diff --git a/codex-rs/core/src/client_common.rs b/codex-rs/core/src/client_common.rs
index 48d0523df..ac285dd16 100644
--- a/codex-rs/core/src/client_common.rs
+++ b/codex-rs/core/src/client_common.rs
@@ -54,7 +54,15 @@ impl Default for Prompt {
 
 impl Prompt {
     pub(crate) fn get_formatted_input(&self) -> Vec<ResponseItem> {
-        self.input
+        Self::format_input(&self.input)
+    }
+
+    /// Converts persisted assistant envelopes into the model-native input items
+    /// used by the Responses API. Callers preparing history for retention must
+    /// use the same conversion as the request builder so encrypted delegated
+    /// messages are not filtered as ordinary assistant text.
+    pub(crate) fn format_input(input: &[ResponseItem]) -> Vec<ResponseItem> {
+        input
             .iter()
             .cloned()
             .map(|item| {
diff --git a/codex-rs/core/src/compact_remote_v2.rs b/codex-rs/core/src/compact_remote_v2.rs
index 0c04d2d18..be0c0122d 100644
--- a/codex-rs/core/src/compact_remote_v2.rs
+++ b/codex-rs/core/src/compact_remote_v2.rs
@@ -16,6 +16,7 @@ use crate::compact_remote::log_remote_compact_failure;
 use crate::compact_remote::process_compacted_history;
 use crate::compact_remote::should_keep_compacted_history_item;
 use crate::compact_remote::trim_function_call_history_to_fit_context_window;
+use crate::context_manager::estimate_item_token_count;
 use crate::hook_runtime::PostCompactHookOutcome;
 use crate::hook_runtime::PreCompactHookOutcome;
 use crate::hook_runtime::run_post_compact_hooks;
@@ -55,6 +56,7 @@ use tracing::info;
 // Mirror the current /responses/compact retained-message default while the
 // server-side path remains the reference implementation.
 const RETAINED_MESSAGE_TOKEN_BUDGET: usize = 64_000;
+const MAX_RETAINED_AGENT_MESSAGE_TOKENS: i64 = 10_000;
 // Compact attempts can run much longer than normal turns, so keep the per-transport
 // retry budget smaller than the general Responses stream retry budget.
 const MAX_REMOTE_COMPACTION_V2_STREAM_RETRIES: u64 = 2;
@@ -392,7 +394,11 @@ async fn run_remote_compact_v2_attempt(
     let prompt_input = history
         .clone()
         .for_prompt(&turn_context.model_info.input_modalities);
-    let mut input = prompt_input.clone();
+    // Apply the same persisted-envelope conversion used by the Responses request
+    // builder before retention. Otherwise encrypted delegated messages are still
+    // assistant envelopes here and the retention role filter drops them.
+    let formatted_prompt_input = Prompt::format_input(&prompt_input);
+    let mut input = formatted_prompt_input.clone();
     input.push(ResponseItem::CompactionTrigger);
     let prompt = Prompt {
         input,
@@ -450,7 +456,7 @@ async fn run_remote_compact_v2_attempt(
     .await;
     Ok(RemoteCompactV2Attempt {
         trace_input_history,
-        prompt_input,
+        prompt_input: formatted_prompt_input,
         compaction_output,
         token_usage,
         owned_client_session,
@@ -616,6 +622,10 @@ fn build_v2_compacted_history(
 }
 
 fn is_retained_for_remote_compaction_v2(item: &ResponseItem) -> bool {
+    if matches!(item, ResponseItem::AgentMessage { .. }) {
+        return estimate_item_token_count(item) <= MAX_RETAINED_AGENT_MESSAGE_TOKENS;
+    }
+
     let ResponseItem::Message { role, .. } = item else {
         return false;
     };
@@ -651,7 +661,7 @@ fn truncate_retained_messages_for_remote_compaction(
 
 fn message_text_token_count(item: &ResponseItem) -> usize {
     let ResponseItem::Message { content, .. } = item else {
-        return 0;
+        return usize::try_from(estimate_item_token_count(item)).unwrap_or(usize::MAX);
     };
 
     content
@@ -676,7 +686,7 @@ fn truncate_message_text_to_token_budget(
         phase,
     } = item
     else {
-        return Some(item);
+        return None;
     };
 
     let mut remaining = max_tokens;
@@ -718,8 +728,10 @@ fn truncate_message_text_to_token_budget(
 #[cfg(test)]
 mod tests {
     use super::*;
+    use codex_protocol::AgentPath;
     use codex_protocol::models::ContentItem;
     use codex_protocol::models::MessagePhase;
+    use codex_protocol::protocol::InterAgentCommunication;
     use pretty_assertions::assert_eq;
     use tokio::sync::mpsc;
     use tokio_util::sync::CancellationToken;
@@ -780,6 +792,35 @@ mod tests {
         );
     }
 
+    #[test]
+    fn build_v2_compacted_history_retains_encrypted_delegated_message_after_formatting() {
+        let communication = InterAgentCommunication::new_encrypted(
+            AgentPath::root(),
+            AgentPath::root().join("worker").expect("valid worker path"),
+            Vec::new(),
+            "encrypted delegated task".to_string(),
+            /*trigger_turn*/ true,
+        );
+        let persisted_envelope: ResponseItem = communication.to_response_input_item().into();
+        assert!(matches!(
+            persisted_envelope,
+            ResponseItem::Message { ref role, .. } if role == "assistant"
+        ));
+
+        let formatted = Prompt::format_input(&[persisted_envelope]);
+        assert!(matches!(
+            formatted.first(),
+            Some(ResponseItem::AgentMessage { .. })
+        ));
+
+        let output = ResponseItem::Compaction {
+            encrypted_content: "new".to_string(),
+        };
+        let history = build_v2_compacted_history(&formatted, output.clone());
+
+        assert_eq!(history, vec![formatted[0].clone(), output]);
+    }
+
     #[test]
     fn build_v2_compacted_history_discards_messages_before_truncating() {
         let old = message("user", "old", /*phase*/ None);
diff --git a/codex-rs/core/src/context_manager/history.rs b/codex-rs/core/src/context_manager/history.rs
index 6086fbcd8..fd8f24b5b 100644
--- a/codex-rs/core/src/context_manager/history.rs
+++ b/codex-rs/core/src/context_manager/history.rs
@@ -612,7 +612,7 @@ fn estimate_encrypted_function_output_length(encoded_len: usize) -> usize {
     encoded_len.saturating_mul(9).div_ceil(16)
 }
 
-fn estimate_item_token_count(item: &ResponseItem) -> i64 {
+pub(crate) fn estimate_item_token_count(item: &ResponseItem) -> i64 {
     let model_visible_bytes = estimate_response_item_model_visible_bytes(item);
     approx_tokens_from_byte_count_i64(model_visible_bytes)
 }
diff --git a/codex-rs/core/src/context_manager/mod.rs b/codex-rs/core/src/context_manager/mod.rs
index 2295c49df..8c12ed939 100644
--- a/codex-rs/core/src/context_manager/mod.rs
+++ b/codex-rs/core/src/context_manager/mod.rs
@@ -4,6 +4,7 @@ pub(crate) mod updates;
 
 pub(crate) use history::ContextManager;
 pub(crate) use history::TotalTokenUsageBreakdown;
+pub(crate) use history::estimate_item_token_count;
 pub(crate) use history::estimate_response_item_model_visible_bytes;
 pub(crate) use history::is_user_turn_boundary;
 pub(crate) use history::truncate_function_output_payload;