Skip to content

armgit / codex

Public Imported from GitHub
0

Lightweight coding agent that runs in your terminal

Rust

Gate experimental context by model capability at session startup (#43147)

pmccrary-oai6af345407d9c
diff --git a/codex-rs/app-server/tests/common/models_cache.rs b/codex-rs/app-server/tests/common/models_cache.rs
index cc69f9951a..1bd6f316d7 100644
--- a/codex-rs/app-server/tests/common/models_cache.rs
+++ b/codex-rs/app-server/tests/common/models_cache.rs
@@ -68,6 +68,7 @@ fn preset_to_info(preset: &ModelPreset, priority: i32) -> ModelInfo {
input_modalities: default_input_modalities(),
used_fallback_model_metadata: false,
supports_search_tool: false,
+ supports_experimental_context: false,
use_responses_lite: false,
guardian: None,
node_repl_auto_review_required: false,
diff --git a/codex-rs/codex-api/tests/models_integration.rs b/codex-rs/codex-api/tests/models_integration.rs
index 288dd123e4..eb3ae763e7 100644
--- a/codex-rs/codex-api/tests/models_integration.rs
+++ b/codex-rs/codex-api/tests/models_integration.rs
@@ -101,6 +101,7 @@ async fn models_client_hits_models_endpoint() {
input_modalities: default_input_modalities(),
used_fallback_model_metadata: false,
supports_search_tool: false,
+ supports_experimental_context: false,
use_responses_lite: false,
guardian: None,
node_repl_auto_review_required: true,
diff --git a/codex-rs/core/src/config/mod.rs b/codex-rs/core/src/config/mod.rs
index eadfc56cea..ebda9778b1 100644
--- a/codex-rs/core/src/config/mod.rs
+++ b/codex-rs/core/src/config/mod.rs
@@ -172,6 +172,7 @@ mod requirements;
mod resolved_permission_profile;
#[cfg(test)]
mod schema;
+mod token_budget_startup;
pub use auth_keyring::bootstrap_auth_config;
pub use auth_keyring::resolve_bootstrap_auth_keyring_backend_kind;
pub use codex_agent_roles::AgentRoleConfig;
@@ -197,6 +198,7 @@ pub use permissions::compile_permission_profile;
pub(crate) use permissions::is_builtin_permission_profile_name;
pub use permissions::resolve_permission_profile;
pub(crate) use resolved_permission_profile::PermissionProfileState;
+pub use token_budget_startup::TokenBudgetStartupConfig;
const DEFAULT_IGNORE_LARGE_UNTRACKED_DIRS: i64 = 200;
const DEFAULT_IGNORE_LARGE_UNTRACKED_FILES: i64 = 10 * 1024 * 1024;
@@ -1053,6 +1055,8 @@ pub struct Config {
/// Context-window token budget configuration, when enabled.
pub token_budget: Option<TokenBudgetConfig>,
+ /// Runtime snapshot of configured token-budget preferences before startup activation.
+ pub token_budget_startup_config: Option<TokenBudgetStartupConfig>,
/// Shared token budget for the root thread and its sub-agents.
pub rollout_budget: Option<RolloutBudgetConfig>,
/// Current-time reminder and clock tool configuration, when enabled.
@@ -4318,6 +4322,7 @@ impl Config {
ghost_snapshot,
multi_agent_v2,
token_budget,
+ token_budget_startup_config: None,
rollout_budget,
current_time_reminder,
sleep_tool_mode,
diff --git a/codex-rs/core/src/config/token_budget_startup.rs b/codex-rs/core/src/config/token_budget_startup.rs
new file mode 100644
index 0000000000..02c9567cb4
--- /dev/null
+++ b/codex-rs/core/src/config/token_budget_startup.rs
@@ -0,0 +1,31 @@
+//! Keeps configured token-budget preferences separate from session startup activation.
+//! Fresh child sessions restore this snapshot before applying their starting model's defaults;
+//! history forks retain their parent's effective activation and the original snapshot.
+
+use super::Config;
+use super::ConstraintResult;
+use super::TokenBudgetConfig;
+use codex_features::Feature;
+
+/// Token-budget preferences before a session applies experimental or model-owned activation.
