mirror of
https://github.com/openai/codex.git
synced 2026-11-20 22:40:37 +00:00
Add experimental context management activation (#42385)
## What changed - Add the under-development `features.context_management.experimental_mode` configuration and schema. - For eligible ChatGPT Plus, Pro, and Pro Lite sessions using the Codex backend, enable token-budget context, history notes, and the `new_context` tool when experimental context management is configured. - Keep the feature disabled for custom providers, provider credentials, non-Codex endpoints, and temporary structured threads. ## Testing - Cover subscription eligibility, backend restrictions, configuration resolution, token-budget guidance, and temporary-thread overrides. GitOrigin-RevId: 19402af19c4d69e09a7cdb3687ee7d233f6a562a
This commit is contained in:
@@ -699,6 +699,9 @@
|
||||
"content_item_kinds": {
|
||||
"type": "boolean"
|
||||
},
|
||||
"context_management": {
|
||||
"$ref": "#/definitions/FeatureToml_for_ContextManagementConfigToml"
|
||||
},
|
||||
"current_time_reminder": {
|
||||
"$ref": "#/definitions/FeatureToml_for_CurrentTimeReminderConfigToml"
|
||||
},
|
||||
@@ -1142,6 +1145,16 @@
|
||||
},
|
||||
"type": "object"
|
||||
},
|
||||
"ContextManagementConfigToml": {
|
||||
"additionalProperties": false,
|
||||
"properties": {
|
||||
"experimental_mode": {
|
||||
"description": "Enables experimental context management.",
|
||||
"type": "boolean"
|
||||
}
|
||||
},
|
||||
"type": "object"
|
||||
},
|
||||
"CurrentTimeReminderConfigToml": {
|
||||
"additionalProperties": false,
|
||||
"properties": {
|
||||
@@ -1255,6 +1268,16 @@
|
||||
}
|
||||
]
|
||||
},
|
||||
"FeatureToml_for_ContextManagementConfigToml": {
|
||||
"anyOf": [
|
||||
{
|
||||
"type": "boolean"
|
||||
},
|
||||
{
|
||||
"$ref": "#/definitions/ContextManagementConfigToml"
|
||||
}
|
||||
]
|
||||
},
|
||||
"FeatureToml_for_CurrentTimeReminderConfigToml": {
|
||||
"anyOf": [
|
||||
{
|
||||
@@ -6014,6 +6037,9 @@
|
||||
"content_item_kinds": {
|
||||
"type": "boolean"
|
||||
},
|
||||
"context_management": {
|
||||
"$ref": "#/definitions/FeatureToml_for_ContextManagementConfigToml"
|
||||
},
|
||||
"current_time_reminder": {
|
||||
"$ref": "#/definitions/FeatureToml_for_CurrentTimeReminderConfigToml"
|
||||
},
|
||||
|
||||
@@ -707,7 +707,11 @@ async fn load_config_resolves_token_budget_config() -> std::io::Result<()> {
|
||||
for (config_toml, expected) in [
|
||||
(
|
||||
"[features]\ntoken_budget = true\n",
|
||||
TokenBudgetConfig::default(),
|
||||
Some(TokenBudgetConfig::default()),
|
||||
),
|
||||
(
|
||||
"features.context_management.experimental_mode = true\n",
|
||||
None,
|
||||
),
|
||||
(
|
||||
r#"
|
||||
@@ -720,14 +724,14 @@ guidance_message = "Preserve important state before compaction."
|
||||
auto_compact_fallback_prompt = " Write notes immediately. "
|
||||
auto_compact_fallback_buffer_tokens = 8000
|
||||
"#,
|
||||
TokenBudgetConfig {
|
||||
Some(TokenBudgetConfig {
|
||||
use_history_notes_extension: true,
|
||||
reminder_threshold_tokens: Some(16_000),
|
||||
reminder_message_template: "Custom reminder: {n_remaining} tokens.".to_string(),
|
||||
guidance_message: Some("Preserve important state before compaction.".to_string()),
|
||||
auto_compact_fallback_prompt: Some("Write notes immediately.".to_string()),
|
||||
auto_compact_fallback_buffer_tokens: Some(8_000),
|
||||
},
|
||||
}),
|
||||
),
|
||||
] {
|
||||
let codex_home = tempdir()?;
|
||||
@@ -739,8 +743,14 @@ auto_compact_fallback_buffer_tokens = 8000
|
||||
)
|
||||
.await?;
|
||||
|
||||
assert!(config.features.enabled(Feature::TokenBudget));
|
||||
assert_eq!(config.token_budget, Some(expected));
|
||||
assert_eq!(
|
||||
config.features.enabled(Feature::TokenBudget),
|
||||
expected.is_some()
|
||||
);
|
||||
if expected.is_none() {
|
||||
assert!(config.features.enabled(Feature::ContextManagement));
|
||||
}
|
||||
assert_eq!(config.token_budget, expected);
|
||||
}
|
||||
Ok(())
|
||||
}
|
||||
|
||||
@@ -2766,7 +2766,7 @@ fn resolve_multi_agent_v2_config(config_toml: &ConfigToml) -> MultiAgentV2Config
|
||||
}
|
||||
}
|
||||
|
||||
fn resolve_token_budget_config(
|
||||
pub(crate) fn resolve_token_budget_config(
|
||||
config_toml: &ConfigToml,
|
||||
features: &ManagedFeatures,
|
||||
) -> std::io::Result<Option<TokenBudgetConfig>> {
|
||||
|
||||
@@ -682,6 +682,8 @@ impl Session {
|
||||
let model_info = models_manager
|
||||
.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.
