Handle inline thinking in Codex Chat conversion

- Split leading <think> blocks from Chat content into Responses reasoning output.
- Keep assistant text output free of inline thinking tags.
- Support streamed inline thinking blocks before normal text deltas.
- Add regression coverage for MiniMax-style inline thinking responses.

src-tauri/src/proxy/providers/streaming_codex_chat.rs

@@ -2,6 +2,7 @@
 
 use super::transform_codex_chat::{
     chat_usage_to_responses_usage, response_id_from_chat_id, response_status_from_finish_reason,
+    split_leading_think_block, strip_leading_think_open_tag,
 };
 use crate::proxy::sse::{strip_sse_field, take_sse_block};
 use bytes::Bytes;
@@ -27,6 +28,20 @@ struct ReasoningItemState {
     done: bool,
 }
 
+#[derive(Debug, Clone, Copy, Default, PartialEq, Eq)]
+enum InlineThinkMode {
+    #[default]
+    Detecting,
+    Reasoning,
+    Text,
+}
+
+#[derive(Debug, Default)]
+struct InlineThinkState {
+    mode: InlineThinkMode,
+    buffer: String,
+}
+
 #[derive(Debug, Default)]
 struct ToolCallState {
     output_index: Option<u32>,
@@ -48,6 +63,7 @@ struct ChatToResponsesState {
     next_output_index: u32,
     text: TextItemState,
     reasoning: ReasoningItemState,
+    inline_think: InlineThinkState,
     tools: BTreeMap<usize, ToolCallState>,
     output_items: Vec<(u32, Value)>,
     latest_usage: Option<Value>,
@@ -65,6 +81,7 @@ impl Default for ChatToResponsesState {
             next_output_index: 0,
             text: TextItemState::default(),
             reasoning: ReasoningItemState::default(),
+            inline_think: InlineThinkState::default(),
             tools: BTreeMap::new(),
             output_items: Vec::new(),
             latest_usage: None,
@@ -110,12 +127,12 @@ impl ChatToResponsesState {
 
             if let Some(content) = delta.get("content").and_then(|v| v.as_str()) {
                 if !content.is_empty() {
-                    events.extend(self.finalize_reasoning());
-                    events.extend(self.push_text_delta(content));
+                    events.extend(self.push_content_delta(content));
                 }
             }
 
             if let Some(tool_calls) = delta.get("tool_calls").and_then(|v| v.as_array()) {
+                events.extend(self.flush_inline_think_at_boundary());
                 events.extend(self.finalize_reasoning());
                 for tool_call in tool_calls {
                     events.extend(self.push_tool_call_delta(tool_call));
@@ -130,6 +147,98 @@ impl ChatToResponsesState {
         events
     }
 
+    fn push_content_delta(&mut self, delta: &str) -> Vec<Bytes> {
+        match self.inline_think.mode {
+            InlineThinkMode::Text => {
+                let mut events = self.finalize_reasoning();
+                events.extend(self.push_text_delta(delta));
+                events
+            }
+            InlineThinkMode::Detecting => {
+                self.inline_think.buffer.push_str(delta);
+                match leading_think_prefix_decision(&self.inline_think.buffer) {
+                    ThinkPrefixDecision::NeedMore => Vec::new(),
+                    ThinkPrefixDecision::Reasoning => {
+                        self.inline_think.mode = InlineThinkMode::Reasoning;
+                        self.drain_complete_inline_think()
+                    }
+                    ThinkPrefixDecision::Text => {
+                        self.inline_think.mode = InlineThinkMode::Text;
+                        let text = std::mem::take(&mut self.inline_think.buffer);
+                        let mut events = self.finalize_reasoning();
+                        events.extend(self.push_text_delta(&text));
+                        events
+                    }
+                }
+            }
+            InlineThinkMode::Reasoning => {
+                self.inline_think.buffer.push_str(delta);
+                self.drain_complete_inline_think()
+            }
+        }
+    }
+
+    fn drain_complete_inline_think(&mut self) -> Vec<Bytes> {
+        let Some((reasoning, answer)) = split_leading_think_block(&self.inline_think.buffer) else {
+            return Vec::new();
+        };
+
+        self.inline_think.mode = InlineThinkMode::Text;
+        self.inline_think.buffer.clear();
+
+        let mut events = Vec::new();
+        if !reasoning.is_empty() {
+            events.extend(self.push_reasoning_delta(&reasoning));
+            events.extend(self.finalize_reasoning());
+        }
+        if !answer.is_empty() {
+            events.extend(self.push_text_delta(&answer));
+        }
+
+        events
+    }
+
+    fn flush_inline_think_at_boundary(&mut self) -> Vec<Bytes> {
+        match self.inline_think.mode {
+            InlineThinkMode::Text => Vec::new(),
+            InlineThinkMode::Detecting => {
+                self.inline_think.mode = InlineThinkMode::Text;
+                let text = std::mem::take(&mut self.inline_think.buffer);
+                if text.is_empty() {
+                    Vec::new()
+                } else {
+                    let mut events = self.finalize_reasoning();
+                    events.extend(self.push_text_delta(&text));
+                    events
+                }
+            }
+            InlineThinkMode::Reasoning => {
+                let buffered = std::mem::take(&mut self.inline_think.buffer);
+                self.inline_think.mode = InlineThinkMode::Text;
+                if let Some((reasoning, answer)) = split_leading_think_block(&buffered) {
+                    let mut events = Vec::new();
+                    if !reasoning.is_empty() {
+                        events.extend(self.push_reasoning_delta(&reasoning));
+                        events.extend(self.finalize_reasoning());
+                    }
+                    if !answer.is_empty() {
+                        events.extend(self.push_text_delta(&answer));
+                    }
+                    return events;
+                }
+
+                let reasoning = strip_leading_think_open_tag(&buffered).unwrap_or(buffered);
+                if reasoning.is_empty() {
+                    Vec::new()
+                } else {
+                    let mut events = self.push_reasoning_delta(&reasoning);
+                    events.extend(self.finalize_reasoning());
+                    events
+                }
+            }
+        }
+    }
+
     fn ensure_response_started(&mut self) -> Vec<Bytes> {
         if self.response_started {
             return Vec::new();
@@ -375,6 +484,7 @@ impl ChatToResponsesState {
         }
 