+#[derive(Debug, Clone, PartialEq)]
+pub struct TokenBudgetStartupConfig {
+ enabled: bool,
+ token_budget: Option<TokenBudgetConfig>,
+}
+
+impl Config {
+ /// Captures configured preferences, removing any inherited startup activation first.
+ pub(crate) fn prepare_token_budget_for_startup(&mut self) -> ConstraintResult<()> {
+ if let Some(snapshot) = self.token_budget_startup_config.as_ref() {
+ self.features
+ .set_enabled(Feature::TokenBudget, snapshot.enabled)?;
+ self.token_budget = snapshot.token_budget.clone();
+ }
+ self.token_budget_startup_config = Some(TokenBudgetStartupConfig {
+ enabled: self.features.enabled(Feature::TokenBudget),
+ token_budget: self.token_budget.clone(),
+ });
+ Ok(())
+ }
+}
diff --git a/codex-rs/core/src/session/mod.rs b/codex-rs/core/src/session/mod.rs
index 712fe63e23..24fd9cf629 100644
--- a/codex-rs/core/src/session/mod.rs
+++ b/codex-rs/core/src/session/mod.rs
@@ -687,11 +687,22 @@ impl Session {
.get_model_info(model.as_str(), &config.to_models_manager_config())
.await;
let auth = auth_manager.auth_cached();
- token_budget::apply_experimental_context(Arc::make_mut(&mut config), auth.as_ref())?;
- // Intentionally resolve `enabled` and `use_history_notes_extension` only at
- // thread startup. Both activation flags stay fixed for this thread runtime,
- // even if the selected model changes later.
- token_budget::apply_model_defaults(Arc::make_mut(&mut config), &model_info);
+ // Forked subagents keep their parent's activation with the copied history.
+ // Fresh children restore configured preferences before applying startup defaults.
+ let inherits_token_budget = matches!(&conversation_history, InitialHistory::Forked(_))
+ && config.token_budget_startup_config.is_some();
+ if !inherits_token_budget {
+ Arc::make_mut(&mut config)
+ .prepare_token_budget_for_startup()
+ .map_err(|err| CodexErr::InvalidRequest(err.to_string()))?;
+ // Resolve activation for this runtime, including when resuming saved history.
+ token_budget::apply_experimental_context(
+ Arc::make_mut(&mut config),
+ auth.as_ref(),
+ &model_info,
+ )?;
+ token_budget::apply_model_defaults(Arc::make_mut(&mut config), &model_info);
+ }
let configured_config = Arc::clone(&config);
let multi_agent_version = config.multi_agent_version_override().or_else(|| {
resolve_multi_agent_version(&conversation_history, inherited_multi_agent_version)
diff --git a/codex-rs/core/src/session/token_budget.rs b/codex-rs/core/src/session/token_budget.rs
index 0aeb66e4f5..d20a17544e 100644
--- a/codex-rs/core/src/session/token_budget.rs
+++ b/codex-rs/core/src/session/token_budget.rs
@@ -21,9 +21,11 @@ fn experimental_context_is_eligible(auth_mode: AuthMode, plan_type: Option<PlanT
pub(super) fn apply_experimental_context(
config: &mut Config,
auth: Option<&CodexAuth>,
+ starting_model: &ModelInfo,
) -> std::io::Result<()> {
let provider = &config.model_provider;
if !config.features.enabled(Feature::ContextManagement)
+ || !starting_model.supports_experimental_context
|| !provider.supports_codex_backend_routes()
|| !provider.requires_openai_auth
|| provider.env_key.is_some()
diff --git a/codex-rs/core/src/tools/handlers/multi_agents_common.rs b/codex-rs/core/src/tools/handlers/multi_agents_common.rs
index 77442ad43b..9cbe72df88 100644
--- a/codex-rs/core/src/tools/handlers/multi_agents_common.rs
+++ b/codex-rs/core/src/tools/handlers/multi_agents_common.rs
@@ -195,6 +195,9 @@ pub(crate) fn build_agent_resume_config(turn: &TurnContext) -> Result<Config, Fu
fn build_agent_shared_config(turn: &TurnContext) -> Result<Config, FunctionCallError> {
let base_config = turn.config.clone();
let mut config = (*base_config).clone();
+ // Preserve activation for history forks without freezing the parent's model-owned prompts.
+ // Fresh child startup restores configured preferences from the retained snapshot.
