diff --git a/codex-rs/codex-api/src/endpoint/compact.rs b/codex-rs/codex-api/src/endpoint/compact.rs
index 2b02ebd0f0..969083bb10 100644
--- a/codex-rs/codex-api/src/endpoint/compact.rs
+++ b/codex-rs/codex-api/src/endpoint/compact.rs
@@ -45,7 +45,7 @@ impl<T: HttpTransport, A: AuthProvider> CompactClient<T, A> {
         }
     }
 
-    pub async fn compact(
+    async fn compact(
         &self,
         body: serde_json::Value,
         extra_headers: HeaderMap,
diff --git a/codex-rs/codex-api/src/endpoint/responses.rs b/codex-rs/codex-api/src/endpoint/responses.rs
index 8b9938b549..0caf280657 100644
--- a/codex-rs/codex-api/src/endpoint/responses.rs
+++ b/codex-rs/codex-api/src/endpoint/responses.rs
@@ -50,10 +50,7 @@ impl<T: HttpTransport, A: AuthProvider> ResponsesClient<T, A> {
         }
     }
 
-    pub async fn stream_request(
-        &self,
-        request: ResponsesRequest,
-    ) -> Result<ResponseStream, ApiError> {
+    async fn stream_request(&self, request: ResponsesRequest) -> Result<ResponseStream, ApiError> {
         self.stream(request.body, request.headers).await
     }
 
@@ -95,6 +92,7 @@ impl<T: HttpTransport, A: AuthProvider> ResponsesClient<T, A> {
         }
     }
 
+    // Pub mainly for testing purpose.
     pub async fn stream(
         &self,
         body: Value,
diff --git a/codex-rs/core/src/client.rs b/codex-rs/core/src/client.rs
index 3c7a645a18..6948e02dc8 100644
--- a/codex-rs/core/src/client.rs
+++ b/codex-rs/core/src/client.rs
@@ -322,9 +322,10 @@ impl ModelClient {
     /// `ResponseItem`s representing the compacted transcript.
     pub async fn compact_conversation_history(
         &self,
-        prompt: &PromptBuilder,
+        input: &[ResponseItem],
+        instructions: &str,
     ) -> Result<Vec<ResponseItem>> {
-        if prompt.input.is_empty() {
+        if input.is_empty() {
             return Ok(Vec::new());
         }
         let auth_manager = self.auth_manager.clone();
@@ -338,10 +339,11 @@ impl ModelClient {
         let client = ApiCompactClient::new(transport, api_provider, api_auth)
             .with_telemetry(Some(request_telemetry));
 
-        let instructions = prompt
-            .get_full_instructions(&self.config.model_family)
-            .into_owned();
-        let payload = prompt.build_compaction_input(&self.config.model, &instructions);
+        let payload = codex_api::CompactionInput {
+            model: &self.config.model,
+            input,
+            instructions,
+        };
 
         let mut extra_headers = ApiHeaderMap::new();
         if let SessionSource::SubAgent(sub) = &self.session_source {
diff --git a/codex-rs/core/src/client_common.rs b/codex-rs/core/src/client_common.rs
index 4d28deea29..27826f81bd 100644
--- a/codex-rs/core/src/client_common.rs
+++ b/codex-rs/core/src/client_common.rs
@@ -5,7 +5,6 @@ use crate::model_family::ModelFamily;
 use crate::model_provider_info::WireApi;
 use crate::tools::spec::create_tools_json_for_chat_completions_api;
 use crate::tools::spec::create_tools_json_for_responses_api;
-use codex_api::CompactionInput;
 use codex_api::Prompt;
 pub use codex_api::common::ResponseEvent;
 use codex_apply_patch::APPLY_PATCH_TOOL_INSTRUCTIONS;
@@ -153,18 +152,6 @@ impl PromptBuilder {
         })
     }
 
-    pub(crate) fn build_compaction_input<'a>(
-        &'a self,
-        model: &'a str,
-        instructions: &'a str,
-    ) -> CompactionInput<'a> {
-        CompactionInput {
-            model,
-            input: &self.input,
-            instructions,
-        }
-    }
-
     pub(crate) fn get_full_instructions<'a>(&'a self, model: &'a ModelFamily) -> Cow<'a, str> {
         let base = self
             .base_instructions_override
diff --git a/codex-rs/core/src/compact_remote.rs b/codex-rs/core/src/compact_remote.rs
index 9809e658b0..5498c6e69a 100644
--- a/codex-rs/core/src/compact_remote.rs
+++ b/codex-rs/core/src/compact_remote.rs
@@ -1,6 +1,6 @@
 use std::sync::Arc;
 
-use crate::client_common::PromptBuilder;
+use crate::PromptBuilder;
 use crate::codex::Session;
 use crate::codex::TurnContext;
 use crate::error::Result as CodexResult;
@@ -9,6 +9,7 @@ use crate::protocol::ContextCompactedEvent;
 use crate::protocol::EventMsg;
 use crate::protocol::RolloutItem;
 use crate::protocol::TaskStartedEvent;
+use codex_apply_patch::APPLY_PATCH_TOOL_INSTRUCTIONS;
 use codex_protocol::models::ResponseItem;
 
 pub(crate) async fn run_inline_remote_auto_compact_task(
@@ -41,14 +42,14 @@ async fn run_remote_compact_task_inner_impl(
     turn_context: &Arc<TurnContext>,
 ) -> CodexResult<()> {
     let mut history = sess.clone_history().await;
-    let mut prompt = PromptBuilder::new()
-        .wire_api(turn_context.client.get_provider().wire_api)
-        .with_input(history.get_history_for_prompt());
-    prompt.base_instructions_override = turn_context.base_instructions.clone();
+    let prompt = PromptBuilder::new()
+        .with_input(history.get_history_for_prompt())
+        .with_base_instructions_override_opt(turn_context.base_instructions.clone())
+        .build(&turn_context.client.get_model_family())?;
 
     let mut new_history = turn_context
         .client
-        .compact_conversation_history(&prompt)
+        .compact_conversation_history(&prompt.input, &prompt.instructions)
         .await?;
     // Required to keep `/undo` available after compaction
     let ghost_snapshots: Vec<ResponseItem> = history