diff --git a/codex-rs/app-server-protocol/schema/precomputed/app-server-exports-stable.json.zst b/codex-rs/app-server-protocol/schema/precomputed/app-server-exports-stable.json.zst
index e6b15206e8..41a1f67de2 100644
Binary files a/codex-rs/app-server-protocol/schema/precomputed/app-server-exports-stable.json.zst and b/codex-rs/app-server-protocol/schema/precomputed/app-server-exports-stable.json.zst differ
diff --git a/codex-rs/app-server/tests/suite/v2/thread_inject_items.rs b/codex-rs/app-server/tests/suite/v2/thread_inject_items.rs
index 5dfd005b8d..972cf70347 100644
--- a/codex-rs/app-server/tests/suite/v2/thread_inject_items.rs
+++ b/codex-rs/app-server/tests/suite/v2/thread_inject_items.rs
@@ -152,7 +152,7 @@ async fn thread_inject_items_adds_raw_response_items_to_thread_history(
                 envelope
                     .metadata
                     .as_ref()
-                    .map(|metadata| metadata.client_authored),
+                    .is_some_and(|metadata| metadata.client_authored),
             )),
             _ => None,
         })
@@ -166,10 +166,10 @@ async fn thread_inject_items_adds_raw_response_items_to_thread_history(
     assert_eq!(
         persisted_injected_items,
         vec![
-            (injected_item.clone(), None),
-            (injected_developer_item.clone(), Some(true)),
-            (marker_shaped_developer_item.clone(), Some(true)),
-            (named_tool_item, None),
+            (injected_item.clone(), false),
+            (injected_developer_item.clone(), true),
+            (marker_shaped_developer_item.clone(), true),
+            (named_tool_item, false),
         ]
     );
 
diff --git a/codex-rs/core/src/context_manager/history.rs b/codex-rs/core/src/context_manager/history.rs
index ee077e315a..799b9eb6a7 100644
--- a/codex-rs/core/src/context_manager/history.rs
+++ b/codex-rs/core/src/context_manager/history.rs
@@ -57,6 +57,7 @@ use codex_utils_output_truncation::approx_bytes_for_tokens;
 use codex_utils_output_truncation::approx_token_count;
 use codex_utils_output_truncation::approx_tokens_from_byte_count_i64;
 use codex_utils_output_truncation::truncate_function_output_payload;
+use codex_utils_output_truncation::with_serialization_allowance;
 use std::num::NonZeroUsize;
 use std::ops::Deref;
 use std::sync::Arc;
@@ -366,9 +367,9 @@ impl ContextManager {
             {
                 // The override already includes the tool's serialization allowance.
                 let policy = metadata
-                    .and_then(|metadata| metadata.fallback_token_limit_override)
+                    .and_then(|metadata| metadata.history_truncation_token_limit)
                     .map(TruncationPolicy::Tokens)
-                    .unwrap_or(policy * 1.2);
+                    .unwrap_or_else(|| with_serialization_allowance(policy));
                 truncate_function_output_payload(output, policy, estimate_audio_token_count);
             }
             if let Some(review_history) = &mut self.review_history
diff --git a/codex-rs/core/src/context_manager/history_tests.rs b/codex-rs/core/src/context_manager/history_tests.rs
index bfbe5d162b..fb43d1f1b5 100644
--- a/codex-rs/core/src/context_manager/history_tests.rs
+++ b/codex-rs/core/src/context_manager/history_tests.rs
@@ -659,7 +659,7 @@ fn annotated_history_apis_preserve_envelopes() {
 #[test_case(Some(100), 85, 100, true; "saved limit has no additional allowance")]
 #[test_case(Some(30_000), 20_000, 30_000, false; "large explicit budget")]
 fn record_annotated_items_preserves_metadata_while_processing_item(
-    fallback_token_limit_override: Option<usize>,
+    history_truncation_token_limit: Option<usize>,
     repeat_count: usize,
     expected_token_limit: usize,
     expected_truncation: bool,
@@ -677,7 +677,7 @@ fn record_annotated_items_preserves_metadata_while_processing_item(
             internal_chat_message_metadata_passthrough: None,
         },
         metadata: Some(CodexHarnessMetadata {
-            fallback_token_limit_override,
+            history_truncation_token_limit,
             ..Default::default()
         }),
     };
diff --git a/codex-rs/core/src/session/mod.rs b/codex-rs/core/src/session/mod.rs
index 15d24f4129..bdd8f24c64 100644
--- a/codex-rs/core/src/session/mod.rs
+++ b/codex-rs/core/src/session/mod.rs
@@ -172,6 +172,7 @@ use codex_thread_store::ThreadPersistenceMetadata;
 use codex_thread_store::ThreadStore;
 use codex_utils_audio::prepare_response_items as prepare_audio_response_items;
 use codex_utils_git_discovery::GitRootDiscovery;
+use codex_utils_output_truncation::with_serialization_allowance;
 use codex_utils_path_uri::PathUri;
 use futures::future::BoxFuture;
 use futures::future::Shared;
@@ -3429,9 +3430,25 @@ impl Session {
         &self,
         turn_context: &TurnContext,
         model_info: &ModelInfo,
-        items: Vec<ResponseItemEnvelope>,
+        mut items: Vec<ResponseItemEnvelope>,
         image_preparations: Vec<ImagePreparationMetadata>,
     ) {
+        // Save the originating history budget for replay.
+        // Preserve any existing tool-specific override.
+        let policy: codex_utils_output_truncation::TruncationPolicy =
+            model_info.truncation_policy.into();
+        for envelope in &mut items {
+            if matches!(
+                envelope.item,
+                ResponseItem::FunctionCallOutput { .. } | ResponseItem::CustomToolCallOutput { .. }
+            ) {
+                envelope
+                    .metadata
+                    .get_or_insert_default()
+                    .history_truncation_token_limit
+                    .get_or_insert_with(|| with_serialization_allowance(policy).token_budget());
+            }
+        }
         let response_items = items
             .iter()
             .map(|envelope| envelope.item.clone())
diff --git a/codex-rs/core/src/session/tests.rs b/codex-rs/core/src/session/tests.rs
index b36f4329a9..b0e69affe8 100644
--- a/codex-rs/core/src/session/tests.rs
+++ b/codex-rs/core/src/session/tests.rs
@@ -2532,10 +2532,17 @@ async fn record_inter_agent_communication_preserves_item_id_in_rollout_and_resum
     );
 }
 