         let mut events = self.ensure_response_started();
+        events.extend(self.flush_inline_think_at_boundary());
         events.extend(self.finalize_reasoning());
         events.extend(self.finalize_text());
         events.extend(self.finalize_tools());
@@ -663,6 +773,29 @@ fn chat_delta_reasoning_text(delta: &Value) -> Option<String> {
     None
 }
 
+enum ThinkPrefixDecision {
+    NeedMore,
+    Reasoning,
+    Text,
+}
+
+fn leading_think_prefix_decision(buffer: &str) -> ThinkPrefixDecision {
+    let trimmed = buffer.trim_start();
+    if trimmed.is_empty() {
+        return ThinkPrefixDecision::NeedMore;
+    }
+
+    if trimmed.starts_with("<think>") {
+        return ThinkPrefixDecision::Reasoning;
+    }
+
+    if "<think>".starts_with(trimmed) {
+        return ThinkPrefixDecision::NeedMore;
+    }
+
+    ThinkPrefixDecision::Text
+}
+
 /// Create a stream that converts Chat Completions SSE chunks into Responses SSE events.
 pub fn create_responses_sse_stream_from_chat<E: std::error::Error + Send + 'static>(
     stream: impl Stream<Item = Result<Bytes, E>> + Send + 'static,
@@ -832,6 +965,24 @@ mod tests {
         assert!(reasoning_pos < message_pos);
     }
 
+    #[tokio::test]
+    async fn converts_inline_think_chat_sse_to_reasoning_without_leaking_tags() {
+        let output = collect(vec![
+            "data: {\"id\":\"chatcmpl_minimax\",\"created\":123,\"model\":\"MiniMax-M2.7\",\"choices\":[{\"delta\":{\"role\":\"assistant\",\"content\":\"<think>\\nNeed\"}}]}\n\n",
+            "data: {\"id\":\"chatcmpl_minimax\",\"created\":123,\"model\":\"MiniMax-M2.7\",\"choices\":[{\"delta\":{\"content\":\" context.</think>\\n\\npong\"},\"finish_reason\":\"stop\"}]}\n\n",
+            "data: {\"id\":\"chatcmpl_minimax\",\"created\":123,\"model\":\"MiniMax-M2.7\",\"choices\":[],\"usage\":{\"prompt_tokens\":4,\"completion_tokens\":6,\"total_tokens\":10,\"completion_tokens_details\":{\"reasoning_tokens\":3}}}\n\n",
+        ])
+        .await;
+
+        assert!(output.contains("event: response.reasoning_summary_text.delta"));
+        assert!(output.contains("Need context."));
+        assert!(output.contains("\"text\":\"pong\""));
+        assert!(output.contains("\"reasoning_tokens\":3"));
+        assert!(!output.contains("<think>"));
+        assert!(!output.contains("</think>"));
+        assert!(output.contains("event: response.completed"));
+    }
+
     #[tokio::test]
     async fn converts_tool_call_chat_sse_to_responses_sse() {
         let output = collect(vec![

src-tauri/src/proxy/providers/transform_codex_chat.rs

@@ -23,6 +23,8 @@ const EXTRA_CHAT_PASSTHROUGH_FIELDS: &[&str] = &[
     "top_logprobs",
     "user",
 ];
+const THINK_OPEN_TAG: &str = "<think>";
+const THINK_CLOSE_TAG: &str = "</think>";
 
