core-agent-ide/codex-rs/core/src/codex.rs
Owen Lin da991bdf3a
feat(otel): Centralize OTEL metric names and shared tag builders (#14117)
This cleans up a bunch of metric plumbing that had started to drift.

The main change is making `codex-otel` the canonical home for shared
metric definitions and metric tag helpers. I moved the `turn/thread`
metric names that were still duplicated into the OTEL metric registry,
added a shared `metrics::tags` module for common tag keys and session
tag construction, and updated `SessionTelemetry` to build its metadata
tags through that shared path.

On the codex-core side, TTFT/TTFM now use the shared metric-name
constants instead of local string definitions. I also switched the
obvious remaining turn/thread metric callsites over to the shared
constants, and added a small helper so TTFT/TTFM can attach an optional
sanitized client.name tag from TurnContext.

This should make follow-on telemetry work less ad hoc:
- one canonical place for metric names
- one canonical place for common metric tag keys/builders
- less duplication between `codex-core` and `codex-otel`
2026-03-09 12:46:42 -07:00

6990 lines
263 KiB
Rust
Raw Blame History

This file contains invisible Unicode characters

This file contains invisible Unicode characters that are indistinguishable to humans but may be processed differently by a computer. If you think that this is intentional, you can safely ignore this warning. Use the Escape button to reveal them.

This file contains Unicode characters that might be confused with other characters. If you think that this is intentional, you can safely ignore this warning. Use the Escape button to reveal them.

use std::collections::HashMap;
use std::collections::HashSet;
use std::fmt::Debug;
use std::path::Path;
use std::path::PathBuf;
use std::sync::Arc;
use std::sync::atomic::AtomicU64;
use crate::AuthManager;
use crate::CodexAuth;
use crate::SandboxState;
use crate::agent::AgentControl;
use crate::agent::AgentStatus;
use crate::agent::agent_status_from_event;
use crate::analytics_client::AnalyticsEventsClient;
use crate::analytics_client::AppInvocation;
use crate::analytics_client::InvocationType;
use crate::analytics_client::build_track_events_context;
use crate::apps::render_apps_section;
use crate::commit_attribution::commit_message_trailer_instruction;
use crate::compact;
use crate::compact::InitialContextInjection;
use crate::compact::run_inline_auto_compact_task;
use crate::compact::should_use_remote_compact_task;
use crate::compact_remote::run_inline_remote_auto_compact_task;
use crate::config::ManagedFeatures;
use crate::connectors;
use crate::exec_policy::ExecPolicyManager;
use crate::features::FEATURES;
use crate::features::Feature;
use crate::features::maybe_push_unstable_features_warning;
#[cfg(test)]
use crate::models_manager::collaboration_mode_presets::CollaborationModesConfig;
use crate::models_manager::manager::ModelsManager;
use crate::parse_command::parse_command;
use crate::parse_turn_item;
use crate::realtime_conversation::RealtimeConversationManager;
use crate::realtime_conversation::handle_audio as handle_realtime_conversation_audio;
use crate::realtime_conversation::handle_close as handle_realtime_conversation_close;
use crate::realtime_conversation::handle_start as handle_realtime_conversation_start;
use crate::realtime_conversation::handle_text as handle_realtime_conversation_text;
use crate::rollout::session_index;
use crate::stream_events_utils::HandleOutputCtx;
use crate::stream_events_utils::handle_non_tool_response_item;
use crate::stream_events_utils::handle_output_item_done;
use crate::stream_events_utils::last_assistant_message_from_item;
use crate::stream_events_utils::raw_assistant_output_text_from_item;
use crate::stream_events_utils::record_completed_response_item;
use crate::terminal;
use crate::truncate::TruncationPolicy;
use crate::turn_metadata::TurnMetadataState;
use crate::util::error_or_panic;
use crate::ws_version_from_features;
use async_channel::Receiver;
use async_channel::Sender;
use chrono::Local;
use chrono::Utc;
use codex_app_server_protocol::McpServerElicitationRequest;
use codex_app_server_protocol::McpServerElicitationRequestParams;
use codex_hooks::HookEvent;
use codex_hooks::HookEventAfterAgent;
use codex_hooks::HookPayload;
use codex_hooks::HookResult;
use codex_hooks::Hooks;
use codex_hooks::HooksConfig;
use codex_network_proxy::NetworkProxy;
use codex_network_proxy::NetworkProxyAuditMetadata;
use codex_network_proxy::normalize_host;
use codex_otel::current_span_trace_id;
use codex_otel::current_span_w3c_trace_context;
use codex_otel::set_parent_from_w3c_trace_context;
use codex_protocol::ThreadId;
use codex_protocol::approvals::ElicitationRequestEvent;
use codex_protocol::approvals::ExecApprovalRequestSkillMetadata;
use codex_protocol::approvals::ExecPolicyAmendment;
use codex_protocol::approvals::NetworkPolicyAmendment;
use codex_protocol::approvals::NetworkPolicyRuleAction;
use codex_protocol::config_types::ModeKind;
use codex_protocol::config_types::Settings;
use codex_protocol::config_types::WebSearchMode;
use codex_protocol::dynamic_tools::DynamicToolResponse;
use codex_protocol::dynamic_tools::DynamicToolSpec;
use codex_protocol::items::PlanItem;
use codex_protocol::items::TurnItem;
use codex_protocol::items::UserMessageItem;
use codex_protocol::mcp::CallToolResult;
use codex_protocol::models::BaseInstructions;
use codex_protocol::models::PermissionProfile;
use codex_protocol::models::format_allow_prefixes;
use codex_protocol::openai_models::ModelInfo;
use codex_protocol::permissions::FileSystemSandboxPolicy;
use codex_protocol::permissions::NetworkSandboxPolicy;
use codex_protocol::protocol::FileChange;
use codex_protocol::protocol::HasLegacyEvent;
use codex_protocol::protocol::ItemCompletedEvent;
use codex_protocol::protocol::ItemStartedEvent;
use codex_protocol::protocol::RawResponseItemEvent;
use codex_protocol::protocol::ReviewRequest;
use codex_protocol::protocol::RolloutItem;
use codex_protocol::protocol::SessionSource;
use codex_protocol::protocol::SubAgentSource;
use codex_protocol::protocol::TurnAbortReason;
use codex_protocol::protocol::TurnContextItem;
use codex_protocol::protocol::TurnContextNetworkItem;
use codex_protocol::protocol::TurnStartedEvent;
use codex_protocol::request_permissions::RequestPermissionsArgs;
use codex_protocol::request_permissions::RequestPermissionsEvent;
use codex_protocol::request_permissions::RequestPermissionsResponse;
use codex_protocol::request_user_input::RequestUserInputArgs;
use codex_protocol::request_user_input::RequestUserInputResponse;
use codex_rmcp_client::ElicitationResponse;
use codex_rmcp_client::OAuthCredentialsStoreMode;
use codex_utils_stream_parser::AssistantTextChunk;
use codex_utils_stream_parser::AssistantTextStreamParser;
use codex_utils_stream_parser::ProposedPlanSegment;
use codex_utils_stream_parser::extract_proposed_plan_text;
use codex_utils_stream_parser::strip_citations;
use futures::future::BoxFuture;
use futures::prelude::*;
use futures::stream::FuturesOrdered;
use rmcp::model::ListResourceTemplatesResult;
use rmcp::model::ListResourcesResult;
use rmcp::model::PaginatedRequestParams;
use rmcp::model::ReadResourceRequestParams;
use rmcp::model::ReadResourceResult;
use rmcp::model::RequestId;
use serde_json;
use serde_json::Value;
use tokio::sync::Mutex;
use tokio::sync::RwLock;
use tokio::sync::oneshot;
use tokio::sync::watch;
use tokio::task::JoinHandle;
use tokio_util::sync::CancellationToken;
use tracing::Instrument;
use tracing::debug;
use tracing::debug_span;
use tracing::error;
use tracing::field;
use tracing::info;
use tracing::info_span;
use tracing::instrument;
use tracing::trace;
use tracing::trace_span;
use tracing::warn;
use uuid::Uuid;
use crate::ModelProviderInfo;
use crate::client::ModelClient;
use crate::client::ModelClientSession;
use crate::client_common::Prompt;
use crate::client_common::ResponseEvent;
use crate::codex_thread::ThreadConfigSnapshot;
use crate::compact::collect_user_messages;
use crate::config::Config;
use crate::config::Constrained;
use crate::config::ConstraintResult;
use crate::config::GhostSnapshotConfig;
use crate::config::StartedNetworkProxy;
use crate::config::resolve_web_search_mode_for_turn;
use crate::config::types::McpServerConfig;
use crate::config::types::ShellEnvironmentPolicy;
use crate::context_manager::ContextManager;
use crate::context_manager::TotalTokenUsageBreakdown;
use crate::environment_context::EnvironmentContext;
use crate::error::CodexErr;
use crate::error::Result as CodexResult;
#[cfg(test)]
use crate::exec::StreamOutput;
use codex_config::CONFIG_TOML_FILE;
mod rollout_reconstruction;
#[cfg(test)]
mod rollout_reconstruction_tests;
#[derive(Debug, PartialEq)]
pub enum SteerInputError {
NoActiveTurn(Vec<UserInput>),
ExpectedTurnMismatch { expected: String, actual: String },
EmptyInput,
}
/// Notes from the previous real user turn.
///
/// Conceptually this is the same role that `previous_model` used to fill, but
/// it can carry other prior-turn settings that matter when constructing
/// sensible state-change diffs or full-context reinjection, such as model
/// switches or detecting a prior `realtime_active -> false` transition.
#[derive(Clone, Debug, PartialEq, Eq)]
pub(crate) struct PreviousTurnSettings {
pub(crate) model: String,
pub(crate) realtime_active: Option<bool>,
}
use crate::exec_policy::ExecPolicyUpdateError;
use crate::feedback_tags;
use crate::file_watcher::FileWatcher;
use crate::file_watcher::FileWatcherEvent;
use crate::git_info::get_git_repo_root;
use crate::instructions::UserInstructions;
use crate::mcp::CODEX_APPS_MCP_SERVER_NAME;
use crate::mcp::McpManager;
use crate::mcp::auth::compute_auth_statuses;
use crate::mcp::maybe_prompt_and_install_mcp_dependencies;
use crate::mcp::with_codex_apps_mcp;
use crate::mcp_connection_manager::McpConnectionManager;
use crate::mcp_connection_manager::codex_apps_tools_cache_key;
use crate::mcp_connection_manager::filter_codex_apps_mcp_tools_only;
use crate::mcp_connection_manager::filter_mcp_tools_by_name;
use crate::mcp_connection_manager::filter_non_codex_apps_mcp_tools_only;
use crate::memories;
use crate::mentions::build_connector_slug_counts;
use crate::mentions::build_skill_name_counts;
use crate::mentions::collect_explicit_app_ids;
use crate::mentions::collect_explicit_plugin_mentions;
use crate::mentions::collect_tool_mentions_from_messages;
use crate::network_policy_decision::execpolicy_network_rule_amendment;
use crate::plugins::PluginsManager;
use crate::plugins::build_plugin_injections;
use crate::project_doc::get_user_instructions;
use crate::protocol::AgentMessageContentDeltaEvent;
use crate::protocol::AgentReasoningSectionBreakEvent;
use crate::protocol::ApplyPatchApprovalRequestEvent;
use crate::protocol::AskForApproval;
use crate::protocol::BackgroundEventEvent;
use crate::protocol::CompactedItem;
use crate::protocol::DeprecationNoticeEvent;
use crate::protocol::ErrorEvent;
use crate::protocol::Event;
use crate::protocol::EventMsg;
use crate::protocol::ExecApprovalRequestEvent;
use crate::protocol::McpServerRefreshConfig;
use crate::protocol::ModelRerouteEvent;
use crate::protocol::ModelRerouteReason;
use crate::protocol::NetworkApprovalContext;
use crate::protocol::Op;
use crate::protocol::PlanDeltaEvent;
use crate::protocol::RateLimitSnapshot;
use crate::protocol::ReasoningContentDeltaEvent;
use crate::protocol::ReasoningRawContentDeltaEvent;
use crate::protocol::RequestUserInputEvent;
use crate::protocol::ReviewDecision;
use crate::protocol::SandboxPolicy;
use crate::protocol::SessionConfiguredEvent;
use crate::protocol::SessionNetworkProxyRuntime;
use crate::protocol::SkillDependencies as ProtocolSkillDependencies;
use crate::protocol::SkillErrorInfo;
use crate::protocol::SkillInterface as ProtocolSkillInterface;
use crate::protocol::SkillMetadata as ProtocolSkillMetadata;
use crate::protocol::SkillToolDependency as ProtocolSkillToolDependency;
use crate::protocol::StreamErrorEvent;
use crate::protocol::Submission;
use crate::protocol::TokenCountEvent;
use crate::protocol::TokenUsage;
use crate::protocol::TokenUsageInfo;
use crate::protocol::TurnDiffEvent;
use crate::protocol::WarningEvent;
use crate::rollout::RolloutRecorder;
use crate::rollout::RolloutRecorderParams;
use crate::rollout::map_session_init_error;
use crate::rollout::metadata;
use crate::rollout::policy::EventPersistenceMode;
use crate::shell;
use crate::shell_snapshot::ShellSnapshot;
use crate::skills::SkillError;
use crate::skills::SkillInjections;
use crate::skills::SkillLoadOutcome;
use crate::skills::SkillMetadata;
use crate::skills::SkillsManager;
use crate::skills::build_skill_injections;
use crate::skills::collect_env_var_dependencies;
use crate::skills::collect_explicit_skill_mentions;
use crate::skills::injection::ToolMentionKind;
use crate::skills::injection::app_id_from_path;
use crate::skills::injection::tool_kind_for_path;
use crate::skills::resolve_skill_dependencies_for_turn;
use crate::state::ActiveTurn;
use crate::state::SessionServices;
use crate::state::SessionState;
use crate::state_db;
use crate::tasks::GhostSnapshotTask;
use crate::tasks::RegularTask;
use crate::tasks::ReviewTask;
use crate::tasks::SessionTask;
use crate::tasks::SessionTaskContext;
use crate::tools::ToolRouter;
use crate::tools::context::SharedTurnDiffTracker;
use crate::tools::handlers::SEARCH_TOOL_BM25_TOOL_NAME;
use crate::tools::js_repl::JsReplHandle;
use crate::tools::js_repl::resolve_compatible_node;
use crate::tools::network_approval::NetworkApprovalService;
use crate::tools::network_approval::build_blocked_request_observer;
use crate::tools::network_approval::build_network_policy_decider;
use crate::tools::parallel::ToolCallRuntime;
use crate::tools::sandboxing::ApprovalStore;
use crate::tools::spec::ToolsConfig;
use crate::tools::spec::ToolsConfigParams;
use crate::turn_diff_tracker::TurnDiffTracker;
use crate::turn_timing::TurnTimingState;
use crate::turn_timing::record_turn_ttfm_metric;
use crate::turn_timing::record_turn_ttft_metric;
use crate::unified_exec::UnifiedExecProcessManager;
use crate::util::backoff;
use crate::windows_sandbox::WindowsSandboxLevelExt;
use codex_async_utils::OrCancelExt;
use codex_otel::SessionTelemetry;
use codex_otel::TelemetryAuthMode;
use codex_otel::metrics::names::THREAD_STARTED_METRIC;
use codex_protocol::config_types::CollaborationMode;
use codex_protocol::config_types::Personality;
use codex_protocol::config_types::ReasoningSummary as ReasoningSummaryConfig;
use codex_protocol::config_types::ServiceTier;
use codex_protocol::config_types::WindowsSandboxLevel;
use codex_protocol::models::ContentItem;
use codex_protocol::models::DeveloperInstructions;
use codex_protocol::models::ResponseInputItem;
use codex_protocol::models::ResponseItem;
use codex_protocol::openai_models::ReasoningEffort as ReasoningEffortConfig;
use codex_protocol::protocol::CodexErrorInfo;
use codex_protocol::protocol::InitialHistory;
use codex_protocol::user_input::UserInput;
use codex_utils_absolute_path::AbsolutePathBuf;
use codex_utils_readiness::Readiness;
use codex_utils_readiness::ReadinessFlag;
/// The high-level interface to the Codex system.
/// It operates as a queue pair where you send submissions and receive events.
pub struct Codex {
pub(crate) tx_sub: Sender<Submission>,
pub(crate) rx_event: Receiver<Event>,
// Last known status of the agent.
pub(crate) agent_status: watch::Receiver<AgentStatus>,
pub(crate) session: Arc<Session>,
}
/// Wrapper returned by [`Codex::spawn`] containing the spawned [`Codex`],
/// the submission id for the initial `ConfigureSession` request and the
/// unique session id.
pub struct CodexSpawnOk {
pub codex: Codex,
pub thread_id: ThreadId,
#[deprecated(note = "use thread_id")]
pub conversation_id: ThreadId,
}
pub(crate) const INITIAL_SUBMIT_ID: &str = "";
pub(crate) const SUBMISSION_CHANNEL_CAPACITY: usize = 512;
const CYBER_VERIFY_URL: &str = "https://chatgpt.com/cyber";
const CYBER_SAFETY_URL: &str = "https://developers.openai.com/codex/concepts/cyber-safety";
impl Codex {
/// Spawn a new [`Codex`] and initialize the session.
#[allow(clippy::too_many_arguments)]
pub(crate) async fn spawn(
mut config: Config,
auth_manager: Arc<AuthManager>,
models_manager: Arc<ModelsManager>,
skills_manager: Arc<SkillsManager>,
plugins_manager: Arc<PluginsManager>,
mcp_manager: Arc<McpManager>,
file_watcher: Arc<FileWatcher>,
conversation_history: InitialHistory,
session_source: SessionSource,
agent_control: AgentControl,
dynamic_tools: Vec<DynamicToolSpec>,
persist_extended_history: bool,
metrics_service_name: Option<String>,
inherited_shell_snapshot: Option<Arc<ShellSnapshot>>,
) -> CodexResult<CodexSpawnOk> {
let (tx_sub, rx_sub) = async_channel::bounded(SUBMISSION_CHANNEL_CAPACITY);
let (tx_event, rx_event) = async_channel::unbounded();
let loaded_plugins = plugins_manager.plugins_for_config(&config);
let loaded_skills = skills_manager.skills_for_config(&config);
for err in &loaded_skills.errors {
error!(
"failed to load skill {}: {}",
err.path.display(),
err.message
);
}
if let SessionSource::SubAgent(SubAgentSource::ThreadSpawn { depth, .. }) = session_source
&& depth >= config.agent_max_depth
{
let _ = config.features.disable(Feature::Collab);
}
if config.features.enabled(Feature::JsRepl)
&& let Err(err) = resolve_compatible_node(config.js_repl_node_path.as_deref()).await
{
let _ = config.features.disable(Feature::JsRepl);
let _ = config.features.disable(Feature::JsReplToolsOnly);
let message = if config.features.enabled(Feature::JsRepl) {
format!(
"`js_repl` remains enabled because enterprise requirements pin it on, but the configured Node runtime is unavailable or incompatible. {err}"
)
} else {
format!(
"Disabled `js_repl` for this session because the configured Node runtime is unavailable or incompatible. {err}"
)
};
warn!("{message}");
config.startup_warnings.push(message);
}
let allowed_skills_for_implicit_invocation =
loaded_skills.allowed_skills_for_implicit_invocation();
let user_instructions = get_user_instructions(
&config,
Some(&allowed_skills_for_implicit_invocation),
Some(loaded_plugins.capability_summaries()),
)
.await;
let exec_policy = if crate::guardian::is_guardian_subagent_source(&session_source) {
// Guardian review should rely on the built-in shell safety checks,
// not on caller-provided exec-policy rules that could shape the
// reviewer or silently auto-approve commands.
ExecPolicyManager::default()
} else {
ExecPolicyManager::load(&config.config_layer_stack)
.await
.map_err(|err| CodexErr::Fatal(format!("failed to load rules: {err}")))?
};
let config = Arc::new(config);
let refresh_strategy = match session_source {
SessionSource::SubAgent(_) => crate::models_manager::manager::RefreshStrategy::Offline,
_ => crate::models_manager::manager::RefreshStrategy::OnlineIfUncached,
};
if config.model.is_none()
|| !matches!(
refresh_strategy,
crate::models_manager::manager::RefreshStrategy::Offline
)
{
let _ = models_manager.list_models(refresh_strategy).await;
}
let model = models_manager
.get_default_model(&config.model, refresh_strategy)
.await;
// Resolve base instructions for the session. Priority order:
// 1. config.base_instructions override
// 2. conversation history => session_meta.base_instructions
// 3. base_instructions for current model
let model_info = models_manager.get_model_info(model.as_str(), &config).await;
let base_instructions = config
.base_instructions
.clone()
.or_else(|| conversation_history.get_base_instructions().map(|s| s.text))
.unwrap_or_else(|| model_info.get_model_instructions(config.personality));
// Respect thread-start tools. When missing (resumed/forked threads), read from the db
// first, then fall back to rollout-file tools.
let persisted_tools = if dynamic_tools.is_empty() {
let thread_id = match &conversation_history {
InitialHistory::Resumed(resumed) => Some(resumed.conversation_id),
InitialHistory::Forked(_) => conversation_history.forked_from_id(),
InitialHistory::New => None,
};
match thread_id {
Some(thread_id) => {
let state_db_ctx = state_db::get_state_db(&config).await;
state_db::get_dynamic_tools(state_db_ctx.as_deref(), thread_id, "codex_spawn")
.await
}
None => None,
}
} else {
None
};
let dynamic_tools = if dynamic_tools.is_empty() {
persisted_tools
.or_else(|| conversation_history.get_dynamic_tools())
.unwrap_or_default()
} else {
dynamic_tools
};
// TODO (aibrahim): Consolidate config.model and config.model_reasoning_effort into config.collaboration_mode
// to avoid extracting these fields separately and constructing CollaborationMode here.
let collaboration_mode = CollaborationMode {
mode: ModeKind::Default,
settings: Settings {
model: model.clone(),
reasoning_effort: config.model_reasoning_effort,
developer_instructions: None,
},
};
let session_configuration = SessionConfiguration {
provider: config.model_provider.clone(),
collaboration_mode,
model_reasoning_summary: config.model_reasoning_summary,
service_tier: config.service_tier,
developer_instructions: config.developer_instructions.clone(),
user_instructions,
personality: config.personality,
base_instructions,
compact_prompt: config.compact_prompt.clone(),
approval_policy: config.permissions.approval_policy.clone(),
sandbox_policy: config.permissions.sandbox_policy.clone(),
file_system_sandbox_policy: config.permissions.file_system_sandbox_policy.clone(),
network_sandbox_policy: config.permissions.network_sandbox_policy,
windows_sandbox_level: WindowsSandboxLevel::from_config(&config),
cwd: config.cwd.clone(),
codex_home: config.codex_home.clone(),
thread_name: None,
original_config_do_not_use: Arc::clone(&config),
metrics_service_name,
app_server_client_name: None,
session_source,
dynamic_tools,
persist_extended_history,
inherited_shell_snapshot,
};
// Generate a unique ID for the lifetime of this Codex session.
let session_source_clone = session_configuration.session_source.clone();
let (agent_status_tx, agent_status_rx) = watch::channel(AgentStatus::PendingInit);
let session_init_span = info_span!("session_init");
let session = Session::new(
session_configuration,
config.clone(),
auth_manager.clone(),
models_manager.clone(),
exec_policy,
tx_event.clone(),
agent_status_tx.clone(),
conversation_history,
session_source_clone,
skills_manager,
plugins_manager,
mcp_manager.clone(),
file_watcher,
agent_control,
)
.instrument(session_init_span)
.await
.map_err(|e| {
error!("Failed to create session: {e:#}");
map_session_init_error(&e, &config.codex_home)
})?;
let thread_id = session.conversation_id;
// This task will run until Op::Shutdown is received.
let session_loop_span = info_span!("session_loop", thread_id = %thread_id);
tokio::spawn(
submission_loop(Arc::clone(&session), config, rx_sub).instrument(session_loop_span),
);
let codex = Codex {
tx_sub,
rx_event,
agent_status: agent_status_rx,
session,
};
#[allow(deprecated)]
Ok(CodexSpawnOk {
codex,
thread_id,
conversation_id: thread_id,
})
}
/// Submit the `op` wrapped in a `Submission` with a unique ID.
pub async fn submit(&self, op: Op) -> CodexResult<String> {
let id = Uuid::now_v7().to_string();
let sub = Submission {
id: id.clone(),
op,
trace: None,
};
self.submit_with_id(sub).await?;
Ok(id)
}
/// Use sparingly: prefer `submit()` so Codex is responsible for generating
/// unique IDs for each submission.
pub async fn submit_with_id(&self, mut sub: Submission) -> CodexResult<()> {
if sub.trace.is_none() {
sub.trace = current_span_w3c_trace_context();
}
self.tx_sub
.send(sub)
.await
.map_err(|_| CodexErr::InternalAgentDied)?;
Ok(())
}
pub async fn next_event(&self) -> CodexResult<Event> {
let event = self
.rx_event
.recv()
.await
.map_err(|_| CodexErr::InternalAgentDied)?;
Ok(event)
}
pub async fn steer_input(
&self,
input: Vec<UserInput>,
expected_turn_id: Option<&str>,
) -> Result<String, SteerInputError> {
self.session.steer_input(input, expected_turn_id).await
}
pub(crate) async fn set_app_server_client_name(
&self,
app_server_client_name: Option<String>,
) -> ConstraintResult<()> {
self.session
.update_settings(SessionSettingsUpdate {
app_server_client_name,
..Default::default()
})
.await
}
pub(crate) async fn agent_status(&self) -> AgentStatus {
self.agent_status.borrow().clone()
}
pub(crate) async fn thread_config_snapshot(&self) -> ThreadConfigSnapshot {
let state = self.session.state.lock().await;
state.session_configuration.thread_config_snapshot()
}
pub(crate) fn state_db(&self) -> Option<state_db::StateDbHandle> {
self.session.state_db()
}
pub(crate) fn enabled(&self, feature: Feature) -> bool {
self.session.enabled(feature)
}
}
/// Context for an initialized model agent
///
/// A session has at most 1 running task at a time, and can be interrupted by user input.
pub(crate) struct Session {
pub(crate) conversation_id: ThreadId,
tx_event: Sender<Event>,
agent_status: watch::Sender<AgentStatus>,
state: Mutex<SessionState>,
/// The set of enabled features should be invariant for the lifetime of the
/// session.
features: ManagedFeatures,
pending_mcp_server_refresh_config: Mutex<Option<McpServerRefreshConfig>>,
pub(crate) conversation: Arc<RealtimeConversationManager>,
pub(crate) active_turn: Mutex<Option<ActiveTurn>>,
pub(crate) services: SessionServices,
js_repl: Arc<JsReplHandle>,
next_internal_sub_id: AtomicU64,
}
#[derive(Clone, Debug)]
pub(crate) struct TurnSkillsContext {
pub(crate) outcome: Arc<SkillLoadOutcome>,
pub(crate) implicit_invocation_seen_skills: Arc<Mutex<HashSet<String>>>,
}
impl TurnSkillsContext {
pub(crate) fn new(outcome: Arc<SkillLoadOutcome>) -> Self {
Self {
outcome,
implicit_invocation_seen_skills: Arc::new(Mutex::new(HashSet::new())),
}
}
}
/// The context needed for a single turn of the thread.