-#[test_case::test_case(false; "plain items")]
-#[test_case::test_case(true; "annotated items")]
+#[test_case::test_case(false, None, false, 5; "plain items")]
+#[test_case::test_case(true, None, false, 5; "annotated items")]
+#[test_case::test_case(true, Some(100), false, 100; "existing tool budget")]
+#[test_case::test_case(false, None, true, 31; "byte policy includes allowance before conversion")]
 #[tokio::test]
-async fn annotated_history_uses_explicit_model_without_a_step(retain_metadata: bool) {
+async fn annotated_history_uses_explicit_model_without_a_step(
+    retain_metadata: bool,
+    saved_budget: Option<usize>,
+    byte_policy: bool,
+    expected_budget: usize,
+) {
     let (session, turn_context, _rx) = make_session_and_context_with_auth_and_config_and_rx(
         CodexAuth::from_api_key("Test API Key"),
         Vec::new(),
@@ -2548,8 +2555,11 @@ async fn annotated_history_uses_explicit_model_without_a_step(retain_metadata: b
     )
     .await;
     let mut model_info = turn_context.model_info().as_ref().clone();
-    model_info.truncation_policy =
-        codex_protocol::openai_models::TruncationPolicyConfig::tokens(/*limit*/ 4);
+    model_info.truncation_policy = if byte_policy {
+        codex_protocol::openai_models::TruncationPolicyConfig::bytes(/*limit*/ 101)
+    } else {
+        codex_protocol::openai_models::TruncationPolicyConfig::tokens(/*limit*/ 4)
+    };
     let text = "diagnostic line\n".repeat(50);
     let mut item = ResponseItem::FunctionCallOutput {
         id: Some(ResponseItemId::with_suffix("fco", "existing")),
@@ -2564,24 +2574,52 @@ async fn annotated_history_uses_explicit_model_without_a_step(retain_metadata: b
     };
     item.set_turn_id_if_missing("existing-turn");
     item.set_create_time_if_missing(123.into());
-    let mut expected = ResponseItemEnvelope {
+    let expected = ResponseItemEnvelope {
         item,
         metadata: retain_metadata.then_some(CodexHarnessMetadata {
             client_authored: true,
+            history_truncation_token_limit: saved_budget,
             ..Default::default()
         }),
     };
+    let mut expected = vec![
+        expected,
+        ResponseItemEnvelope {
+            item: ResponseItem::CustomToolCallOutput {
+                id: Some(ResponseItemId::with_suffix("ctco", "existing")),
+                call_id: "call-custom".to_string(),
+                name: None,
+                output: FunctionCallOutputPayload::from_text(text.clone()),
+                internal_chat_message_metadata_passthrough: None,
+            },
+            metadata: retain_metadata.then_some(CodexHarnessMetadata {
+                client_authored: true,
+                history_truncation_token_limit: saved_budget,
+                ..Default::default()
+            }),
+        },
+    ];
+    expected[1].item.set_turn_id_if_missing("existing-turn");
+    expected[1].item.set_create_time_if_missing(123.into());
     session
-        .record_annotated_conversation_items(&turn_context, &model_info, vec![expected.clone()])
+        .record_annotated_conversation_items(&turn_context, &model_info, expected.clone())
         .await;
-    let ResponseItem::FunctionCallOutput { output, .. } = &mut expected.item else {
-        unreachable!("fixture is a tool output");
-    };
-    output.body = FunctionCallOutputBody::Text(codex_utils_output_truncation::truncate_text(
-        &text,
-        codex_utils_output_truncation::TruncationPolicy::Tokens(4) * 1.2,
-    ));
-    assert_eq!(session.clone_history().await.annotated_items(), &[expected]);
+    for envelope in &mut expected {
+        envelope
+            .metadata
+            .get_or_insert_default()
+            .history_truncation_token_limit = Some(expected_budget);
+        let (ResponseItem::FunctionCallOutput { output, .. }
+        | ResponseItem::CustomToolCallOutput { output, .. }) = &mut envelope.item
+        else {
+            unreachable!("fixture is a tool output");
+        };
+        output.body = FunctionCallOutputBody::Text(codex_utils_output_truncation::truncate_text(
+            &text,
+            codex_utils_output_truncation::TruncationPolicy::Tokens(expected_budget),
+        ));
+    }
+    assert_eq!(session.clone_history().await.annotated_items(), &expected);
 }
 