 /// Convert an OpenAI Responses request into an OpenAI Chat Completions request.
 pub fn responses_to_chat_completions(body: Value) -> Result<Value, ProxyError> {
@@ -421,11 +423,20 @@ fn chat_reasoning_text(message: &Value) -> Option<String> {
         }
     }
 
-    let reasoning = message.get("reasoning")?;
-    for key in ["content", "text", "summary"] {
-        if let Some(text) = reasoning.get(key).and_then(|v| v.as_str()) {
-            if !text.is_empty() {
-                return Some(text.to_string());
+    if let Some(reasoning) = message.get("reasoning") {
+        for key in ["content", "text", "summary"] {
+            if let Some(text) = reasoning.get(key).and_then(|v| v.as_str()) {
+                if !text.is_empty() {
+                    return Some(text.to_string());
+                }
+            }
+        }
+    }
+
+    if let Some(content) = message.get("content").and_then(|v| v.as_str()) {
+        if let Some((reasoning, _answer)) = split_leading_think_block(content) {
+            if !reasoning.is_empty() {
+                return Some(reasoning);
             }
         }
     }
@@ -437,6 +448,9 @@ fn chat_message_to_response_output_item(message: &Value, response_id: &str) -> O
     let mut content = Vec::new();
 
     if let Some(text) = message.get("content").and_then(|v| v.as_str()) {
+        let text = split_leading_think_block(text)
+            .map(|(_reasoning, answer)| answer)
+            .unwrap_or_else(|| text.to_string());
         if !text.is_empty() {
             content.push(json!({
                 "type": "output_text",
@@ -496,6 +510,36 @@ fn chat_message_to_response_output_item(message: &Value, response_id: &str) -> O
     }))
 }
 
+pub(crate) fn split_leading_think_block(text: &str) -> Option<(String, String)> {
+    let leading_ws_len = text.len() - text.trim_start().len();
+    let after_ws = &text[leading_ws_len..];
+    if !after_ws.starts_with(THINK_OPEN_TAG) {
+        return None;
+    }
+
+    let body_start = leading_ws_len + THINK_OPEN_TAG.len();
+    let close_relative = text[body_start..].find(THINK_CLOSE_TAG)?;
+    let close_start = body_start + close_relative;
+    let answer_start = close_start + THINK_CLOSE_TAG.len();
+
+    Some((
+        text[body_start..close_start].trim().to_string(),
+        strip_think_answer_separator(&text[answer_start..]).to_string(),
+    ))
+}
+
+pub(crate) fn strip_leading_think_open_tag(text: &str) -> Option<String> {
+    let leading_ws_len = text.len() - text.trim_start().len();
+    let after_ws = &text[leading_ws_len..];
+    after_ws
+        .strip_prefix(THINK_OPEN_TAG)
+        .map(|value| value.trim().to_string())
+}
+
+fn strip_think_answer_separator(text: &str) -> &str {
+    text.trim_start_matches(['\r', '\n', '\t', ' '])
+}
+
 fn chat_tool_calls_to_response_output_items(message: &Value) -> Vec<Value> {
     let mut output = Vec::new();
 
@@ -788,6 +832,43 @@ mod tests {
         assert_eq!(result["usage"]["input_tokens_details"]["cached_tokens"], 3);
     }
 
+    #[test]
+    fn chat_response_to_responses_splits_inline_think_content() {
+        let input = json!({
+            "id": "chatcmpl_think",
+            "object": "chat.completion",
+            "created": 123,
+            "model": "MiniMax-M2.7",
+            "choices": [{
+                "message": {
+                    "role": "assistant",
+                    "content": "<think>\nI should answer with pong.\n</think>\n\npong"
+                },
+                "finish_reason": "stop"
+            }],
+            "usage": {
+                "prompt_tokens": 10,
+                "completion_tokens": 20,
+                "total_tokens": 30,
+                "completion_tokens_details": {"reasoning_tokens": 18}
+            }
+        });
+
+        let result = chat_completion_to_response(input).unwrap();
+
+        assert_eq!(result["output"][0]["type"], "reasoning");
+        assert_eq!(
+            result["output"][0]["summary"][0]["text"],
+            "I should answer with pong."
+        );
+        assert_eq!(result["output"][1]["type"], "message");
+        assert_eq!(result["output"][1]["content"][0]["text"], "pong");
+        assert_eq!(
+            result["usage"]["output_tokens_details"]["reasoning_tokens"],
+            18
+        );
+    }
+
     #[test]
     fn chat_response_length_maps_to_incomplete_response() {
         let input = json!({