#[derive(Debug)]
pub(crate) struct TurnContext {
pub(crate) sub_id: String,
pub(crate) trace_id: Option<String>,
pub(crate) realtime_active: bool,
pub(crate) config: Arc<Config>,
pub(crate) auth_manager: Option<Arc<AuthManager>>,
pub(crate) model_info: ModelInfo,
pub(crate) session_telemetry: SessionTelemetry,
pub(crate) provider: ModelProviderInfo,
pub(crate) reasoning_effort: Option<ReasoningEffortConfig>,
pub(crate) reasoning_summary: ReasoningSummaryConfig,
pub(crate) session_source: SessionSource,
/// The session's current working directory. All relative paths provided by
/// the model as well as sandbox policies are resolved against this path
/// instead of `std::env::current_dir()`.
pub(crate) cwd: PathBuf,
pub(crate) current_date: Option<String>,
pub(crate) timezone: Option<String>,
pub(crate) app_server_client_name: Option<String>,
pub(crate) developer_instructions: Option<String>,
pub(crate) compact_prompt: Option<String>,
pub(crate) user_instructions: Option<String>,
pub(crate) collaboration_mode: CollaborationMode,
pub(crate) personality: Option<Personality>,
pub(crate) approval_policy: Constrained<AskForApproval>,
pub(crate) sandbox_policy: Constrained<SandboxPolicy>,
pub(crate) file_system_sandbox_policy: FileSystemSandboxPolicy,
pub(crate) network_sandbox_policy: NetworkSandboxPolicy,
pub(crate) network: Option<NetworkProxy>,
pub(crate) windows_sandbox_level: WindowsSandboxLevel,
pub(crate) shell_environment_policy: ShellEnvironmentPolicy,
pub(crate) tools_config: ToolsConfig,
pub(crate) features: ManagedFeatures,
pub(crate) ghost_snapshot: GhostSnapshotConfig,
pub(crate) final_output_json_schema: Option<Value>,
pub(crate) codex_linux_sandbox_exe: Option<PathBuf>,
pub(crate) tool_call_gate: Arc<ReadinessFlag>,
pub(crate) truncation_policy: TruncationPolicy,
pub(crate) js_repl: Arc<JsReplHandle>,
pub(crate) dynamic_tools: Vec<DynamicToolSpec>,
pub(crate) turn_metadata_state: Arc<TurnMetadataState>,
pub(crate) turn_skills: TurnSkillsContext,
pub(crate) turn_timing_state: Arc<TurnTimingState>,
}
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
})
}
pub(crate) async fn with_model(&self, model: String, models_manager: &ModelsManager) -> Self {
let mut config = (*self.config).clone();
config.model = Some(model.clone());
let model_info = models_manager.get_model_info(model.as_str(), &config).await;
let truncation_policy = model_info.truncation_policy.into();
let supported_reasoning_levels = model_info
.supported_reasoning_levels
.iter()
.map(|preset| preset.effort)
.collect::<Vec<_>>();
let reasoning_effort = if let Some(current_reasoning_effort) = self.reasoning_effort {
if supported_reasoning_levels.contains(&current_reasoning_effort) {
Some(current_reasoning_effort)
} else {
supported_reasoning_levels
.get(supported_reasoning_levels.len().saturating_sub(1) / 2)
.copied()
.or(model_info.default_reasoning_level)
}
} else {
supported_reasoning_levels
.get(supported_reasoning_levels.len().saturating_sub(1) / 2)
.copied()
.or(model_info.default_reasoning_level)
};
config.model_reasoning_effort = reasoning_effort;
let collaboration_mode =
self.collaboration_mode
.with_updates(Some(model.clone()), Some(reasoning_effort), None);
let features = self.features.clone();
let tools_config = ToolsConfig::new(&ToolsConfigParams {
model_info: &model_info,
features: &features,
web_search_mode: self.tools_config.web_search_mode,
session_source: self.session_source.clone(),
})
.with_web_search_config(self.tools_config.web_search_config.clone())
.with_allow_login_shell(self.tools_config.allow_login_shell)
.with_agent_roles(config.agent_roles.clone());
Self {
sub_id: self.sub_id.clone(),
trace_id: self.trace_id.clone(),
realtime_active: self.realtime_active,
config: Arc::new(config),
auth_manager: self.auth_manager.clone(),
model_info: model_info.clone(),
session_telemetry: self
.session_telemetry
.clone()
.with_model(model.as_str(), model_info.slug.as_str()),
provider: self.provider.clone(),
reasoning_effort,
reasoning_summary: self.reasoning_summary,
session_source: self.session_source.clone(),
cwd: self.cwd.clone(),
current_date: self.current_date.clone(),
timezone: self.timezone.clone(),
app_server_client_name: self.app_server_client_name.clone(),
developer_instructions: self.developer_instructions.clone(),
compact_prompt: self.compact_prompt.clone(),
user_instructions: self.user_instructions.clone(),
collaboration_mode,
personality: self.personality,
approval_policy: self.approval_policy.clone(),
sandbox_policy: self.sandbox_policy.clone(),
file_system_sandbox_policy: self.file_system_sandbox_policy.clone(),
network_sandbox_policy: self.network_sandbox_policy,
network: self.network.clone(),
windows_sandbox_level: self.windows_sandbox_level,
shell_environment_policy: self.shell_environment_policy.clone(),
tools_config,
features,
ghost_snapshot: self.ghost_snapshot.clone(),
final_output_json_schema: self.final_output_json_schema.clone(),
codex_linux_sandbox_exe: self.codex_linux_sandbox_exe.clone(),
tool_call_gate: Arc::new(ReadinessFlag::new()),
truncation_policy,
js_repl: Arc::clone(&self.js_repl),
dynamic_tools: self.dynamic_tools.clone(),
turn_metadata_state: self.turn_metadata_state.clone(),
turn_skills: self.turn_skills.clone(),
turn_timing_state: Arc::clone(&self.turn_timing_state),
}
}
pub(crate) fn resolve_path(&self, path: Option<String>) -> PathBuf {
path.as_ref()
.map(PathBuf::from)
.map_or_else(|| self.cwd.clone(), |p| self.cwd.join(p))
}
pub(crate) fn compact_prompt(&self) -> &str {
self.compact_prompt
.as_deref()
.unwrap_or(compact::SUMMARIZATION_PROMPT)
}
pub(crate) fn to_turn_context_item(&self) -> TurnContextItem {
TurnContextItem {
turn_id: Some(self.sub_id.clone()),
trace_id: self.trace_id.clone(),
cwd: self.cwd.clone(),
current_date: self.current_date.clone(),
timezone: self.timezone.clone(),
approval_policy: self.approval_policy.value(),
sandbox_policy: self.sandbox_policy.get().clone(),
network: self.turn_context_network_item(),
model: self.model_info.slug.clone(),
personality: self.personality,
collaboration_mode: Some(self.collaboration_mode.clone()),
realtime_active: Some(self.realtime_active),
effort: self.reasoning_effort,
summary: self.reasoning_summary,
user_instructions: self.user_instructions.clone(),
developer_instructions: self.developer_instructions.clone(),
final_output_json_schema: self.final_output_json_schema.clone(),
truncation_policy: Some(self.truncation_policy.into()),
}
}
fn turn_context_network_item(&self) -> Option<TurnContextNetworkItem> {
let network = self
.config
.config_layer_stack
.requirements()
.network
.as_ref()?;
Some(TurnContextNetworkItem {
allowed_domains: network.allowed_domains.clone().unwrap_or_default(),
denied_domains: network.denied_domains.clone().unwrap_or_default(),
})
}
}
fn local_time_context() -> (String, String) {
match iana_time_zone::get_timezone() {
Ok(timezone) => (Local::now().format("%Y-%m-%d").to_string(), timezone),
Err(_) => (
Utc::now().format("%Y-%m-%d").to_string(),
"Etc/UTC".to_string(),
),
}
}
#[derive(Clone)]
pub(crate) struct SessionConfiguration {
/// Provider identifier ("openai", "openrouter", ...).
provider: ModelProviderInfo,
collaboration_mode: CollaborationMode,
model_reasoning_summary: Option<ReasoningSummaryConfig>,
service_tier: Option<ServiceTier>,
/// Developer instructions that supplement the base instructions.
developer_instructions: Option<String>,
/// Model instructions that are appended to the base instructions.
user_instructions: Option<String>,
/// Personality preference for the model.
personality: Option<Personality>,
/// Base instructions for the session.
base_instructions: String,
/// Compact prompt override.
compact_prompt: Option<String>,
/// When to escalate for approval for execution
approval_policy: Constrained<AskForApproval>,
/// How to sandbox commands executed in the system
sandbox_policy: Constrained<SandboxPolicy>,
file_system_sandbox_policy: FileSystemSandboxPolicy,
network_sandbox_policy: NetworkSandboxPolicy,
windows_sandbox_level: WindowsSandboxLevel,
/// Working directory that should be treated as the *root* of the
/// session. All relative paths supplied by the model as well as the
/// execution sandbox are resolved against this directory **instead**
/// of the process-wide current working directory. CLI front-ends are
/// expected to expand this to an absolute path before sending the
/// `ConfigureSession` operation so that the business-logic layer can
/// operate deterministically.
cwd: PathBuf,
/// Directory containing all Codex state for this session.
codex_home: PathBuf,
/// Optional user-facing name for the thread, updated during the session.
thread_name: Option<String>,
// TODO(pakrym): Remove config from here
original_config_do_not_use: Arc<Config>,
/// Optional service name tag for session metrics.
metrics_service_name: Option<String>,
app_server_client_name: Option<String>,
/// Source of the session (cli, vscode, exec, mcp, ...)
session_source: SessionSource,
dynamic_tools: Vec<DynamicToolSpec>,
persist_extended_history: bool,
inherited_shell_snapshot: Option<Arc<ShellSnapshot>>,
}
impl SessionConfiguration {
pub(crate) fn codex_home(&self) -> &PathBuf {
&self.codex_home
}
fn thread_config_snapshot(&self) -> ThreadConfigSnapshot {
ThreadConfigSnapshot {
model: self.collaboration_mode.model().to_string(),
model_provider_id: self.original_config_do_not_use.model_provider_id.clone(),
service_tier: self.service_tier,
approval_policy: self.approval_policy.value(),
sandbox_policy: self.sandbox_policy.get().clone(),
cwd: self.cwd.clone(),
ephemeral: self.original_config_do_not_use.ephemeral,
reasoning_effort: self.collaboration_mode.reasoning_effort(),
personality: self.personality,
session_source: self.session_source.clone(),
}
}
pub(crate) fn apply(&self, updates: &SessionSettingsUpdate) -> ConstraintResult<Self> {
let mut next_configuration = self.clone();
if let Some(collaboration_mode) = updates.collaboration_mode.clone() {
next_configuration.collaboration_mode = collaboration_mode;
}
if let Some(summary) = updates.reasoning_summary {
next_configuration.model_reasoning_summary = Some(summary);
}
if let Some(service_tier) = updates.service_tier {
next_configuration.service_tier = service_tier;
}
if let Some(personality) = updates.personality {
next_configuration.personality = Some(personality);
}
if let Some(approval_policy) = updates.approval_policy {
next_configuration.approval_policy.set(approval_policy)?;
}
if let Some(sandbox_policy) = updates.sandbox_policy.clone() {
next_configuration.sandbox_policy.set(sandbox_policy)?;
next_configuration.file_system_sandbox_policy =
FileSystemSandboxPolicy::from(next_configuration.sandbox_policy.get());
next_configuration.network_sandbox_policy =
NetworkSandboxPolicy::from(next_configuration.sandbox_policy.get());
}
if let Some(windows_sandbox_level) = updates.windows_sandbox_level {
next_configuration.windows_sandbox_level = windows_sandbox_level;
}
if let Some(cwd) = updates.cwd.clone() {
next_configuration.cwd = cwd;
}
if let Some(app_server_client_name) = updates.app_server_client_name.clone() {
next_configuration.app_server_client_name = Some(app_server_client_name);
}
Ok(next_configuration)
}
}
#[derive(Default, Clone)]
pub(crate) struct SessionSettingsUpdate {
pub(crate) cwd: Option<PathBuf>,
pub(crate) approval_policy: Option<AskForApproval>,
pub(crate) sandbox_policy: Option<SandboxPolicy>,
pub(crate) windows_sandbox_level: Option<WindowsSandboxLevel>,
pub(crate) collaboration_mode: Option<CollaborationMode>,
pub(crate) reasoning_summary: Option<ReasoningSummaryConfig>,
pub(crate) service_tier: Option<Option<ServiceTier>>,
pub(crate) final_output_json_schema: Option<Option<Value>>,
pub(crate) personality: Option<Personality>,
pub(crate) app_server_client_name: Option<String>,
}
impl Session {
/// Builds the `x-codex-beta-features` header value for this session.
///
/// `ModelClient` is session-scoped and intentionally does not depend on the full `Config`, so
/// we precompute the comma-separated list of enabled experimental feature keys at session
/// creation time and thread it into the client.
fn build_model_client_beta_features_header(config: &Config) -> Option<String> {
let beta_features_header = FEATURES
.iter()
.filter_map(|spec| {
if spec.stage.experimental_menu_description().is_some()
&& config.features.enabled(spec.id)
{
Some(spec.key)
} else {
None
}
})
.collect::<Vec<_>>()
.join(",");
if beta_features_header.is_empty() {
None
} else {
Some(beta_features_header)
}
}
async fn start_managed_network_proxy(
spec: &crate::config::NetworkProxySpec,
sandbox_policy: &SandboxPolicy,
network_policy_decider: Option<Arc<dyn codex_network_proxy::NetworkPolicyDecider>>,
blocked_request_observer: Option<Arc<dyn codex_network_proxy::BlockedRequestObserver>>,
managed_network_requirements_enabled: bool,
audit_metadata: NetworkProxyAuditMetadata,
) -> anyhow::Result<(StartedNetworkProxy, SessionNetworkProxyRuntime)> {
let network_proxy = spec
.start_proxy(
sandbox_policy,
network_policy_decider,
blocked_request_observer,
managed_network_requirements_enabled,
audit_metadata,
)
.await
.map_err(|err| anyhow::anyhow!("failed to start managed network proxy: {err}"))?;
let session_network_proxy = {
let proxy = network_proxy.proxy();
SessionNetworkProxyRuntime {
http_addr: proxy.http_addr().to_string(),
socks_addr: proxy.socks_addr().to_string(),
}
};
Ok((network_proxy, session_network_proxy))
}
/// Don't expand the number of mutated arguments on config. We are in the process of getting rid of it.
pub(crate) fn build_per_turn_config(session_configuration: &SessionConfiguration) -> Config {
// todo(aibrahim): store this state somewhere else so we don't need to mut config
let config = session_configuration.original_config_do_not_use.clone();
let mut per_turn_config = (*config).clone();
per_turn_config.model_reasoning_effort =
session_configuration.collaboration_mode.reasoning_effort();
per_turn_config.model_reasoning_summary = session_configuration.model_reasoning_summary;
per_turn_config.service_tier = session_configuration.service_tier;
per_turn_config.personality = session_configuration.personality;
let resolved_web_search_mode = resolve_web_search_mode_for_turn(
&per_turn_config.web_search_mode,
session_configuration.sandbox_policy.get(),
);
if let Err(err) = per_turn_config
.web_search_mode
.set(resolved_web_search_mode)
{
let fallback_value = per_turn_config.web_search_mode.value();
tracing::warn!(
error = %err,
?resolved_web_search_mode,
?fallback_value,
"resolved web_search_mode is disallowed by requirements; keeping constrained value"
);
}
per_turn_config.features = config.features.clone();
per_turn_config
}
pub(crate) async fn codex_home(&self) -> PathBuf {
let state = self.state.lock().await;
state.session_configuration.codex_home().clone()
}
fn start_file_watcher_listener(self: &Arc<Self>) {
let mut rx = self.services.file_watcher.subscribe();
let weak_sess = Arc::downgrade(self);
tokio::spawn(async move {
loop {
match rx.recv().await {
Ok(FileWatcherEvent::SkillsChanged { .. }) => {
let Some(sess) = weak_sess.upgrade() else {
break;
};
let event = Event {
id: sess.next_internal_sub_id(),
msg: EventMsg::SkillsUpdateAvailable,
};
sess.send_event_raw(event).await;
}
Err(tokio::sync::broadcast::error::RecvError::Closed) => break,
Err(tokio::sync::broadcast::error::RecvError::Lagged(_)) => continue,
}
}
});
}
#[allow(clippy::too_many_arguments)]
fn make_turn_context(
auth_manager: Option<Arc<AuthManager>>,
session_telemetry: &SessionTelemetry,
provider: ModelProviderInfo,
session_configuration: &SessionConfiguration,
per_turn_config: Config,
model_info: ModelInfo,
network: Option<NetworkProxy>,
sub_id: String,
js_repl: Arc<JsReplHandle>,
skills_outcome: Arc<SkillLoadOutcome>,
) -> TurnContext {
let reasoning_effort = session_configuration.collaboration_mode.reasoning_effort();
let reasoning_summary = session_configuration
.model_reasoning_summary
.unwrap_or(model_info.default_reasoning_summary);
let session_telemetry = session_telemetry.clone().with_model(
session_configuration.collaboration_mode.model(),
model_info.slug.as_str(),
);
let session_source = session_configuration.session_source.clone();
let auth_manager_for_context = auth_manager;
let provider_for_context = provider;
let session_telemetry_for_context = session_telemetry;
let per_turn_config = Arc::new(per_turn_config);
let tools_config = ToolsConfig::new(&ToolsConfigParams {
model_info: &model_info,
features: &per_turn_config.features,
web_search_mode: Some(per_turn_config.web_search_mode.value()),
session_source: session_source.clone(),
})
.with_web_search_config(per_turn_config.web_search_config.clone())
.with_allow_login_shell(per_turn_config.permissions.allow_login_shell)
.with_agent_roles(per_turn_config.agent_roles.clone());
let cwd = session_configuration.cwd.clone();
let turn_metadata_state = Arc::new(TurnMetadataState::new(
sub_id.clone(),
cwd.clone(),
session_configuration.sandbox_policy.get(),
session_configuration.windows_sandbox_level,
per_turn_config
.features
.enabled(Feature::UseLinuxSandboxBwrap),
));
let (current_date, timezone) = local_time_context();
TurnContext {
sub_id,
trace_id: current_span_trace_id(),
realtime_active: false,
config: per_turn_config.clone(),
auth_manager: auth_manager_for_context,
model_info: model_info.clone(),
session_telemetry: session_telemetry_for_context,
provider: provider_for_context,
reasoning_effort,
reasoning_summary,
session_source,
cwd,
current_date: Some(current_date),
timezone: Some(timezone),
app_server_client_name: session_configuration.app_server_client_name.clone(),
developer_instructions: session_configuration.developer_instructions.clone(),
compact_prompt: session_configuration.compact_prompt.clone(),
user_instructions: session_configuration.user_instructions.clone(),
collaboration_mode: session_configuration.collaboration_mode.clone(),
personality: session_configuration.personality,
approval_policy: session_configuration.approval_policy.clone(),
sandbox_policy: session_configuration.sandbox_policy.clone(),
file_system_sandbox_policy: session_configuration.file_system_sandbox_policy.clone(),
network_sandbox_policy: session_configuration.network_sandbox_policy,
network,
windows_sandbox_level: session_configuration.windows_sandbox_level,
shell_environment_policy: per_turn_config.permissions.shell_environment_policy.clone(),
tools_config,
features: per_turn_config.features.clone(),
ghost_snapshot: per_turn_config.ghost_snapshot.clone(),
final_output_json_schema: None,
codex_linux_sandbox_exe: per_turn_config.codex_linux_sandbox_exe.clone(),
tool_call_gate: Arc::new(ReadinessFlag::new()),
truncation_policy: model_info.truncation_policy.into(),
js_repl,
dynamic_tools: session_configuration.dynamic_tools.clone(),
turn_metadata_state,
turn_skills: TurnSkillsContext::new(skills_outcome),
turn_timing_state: Arc::new(TurnTimingState::default()),
}
}
#[allow(clippy::too_many_arguments)]
async fn new(
mut session_configuration: SessionConfiguration,
config: Arc<Config>,
auth_manager: Arc<AuthManager>,
models_manager: Arc<ModelsManager>,
exec_policy: ExecPolicyManager,
tx_event: Sender<Event>,
agent_status: watch::Sender<AgentStatus>,
initial_history: InitialHistory,
session_source: SessionSource,
skills_manager: Arc<SkillsManager>,
plugins_manager: Arc<PluginsManager>,
mcp_manager: Arc<McpManager>,
file_watcher: Arc<FileWatcher>,
agent_control: AgentControl,
) -> anyhow::Result<Arc<Self>> {
debug!(
"Configuring session: model={}; provider={:?}",
session_configuration.collaboration_mode.model(),
session_configuration.provider
);
if !session_configuration.cwd.is_absolute() {
return Err(anyhow::anyhow!(
"cwd is not absolute: {:?}",
session_configuration.cwd
));
}
let forked_from_id = initial_history.forked_from_id();
let (conversation_id, rollout_params) = match &initial_history {
InitialHistory::New | InitialHistory::Forked(_) => {
let conversation_id = ThreadId::default();
(
conversation_id,
RolloutRecorderParams::new(
conversation_id,
forked_from_id,
session_source,
BaseInstructions {
text: session_configuration.base_instructions.clone(),
},
session_configuration.dynamic_tools.clone(),
if session_configuration.persist_extended_history {
EventPersistenceMode::Extended
} else {
EventPersistenceMode::Limited
},
),
)
}
InitialHistory::Resumed(resumed_history) => (
resumed_history.conversation_id,
RolloutRecorderParams::resume(
resumed_history.rollout_path.clone(),
if session_configuration.persist_extended_history {
EventPersistenceMode::Extended
} else {
EventPersistenceMode::Limited
},
),
),
};
let state_builder = match &initial_history {
InitialHistory::Resumed(resumed) => metadata::builder_from_items(
resumed.history.as_slice(),
resumed.rollout_path.as_path(),
),
InitialHistory::New | InitialHistory::Forked(_) => None,
};
// Kick off independent async setup tasks in parallel to reduce startup latency.
//
// - initialize RolloutRecorder with new or resumed session info
// - perform default shell discovery
// - load history metadata (skipped for subagents)
let rollout_fut = async {
if config.ephemeral {
Ok::<_, anyhow::Error>((None, None))
} else {
let state_db_ctx = state_db::init(&config).await;
let rollout_recorder = RolloutRecorder::new(
&config,
rollout_params,
state_db_ctx.clone(),
state_builder.clone(),
)
.await?;
Ok((Some(rollout_recorder), state_db_ctx))
}
};
let history_meta_fut = async {
if matches!(
session_configuration.session_source,
SessionSource::SubAgent(_)
) {
(0, 0)
} else {
crate::message_history::history_metadata(&config).await
}
};
let auth_manager_clone = Arc::clone(&auth_manager);
let config_for_mcp = Arc::clone(&config);
let mcp_manager_for_mcp = Arc::clone(&mcp_manager);
let auth_and_mcp_fut = async move {
let auth = auth_manager_clone.auth().await;
let mcp_servers = mcp_manager_for_mcp.effective_servers(&config_for_mcp, auth.as_ref());
let auth_statuses = compute_auth_statuses(
mcp_servers.iter(),
config_for_mcp.mcp_oauth_credentials_store_mode,
)
.await;
(auth, mcp_servers, auth_statuses)
};
// Join all independent futures.
let (
rollout_recorder_and_state_db,
(history_log_id, history_entry_count),
(auth, mcp_servers, auth_statuses),
) = tokio::join!(rollout_fut, history_meta_fut, auth_and_mcp_fut);
let (rollout_recorder, state_db_ctx) = rollout_recorder_and_state_db.map_err(|e| {
error!("failed to initialize rollout recorder: {e:#}");
e
})?;
let rollout_path = rollout_recorder
.as_ref()
.map(|rec| rec.rollout_path.clone());
let mut post_session_configured_events = Vec::<Event>::new();
for usage in config.features.legacy_feature_usages() {
post_session_configured_events.push(Event {
id: INITIAL_SUBMIT_ID.to_owned(),
msg: EventMsg::DeprecationNotice(DeprecationNoticeEvent {
summary: usage.summary.clone(),
details: usage.details.clone(),
}),
});
}
if crate::config::uses_deprecated_instructions_file(&config.config_layer_stack) {
post_session_configured_events.push(Event {
id: INITIAL_SUBMIT_ID.to_owned(),
msg: EventMsg::DeprecationNotice(DeprecationNoticeEvent {
summary: "`experimental_instructions_file` is deprecated and ignored. Use `model_instructions_file` instead."
