Show reasoner model's thinking in a <<< thinking section

py/chat.py

@@ -30,7 +30,7 @@ def initialize_chat_window():
         vim.command("redraw")
 
         file_content = vim.eval('trim(join(getline(1, "$"), "\n"))')
-        role_lines = re.findall(r'(^>>> user|^>>> system|^<<< assistant).*', file_content, flags=re.MULTILINE)
+        role_lines = re.findall(r'(^>>> user|^>>> system|^<<< thinking|^<<< assistant).*', file_content, flags=re.MULTILINE)
         if not role_lines[-1].startswith(">>> user"):
             # last role is not user, most likely completion was cancelled before
             vim.command("normal! o")
@@ -58,14 +58,27 @@ def initialize_chat_window():
     try:
         last_content = messages[-1]["content"][-1]
         if last_content['type'] != 'text' or last_content['text']:
-            vim.command("normal! Go\n<<< assistant\n\n")
             vim.command("redraw")
 
             print('Answering...')
-            vim.command("redraw")
 
-            text_chunks = make_chat_text_chunks(messages, options)
-            render_text_chunks(text_chunks)
+            def _chunks_to_sections(chunks):
+                first_thinking_chunk = True
+                first_content_chunk = True
+                for chunk in chunks:
+                    if chunk['thinking'] is not None:
+                        if first_thinking_chunk:
+                            first_thinking_chunk = False
+                            vim.command("normal! Go\n<<< thinking\n\n")
+                        yield chunk['thinking']
+                    if chunk['content'] is not None:
+                        if first_content_chunk:
+                            first_content_chunk = False
+                            vim.command("normal! Go\n<<< assistant\n\n")
+                        yield chunk['content']
+
+            chunks = make_chat_text_chunks(messages, options)
+            render_text_chunks(_chunks_to_sections(chunks))
 
             vim.command("normal! a\n\n>>> user\n\n")
             vim.command("redraw")

py/complete.py

@@ -34,7 +34,10 @@ def chat_engine(prompt):
         chat_content = f"{initial_prompt}\n\n>>> user\n\n{prompt}".strip()
         messages = parse_chat_messages(chat_content)
         print_debug("[engine-chat] text:\n" + chat_content)
-        return make_chat_text_chunks(messages, config_options)
+        return filter(
+            lambda c: c,
+            map(lambda c: c.get("content"), make_chat_text_chunks(messages, config_options)),
+        )
 
     engines = {"chat": chat_engine, "complete": complete_engine}
 

py/utils.py

@@ -158,6 +158,8 @@ def parse_chat_messages(chat_content):
             case '>>> system':
                 messages.append({'role': 'system', 'content': [{ 'type': 'text', 'text': '' }]})
                 current_type = 'system'
+            case '<<< thinking':
+                current_type = 'thinking'
             case '<<< assistant':
                 messages.append({'role': 'assistant', 'content': [{ 'type': 'text', 'text': '' }]})
                 current_type = 'assistant'
@@ -315,34 +317,33 @@ def make_chat_text_chunks(messages, config_options):
     openai_options = make_openai_options(config_options)
     http_options = make_http_options(config_options)
 
+    def _flatten_content(messages):
+        """Some providers like api.deepseek.com & api.groq.com expect a flat 'content' field."""
+        for message in messages:
+            match message['role']:
+                case 'system' | 'assistant':
+                    message['content'] = '\n'.join(map(lambda c: c['text'], message['content']))
+        return messages
+
     request = {
-        'messages': messages,
+        'messages': _flatten_content(messages),
         **openai_options
     }
     print_debug("[engine-chat] request: {}", request)
     url = config_options['endpoint_url']
     response = openai_request(url, request, http_options)
+    _choice_key = 'delta' if openai_options['stream'] else 'message'
 
-    def _choices(resp):
-        choices = resp.get('choices', [{}])
-
-        # NOTE choices may exist in the response, but be an empty list.
-        if not choices:
-            return [{}]
-
-        return choices
+    def _get_delta(resp):
+        choices = resp.get('choices') or [{}]
+        return choices[0].get(_choice_key, {})
 
-    def map_chunk_no_stream(resp):
+    def _map_chunk(resp):
         print_debug("[engine-chat] response: {}", resp)
-        return _choices(resp)[0].get('message', {}).get('content', '')
-
-    def map_chunk_stream(resp):
-        print_debug("[engine-chat] response: {}", resp)
-        return _choices(resp)[0].get('delta', {}).get('content', '')
-
-    map_chunk = map_chunk_stream if openai_options['stream'] else map_chunk_no_stream
+        delta = _get_delta(resp)
+        return {'thinking': delta.get('reasoning_content'), 'content': delta.get('content')}
 
-    return map(map_chunk, response)
+    return map(_map_chunk, response)
 
 def read_role_files():
     plugin_root = vim.eval("s:plugin_root")

syntax/aichat.vim

@@ -1,6 +1,7 @@
 syntax match aichatRole ">>> system"
 syntax match aichatRole ">>> user"
 syntax match aichatRole ">>> include"
+syntax match aichatRole "<<< thinking"
 syntax match aichatRole "<<< assistant"
 
 highlight default link aichatRole Comment