fix: handle reasoning_content for Kimi/thinking models (#7252)

Signed-off-by: clayarnoldg2m <carnold@g2m.ai>
Co-authored-by: Claude Opus 4.6 <noreply@anthropic.com>

crates/goose/src/agents/agent.rs

@@ -1346,8 +1346,9 @@ impl Agent {
                                     }
                                 }
 
-                                // Preserve thinking content from the original response
+                                // Preserve thinking/reasoning content from the original response
                                 // Gemini (and other thinking models) require thinking to be echoed back
+                                // Kimi/DeepSeek require reasoning_content on assistant tool call messages
                                 let thinking_content: Vec<MessageContent> = response.content.iter()
                                     .filter(|c| matches!(c, MessageContent::Thinking(_)))
                                     .cloned()
@@ -1361,10 +1362,25 @@ impl Agent {
                                     messages_to_add.push(thinking_msg);
                                 }
 
+                                // Collect reasoning content to attach to tool request messages
+                                let reasoning_content: Vec<MessageContent> = response.content.iter()
+                                    .filter(|c| matches!(c, MessageContent::Reasoning(_)))
+                                    .cloned()
+                                    .collect();
+
                                 for (idx, request) in frontend_requests.iter().chain(remaining_requests.iter()).enumerate() {
                                     if request.tool_call.is_ok() {
-                                        let request_msg = Message::assistant()
+                                        let mut request_msg = Message::assistant()
-                                            .with_id(format!("msg_{}", Uuid::new_v4()))
+                                            .with_id(format!("msg_{}", Uuid::new_v4()));
+
+                                        // Attach reasoning content to EVERY split tool request message.
+                                        // Providers like Kimi require reasoning_content on all assistant
+                                        // messages with tool_calls when thinking mode is enabled.
+                                        for rc in &reasoning_content {
+                                            request_msg = request_msg.with_content(rc.clone());
+                                        }
+
+                                        request_msg = request_msg
                                             .with_tool_request_with_metadata(
                                                 request.id.clone(),
                                                 request.tool_call.clone(),

crates/goose/src/agents/platform_extensions/apps.rs

@@ -290,8 +290,17 @@ impl AppsManagerClient {
         let messages = vec![Message::user().with_text(&user_prompt)];
         let tools = vec![Self::create_app_content_tool()];
 
+        let mut model_config = provider.get_model_config();
+        model_config.max_tokens = Some(16384);
+
         let (response, _usage) = provider
-            .complete(session_id, &system_prompt, &messages, &tools)
+            .complete_with_model(
+                Some(session_id),
+                &model_config,
+                &system_prompt,
+                &messages,
+                &tools,
+            )
             .await
             .map_err(|e| format!("LLM call failed: {}", e))?;
 
@@ -321,8 +330,17 @@ impl AppsManagerClient {
         let messages = vec![Message::user().with_text(&user_prompt)];
         let tools = vec![Self::update_app_content_tool()];
 
+        let mut model_config = provider.get_model_config();
+        model_config.max_tokens = Some(16384);
+
         let (response, _usage) = provider
-            .complete(session_id, &system_prompt, &messages, &tools)
+            .complete_with_model(
+                Some(session_id),
+                &model_config,
+                &system_prompt,
+                &messages,
+                &tools,
+            )
             .await
             .map_err(|e| format!("LLM call failed: {}", e))?;
 

crates/goose/src/providers/formats/openai.rs

@@ -82,7 +82,7 @@ pub fn format_messages(messages: &[Message], image_format: &ImageFormat) -> Vec<
         let mut output = Vec::new();
         let mut content_array = Vec::new();
         let mut text_array = Vec::new();
-        let mut reasoning_text: Option<String> = None;
+        let mut reasoning_text = String::new();
 
         for content in &message.content {
             match content {
@@ -116,7 +116,7 @@ pub fn format_messages(messages: &[Message], image_format: &ImageFormat) -> Vec<
                     continue;
                 }
                 MessageContent::Reasoning(r) => {
-                    reasoning_text = Some(r.text.clone());
+                    reasoning_text.push_str(&r.text);
                 }
                 MessageContent::ToolRequest(request) => match &request.tool_call {
                     Ok(tool_call) => {
@@ -278,15 +278,11 @@ pub fn format_messages(messages: &[Message], image_format: &ImageFormat) -> Vec<
             converted["content"] = json!(null);
         }
 