.to_string(),
details: Some(
"Move the setting to `model_instructions_file` in config.toml (or under a profile) to load instructions from a file."
.to_string(),
),
}),
});
}
for message in &config.startup_warnings {
post_session_configured_events.push(Event {
id: "".to_owned(),
msg: EventMsg::Warning(WarningEvent {
message: message.clone(),
}),
});
}
maybe_push_unstable_features_warning(&config, &mut post_session_configured_events);
if config.permissions.approval_policy.value() == AskForApproval::OnFailure {
post_session_configured_events.push(Event {
id: "".to_owned(),
msg: EventMsg::Warning(WarningEvent {
message: "`on-failure` approval policy is deprecated and will be removed in a future release. Use `on-request` for interactive approvals or `never` for non-interactive runs.".to_string(),
}),
});
}
let auth = auth.as_ref();
let auth_mode = auth.map(CodexAuth::auth_mode).map(TelemetryAuthMode::from);
let account_id = auth.and_then(CodexAuth::get_account_id);
let account_email = auth.and_then(CodexAuth::get_account_email);
let originator = crate::default_client::originator().value;
let terminal_type = terminal::user_agent();
let session_model = session_configuration.collaboration_mode.model().to_string();
let mut session_telemetry = SessionTelemetry::new(
conversation_id,
session_model.as_str(),
session_model.as_str(),
account_id.clone(),
account_email.clone(),
auth_mode,
originator.clone(),
config.otel.log_user_prompt,
terminal_type.clone(),
session_configuration.session_source.clone(),
);
if let Some(service_name) = session_configuration.metrics_service_name.as_deref() {
session_telemetry = session_telemetry.with_metrics_service_name(service_name);
}
let network_proxy_audit_metadata = NetworkProxyAuditMetadata {
conversation_id: Some(conversation_id.to_string()),
app_version: Some(env!("CARGO_PKG_VERSION").to_string()),
user_account_id: account_id,
auth_mode: auth_mode.map(|mode| mode.to_string()),
originator: Some(originator),
user_email: account_email,
terminal_type: Some(terminal_type),
model: Some(session_model.clone()),
slug: Some(session_model),
};
config.features.emit_metrics(&session_telemetry);
session_telemetry.counter(
THREAD_STARTED_METRIC,
1,
&[(
"is_git",
if get_git_repo_root(&session_configuration.cwd).is_some() {
"true"
} else {
"false"
},
)],
);
session_telemetry.conversation_starts(
config.model_provider.name.as_str(),
session_configuration.collaboration_mode.reasoning_effort(),
config
.model_reasoning_summary
.unwrap_or(ReasoningSummaryConfig::Auto),
config.model_context_window,
config.model_auto_compact_token_limit,
config.permissions.approval_policy.value(),
config.permissions.sandbox_policy.get().clone(),
mcp_servers.keys().map(String::as_str).collect(),
config.active_profile.clone(),
);
let use_zsh_fork_shell = config.features.enabled(Feature::ShellZshFork);
let mut default_shell = if use_zsh_fork_shell {
let zsh_path = config.zsh_path.as_ref().ok_or_else(|| {
anyhow::anyhow!(
"zsh fork feature enabled, but `zsh_path` is not configured; set `zsh_path` in config.toml"
)
})?;
let zsh_path = zsh_path.to_path_buf();
shell::get_shell(shell::ShellType::Zsh, Some(&zsh_path)).ok_or_else(|| {
anyhow::anyhow!(
"zsh fork feature enabled, but zsh_path `{}` is not usable; set `zsh_path` to a valid zsh executable",
zsh_path.display()
)
})?
} else {
shell::default_user_shell()
};
// Create the mutable state for the Session.
let shell_snapshot_tx = if config.features.enabled(Feature::ShellSnapshot) {
if let Some(snapshot) = session_configuration.inherited_shell_snapshot.clone() {
let (tx, rx) = watch::channel(Some(snapshot));
default_shell.shell_snapshot = rx;
tx
} else {
ShellSnapshot::start_snapshotting(
config.codex_home.clone(),
conversation_id,
session_configuration.cwd.clone(),
&mut default_shell,
session_telemetry.clone(),
)
}
} else {
let (tx, rx) = watch::channel(None);
default_shell.shell_snapshot = rx;
tx
};
let thread_name =
match session_index::find_thread_name_by_id(&config.codex_home, &conversation_id).await
{
Ok(name) => name,
Err(err) => {
warn!("Failed to read session index for thread name: {err}");
None
}
};
session_configuration.thread_name = thread_name.clone();
let state = SessionState::new(session_configuration.clone());
let managed_network_requirements_enabled = config.managed_network_requirements_enabled();
let network_approval = Arc::new(NetworkApprovalService::default());
// The managed proxy can call back into core for allowlist-miss decisions.
let network_policy_decider_session = if managed_network_requirements_enabled {
config
.permissions
.network
.as_ref()
.map(|_| Arc::new(RwLock::new(std::sync::Weak::<Session>::new())))
} else {
None
};
let blocked_request_observer = if managed_network_requirements_enabled {
config
.permissions
.network
.as_ref()
.map(|_| build_blocked_request_observer(Arc::clone(&network_approval)))
} else {
None
};
let network_policy_decider =
network_policy_decider_session
.as_ref()
.map(|network_policy_decider_session| {
build_network_policy_decider(
Arc::clone(&network_approval),
Arc::clone(network_policy_decider_session),
)
});
let (network_proxy, session_network_proxy) =
if let Some(spec) = config.permissions.network.as_ref() {
let (network_proxy, session_network_proxy) = Self::start_managed_network_proxy(
spec,
config.permissions.sandbox_policy.get(),
network_policy_decider.as_ref().map(Arc::clone),
blocked_request_observer.as_ref().map(Arc::clone),
managed_network_requirements_enabled,
network_proxy_audit_metadata,
)
.await?;
(Some(network_proxy), Some(session_network_proxy))
} else {
(None, None)
};
let services = SessionServices {
// Initialize the MCP connection manager with an uninitialized
// instance. It will be replaced with one created via
// McpConnectionManager::new() once all its constructor args are
// available. This also ensures `SessionConfigured` is emitted
// before any MCP-related events. It is reasonable to consider
// changing this to use Option or OnceCell, though the current
// setup is straightforward enough and performs well.
mcp_connection_manager: Arc::new(RwLock::new(McpConnectionManager::new_uninitialized(
&config.permissions.approval_policy,
))),
mcp_startup_cancellation_token: Mutex::new(CancellationToken::new()),
unified_exec_manager: UnifiedExecProcessManager::new(
config.background_terminal_max_timeout,
),
shell_zsh_path: config.zsh_path.clone(),
main_execve_wrapper_exe: config.main_execve_wrapper_exe.clone(),
analytics_events_client: AnalyticsEventsClient::new(
Arc::clone(&config),
Arc::clone(&auth_manager),
),
hooks: Hooks::new(HooksConfig {
legacy_notify_argv: config.notify.clone(),
}),
rollout: Mutex::new(rollout_recorder),
user_shell: Arc::new(default_shell),
shell_snapshot_tx,
show_raw_agent_reasoning: config.show_raw_agent_reasoning,
exec_policy,
auth_manager: Arc::clone(&auth_manager),
session_telemetry,
models_manager: Arc::clone(&models_manager),
tool_approvals: Mutex::new(ApprovalStore::default()),
execve_session_approvals: RwLock::new(HashMap::new()),
skills_manager,
plugins_manager: Arc::clone(&plugins_manager),
mcp_manager: Arc::clone(&mcp_manager),
file_watcher,
agent_control,
network_proxy,
network_approval: Arc::clone(&network_approval),
state_db: state_db_ctx.clone(),
model_client: ModelClient::new(
Some(Arc::clone(&auth_manager)),
conversation_id,
session_configuration.provider.clone(),
session_configuration.session_source.clone(),
config.model_verbosity,
ws_version_from_features(config.as_ref()),
config.features.enabled(Feature::EnableRequestCompression),
config.features.enabled(Feature::RuntimeMetrics),
Self::build_model_client_beta_features_header(config.as_ref()),
),
};
let js_repl = Arc::new(JsReplHandle::with_node_path(
config.js_repl_node_path.clone(),
config.js_repl_node_module_dirs.clone(),
));
let sess = Arc::new(Session {
conversation_id,
tx_event: tx_event.clone(),
agent_status,
state: Mutex::new(state),
features: config.features.clone(),
pending_mcp_server_refresh_config: Mutex::new(None),
conversation: Arc::new(RealtimeConversationManager::new()),
active_turn: Mutex::new(None),
services,
js_repl,
next_internal_sub_id: AtomicU64::new(0),
});
if let Some(network_policy_decider_session) = network_policy_decider_session {
let mut guard = network_policy_decider_session.write().await;
*guard = Arc::downgrade(&sess);
}
// Dispatch the SessionConfiguredEvent first and then report any errors.
// If resuming, include converted initial messages in the payload so UIs can render them immediately.
let initial_messages = initial_history.get_event_msgs();
let events = std::iter::once(Event {
id: INITIAL_SUBMIT_ID.to_owned(),
msg: EventMsg::SessionConfigured(SessionConfiguredEvent {
session_id: conversation_id,
forked_from_id,
thread_name: session_configuration.thread_name.clone(),
model: session_configuration.collaboration_mode.model().to_string(),
model_provider_id: config.model_provider_id.clone(),
service_tier: session_configuration.service_tier,
approval_policy: session_configuration.approval_policy.value(),
sandbox_policy: session_configuration.sandbox_policy.get().clone(),
cwd: session_configuration.cwd.clone(),
reasoning_effort: session_configuration.collaboration_mode.reasoning_effort(),
history_log_id,
history_entry_count,
initial_messages,
network_proxy: session_network_proxy,
rollout_path,
}),
})
.chain(post_session_configured_events.into_iter());
for event in events {
sess.send_event_raw(event).await;
}
// Start the watcher after SessionConfigured so it cannot emit earlier events.
sess.start_file_watcher_listener();
// Construct sandbox_state before MCP startup so it can be sent to each
// MCP server immediately after it becomes ready (avoiding blocking).
let sandbox_state = SandboxState {
sandbox_policy: session_configuration.sandbox_policy.get().clone(),
codex_linux_sandbox_exe: config.codex_linux_sandbox_exe.clone(),
sandbox_cwd: session_configuration.cwd.clone(),
use_linux_sandbox_bwrap: config.features.enabled(Feature::UseLinuxSandboxBwrap),
};
let mut required_mcp_servers: Vec<String> = mcp_servers
.iter()
.filter(|(_, server)| server.enabled && server.required)
.map(|(name, _)| name.clone())
.collect();
required_mcp_servers.sort();
let tool_plugin_provenance = mcp_manager.tool_plugin_provenance(config.as_ref());
{
let mut cancel_guard = sess.services.mcp_startup_cancellation_token.lock().await;
cancel_guard.cancel();
*cancel_guard = CancellationToken::new();
}
let (mcp_connection_manager, cancel_token) = McpConnectionManager::new(
&mcp_servers,
config.mcp_oauth_credentials_store_mode,
auth_statuses.clone(),
&session_configuration.approval_policy,
tx_event.clone(),
sandbox_state,
config.codex_home.clone(),
codex_apps_tools_cache_key(auth),
tool_plugin_provenance,
)
.await;
{
let mut manager_guard = sess.services.mcp_connection_manager.write().await;
*manager_guard = mcp_connection_manager;
}
{
let mut cancel_guard = sess.services.mcp_startup_cancellation_token.lock().await;
if cancel_guard.is_cancelled() {
cancel_token.cancel();
}
*cancel_guard = cancel_token;
}
if !required_mcp_servers.is_empty() {
let failures = sess
.services
.mcp_connection_manager
.read()
.await
.required_startup_failures(&required_mcp_servers)
.await;
if !failures.is_empty() {
let details = failures
.iter()
.map(|failure| format!("{}: {}", failure.server, failure.error))
.collect::<Vec<_>>()
.join("; ");
return Err(anyhow::anyhow!(
"required MCP servers failed to initialize: {details}"
));
}
}
sess.schedule_startup_prewarm(session_configuration.base_instructions.clone())
.await;
// record_initial_history can emit events. We record only after the SessionConfiguredEvent is emitted.
sess.record_initial_history(initial_history).await;
memories::start_memories_startup_task(
&sess,
Arc::clone(&config),
&session_configuration.session_source,
);
Ok(sess)
}
pub(crate) fn get_tx_event(&self) -> Sender<Event> {
self.tx_event.clone()
}
pub(crate) fn state_db(&self) -> Option<state_db::StateDbHandle> {
self.services.state_db.clone()
}
/// Ensure rollout file writes are durably flushed.
pub(crate) async fn flush_rollout(&self) {
let recorder = {
let guard = self.services.rollout.lock().await;
guard.clone()
};
if let Some(rec) = recorder
&& let Err(e) = rec.flush().await
{
warn!("failed to flush rollout recorder: {e}");
}
}
pub(crate) async fn ensure_rollout_materialized(&self) {
let recorder = {
let guard = self.services.rollout.lock().await;
guard.clone()
};
if let Some(rec) = recorder
&& let Err(e) = rec.persist().await
{
warn!("failed to materialize rollout recorder: {e}");
}
}
fn next_internal_sub_id(&self) -> String {
let id = self
.next_internal_sub_id
.fetch_add(1, std::sync::atomic::Ordering::SeqCst);
format!("auto-compact-{id}")
}
pub(crate) async fn route_realtime_text_input(self: &Arc<Self>, text: String) {
handlers::user_input_or_turn(
self,
self.next_internal_sub_id(),
Op::UserInput {
items: vec![UserInput::Text {
text,
text_elements: Vec::new(),
}],
final_output_json_schema: None,
},
)
.await;
}
pub(crate) async fn get_total_token_usage(&self) -> i64 {
let state = self.state.lock().await;
state.get_total_token_usage(state.server_reasoning_included())
}
pub(crate) async fn get_total_token_usage_breakdown(&self) -> TotalTokenUsageBreakdown {
let state = self.state.lock().await;
state.history.get_total_token_usage_breakdown()
}
pub(crate) async fn total_token_usage(&self) -> Option<TokenUsage> {
let state = self.state.lock().await;
state.token_info().map(|info| info.total_token_usage)
}
pub(crate) async fn get_estimated_token_count(
&self,
turn_context: &TurnContext,
) -> Option<i64> {
let state = self.state.lock().await;
state.history.estimate_token_count(turn_context)
}
pub(crate) async fn get_base_instructions(&self) -> BaseInstructions {
let state = self.state.lock().await;
BaseInstructions {
text: state.session_configuration.base_instructions.clone(),
}
}
pub(crate) async fn merge_mcp_tool_selection(&self, tool_names: Vec<String>) -> Vec<String> {
let mut state = self.state.lock().await;
state.merge_mcp_tool_selection(tool_names)
}
pub(crate) async fn set_mcp_tool_selection(&self, tool_names: Vec<String>) {
let mut state = self.state.lock().await;
state.set_mcp_tool_selection(tool_names);
}
pub(crate) async fn get_mcp_tool_selection(&self) -> Option<Vec<String>> {
let state = self.state.lock().await;
state.get_mcp_tool_selection()
}
pub(crate) async fn clear_mcp_tool_selection(&self) {
let mut state = self.state.lock().await;
state.clear_mcp_tool_selection();
}
// Merges connector IDs into the session-level explicit connector selection.
pub(crate) async fn merge_connector_selection(
&self,
connector_ids: HashSet<String>,
) -> HashSet<String> {
let mut state = self.state.lock().await;
state.merge_connector_selection(connector_ids)
}
// Returns the connector IDs currently selected for this session.
pub(crate) async fn get_connector_selection(&self) -> HashSet<String> {
let state = self.state.lock().await;
state.get_connector_selection()
}
// Clears connector IDs that were accumulated for explicit selection.
pub(crate) async fn clear_connector_selection(&self) {
let mut state = self.state.lock().await;
state.clear_connector_selection();
}
async fn record_initial_history(&self, conversation_history: InitialHistory) {
let turn_context = self.new_default_turn().await;
self.clear_mcp_tool_selection().await;
let is_subagent = {
let state = self.state.lock().await;
matches!(
state.session_configuration.session_source,
SessionSource::SubAgent(_)
)
};
match conversation_history {
InitialHistory::New => {
// Build and record initial items (user instructions + environment context)
// TODO(ccunningham): Defer initial context insertion until the first real turn
// starts so it reflects the actual first-turn settings (permissions, etc.) and
// we do not emit model-visible "diff" updates before the first user message.
let items = self.build_initial_context(&turn_context).await;
self.record_conversation_items(&turn_context, &items).await;
{
let mut state = self.state.lock().await;
state.set_reference_context_item(Some(turn_context.to_turn_context_item()));
}
self.set_previous_turn_settings(None).await;
// Ensure initial items are visible to immediate readers (e.g., tests, forks).
if !is_subagent {
self.flush_rollout().await;
}
}
InitialHistory::Resumed(resumed_history) => {
let rollout_items = resumed_history.history;
let restored_tool_selection =
Self::extract_mcp_tool_selection_from_rollout(&rollout_items);
let reconstructed_rollout = self
.reconstruct_history_from_rollout(&turn_context, &rollout_items)
.await;
let previous_turn_settings = reconstructed_rollout.previous_turn_settings.clone();
self.set_previous_turn_settings(previous_turn_settings.clone())
.await;
{
let mut state = self.state.lock().await;
state.set_reference_context_item(reconstructed_rollout.reference_context_item);
}
// If resuming, warn when the last recorded model differs from the current one.
let curr: &str = turn_context.model_info.slug.as_str();
if let Some(prev) = previous_turn_settings
.as_ref()
.map(|settings| settings.model.as_str())
.filter(|model| *model != curr)
{
warn!("resuming session with different model: previous={prev}, current={curr}");
self.send_event(
&turn_context,
EventMsg::Warning(WarningEvent {
message: format!(
"This session was recorded with model `{prev}` but is resuming with `{curr}`. \
Consider switching back to `{prev}` as it may affect Codex performance."
),
}),
)
.await;
}
// Always add response items to conversation history
let reconstructed_history = reconstructed_rollout.history;
if !reconstructed_history.is_empty() {
self.record_into_history(&reconstructed_history, &turn_context)
.await;
}
// Seed usage info from the recorded rollout so UIs can show token counts
// immediately on resume/fork.
if let Some(info) = Self::last_token_info_from_rollout(&rollout_items) {
let mut state = self.state.lock().await;
state.set_token_info(Some(info));
}
if let Some(selected_tools) = restored_tool_selection {
self.set_mcp_tool_selection(selected_tools).await;
}
// Defer seeding the session's initial context until the first turn starts so
// turn/start overrides can be merged before we write to the rollout.
if !is_subagent {
self.flush_rollout().await;
}
}
InitialHistory::Forked(rollout_items) => {
let restored_tool_selection =
Self::extract_mcp_tool_selection_from_rollout(&rollout_items);
let reconstructed_rollout = self
.reconstruct_history_from_rollout(&turn_context, &rollout_items)
.await;
self.set_previous_turn_settings(
reconstructed_rollout.previous_turn_settings.clone(),
)
.await;
{
let mut state = self.state.lock().await;
state.set_reference_context_item(
reconstructed_rollout.reference_context_item.clone(),
);
}
// Always add response items to conversation history
let reconstructed_history = reconstructed_rollout.history;
if !reconstructed_history.is_empty() {
self.record_into_history(&reconstructed_history, &turn_context)
.await;
}
// Seed usage info from the recorded rollout so UIs can show token counts
// immediately on resume/fork.
if let Some(info) = Self::last_token_info_from_rollout(&rollout_items) {
let mut state = self.state.lock().await;
state.set_token_info(Some(info));
}
if let Some(selected_tools) = restored_tool_selection {
self.set_mcp_tool_selection(selected_tools).await;
}
// If persisting, persist all rollout items as-is (recorder filters)
if !rollout_items.is_empty() {
self.persist_rollout_items(&rollout_items).await;
}
// Append the current session's initial context after the reconstructed history.
let initial_context = self.build_initial_context(&turn_context).await;
self.record_conversation_items(&turn_context, &initial_context)
.await;
{
let mut state = self.state.lock().await;
state.set_reference_context_item(Some(turn_context.to_turn_context_item()));
}
// Forked threads should remain file-backed immediately after startup.
self.ensure_rollout_materialized().await;
// Flush after seeding history and any persisted rollout copy.