 #[tokio::test]
diff --git a/codex-rs/core/src/tools/context.rs b/codex-rs/core/src/tools/context.rs
index 72f9481ff3..5e023fd4c0 100644
--- a/codex-rs/core/src/tools/context.rs
+++ b/codex-rs/core/src/tools/context.rs
@@ -21,6 +21,7 @@ use codex_utils_output_truncation::approx_token_count;
 use codex_utils_output_truncation::formatted_truncate_text;
 use codex_utils_output_truncation::truncate_function_output_payload;
 use codex_utils_output_truncation::truncate_text;
+use codex_utils_output_truncation::with_serialization_allowance;
 use serde::Serialize;
 use serde_json::Value as JsonValue;
 use std::num::NonZeroUsize;
@@ -124,7 +125,7 @@ impl ToolOutput for McpToolOutput {
     }
 
     fn fallback_token_limit_override(&self) -> Option<usize> {
-        Some((self.truncation_policy * 1.2).token_budget())
+        Some(with_serialization_allowance(self.truncation_policy).token_budget())
     }
 
     fn to_response_item(&self, call_id: &str, _payload: &ToolPayload) -> ResponseInputItem {
@@ -173,7 +174,7 @@ impl McpToolOutput {
         // History receives this budget in tokens. Code Mode keeps the raw result.
         truncate_function_output_payload(
             &mut payload,
-            self.truncation_policy * 1.2,
+            with_serialization_allowance(self.truncation_policy),
             estimate_audio_token_count,
         );
         payload
@@ -515,7 +516,7 @@ impl ExecCommandToolOutput {
 