+ config.token_budget = turn.configured_token_budget.clone();
config.model = Some(turn.model_info().slug.clone());
config.model_provider = turn.provider.info().clone();
config.model_reasoning_effort = turn
diff --git a/codex-rs/core/src/tools/handlers/multi_agents_tests.rs b/codex-rs/core/src/tools/handlers/multi_agents_tests.rs
index 45d4b316d7..f9c7b48924 100644
--- a/codex-rs/core/src/tools/handlers/multi_agents_tests.rs
+++ b/codex-rs/core/src/tools/handlers/multi_agents_tests.rs
@@ -4371,8 +4371,10 @@ async fn tool_handlers_cascade_close_and_resume_and_keep_explicitly_closed_subtr
assert_eq!(shutdown_report.timed_out, Vec::<ThreadId>::new());
}
+#[test_case::test_case(false; "inactive_parent")]
+#[test_case::test_case(true; "active_parent")]
#[tokio::test]
-async fn build_agent_spawn_config_uses_turn_context_values() {
+async fn build_agent_spawn_config_uses_turn_context_values(parent_enabled: bool) {
fn pick_allowed_sandbox_policy(
permissions: &crate::config::Permissions,
base: SandboxPolicy,
@@ -4443,8 +4445,26 @@ async fn build_agent_spawn_config_uses_turn_context_values() {
.set(AskForApproval::OnRequest)
.expect("approval policy set");
+ let parent_config = Arc::make_mut(&mut turn.config);
+ parent_config
+ .prepare_token_budget_for_startup()
+ .expect("capture configured token budget");
+ parent_config
+ .features
+ .set_enabled(Feature::TokenBudget, parent_enabled)
+ .expect("set parent experimental context");
+ parent_config
+ .token_budget
+ .get_or_insert_default()
+ .use_history_notes_extension = parent_enabled;
+ let mut expected = parent_config.clone();
+ turn.configured_token_budget = expected.token_budget.clone();
+ Arc::make_mut(&mut turn.config)
+ .token_budget
+ .get_or_insert_default()
+ .guidance_message = Some("Parent model's resolved guidance.".to_string());
+
let config = build_agent_spawn_config(&base_instructions, &turn).expect("spawn config");
- let mut expected = (*turn.config).clone();
expected.base_instructions_provenance = base_instructions.provenance.clone();
expected.base_instructions = Some(base_instructions.text);
expected.model = Some(turn.model_info().slug.clone());
diff --git a/codex-rs/core/tests/suite/auto_review.rs b/codex-rs/core/tests/suite/auto_review.rs
index b6fae525e5..75575a61bb 100644
--- a/codex-rs/core/tests/suite/auto_review.rs
+++ b/codex-rs/core/tests/suite/auto_review.rs
@@ -622,6 +622,7 @@ fn remote_model_with_auto_review_override(slug: &str, review_model: &str) -> Mod
input_modalities: default_input_modalities(),
used_fallback_model_metadata: false,
supports_search_tool: false,
+ supports_experimental_context: false,
use_responses_lite: false,
node_repl_auto_review_required: false,
node_repl_disabled: false,
diff --git a/codex-rs/core/tests/suite/model_switching.rs b/codex-rs/core/tests/suite/model_switching.rs
index 025d93ded9..8990d0496a 100644
--- a/codex-rs/core/tests/suite/model_switching.rs
+++ b/codex-rs/core/tests/suite/model_switching.rs
@@ -120,6 +120,7 @@ fn test_model_info(
input_modalities,
used_fallback_model_metadata: false,
supports_search_tool: false,
+ supports_experimental_context: false,
use_responses_lite: false,
guardian: None,
node_repl_auto_review_required: false,
@@ -1358,6 +1359,7 @@ async fn model_switch_to_smaller_model_updates_token_context_window() -> Result<
input_modalities: default_input_modalities(),
used_fallback_model_metadata: false,
supports_search_tool: false,
+ supports_experimental_context: false,
use_responses_lite: false,
guardian: None,