if !is_subagent {
self.flush_rollout().await;
}
}
}
}
fn last_token_info_from_rollout(rollout_items: &[RolloutItem]) -> Option<TokenUsageInfo> {
rollout_items.iter().rev().find_map(|item| match item {
RolloutItem::EventMsg(EventMsg::TokenCount(ev)) => ev.info.clone(),
_ => None,
})
}
fn extract_mcp_tool_selection_from_rollout(
rollout_items: &[RolloutItem],
) -> Option<Vec<String>> {
let mut search_call_ids = HashSet::new();
let mut active_selected_tools: Option<Vec<String>> = None;
for item in rollout_items {
let RolloutItem::ResponseItem(response_item) = item else {
continue;
};
match response_item {
ResponseItem::FunctionCall { name, call_id, .. } => {
if name == SEARCH_TOOL_BM25_TOOL_NAME {
search_call_ids.insert(call_id.clone());
}
}
ResponseItem::FunctionCallOutput { call_id, output } => {
if !search_call_ids.contains(call_id) {
continue;
}
let Some(content) = output.body.to_text() else {
continue;
};
let Ok(payload) = serde_json::from_str::<Value>(&content) else {
continue;
};
let Some(selected_tools) = payload
.get("active_selected_tools")
.and_then(Value::as_array)
else {
continue;
};
let Some(selected_tools) = selected_tools
.iter()
.map(|value| value.as_str().map(str::to_string))
.collect::<Option<Vec<_>>>()
else {
continue;
};
active_selected_tools = Some(selected_tools);
}
_ => {}
}
}
active_selected_tools
}
async fn previous_turn_settings(&self) -> Option<PreviousTurnSettings> {
let state = self.state.lock().await;
state.previous_turn_settings()
}
pub(crate) async fn set_previous_turn_settings(
&self,
previous_turn_settings: Option<PreviousTurnSettings>,
) {
let mut state = self.state.lock().await;
state.set_previous_turn_settings(previous_turn_settings);
}
fn maybe_refresh_shell_snapshot_for_cwd(
&self,
previous_cwd: &Path,
next_cwd: &Path,
codex_home: &Path,
session_source: &SessionSource,
) {
if previous_cwd == next_cwd {
return;
}
if !self.features.enabled(Feature::ShellSnapshot) {
return;
}
if matches!(
session_source,
SessionSource::SubAgent(SubAgentSource::ThreadSpawn { .. })
) {
return;
}
ShellSnapshot::refresh_snapshot(
codex_home.to_path_buf(),
self.conversation_id,
next_cwd.to_path_buf(),
self.services.user_shell.as_ref().clone(),
self.services.shell_snapshot_tx.clone(),
self.services.session_telemetry.clone(),
);
}
pub(crate) async fn update_settings(
&self,
updates: SessionSettingsUpdate,
) -> ConstraintResult<()> {
let mut state = self.state.lock().await;
match state.session_configuration.apply(&updates) {
Ok(updated) => {
let previous_cwd = state.session_configuration.cwd.clone();
let next_cwd = updated.cwd.clone();
let codex_home = updated.codex_home.clone();
let session_source = updated.session_source.clone();
state.session_configuration = updated;
drop(state);
self.maybe_refresh_shell_snapshot_for_cwd(
&previous_cwd,
&next_cwd,
&codex_home,
&session_source,
);
Ok(())
}
Err(err) => {
warn!("rejected session settings update: {err}");
Err(err)
}
}
}
pub(crate) async fn new_turn_with_sub_id(
&self,
sub_id: String,
updates: SessionSettingsUpdate,
) -> ConstraintResult<Arc<TurnContext>> {
let (
session_configuration,
sandbox_policy_changed,
previous_cwd,
codex_home,
session_source,
) = {
let mut state = self.state.lock().await;
match state.session_configuration.clone().apply(&updates) {
Ok(next) => {
let previous_cwd = state.session_configuration.cwd.clone();
let sandbox_policy_changed =
state.session_configuration.sandbox_policy != next.sandbox_policy;
let codex_home = next.codex_home.clone();
let session_source = next.session_source.clone();
state.session_configuration = next.clone();
(
next,
sandbox_policy_changed,
previous_cwd,
codex_home,
session_source,
)
}
Err(err) => {
drop(state);
self.send_event_raw(Event {
id: sub_id.clone(),
msg: EventMsg::Error(ErrorEvent {
message: err.to_string(),
codex_error_info: Some(CodexErrorInfo::BadRequest),
}),
})
.await;
return Err(err);
}
}
};
self.maybe_refresh_shell_snapshot_for_cwd(
&previous_cwd,
&session_configuration.cwd,
&codex_home,
&session_source,
);
Ok(self
.new_turn_from_configuration(
sub_id,
session_configuration,
updates.final_output_json_schema,
sandbox_policy_changed,
)
.await)
}
async fn new_turn_from_configuration(
&self,
sub_id: String,
session_configuration: SessionConfiguration,
final_output_json_schema: Option<Option<Value>>,
sandbox_policy_changed: bool,
) -> Arc<TurnContext> {
let per_turn_config = Self::build_per_turn_config(&session_configuration);
self.services
.mcp_connection_manager
.read()
.await
.set_approval_policy(&session_configuration.approval_policy);
if sandbox_policy_changed {
let sandbox_state = SandboxState {
sandbox_policy: per_turn_config.permissions.sandbox_policy.get().clone(),
codex_linux_sandbox_exe: per_turn_config.codex_linux_sandbox_exe.clone(),
sandbox_cwd: per_turn_config.cwd.clone(),
use_linux_sandbox_bwrap: per_turn_config
.features
.enabled(Feature::UseLinuxSandboxBwrap),
};
if let Err(e) = self
.services
.mcp_connection_manager
.read()
.await
.notify_sandbox_state_change(&sandbox_state)
.await
{
warn!("Failed to notify sandbox state change to MCP servers: {e:#}");
}
}
let model_info = self
.services
.models_manager
.get_model_info(
session_configuration.collaboration_mode.model(),
&per_turn_config,
)
.await;
let skills_outcome = Arc::new(
self.services
.skills_manager
.skills_for_cwd(&session_configuration.cwd, false)
.await,
);
let mut turn_context: TurnContext = Self::make_turn_context(
Some(Arc::clone(&self.services.auth_manager)),
&self.services.session_telemetry,
session_configuration.provider.clone(),
&session_configuration,
per_turn_config,
model_info,
self.services
.network_proxy
.as_ref()
.map(StartedNetworkProxy::proxy),
sub_id,
Arc::clone(&self.js_repl),
skills_outcome,
);
turn_context.realtime_active = self.conversation.running_state().await.is_some();
if let Some(final_schema) = final_output_json_schema {
turn_context.final_output_json_schema = final_schema;
}
let turn_context = Arc::new(turn_context);
turn_context.turn_metadata_state.spawn_git_enrichment_task();
turn_context
}
pub(crate) async fn maybe_emit_unknown_model_warning_for_turn(&self, tc: &TurnContext) {
if tc.model_info.used_fallback_model_metadata {
self.send_event(
tc,
EventMsg::Warning(WarningEvent {
message: format!(
"Model metadata for `{}` not found. Defaulting to fallback metadata; this can degrade performance and cause issues.",
tc.model_info.slug
),
}),
)
.await;
}
}
pub(crate) async fn new_default_turn(&self) -> Arc<TurnContext> {
self.new_default_turn_with_sub_id(self.next_internal_sub_id())
.await
}
pub(crate) async fn take_startup_regular_task(&self) -> Option<RegularTask> {
let startup_regular_task = {
let mut state = self.state.lock().await;
state.take_startup_regular_task()
};
let startup_regular_task = startup_regular_task?;
match startup_regular_task.await {
Ok(Ok(regular_task)) => Some(regular_task),
Ok(Err(err)) => {
warn!("startup websocket prewarm setup failed: {err:#}");
None
}
Err(err) => {
warn!("startup websocket prewarm setup join failed: {err}");
None
}
}
}
async fn schedule_startup_prewarm(self: &Arc<Self>, base_instructions: String) {
let sess = Arc::clone(self);
let startup_regular_task: JoinHandle<CodexResult<RegularTask>> =
tokio::spawn(
async move { sess.schedule_startup_prewarm_inner(base_instructions).await },
);
let mut state = self.state.lock().await;
state.set_startup_regular_task(startup_regular_task);
}
async fn schedule_startup_prewarm_inner(
self: &Arc<Self>,
base_instructions: String,
) -> CodexResult<RegularTask> {
let startup_turn_context = self
.new_default_turn_with_sub_id(INITIAL_SUBMIT_ID.to_owned())
.await;
let startup_cancellation_token = CancellationToken::new();
let startup_router = built_tools(
self,
startup_turn_context.as_ref(),
&[],
&HashSet::new(),
None,
&startup_cancellation_token,
)
.await?;
let startup_prompt = build_prompt(
Vec::new(),
startup_router.as_ref(),
startup_turn_context.as_ref(),
BaseInstructions {
text: base_instructions,
},
);
let startup_turn_metadata_header = startup_turn_context
.turn_metadata_state
.current_header_value();
RegularTask::with_startup_prewarm(
self.services.model_client.clone(),
startup_prompt,
startup_turn_context,
startup_turn_metadata_header,
)
.await
}
pub(crate) async fn get_config(&self) -> std::sync::Arc<Config> {
let state = self.state.lock().await;
state
.session_configuration
.original_config_do_not_use
.clone()
}
pub(crate) async fn provider(&self) -> ModelProviderInfo {
let state = self.state.lock().await;
state.session_configuration.provider.clone()
}
pub(crate) async fn reload_user_config_layer(&self) {
let config_toml_path = {
let state = self.state.lock().await;
state
.session_configuration
.codex_home
.join(CONFIG_TOML_FILE)
};
let user_config = match std::fs::read_to_string(&config_toml_path) {
Ok(contents) => match toml::from_str::<toml::Value>(&contents) {
Ok(config) => config,
Err(err) => {
warn!("failed to parse user config while reloading layer: {err}");
return;
}
},
Err(err) if err.kind() == std::io::ErrorKind::NotFound => {
toml::Value::Table(Default::default())
}
Err(err) => {
warn!("failed to read user config while reloading layer: {err}");
return;
}
};
let config_toml_path = match AbsolutePathBuf::try_from(config_toml_path) {
Ok(path) => path,
Err(err) => {
warn!("failed to resolve user config path while reloading layer: {err}");
return;
}
};
let mut state = self.state.lock().await;
let mut config = (*state.session_configuration.original_config_do_not_use).clone();
config.config_layer_stack = config
.config_layer_stack
.with_user_config(&config_toml_path, user_config);
state.session_configuration.original_config_do_not_use = Arc::new(config);
self.services.skills_manager.clear_cache();
self.services.plugins_manager.clear_cache();
}
pub(crate) async fn new_default_turn_with_sub_id(&self, sub_id: String) -> Arc<TurnContext> {
let session_configuration = {
let state = self.state.lock().await;
state.session_configuration.clone()
};
self.new_turn_from_configuration(sub_id, session_configuration, None, false)
.await
}
async fn build_settings_update_items(
&self,
reference_context_item: Option<&TurnContextItem>,
current_context: &TurnContext,
) -> Vec<ResponseItem> {
// TODO: Make context updates a pure diff of persisted previous/current TurnContextItem
// state so replay/backtracking is deterministic. Runtime inputs that affect model-visible
// context (shell, exec policy, feature gates, previous-turn bridge) should be persisted
// state or explicit non-state replay events.
let previous_turn_settings = {
let state = self.state.lock().await;
state.previous_turn_settings()
};
let shell = self.user_shell();
let exec_policy = self.services.exec_policy.current();
crate::context_manager::updates::build_settings_update_items(
reference_context_item,
previous_turn_settings.as_ref(),
current_context,
shell.as_ref(),
exec_policy.as_ref(),
self.features.enabled(Feature::Personality),
)
}
/// Persist the event to rollout and send it to clients.
pub(crate) async fn send_event(&self, turn_context: &TurnContext, msg: EventMsg) {
let legacy_source = msg.clone();
let event = Event {
id: turn_context.sub_id.clone(),
msg,
};
self.send_event_raw(event).await;
self.maybe_mirror_event_text_to_realtime(&legacy_source)
.await;
self.maybe_clear_realtime_handoff_for_event(&legacy_source)
.await;
let show_raw_agent_reasoning = self.show_raw_agent_reasoning();
for legacy in legacy_source.as_legacy_events(show_raw_agent_reasoning) {
let legacy_event = Event {
id: turn_context.sub_id.clone(),
msg: legacy,
};
self.send_event_raw(legacy_event).await;
}
}
async fn maybe_mirror_event_text_to_realtime(&self, msg: &EventMsg) {
let Some(text) = realtime_text_for_event(msg) else {
return;
};
if self.conversation.running_state().await.is_none()
|| self.conversation.active_handoff_id().await.is_none()
{
return;
}
if let Err(err) = self.conversation.handoff_out(text).await {
debug!("failed to mirror event text to realtime conversation: {err}");
}
}
async fn maybe_clear_realtime_handoff_for_event(&self, msg: &EventMsg) {
if !matches!(msg, EventMsg::TurnComplete(_)) {
return;
}
self.conversation.clear_active_handoff().await;
}
pub(crate) async fn send_event_raw(&self, event: Event) {
// Record the last known agent status.
if let Some(status) = agent_status_from_event(&event.msg) {
self.agent_status.send_replace(status);
}
// Persist the event into rollout (recorder filters as needed)
let rollout_items = vec![RolloutItem::EventMsg(event.msg.clone())];
self.persist_rollout_items(&rollout_items).await;
if let Err(e) = self.tx_event.send(event).await {
debug!("dropping event because channel is closed: {e}");
}
}
/// Persist the event to the rollout file, flush it, and only then deliver it to clients.
///
/// Most events can be delivered immediately after queueing the rollout write, but some
/// clients (e.g. app-server thread/rollback) re-read the rollout file synchronously on
/// receipt of the event and depend on the marker already being visible on disk.
pub(crate) async fn send_event_raw_flushed(&self, event: Event) {
// Record the last known agent status.
if let Some(status) = agent_status_from_event(&event.msg) {
self.agent_status.send_replace(status);
}
self.persist_rollout_items(&[RolloutItem::EventMsg(event.msg.clone())])
.await;
self.flush_rollout().await;
if let Err(e) = self.tx_event.send(event).await {
debug!("dropping event because channel is closed: {e}");
}
}
pub(crate) async fn emit_turn_item_started(&self, turn_context: &TurnContext, item: &TurnItem) {
self.send_event(
turn_context,
EventMsg::ItemStarted(ItemStartedEvent {
thread_id: self.conversation_id,
turn_id: turn_context.sub_id.clone(),
item: item.clone(),
}),
)
.await;
}
pub(crate) async fn emit_turn_item_completed(
&self,
turn_context: &TurnContext,
item: TurnItem,
) {
record_turn_ttfm_metric(turn_context, &item).await;
self.send_event(
turn_context,
EventMsg::ItemCompleted(ItemCompletedEvent {
thread_id: self.conversation_id,
turn_id: turn_context.sub_id.clone(),
item,
}),
)
.await;
}
/// Adds an execpolicy amendment to both the in-memory and on-disk policies so future
/// commands can use the newly approved prefix.
pub(crate) async fn persist_execpolicy_amendment(
&self,
amendment: &ExecPolicyAmendment,
) -> Result<(), ExecPolicyUpdateError> {
let codex_home = self
.state
.lock()
.await
.session_configuration
.codex_home()
.clone();
self.services
.exec_policy
.append_amendment_and_update(&codex_home, amendment)
.await?;
Ok(())
}
pub(crate) async fn turn_context_for_sub_id(&self, sub_id: &str) -> Option<Arc<TurnContext>> {
let active = self.active_turn.lock().await;
active
.as_ref()
.and_then(|turn| turn.tasks.get(sub_id))
.map(|task| Arc::clone(&task.turn_context))
}
async fn active_turn_context_and_cancellation_token(
&self,
) -> Option<(Arc<TurnContext>, CancellationToken)> {
let active = self.active_turn.lock().await;
let (_, task) = active.as_ref()?.tasks.first()?;
Some((
Arc::clone(&task.turn_context),
task.cancellation_token.child_token(),
))
}
pub(crate) async fn record_execpolicy_amendment_message(
&self,
sub_id: &str,
amendment: &ExecPolicyAmendment,
) {
let Some(prefixes) = format_allow_prefixes(vec![amendment.command.clone()]) else {
warn!("execpolicy amendment for {sub_id} had no command prefix");
return;
};
let text = format!("Approved command prefix saved:\n{prefixes}");
let message: ResponseItem = DeveloperInstructions::new(text.clone()).into();
if let Some(turn_context) = self.turn_context_for_sub_id(sub_id).await {
self.record_conversation_items(&turn_context, std::slice::from_ref(&message))
.await;
return;
}
if self
.inject_response_items(vec![ResponseInputItem::Message {
role: "developer".to_string(),
content: vec![ContentItem::InputText { text }],
}])
.await
.is_err()
{
warn!("no active turn found to record execpolicy amendment message for {sub_id}");
}
}
pub(crate) async fn persist_network_policy_amendment(
&self,
amendment: &NetworkPolicyAmendment,
network_approval_context: &NetworkApprovalContext,
) -> anyhow::Result<()> {
let host =
Self::validated_network_policy_amendment_host(amendment, network_approval_context)?;
let codex_home = self
.state
.lock()
.await
.session_configuration
.codex_home()
.clone();
let execpolicy_amendment =
execpolicy_network_rule_amendment(amendment, network_approval_context, &host);
if let Some(started_network_proxy) = self.services.network_proxy.as_ref() {
let proxy = started_network_proxy.proxy();
match amendment.action {
NetworkPolicyRuleAction::Allow => proxy
.add_allowed_domain(&host)
.await
.map_err(|err| anyhow::anyhow!("failed to update runtime allowlist: {err}"))?,
NetworkPolicyRuleAction::Deny => proxy
.add_denied_domain(&host)
.await
.map_err(|err| anyhow::anyhow!("failed to update runtime denylist: {err}"))?,
}
}
self.services
.exec_policy
.append_network_rule_and_update(
&codex_home,
&host,
execpolicy_amendment.protocol,
execpolicy_amendment.decision,
Some(execpolicy_amendment.justification),
)
.await
.map_err(|err| {
anyhow::anyhow!("failed to persist network policy amendment to execpolicy: {err}")
})?;
Ok(())
}
fn validated_network_policy_amendment_host(
amendment: &NetworkPolicyAmendment,
network_approval_context: &NetworkApprovalContext,
) -> anyhow::Result<String> {
let approved_host = normalize_host(&network_approval_context.host);
let amendment_host = normalize_host(&amendment.host);
if amendment_host != approved_host {
return Err(anyhow::anyhow!(
"network policy amendment host '{}' does not match approved host '{}'",
amendment.host,
network_approval_context.host
));
}
Ok(approved_host)
}
pub(crate) async fn record_network_policy_amendment_message(
&self,
sub_id: &str,
amendment: &NetworkPolicyAmendment,
) {
let (action, list_name) = match amendment.action {
NetworkPolicyRuleAction::Allow => ("Allowed", "allowlist"),
NetworkPolicyRuleAction::Deny => ("Denied", "denylist"),
};
let text = format!(
"{action} network rule saved in execpolicy ({list_name}): {}",
amendment.host
);
let message: ResponseItem = DeveloperInstructions::new(text.clone()).into();
if let Some(turn_context) = self.turn_context_for_sub_id(sub_id).await {
self.record_conversation_items(&turn_context, std::slice::from_ref(&message))
.await;
return;
}
if self
.inject_response_items(vec![ResponseInputItem::Message {
role: "developer".to_string(),
content: vec![ContentItem::InputText { text }],
}])
.await
.is_err()
{
warn!("no active turn found to record network policy amendment message for {sub_id}");
}
}
/// Emit an exec approval request event and await the user's decision.
///
/// The request is keyed by `call_id` + `approval_id` so matching responses
/// are delivered to the correct in-flight turn. If the pending approval is
/// cleared before a response arrives, treat it as an abort so interrupted
/// turns do not continue on a synthetic denial.
///
/// Note that if `available_decisions` is `None`, then the other fields will
/// be used to derive the available decisions via
/// [ExecApprovalRequestEvent::default_available_decisions].
#[allow(clippy::too_many_arguments)]
pub async fn request_command_approval(
&self,
turn_context: &TurnContext,
call_id: String,
approval_id: Option<String>,
command: Vec<String>,
cwd: PathBuf,
reason: Option<String>,
network_approval_context: Option<NetworkApprovalContext>,
proposed_execpolicy_amendment: Option<ExecPolicyAmendment>,
additional_permissions: Option<PermissionProfile>,
skill_metadata: Option<ExecApprovalRequestSkillMetadata>,
available_decisions: Option<Vec<ReviewDecision>>,
) -> ReviewDecision {
// command-level approvals use `call_id`.
// `approval_id` is only present for subcommand callbacks (execve intercept)
let effective_approval_id = approval_id.clone().unwrap_or_else(|| call_id.clone());
// Add the tx_approve callback to the map before sending the request.
let (tx_approve, rx_approve) = oneshot::channel();
let prev_entry = {
let mut active = self.active_turn.lock().await;
match active.as_mut() {
Some(at) => {
let mut ts = at.turn_state.lock().await;
ts.insert_pending_approval(effective_approval_id.clone(), tx_approve)
}
None => None,
}
};
if prev_entry.is_some() {
warn!("Overwriting existing pending approval for call_id: {effective_approval_id}");
}
let parsed_cmd = parse_command(&command);
let proposed_network_policy_amendments = network_approval_context.as_ref().map(|context| {
vec![
NetworkPolicyAmendment {
host: context.host.clone(),
action: NetworkPolicyRuleAction::Allow,
},
NetworkPolicyAmendment {
host: context.host.clone(),
action: NetworkPolicyRuleAction::Deny,
},
]
});
let available_decisions = available_decisions.unwrap_or_else(|| {
ExecApprovalRequestEvent::default_available_decisions(
network_approval_context.as_ref(),
proposed_execpolicy_amendment.as_ref(),
proposed_network_policy_amendments.as_deref(),
additional_permissions.as_ref(),
)
});
let event = EventMsg::ExecApprovalRequest(ExecApprovalRequestEvent {
call_id,
approval_id,
turn_id: turn_context.sub_id.clone(),
command,
cwd,
reason,
network_approval_context,
proposed_execpolicy_amendment,
proposed_network_policy_amendments,
additional_permissions,
skill_metadata,
available_decisions: Some(available_decisions),
parsed_cmd,
});
self.send_event(turn_context, event).await;
rx_approve.await.unwrap_or(ReviewDecision::Abort)
}
pub async fn request_patch_approval(
&self,
turn_context: &TurnContext,
call_id: String,
changes: HashMap<PathBuf, FileChange>,
reason: Option<String>,
grant_root: Option<PathBuf>,
) -> oneshot::Receiver<ReviewDecision> {
// Add the tx_approve callback to the map before sending the request.
let (tx_approve, rx_approve) = oneshot::channel();
let approval_id = call_id.clone();
let prev_entry = {
let mut active = self.active_turn.lock().await;
match active.as_mut() {
Some(at) => {
let mut ts = at.turn_state.lock().await;
ts.insert_pending_approval(approval_id.clone(), tx_approve)
}
None => None,
}
};
if prev_entry.is_some() {
warn!("Overwriting existing pending approval for call_id: {approval_id}");
}
let event = EventMsg::ApplyPatchApprovalRequest(ApplyPatchApprovalRequestEvent {
call_id,
turn_id: turn_context.sub_id.clone(),
changes,
reason,
grant_root,
});
self.send_event(turn_context, event).await;
rx_approve
}
pub async fn request_permissions(
&self,
turn_context: &TurnContext,
call_id: String,
args: RequestPermissionsArgs,
) -> Option<RequestPermissionsResponse> {
let (tx_response, rx_response) = oneshot::channel();
let prev_entry = {
let mut active = self.active_turn.lock().await;
match active.as_mut() {
Some(at) => {
let mut ts = at.turn_state.lock().await;
ts.insert_pending_request_permissions(call_id.clone(), tx_response)
}
None => None,
}
};
if prev_entry.is_some() {
warn!("Overwriting existing pending request_permissions for call_id: {call_id}");
}
let event = EventMsg::RequestPermissions(RequestPermissionsEvent {
call_id,
turn_id: turn_context.sub_id.clone(),
reason: args.reason,
permissions: args.permissions,
});
self.send_event(turn_context, event).await;
rx_response.await.ok()
}
pub async fn request_user_input(
&self,
turn_context: &TurnContext,
call_id: String,
args: RequestUserInputArgs,
) -> Option<RequestUserInputResponse> {
let sub_id = turn_context.sub_id.clone();
let (tx_response, rx_response) = oneshot::channel();
let event_id = sub_id.clone();
let prev_entry = {
let mut active = self.active_turn.lock().await;
match active.as_mut() {
Some(at) => {
let mut ts = at.turn_state.lock().await;
ts.insert_pending_user_input(sub_id, tx_response)
}
None => None,
}
};
if prev_entry.is_some() {
warn!("Overwriting existing pending user input for sub_id: {event_id}");
}
let event = EventMsg::RequestUserInput(RequestUserInputEvent {
call_id,
turn_id: turn_context.sub_id.clone(),
questions: args.questions,
});
self.send_event(turn_context, event).await;
rx_response.await.ok()
}
pub async fn request_mcp_server_elicitation(
&self,
turn_context: &TurnContext,
request_id: RequestId,
params: McpServerElicitationRequestParams,
) -> Option<ElicitationResponse> {
let server_name = params.server_name.clone();
let request = match params.request {
McpServerElicitationRequest::Form {
meta,
message,
requested_schema,
} => {
let requested_schema = match serde_json::to_value(requested_schema) {
Ok(requested_schema) => requested_schema,
Err(err) => {
warn!(
"failed to serialize MCP elicitation schema for server_name: {server_name}, request_id: {request_id}: {err:#}"
);
return None;
}
};
codex_protocol::approvals::ElicitationRequest::Form {
meta,
message,
requested_schema,
}
}
McpServerElicitationRequest::Url {
meta,
message,
url,
elicitation_id,
} => codex_protocol::approvals::ElicitationRequest::Url {
meta,
message,
url,
elicitation_id,
},
};
let (tx_response, rx_response) = oneshot::channel();
let prev_entry = {
let mut active = self.active_turn.lock().await;
match active.as_mut() {
Some(at) => {
let mut ts = at.turn_state.lock().await;
ts.insert_pending_elicitation(
server_name.clone(),
request_id.clone(),
tx_response,
)
}
None => None,
}
};
if prev_entry.is_some() {
warn!(
"Overwriting existing pending elicitation for server_name: {server_name}, request_id: {request_id}"
);
}
let id = match request_id {
rmcp::model::NumberOrString::String(value) => {
codex_protocol::mcp::RequestId::String(value.to_string())
}
rmcp::model::NumberOrString::Number(value) => {
codex_protocol::mcp::RequestId::Integer(value)
}
};
let event = EventMsg::ElicitationRequest(ElicitationRequestEvent {
turn_id: params.turn_id,
server_name,
id,
request,
});
self.send_event(turn_context, event).await;
rx_response.await.ok()
}
pub async fn notify_user_input_response(
&self,
sub_id: &str,
response: RequestUserInputResponse,
) {
let entry = {
let mut active = self.active_turn.lock().await;
match active.as_mut() {
Some(at) => {
let mut ts = at.turn_state.lock().await;
ts.remove_pending_user_input(sub_id)
}
None => None,
}
};
match entry {
Some(tx_response) => {
tx_response.send(response).ok();
}
None => {
warn!("No pending user input found for sub_id: {sub_id}");
}
}
}
pub async fn notify_request_permissions_response(
&self,
call_id: &str,
response: RequestPermissionsResponse,
) {
let entry = {
let mut active = self.active_turn.lock().await;
match active.as_mut() {
Some(at) => {
let mut ts = at.turn_state.lock().await;
let entry = ts.remove_pending_request_permissions(call_id);
if entry.is_some() && !response.permissions.is_empty() {
ts.record_granted_permissions(response.permissions.clone());
}
entry
}
None => None,
}
};
match entry {
Some(tx_response) => {
tx_response.send(response).ok();
}
None => {
warn!("No pending request_permissions found for call_id: {call_id}");
}
}
}
pub(crate) async fn granted_turn_permissions(&self) -> Option<PermissionProfile> {
let active = self.active_turn.lock().await;
let active = active.as_ref()?;
let ts = active.turn_state.lock().await;
ts.granted_permissions()
}
pub async fn notify_dynamic_tool_response(&self, call_id: &str, response: DynamicToolResponse) {
let entry = {
let mut active = self.active_turn.lock().await;
match active.as_mut() {
Some(at) => {
let mut ts = at.turn_state.lock().await;
ts.remove_pending_dynamic_tool(call_id)
}
None => None,
}
};
match entry {
Some(tx_response) => {
tx_response.send(response).ok();
}
None => {
warn!("No pending dynamic tool call found for call_id: {call_id}");
}
}
}
pub async fn notify_approval(&self, approval_id: &str, decision: ReviewDecision) {
let entry = {
let mut active = self.active_turn.lock().await;
match active.as_mut() {
Some(at) => {
let mut ts = at.turn_state.lock().await;
ts.remove_pending_approval(approval_id)
}
None => None,
}
};
match entry {
Some(tx_approve) => {
tx_approve.send(decision).ok();
}
None => {
warn!("No pending approval found for call_id: {approval_id}");
}
}
}
pub async fn resolve_elicitation(
&self,
server_name: String,
id: RequestId,
response: ElicitationResponse,
) -> anyhow::Result<()> {
let entry = {
let mut active = self.active_turn.lock().await;
match active.as_mut() {
Some(at) => {
let mut ts = at.turn_state.lock().await;
ts.remove_pending_elicitation(&server_name, &id)
}
None => None,
}
};
if let Some(tx_response) = entry {
tx_response
.send(response)
.map_err(|e| anyhow::anyhow!("failed to send elicitation response: {e:?}"))?;
return Ok(());
}
self.services
.mcp_connection_manager
.read()
.await
.resolve_elicitation(server_name, id, response)
.await
}
/// Records input items: always append to conversation history and
/// persist these response items to rollout.
pub(crate) async fn record_conversation_items(
&self,
turn_context: &TurnContext,
items: &[ResponseItem],
) {
self.record_into_history(items, turn_context).await;
self.persist_rollout_response_items(items).await;
self.send_raw_response_items(turn_context, items).await;
}
/// Append ResponseItems to the in-memory conversation history only.
