fix(core): align tool output caps with model policy

code-rs/core/src/codex/streaming.rs

@@ -279,6 +279,8 @@ pub(super) async fn submission_loop(
                     .unwrap_or(false);
 
                 let old_model_family = updated_config.model_family.clone();
+                let old_tool_output_max_bytes = updated_config.tool_output_max_bytes;
+                let old_default_tool_output_max_bytes = old_model_family.tool_output_max_bytes();
 
                 updated_config.model = model.clone();
                 updated_config.model_explicit = model_explicit;
@@ -300,6 +302,9 @@ pub(super) async fn submission_loop(
                 updated_config.model_family = find_family_for_model(&updated_config.model)
                     .unwrap_or_else(|| derive_default_model_family(&updated_config.model));
 
+                let new_default_tool_output_max_bytes =
+                    updated_config.model_family.tool_output_max_bytes();
+
                 let old_context_window = old_model_family.context_window;
                 let new_context_window = updated_config.model_family.context_window;
                 let old_max_tokens = old_model_family.max_output_tokens;
@@ -323,6 +328,10 @@ pub(super) async fn submission_loop(
                     new_auto_compact,
                 );
 
+                if old_tool_output_max_bytes == old_default_tool_output_max_bytes {
+                    updated_config.tool_output_max_bytes = new_default_tool_output_max_bytes;
+                }
+
                 let skills_outcome =
                     updated_config.skills_enabled.then(|| load_skills(&updated_config));
                 if let Some(outcome) = &skills_outcome {

code-rs/core/src/config.rs

@@ -109,9 +109,6 @@ pub const GPT_5_CODEX_MEDIUM_MODEL: &str = "gpt-5.2-codex";
 /// files are *silently truncated* to this size so we do not take up too much of
 /// the context window.
 pub(crate) const PROJECT_DOC_MAX_BYTES: usize = 32 * 1024; // 32 KiB
-/// Maximum number of bytes of tool output included in a model request.
-pub(crate) const DEFAULT_TOOL_OUTPUT_MAX_BYTES: usize = 32 * 1024; // 32 KiB
-
 pub(crate) const CONFIG_TOML_FILE: &str = "config.toml";
 
 const DEFAULT_RESPONSES_ORIGINATOR_HEADER: &str = "code_cli_rs";
@@ -1081,6 +1078,7 @@ impl Config {
 
         let model_family =
             find_family_for_model(&model).unwrap_or_else(|| derive_default_model_family(&model));
+        let default_tool_output_max_bytes = model_family.tool_output_max_bytes();
 
         // Chat model reasoning effort (used when other flows follow the chat model).
         let preferred_model_reasoning_effort = config_profile
@@ -1373,7 +1371,7 @@ impl Config {
             project_doc_max_bytes: cfg.project_doc_max_bytes.unwrap_or(PROJECT_DOC_MAX_BYTES),
             tool_output_max_bytes: cfg
                 .tool_output_max_bytes
-                .unwrap_or(DEFAULT_TOOL_OUTPUT_MAX_BYTES),
+                .unwrap_or(default_tool_output_max_bytes),
             project_doc_fallback_filenames: cfg
                 .project_doc_fallback_filenames
                 .unwrap_or_default()
@@ -1664,7 +1662,10 @@ persistence = "none"
             ConfigOverrides::default(),
             code_home.path().to_path_buf(),
         )?;
-        assert_eq!(default_config.tool_output_max_bytes, DEFAULT_TOOL_OUTPUT_MAX_BYTES);
+        assert_eq!(
+            default_config.tool_output_max_bytes,
+            default_config.model_family.tool_output_max_bytes()
+        );
 
         let cfg = toml::from_str::<ConfigToml>("tool_output_max_bytes = 65536")
             .expect("TOML should deserialize");

code-rs/core/src/exec_command/session_manager.rs

@@ -23,6 +23,8 @@ use crate::exec_command::exec_command_session::ExecCommandSession;
 use crate::exec_command::session_id::SessionId;
 use code_protocol::models::FunctionCallOutputPayload;
 
+const EXEC_COMMAND_OUTPUT_MAX_BYTES: u64 = 1024 * 1024;
+
 #[derive(Debug, Default)]
 pub struct SessionManager {
     next_session_id: AtomicU32,
@@ -271,7 +273,10 @@ impl SessionManager {
 
         // Collect output until either timeout expires or process exits.
         // Enforce the byte cap incrementally so runaway commands cannot exhaust memory.
-        let cap_bytes_u64 = params.max_output_tokens.saturating_mul(4);
+        let cap_bytes_u64 = params
+            .max_output_tokens
+            .saturating_mul(4)
+            .min(EXEC_COMMAND_OUTPUT_MAX_BYTES);
         let cap_bytes: usize = cap_bytes_u64.min(usize::MAX as u64) as usize;
         let mut collector = TruncatingCollector::new(cap_bytes);
 
@@ -367,7 +372,9 @@ impl SessionManager {
             return Err("failed to write to stdin".to_string());
         }
 
