feat(providers): support reasoning_content fallback for thinking models

Reasoning/thinking models (Qwen3, GLM-4, DeepSeek, etc.) may return
output in `reasoning_content` instead of `content`. Add automatic
fallback for both OpenAI and OpenAI-compatible providers, including
streaming SSE support.

Changes:
- Add `reasoning_content` field to response structs in both providers
- Add `effective_content()` helper that prefers `content` but falls
  back to `reasoning_content` when content is empty/null/missing
- Update all extraction sites to use `effective_content()`
- Add streaming SSE fallback for `reasoning_content` chunks
- Add 16 focused unit tests covering all edge cases

Tested end-to-end against GLM-4.7-flash via local LLM server.

src/providers/compatible.rs

@@ -171,10 +171,26 @@ struct Choice {
 struct ResponseMessage {
     #[serde(default)]
     content: Option<String>,
+    /// Reasoning/thinking models (e.g. Qwen3, GLM-4) may return their output
+    /// in `reasoning_content` instead of `content`. Used as automatic fallback.
+    #[serde(default)]
+    reasoning_content: Option<String>,
     #[serde(default)]
     tool_calls: Option<Vec<ToolCall>>,
 }
 
+impl ResponseMessage {
+    /// Extract text content, falling back to `reasoning_content` when `content`
+    /// is missing or empty. Reasoning/thinking models (Qwen3, GLM-4, etc.)
+    /// often return their output solely in `reasoning_content`.
+    fn effective_content(&self) -> String {
+        match &self.content {
+            Some(c) if !c.is_empty() => c.clone(),
+            _ => self.reasoning_content.clone().unwrap_or_default(),
+        }
+    }
+}
+
 #[derive(Debug, Deserialize, Serialize)]
 struct ToolCall {
     #[serde(rename = "type")]
@@ -245,6 +261,9 @@ struct StreamChoice {
 struct StreamDelta {
     #[serde(default)]
     content: Option<String>,
+    /// Reasoning/thinking models may stream output via `reasoning_content`.
+    #[serde(default)]
+    reasoning_content: Option<String>,
 }
 
 /// Parse SSE (Server-Sent Events) stream from OpenAI-compatible providers.
@@ -274,6 +293,10 @@ fn parse_sse_line(line: &str) -> StreamResult<Option<String>> {
             if let Some(content) = &choice.delta.content {
                 return Ok(Some(content.clone()));
             }
+            // Fallback to reasoning_content for thinking models
+            if let Some(reasoning) = &choice.delta.reasoning_content {
+                return Ok(Some(reasoning.clone()));
+            }
         }
     }
 