node_repl_auto_review_required: false,
diff --git a/codex-rs/core/tests/suite/models_cache_ttl.rs b/codex-rs/core/tests/suite/models_cache_ttl.rs
index 7dea480f5b..701ada2f8c 100644
--- a/codex-rs/core/tests/suite/models_cache_ttl.rs
+++ b/codex-rs/core/tests/suite/models_cache_ttl.rs
@@ -538,6 +538,7 @@ fn test_remote_model(slug: &str, priority: i32) -> ModelInfo {
input_modalities: default_input_modalities(),
used_fallback_model_metadata: false,
supports_search_tool: false,
+ supports_experimental_context: false,
use_responses_lite: false,
guardian: None,
node_repl_auto_review_required: false,
diff --git a/codex-rs/core/tests/suite/personality.rs b/codex-rs/core/tests/suite/personality.rs
index 3de3c34eec..eba1485915 100644
--- a/codex-rs/core/tests/suite/personality.rs
+++ b/codex-rs/core/tests/suite/personality.rs
@@ -672,6 +672,7 @@ async fn remote_model_friendly_personality_instructions_with_feature() -> anyhow
input_modalities: default_input_modalities(),
used_fallback_model_metadata: false,
supports_search_tool: false,
+ supports_experimental_context: false,
use_responses_lite: false,
guardian: None,
node_repl_auto_review_required: false,
@@ -803,6 +804,7 @@ async fn user_turn_personality_remote_model_template_includes_update_message() -
input_modalities: default_input_modalities(),
used_fallback_model_metadata: false,
supports_search_tool: false,
+ supports_experimental_context: false,
use_responses_lite: false,
guardian: None,
node_repl_auto_review_required: false,
diff --git a/codex-rs/core/tests/suite/remote_models.rs b/codex-rs/core/tests/suite/remote_models.rs
index 545ca6c7d0..ba6d217c6a 100644
--- a/codex-rs/core/tests/suite/remote_models.rs
+++ b/codex-rs/core/tests/suite/remote_models.rs
@@ -609,6 +609,7 @@ async fn remote_models_remote_model_uses_unified_exec() -> Result<()> {
input_modalities: default_input_modalities(),
used_fallback_model_metadata: false,
supports_search_tool: false,
+ supports_experimental_context: false,
use_responses_lite: false,
guardian: None,
node_repl_auto_review_required: false,
@@ -879,6 +880,7 @@ async fn remote_models_apply_legacy_instructions() -> Result<()> {
input_modalities: default_input_modalities(),
used_fallback_model_metadata: false,
supports_search_tool: false,
+ supports_experimental_context: false,
use_responses_lite: false,
guardian: None,
node_repl_auto_review_required: false,
@@ -1467,6 +1469,7 @@ fn test_remote_model_with_policy(
input_modalities: default_input_modalities(),
used_fallback_model_metadata: false,
supports_search_tool: false,
+ supports_experimental_context: false,
use_responses_lite: false,
guardian: None,
node_repl_auto_review_required: false,
diff --git a/codex-rs/core/tests/suite/rmcp_client.rs b/codex-rs/core/tests/suite/rmcp_client.rs
index 1c2412902d..ce4c1c66b7 100644
--- a/codex-rs/core/tests/suite/rmcp_client.rs
+++ b/codex-rs/core/tests/suite/rmcp_client.rs
@@ -2770,6 +2770,7 @@ async fn stdio_image_responses_are_sanitized_for_text_only_model() -> anyhow::Re
input_modalities: vec![InputModality::Text],
used_fallback_model_metadata: false,
supports_search_tool: false,
+ supports_experimental_context: false,
use_responses_lite: false,
guardian: None,
node_repl_auto_review_required: false,
diff --git a/codex-rs/core/tests/suite/spawn_agent_description.rs b/codex-rs/core/tests/suite/spawn_agent_description.rs
index 712e35bb36..5d451683b7 100644
--- a/codex-rs/core/tests/suite/spawn_agent_description.rs
+++ b/codex-rs/core/tests/suite/spawn_agent_description.rs
