diff --git a/codex-rs/core/src/codex.rs b/codex-rs/core/src/codex.rs
index bf2aeb4161..7edc90337a 100644
--- a/codex-rs/core/src/codex.rs
+++ b/codex-rs/core/src/codex.rs
@@ -539,10 +539,7 @@ pub(crate) struct TurnContext {
 }
 impl TurnContext {
     pub(crate) fn model_context_window(&self) -> Option<i64> {
-        let effective_context_window_percent = self.model_info.effective_context_window_percent;
-        self.model_info.context_window.map(|context_window| {
-            context_window.saturating_mul(effective_context_window_percent) / 100
-        })
+        effective_model_context_window(&self.model_info)
     }
 
     pub(crate) fn resolve_path(&self, path: Option<String>) -> PathBuf {
@@ -684,6 +681,13 @@ impl SessionConfiguration {
     }
 }
 
+fn effective_model_context_window(model_info: &ModelInfo) -> Option<i64> {
+    let effective_context_window_percent = model_info.effective_context_window_percent;
+    model_info
+        .context_window
+        .map(|context_window| context_window.saturating_mul(effective_context_window_percent) / 100)
+}
+
 #[derive(Default, Clone)]
 pub(crate) struct SessionSettingsUpdate {
     pub(crate) cwd: Option<PathBuf>,
@@ -1044,10 +1048,14 @@ impl Session {
                 }
             };
         session_configuration.thread_name = thread_name.clone();
+        let model_info = models_manager
+            .get_model_info(session_configuration.collaboration_mode.model(), &config)
+            .await;
         let state = SessionState::new(
             session_configuration.clone(),
             conversation_id.to_string(),
             config.codex_home.clone(),
+            effective_model_context_window(&model_info),
         );
 
         let services = SessionServices {
@@ -2036,6 +2044,7 @@ impl Session {
         let mut history = ContextManager::new(
             Arc::new(self.conversation_id.to_string()),
             Arc::new(turn_context.config.codex_home.clone()),
+            turn_context.model_context_window(),
         );
         for item in rollout_items {
             match item {
@@ -5529,6 +5538,7 @@ mod tests {
             session_configuration,
             conversation_id.to_string(),
             config.codex_home.clone(),
+            effective_model_context_window(&model_info),
         );
         let initial = RateLimitSnapshot {
             primary: Some(RateLimitWindow {
@@ -5617,6 +5627,7 @@ mod tests {
             session_configuration,
             conversation_id.to_string(),
             config.codex_home.clone(),
+            effective_model_context_window(&model_info),
         );
         let initial = RateLimitSnapshot {
             primary: Some(RateLimitWindow {
@@ -5906,6 +5917,7 @@ mod tests {
             session_configuration.clone(),
             conversation_id.to_string(),
             config.codex_home.clone(),
+            effective_model_context_window(&model_info),
         );
         mark_state_initial_context_seeded(&mut state);
         let skills_manager = Arc::new(SkillsManager::new(config.codex_home.clone()));
@@ -6042,6 +6054,7 @@ mod tests {
             session_configuration.clone(),
             conversation_id.to_string(),
             config.codex_home.clone(),
+            effective_model_context_window(&model_info),
         );
         mark_state_initial_context_seeded(&mut state);
         let skills_manager = Arc::new(SkillsManager::new(config.codex_home.clone()));
@@ -6544,6 +6557,7 @@ mod tests {
         let mut live_history = ContextManager::new(
             Arc::new(session.conversation_id.to_string()),
             Arc::new(turn_context.config.codex_home.clone()),
+            turn_context.model_context_window(),
         );
 
         let initial_context = session.build_initial_context(turn_context).await;
diff --git a/codex-rs/core/src/context_manager/context_discoverable.rs b/codex-rs/core/src/context_manager/context_discoverable.rs
index 0a5f490cf1..8d3499f1f8 100644
--- a/codex-rs/core/src/context_manager/context_discoverable.rs
+++ b/codex-rs/core/src/context_manager/context_discoverable.rs
@@ -92,8 +92,7 @@ impl ContextDiscoverable for Message {
         };
 
         let estimated_tokens = i64::try_from(approx_token_count(&user_text)).unwrap_or(i64::MAX);
-        let offload_threshold = context_window_tokens.saturating_mul(95) / 100;
-        if current_usage_tokens.saturating_add(estimated_tokens) <= offload_threshold {
+        if current_usage_tokens.saturating_add(estimated_tokens) <= context_window_tokens {
             return ResponseItem::Message(self.clone());
         }
 
diff --git a/codex-rs/core/src/context_manager/history.rs b/codex-rs/core/src/context_manager/history.rs
index 93ad67e87a..bfb19c5d6d 100644
--- a/codex-rs/core/src/context_manager/history.rs
+++ b/codex-rs/core/src/context_manager/history.rs
@@ -29,10 +29,14 @@ pub(crate) struct ContextManager {
 }
 