     fn response_text(&self) -> String {
         let header = self.response_header();
-        let output_budget = (self.truncation_policy * 1.2)
+        let output_budget = with_serialization_allowance(self.truncation_policy)
             .byte_budget()
             .saturating_sub(header.len().saturating_add(/*rhs*/ 1));
         let mut policy = self.model_output_policy();
diff --git a/codex-rs/core/src/tools/handlers/mcp_resource.rs b/codex-rs/core/src/tools/handlers/mcp_resource.rs
index 5709d05375..d6577f30d6 100644
--- a/codex-rs/core/src/tools/handlers/mcp_resource.rs
+++ b/codex-rs/core/src/tools/handlers/mcp_resource.rs
@@ -13,6 +13,7 @@ use codex_protocol::mcp::CallToolResult;
 use codex_protocol::models::function_call_output_content_items_to_text;
 use codex_protocol::protocol::TruncationPolicy;
 use codex_utils_output_truncation::truncate_text;
+use codex_utils_output_truncation::with_serialization_allowance;
 use rmcp::model::ListResourceTemplatesResult;
 use rmcp::model::ListResourcesResult;
 use rmcp::model::PaginatedRequestParams;
@@ -362,7 +363,7 @@ where
     })?;
     // Match regular MCP tool outputs by bounding the copy persisted to the
     // rollout and injected into model context.
-    let content = truncate_text(&content, truncation_policy * 1.2);
+    let content = truncate_text(&content, with_serialization_allowance(truncation_policy));
 
     Ok(FunctionToolOutput::from_text(content, Some(true)))
 }
diff --git a/codex-rs/core/src/tools/registry.rs b/codex-rs/core/src/tools/registry.rs
index fb2ccbb36e..4e7d5a1a89 100644
--- a/codex-rs/core/src/tools/registry.rs
+++ b/codex-rs/core/src/tools/registry.rs
@@ -203,7 +203,7 @@ impl AnyToolResult {
             metadata: result
                 .fallback_token_limit_override()
                 .map(|limit| CodexHarnessMetadata {
-                    fallback_token_limit_override: Some(limit),
+                    history_truncation_token_limit: Some(limit),
                     ..Default::default()
                 }),
         }
diff --git a/codex-rs/core/src/tools/registry_tests.rs b/codex-rs/core/src/tools/registry_tests.rs
index 9808432f86..edc68df33b 100644
--- a/codex-rs/core/src/tools/registry_tests.rs
+++ b/codex-rs/core/src/tools/registry_tests.rs
@@ -629,7 +629,7 @@ fn post_tool_use_feedback_output_preserves_fallback_token_limit_override(
                 output: FunctionCallOutputPayload::from_text("hook feedback".to_string()),
             }),
             metadata: Some(CodexHarnessMetadata {
-                fallback_token_limit_override: Some(expected_token_limit),
+                history_truncation_token_limit: Some(expected_token_limit),
                 ..Default::default()
             }),
         }
diff --git a/codex-rs/core/tests/suite/step_settings.rs b/codex-rs/core/tests/suite/step_settings.rs
index 998b4173ba..76c6083673 100644
--- a/codex-rs/core/tests/suite/step_settings.rs
+++ b/codex-rs/core/tests/suite/step_settings.rs
@@ -3,6 +3,7 @@ use base64::Engine;
 use base64::engine::general_purpose::STANDARD as BASE64_STANDARD;
 use codex_config::McpServerConfig;
 use codex_core::CodexThread;
+use codex_core::ForkSnapshot;
 use codex_core::StartThreadOptions;
 use codex_core::TurnInputRequest;
 use codex_core::config::Config;
@@ -29,6 +30,7 @@ use codex_protocol::dynamic_tools::DynamicToolCallOutputContentItem;
 use codex_protocol::dynamic_tools::DynamicToolFunctionSpec;
 use codex_protocol::dynamic_tools::DynamicToolResponse;
 use codex_protocol::dynamic_tools::DynamicToolSpec;
+use codex_protocol::mcp::ClientMcpExtensions;
 use codex_protocol::models::PermissionProfile;
 use codex_protocol::models::ResponseItem;
 use codex_protocol::openai_models::ApplyPatchToolType;
@@ -83,6 +85,7 @@ use core_test_support::responses::ev_function_call_with_namespace;
 use core_test_support::responses::ev_response_created;
 use core_test_support::responses::mount_models_once;
 use core_test_support::responses::mount_response_sequence;
+use core_test_support::responses::mount_sse_once;
 use core_test_support::responses::mount_sse_sequence;
 use core_test_support::responses::sse;
 use core_test_support::responses::sse_completed;
@@ -286,7 +289,7 @@ fn request_turn_id(request: &ResponsesRequest) -> String {
 