-        let cap_bytes_u64 = max_output_tokens.saturating_mul(4);
+        let cap_bytes_u64 = max_output_tokens
+            .saturating_mul(4)
+            .min(EXEC_COMMAND_OUTPUT_MAX_BYTES);
         let cap_bytes: usize = cap_bytes_u64.min(usize::MAX as u64) as usize;
 
         // Collect output up to yield_time_ms, truncating to max_output_tokens bytes.

code-rs/core/src/model_family.rs

@@ -1,5 +1,6 @@
 use crate::config_types::ReasoningEffort;
 use crate::tool_apply_patch::ApplyPatchToolType;
+use code_protocol::protocol::TruncationPolicy;
 
 /// The `instructions` field in the payload sent to a model should always start
 /// with this content.
@@ -21,7 +22,7 @@ const CONTEXT_WINDOW_1M: u64 = 1_047_576;
 const MAX_OUTPUT_DEFAULT: u64 = 128_000;
 
 /// A model family is a group of models that share certain characteristics.
-#[derive(Debug, Clone, PartialEq, Eq, Hash)]
+#[derive(Debug, Clone, PartialEq, Eq)]
 pub struct ModelFamily {
     /// The full model slug used to derive this model family, e.g.
     /// "gpt-4.1-2025-04-14".
@@ -40,6 +41,9 @@ pub struct ModelFamily {
     /// Maximum number of output tokens that can be generated for the model.
     pub max_output_tokens: Option<u64>,
 
+    /// Truncation policy to apply when recording tool outputs in the model context.
+    pub truncation_policy: TruncationPolicy,
+
     /// Token threshold where we should automatically compact history.
     auto_compact_token_limit: Option<i64>,
 
@@ -80,6 +84,7 @@ macro_rules! model_family {
             needs_special_apply_patch_instructions: false,
             context_window: Some(CONTEXT_WINDOW_272K),
             max_output_tokens: Some(MAX_OUTPUT_DEFAULT),
+            truncation_policy: TruncationPolicy::Bytes(10_000),
             auto_compact_token_limit: None,
             supports_reasoning_summaries: false,
             default_reasoning_effort: None,
@@ -158,6 +163,7 @@ pub fn find_family_for_model(slug: &str) -> Option<ModelFamily> {
             apply_patch_tool_type: Some(ApplyPatchToolType::Freeform),
             supports_parallel_tool_calls: true,
             default_reasoning_effort: Some(ReasoningEffort::Medium),
+            truncation_policy: TruncationPolicy::Tokens(10_000),
         )
     } else if slug.starts_with("exp-codex") || slug.starts_with("codex-1p") {
         // Same defaults as gpt-5.2-codex.
@@ -167,6 +173,7 @@ pub fn find_family_for_model(slug: &str) -> Option<ModelFamily> {
             base_instructions: GPT_5_2_CODEX_INSTRUCTIONS.to_string(),
             apply_patch_tool_type: Some(ApplyPatchToolType::Freeform),
             supports_parallel_tool_calls: true,
+            truncation_policy: TruncationPolicy::Tokens(10_000),
         )
     } else if slug.starts_with("exp-") {
         model_family!(
@@ -175,6 +182,7 @@ pub fn find_family_for_model(slug: &str) -> Option<ModelFamily> {
             apply_patch_tool_type: Some(ApplyPatchToolType::Freeform),
             supports_parallel_tool_calls: true,
             default_reasoning_effort: Some(ReasoningEffort::Medium),
+            truncation_policy: TruncationPolicy::Bytes(10_000),
         )
     } else if slug.starts_with("gpt-5.1-codex-max") {
         model_family!(
@@ -184,6 +192,7 @@ pub fn find_family_for_model(slug: &str) -> Option<ModelFamily> {
             apply_patch_tool_type: Some(ApplyPatchToolType::Freeform),
             context_window: Some(CONTEXT_WINDOW_272K),
             max_output_tokens: Some(MAX_OUTPUT_DEFAULT),
+            truncation_policy: TruncationPolicy::Tokens(10_000),
         )
     } else if slug.starts_with("codex-")
         || slug.starts_with("gpt-5-codex")
@@ -196,6 +205,7 @@ pub fn find_family_for_model(slug: &str) -> Option<ModelFamily> {
             apply_patch_tool_type: Some(ApplyPatchToolType::Freeform),
             context_window: Some(CONTEXT_WINDOW_272K),
             max_output_tokens: Some(MAX_OUTPUT_DEFAULT),
+            truncation_policy: TruncationPolicy::Tokens(10_000),
         )
     } else if slug.starts_with("gpt-5.2-codex") {
         // Same defaults as gpt-5.1-codex-max.
@@ -207,6 +217,7 @@ pub fn find_family_for_model(slug: &str) -> Option<ModelFamily> {
             supports_parallel_tool_calls: true,
             context_window: Some(CONTEXT_WINDOW_272K),
             max_output_tokens: Some(MAX_OUTPUT_DEFAULT),
+            truncation_policy: TruncationPolicy::Tokens(10_000),
         )
     } else if slug.starts_with("bengalfox") {
         model_family!(
@@ -217,6 +228,7 @@ pub fn find_family_for_model(slug: &str) -> Option<ModelFamily> {
             supports_parallel_tool_calls: true,
             context_window: Some(CONTEXT_WINDOW_272K),
             max_output_tokens: Some(MAX_OUTPUT_DEFAULT),
+            truncation_policy: TruncationPolicy::Tokens(10_000),
         )
     } else if slug.starts_with("gpt-5.2") {
         model_family!(
@@ -228,6 +240,7 @@ pub fn find_family_for_model(slug: &str) -> Option<ModelFamily> {
             supports_parallel_tool_calls: true,
             context_window: Some(CONTEXT_WINDOW_272K),
             max_output_tokens: Some(MAX_OUTPUT_DEFAULT),
+            truncation_policy: TruncationPolicy::Bytes(10_000),
         )
     } else if slug.starts_with("boomslang") {
         model_family!(
@@ -239,6 +252,7 @@ pub fn find_family_for_model(slug: &str) -> Option<ModelFamily> {
             supports_parallel_tool_calls: true,
             context_window: Some(CONTEXT_WINDOW_272K),
             max_output_tokens: Some(MAX_OUTPUT_DEFAULT),
+            truncation_policy: TruncationPolicy::Bytes(10_000),
         )
     } else if slug.starts_with("gpt-5.1") {
         model_family!(
@@ -250,6 +264,7 @@ pub fn find_family_for_model(slug: &str) -> Option<ModelFamily> {
             supports_parallel_tool_calls: true,
             context_window: Some(CONTEXT_WINDOW_272K),
             max_output_tokens: Some(MAX_OUTPUT_DEFAULT),
+            truncation_policy: TruncationPolicy::Bytes(10_000),
         )
     } else if slug.starts_with("gpt-5") {
         model_family!(
@@ -259,6 +274,7 @@ pub fn find_family_for_model(slug: &str) -> Option<ModelFamily> {
             base_instructions: BASE_INSTRUCTIONS_WITH_APPLY_PATCH.to_string(),
             context_window: Some(CONTEXT_WINDOW_272K),
             max_output_tokens: Some(MAX_OUTPUT_DEFAULT),
+            truncation_policy: TruncationPolicy::Bytes(10_000),
         )
     } else {
         None
@@ -272,6 +288,7 @@ pub fn derive_default_model_family(model: &str) -> ModelFamily {
         needs_special_apply_patch_instructions: false,
         context_window: None,
         max_output_tokens: None,
+        truncation_policy: TruncationPolicy::Bytes(10_000),
         auto_compact_token_limit: None,
         supports_reasoning_summaries: false,
         default_reasoning_effort: None,
@@ -293,6 +310,17 @@ impl ModelFamily {
         self.auto_compact_token_limit = limit;
     }
 
