fix: return thinking as reasoning output item per OpenAI spec

Thinking content was previously returned as non-standard top-level
fields (thinking, thinking_duration). Now follows the official OpenAI
Responses API format:

- Reasoning appears as a 'type: reasoning' item in the output array
  with summary[].text containing the thinking content
- Message item follows after the reasoning item
- thinking_signature kept as proxy extension (internal multi-turn data)
- Removed ResponseOutput/OutputContent structs in favor of
  serde_json::Value for polymorphic output items

src/api/types.rs

@@ -76,7 +76,9 @@ pub(crate) struct ResponsesResponse {
     #[serde(serialize_with = "serialize_option_u64")]
     pub max_output_tokens: Option<u64>,
     pub model: String,
-    pub output: Vec<ResponseOutput>,
+    /// Output items — can contain both `"reasoning"` and `"message"` items.
+    /// Uses serde_json::Value because reasoning and message items have different shapes.
+    pub output: Vec<serde_json::Value>,
     pub parallel_tool_calls: bool,
     pub previous_response_id: Option<String>,
     pub reasoning: Reasoning,
@@ -91,34 +93,10 @@ pub(crate) struct ResponsesResponse {
     pub user: Option<String>,
     pub metadata: serde_json::Value,
     /// Proxy extension: opaque thinking verification signature.
-    /// Present for all models. Required for multi-turn chaining with thinking models.
+    /// Required for multi-turn chaining with thinking models.
+    /// Not part of the official OpenAI spec — internal proxy data.
     #[serde(skip_serializing_if = "Option::is_none")]
     pub thinking_signature: Option<String>,
-    /// Proxy extension: the model's internal reasoning/thinking content.
-    /// Available for all models (Opus, Gemini Flash, Gemini Pro).
-    #[serde(skip_serializing_if = "Option::is_none")]
-    pub thinking: Option<String>,
-    /// Proxy extension: time spent thinking (e.g. "0.041999832s").
-    #[serde(skip_serializing_if = "Option::is_none")]
-    pub thinking_duration: Option<String>,
-}
-
-#[derive(Serialize, Clone)]
-pub(crate) struct ResponseOutput {
-    #[serde(rename = "type")]
-    pub output_type: &'static str,
-    pub id: String,
-    pub status: &'static str,
-    pub role: &'static str,
-    pub content: Vec<OutputContent>,
-}
-
-#[derive(Serialize, Clone)]
-pub(crate) struct OutputContent {
-    #[serde(rename = "type")]
-    pub content_type: &'static str,
-    pub text: String,
-    pub annotations: Vec<serde_json::Value>,
 }
 
 #[derive(Serialize, Clone)]
@@ -201,6 +179,47 @@ impl Default for TextFormat {
     }
 }
 
+// ─── Output item builders ────────────────────────────────────────────────────
+
+/// Build a reasoning output item (goes BEFORE the message item in `output`).
+/// Matches: https://platform.openai.com/docs/api-reference/responses
+pub fn build_reasoning_output(thinking_text: &str) -> serde_json::Value {
+    serde_json::json!({
+        "id": format!("rs_{}", uuid::Uuid::new_v4().to_string().replace('-', "")),
+        "type": "reasoning",
+        "summary": [{
+            "type": "summary_text",
+            "text": thinking_text,
+        }],
+    })
+}
+
+/// Build a message output item.
+pub fn build_message_output(msg_id: &str, text: &str) -> serde_json::Value {
+    serde_json::json!({
+        "type": "message",
+        "id": msg_id,
+        "status": "completed",
+        "role": "assistant",
+        "content": [{
+            "type": "output_text",
+            "text": text,
+            "annotations": [],
+        }],
+    })
+}
+
+/// Build an in-progress message output item (no content yet, for streaming).
+pub fn build_message_output_in_progress(msg_id: &str) -> serde_json::Value {
+    serde_json::json!({
+        "type": "message",
+        "id": msg_id,
+        "status": "in_progress",
+        "role": "assistant",
+        "content": [],
+    })
+}
+
 // ─── Helpers ─────────────────────────────────────────────────────────────────
 
 /// Serialize Option<u64> as either the number or JSON null (not omitted).