@@ -529,10 +552,10 @@ impl Provider for OpenAiCompatibleProvider {
                         .map_or(false, |t| !t.is_empty())
                 {
                     serde_json::to_string(&c.message)
-                        .unwrap_or_else(|_| c.message.content.unwrap_or_default())
+                        .unwrap_or_else(|_| c.message.effective_content())
                 } else {
-                    // No tool calls, return content as-is
+                    // No tool calls, return content (with reasoning_content fallback)
-                    c.message.content.unwrap_or_default()
+                    c.message.effective_content()
                 }
             })
             .ok_or_else(|| anyhow::anyhow!("No response from {}", self.name))
@@ -617,10 +640,10 @@ impl Provider for OpenAiCompatibleProvider {
                         .map_or(false, |t| !t.is_empty())
                 {
                     serde_json::to_string(&c.message)
-                        .unwrap_or_else(|_| c.message.content.unwrap_or_default())
+                        .unwrap_or_else(|_| c.message.effective_content())
                 } else {
-                    // No tool calls, return content as-is
+                    // No tool calls, return content (with reasoning_content fallback)
-                    c.message.content.unwrap_or_default()
+                    c.message.effective_content()
                 }
             })
             .ok_or_else(|| anyhow::anyhow!("No response from {}", self.name))
@@ -1150,4 +1173,96 @@ mod tests {
         let result = provider.warmup().await;
         assert!(result.is_ok());
     }
+
+    // ══════════════════════════════════════════════════════════
+    // Reasoning model fallback tests (reasoning_content)
+    // ══════════════════════════════════════════════════════════
+
+    #[test]
+    fn reasoning_content_fallback_when_content_empty() {
+        // Reasoning models (Qwen3, GLM-4) return content: "" with reasoning_content populated
+        let json = r#"{"choices":[{"message":{"content":"","reasoning_content":"Thinking output here"}}]}"#;
+        let resp: ApiChatResponse = serde_json::from_str(json).unwrap();
+        let msg = &resp.choices[0].message;
+        assert_eq!(msg.effective_content(), "Thinking output here");
+    }
+
+    #[test]
+    fn reasoning_content_fallback_when_content_null() {
+        // Some models may return content: null with reasoning_content
+        let json =
+            r#"{"choices":[{"message":{"content":null,"reasoning_content":"Fallback text"}}]}"#;
+        let resp: ApiChatResponse = serde_json::from_str(json).unwrap();
+        let msg = &resp.choices[0].message;
+        assert_eq!(msg.effective_content(), "Fallback text");
+    }
+
+    #[test]
+    fn reasoning_content_fallback_when_content_missing() {
+        // content field absent entirely, reasoning_content present
+        let json = r#"{"choices":[{"message":{"reasoning_content":"Only reasoning"}}]}"#;
+        let resp: ApiChatResponse = serde_json::from_str(json).unwrap();
+        let msg = &resp.choices[0].message;
+        assert_eq!(msg.effective_content(), "Only reasoning");
+    }
+
+    #[test]
+    fn reasoning_content_not_used_when_content_present() {
+        // Normal model: content populated, reasoning_content should be ignored
+        let json = r#"{"choices":[{"message":{"content":"Normal response","reasoning_content":"Should be ignored"}}]}"#;
+        let resp: ApiChatResponse = serde_json::from_str(json).unwrap();
+        let msg = &resp.choices[0].message;
+        assert_eq!(msg.effective_content(), "Normal response");
+    }
+
+    #[test]
+    fn reasoning_content_both_absent_returns_empty() {
+        // Neither content nor reasoning_content — returns empty string
+        let json = r#"{"choices":[{"message":{}}]}"#;
+        let resp: ApiChatResponse = serde_json::from_str(json).unwrap();
+        let msg = &resp.choices[0].message;
+        assert_eq!(msg.effective_content(), "");
+    }
+
+    #[test]
+    fn reasoning_content_ignored_by_normal_models() {
+        // Standard response without reasoning_content still works
+        let json = r#"{"choices":[{"message":{"content":"Hello from Venice!"}}]}"#;
+        let resp: ApiChatResponse = serde_json::from_str(json).unwrap();
+        let msg = &resp.choices[0].message;
+        assert!(msg.reasoning_content.is_none());
+        assert_eq!(msg.effective_content(), "Hello from Venice!");
+    }
+
+    // ══════════════════════════════════════════════════════════
+    // SSE streaming reasoning_content fallback tests
+    // ══════════════════════════════════════════════════════════
+
+    #[test]
+    fn parse_sse_line_with_content() {
+        let line = r#"data: {"choices":[{"delta":{"content":"hello"}}]}"#;
+        let result = parse_sse_line(line).unwrap();
+        assert_eq!(result, Some("hello".to_string()));
+    }
+
+    #[test]
+    fn parse_sse_line_with_reasoning_content() {
+        let line = r#"data: {"choices":[{"delta":{"reasoning_content":"thinking..."}}]}"#;
+        let result = parse_sse_line(line).unwrap();
+        assert_eq!(result, Some("thinking...".to_string()));
+    }
+
+    #[test]
+    fn parse_sse_line_with_both_prefers_content() {
+        let line = r#"data: {"choices":[{"delta":{"content":"real answer","reasoning_content":"thinking..."}}]}"#;
+        let result = parse_sse_line(line).unwrap();
+        assert_eq!(result, Some("real answer".to_string()));
+    }
+
+    #[test]
+    fn parse_sse_line_done_sentinel() {
+        let line = "data: [DONE]";
+        let result = parse_sse_line(line).unwrap();
+        assert_eq!(result, None);
+    }
 }

src/providers/openai.rs

@@ -37,7 +37,20 @@ struct Choice {
 
 #[derive(Debug, Deserialize)]
 struct ResponseMessage {
-    content: String,
+    #[serde(default)]
+    content: Option<String>,
+    /// Reasoning/thinking models may return output in `reasoning_content`.
+    #[serde(default)]
+    reasoning_content: Option<String>,
+}
+
+impl ResponseMessage {
+    fn effective_content(&self) -> String {
+        match &self.content {
+            Some(c) if !c.is_empty() => c.clone(),
+            _ => self.reasoning_content.clone().unwrap_or_default(),
+        }
+    }
 }
 
 #[derive(Debug, Serialize)]
@@ -105,10 +118,22 @@ struct NativeChoice {
 struct NativeResponseMessage {
     #[serde(default)]
     content: Option<String>,
+    /// Reasoning/thinking models may return output in `reasoning_content`.
+    #[serde(default)]
+    reasoning_content: Option<String>,
     #[serde(default)]
     tool_calls: Option<Vec<NativeToolCall>>,
 }
 
+impl NativeResponseMessage {
+    fn effective_content(&self) -> Option<String> {
+        match &self.content {
+            Some(c) if !c.is_empty() => Some(c.clone()),
+            _ => self.reasoning_content.clone(),
+        }
+    }
+}
+
 impl OpenAiProvider {
     pub fn new(credential: Option<&str>) -> Self {
         Self {
@@ -205,6 +230,7 @@ impl OpenAiProvider {
     }
 
     fn parse_native_response(message: NativeResponseMessage) -> ProviderChatResponse {
+        let text = message.effective_content();
         let tool_calls = message
             .tool_calls
             .unwrap_or_default()
@@ -216,10 +242,7 @@ impl OpenAiProvider {
             })
             .collect::<Vec<_>>();
 