+    pub fn tool_output_max_bytes(&self) -> usize {
+        match self.truncation_policy {
+            TruncationPolicy::Bytes(limit) => limit,
+            TruncationPolicy::Tokens(limit) => limit.saturating_mul(4),
+        }
+    }
+
+    pub fn set_truncation_policy(&mut self, policy: TruncationPolicy) {
+        self.truncation_policy = policy;
+    }
+
     const fn default_auto_compact_limit(context_window: u64) -> i64 {
         // Match upstream behaviour: 90% of the context window.
         ((context_window as i64) * 9) / 10

code-rs/core/src/remote_models/mod.rs

@@ -8,6 +8,7 @@ use code_protocol::openai_models::ApplyPatchToolType as ProtocolApplyPatchToolTy
 use code_protocol::openai_models::ModelInfo;
 use code_protocol::openai_models::ModelsResponse;
 use code_protocol::openai_models::ReasoningEffort as ProtocolReasoningEffort;
+use code_protocol::openai_models::TruncationMode as ProtocolTruncationMode;
 use reqwest::header;
 use reqwest::Method;
 use reqwest::Url;
@@ -352,6 +353,8 @@ pub fn apply_model_info_overrides(info: &ModelInfo, mut family: ModelFamily) ->
         family.set_auto_compact_token_limit(Some(limit));
     }
 
+    family.set_truncation_policy(map_truncation_policy(&info.truncation_policy));
+
     family.supports_reasoning_summaries = info.supports_reasoning_summaries;
     family.supports_parallel_tool_calls = info.supports_parallel_tool_calls;
     if let Some(effort) = info.default_reasoning_level {
@@ -380,6 +383,16 @@ fn map_reasoning_effort(effort: ProtocolReasoningEffort) -> crate::config_types:
     }
 }
 
+fn map_truncation_policy(
+    policy: &code_protocol::openai_models::TruncationPolicyConfig,
+) -> code_protocol::protocol::TruncationPolicy {
+    let limit = usize::try_from(policy.limit).unwrap_or(usize::MAX);
+    match policy.mode {
+        ProtocolTruncationMode::Bytes => code_protocol::protocol::TruncationPolicy::Bytes(limit),
+        ProtocolTruncationMode::Tokens => code_protocol::protocol::TruncationPolicy::Tokens(limit),
+    }
+}
+
 /// Convert the build's version triple into a whole semver string.
 fn format_client_version_to_whole() -> String {
     format!(