pub(crate) async fn record_into_history(
&self,
items: &[ResponseItem],
turn_context: &TurnContext,
) {
let mut state = self.state.lock().await;
state.record_items(items.iter(), turn_context.truncation_policy);
}
pub(crate) async fn record_model_warning(&self, message: impl Into<String>, ctx: &TurnContext) {
self.services
.session_telemetry
.counter("codex.model_warning", 1, &[]);
let item = ResponseItem::Message {
id: None,
role: "user".to_string(),
content: vec![ContentItem::InputText {
text: format!("Warning: {}", message.into()),
}],
end_turn: None,
phase: None,
};
self.record_conversation_items(ctx, &[item]).await;
}
async fn maybe_warn_on_server_model_mismatch(
self: &Arc<Self>,
turn_context: &Arc<TurnContext>,
server_model: String,
) -> bool {
let requested_model = turn_context.model_info.slug.clone();
let server_model_normalized = server_model.to_ascii_lowercase();
let requested_model_normalized = requested_model.to_ascii_lowercase();
if server_model_normalized == requested_model_normalized {
info!("server reported model {server_model} (matches requested model)");
return false;
}
warn!("server reported model {server_model} while requested model was {requested_model}");
let warning_message = format!(
"Your account was flagged for potentially high-risk cyber activity and this request was routed to gpt-5.2 as a fallback. To regain access to gpt-5.3-codex, apply for trusted access: {CYBER_VERIFY_URL} or learn more: {CYBER_SAFETY_URL}"
);
self.send_event(
turn_context,
EventMsg::ModelReroute(ModelRerouteEvent {
from_model: requested_model.clone(),
to_model: server_model.clone(),
reason: ModelRerouteReason::HighRiskCyberActivity,
}),
)
.await;
self.send_event(
turn_context,
EventMsg::Warning(WarningEvent {
message: warning_message.clone(),
}),
)
.await;
self.record_model_warning(warning_message, turn_context)
.await;
true
}
pub(crate) async fn replace_history(
&self,
items: Vec<ResponseItem>,
reference_context_item: Option<TurnContextItem>,
) {
let mut state = self.state.lock().await;
state.replace_history(items, reference_context_item);
}
pub(crate) async fn replace_compacted_history(
&self,
items: Vec<ResponseItem>,
reference_context_item: Option<TurnContextItem>,
compacted_item: CompactedItem,
) {
self.replace_history(items, reference_context_item.clone())
.await;
self.persist_rollout_items(&[RolloutItem::Compacted(compacted_item)])
.await;
if let Some(turn_context_item) = reference_context_item {
self.persist_rollout_items(&[RolloutItem::TurnContext(turn_context_item)])
.await;
}
}
async fn persist_rollout_response_items(&self, items: &[ResponseItem]) {
let rollout_items: Vec<RolloutItem> = items
.iter()
.cloned()
.map(RolloutItem::ResponseItem)
.collect();
self.persist_rollout_items(&rollout_items).await;
}
pub fn enabled(&self, feature: Feature) -> bool {
self.features.enabled(feature)
}
pub(crate) fn features(&self) -> ManagedFeatures {
self.features.clone()
}
pub(crate) async fn collaboration_mode(&self) -> CollaborationMode {
let state = self.state.lock().await;
state.session_configuration.collaboration_mode.clone()
}
async fn send_raw_response_items(&self, turn_context: &TurnContext, items: &[ResponseItem]) {
for item in items {
self.send_event(
turn_context,
EventMsg::RawResponseItem(RawResponseItemEvent { item: item.clone() }),
)
.await;
}
}
pub(crate) async fn build_initial_context(
&self,
turn_context: &TurnContext,
) -> Vec<ResponseItem> {
let mut developer_sections = Vec::<String>::with_capacity(8);
let mut contextual_user_sections = Vec::<String>::with_capacity(2);
let shell = self.user_shell();
let (reference_context_item, previous_turn_settings, collaboration_mode, base_instructions) = {
let state = self.state.lock().await;
(
state.reference_context_item(),
state.previous_turn_settings(),
state.session_configuration.collaboration_mode.clone(),
state.session_configuration.base_instructions.clone(),
)
};
if let Some(model_switch_message) =
crate::context_manager::updates::build_model_instructions_update_item(
previous_turn_settings.as_ref(),
turn_context,
)
{
developer_sections.push(model_switch_message.into_text());
}
developer_sections.push(
DeveloperInstructions::from_policy(
turn_context.sandbox_policy.get(),
turn_context.approval_policy.value(),
self.services.exec_policy.current().as_ref(),
&turn_context.cwd,
turn_context.features.enabled(Feature::RequestPermissions),
)
.into_text(),
);
if let Some(developer_instructions) = turn_context.developer_instructions.as_deref() {
developer_sections.push(developer_instructions.to_string());
}
// Add developer instructions for memories.
if turn_context.features.enabled(Feature::MemoryTool)
&& turn_context.config.memories.use_memories
&& let Some(memory_prompt) =
build_memory_tool_developer_instructions(&turn_context.config.codex_home).await
{
developer_sections.push(memory_prompt);
}
// Add developer instructions from collaboration_mode if they exist and are non-empty
if let Some(collab_instructions) =
DeveloperInstructions::from_collaboration_mode(&collaboration_mode)
{
developer_sections.push(collab_instructions.into_text());
}
if let Some(realtime_update) = crate::context_manager::updates::build_initial_realtime_item(
reference_context_item.as_ref(),
previous_turn_settings.as_ref(),
turn_context,
) {
developer_sections.push(realtime_update.into_text());
}
if self.features.enabled(Feature::Personality)
&& let Some(personality) = turn_context.personality
{
let model_info = turn_context.model_info.clone();
let has_baked_personality = model_info.supports_personality()
&& base_instructions == model_info.get_model_instructions(Some(personality));
if !has_baked_personality
&& let Some(personality_message) =
crate::context_manager::updates::personality_message_for(
&model_info,
personality,
)
{
developer_sections.push(
DeveloperInstructions::personality_spec_message(personality_message)
.into_text(),
);
}
}
if turn_context.features.enabled(Feature::Apps) {
developer_sections.push(render_apps_section());
}
if turn_context.features.enabled(Feature::CodexGitCommit)
&& let Some(commit_message_instruction) = commit_message_trailer_instruction(
turn_context.config.commit_attribution.as_deref(),
)
{
developer_sections.push(commit_message_instruction);
}
if let Some(user_instructions) = turn_context.user_instructions.as_deref() {
contextual_user_sections.push(
UserInstructions {
text: user_instructions.to_string(),
directory: turn_context.cwd.to_string_lossy().into_owned(),
}
.serialize_to_text(),
);
}
let subagents = self
.services
.agent_control
.format_environment_context_subagents(self.conversation_id)
.await;
contextual_user_sections.push(
EnvironmentContext::from_turn_context(turn_context, shell.as_ref())
.with_subagents(subagents)
.serialize_to_xml(),
);
let mut items = Vec::with_capacity(2);
if let Some(developer_message) =
crate::context_manager::updates::build_developer_update_item(developer_sections)
{
items.push(developer_message);
}
if let Some(contextual_user_message) =
crate::context_manager::updates::build_contextual_user_message(contextual_user_sections)
{
items.push(contextual_user_message);
}
items
}
pub(crate) async fn persist_rollout_items(&self, items: &[RolloutItem]) {
let recorder = {
let guard = self.services.rollout.lock().await;
guard.clone()
};
if let Some(rec) = recorder
&& let Err(e) = rec.record_items(items).await
{
error!("failed to record rollout items: {e:#}");
}
}
pub(crate) async fn clone_history(&self) -> ContextManager {
let state = self.state.lock().await;
state.clone_history()
}
pub(crate) async fn reference_context_item(&self) -> Option<TurnContextItem> {
let state = self.state.lock().await;
state.reference_context_item()
}
/// Persist the latest turn context snapshot for the first real user turn and for
/// steady-state turns that emit model-visible context updates.
///
/// When the reference snapshot is missing, this injects full initial context. Otherwise, it
/// emits only settings diff items.
///
/// If full context is injected and a model switch occurred, this prepends the
/// `<model_switch>` developer message so model-specific instructions are not lost.
///
/// This is the normal runtime path that establishes a new `reference_context_item`.
/// Mid-turn compaction is the other path that can re-establish that baseline when it
/// reinjects full initial context into replacement history. Other non-regular tasks
/// intentionally do not update the baseline.
pub(crate) async fn record_context_updates_and_set_reference_context_item(
&self,
turn_context: &TurnContext,
) {
let reference_context_item = {
let state = self.state.lock().await;
state.reference_context_item()
};
let should_inject_full_context = reference_context_item.is_none();
let context_items = if should_inject_full_context {
self.build_initial_context(turn_context).await
} else {
// Steady-state path: append only context diffs to minimize token overhead.
self.build_settings_update_items(reference_context_item.as_ref(), turn_context)
.await
};
let turn_context_item = turn_context.to_turn_context_item();
if !context_items.is_empty() {
self.record_conversation_items(turn_context, &context_items)
.await;
}
// Persist one `TurnContextItem` per real user turn so resume/lazy replay can recover the
// latest durable baseline even when this turn emitted no model-visible context diffs.
self.persist_rollout_items(&[RolloutItem::TurnContext(turn_context_item.clone())])
.await;
// Advance the in-memory diff baseline even when this turn emitted no model-visible
// context items. This keeps later runtime diffing aligned with the current turn state.
let mut state = self.state.lock().await;
state.set_reference_context_item(Some(turn_context_item));
}
pub(crate) async fn update_token_usage_info(
&self,
turn_context: &TurnContext,
token_usage: Option<&TokenUsage>,
) {
if let Some(token_usage) = token_usage {
let mut state = self.state.lock().await;
state.update_token_info_from_usage(token_usage, turn_context.model_context_window());
}
self.send_token_count_event(turn_context).await;
}
pub(crate) async fn recompute_token_usage(&self, turn_context: &TurnContext) {
let history = self.clone_history().await;
let base_instructions = self.get_base_instructions().await;
let Some(estimated_total_tokens) =
history.estimate_token_count_with_base_instructions(&base_instructions)
else {
return;
};
{
let mut state = self.state.lock().await;
let mut info = state.token_info().unwrap_or(TokenUsageInfo {
total_token_usage: TokenUsage::default(),
last_token_usage: TokenUsage::default(),
model_context_window: None,
});
info.last_token_usage = TokenUsage {
input_tokens: 0,
cached_input_tokens: 0,
output_tokens: 0,
reasoning_output_tokens: 0,
total_tokens: estimated_total_tokens.max(0),
};
if let Some(model_context_window) = turn_context.model_context_window() {
info.model_context_window = Some(model_context_window);
}
state.set_token_info(Some(info));
}
self.send_token_count_event(turn_context).await;
}
pub(crate) async fn update_rate_limits(
&self,
turn_context: &TurnContext,
new_rate_limits: RateLimitSnapshot,
) {
{
let mut state = self.state.lock().await;
state.set_rate_limits(new_rate_limits);
}
self.send_token_count_event(turn_context).await;
}
pub(crate) async fn mcp_dependency_prompted(&self) -> HashSet<String> {
let state = self.state.lock().await;
state.mcp_dependency_prompted()
}
pub(crate) async fn record_mcp_dependency_prompted<I>(&self, names: I)
where
I: IntoIterator<Item = String>,
{
let mut state = self.state.lock().await;
state.record_mcp_dependency_prompted(names);
}
pub async fn dependency_env(&self) -> HashMap<String, String> {
let state = self.state.lock().await;
state.dependency_env()
}
pub async fn set_dependency_env(&self, values: HashMap<String, String>) {
let mut state = self.state.lock().await;
state.set_dependency_env(values);
}
pub(crate) async fn set_server_reasoning_included(&self, included: bool) {
let mut state = self.state.lock().await;
state.set_server_reasoning_included(included);
}
async fn send_token_count_event(&self, turn_context: &TurnContext) {
let (info, rate_limits) = {
let state = self.state.lock().await;
state.token_info_and_rate_limits()
};
let event = EventMsg::TokenCount(TokenCountEvent { info, rate_limits });
self.send_event(turn_context, event).await;
}
pub(crate) async fn set_total_tokens_full(&self, turn_context: &TurnContext) {
if let Some(context_window) = turn_context.model_context_window() {
let mut state = self.state.lock().await;
state.set_token_usage_full(context_window);
}
self.send_token_count_event(turn_context).await;
}
pub(crate) async fn record_response_item_and_emit_turn_item(
&self,
turn_context: &TurnContext,
response_item: ResponseItem,
) {
// Add to conversation history and persist response item to rollout.
self.record_conversation_items(turn_context, std::slice::from_ref(&response_item))
.await;
// Derive a turn item and emit lifecycle events if applicable.
if let Some(item) = parse_turn_item(&response_item) {
self.emit_turn_item_started(turn_context, &item).await;
self.emit_turn_item_completed(turn_context, item).await;
}
}
pub(crate) async fn record_user_prompt_and_emit_turn_item(
&self,
turn_context: &TurnContext,
input: &[UserInput],
response_item: ResponseItem,
) {
// Persist the user message to history, but emit the turn item from `UserInput` so
// UI-only `text_elements` are preserved. `ResponseItem::Message` does not carry
// those spans, and `record_response_item_and_emit_turn_item` would drop them.
self.record_conversation_items(turn_context, std::slice::from_ref(&response_item))
.await;
let turn_item = TurnItem::UserMessage(UserMessageItem::new(input));
self.emit_turn_item_started(turn_context, &turn_item).await;
self.emit_turn_item_completed(turn_context, turn_item).await;
self.ensure_rollout_materialized().await;
}
pub(crate) async fn notify_background_event(
&self,
turn_context: &TurnContext,
message: impl Into<String>,
) {
let event = EventMsg::BackgroundEvent(BackgroundEventEvent {
message: message.into(),
});
self.send_event(turn_context, event).await;
}
pub(crate) async fn notify_stream_error(
&self,
turn_context: &TurnContext,
message: impl Into<String>,
codex_error: CodexErr,
) {
let additional_details = codex_error.to_string();
let codex_error_info = CodexErrorInfo::ResponseStreamDisconnected {
http_status_code: codex_error.http_status_code_value(),
};
let event = EventMsg::StreamError(StreamErrorEvent {
message: message.into(),
codex_error_info: Some(codex_error_info),
additional_details: Some(additional_details),
});
self.send_event(turn_context, event).await;
}
async fn maybe_start_ghost_snapshot(
self: &Arc<Self>,
turn_context: Arc<TurnContext>,
cancellation_token: CancellationToken,
) {
if !self.enabled(Feature::GhostCommit) {
return;
}
let token = match turn_context.tool_call_gate.subscribe().await {
Ok(token) => token,
Err(err) => {
warn!("failed to subscribe to ghost snapshot readiness: {err}");
return;
}
};
info!("spawning ghost snapshot task");
let task = GhostSnapshotTask::new(token);
Arc::new(task)
.run(
Arc::new(SessionTaskContext::new(self.clone())),
turn_context.clone(),
Vec::new(),
cancellation_token,
)
.await;
}
/// Inject additional user input into the currently active turn.
///
/// Returns the active turn id when accepted.
pub async fn steer_input(
&self,
input: Vec<UserInput>,
expected_turn_id: Option<&str>,
) -> Result<String, SteerInputError> {
if input.is_empty() {
return Err(SteerInputError::EmptyInput);
}
let mut active = self.active_turn.lock().await;
let Some(active_turn) = active.as_mut() else {
return Err(SteerInputError::NoActiveTurn(input));
};
let Some((active_turn_id, _)) = active_turn.tasks.first() else {
return Err(SteerInputError::NoActiveTurn(input));
};
if let Some(expected_turn_id) = expected_turn_id
&& expected_turn_id != active_turn_id
{
return Err(SteerInputError::ExpectedTurnMismatch {
expected: expected_turn_id.to_string(),
actual: active_turn_id.clone(),
});
}
let mut turn_state = active_turn.turn_state.lock().await;
turn_state.push_pending_input(input.into());
Ok(active_turn_id.clone())
}
/// Returns the input if there was no task running to inject into
pub async fn inject_response_items(
&self,
input: Vec<ResponseInputItem>,
) -> Result<(), Vec<ResponseInputItem>> {
let mut active = self.active_turn.lock().await;
match active.as_mut() {
Some(at) => {
let mut ts = at.turn_state.lock().await;
for item in input {
ts.push_pending_input(item);
}
Ok(())
}
None => Err(input),
}
}
pub async fn get_pending_input(&self) -> Vec<ResponseInputItem> {
let mut active = self.active_turn.lock().await;
match active.as_mut() {
Some(at) => {
let mut ts = at.turn_state.lock().await;
ts.take_pending_input()
}
None => Vec::with_capacity(0),
}
}
pub async fn has_pending_input(&self) -> bool {
let active = self.active_turn.lock().await;
match active.as_ref() {
Some(at) => {
let ts = at.turn_state.lock().await;
ts.has_pending_input()
}
None => false,
}
}
pub async fn list_resources(
&self,
server: &str,
params: Option<PaginatedRequestParams>,
) -> anyhow::Result<ListResourcesResult> {
self.services
.mcp_connection_manager
.read()
.await
.list_resources(server, params)
.await
}
pub async fn list_resource_templates(
&self,
server: &str,
params: Option<PaginatedRequestParams>,
) -> anyhow::Result<ListResourceTemplatesResult> {
self.services
.mcp_connection_manager
.read()
.await
.list_resource_templates(server, params)
.await
}
pub async fn read_resource(
&self,
server: &str,
params: ReadResourceRequestParams,
) -> anyhow::Result<ReadResourceResult> {
self.services
.mcp_connection_manager
.read()
.await
.read_resource(server, params)
.await
}
pub async fn call_tool(
&self,
server: &str,
tool: &str,
arguments: Option<serde_json::Value>,
) -> anyhow::Result<CallToolResult> {
self.services
.mcp_connection_manager
.read()
.await
.call_tool(server, tool, arguments)
.await
}
pub(crate) async fn parse_mcp_tool_name(&self, tool_name: &str) -> Option<(String, String)> {
self.services
.mcp_connection_manager
.read()
.await
.parse_tool_name(tool_name)
.await
}
pub async fn interrupt_task(self: &Arc<Self>) {
info!("interrupt received: abort current task, if any");
let has_active_turn = { self.active_turn.lock().await.is_some() };
if has_active_turn {
self.abort_all_tasks(TurnAbortReason::Interrupted).await;
} else {
self.cancel_mcp_startup().await;
}
}
pub(crate) fn hooks(&self) -> &Hooks {
&self.services.hooks
}
pub(crate) fn user_shell(&self) -> Arc<shell::Shell> {
Arc::clone(&self.services.user_shell)
}
async fn refresh_mcp_servers_inner(
&self,
turn_context: &TurnContext,
mcp_servers: HashMap<String, McpServerConfig>,
store_mode: OAuthCredentialsStoreMode,
) {
let auth = self.services.auth_manager.auth().await;
let config = self.get_config().await;
let tool_plugin_provenance = self
.services
.mcp_manager
.tool_plugin_provenance(config.as_ref());
let mcp_servers = with_codex_apps_mcp(
mcp_servers,
self.features.enabled(Feature::Apps),
auth.as_ref(),
config.as_ref(),
);
let auth_statuses = compute_auth_statuses(mcp_servers.iter(), store_mode).await;
let sandbox_state = SandboxState {
sandbox_policy: turn_context.sandbox_policy.get().clone(),
codex_linux_sandbox_exe: turn_context.codex_linux_sandbox_exe.clone(),
sandbox_cwd: turn_context.cwd.clone(),
use_linux_sandbox_bwrap: turn_context.features.enabled(Feature::UseLinuxSandboxBwrap),
};
{
let mut guard = self.services.mcp_startup_cancellation_token.lock().await;
guard.cancel();
*guard = CancellationToken::new();
}
let (refreshed_manager, cancel_token) = McpConnectionManager::new(
&mcp_servers,
store_mode,
auth_statuses,
&turn_context.config.permissions.approval_policy,
self.get_tx_event(),
sandbox_state,
config.codex_home.clone(),
codex_apps_tools_cache_key(auth.as_ref()),
tool_plugin_provenance,
)
.await;
{
let mut guard = self.services.mcp_startup_cancellation_token.lock().await;
if guard.is_cancelled() {
cancel_token.cancel();
}
*guard = cancel_token;
}
let mut manager = self.services.mcp_connection_manager.write().await;
*manager = refreshed_manager;
}
async fn refresh_mcp_servers_if_requested(&self, turn_context: &TurnContext) {
let refresh_config = { self.pending_mcp_server_refresh_config.lock().await.take() };
let Some(refresh_config) = refresh_config else {
return;
};
let McpServerRefreshConfig {
mcp_servers,
mcp_oauth_credentials_store_mode,
} = refresh_config;
let mcp_servers =
match serde_json::from_value::<HashMap<String, McpServerConfig>>(mcp_servers) {
Ok(servers) => servers,
Err(err) => {
warn!("failed to parse MCP server refresh config: {err}");
return;
}
};
let store_mode = match serde_json::from_value::<OAuthCredentialsStoreMode>(
mcp_oauth_credentials_store_mode,
) {
Ok(mode) => mode,
Err(err) => {
warn!("failed to parse MCP OAuth refresh config: {err}");
return;
}
};
self.refresh_mcp_servers_inner(turn_context, mcp_servers, store_mode)
.await;
}
pub(crate) async fn refresh_mcp_servers_now(
&self,
turn_context: &TurnContext,
mcp_servers: HashMap<String, McpServerConfig>,
store_mode: OAuthCredentialsStoreMode,
) {
self.refresh_mcp_servers_inner(turn_context, mcp_servers, store_mode)
.await;
}
#[cfg(test)]
async fn mcp_startup_cancellation_token(&self) -> CancellationToken {
self.services
.mcp_startup_cancellation_token
.lock()
.await
.clone()
}
fn show_raw_agent_reasoning(&self) -> bool {
self.services.show_raw_agent_reasoning
}
async fn cancel_mcp_startup(&self) {
self.services
.mcp_startup_cancellation_token
.lock()
.await
.cancel();
}
}
async fn submission_loop(sess: Arc<Session>, config: Arc<Config>, rx_sub: Receiver<Submission>) {
// To break out of this loop, send Op::Shutdown.
while let Ok(sub) = rx_sub.recv().await {
debug!(?sub, "Submission");
let dispatch_span = submission_dispatch_span(&sub);
let should_exit = async {
match sub.op.clone() {
Op::Interrupt => {
handlers::interrupt(&sess).await;
false
}
Op::CleanBackgroundTerminals => {
handlers::clean_background_terminals(&sess).await;
false
}
Op::RealtimeConversationStart(params) => {
if let Err(err) =
handle_realtime_conversation_start(&sess, sub.id.clone(), params).await
{
sess.send_event_raw(Event {
id: sub.id.clone(),
msg: EventMsg::Error(ErrorEvent {
message: err.to_string(),
codex_error_info: Some(CodexErrorInfo::Other),
}),
})
.await;
}
false
}
Op::RealtimeConversationAudio(params) => {
handle_realtime_conversation_audio(&sess, sub.id.clone(), params).await;
false
}
Op::RealtimeConversationText(params) => {
handle_realtime_conversation_text(&sess, sub.id.clone(), params).await;
false
}
Op::RealtimeConversationClose => {
handle_realtime_conversation_close(&sess, sub.id.clone()).await;
false
}
Op::OverrideTurnContext {
cwd,
approval_policy,
sandbox_policy,
windows_sandbox_level,
model,
effort,
summary,
service_tier,
collaboration_mode,
personality,
} => {
let collaboration_mode = if let Some(collab_mode) = collaboration_mode {
collab_mode
} else {
let state = sess.state.lock().await;
state.session_configuration.collaboration_mode.with_updates(
model.clone(),
effort,
None,
)
};
handlers::override_turn_context(
&sess,
sub.id.clone(),
SessionSettingsUpdate {
cwd,
approval_policy,
sandbox_policy,
windows_sandbox_level,
collaboration_mode: Some(collaboration_mode),
reasoning_summary: summary,
service_tier,
personality,
..Default::default()
},
)
.await;
false
}
Op::UserInput { .. } | Op::UserTurn { .. } => {
handlers::user_input_or_turn(&sess, sub.id.clone(), sub.op).await;
false
}
Op::ExecApproval {
id: approval_id,
turn_id,
decision,
} => {
handlers::exec_approval(&sess, approval_id, turn_id, decision).await;
false
}
Op::PatchApproval { id, decision } => {
handlers::patch_approval(&sess, id, decision).await;
false
}
Op::UserInputAnswer { id, response } => {
handlers::request_user_input_response(&sess, id, response).await;
false
}
Op::RequestPermissionsResponse { id, response } => {
handlers::request_permissions_response(&sess, id, response).await;
false
}
Op::DynamicToolResponse { id, response } => {
handlers::dynamic_tool_response(&sess, id, response).await;
false
}
Op::AddToHistory { text } => {
handlers::add_to_history(&sess, &config, text).await;
false
}
Op::GetHistoryEntryRequest { offset, log_id } => {
handlers::get_history_entry_request(
&sess,
&config,
sub.id.clone(),
offset,
log_id,
)
.await;
false
}
Op::ListMcpTools => {
handlers::list_mcp_tools(&sess, &config, sub.id.clone()).await;
false
}
Op::RefreshMcpServers { config } => {
handlers::refresh_mcp_servers(&sess, config).await;
false
}
Op::ReloadUserConfig => {
handlers::reload_user_config(&sess).await;
false
}
Op::ListCustomPrompts => {
handlers::list_custom_prompts(&sess, sub.id.clone()).await;
false
}
Op::ListSkills { cwds, force_reload } => {
handlers::list_skills(&sess, sub.id.clone(), cwds, force_reload).await;
false
}
Op::ListRemoteSkills {
hazelnut_scope,
product_surface,
enabled,
} => {
handlers::list_remote_skills(
&sess,
&config,
sub.id.clone(),
hazelnut_scope,
product_surface,
enabled,
)
.await;
false
}
Op::DownloadRemoteSkill { hazelnut_id } => {
handlers::export_remote_skill(&sess, &config, sub.id.clone(), hazelnut_id)
.await;
false
}
Op::Undo => {
handlers::undo(&sess, sub.id.clone()).await;
false
}
Op::Compact => {
handlers::compact(&sess, sub.id.clone()).await;
false
}
Op::DropMemories => {
handlers::drop_memories(&sess, &config, sub.id.clone()).await;
false
}
Op::UpdateMemories => {
handlers::update_memories(&sess, &config, sub.id.clone()).await;
false
}
Op::ThreadRollback { num_turns } => {
handlers::thread_rollback(&sess, sub.id.clone(), num_turns).await;
false
}
Op::SetThreadName { name } => {
handlers::set_thread_name(&sess, sub.id.clone(), name).await;
false
}
Op::RunUserShellCommand { command } => {
handlers::run_user_shell_command(&sess, sub.id.clone(), command).await;
false
}
Op::ResolveElicitation {
server_name,
request_id,
decision,
content,
meta,
} => {
handlers::resolve_elicitation(
&sess,
server_name,
request_id,
decision,
content,
meta,
)
.await;
false
}
Op::Shutdown => handlers::shutdown(&sess, sub.id.clone()).await,
Op::Review { review_request } => {
handlers::review(&sess, &config, sub.id.clone(), review_request).await;
false
}
_ => false, // Ignore unknown ops; enum is non_exhaustive to allow extensions.
