feat: structured multi-agent output (#15515)

Send input now sends messages as assistant message and with this format:

```
author: /root/worker_a
recipient: /root/worker_a/tester
other_recipients: []
Content: bla bla bla. Actual content. Only text for now
```

codex-rs/core/src/context_manager/history.rs

@@ -1,3 +1,4 @@
+use crate::agent::inter_agent_instruction::InterAgentInstruction;
 use crate::codex::TurnContext;
 use crate::context_manager::normalize;
 use crate::event_mapping::is_contextual_user_message_content;
@@ -204,9 +205,10 @@ impl ContextManager {
         }
     }
 
-    /// Drop the last `num_turns` user turns from this history.
+    /// Drop the last `num_turns` instruction turns from this history.
     ///
-    /// "User turns" are identified as `ResponseItem::Message` entries whose role is `"user"`.
+    /// Instruction turns are history messages that should behave like a new prompt boundary:
+    /// ordinary user messages and structured assistant inter-agent instructions.
     ///
     /// This mirrors thread-rollback semantics:
     /// - `num_turns == 0` is a no-op
@@ -248,7 +250,7 @@ impl ContextManager {
     }
 
     fn get_non_last_reasoning_items_tokens(&self) -> i64 {
-        // Get reasoning items excluding all the ones after the last user message.
+        // Get reasoning items excluding all the ones after the last instruction boundary.
         let Some(last_user_index) = self.items.iter().rposition(is_user_turn_boundary) else {
             return 0;
         };
@@ -631,7 +633,12 @@ pub(crate) fn is_user_turn_boundary(item: &ResponseItem) -> bool {
         return false;
     };
 
-    role == "user" && !is_contextual_user_message_content(content)
+    (role == "user" && !is_contextual_user_message_content(content))
+        || (role == "assistant" && is_inter_agent_instruction_content(content))
+}
+
+fn is_inter_agent_instruction_content(content: &[ContentItem]) -> bool {
+    InterAgentInstruction::is_message_content(content)
 }
 
 fn user_message_positions(items: &[ResponseItem]) -> Vec<usize> {

codex-rs/core/src/context_manager/history_tests.rs

@@ -38,6 +38,18 @@ fn assistant_msg(text: &str) -> ResponseItem {
     }
 }
 
+fn inter_agent_assistant_msg(text: &str) -> ResponseItem {
+    ResponseItem::Message {
+        id: None,
+        role: "assistant".to_string(),
+        content: vec![ContentItem::OutputText {
+            text: text.to_string(),
+        }],
+        end_turn: None,
+        phase: None,
+    }
+}
+
 fn create_history_with_items(items: Vec<ResponseItem>) -> ContextManager {
     let mut h = ContextManager::new();
     // Use a generous but fixed token budget; tests only rely on truncation
@@ -225,6 +237,35 @@ fn items_after_last_model_generated_tokens_are_zero_without_model_generated_item
     );
 }
 
+#[test]
+fn inter_agent_assistant_messages_are_turn_boundaries() {
+    let item = inter_agent_assistant_msg(
+        "author: /root\nrecipient: /root/worker\nother_recipients: []\nContent: continue",
+    );
+
+    assert!(is_user_turn_boundary(&item));
+}
+
+#[test]
+fn drop_last_n_user_turns_treats_inter_agent_assistant_messages_as_instruction_turns() {
+    let first_turn = user_input_text_msg("first");
+    let first_reply = assistant_msg("done");
+    let inter_agent_turn = inter_agent_assistant_msg(
+        "author: /root\nrecipient: /root/worker\nother_recipients: []\nContent: continue",
+    );
+    let inter_agent_reply = assistant_msg("worker reply");
+    let mut history = create_history_with_items(vec![
+        first_turn.clone(),
+        first_reply.clone(),
+        inter_agent_turn,
+        inter_agent_reply,
+    ]);
+
+    history.drop_last_n_user_turns(1);
+
+    assert_eq!(history.raw_items(), &vec![first_turn, first_reply]);
+}
+
 #[test]
 fn total_token_usage_includes_all_items_after_last_model_generated_item() {
     let mut history = create_history_with_items(vec![assistant_msg("already counted by API")]);