 // Dynamic tools return the original payload, so handler truncation cannot hide a recorder bug.
 #[tokio::test(flavor = "multi_thread", worker_threads = 2)]
-async fn tool_result_history_keeps_originating_model_across_switch() -> Result<()> {
+async fn tool_result_history_keeps_originating_model_across_switch_and_replay() -> Result<()> {
     skip_if_no_network!(Ok(()));
     let server = start_mock_server().await;
     let mut test = step_settings_test()
@@ -438,10 +441,8 @@ async fn tool_result_history_keeps_originating_model_across_switch() -> Result<(
 
     // Persistence and raw notifications retain the prepared, untruncated payload in append order.
     test.codex.shutdown_and_wait().await?;
-    let history = codex_rollout::RolloutRecorder::get_rollout_history(
-        &test.codex.rollout_path().expect("rollout path"),
-    )
-    .await?;
+    let rollout_path = test.codex.rollout_path().expect("rollout path");
+    let history = codex_rollout::RolloutRecorder::get_rollout_history(&rollout_path).await?;
     let saved_outputs = history
         .get_rollout_items()
         .iter()
@@ -449,12 +450,183 @@ async fn tool_result_history_keeps_originating_model_across_switch() -> Result<(
             RolloutItem::ResponseItem(envelope)
                 if matches!(&envelope.item, ResponseItem::FunctionCallOutput { .. }) =>
             {
-                Some(serde_json::to_value(&envelope.item).expect("saved output"))
+                Some((
+                    serde_json::to_value(&envelope.item).expect("saved output"),
+                    envelope
+                        .metadata
+                        .as_ref()
+                        .and_then(|metadata| metadata.history_truncation_token_limit),
+                ))
             }
             _ => None,
         })
         .collect::<Vec<_>>();
-    assert_eq!(saved_outputs, raw_outputs);
+    assert_eq!(
+        saved_outputs,
+        raw_outputs
+            .into_iter()
+            .zip([Some(120), Some(480)])
+            .collect::<Vec<_>>()
+    );
+
+    // Resume and fork under both limits. A's output must not grow under B, and
+    // B's output must not shrink under A; image preparation and item IDs also survive.
+    for model in [MODEL_A, MODEL_B] {
+        let mut replay_config = test.config.clone();
+        replay_config.model = Some(model.to_string());
+        let resumed = test
+            .thread_manager
+            .resume_thread_from_rollout(
+                replay_config.clone(),
+                rollout_path.clone(),
+                codex_core::test_support::auth_manager_from_auth(CodexAuth::from_api_key("dummy")),
+                /*parent_trace*/ None,
+                ClientMcpExtensions::default(),
+            )
+            .await?
+            .thread;
+        let forked = test
+            .thread_manager
+            .fork_thread(
+                ForkSnapshot::Interrupted,
+                StartThreadOptions::new(replay_config),
+                rollout_path.clone(),
+            )
+            .await?
+            .thread;
+        for thread in [resumed, forked] {
+            let replay = mount_sse_once(&server, sse_completed("resp-replay")).await;
+            thread
+                .start_or_steer_turn(TurnInputRequest::user_input(vec![UserInput::Text {
+                    text: "review saved diagnostics".to_string(),
+                    text_elements: Vec::new(),
+                }]))
+                .await?;
+            wait_for_event(&thread, |event| matches!(event, EventMsg::TurnComplete(_))).await;
+            let request = replay.single_request();
+            assert_eq!(request.body_json()["model"], model);
+            for call_id in ["call-a", "call-b"] {
+                assert_eq!(
+                    request.function_call_output(call_id),
+                    requests[2].function_call_output(call_id)
+                );
+            }
+            thread.shutdown_and_wait().await?;
+        }
+    }
+    Ok(())
+}
+
+#[tokio::test(flavor = "multi_thread", worker_threads = 2)]
+async fn custom_tool_output_replay_preserves_originating_budget() -> Result<()> {
+    skip_if_no_network!(Ok(()));
+
+    let server = start_mock_server().await;