}
}
.instrument(dispatch_span)
.await;
if should_exit {
break;
}
}
debug!("Agent loop exited");
}
fn submission_dispatch_span(sub: &Submission) -> tracing::Span {
let dispatch_span = match &sub.op {
Op::RealtimeConversationAudio(_) => {
debug_span!("submission_dispatch", submission.id = sub.id.as_str())
}
_ => info_span!("submission_dispatch", submission.id = sub.id.as_str()),
};
if let Some(trace) = sub.trace.as_ref()
&& !set_parent_from_w3c_trace_context(&dispatch_span, trace)
{
warn!(
submission.id = sub.id.as_str(),
"ignoring invalid submission trace carrier"
);
}
dispatch_span
}
/// Operation handlers
mod handlers {
use crate::codex::Session;
use crate::codex::SessionSettingsUpdate;
use crate::codex::SteerInputError;
use crate::codex::spawn_review_thread;
use crate::config::Config;
use crate::mcp::auth::compute_auth_statuses;
use crate::mcp::collect_mcp_snapshot_from_manager;
use crate::review_prompts::resolve_review_request;
use crate::rollout::RolloutRecorder;
use crate::rollout::session_index;
use crate::tasks::CompactTask;
use crate::tasks::UndoTask;
use crate::tasks::UserShellCommandMode;
use crate::tasks::UserShellCommandTask;
use crate::tasks::execute_user_shell_command;
use codex_protocol::custom_prompts::CustomPrompt;
use codex_protocol::protocol::CodexErrorInfo;
use codex_protocol::protocol::ErrorEvent;
use codex_protocol::protocol::Event;
use codex_protocol::protocol::EventMsg;
use codex_protocol::protocol::ListCustomPromptsResponseEvent;
use codex_protocol::protocol::ListRemoteSkillsResponseEvent;
use codex_protocol::protocol::ListSkillsResponseEvent;
use codex_protocol::protocol::McpServerRefreshConfig;
use codex_protocol::protocol::Op;
use codex_protocol::protocol::RemoteSkillDownloadedEvent;
use codex_protocol::protocol::RemoteSkillHazelnutScope;
use codex_protocol::protocol::RemoteSkillProductSurface;
use codex_protocol::protocol::RemoteSkillSummary;
use codex_protocol::protocol::ReviewDecision;
use codex_protocol::protocol::ReviewRequest;
use codex_protocol::protocol::RolloutItem;
use codex_protocol::protocol::SkillsListEntry;
use codex_protocol::protocol::ThreadNameUpdatedEvent;
use codex_protocol::protocol::ThreadRolledBackEvent;
use codex_protocol::protocol::TurnAbortReason;
use codex_protocol::protocol::WarningEvent;
use codex_protocol::request_permissions::RequestPermissionsResponse;
use codex_protocol::request_user_input::RequestUserInputResponse;
use crate::context_manager::is_user_turn_boundary;
use codex_protocol::config_types::CollaborationMode;
use codex_protocol::config_types::ModeKind;
use codex_protocol::config_types::Settings;
use codex_protocol::dynamic_tools::DynamicToolResponse;
use codex_protocol::mcp::RequestId as ProtocolRequestId;
use codex_protocol::user_input::UserInput;
use codex_rmcp_client::ElicitationAction;
use codex_rmcp_client::ElicitationResponse;
use serde_json::Value;
use std::path::PathBuf;
use std::sync::Arc;
use tracing::info;
use tracing::warn;
pub async fn interrupt(sess: &Arc<Session>) {
sess.interrupt_task().await;
}
pub async fn clean_background_terminals(sess: &Arc<Session>) {
sess.close_unified_exec_processes().await;
}
pub async fn override_turn_context(
sess: &Session,
sub_id: String,
updates: SessionSettingsUpdate,
) {
if let Err(err) = sess.update_settings(updates).await {
sess.send_event_raw(Event {
id: sub_id,
msg: EventMsg::Error(ErrorEvent {
message: err.to_string(),
codex_error_info: Some(CodexErrorInfo::BadRequest),
}),
})
.await;
}
}
pub async fn user_input_or_turn(sess: &Arc<Session>, sub_id: String, op: Op) {
let (items, updates) = match op {
Op::UserTurn {
cwd,
approval_policy,
sandbox_policy,
model,
effort,
summary,
service_tier,
final_output_json_schema,
items,
collaboration_mode,
personality,
} => {
let collaboration_mode = collaboration_mode.or_else(|| {
Some(CollaborationMode {
mode: ModeKind::Default,
settings: Settings {
model: model.clone(),
reasoning_effort: effort,
developer_instructions: None,
},
})
});
(
items,
SessionSettingsUpdate {
cwd: Some(cwd),
approval_policy: Some(approval_policy),
sandbox_policy: Some(sandbox_policy),
windows_sandbox_level: None,
collaboration_mode,
reasoning_summary: summary,
service_tier,
final_output_json_schema: Some(final_output_json_schema),
personality,
app_server_client_name: None,
},
)
}
Op::UserInput {
items,
final_output_json_schema,
} => (
items,
SessionSettingsUpdate {
final_output_json_schema: Some(final_output_json_schema),
..Default::default()
},
),
_ => unreachable!(),
};
let Ok(current_context) = sess.new_turn_with_sub_id(sub_id, updates).await else {
// new_turn_with_sub_id already emits the error event.
return;
};
sess.maybe_emit_unknown_model_warning_for_turn(current_context.as_ref())
.await;
current_context.session_telemetry.user_prompt(&items);
// Attempt to inject input into current task.
if let Err(SteerInputError::NoActiveTurn(items)) = sess.steer_input(items, None).await {
sess.refresh_mcp_servers_if_requested(&current_context)
.await;
let regular_task = sess.take_startup_regular_task().await.unwrap_or_default();
sess.spawn_task(Arc::clone(&current_context), items, regular_task)
.await;
}
}
pub async fn run_user_shell_command(sess: &Arc<Session>, sub_id: String, command: String) {
if let Some((turn_context, cancellation_token)) =
sess.active_turn_context_and_cancellation_token().await
{
let session = Arc::clone(sess);
tokio::spawn(async move {
execute_user_shell_command(
session,
turn_context,
command,
cancellation_token,
UserShellCommandMode::ActiveTurnAuxiliary,
)
.await;
});
return;
}
let turn_context = sess.new_default_turn_with_sub_id(sub_id).await;
sess.spawn_task(
Arc::clone(&turn_context),
Vec::new(),
UserShellCommandTask::new(command),
)
.await;
}
pub async fn resolve_elicitation(
sess: &Arc<Session>,
server_name: String,
request_id: ProtocolRequestId,
decision: codex_protocol::approvals::ElicitationAction,
content: Option<Value>,
meta: Option<Value>,
) {
let action = match decision {
codex_protocol::approvals::ElicitationAction::Accept => ElicitationAction::Accept,
codex_protocol::approvals::ElicitationAction::Decline => ElicitationAction::Decline,
codex_protocol::approvals::ElicitationAction::Cancel => ElicitationAction::Cancel,
};
let content = match action {
// Preserve the legacy fallback for clients that only send an action.
ElicitationAction::Accept => Some(content.unwrap_or_else(|| serde_json::json!({}))),
ElicitationAction::Decline | ElicitationAction::Cancel => None,
};
let response = ElicitationResponse {
action,
content,
meta,
};
let request_id = match request_id {
ProtocolRequestId::String(value) => {
rmcp::model::NumberOrString::String(std::sync::Arc::from(value))
}
ProtocolRequestId::Integer(value) => rmcp::model::NumberOrString::Number(value),
};
if let Err(err) = sess
.resolve_elicitation(server_name, request_id, response)
.await
{
warn!(
error = %err,
"failed to resolve elicitation request in session"
);
}
}
/// Propagate a user's exec approval decision to the session.
/// Also optionally applies an execpolicy amendment.
pub async fn exec_approval(
sess: &Arc<Session>,
approval_id: String,
turn_id: Option<String>,
decision: ReviewDecision,
) {
let event_turn_id = turn_id.unwrap_or_else(|| approval_id.clone());
if let ReviewDecision::ApprovedExecpolicyAmendment {
proposed_execpolicy_amendment,
} = &decision
{
match sess
.persist_execpolicy_amendment(proposed_execpolicy_amendment)
.await
{
Ok(()) => {
sess.record_execpolicy_amendment_message(
&event_turn_id,
proposed_execpolicy_amendment,
)
.await;
}
Err(err) => {
let message = format!("Failed to apply execpolicy amendment: {err}");
tracing::warn!("{message}");
let warning = EventMsg::Warning(WarningEvent { message });
sess.send_event_raw(Event {
id: event_turn_id.clone(),
msg: warning,
})
.await;
}
}
}
match decision {
ReviewDecision::Abort => {
sess.interrupt_task().await;
}
other => sess.notify_approval(&approval_id, other).await,
}
}
pub async fn patch_approval(sess: &Arc<Session>, id: String, decision: ReviewDecision) {
match decision {
ReviewDecision::Abort => {
sess.interrupt_task().await;
}
other => sess.notify_approval(&id, other).await,
}
}
pub async fn request_user_input_response(
sess: &Arc<Session>,
id: String,
response: RequestUserInputResponse,
) {
sess.notify_user_input_response(&id, response).await;
}
pub async fn request_permissions_response(
sess: &Arc<Session>,
id: String,
response: RequestPermissionsResponse,
) {
sess.notify_request_permissions_response(&id, response)
.await;
}
pub async fn dynamic_tool_response(
sess: &Arc<Session>,
id: String,
response: DynamicToolResponse,
) {
sess.notify_dynamic_tool_response(&id, response).await;
}
pub async fn add_to_history(sess: &Arc<Session>, config: &Arc<Config>, text: String) {
let id = sess.conversation_id;
let config = Arc::clone(config);
tokio::spawn(async move {
if let Err(e) = crate::message_history::append_entry(&text, &id, &config).await {
warn!("failed to append to message history: {e}");
}
});
}
pub async fn get_history_entry_request(
sess: &Arc<Session>,
config: &Arc<Config>,
sub_id: String,
offset: usize,
log_id: u64,
) {
let config = Arc::clone(config);
let sess_clone = Arc::clone(sess);
tokio::spawn(async move {
// Run lookup in blocking thread because it does file IO + locking.
let entry_opt = tokio::task::spawn_blocking(move || {
crate::message_history::lookup(log_id, offset, &config)
})
.await
.unwrap_or(None);
let event = Event {
id: sub_id,
msg: EventMsg::GetHistoryEntryResponse(
crate::protocol::GetHistoryEntryResponseEvent {
offset,
log_id,
entry: entry_opt.map(|e| codex_protocol::message_history::HistoryEntry {
conversation_id: e.session_id,
ts: e.ts,
text: e.text,
}),
},
),
};
sess_clone.send_event_raw(event).await;
});
}
pub async fn refresh_mcp_servers(sess: &Arc<Session>, refresh_config: McpServerRefreshConfig) {
let mut guard = sess.pending_mcp_server_refresh_config.lock().await;
*guard = Some(refresh_config);
}
pub async fn reload_user_config(sess: &Arc<Session>) {
sess.reload_user_config_layer().await;
}
pub async fn list_mcp_tools(sess: &Session, config: &Arc<Config>, sub_id: String) {
let mcp_connection_manager = sess.services.mcp_connection_manager.read().await;
let auth = sess.services.auth_manager.auth().await;
let mcp_servers = sess
.services
.mcp_manager
.effective_servers(config, auth.as_ref());
let snapshot = collect_mcp_snapshot_from_manager(
&mcp_connection_manager,
compute_auth_statuses(mcp_servers.iter(), config.mcp_oauth_credentials_store_mode)
.await,
)
.await;
let event = Event {
id: sub_id,
msg: EventMsg::McpListToolsResponse(snapshot),
};
sess.send_event_raw(event).await;
}
pub async fn list_custom_prompts(sess: &Session, sub_id: String) {
let custom_prompts: Vec<CustomPrompt> =
if let Some(dir) = crate::custom_prompts::default_prompts_dir() {
crate::custom_prompts::discover_prompts_in(&dir).await
} else {
Vec::new()
};
let event = Event {
id: sub_id,
msg: EventMsg::ListCustomPromptsResponse(ListCustomPromptsResponseEvent {
custom_prompts,
}),
};
sess.send_event_raw(event).await;
}
pub async fn list_skills(
sess: &Session,
sub_id: String,
cwds: Vec<PathBuf>,
force_reload: bool,
) {
let cwds = if cwds.is_empty() {
let state = sess.state.lock().await;
vec![state.session_configuration.cwd.clone()]
} else {
cwds
};
let skills_manager = &sess.services.skills_manager;
let mut skills = Vec::new();
for cwd in cwds {
let outcome = skills_manager.skills_for_cwd(&cwd, force_reload).await;
let errors = super::errors_to_info(&outcome.errors);
let skills_metadata = super::skills_to_info(&outcome.skills, &outcome.disabled_paths);
skills.push(SkillsListEntry {
cwd,
skills: skills_metadata,
errors,
});
}
let event = Event {
id: sub_id,
msg: EventMsg::ListSkillsResponse(ListSkillsResponseEvent { skills }),
};
sess.send_event_raw(event).await;
}
pub async fn list_remote_skills(
sess: &Session,
config: &Arc<Config>,
sub_id: String,
hazelnut_scope: RemoteSkillHazelnutScope,
product_surface: RemoteSkillProductSurface,
enabled: Option<bool>,
) {
let auth = sess.services.auth_manager.auth().await;
let response = crate::skills::remote::list_remote_skills(
config,
auth.as_ref(),
hazelnut_scope,
product_surface,
enabled,
)
.await
.map(|skills| {
skills
.into_iter()
.map(|skill| RemoteSkillSummary {
id: skill.id,
name: skill.name,
description: skill.description,
})
.collect::<Vec<_>>()
});
match response {
Ok(skills) => {
let event = Event {
id: sub_id,
msg: EventMsg::ListRemoteSkillsResponse(ListRemoteSkillsResponseEvent {
skills,
}),
};
sess.send_event_raw(event).await;
}
Err(err) => {
let event = Event {
id: sub_id,
msg: EventMsg::Error(ErrorEvent {
message: format!("failed to list remote skills: {err}"),
codex_error_info: Some(CodexErrorInfo::Other),
}),
};
sess.send_event_raw(event).await;
}
}
}
pub async fn export_remote_skill(
sess: &Session,
config: &Arc<Config>,
sub_id: String,
hazelnut_id: String,
) {
let auth = sess.services.auth_manager.auth().await;
match crate::skills::remote::export_remote_skill(
config,
auth.as_ref(),
hazelnut_id.as_str(),
)
.await
{
Ok(result) => {
let id = result.id;
let event = Event {
id: sub_id,
msg: EventMsg::RemoteSkillDownloaded(RemoteSkillDownloadedEvent {
id: id.clone(),
name: id,
path: result.path,
}),
};
sess.send_event_raw(event).await;
}
Err(err) => {
let event = Event {
id: sub_id,
msg: EventMsg::Error(ErrorEvent {
message: format!("failed to export remote skill {hazelnut_id}: {err}"),
codex_error_info: Some(CodexErrorInfo::Other),
}),
};
sess.send_event_raw(event).await;
}
}
}
pub async fn undo(sess: &Arc<Session>, sub_id: String) {
let turn_context = sess.new_default_turn_with_sub_id(sub_id).await;
sess.spawn_task(turn_context, Vec::new(), UndoTask::new())
.await;
}
pub async fn compact(sess: &Arc<Session>, sub_id: String) {
let turn_context = sess.new_default_turn_with_sub_id(sub_id).await;
sess.spawn_task(
Arc::clone(&turn_context),
vec![UserInput::Text {
text: turn_context.compact_prompt().to_string(),
// Compaction prompt is synthesized; no UI element ranges to preserve.
text_elements: Vec::new(),
}],
CompactTask,
)
.await;
}
pub async fn drop_memories(sess: &Arc<Session>, config: &Arc<Config>, sub_id: String) {
let mut errors = Vec::new();
if let Some(state_db) = sess.services.state_db.as_deref() {
if let Err(err) = state_db.clear_memory_data().await {
errors.push(format!("failed clearing memory rows from state db: {err}"));
}
} else {
errors.push("state db unavailable; memory rows were not cleared".to_string());
}
let memory_root = crate::memories::memory_root(&config.codex_home);
if let Err(err) = crate::memories::clear_memory_root_contents(&memory_root).await {
errors.push(format!(
"failed clearing memory directory {}: {err}",
memory_root.display()
));
}
if errors.is_empty() {
sess.send_event_raw(Event {
id: sub_id,
msg: EventMsg::Warning(WarningEvent {
message: format!(
"Dropped memories at {} and cleared memory rows from state db.",
memory_root.display()
),
}),
})
.await;
return;
}
sess.send_event_raw(Event {
id: sub_id,
msg: EventMsg::Error(ErrorEvent {
message: format!("Memory drop completed with errors: {}", errors.join("; ")),
codex_error_info: Some(CodexErrorInfo::Other),
}),
})
.await;
}
pub async fn update_memories(sess: &Arc<Session>, config: &Arc<Config>, sub_id: String) {
let session_source = {
let state = sess.state.lock().await;
state.session_configuration.session_source.clone()
};
crate::memories::start_memories_startup_task(sess, Arc::clone(config), &session_source);
sess.send_event_raw(Event {
id: sub_id.clone(),
msg: EventMsg::Warning(WarningEvent {
message: "Memory update triggered.".to_string(),
}),
})
.await;
}
pub async fn thread_rollback(sess: &Arc<Session>, sub_id: String, num_turns: u32) {
if num_turns == 0 {
sess.send_event_raw(Event {
id: sub_id,
msg: EventMsg::Error(ErrorEvent {
message: "num_turns must be >= 1".to_string(),
codex_error_info: Some(CodexErrorInfo::ThreadRollbackFailed),
}),
})
.await;
return;
}
let has_active_turn = { sess.active_turn.lock().await.is_some() };
if has_active_turn {
sess.send_event_raw(Event {
id: sub_id,
msg: EventMsg::Error(ErrorEvent {
message: "Cannot rollback while a turn is in progress.".to_string(),
codex_error_info: Some(CodexErrorInfo::ThreadRollbackFailed),
}),
})
.await;
return;
}
let turn_context = sess.new_default_turn_with_sub_id(sub_id).await;
let rollout_path = {
let recorder = {
let guard = sess.services.rollout.lock().await;
guard.clone()
};
let Some(recorder) = recorder else {
sess.send_event_raw(Event {
id: turn_context.sub_id.clone(),
msg: EventMsg::Error(ErrorEvent {
message: "thread rollback requires a persisted rollout path".to_string(),
codex_error_info: Some(CodexErrorInfo::ThreadRollbackFailed),
}),
})
.await;
return;
};
recorder.rollout_path().to_path_buf()
};
if let Some(recorder) = {
let guard = sess.services.rollout.lock().await;
guard.clone()
} && let Err(err) = recorder.flush().await
{
sess.send_event_raw(Event {
id: turn_context.sub_id.clone(),
msg: EventMsg::Error(ErrorEvent {
message: format!(
"failed to flush rollout `{}` for rollback replay: {err}",
rollout_path.display()
),
codex_error_info: Some(CodexErrorInfo::ThreadRollbackFailed),
}),
})
.await;
return;
}
let initial_history =
match RolloutRecorder::get_rollout_history(rollout_path.as_path()).await {
Ok(history) => history,
Err(err) => {
sess.send_event_raw(Event {
id: turn_context.sub_id.clone(),
msg: EventMsg::Error(ErrorEvent {
message: format!(
"failed to load rollout `{}` for rollback replay: {err}",
rollout_path.display()
),
codex_error_info: Some(CodexErrorInfo::ThreadRollbackFailed),
}),
})
.await;
return;
}
};
let rollback_event = ThreadRolledBackEvent { num_turns };
let replay_items = initial_history
.get_rollout_items()
.into_iter()
.chain(std::iter::once(RolloutItem::EventMsg(
EventMsg::ThreadRolledBack(rollback_event.clone()),
)))
.collect::<Vec<_>>();
let reconstructed = sess
.reconstruct_history_from_rollout(turn_context.as_ref(), replay_items.as_slice())
.await;
sess.replace_history(
reconstructed.history,
reconstructed.reference_context_item.clone(),
)
.await;
sess.set_previous_turn_settings(reconstructed.previous_turn_settings)
.await;
sess.recompute_token_usage(turn_context.as_ref()).await;
sess.send_event_raw_flushed(Event {
id: turn_context.sub_id.clone(),
msg: EventMsg::ThreadRolledBack(rollback_event),
})
.await;
}
/// Persists the thread name in the session index, updates in-memory state, and emits
/// a `ThreadNameUpdated` event on success.
///
/// This appends the name to `CODEX_HOME/sessions_index.jsonl` via `session_index::append_thread_name` for the
/// current `thread_id`, then updates `SessionConfiguration::thread_name`.