|
||||
|
||||
@@ -2,10 +2,59 @@ use super::session::Session;
|
||||
use super::turn_context::TurnContext;
|
||||
use crate::config::Config;
|
||||
use crate::config::TokenBudgetConfig;
|
||||
use crate::config::resolve_token_budget_config;
|
||||
use crate::context::ContextualUserFragment;
|
||||
use codex_features::Feature;
|
||||
use codex_login::CodexAuth;
|
||||
use codex_protocol::account::PlanType;
|
||||
use codex_protocol::auth::AuthMode;
|
||||
use codex_protocol::openai_models::ModelInfo;
|
||||
|
||||
fn experimental_context_is_eligible(auth_mode: AuthMode, plan_type: Option<PlanType>) -> bool {
|
||||
auth_mode == AuthMode::Chatgpt
|
||||
&& matches!(
|
||||
plan_type,
|
||||
Some(PlanType::Plus | PlanType::Pro | PlanType::ProLite)
|
||||
)
|
||||
}
|
||||
|
||||
pub(super) fn apply_experimental_context(
|
||||
config: &mut Config,
|
||||
auth: Option<&CodexAuth>,
|
||||
) -> std::io::Result<()> {
|
||||
let provider = &config.model_provider;
|
||||
if !config.features.enabled(Feature::ContextManagement)
|
||||
|| !provider.supports_codex_backend_routes()
|
||||
|| !provider.requires_openai_auth
|
||||
|| provider.env_key.is_some()
|
||||
|| provider.experimental_bearer_token.is_some()
|
||||
|| provider.auth.is_some()
|
||||
|| provider.aws.is_some()
|
||||
|| !auth.is_some_and(|auth| {
|
||||
experimental_context_is_eligible(auth.auth_mode(), auth.account_plan_type())
|
||||
})
|
||||
|| config.features.enable(Feature::TokenBudget).is_err()
|
||||
|| !config.features.enabled(Feature::TokenBudget)
|
||||
{
|
||||
return Ok(());
|
||||
}
|
||||
|
||||
if config.token_budget.is_none() {
|
||||
let config_toml = config
|
||||
.config_layer_stack
|
||||
.effective_config()
|
||||
.try_into()
|
||||
.map_err(|err| std::io::Error::new(std::io::ErrorKind::InvalidData, err))?;
|
||||
config.token_budget = resolve_token_budget_config(&config_toml, &config.features)?;
|
||||
}
|
||||
|
||||
config
|
||||
.token_budget
|
||||
.get_or_insert_default()
|
||||
.use_history_notes_extension = true;
|
||||
Ok(())
|
||||
}
|
||||
|
||||
/// Detects explicit preferences before model defaults are applied to the turn config.
|
||||
pub(super) fn has_explicit_settings(config: &Config) -> bool {
|
||||
config
|
||||
@@ -163,3 +212,27 @@ pub(super) async fn maybe_record(
|
||||
sess.record_conversation_items(turn_context, std::slice::from_ref(&response_item))
|
||||
.await;
|
||||
}
|
||||
|
||||
#[cfg(test)]
|
||||
mod tests {
|
||||
use super::experimental_context_is_eligible;
|
||||
use codex_protocol::account::PlanType;
|
||||
use codex_protocol::auth::AuthMode;
|
||||
|
||||
#[test]
|
||||
fn experimental_context_requires_eligible_chatgpt_subscription() {
|
||||
for (auth_mode, plan_type, expected) in [
|
||||
(AuthMode::Chatgpt, PlanType::Plus, true),
|
||||
(AuthMode::Chatgpt, PlanType::Pro, true),
|
||||
(AuthMode::Chatgpt, PlanType::ProLite, true),
|
||||
(AuthMode::Chatgpt, PlanType::Free, false),
|
||||
(AuthMode::Chatgpt, PlanType::Enterprise, false),
|
||||
(AuthMode::ApiKey, PlanType::Pro, false),
|
||||
] {
|
||||
assert_eq!(
|
||||
experimental_context_is_eligible(auth_mode, Some(plan_type)),
|
||||
expected
|
||||
);
|
||||
}
|
||||
}
|
||||
}
|
||||
|
||||
@@ -6,6 +6,7 @@ use codex_core::config::Config;
|
||||
use codex_core::config::TokenBudgetConfig;
|
||||
use codex_extension_api::ExtensionRegistryBuilder;
|
||||
use codex_features::Feature;
|
||||
use codex_login::CodexAuth;
|
||||
use codex_model_provider_info::built_in_model_providers;
|
||||
use codex_protocol::config_types::AutoCompactTokenLimitScope;
|
||||
use codex_protocol::items::TurnItem;
|
||||
@@ -235,8 +236,12 @@ async fn token_budget_context_is_only_emitted_with_full_context() -> Result<()>
