // streaming where tokens don't start with these patterns.
+        if (chunk && reasoningBuffer && !reasoningBuffer.endsWith('\n')
+          && /^(\*\*|#{1,6}\s|[-*]\s|\d+\.\s)/.test(chunk)) {
+          reasoningBuffer += '\n';
+        }
+        reasoningBuffer += chunk;
         break;
       }