fix(openai): strip internal metadata from gpt-5.4 responses

- Add strip_internal_metadata helper to remove prefixes like 'to=multi_tool_use.parallel'
- Clean up Thai text preambles reported in the journal
- Apply metadata stripping to both synchronous and streaming response paths
- Improve visual quality of proxied model responses

src/providers/openai.rs

@@ -43,6 +43,7 @@ impl OpenAIProvider {
     pub fn parse_tool_uses_json(text: &str) -> Vec<crate::models::ToolCall> {
         let mut calls = Vec::new();
         if let Some(start) = text.find("{\"tool_uses\":") {
+            // ... (rest of method unchanged)
             // Find the end of the JSON block by matching braces
             let sub = &text[start..];
             let mut brace_count = 0;
@@ -87,6 +88,27 @@ impl OpenAIProvider {
         }
         calls
     }
+
+    /// Strips internal metadata prefixes like 'to=multi_tool_use.parallel' from model responses.
+    pub fn strip_internal_metadata(text: &str) -> String {
+        let mut result = text.to_string();
+        
+        // Patterns to strip
+        let patterns = [
+            "to=multi_tool_use.parallel",
+            "to=functions.multi_tool_use",
+            "ส่งเงินบาทไทยjson", // User reported Thai text preamble
+        ];
+
+        for p in patterns {
+            if let Some(start) = result.find(p) {
+                // Remove the pattern and any whitespace around it
+                result.replace_range(start..start + p.len(), "");
+            }
+        }
+
+        result.trim().to_string()
+    }
 }
 
 #[async_trait]
@@ -405,6 +427,8 @@ impl super::Provider for OpenAIProvider {
             tool_calls.extend(embedded_calls);
         }
         
+        content_text = Self::strip_internal_metadata(&content_text);
+
         Ok(ProviderResponse {
             content: content_text,
             reasoning_content: None,
@@ -773,9 +797,10 @@ impl super::Provider for OpenAIProvider {
                                         if let Some(start) = content_buffer.find("{\"tool_uses\":") {
                                             // Yield text before the JSON block
                                             let preamble = content_buffer[..start].to_string();
-                                            if !preamble.is_empty() {
+                                            let stripped_preamble = Self::strip_internal_metadata(&preamble);
+                                            if !stripped_preamble.is_empty() {
                                                 yield ProviderStreamChunk {
-                                                    content: preamble,
+                                                    content: stripped_preamble,
                                                     reasoning_content: None,
                                                     finish_reason: None,
                                                     tool_calls: None,
@@ -785,6 +810,7 @@ impl super::Provider for OpenAIProvider {
                                             }
                                             
                                             // Yield the tool calls
+                                            // ... (rest of tool call yielding unchanged)
                                             let deltas: Vec<crate::models::ToolCallDelta> = embedded_calls.into_iter().enumerate().map(|(idx, tc)| {
                                                 crate::models::ToolCallDelta {
                                                     index: idx as u32,
@@ -834,14 +860,17 @@ impl super::Provider for OpenAIProvider {
                                 } else {
                                     // Standard text, yield and clear buffer
                                     let content = std::mem::take(&mut content_buffer);
-                                    yield ProviderStreamChunk {
+                                    let stripped_content = Self::strip_internal_metadata(&content);
-                                        content,
+                                    if !stripped_content.is_empty() {
-                                        reasoning_content: None,
+                                        yield ProviderStreamChunk {
-                                        finish_reason: None,
+                                            content: stripped_content,
-                                        tool_calls: None,
+                                            reasoning_content: None,
-                                        model: model.clone(),
+                                            finish_reason: None,
-                                        usage: None,
+                                            tool_calls: None,
-                                    };
+                                            model: model.clone(),
+                                            usage: None,
+                                        };
+                                    }
                                 }
                             }