-        // DeepSeek requires reasoning_content field when tool_calls are present
+        // Include reasoning_content only when non-empty.
-        // Set it to the captured reasoning text, or empty string if not present
+        // Kimi rejects empty reasoning_content (""), so we must omit it entirely
-        if converted.get("tool_calls").is_some() {
+        // when there's no reasoning to send.
-            let reasoning = reasoning_text.unwrap_or_default();
+        if !reasoning_text.is_empty() {
-            converted["reasoning_content"] = json!(reasoning);
+            converted["reasoning_content"] = json!(reasoning_text);
-        } else if let Some(reasoning) = reasoning_text {
-            if !reasoning.is_empty() {
-                converted["reasoning_content"] = json!(reasoning);
-            }
         }
 
         if converted.get("content").is_some() || converted.get("tool_calls").is_some() {
@@ -542,6 +538,7 @@ where
         use futures::StreamExt;
 
         let mut accumulated_reasoning: Vec<Value> = Vec::new();
+        let mut accumulated_reasoning_content = String::new();
 
         'outer: while let Some(response) = stream.next().await {
             if response.as_ref().is_ok_and(|s| s == "data: [DONE]") {
@@ -562,6 +559,9 @@ where
                 if let Some(details) = &chunk.choices[0].delta.reasoning_details {
                     accumulated_reasoning.extend(details.iter().cloned());
                 }
+                if let Some(rc) = &chunk.choices[0].delta.reasoning_content {
+                    accumulated_reasoning_content.push_str(rc);
+                }
             }
 
             let mut usage = extract_usage_with_output_tokens(&chunk);
@@ -602,6 +602,9 @@ where
                                     if let Some(details) = &tool_chunk.choices[0].delta.reasoning_details {
                                         accumulated_reasoning.extend(details.iter().cloned());
                                     }
+                                    if let Some(rc) = &tool_chunk.choices[0].delta.reasoning_content {
+                                        accumulated_reasoning_content.push_str(rc);
+                                    }
                                     if let Some(delta_tool_calls) = &tool_chunk.choices[0].delta.tool_calls {
                                         for delta_call in delta_tool_calls {
                                             if let Some(index) = delta_call.index {
@@ -642,6 +645,10 @@ where
                 };
 
                 let mut contents = Vec::new();
+                if !accumulated_reasoning_content.is_empty() {
+                    contents.push(MessageContent::reasoning(&accumulated_reasoning_content));
+                    accumulated_reasoning_content.clear();
+                }
                 let mut sorted_indices: Vec<_> = tool_call_data.keys().cloned().collect();
                 sorted_indices.sort();
 

crates/goose/src/providers/gemini_cli.rs

@@ -221,10 +221,18 @@ impl GeminiCliProvider {
 
     fn parse_stream_json_response(events: &[Value]) -> Result<(Message, Usage), ProviderError> {
         let mut all_text_content = Vec::new();
+        let mut all_thinking_content = Vec::new();
         let mut usage = Usage::default();
 
         for parsed in events {
             match parsed.get("type").and_then(|t| t.as_str()) {
+                Some("thinking") => {
+                    if let Some(content) = parsed.get("content").and_then(|c| c.as_str()) {
+                        if !content.is_empty() {
+                            all_thinking_content.push(content.to_string());
+                        }
+                    }
+                }
                 Some("message") => {
                     if parsed.get("role").and_then(|r| r.as_str()) == Some("assistant") {
                         if let Some(content) = parsed.get("content").and_then(|c| c.as_str()) {
@@ -253,11 +261,16 @@ impl GeminiCliProvider {
             ));
         }
 