///
/// Returns an error event if the name is empty or session persistence is disabled.
pub async fn set_thread_name(sess: &Arc<Session>, sub_id: String, name: String) {
let Some(name) = crate::util::normalize_thread_name(&name) else {
let event = Event {
id: sub_id,
msg: EventMsg::Error(ErrorEvent {
message: "Thread name cannot be empty.".to_string(),
codex_error_info: Some(CodexErrorInfo::BadRequest),
}),
};
sess.send_event_raw(event).await;
return;
};
let persistence_enabled = {
let rollout = sess.services.rollout.lock().await;
rollout.is_some()
};
if !persistence_enabled {
let event = Event {
id: sub_id,
msg: EventMsg::Error(ErrorEvent {
message: "Session persistence is disabled; cannot rename thread.".to_string(),
codex_error_info: Some(CodexErrorInfo::Other),
}),
};
sess.send_event_raw(event).await;
return;
};
let codex_home = sess.codex_home().await;
if let Err(e) =
session_index::append_thread_name(&codex_home, sess.conversation_id, &name).await
{
let event = Event {
id: sub_id,
msg: EventMsg::Error(ErrorEvent {
message: format!("Failed to set thread name: {e}"),
codex_error_info: Some(CodexErrorInfo::Other),
}),
};
sess.send_event_raw(event).await;
return;
}
{
let mut state = sess.state.lock().await;
state.session_configuration.thread_name = Some(name.clone());
}
sess.send_event_raw(Event {
id: sub_id,
msg: EventMsg::ThreadNameUpdated(ThreadNameUpdatedEvent {
thread_id: sess.conversation_id,
thread_name: Some(name),
}),
})
.await;
}
pub async fn shutdown(sess: &Arc<Session>, sub_id: String) -> bool {
sess.abort_all_tasks(TurnAbortReason::Interrupted).await;
let _ = sess.conversation.shutdown().await;
sess.services
.unified_exec_manager
.terminate_all_processes()
.await;
info!("Shutting down Codex instance");
let history = sess.clone_history().await;
let turn_count = history
.raw_items()
.iter()
.filter(|item| is_user_turn_boundary(item))
.count();
sess.services.session_telemetry.counter(
"codex.conversation.turn.count",
i64::try_from(turn_count).unwrap_or(0),
&[],
);
// Gracefully flush and shutdown rollout recorder on session end so tests
// that inspect the rollout file do not race with the background writer.
let recorder_opt = {
let mut guard = sess.services.rollout.lock().await;
guard.take()
};
if let Some(rec) = recorder_opt
&& let Err(e) = rec.shutdown().await
{
warn!("failed to shutdown rollout recorder: {e}");
let event = Event {
id: sub_id.clone(),
msg: EventMsg::Error(ErrorEvent {
message: "Failed to shutdown rollout recorder".to_string(),
codex_error_info: Some(CodexErrorInfo::Other),
}),
};
sess.send_event_raw(event).await;
}
let event = Event {
id: sub_id,
msg: EventMsg::ShutdownComplete,
};
sess.send_event_raw(event).await;
true
}
pub async fn review(
sess: &Arc<Session>,
config: &Arc<Config>,
sub_id: String,
review_request: ReviewRequest,
) {
let turn_context = sess.new_default_turn_with_sub_id(sub_id.clone()).await;
sess.maybe_emit_unknown_model_warning_for_turn(turn_context.as_ref())
.await;
sess.refresh_mcp_servers_if_requested(&turn_context).await;
match resolve_review_request(review_request, turn_context.cwd.as_path()) {
Ok(resolved) => {
spawn_review_thread(
Arc::clone(sess),
Arc::clone(config),
turn_context.clone(),
sub_id,
resolved,
)
.await;
}
Err(err) => {
let event = Event {
id: sub_id,
msg: EventMsg::Error(ErrorEvent {
message: err.to_string(),
codex_error_info: Some(CodexErrorInfo::Other),
}),
};
sess.send_event(&turn_context, event.msg).await;
}
}
}
}
/// Spawn a review thread using the given prompt.
async fn spawn_review_thread(
sess: Arc<Session>,
config: Arc<Config>,
parent_turn_context: Arc<TurnContext>,
sub_id: String,
resolved: crate::review_prompts::ResolvedReviewRequest,
) {
let model = config
.review_model
.clone()
.unwrap_or_else(|| parent_turn_context.model_info.slug.clone());
let review_model_info = sess
.services
.models_manager
.get_model_info(&model, &config)
.await;
// For reviews, disable web_search and view_image regardless of global settings.
let mut review_features = sess.features.clone();
let _ = review_features.disable(crate::features::Feature::WebSearchRequest);
let _ = review_features.disable(crate::features::Feature::WebSearchCached);
let review_web_search_mode = WebSearchMode::Disabled;
let tools_config = ToolsConfig::new(&ToolsConfigParams {
model_info: &review_model_info,
features: &review_features,
web_search_mode: Some(review_web_search_mode),
session_source: parent_turn_context.session_source.clone(),
})
.with_web_search_config(None)
.with_allow_login_shell(config.permissions.allow_login_shell)
.with_agent_roles(config.agent_roles.clone());
let review_prompt = resolved.prompt.clone();
let provider = parent_turn_context.provider.clone();
let auth_manager = parent_turn_context.auth_manager.clone();
let model_info = review_model_info.clone();
// Build perturn client with the requested model/family.
let mut per_turn_config = (*config).clone();
per_turn_config.model = Some(model.clone());
per_turn_config.features = review_features.clone();
if let Err(err) = per_turn_config.web_search_mode.set(review_web_search_mode) {
let fallback_value = per_turn_config.web_search_mode.value();
tracing::warn!(
error = %err,
?review_web_search_mode,
?fallback_value,
"review web_search_mode is disallowed by requirements; keeping constrained value"
);
}
let session_telemetry = parent_turn_context
.session_telemetry
.clone()
.with_model(model.as_str(), review_model_info.slug.as_str());
let auth_manager_for_context = auth_manager.clone();
let provider_for_context = provider.clone();
let session_telemetry_for_context = session_telemetry.clone();
let reasoning_effort = per_turn_config.model_reasoning_effort;
let reasoning_summary = per_turn_config
.model_reasoning_summary
.unwrap_or(model_info.default_reasoning_summary);
let session_source = parent_turn_context.session_source.clone();
let per_turn_config = Arc::new(per_turn_config);
let review_turn_id = sub_id.to_string();
let turn_metadata_state = Arc::new(TurnMetadataState::new(
review_turn_id.clone(),
parent_turn_context.cwd.clone(),
parent_turn_context.sandbox_policy.get(),
parent_turn_context.windows_sandbox_level,
parent_turn_context
.features
.enabled(Feature::UseLinuxSandboxBwrap),
));
let review_turn_context = TurnContext {
sub_id: review_turn_id,
trace_id: current_span_trace_id(),
realtime_active: parent_turn_context.realtime_active,
config: per_turn_config,
auth_manager: auth_manager_for_context,
model_info: model_info.clone(),
session_telemetry: session_telemetry_for_context,
provider: provider_for_context,
reasoning_effort,
reasoning_summary,
session_source,
tools_config,
features: parent_turn_context.features.clone(),
ghost_snapshot: parent_turn_context.ghost_snapshot.clone(),
current_date: parent_turn_context.current_date.clone(),
timezone: parent_turn_context.timezone.clone(),
app_server_client_name: parent_turn_context.app_server_client_name.clone(),
developer_instructions: None,
user_instructions: None,
compact_prompt: parent_turn_context.compact_prompt.clone(),
collaboration_mode: parent_turn_context.collaboration_mode.clone(),
personality: parent_turn_context.personality,
approval_policy: parent_turn_context.approval_policy.clone(),
sandbox_policy: parent_turn_context.sandbox_policy.clone(),
file_system_sandbox_policy: parent_turn_context.file_system_sandbox_policy.clone(),
network_sandbox_policy: parent_turn_context.network_sandbox_policy,
network: parent_turn_context.network.clone(),
windows_sandbox_level: parent_turn_context.windows_sandbox_level,
shell_environment_policy: parent_turn_context.shell_environment_policy.clone(),
cwd: parent_turn_context.cwd.clone(),
final_output_json_schema: None,
codex_linux_sandbox_exe: parent_turn_context.codex_linux_sandbox_exe.clone(),
tool_call_gate: Arc::new(ReadinessFlag::new()),
js_repl: Arc::clone(&sess.js_repl),
dynamic_tools: parent_turn_context.dynamic_tools.clone(),
truncation_policy: model_info.truncation_policy.into(),
turn_metadata_state,
turn_skills: TurnSkillsContext::new(parent_turn_context.turn_skills.outcome.clone()),
turn_timing_state: Arc::new(TurnTimingState::default()),
};
// Seed the child task with the review prompt as the initial user message.
let input: Vec<UserInput> = vec![UserInput::Text {
text: review_prompt,
// Review prompt is synthesized; no UI element ranges to preserve.
text_elements: Vec::new(),
}];
let tc = Arc::new(review_turn_context);
tc.turn_metadata_state.spawn_git_enrichment_task();
// TODO(ccunningham): Review turns currently rely on `spawn_task` for TurnComplete but do not
// emit a parent TurnStarted. Consider giving review a full parent turn lifecycle
// (TurnStarted + TurnComplete) for consistency with other standalone tasks.
sess.spawn_task(tc.clone(), input, ReviewTask::new()).await;
// Announce entering review mode so UIs can switch modes.
let review_request = ReviewRequest {
target: resolved.target,
user_facing_hint: Some(resolved.user_facing_hint),
};
sess.send_event(&tc, EventMsg::EnteredReviewMode(review_request))
.await;
}
fn skills_to_info(
skills: &[SkillMetadata],
disabled_paths: &HashSet<PathBuf>,
) -> Vec<ProtocolSkillMetadata> {
skills
.iter()
.map(|skill| ProtocolSkillMetadata {
name: skill.name.clone(),
description: skill.description.clone(),
short_description: skill.short_description.clone(),
interface: skill
.interface
.clone()
.map(|interface| ProtocolSkillInterface {
display_name: interface.display_name,
short_description: interface.short_description,
icon_small: interface.icon_small,
icon_large: interface.icon_large,
brand_color: interface.brand_color,
default_prompt: interface.default_prompt,
}),
dependencies: skill.dependencies.clone().map(|dependencies| {
ProtocolSkillDependencies {
tools: dependencies
.tools
.into_iter()
.map(|tool| ProtocolSkillToolDependency {
r#type: tool.r#type,
value: tool.value,
description: tool.description,
transport: tool.transport,
command: tool.command,
url: tool.url,
})
.collect(),
}
}),
path: skill.path_to_skills_md.clone(),
scope: skill.scope,
enabled: !disabled_paths.contains(&skill.path_to_skills_md),
})
.collect()
}
fn errors_to_info(errors: &[SkillError]) -> Vec<SkillErrorInfo> {
errors
.iter()
.map(|err| SkillErrorInfo {
path: err.path.clone(),
message: err.message.clone(),
})
.collect()
}
/// Takes a user message as input and runs a loop where, at each sampling request, the model
/// replies with either:
///
/// - requested function calls
/// - an assistant message
///
/// While it is possible for the model to return multiple of these items in a
/// single sampling request, in practice, we generally one item per sampling request:
///
/// - If the model requests a function call, we execute it and send the output
/// back to the model in the next sampling request.
/// - If the model sends only an assistant message, we record it in the
/// conversation history and consider the turn complete.
///
pub(crate) async fn run_turn(
sess: Arc<Session>,
turn_context: Arc<TurnContext>,
input: Vec<UserInput>,
prewarmed_client_session: Option<ModelClientSession>,
cancellation_token: CancellationToken,
) -> Option<String> {
if input.is_empty() {
return None;
}
let model_info = turn_context.model_info.clone();
let auto_compact_limit = model_info.auto_compact_token_limit().unwrap_or(i64::MAX);
let event = EventMsg::TurnStarted(TurnStartedEvent {
turn_id: turn_context.sub_id.clone(),
model_context_window: turn_context.model_context_window(),
collaboration_mode_kind: turn_context.collaboration_mode.mode,
});
sess.send_event(&turn_context, event).await;
// TODO(ccunningham): Pre-turn compaction runs before context updates and the
// new user message are recorded. Estimate pending incoming items (context
// diffs/full reinjection + user input) and trigger compaction preemptively
// when they would push the thread over the compaction threshold.
if run_pre_sampling_compact(&sess, &turn_context)
.await
.is_err()
{
error!("Failed to run pre-sampling compact");
return None;
}
let skills_outcome = Some(turn_context.turn_skills.outcome.as_ref());
sess.record_context_updates_and_set_reference_context_item(turn_context.as_ref())
.await;
let loaded_plugins = sess
.services
.plugins_manager
.plugins_for_config(&turn_context.config);
// Structured plugin:// mentions are resolved from the current session's
// enabled plugins, then converted into turn-scoped guidance below.
let mentioned_plugins =
collect_explicit_plugin_mentions(&input, loaded_plugins.capability_summaries());
let mcp_tools =
if turn_context.config.features.enabled(Feature::Apps) || !mentioned_plugins.is_empty() {
// Plugin mentions need raw MCP/app inventory even when app tools
// are normally hidden so we can describe the plugin's currently
// usable capabilities for this turn.
match sess
.services
.mcp_connection_manager
.read()
.await
.list_all_tools()
.or_cancel(&cancellation_token)
.await
{
Ok(mcp_tools) => mcp_tools,
Err(_) if turn_context.config.features.enabled(Feature::Apps) => return None,
Err(_) => HashMap::new(),
}
} else {
HashMap::new()
};
let available_connectors = if turn_context.config.features.enabled(Feature::Apps) {
let connectors = connectors::merge_plugin_apps_with_accessible(
loaded_plugins.effective_apps(),
connectors::accessible_connectors_from_mcp_tools(&mcp_tools),
);
connectors::with_app_enabled_state(connectors, &turn_context.config)
} else {
Vec::new()
};
let connector_slug_counts = build_connector_slug_counts(&available_connectors);
let skill_name_counts_lower = skills_outcome
.as_ref()
.map_or_else(HashMap::new, |outcome| {
build_skill_name_counts(&outcome.skills, &outcome.disabled_paths).1
});
let mentioned_skills = skills_outcome.as_ref().map_or_else(Vec::new, |outcome| {
collect_explicit_skill_mentions(
&input,
&outcome.skills,
&outcome.disabled_paths,
&connector_slug_counts,
)
});
let config = turn_context.config.clone();
if config
.features
.enabled(Feature::SkillEnvVarDependencyPrompt)
{
let env_var_dependencies = collect_env_var_dependencies(&mentioned_skills);
resolve_skill_dependencies_for_turn(&sess, &turn_context, &env_var_dependencies).await;
}
maybe_prompt_and_install_mcp_dependencies(
sess.as_ref(),
turn_context.as_ref(),
&cancellation_token,
&mentioned_skills,
)
.await;
let session_telemetry = turn_context.session_telemetry.clone();
let thread_id = sess.conversation_id.to_string();
let tracking = build_track_events_context(
turn_context.model_info.slug.clone(),
thread_id,
turn_context.sub_id.clone(),
);
let SkillInjections {
items: skill_items,
warnings: skill_warnings,
} = build_skill_injections(
&mentioned_skills,
Some(&session_telemetry),
&sess.services.analytics_events_client,
tracking.clone(),
)
.await;
for message in skill_warnings {
sess.send_event(&turn_context, EventMsg::Warning(WarningEvent { message }))
.await;
}
let plugin_items =
build_plugin_injections(&mentioned_plugins, &mcp_tools, &available_connectors);
let mut explicitly_enabled_connectors = collect_explicit_app_ids(&input);
explicitly_enabled_connectors.extend(collect_explicit_app_ids_from_skill_items(
&skill_items,
&available_connectors,
&skill_name_counts_lower,
));
// Explicit plugin mentions can make a plugin's enabled apps callable for
// this turn without persisting those connectors as sticky user selections.
let mut turn_enabled_connectors = explicitly_enabled_connectors.clone();
turn_enabled_connectors.extend(
mentioned_plugins
.iter()
.flat_map(|plugin| plugin.app_connector_ids.iter())
.map(|connector_id| connector_id.0.clone())
.filter(|connector_id| {
available_connectors
.iter()
.any(|connector| connector.is_enabled && connector.id == *connector_id)
}),
);
let connector_names_by_id = available_connectors
.iter()
.map(|connector| (connector.id.as_str(), connector.name.as_str()))
.collect::<HashMap<&str, &str>>();
let mentioned_app_invocations = explicitly_enabled_connectors
.iter()
.map(|connector_id| AppInvocation {
connector_id: Some(connector_id.clone()),
app_name: connector_names_by_id
.get(connector_id.as_str())
.map(|name| (*name).to_string()),
invocation_type: Some(InvocationType::Explicit),
})
.collect::<Vec<_>>();
sess.services
.analytics_events_client
.track_app_mentioned(tracking.clone(), mentioned_app_invocations);
sess.merge_connector_selection(explicitly_enabled_connectors.clone())
.await;
let initial_input_for_turn: ResponseInputItem = ResponseInputItem::from(input.clone());
let response_item: ResponseItem = initial_input_for_turn.clone().into();
sess.record_user_prompt_and_emit_turn_item(turn_context.as_ref(), &input, response_item)
.await;
// Track the previous-turn baseline from the regular user-turn path only so
// standalone tasks (compact/shell/review/undo) cannot suppress future
// model/realtime injections.
sess.set_previous_turn_settings(Some(PreviousTurnSettings {
model: turn_context.model_info.slug.clone(),
realtime_active: Some(turn_context.realtime_active),
}))
.await;
if !skill_items.is_empty() {
sess.record_conversation_items(&turn_context, &skill_items)
.await;
}
if !plugin_items.is_empty() {
sess.record_conversation_items(&turn_context, &plugin_items)
.await;
}
sess.maybe_start_ghost_snapshot(Arc::clone(&turn_context), cancellation_token.child_token())
.await;
let mut last_agent_message: Option<String> = None;
// Although from the perspective of codex.rs, TurnDiffTracker has the lifecycle of a Task which contains
// many turns, from the perspective of the user, it is a single turn.
let turn_diff_tracker = Arc::new(tokio::sync::Mutex::new(TurnDiffTracker::new()));
let mut server_model_warning_emitted_for_turn = false;
// `ModelClientSession` is turn-scoped and caches WebSocket + sticky routing state, so we reuse
// one instance across retries within this turn.
let mut client_session =
prewarmed_client_session.unwrap_or_else(|| sess.services.model_client.new_session());
loop {
// Note that pending_input would be something like a message the user
// submitted through the UI while the model was running. Though the UI
// may support this, the model might not.
let pending_response_items = sess
.get_pending_input()
.await
.into_iter()
.map(ResponseItem::from)
.collect::<Vec<ResponseItem>>();
if !pending_response_items.is_empty() {
for response_item in pending_response_items {
if let Some(TurnItem::UserMessage(user_message)) = parse_turn_item(&response_item) {
// todo(aibrahim): move pending input to be UserInput only to keep TextElements. context: https://github.com/openai/codex/pull/10656#discussion_r2765522480
sess.record_user_prompt_and_emit_turn_item(
turn_context.as_ref(),
&user_message.content,
response_item,
)
.await;
} else {
sess.record_conversation_items(
&turn_context,
std::slice::from_ref(&response_item),
)
.await;
}
}
}
// Construct the input that we will send to the model.
let sampling_request_input: Vec<ResponseItem> = {
sess.clone_history()
.await
.for_prompt(&turn_context.model_info.input_modalities)
};
let sampling_request_input_messages = sampling_request_input
.iter()
.filter_map(|item| match parse_turn_item(item) {
Some(TurnItem::UserMessage(user_message)) => Some(user_message),
_ => None,
})
.map(|user_message| user_message.message())
.collect::<Vec<String>>();
let turn_metadata_header = turn_context.turn_metadata_state.current_header_value();
match run_sampling_request(
Arc::clone(&sess),
Arc::clone(&turn_context),
Arc::clone(&turn_diff_tracker),
&mut client_session,
turn_metadata_header.as_deref(),
sampling_request_input,
&turn_enabled_connectors,
skills_outcome,
&mut server_model_warning_emitted_for_turn,
cancellation_token.child_token(),
)
.await
{
Ok(sampling_request_output) => {
let SamplingRequestResult {
needs_follow_up,
last_agent_message: sampling_request_last_agent_message,
} = sampling_request_output;
let total_usage_tokens = sess.get_total_token_usage().await;
let token_limit_reached = total_usage_tokens >= auto_compact_limit;
let estimated_token_count =
sess.get_estimated_token_count(turn_context.as_ref()).await;
trace!(
turn_id = %turn_context.sub_id,
total_usage_tokens,
estimated_token_count = ?estimated_token_count,
auto_compact_limit,
token_limit_reached,
needs_follow_up,
"post sampling token usage"
);
// as long as compaction works well in getting us way below the token limit, we shouldn't worry about being in an infinite loop.
if token_limit_reached && needs_follow_up {
if run_auto_compact(
&sess,
&turn_context,
InitialContextInjection::BeforeLastUserMessage,
)
.await
.is_err()
{
return None;
}
continue;
}
if !needs_follow_up {
last_agent_message = sampling_request_last_agent_message;
let hook_outcomes = sess
.hooks()
.dispatch(HookPayload {
session_id: sess.conversation_id,
cwd: turn_context.cwd.clone(),
client: turn_context.app_server_client_name.clone(),
triggered_at: chrono::Utc::now(),
hook_event: HookEvent::AfterAgent {
event: HookEventAfterAgent {
thread_id: sess.conversation_id,
turn_id: turn_context.sub_id.clone(),
input_messages: sampling_request_input_messages,
last_assistant_message: last_agent_message.clone(),
},
},
})
.await;
let mut abort_message = None;
for hook_outcome in hook_outcomes {
let hook_name = hook_outcome.hook_name;
match hook_outcome.result {
HookResult::Success => {}
HookResult::FailedContinue(error) => {
warn!(
turn_id = %turn_context.sub_id,
hook_name = %hook_name,
error = %error,
"after_agent hook failed; continuing"
);
}
HookResult::FailedAbort(error) => {
let message = format!(
"after_agent hook '{hook_name}' failed and aborted turn completion: {error}"
);
warn!(
turn_id = %turn_context.sub_id,
hook_name = %hook_name,
error = %error,
"after_agent hook failed; aborting operation"
);
if abort_message.is_none() {
abort_message = Some(message);
}
}
}
}
if let Some(message) = abort_message {
sess.send_event(
&turn_context,
EventMsg::Error(ErrorEvent {
message,
codex_error_info: None,
}),
)
.await;
return None;
}
break;
}
continue;
}
Err(CodexErr::TurnAborted) => {
// Aborted turn is reported via a different event.
break;
}
Err(CodexErr::InvalidImageRequest()) => {
let mut state = sess.state.lock().await;
error_or_panic(
"Invalid image detected; sanitizing tool output to prevent poisoning",
);
if state.history.replace_last_turn_images("Invalid image") {
continue;
}
let event = EventMsg::Error(ErrorEvent {
message: "Invalid image in your last message. Please remove it and try again."
.to_string(),
codex_error_info: Some(CodexErrorInfo::BadRequest),
});
sess.send_event(&turn_context, event).await;
break;
}
Err(e) => {
info!("Turn error: {e:#}");
let event = EventMsg::Error(e.to_error_event(None));
sess.send_event(&turn_context, event).await;
// let the user continue the conversation
break;
}
}
}
last_agent_message
}
async fn run_pre_sampling_compact(
sess: &Arc<Session>,
turn_context: &Arc<TurnContext>,
) -> CodexResult<()> {
let total_usage_tokens_before_compaction = sess.get_total_token_usage().await;
maybe_run_previous_model_inline_compact(
sess,
turn_context,
total_usage_tokens_before_compaction,
)
.await?;
let total_usage_tokens = sess.get_total_token_usage().await;
let auto_compact_limit = turn_context
.model_info
.auto_compact_token_limit()
.unwrap_or(i64::MAX);
// Compact if the total usage tokens are greater than the auto compact limit
if total_usage_tokens >= auto_compact_limit {
run_auto_compact(sess, turn_context, InitialContextInjection::DoNotInject).await?;
}
Ok(())
}
/// Runs pre-sampling compaction against the previous model when switching to a smaller
/// context-window model.
///
/// Returns `Ok(true)` when compaction ran successfully, `Ok(false)` when compaction was skipped
/// because the model/context-window preconditions were not met, and `Err(_)` only when compaction
/// was attempted and failed.
async fn maybe_run_previous_model_inline_compact(
sess: &Arc<Session>,
turn_context: &Arc<TurnContext>,
total_usage_tokens: i64,
) -> CodexResult<bool> {
let Some(previous_turn_settings) = sess.previous_turn_settings().await else {
return Ok(false);
};
let previous_model_turn_context = Arc::new(
turn_context
.with_model(previous_turn_settings.model, &sess.services.models_manager)
.await,
);
let Some(old_context_window) = previous_model_turn_context.model_context_window() else {
return Ok(false);
};
let Some(new_context_window) = turn_context.model_context_window() else {
return Ok(false);
};
let new_auto_compact_limit = turn_context
.model_info
.auto_compact_token_limit()
.unwrap_or(i64::MAX);
let should_run = total_usage_tokens > new_auto_compact_limit
&& previous_model_turn_context.model_info.slug != turn_context.model_info.slug
&& old_context_window > new_context_window;
if should_run {
run_auto_compact(
sess,
&previous_model_turn_context,
InitialContextInjection::DoNotInject,
)
.await?;
return Ok(true);
}
Ok(false)
}
async fn run_auto_compact(
sess: &Arc<Session>,
turn_context: &Arc<TurnContext>,
initial_context_injection: InitialContextInjection,
) -> CodexResult<()> {
if should_use_remote_compact_task(&turn_context.provider) {
run_inline_remote_auto_compact_task(
Arc::clone(sess),
Arc::clone(turn_context),
initial_context_injection,
)
.await?;
} else {
run_inline_auto_compact_task(
Arc::clone(sess),
Arc::clone(turn_context),
initial_context_injection,
)
.await?;
}
Ok(())
}
fn collect_explicit_app_ids_from_skill_items(
skill_items: &[ResponseItem],
connectors: &[connectors::AppInfo],
skill_name_counts_lower: &HashMap<String, usize>,
) -> HashSet<String> {
if skill_items.is_empty() || connectors.is_empty() {
return HashSet::new();
}
let skill_messages = skill_items
.iter()
.filter_map(|item| match item {
ResponseItem::Message { content, .. } => {
content.iter().find_map(|content_item| match content_item {
ContentItem::InputText { text } => Some(text.clone()),
_ => None,
})
}
_ => None,
})
.collect::<Vec<String>>();
if skill_messages.is_empty() {
return HashSet::new();
}
let mentions = collect_tool_mentions_from_messages(&skill_messages);
let mention_names_lower = mentions
.plain_names
.iter()
.map(|name| name.to_ascii_lowercase())
.collect::<HashSet<String>>();
let mut connector_ids = mentions
.paths
.iter()
.filter(|path| tool_kind_for_path(path) == ToolMentionKind::App)
.filter_map(|path| app_id_from_path(path).map(str::to_string))
.collect::<HashSet<String>>();
let connector_slug_counts = build_connector_slug_counts(connectors);
for connector in connectors {
let slug = connectors::connector_mention_slug(connector);
let connector_count = connector_slug_counts.get(&slug).copied().unwrap_or(0);
let skill_count = skill_name_counts_lower.get(&slug).copied().unwrap_or(0);
if connector_count == 1 && skill_count == 0 && mention_names_lower.contains(&slug) {
connector_ids.insert(connector.id.clone());
}
}
connector_ids
}
fn filter_connectors_for_input(
connectors: &[connectors::AppInfo],
input: &[ResponseItem],
explicitly_enabled_connectors: &HashSet<String>,
skill_name_counts_lower: &HashMap<String, usize>,
) -> Vec<connectors::AppInfo> {
let connectors: Vec<connectors::AppInfo> = connectors
.iter()
.filter(|connector| connector.is_enabled)
.cloned()
.collect::<Vec<_>>();
if connectors.is_empty() {
return Vec::new();
}
let user_messages = collect_user_messages(input);
if user_messages.is_empty() && explicitly_enabled_connectors.is_empty() {
return Vec::new();
}
let mentions = collect_tool_mentions_from_messages(&user_messages);
let mention_names_lower = mentions
.plain_names
.iter()
.map(|name| name.to_ascii_lowercase())
.collect::<HashSet<String>>();
let connector_slug_counts = build_connector_slug_counts(&connectors);
let mut allowed_connector_ids = explicitly_enabled_connectors.clone();
for path in mentions
.paths
.iter()
.filter(|path| tool_kind_for_path(path) == ToolMentionKind::App)
{
if let Some(connector_id) = app_id_from_path(path) {
allowed_connector_ids.insert(connector_id.to_string());
}
}
connectors
.into_iter()
.filter(|connector| {
connector_inserted_in_messages(
connector,
&mention_names_lower,
&allowed_connector_ids,
&connector_slug_counts,
skill_name_counts_lower,
)
})
.collect()
}
fn connector_inserted_in_messages(
connector: &connectors::AppInfo,
mention_names_lower: &HashSet<String>,
allowed_connector_ids: &HashSet<String>,
connector_slug_counts: &HashMap<String, usize>,
skill_name_counts_lower: &HashMap<String, usize>,
) -> bool {
if allowed_connector_ids.contains(&connector.id) {
return true;
}
let mention_slug = connectors::connector_mention_slug(connector);
let connector_count = connector_slug_counts
.get(&mention_slug)
.copied()
.unwrap_or(0);
let skill_count = skill_name_counts_lower
.get(&mention_slug)
.copied()
.unwrap_or(0);
connector_count == 1 && skill_count == 0 && mention_names_lower.contains(&mention_slug)
}
fn filter_codex_apps_mcp_tools(
mcp_tools: &HashMap<String, crate::mcp_connection_manager::ToolInfo>,
connectors: &[connectors::AppInfo],
config: &Config,
) -> HashMap<String, crate::mcp_connection_manager::ToolInfo> {
let allowed: HashSet<&str> = connectors
.iter()
.map(|connector| connector.id.as_str())
.collect();
mcp_tools
.iter()
.filter(|(_, tool)| {
if tool.server_name != CODEX_APPS_MCP_SERVER_NAME {
return true;
}
let Some(connector_id) = codex_apps_connector_id(tool) else {
return false;
};
allowed.contains(connector_id) && connectors::codex_app_tool_is_enabled(config, tool)
})
.map(|(name, tool)| (name.clone(), tool.clone()))
.collect()
}
fn codex_apps_connector_id(tool: &crate::mcp_connection_manager::ToolInfo) -> Option<&str> {
tool.connector_id.as_deref()
}
fn build_prompt(
input: Vec<ResponseItem>,
router: &ToolRouter,
turn_context: &TurnContext,
base_instructions: BaseInstructions,
) -> Prompt {
Prompt {
input,
tools: router.specs(),
parallel_tool_calls: turn_context.model_info.supports_parallel_tool_calls,
base_instructions,
personality: turn_context.personality,
output_schema: turn_context.final_output_json_schema.clone(),
}
}
#[allow(clippy::too_many_arguments)]
#[instrument(level = "trace",
skip_all,
fields(
turn_id = %turn_context.sub_id,
model = %turn_context.model_info.slug,
cwd = %turn_context.cwd.display()
)
)]
async fn run_sampling_request(
sess: Arc<Session>,
turn_context: Arc<TurnContext>,
turn_diff_tracker: SharedTurnDiffTracker,
client_session: &mut ModelClientSession,
turn_metadata_header: Option<&str>,
input: Vec<ResponseItem>,
explicitly_enabled_connectors: &HashSet<String>,
skills_outcome: Option<&SkillLoadOutcome>,
server_model_warning_emitted_for_turn: &mut bool,
cancellation_token: CancellationToken,
) -> CodexResult<SamplingRequestResult> {
let router = built_tools(
sess.as_ref(),
turn_context.as_ref(),
&input,
explicitly_enabled_connectors,
skills_outcome,
&cancellation_token,
)
.await?;
let base_instructions = sess.get_base_instructions().await;
let prompt = build_prompt(
input,
router.as_ref(),
turn_context.as_ref(),
base_instructions,
);
let mut retries = 0;
loop {
let err = match try_run_sampling_request(
Arc::clone(&router),
Arc::clone(&sess),
Arc::clone(&turn_context),
client_session,
turn_metadata_header,
Arc::clone(&turn_diff_tracker),
server_model_warning_emitted_for_turn,
&prompt,
cancellation_token.child_token(),
)
.await
{
Ok(output) => {
return Ok(output);
}
Err(CodexErr::ContextWindowExceeded) => {
sess.set_total_tokens_full(&turn_context).await;
return Err(CodexErr::ContextWindowExceeded);
}
Err(CodexErr::UsageLimitReached(e)) => {
let rate_limits = e.rate_limits.clone();
if let Some(rate_limits) = rate_limits {
sess.update_rate_limits(&turn_context, *rate_limits).await;
}
return Err(CodexErr::UsageLimitReached(e));
}
Err(err) => err,
};
if !err.is_retryable() {
return Err(err);
}
// Use the configured provider-specific stream retry budget.