+    let test = step_settings_test()
+        .with_model(MODEL_B)
+        .with_config(|config| {
+            config.features.enable(Feature::CodeMode).unwrap();
+            for model in &mut config.model_catalog.as_mut().expect("models").models {
+                model.truncation_policy =
+                    TruncationPolicyConfig::tokens(if model.slug == MODEL_B { 400 } else { 100 });
+                model.tool_mode = Some(ToolMode::CodeModeOnly);
+                model.use_responses_lite = false;
+            }
+        })
+        .build_with_auto_env(&server)
+        .await?;
+    let responses = mount_sse_sequence(
+        &server,
+        vec![
+            sse(vec![
+                ev_response_created("resp-custom"),
+                ev_custom_tool_call(
+                    "call-custom",
+                    "exec",
+                    "text('diagnostic line\\n'.repeat(500));",
+                ),
+                ev_completed("resp-custom"),
+            ]),
+            sse_completed("resp-done"),
+        ],
+    )
+    .await;
+    test.submit_text_turn("collect diagnostics").await?;
+    let live_output = responses.requests()[1].custom_tool_call_output("call-custom");
+    let text = "diagnostic line\n".repeat(500);
+    let live_text = live_output["output"][1]["text"]
+        .as_str()
+        .expect("bounded custom output");
+    assert!(live_text.len() < text.len());
+    // A's smaller history budget would truncate this result again without the saved budget.
+    assert_ne!(
+        truncate_text(live_text, TruncationPolicy::Tokens(120)),
+        live_text
+    );
+
+    test.codex.shutdown_and_wait().await?;
+    let rollout_path = test.codex.rollout_path().expect("rollout path");
+    let history = codex_rollout::RolloutRecorder::get_rollout_history(&rollout_path).await?;
+    let saved = history
+        .get_rollout_items()
+        .iter()
+        .find_map(|item| match item {
+            RolloutItem::ResponseItem(envelope)
+                if matches!(&envelope.item, ResponseItem::CustomToolCallOutput { call_id, .. } if call_id == "call-custom") =>
+            {
+                Some(envelope)
+            }
+            _ => None,
+        })
+        .expect("saved custom output");
+    assert_eq!(
+        saved
+            .metadata
+            .as_ref()
+            .and_then(|metadata| metadata.history_truncation_token_limit),
+        Some(480)
+    );
+    assert_eq!(
+        serde_json::to_value(&saved.item)?["output"][1]["text"],
+        text
+    );
+
+    let mut replay_config = test.config.clone();
+    replay_config.model = Some(MODEL_A.to_string());
+    let resumed = test
+        .thread_manager
+        .resume_thread_from_rollout(
+            replay_config.clone(),
+            rollout_path.clone(),
+            codex_core::test_support::auth_manager_from_auth(CodexAuth::from_api_key("dummy")),
+            /*parent_trace*/ None,
+            ClientMcpExtensions::default(),
+        )
+        .await?
+        .thread;
+    let forked = test
+        .thread_manager
+        .fork_thread(
+            ForkSnapshot::Interrupted,
+            StartThreadOptions::new(replay_config),
+            rollout_path,
+        )
+        .await?
+        .thread;
+    for thread in [resumed, forked] {
+        let replay = mount_sse_once(&server, sse_completed("resp-replay")).await;
+        thread
+            .start_or_steer_turn(TurnInputRequest::user_input(vec![UserInput::Text {
+                text: "review saved diagnostics".to_string(),
+                text_elements: Vec::new(),
+            }]))
+            .await?;
+        wait_for_event(&thread, |event| matches!(event, EventMsg::TurnComplete(_))).await;
+        let request = replay.single_request();
+        assert_eq!(request.body_json()["model"], MODEL_A);
+        assert_eq!(request.custom_tool_call_output("call-custom"), live_output);
+        thread.shutdown_and_wait().await?;
+    }
     Ok(())
 }
 
diff --git a/codex-rs/history/src/lib.rs b/codex-rs/history/src/lib.rs
index 258c32291f..77e1b66277 100644
--- a/codex-rs/history/src/lib.rs
+++ b/codex-rs/history/src/lib.rs
@@ -50,10 +50,14 @@ pub struct CodexHarnessMetadata {
     #[serde(default)]
     pub client_authored: bool,
 