-        let message = Message::new(
+        let mut content = Vec::new();
-            Role::Assistant,
+
-            chrono::Utc::now().timestamp(),
+        let combined_thinking = all_thinking_content.join("");
-            vec![MessageContent::text(combined_text)],
+        if !combined_thinking.is_empty() {
-        );
+            content.push(MessageContent::thinking(combined_thinking, String::new()));
+        }
+
+        content.push(MessageContent::text(combined_text));
+
+        let message = Message::new(Role::Assistant, chrono::Utc::now().timestamp(), content);
 
         Ok((message, usage))
     }
@@ -395,6 +408,44 @@ mod tests {
         assert!(GeminiCliProvider::parse_stream_json_response(&empty).is_err());
     }
 
+    #[test]
+    fn test_parse_thinking_blocks() {
+        let events = vec![
+            json!({"type":"init","session_id":"abc","model":"gemini-2.5-pro"}),
+            json!({"type":"thinking","content":"Let me reason about this...","delta":true}),
+            json!({"type":"thinking","content":" Step 1: analyze the problem.","delta":true}),
+            json!({"type":"message","role":"assistant","content":"Here is the answer.","delta":true}),
+            json!({"type":"result","status":"success","stats":{"input_tokens":30,"output_tokens":15,"total_tokens":45}}),
+        ];
+        let (message, usage) = GeminiCliProvider::parse_stream_json_response(&events).unwrap();
+        assert_eq!(message.role, Role::Assistant);
+
+        // Should have thinking content followed by text content
+        assert_eq!(message.content.len(), 2);
+        let thinking = message.content[0]
+            .as_thinking()
+            .expect("first content should be thinking");
+        assert_eq!(
+            thinking.thinking,
+            "Let me reason about this... Step 1: analyze the problem."
+        );
+        assert_eq!(message.as_concat_text(), "Here is the answer.");
+        assert_eq!(usage.input_tokens, Some(30));
+        assert_eq!(usage.output_tokens, Some(15));
+    }
+
+    #[test]
+    fn test_parse_no_thinking_blocks() {
+        // When there's no thinking, message should only have text content
+        let events = vec![
+            json!({"type":"message","role":"assistant","content":"Direct answer.","delta":true}),
+            json!({"type":"result","status":"success","stats":{"input_tokens":10,"output_tokens":5,"total_tokens":15}}),
+        ];
+        let (message, _usage) = GeminiCliProvider::parse_stream_json_response(&events).unwrap();
+        assert_eq!(message.content.len(), 1);
+        assert_eq!(message.as_concat_text(), "Direct answer.");
+    }
+
     #[test]
     fn test_build_prompt_first_and_resume() {
         let provider = make_provider();

crates/goose/tests/providers.rs

@@ -179,15 +179,17 @@ impl ProviderTester {
             .complete(session_id, "You are a helpful assistant.", &[message], &[])
             .await?;
 
-        assert_eq!(
+        assert!(
-            response.content.len(),
+            !response.content.is_empty(),
-            1,
+            "Expected at least one content item in response"
-            "Expected single content item in response"
         );
 
         assert!(
-            matches!(response.content[0], MessageContent::Text(_)),
+            response
-            "Expected text response"
+                .content
+                .iter()
+                .any(|c| matches!(c, MessageContent::Text(_))),
+            "Expected at least one text content item in response"
         );
 
         println!(

ui/desktop/src/components/settings/providers/modal/ProviderConfiguationModal.tsx

@@ -105,8 +105,15 @@ export default function ProviderConfigurationModal({
 
     const toSubmit = Object.fromEntries(
       Object.entries(configValues)
-        .filter(([_k, entry]) => !!entry.value)
+        .filter(
-        .map(([k, entry]) => [k, entry.value || ''])
+          ([_k, entry]) =>
+            !!entry.value ||
+            (entry.serverValue != null && typeof entry.serverValue === 'string')
+        )
+        .map(([k, entry]) => [
+          k,
+          entry.value ?? (typeof entry.serverValue === 'string' ? entry.serverValue : ''),
+        ])
     );
 
     try {