diff --git a/src/providers/openai.rs b/src/providers/openai.rs
index 0e5c5839..08d26053 100644
--- a/src/providers/openai.rs
+++ b/src/providers/openai.rs
@@ -149,11 +149,34 @@ impl super::Provider for OpenAIProvider {
             }));
         }
 
+        let mut body = serde_json::json!({
+            "model": request.model,
+            "input": input_parts,
+        });
+
+        // Add standard parameters
+        if let Some(temp) = request.temperature {
+            body["temperature"] = serde_json::json!(temp);
+        }
+        
+        // Newer models (gpt-5, o1) prefer max_completion_tokens
+        if let Some(max_tokens) = request.max_tokens {
+            if request.model.contains("gpt-5") || request.model.starts_with("o1-") || request.model.starts_with("o3-") {
+                body["max_completion_tokens"] = serde_json::json!(max_tokens);
+            } else {
+                body["max_tokens"] = serde_json::json!(max_tokens);
+            }
+        }
+
+        if let Some(tools) = &request.tools {
+            body["tools"] = serde_json::json!(tools);
+        }
+
         let resp = self
             .client
             .post(format!("{}/responses", self.config.base_url))
             .header("Authorization", format!("Bearer {}", self.api_key))
-            .json(&serde_json::json!({ "model": request.model, "input": input_parts }))
+            .json(&body)
             .send()
             .await
             .map_err(|e| AppError::ProviderError(e.to_string()))?;
@@ -173,8 +196,8 @@ impl super::Provider for OpenAIProvider {
         // Normalize Responses API output into ProviderResponse
         let mut content_text = String::new();
         if let Some(output) = resp_json.get("output").and_then(|o| o.as_array()) {
-            if let Some(first) = output.get(0) {
-                if let Some(contents) = first.get("content").and_then(|c| c.as_array()) {
+            for out in output {
+                if let Some(contents) = out.get("content").and_then(|c| c.as_array()) {
                     for item in contents {
                         if let Some(text) = item.get("text").and_then(|t| t.as_str()) {
                             if !content_text.is_empty() { content_text.push_str("\n"); }
@@ -388,12 +411,27 @@ impl super::Provider for OpenAIProvider {
             }));
         }
 
-        let body = serde_json::json!({
+        let mut body = serde_json::json!({
             "model": request.model,
             "input": input_parts,
-            "stream": true
+            "stream": true,
+            "stream_options": { "include_usage": true }
         });
 
+        // Add standard parameters
+        if let Some(temp) = request.temperature {
+            body["temperature"] = serde_json::json!(temp);
+        }
+        
+        // Newer models (gpt-5, o1) prefer max_completion_tokens
+        if let Some(max_tokens) = request.max_tokens {
+            if request.model.contains("gpt-5") || request.model.starts_with("o1-") || request.model.starts_with("o3-") {
+                body["max_completion_tokens"] = serde_json::json!(max_tokens);
+            } else {
+                body["max_tokens"] = serde_json::json!(max_tokens);
+            }
+        }
+
         let url = format!("{}/responses", self.config.base_url);
         let api_key = self.api_key.clone();
         let model = request.model.clone();
@@ -420,32 +458,33 @@ impl super::Provider for OpenAIProvider {
                         let chunk: serde_json::Value = serde_json::from_str(&msg.data)
                             .map_err(|e| AppError::ProviderError(format!("Failed to parse Responses stream chunk: {}", e)))?;
 
-                        // Try standard OpenAI parsing first
+                        // Try standard OpenAI parsing first (choices/usage)
                         if let Some(p_chunk) = helpers::parse_openai_stream_chunk(&chunk, &model, None) {
                             yield p_chunk?;
                         } else {
                             // Responses API specific parsing for streaming
-                            // Often it follows a similar structure to the non-streaming response but in chunks
                             let mut content = String::new();
                             
                             // Check for output[0].content[0].text (similar to non-stream)
                             if let Some(output) = chunk.get("output").and_then(|o| o.as_array()) {
-                                if let Some(first) = output.get(0) {
-                                    if let Some(contents) = first.get("content").and_then(|c| c.as_array()) {
+                                for out in output {
+                                    if let Some(contents) = out.get("content").and_then(|c| c.as_array()) {
                                         for item in contents {
                                             if let Some(text) = item.get("text").and_then(|t| t.as_str()) {
                                                 content.push_str(text);
+                                            } else if let Some(delta) = item.get("delta").and_then(|d| d.get("text")).and_then(|t| t.as_str()) {
+                                                content.push_str(delta);
                                             }
                                         }
                                     }
                                 }
                             }
                             
-                            // Check for candidates[0].content.parts[0].text (Gemini-like, which OpenAI sometimes uses for v1/responses)
+                            // Check for candidates[0].content.parts[0].text
                             if content.is_empty() {
                                 if let Some(cands) = chunk.get("candidates").and_then(|c| c.as_array()) {
-                                    if let Some(c0) = cands.get(0) {
-                                        if let Some(content_obj) = c0.get("content") {
+                                    for c in cands {
+                                        if let Some(content_obj) = c.get("content") {
                                             if let Some(parts) = content_obj.get("parts").and_then(|p| p.as_array()) {
                                                 for p in parts {
                                                     if let Some(t) = p.get("text").and_then(|v| v.as_str()) {
@@ -488,6 +527,8 @@ impl super::Provider for OpenAIProvider {
                                 Err(AppError::ProviderError(format!("OpenAI Responses API error ({}): {}", status, error_body)))?;
                             }
                             Ok(_) => {
+                                // If the probe returned 200, but the stream ended, it might be a silent failure or timeout.
+                                tracing::warn!("Responses stream ended prematurely (probe returned 200)");
                                 Err(AppError::ProviderError(format!("Responses stream error (probe returned 200): {}", e)))?;
                             }
                             Err(probe_err) => {