 impl ContextManager {
-    pub(crate) fn new(thread_id: Arc<String>, codex_home: Arc<PathBuf>) -> Self {
+    pub(crate) fn new(
+        thread_id: Arc<String>,
+        codex_home: Arc<PathBuf>,
+        model_context_window: Option<i64>,
+    ) -> Self {
         Self {
             items: Vec::new(),
-            token_info: TokenUsageInfo::new_or_append(&None, &None, None),
+            token_info: TokenUsageInfo::new_or_append(&None, &None, model_context_window),
             thread_id,
             codex_home,
         }
diff --git a/codex-rs/core/src/context_manager/history_tests.rs b/codex-rs/core/src/context_manager/history_tests.rs
index 64174cd56c..5925f68175 100644
--- a/codex-rs/core/src/context_manager/history_tests.rs
+++ b/codex-rs/core/src/context_manager/history_tests.rs
@@ -39,6 +39,7 @@ fn create_history_with_items(items: Vec<ResponseItem>) -> ContextManager {
     let mut h = ContextManager::new(
         Arc::new("test-thread".to_string()),
         Arc::new(PathBuf::from("/tmp/test-codex-home")),
+        None,
     );
     // Use a generous but fixed token budget; tests only rely on truncation
     // behavior, not on a specific model's token limit.
@@ -139,6 +140,7 @@ fn filters_non_api_messages() {
     let mut h = ContextManager::new(
         Arc::new("test-thread".to_string()),
         Arc::new(PathBuf::from("/tmp/test-codex-home")),
+        None,
     );
     let policy = TruncationPolicy::Tokens(10_000);
     // System message is not API messages; Other is ignored.
@@ -622,6 +624,7 @@ fn record_items_truncates_function_call_output_content() {
     let mut history = ContextManager::new(
         Arc::new("test-thread".to_string()),
         Arc::new(PathBuf::from("/tmp/test-codex-home")),
+        None,
     );
     // Any reasonably small token budget works; the test only cares that
     // truncation happens and the marker is present.
@@ -664,6 +667,7 @@ fn record_items_truncates_custom_tool_call_output_content() {
     let mut history = ContextManager::new(
         Arc::new("test-thread".to_string()),
         Arc::new(PathBuf::from("/tmp/test-codex-home")),
+        None,
     );
     let policy = TruncationPolicy::Tokens(1_000);
     let line = "custom output that is very long\n";
@@ -700,6 +704,7 @@ fn record_items_respects_custom_token_limit() {
     let mut history = ContextManager::new(
         Arc::new("test-thread".to_string()),
         Arc::new(PathBuf::from("/tmp/test-codex-home")),
+        None,
     );
     let policy = TruncationPolicy::Tokens(10);
     let long_output = "tokenized content repeated many times ".repeat(200);
@@ -733,6 +738,7 @@ fn record_items_with_discoverability_offloads_large_user_message() {
     let mut history = ContextManager::new(
         Arc::new("thread-1".to_string()),
         Arc::new(codex_home.path().to_path_buf()),
+        None,
     );
     let policy = TruncationPolicy::Tokens(10_000);
     let large_text = "large user prompt ".repeat(200);
@@ -766,6 +772,7 @@ fn record_items_with_discoverability_keeps_small_user_message_in_history() {
     let mut history = ContextManager::new(
         Arc::new("thread-2".to_string()),
         Arc::new(codex_home.path().to_path_buf()),
+        None,
     );
     let policy = TruncationPolicy::Tokens(10_000);
     let item = user_input_text_msg("small prompt");
@@ -782,6 +789,7 @@ fn record_items_with_discoverability_does_not_change_non_user_messages() {
     let mut history = ContextManager::new(
         Arc::new("thread-3".to_string()),
         Arc::new(codex_home.path().to_path_buf()),
+        None,
     );
     let policy = TruncationPolicy::Tokens(10_000);
     let item = assistant_msg(&"assistant reply ".repeat(300));
@@ -798,6 +806,7 @@ fn record_items_with_discoverability_ignores_image_only_user_messages() {
     let mut history = ContextManager::new(
         Arc::new("thread-4".to_string()),
         Arc::new(codex_home.path().to_path_buf()),
+        None,
     );
     let policy = TruncationPolicy::Tokens(10_000);
     let item = ResponseItem::Message(codex_protocol::models::Message {
@@ -822,6 +831,7 @@ fn record_items_with_discoverability_offloads_once_above_ninety_five_percent_win
     let mut history = ContextManager::new(
         Arc::new("thread-95".to_string()),
         Arc::new(codex_home.path().to_path_buf()),
+        None,
     );
     let policy = TruncationPolicy::Tokens(10_000);
     let text = "near context threshold ".repeat(100);
diff --git a/codex-rs/core/src/state/session.rs b/codex-rs/core/src/state/session.rs
index d6c3ee4eaf..7d87312f4f 100644
--- a/codex-rs/core/src/state/session.rs
+++ b/codex-rs/core/src/state/session.rs
@@ -36,8 +36,13 @@ impl SessionState {
         session_configuration: SessionConfiguration,
         thread_id: String,
         codex_home: PathBuf,
+        model_context_window: Option<i64>,
     ) -> Self {
-        let history = ContextManager::new(Arc::new(thread_id), Arc::new(codex_home));
+        let history = ContextManager::new(
+            Arc::new(thread_id),
+            Arc::new(codex_home),
+            model_context_window,
+        );
         Self {
             session_configuration,
             history,