[prompts] Better user_instructions handling (#1836)

## Summary
Our recent change in #1737 can sometimes lead to the model confusing
AGENTS.md context as part of the message. But a little prompting and
formatting can help fix this!

## Testing
- Ran locally with a few different prompts to verify the model
behaves well.
- Updated unit tests

codex-rs/core/prompt.md

@@ -9,6 +9,8 @@ You MUST adhere to the following criteria when executing the task:
 - Analyzing code for vulnerabilities is allowed.
 - Showing user code and tool call details is allowed.
 - User instructions may overwrite the _CODING GUIDELINES_ section in this developer message.
+- `user_instructions` are not part of the user's request, but guidance for how to complete the task.
+- Do not cite `user_instructions` back to the user unless a specific piece is relevant.
 - Do not use \`ls -R\`, \`find\`, or \`grep\` - these are slow in large repos. Use \`rg\` and \`rg --files\`.
 - Use \`apply_patch\` to edit files: {"command":["apply_patch","*** Begin Patch\\n*** Update File: path/to/file.py\\n@@ def example():\\n- pass\\n+ return 123\\n*** End Patch"]}
 - If completing the user's task requires writing or modifying files:

codex-rs/core/src/chat_completions.rs

@@ -40,7 +40,7 @@ pub(crate) async fn stream_chat_completions(
     let full_instructions = prompt.get_full_instructions(model);
     messages.push(json!({"role": "system", "content": full_instructions}));
 
-    if let Some(instr) = &prompt.user_instructions {
+    if let Some(instr) = &prompt.get_formatted_user_instructions() {
         messages.push(json!({"role": "user", "content": instr}));
     }
 

codex-rs/core/src/client.rs

@@ -144,11 +144,11 @@ impl ModelClient {
         };
 
         let mut input_with_instructions = Vec::with_capacity(prompt.input.len() + 1);
-        if let Some(ui) = &prompt.user_instructions {
+        if let Some(ui) = prompt.get_formatted_user_instructions() {
             input_with_instructions.push(ResponseItem::Message {
                 id: None,
                 role: "user".to_string(),
-                content: vec![ContentItem::InputText { text: ui.clone() }],
+                content: vec![ContentItem::InputText { text: ui }],
             });
         }
         input_with_instructions.extend(prompt.input.clone());

codex-rs/core/src/client_common.rs

@@ -17,6 +17,10 @@ use tokio::sync::mpsc;
 /// with this content.
 const BASE_INSTRUCTIONS: &str = include_str!("../prompt.md");
 
+/// wraps user instructions message in a tag for the model to parse more easily.
+const USER_INSTRUCTIONS_START: &str = "<user_instructions>\n\n";
+const USER_INSTRUCTIONS_END: &str = "\n\n</user_instructions>";
+
 /// API request payload for a single model turn.
 #[derive(Default, Debug, Clone)]
 pub struct Prompt {
@@ -49,6 +53,12 @@ impl Prompt {
         }
         Cow::Owned(sections.join("\n"))
     }
+
+    pub(crate) fn get_formatted_user_instructions(&self) -> Option<String> {
+        self.user_instructions
+            .as_ref()
+            .map(|ui| format!("{USER_INSTRUCTIONS_START}{ui}{USER_INSTRUCTIONS_END}"))
+    }
 }
 
 #[derive(Debug)]

codex-rs/core/tests/client.rs

@@ -376,7 +376,13 @@ async fn includes_user_instructions_message_in_request() {
         request_body["input"][0]["content"][0]["text"]
             .as_str()
             .unwrap()
-            .starts_with("be nice")
+            .starts_with("<user_instructions>\n\nbe nice")
+    );
+    assert!(
+        request_body["input"][0]["content"][0]["text"]
+            .as_str()
+            .unwrap()
+            .ends_with("</user_instructions>")
     );
 }