@@ -90,6 +90,7 @@ fn test_model_info(
input_modalities: default_input_modalities(),
used_fallback_model_metadata: false,
supports_search_tool: false,
+ supports_experimental_context: false,
use_responses_lite: false,
guardian: None,
node_repl_auto_review_required: false,
diff --git a/codex-rs/core/tests/suite/token_budget.rs b/codex-rs/core/tests/suite/token_budget.rs
index 0e259ed21c..3828adf59a 100644
--- a/codex-rs/core/tests/suite/token_budget.rs
+++ b/codex-rs/core/tests/suite/token_budget.rs
@@ -256,6 +256,9 @@ async fn token_budget_guidance_precedes_standalone_context_window(
let guidance_message = "Preserve important state before compaction.";
let backend_url = format!("{}/backend-api/codex", server.uri());
let test = test_codex()
+ .with_model_info_override("gpt-5.2", |model| {
+ model.supports_experimental_context = true;
+ })
.with_auth(CodexAuth::from_external_chatgpt_tokens(
"header.e30.signature",
"account-123",
@@ -304,15 +307,17 @@ async fn token_budget_guidance_precedes_standalone_context_window(
Ok(())
}
-#[test_case("OpenAI", "/backend-api/codex", None, true; "codex_backend")]
-#[test_case("Custom", "/backend-api/codex", None, false; "custom_provider")]
-#[test_case("OpenAI", "/v1", None, false; "non_codex_endpoint")]
-#[test_case("OpenAI", "/backend-api/codex", Some("test-provider-token"), false; "provider_credentials")]
+#[test_case("OpenAI", "/backend-api/codex", None, true, true; "codex_backend")]
+#[test_case("OpenAI", "/backend-api/codex", None, false, false; "unsupported_model")]
+#[test_case("Custom", "/backend-api/codex", None, true, false; "custom_provider")]
+#[test_case("OpenAI", "/v1", None, true, false; "non_codex_endpoint")]
+#[test_case("OpenAI", "/backend-api/codex", Some("test-provider-token"), true, false; "provider_credentials")]
#[tokio::test(flavor = "multi_thread", worker_threads = 2)]
-async fn experimental_context_requires_codex_backend(
+async fn experimental_context_requires_capable_model_and_codex_backend(
provider_name: &'static str,
base_path: &'static str,
bearer_token: Option<&'static str>,
+ supports_context: bool,
expected_enabled: bool,
) -> Result<()> {
skip_if_no_network!(Ok(()));
@@ -321,6 +326,9 @@ async fn experimental_context_requires_codex_backend(
let response = mount_sse_once(&server, sse_completed("resp-1")).await;
let base_url = format!("{}{base_path}", server.uri());
let test = test_codex()
+ .with_model_info_override("gpt-5.2", move |model| {
+ model.supports_experimental_context = supports_context;
+ })
.with_auth(CodexAuth::from_external_chatgpt_tokens(
"header.e30.signature",
"account-123",
@@ -446,8 +454,10 @@ async fn token_budget_explicit_default_template_overrides_model_defaults() -> Re
Ok(())
}
+#[test_case(Feature::TokenBudget; "explicit_activation")]
+#[test_case(Feature::ContextManagement; "experimental_activation")]
#[tokio::test(flavor = "multi_thread", worker_threads = 2)]
-async fn token_budget_defaults_follow_the_active_model() -> Result<()> {
+async fn token_budget_defaults_follow_the_active_model(activation: Feature) -> Result<()> {
skip_if_no_network!(Ok(()));
let server = start_mock_server().await;
@@ -456,8 +466,15 @@ async fn token_budget_defaults_follow_the_active_model() -> Result<()> {
vec![sse_completed("resp-1"), sse_completed("resp-2")],
)
.await;
+ let backend_url = format!("{}/backend-api/codex", server.uri());
let test = test_codex()
+ .with_auth(CodexAuth::from_external_chatgpt_tokens(
+ "header.e30.signature",
+ "account-123",
+ Some("plus"),
+ )?)