-        ProviderChatResponse {
+        ProviderChatResponse { text, tool_calls }
-            text: message.content,
-            tool_calls,
-        }
     }
 }
 
@@ -274,7 +297,7 @@ impl Provider for OpenAiProvider {
             .choices
             .into_iter()
             .next()
-            .map(|c| c.message.content)
+            .map(|c| c.message.effective_content())
             .ok_or_else(|| anyhow::anyhow!("No response from OpenAI"))
     }
 
@@ -417,7 +440,7 @@ mod tests {
         let json = r#"{"choices":[{"message":{"content":"Hi!"}}]}"#;
         let resp: ChatResponse = serde_json::from_str(json).unwrap();
         assert_eq!(resp.choices.len(), 1);
-        assert_eq!(resp.choices[0].message.content, "Hi!");
+        assert_eq!(resp.choices[0].message.effective_content(), "Hi!");
     }
 
     #[test]
@@ -432,14 +455,14 @@ mod tests {
         let json = r#"{"choices":[{"message":{"content":"A"}},{"message":{"content":"B"}}]}"#;
         let resp: ChatResponse = serde_json::from_str(json).unwrap();
         assert_eq!(resp.choices.len(), 2);
-        assert_eq!(resp.choices[0].message.content, "A");
+        assert_eq!(resp.choices[0].message.effective_content(), "A");
     }
 
     #[test]
     fn response_with_unicode() {
         let json = r#"{"choices":[{"message":{"content":"こんにちは 🦀"}}]}"#;
         let resp: ChatResponse = serde_json::from_str(json).unwrap();
-        assert_eq!(resp.choices[0].message.content, "こんにちは 🦀");
+        assert_eq!(resp.choices[0].message.effective_content(), "こんにちは 🦀");
     }
 
     #[test]
@@ -447,7 +470,10 @@ mod tests {
         let long = "x".repeat(100_000);
         let json = format!(r#"{{"choices":[{{"message":{{"content":"{long}"}}}}]}}"#);
         let resp: ChatResponse = serde_json::from_str(&json).unwrap();
-        assert_eq!(resp.choices[0].message.content.len(), 100_000);
+        assert_eq!(
+            resp.choices[0].message.content.as_ref().unwrap().len(),
+            100_000
+        );
     }
 
     #[tokio::test]
@@ -456,4 +482,48 @@ mod tests {
         let result = provider.warmup().await;
         assert!(result.is_ok());
     }
+
+    // ══════════════════════════════════════════════════════════
+    // Reasoning model fallback tests (reasoning_content)
+    // ══════════════════════════════════════════════════════════
+
+    #[test]
+    fn reasoning_content_fallback_empty_content() {
+        let json = r#"{"choices":[{"message":{"content":"","reasoning_content":"Thinking..."}}]}"#;
+        let resp: ChatResponse = serde_json::from_str(json).unwrap();
+        assert_eq!(resp.choices[0].message.effective_content(), "Thinking...");
+    }
+
+    #[test]
+    fn reasoning_content_fallback_null_content() {
+        let json =
+            r#"{"choices":[{"message":{"content":null,"reasoning_content":"Thinking..."}}]}"#;
+        let resp: ChatResponse = serde_json::from_str(json).unwrap();
+        assert_eq!(resp.choices[0].message.effective_content(), "Thinking...");
+    }
+
+    #[test]
+    fn reasoning_content_not_used_when_content_present() {
+        let json = r#"{"choices":[{"message":{"content":"Hello","reasoning_content":"Ignored"}}]}"#;
+        let resp: ChatResponse = serde_json::from_str(json).unwrap();
+        assert_eq!(resp.choices[0].message.effective_content(), "Hello");
+    }
+
+    #[test]
+    fn native_response_reasoning_content_fallback() {
+        let json =
+            r#"{"choices":[{"message":{"content":"","reasoning_content":"Native thinking"}}]}"#;
+        let resp: NativeChatResponse = serde_json::from_str(json).unwrap();
+        let msg = &resp.choices[0].message;
+        assert_eq!(msg.effective_content(), Some("Native thinking".to_string()));
+    }
+
+    #[test]
+    fn native_response_reasoning_content_ignored_when_content_present() {
+        let json =
+            r#"{"choices":[{"message":{"content":"Real answer","reasoning_content":"Ignored"}}]}"#;
+        let resp: NativeChatResponse = serde_json::from_str(json).unwrap();
+        let msg = &resp.choices[0].message;
+        assert_eq!(msg.effective_content(), Some("Real answer".to_string()));
+    }
 }