|
||||
Ok(())
|
||||
}
|
||||
|
||||
#[test_case("features.token_budget.enabled = true"; "token_budget")]
|
||||
#[test_case("features.context_management.experimental_mode = true"; "experimental_mode")]
|
||||
#[tokio::test(flavor = "multi_thread", worker_threads = 2)]
|
||||
async fn token_budget_guidance_precedes_standalone_context_window() -> Result<()> {
|
||||
async fn token_budget_guidance_precedes_standalone_context_window(
|
||||
activation: &'static str,
|
||||
) -> Result<()> {
|
||||
skip_if_no_network!(Ok(()));
|
||||
|
||||
let server = start_mock_server().await;
|
||||
@@ -249,18 +254,26 @@ async fn token_budget_guidance_precedes_standalone_context_window() -> Result<()
|
||||
)
|
||||
.await;
|
||||
let guidance_message = "Preserve important state before compaction.";
|
||||
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_pre_build_hook(move |home| {
|
||||
std::fs::write(
|
||||
home.join("config.toml"),
|
||||
format!(
|
||||
"{activation}\nfeatures.token_budget.guidance_message = {guidance_message:?}\n"
|
||||
),
|
||||
)
|
||||
.expect("write token-budget guidance configuration");
|
||||
})
|
||||
.with_config(move |config| {
|
||||
config.model_provider.base_url = Some(backend_url);
|
||||
config.update_plan_enabled = true;
|
||||
config.model_context_window = Some(CONFIGURED_CONTEXT_WINDOW);
|
||||
config.token_budget = Some(TokenBudgetConfig {
|
||||
guidance_message: Some(guidance_message.to_string()),
|
||||
..TokenBudgetConfig::default()
|
||||
});
|
||||
config
|
||||
.features
|
||||
.enable(Feature::TokenBudget)
|
||||
.expect("test config should allow token budget");
|
||||
})
|
||||
.build_with_auto_env(&server)
|
||||
.await?;
|
||||
@@ -291,6 +304,59 @@ async fn token_budget_guidance_precedes_standalone_context_window() -> Result<()
|
||||
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")]
|
||||
#[tokio::test(flavor = "multi_thread", worker_threads = 2)]
|
||||
async fn experimental_context_requires_codex_backend(
|
||||
provider_name: &'static str,
|
||||
base_path: &'static str,
|
||||
bearer_token: Option<&'static str>,
|
||||
expected_enabled: bool,
|
||||
) -> Result<()> {
|
||||
skip_if_no_network!(Ok(()));
|
||||
|
||||
let server = start_mock_server().await;
|
||||
let response = mount_sse_once(&server, sse_completed("resp-1")).await;
|
||||
let base_url = format!("{}{base_path}", server.uri());
|
||||
let test = test_codex()
|
||||
.with_auth(CodexAuth::from_external_chatgpt_tokens(
|
||||
"header.e30.signature",
|
||||
"account-123",
|
||||
Some("plus"),
|
||||
)?)
|
||||
.with_config(move |config| {
|
||||
config.model_provider.name = provider_name.to_string();
|
||||
config.model_provider.base_url = Some(base_url);
|
||||
config.model_provider.experimental_bearer_token = bearer_token.map(Into::into);
|
||||
config.model_context_window = Some(CONFIGURED_CONTEXT_WINDOW);
|
||||
config
|
||||
.features
|
||||
.enable(Feature::ContextManagement)
|
||||
.expect("test config should allow experimental context");
|
||||
})
|
||||
.build_with_auto_env(&server)
|
||||
.await?;
|
||||
|
||||
test.submit_turn("inspect experimental context activation")
|
||||
.await?;
|
||||
|
||||
let request = response.single_request();
|
||||
assert_eq!(
|
||||
tool_names(&request)
|
||||
.iter()
|
||||
.any(|name| name == "new_context"),
|
||||
expected_enabled
|
||||
);
|
||||
assert_eq!(
|
||||
!token_budget_contexts(&request).is_empty(),
|
||||
expected_enabled
|
||||
);
|
||||
|
||||
Ok(())
|
||||
}
|
||||
|
||||
#[tokio::test(flavor = "multi_thread", worker_threads = 2)]
|
||||
async fn token_budget_uses_model_message_defaults() -> Result<()> {
|
||||
skip_if_no_network!(Ok(()));
|
||||
|
||||
Reference in New Issue
Block a user