.with_model_info_override("gpt-5.2", |model_info| {
+ model_info.supports_experimental_context = true;
let mut defaults = model_token_budget_config();
defaults.guidance_message = "Use first-model context-window guidance.".to_string();
model_info
@@ -476,11 +493,12 @@ async fn token_budget_defaults_follow_the_active_model() -> Result<()> {
.token_budget = Some(defaults);
})
.with_model("gpt-5.2")
- .with_config(|config| {
+ .with_config(move |config| {
+ config.model_provider.base_url = Some(backend_url);
config.model_context_window = Some(CONFIGURED_CONTEXT_WINDOW);
config
.features
- .enable(Feature::TokenBudget)
+ .enable(activation)
.expect("test config should allow token budget");
})
.build_with_auto_env(&server)
@@ -508,6 +526,7 @@ async fn token_budget_defaults_follow_the_active_model() -> Result<()> {
let requests = responses.requests();
assert_eq!(requests.len(), 2);
+ assert_eq!(requests[1].body_json()["model"], "gpt-5.4");
assert!(requests[0].body_contains_text("Use first-model context-window guidance."));
assert!(
requests[1].body_contains_text("Use first-model context-window guidance."),
diff --git a/codex-rs/core/tests/suite/view_image.rs b/codex-rs/core/tests/suite/view_image.rs
index a919b19159..0256da4ec2 100644
--- a/codex-rs/core/tests/suite/view_image.rs
+++ b/codex-rs/core/tests/suite/view_image.rs
@@ -1622,6 +1622,7 @@ async fn view_image_tool_returns_unsupported_message_for_text_only_model() -> an
input_modalities: vec![InputModality::Text],
used_fallback_model_metadata: false,
supports_search_tool: false,
+ supports_experimental_context: false,
use_responses_lite: false,
guardian: None,
node_repl_auto_review_required: false,
diff --git a/codex-rs/models-manager/models.json b/codex-rs/models-manager/models.json
index 698da6fb7a..c9b4d6ce6e 100644
--- a/codex-rs/models-manager/models.json
+++ b/codex-rs/models-manager/models.json
@@ -2,6 +2,7 @@
"models": [
{
"slug": "gpt-6-astra",
+ "supports_experimental_context": true,
"prefer_websockets": true,
"support_verbosity": true,
"default_verbosity": "low",
diff --git a/codex-rs/models-manager/src/model_info.rs b/codex-rs/models-manager/src/model_info.rs
index fd9e2615c1..ca10069931 100644
--- a/codex-rs/models-manager/src/model_info.rs
+++ b/codex-rs/models-manager/src/model_info.rs
@@ -178,6 +178,7 @@ pub fn model_info_from_slug(slug: &str) -> ModelInfo {
input_modalities: default_input_modalities(),
used_fallback_model_metadata: true, // this is the fallback model metadata
supports_search_tool: false,
+ supports_experimental_context: false,
use_responses_lite: false,
guardian: None,
node_repl_auto_review_required: false,
diff --git a/codex-rs/protocol/src/openai_models.rs b/codex-rs/protocol/src/openai_models.rs
index f5336ddc5a..8c021b17c2 100644
--- a/codex-rs/protocol/src/openai_models.rs
+++ b/codex-rs/protocol/src/openai_models.rs
@@ -470,6 +470,9 @@ pub struct ModelInfo {
pub used_fallback_model_metadata: bool,
#[serde(default)]
pub supports_search_tool: bool,
+ /// Whether experimental context management may be activated at session startup.
+ #[serde(default)]
+ pub supports_experimental_context: bool,
#[serde(default)]
pub use_responses_lite: bool,
#[serde(default)]
@@ -1011,6 +1014,7 @@ mod tests {
input_modalities: default_input_modalities(),
used_fallback_model_metadata: false,
supports_search_tool: false,
+ supports_experimental_context: false,
use_responses_lite: false,
guardian: None,
node_repl_auto_review_required: false,
@@ -1772,6 +1776,7 @@ mod tests {
assert!(!model.supports_image_detail_original);
assert_eq!(model.web_search_tool_type, WebSearchToolType::Text);
assert!(!model.supports_search_tool);
+ assert!(!model.supports_experimental_context);
assert!(!model.use_responses_lite);
assert!(!model.node_repl_auto_review_required);
assert!(!model.node_repl_disabled);
diff --git a/codex-rs/thread-manager-sample/src/main.rs b/codex-rs/thread-manager-sample/src/main.rs
index d13b3379d5..e99c62c933 100644
--- a/codex-rs/thread-manager-sample/src/main.rs
+++ b/codex-rs/thread-manager-sample/src/main.rs
@@ -308,6 +308,7 @@ fn new_config(model: Option<String>, arg0_paths: Arg0DispatchPaths) -> anyhow::R
multi_agent_v2: MultiAgentV2Config::default(),
max_goal_token_budget: None,
token_budget: None,
+ token_budget_startup_config: None,
rollout_budget: None,
current_time_reminder: None,
sleep_tool_mode: Default::default(),