-    /// Overrides history's fallback truncation budget, including on resume.
-    /// Measured in tokens, with any tool-specific allowance already included.
-    #[serde(default, skip_serializing_if = "Option::is_none")]
-    pub fallback_token_limit_override: Option<usize>,
+    /// The originating history budget, including any tool-specific allowance.
+    /// Measured in tokens and reused when replaying persisted history.
+    #[serde(
+        default,
+        rename = "fallback_token_limit_override",
+        skip_serializing_if = "Option::is_none"
+    )]
+    pub history_truncation_token_limit: Option<usize>,
 
     /// Whether a response configuration update was created by the Codex harness itself.
     #[serde(default, skip_serializing_if = "std::ops::Not::not")]
diff --git a/codex-rs/history/src/tests.rs b/codex-rs/history/src/tests.rs
index c9a6597201..b4d922907a 100644
--- a/codex-rs/history/src/tests.rs
+++ b/codex-rs/history/src/tests.rs
@@ -78,7 +78,7 @@ fn response_item_envelope_stores_metadata_beside_rollout_payload() -> Result<()>
             item: response_item.clone(),
             metadata: Some(CodexHarnessMetadata {
                 client_authored: true,
-                fallback_token_limit_override: Some(20_000),
+                history_truncation_token_limit: Some(20_000),
                 inherited_user_message: true,
                 ..Default::default()
             }),
@@ -110,7 +110,7 @@ fn response_item_envelope_stores_metadata_beside_rollout_payload() -> Result<()>
         envelope.metadata,
         Some(CodexHarnessMetadata {
             client_authored: true,
-            fallback_token_limit_override: Some(20_000),
+            history_truncation_token_limit: Some(20_000),
             inherited_user_message: true,
             ..Default::default()
         })
diff --git a/codex-rs/tools/src/tool_call.rs b/codex-rs/tools/src/tool_call.rs
index 3ba56a9a4e..65336a789f 100644
--- a/codex-rs/tools/src/tool_call.rs
+++ b/codex-rs/tools/src/tool_call.rs
@@ -8,6 +8,7 @@ use codex_protocol::models::ResponseItem;
 use codex_protocol::protocol::EventMsg;
 use codex_utils_absolute_path::AbsolutePathBuf;
 use codex_utils_output_truncation::TruncationPolicy;
+use codex_utils_output_truncation::with_serialization_allowance;
 use std::future::Future;
 use std::marker::PhantomData;
 use std::pin::Pin;
@@ -147,9 +148,8 @@ impl ToolCall<'_> {
     /// Callers must include serialization overhead when fitting a response to this budget.
     pub fn response_byte_budget(&self, max_response_bytes: usize) -> usize {
         match &self.source {
-            ToolCallSource::Direct => {
-                max_response_bytes.min((self.truncation_policy * 1.2).byte_budget())
-            }
+            ToolCallSource::Direct => max_response_bytes
+                .min(with_serialization_allowance(self.truncation_policy).byte_budget()),
             ToolCallSource::CodeMode {
                 cell_id: _,
                 runtime_tool_call_id: _,
diff --git a/codex-rs/utils/output-truncation/src/lib.rs b/codex-rs/utils/output-truncation/src/lib.rs
index e612457396..f4f502ab28 100644
--- a/codex-rs/utils/output-truncation/src/lib.rs
+++ b/codex-rs/utils/output-truncation/src/lib.rs
@@ -11,6 +11,12 @@ use codex_utils_string::truncate_middle_with_token_budget;
 
 pub use codex_protocol::protocol::TruncationPolicy;
 
+/// Adds the existing 20% allowance for serialization and headers.
+/// Saved history budgets already include this allowance.
+pub fn with_serialization_allowance(policy: TruncationPolicy) -> TruncationPolicy {
+    policy * 1.2
+}
+
 pub fn formatted_truncate_text(content: &str, policy: TruncationPolicy) -> String {
     if content.len() <= policy.byte_budget() {
         return content.to_string();