let max_retries = turn_context.provider.stream_max_retries();
if retries >= max_retries
&& client_session.try_switch_fallback_transport(
&turn_context.session_telemetry,
&turn_context.model_info,
)
{
sess.send_event(
&turn_context,
EventMsg::Warning(WarningEvent {
message: format!("Falling back from WebSockets to HTTPS transport. {err:#}"),
}),
)
.await;
retries = 0;
continue;
}
if retries < max_retries {
retries += 1;
let delay = match &err {
CodexErr::Stream(_, requested_delay) => {
requested_delay.unwrap_or_else(|| backoff(retries))
}
_ => backoff(retries),
};
warn!(
"stream disconnected - retrying sampling request ({retries}/{max_retries} in {delay:?})...",
);
// In release builds, hide the first websocket retry notification to reduce noisy
// transient reconnect messages. In debug builds, keep full visibility for diagnosis.
let report_error = retries > 1
|| cfg!(debug_assertions)
|| !sess
.services
.model_client
.responses_websocket_enabled(&turn_context.model_info);
if report_error {
// Surface retry information to any UI/frontend so the
// user understands what is happening instead of staring
// at a seemingly frozen screen.
sess.notify_stream_error(
&turn_context,
format!("Reconnecting... {retries}/{max_retries}"),
err,
)
.await;
}
tokio::time::sleep(delay).await;
} else {
return Err(err);
}
}
}
async fn built_tools(
sess: &Session,
turn_context: &TurnContext,
input: &[ResponseItem],
explicitly_enabled_connectors: &HashSet<String>,
skills_outcome: Option<&SkillLoadOutcome>,
cancellation_token: &CancellationToken,
) -> CodexResult<Arc<ToolRouter>> {
let mcp_connection_manager = sess.services.mcp_connection_manager.read().await;
let has_mcp_servers = mcp_connection_manager.has_servers();
let mut mcp_tools = mcp_connection_manager
.list_all_tools()
.or_cancel(cancellation_token)
.await?;
drop(mcp_connection_manager);
let loaded_plugins = sess
.services
.plugins_manager
.plugins_for_config(&turn_context.config);
let mut effective_explicitly_enabled_connectors = explicitly_enabled_connectors.clone();
effective_explicitly_enabled_connectors.extend(sess.get_connector_selection().await);
let connectors = if turn_context.features.enabled(Feature::Apps) {
let connectors = connectors::merge_plugin_apps_with_accessible(
loaded_plugins.effective_apps(),
connectors::accessible_connectors_from_mcp_tools(&mcp_tools),
);
Some(connectors::with_app_enabled_state(
connectors,
&turn_context.config,
))
} else {
None
};
// Keep the connector-grouped app view around for the router even though
// app tools only become prompt-visible after explicit selection/discovery.
let app_tools = connectors.as_ref().map(|connectors| {
filter_codex_apps_mcp_tools(&mcp_tools, connectors, &turn_context.config)
});
if let Some(connectors) = connectors.as_ref() {
let skill_name_counts_lower = skills_outcome.map_or_else(HashMap::new, |outcome| {
build_skill_name_counts(&outcome.skills, &outcome.disabled_paths).1
});
let explicitly_enabled = filter_connectors_for_input(
connectors,
input,
&effective_explicitly_enabled_connectors,
&skill_name_counts_lower,
);
let mut selected_mcp_tools = filter_non_codex_apps_mcp_tools_only(&mcp_tools);
if let Some(selected_tools) = sess.get_mcp_tool_selection().await {
selected_mcp_tools.extend(filter_mcp_tools_by_name(&mcp_tools, &selected_tools));
}
selected_mcp_tools.extend(filter_codex_apps_mcp_tools_only(
&mcp_tools,
explicitly_enabled.as_ref(),
));
mcp_tools =
connectors::filter_codex_apps_tools_by_policy(selected_mcp_tools, &turn_context.config);
}
Ok(Arc::new(ToolRouter::from_config(
&turn_context.tools_config,
has_mcp_servers.then(|| {
mcp_tools
.into_iter()
.map(|(name, tool)| (name, tool.tool))
.collect()
}),
app_tools,
turn_context.dynamic_tools.as_slice(),
)))
}
#[derive(Debug)]
struct SamplingRequestResult {
needs_follow_up: bool,
last_agent_message: Option<String>,
}
/// Ephemeral per-response state for streaming a single proposed plan.
/// This is intentionally not persisted or stored in session/state since it
/// only exists while a response is actively streaming. The final plan text
/// is extracted from the completed assistant message.
/// Tracks a single proposed plan item across a streaming response.
struct ProposedPlanItemState {
item_id: String,
started: bool,
completed: bool,
}
/// Aggregated state used only while streaming a plan-mode response.
/// Includes per-item parsers, deferred agent message bookkeeping, and the plan item lifecycle.
struct PlanModeStreamState {
/// Agent message items started by the model but deferred until we see non-plan text.
pending_agent_message_items: HashMap<String, TurnItem>,
/// Agent message items whose start notification has been emitted.
started_agent_message_items: HashSet<String>,
/// Leading whitespace buffered until we see non-whitespace text for an item.
leading_whitespace_by_item: HashMap<String, String>,
/// Tracks plan item lifecycle while streaming plan output.
plan_item_state: ProposedPlanItemState,
}
impl PlanModeStreamState {
fn new(turn_id: &str) -> Self {
Self {
pending_agent_message_items: HashMap::new(),
started_agent_message_items: HashSet::new(),
leading_whitespace_by_item: HashMap::new(),
plan_item_state: ProposedPlanItemState::new(turn_id),
}
}
}
#[derive(Debug, Default)]
struct AssistantMessageStreamParsers {
plan_mode: bool,
parsers_by_item: HashMap<String, AssistantTextStreamParser>,
}
type ParsedAssistantTextDelta = AssistantTextChunk;
impl AssistantMessageStreamParsers {
fn new(plan_mode: bool) -> Self {
Self {
plan_mode,
parsers_by_item: HashMap::new(),
}
}
fn parser_mut(&mut self, item_id: &str) -> &mut AssistantTextStreamParser {
let plan_mode = self.plan_mode;
self.parsers_by_item
.entry(item_id.to_string())
.or_insert_with(|| AssistantTextStreamParser::new(plan_mode))
}
fn seed_item_text(&mut self, item_id: &str, text: &str) -> ParsedAssistantTextDelta {
if text.is_empty() {
return ParsedAssistantTextDelta::default();
}
self.parser_mut(item_id).push_str(text)
}
fn parse_delta(&mut self, item_id: &str, delta: &str) -> ParsedAssistantTextDelta {
self.parser_mut(item_id).push_str(delta)
}
fn finish_item(&mut self, item_id: &str) -> ParsedAssistantTextDelta {
let Some(mut parser) = self.parsers_by_item.remove(item_id) else {
return ParsedAssistantTextDelta::default();
};
parser.finish()
}
fn drain_finished(&mut self) -> Vec<(String, ParsedAssistantTextDelta)> {
let parsers_by_item = std::mem::take(&mut self.parsers_by_item);
parsers_by_item
.into_iter()
.map(|(item_id, mut parser)| (item_id, parser.finish()))
.collect()
}
}
impl ProposedPlanItemState {
fn new(turn_id: &str) -> Self {
Self {
item_id: format!("{turn_id}-plan"),
started: false,
completed: false,
}
}
async fn start(&mut self, sess: &Session, turn_context: &TurnContext) {
if self.started || self.completed {
return;
}
self.started = true;
let item = TurnItem::Plan(PlanItem {
id: self.item_id.clone(),
text: String::new(),
});
sess.emit_turn_item_started(turn_context, &item).await;
}
async fn push_delta(&mut self, sess: &Session, turn_context: &TurnContext, delta: &str) {
if self.completed {
return;
}
if delta.is_empty() {
return;
}
let event = PlanDeltaEvent {
thread_id: sess.conversation_id.to_string(),
turn_id: turn_context.sub_id.clone(),
item_id: self.item_id.clone(),
delta: delta.to_string(),
};
sess.send_event(turn_context, EventMsg::PlanDelta(event))
.await;
}
async fn complete_with_text(
&mut self,
sess: &Session,
turn_context: &TurnContext,
text: String,
) {
if self.completed || !self.started {
return;
}
self.completed = true;
let item = TurnItem::Plan(PlanItem {
id: self.item_id.clone(),
text,
});
sess.emit_turn_item_completed(turn_context, item).await;
}
}
/// In plan mode we defer agent message starts until the parser emits non-plan
/// text. The parser buffers each line until it can rule out a tag prefix, so
/// plan-only outputs never show up as empty assistant messages.
async fn maybe_emit_pending_agent_message_start(
sess: &Session,
turn_context: &TurnContext,
state: &mut PlanModeStreamState,
item_id: &str,
) {
if state.started_agent_message_items.contains(item_id) {
return;
}
if let Some(item) = state.pending_agent_message_items.remove(item_id) {
sess.emit_turn_item_started(turn_context, &item).await;
state
.started_agent_message_items
.insert(item_id.to_string());
}
}
/// Agent messages are text-only today; concatenate all text entries.
fn agent_message_text(item: &codex_protocol::items::AgentMessageItem) -> String {
item.content
.iter()
.map(|entry| match entry {
codex_protocol::items::AgentMessageContent::Text { text } => text.as_str(),
})
.collect()
}
fn realtime_text_for_event(msg: &EventMsg) -> Option<String> {
match msg {
EventMsg::AgentMessage(event) => Some(event.message.clone()),
EventMsg::ItemCompleted(event) => match &event.item {
TurnItem::AgentMessage(item) => Some(agent_message_text(item)),
_ => None,
},
EventMsg::Error(_)
| EventMsg::Warning(_)
| EventMsg::RealtimeConversationStarted(_)
| EventMsg::RealtimeConversationRealtime(_)
| EventMsg::RealtimeConversationClosed(_)
| EventMsg::ModelReroute(_)
| EventMsg::ContextCompacted(_)
| EventMsg::ThreadRolledBack(_)
| EventMsg::TurnStarted(_)
| EventMsg::TurnComplete(_)
| EventMsg::TokenCount(_)
| EventMsg::UserMessage(_)
| EventMsg::AgentMessageDelta(_)
| EventMsg::AgentReasoning(_)
| EventMsg::AgentReasoningDelta(_)
| EventMsg::AgentReasoningRawContent(_)
| EventMsg::AgentReasoningRawContentDelta(_)
| EventMsg::AgentReasoningSectionBreak(_)
| EventMsg::SessionConfigured(_)
| EventMsg::ThreadNameUpdated(_)
| EventMsg::McpStartupUpdate(_)
| EventMsg::McpStartupComplete(_)
| EventMsg::McpToolCallBegin(_)
| EventMsg::McpToolCallEnd(_)
| EventMsg::WebSearchBegin(_)
| EventMsg::WebSearchEnd(_)
| EventMsg::ExecCommandBegin(_)
| EventMsg::ExecCommandOutputDelta(_)
| EventMsg::TerminalInteraction(_)
| EventMsg::ExecCommandEnd(_)
| EventMsg::PatchApplyBegin(_)
| EventMsg::PatchApplyEnd(_)
| EventMsg::ViewImageToolCall(_)
| EventMsg::ImageGenerationBegin(_)
| EventMsg::ImageGenerationEnd(_)
| EventMsg::ExecApprovalRequest(_)
| EventMsg::RequestPermissions(_)
| EventMsg::RequestUserInput(_)
| EventMsg::DynamicToolCallRequest(_)
| EventMsg::DynamicToolCallResponse(_)
| EventMsg::ElicitationRequest(_)
| EventMsg::ApplyPatchApprovalRequest(_)
| EventMsg::DeprecationNotice(_)
| EventMsg::BackgroundEvent(_)
| EventMsg::UndoStarted(_)
| EventMsg::UndoCompleted(_)
| EventMsg::StreamError(_)
| EventMsg::TurnDiff(_)
| EventMsg::GetHistoryEntryResponse(_)
| EventMsg::McpListToolsResponse(_)
| EventMsg::ListCustomPromptsResponse(_)
| EventMsg::ListSkillsResponse(_)
| EventMsg::ListRemoteSkillsResponse(_)
| EventMsg::RemoteSkillDownloaded(_)
| EventMsg::SkillsUpdateAvailable
| EventMsg::PlanUpdate(_)
| EventMsg::TurnAborted(_)
| EventMsg::ShutdownComplete
| EventMsg::EnteredReviewMode(_)
| EventMsg::ExitedReviewMode(_)
| EventMsg::RawResponseItem(_)
| EventMsg::ItemStarted(_)
| EventMsg::AgentMessageContentDelta(_)
| EventMsg::PlanDelta(_)
| EventMsg::ReasoningContentDelta(_)
| EventMsg::ReasoningRawContentDelta(_)
| EventMsg::CollabAgentSpawnBegin(_)
| EventMsg::CollabAgentSpawnEnd(_)
| EventMsg::CollabAgentInteractionBegin(_)
| EventMsg::CollabAgentInteractionEnd(_)
| EventMsg::CollabWaitingBegin(_)
| EventMsg::CollabWaitingEnd(_)
| EventMsg::CollabCloseBegin(_)
| EventMsg::CollabCloseEnd(_)
| EventMsg::CollabResumeBegin(_)
| EventMsg::CollabResumeEnd(_) => None,
}
}
/// Split the stream into normal assistant text vs. proposed plan content.
/// Normal text becomes AgentMessage deltas; plan content becomes PlanDelta +
/// TurnItem::Plan.
async fn handle_plan_segments(
sess: &Session,
turn_context: &TurnContext,
state: &mut PlanModeStreamState,
item_id: &str,
segments: Vec<ProposedPlanSegment>,
) {
for segment in segments {
match segment {
ProposedPlanSegment::Normal(delta) => {
if delta.is_empty() {
continue;
}
let has_non_whitespace = delta.chars().any(|ch| !ch.is_whitespace());
if !has_non_whitespace && !state.started_agent_message_items.contains(item_id) {
let entry = state
.leading_whitespace_by_item
.entry(item_id.to_string())
.or_default();
entry.push_str(&delta);
continue;
}
let delta = if !state.started_agent_message_items.contains(item_id) {
if let Some(prefix) = state.leading_whitespace_by_item.remove(item_id) {
format!("{prefix}{delta}")
} else {
delta
}
} else {
delta
};
maybe_emit_pending_agent_message_start(sess, turn_context, state, item_id).await;
let event = AgentMessageContentDeltaEvent {
thread_id: sess.conversation_id.to_string(),
turn_id: turn_context.sub_id.clone(),
item_id: item_id.to_string(),
delta,
};
sess.send_event(turn_context, EventMsg::AgentMessageContentDelta(event))
.await;
}
ProposedPlanSegment::ProposedPlanStart => {
if !state.plan_item_state.completed {
state.plan_item_state.start(sess, turn_context).await;
}
}
ProposedPlanSegment::ProposedPlanDelta(delta) => {
if !state.plan_item_state.completed {
if !state.plan_item_state.started {
state.plan_item_state.start(sess, turn_context).await;
}
state
.plan_item_state
.push_delta(sess, turn_context, &delta)
.await;
}
}
ProposedPlanSegment::ProposedPlanEnd => {}
}
}
}
async fn emit_streamed_assistant_text_delta(
sess: &Session,
turn_context: &TurnContext,
plan_mode_state: Option<&mut PlanModeStreamState>,
item_id: &str,
parsed: ParsedAssistantTextDelta,
) {
if parsed.is_empty() {
return;
}
if !parsed.citations.is_empty() {
// Citation extraction is intentionally local for now; we strip citations from display text
// but do not yet surface them in protocol events.
let _citations = parsed.citations;
}
if let Some(state) = plan_mode_state {
if !parsed.plan_segments.is_empty() {
handle_plan_segments(sess, turn_context, state, item_id, parsed.plan_segments).await;
}
return;
}
if parsed.visible_text.is_empty() {
return;
}
let event = AgentMessageContentDeltaEvent {
thread_id: sess.conversation_id.to_string(),
turn_id: turn_context.sub_id.clone(),
item_id: item_id.to_string(),
delta: parsed.visible_text,
};
sess.send_event(turn_context, EventMsg::AgentMessageContentDelta(event))
.await;
}
/// Flush buffered assistant text parser state when an assistant message item ends.
async fn flush_assistant_text_segments_for_item(
sess: &Session,
turn_context: &TurnContext,
plan_mode_state: Option<&mut PlanModeStreamState>,
parsers: &mut AssistantMessageStreamParsers,
item_id: &str,
) {
let parsed = parsers.finish_item(item_id);
emit_streamed_assistant_text_delta(sess, turn_context, plan_mode_state, item_id, parsed).await;
}
/// Flush any remaining buffered assistant text parser state at response completion.
async fn flush_assistant_text_segments_all(
sess: &Session,
turn_context: &TurnContext,
mut plan_mode_state: Option<&mut PlanModeStreamState>,
parsers: &mut AssistantMessageStreamParsers,
) {
for (item_id, parsed) in parsers.drain_finished() {
emit_streamed_assistant_text_delta(
sess,
turn_context,
plan_mode_state.as_deref_mut(),
&item_id,
parsed,
)
.await;
}
}
/// Emit completion for plan items by parsing the finalized assistant message.
async fn maybe_complete_plan_item_from_message(
sess: &Session,
turn_context: &TurnContext,
state: &mut PlanModeStreamState,
item: &ResponseItem,
) {
if let ResponseItem::Message { role, content, .. } = item
&& role == "assistant"
{
let mut text = String::new();
for entry in content {
if let ContentItem::OutputText { text: chunk } = entry {
text.push_str(chunk);
}
}
if let Some(plan_text) = extract_proposed_plan_text(&text) {
let (plan_text, _citations) = strip_citations(&plan_text);
if !state.plan_item_state.started {
state.plan_item_state.start(sess, turn_context).await;
}
state
.plan_item_state
.complete_with_text(sess, turn_context, plan_text)
.await;
}
}
}
/// Emit a completed agent message in plan mode, respecting deferred starts.
async fn emit_agent_message_in_plan_mode(
sess: &Session,
turn_context: &TurnContext,
agent_message: codex_protocol::items::AgentMessageItem,
state: &mut PlanModeStreamState,
) {
let agent_message_id = agent_message.id.clone();
let text = agent_message_text(&agent_message);
if text.trim().is_empty() {
state.pending_agent_message_items.remove(&agent_message_id);
state.started_agent_message_items.remove(&agent_message_id);
return;
}
maybe_emit_pending_agent_message_start(sess, turn_context, state, &agent_message_id).await;
if !state
.started_agent_message_items
.contains(&agent_message_id)
{
let start_item = state
.pending_agent_message_items
.remove(&agent_message_id)
.unwrap_or_else(|| {
TurnItem::AgentMessage(codex_protocol::items::AgentMessageItem {
id: agent_message_id.clone(),
content: Vec::new(),
phase: None,
})
});
sess.emit_turn_item_started(turn_context, &start_item).await;
state
.started_agent_message_items
.insert(agent_message_id.clone());
}
sess.emit_turn_item_completed(turn_context, TurnItem::AgentMessage(agent_message))
.await;
state.started_agent_message_items.remove(&agent_message_id);
}
/// Emit completion for a plan-mode turn item, handling agent messages specially.
async fn emit_turn_item_in_plan_mode(
sess: &Session,
turn_context: &TurnContext,
turn_item: TurnItem,
previously_active_item: Option<&TurnItem>,
state: &mut PlanModeStreamState,
) {
match turn_item {
TurnItem::AgentMessage(agent_message) => {
emit_agent_message_in_plan_mode(sess, turn_context, agent_message, state).await;
}
_ => {
if previously_active_item.is_none() {
sess.emit_turn_item_started(turn_context, &turn_item).await;
}
sess.emit_turn_item_completed(turn_context, turn_item).await;
}
}
}
/// Handle a completed assistant response item in plan mode, returning true if handled.
async fn handle_assistant_item_done_in_plan_mode(
sess: &Session,
turn_context: &TurnContext,
item: &ResponseItem,
state: &mut PlanModeStreamState,
previously_active_item: Option<&TurnItem>,
last_agent_message: &mut Option<String>,
) -> bool {
if let ResponseItem::Message { role, .. } = item
&& role == "assistant"
{
maybe_complete_plan_item_from_message(sess, turn_context, state, item).await;
if let Some(turn_item) =
handle_non_tool_response_item(item, true, Some(&turn_context.cwd)).await
{
emit_turn_item_in_plan_mode(
sess,
turn_context,
turn_item,
previously_active_item,
state,
)
.await;
}
record_completed_response_item(sess, turn_context, item).await;
if let Some(agent_message) = last_assistant_message_from_item(item, true) {
*last_agent_message = Some(agent_message);
}
return true;
}
false
}
async fn drain_in_flight(
in_flight: &mut FuturesOrdered<BoxFuture<'static, CodexResult<ResponseInputItem>>>,
sess: Arc<Session>,
turn_context: Arc<TurnContext>,
) -> CodexResult<()> {
while let Some(res) = in_flight.next().await {
match res {
Ok(response_input) => {
sess.record_conversation_items(&turn_context, &[response_input.into()])
.await;
}
Err(err) => {
error_or_panic(format!("in-flight tool future failed during drain: {err}"));
}
}
}
Ok(())
}
#[allow(clippy::too_many_arguments)]
#[instrument(level = "trace",
skip_all,
fields(
turn_id = %turn_context.sub_id,
model = %turn_context.model_info.slug
)
)]
async fn try_run_sampling_request(
router: Arc<ToolRouter>,
sess: Arc<Session>,
turn_context: Arc<TurnContext>,
client_session: &mut ModelClientSession,
turn_metadata_header: Option<&str>,
turn_diff_tracker: SharedTurnDiffTracker,
server_model_warning_emitted_for_turn: &mut bool,
prompt: &Prompt,
cancellation_token: CancellationToken,
) -> CodexResult<SamplingRequestResult> {
feedback_tags!(
model = turn_context.model_info.slug.clone(),
approval_policy = turn_context.approval_policy.value(),
sandbox_policy = turn_context.sandbox_policy.get(),
effort = turn_context.reasoning_effort,
auth_mode = sess.services.auth_manager.auth_mode(),
features = sess.features.enabled_features(),
);
let mut stream = client_session
.stream(
prompt,
&turn_context.model_info,
&turn_context.session_telemetry,
turn_context.reasoning_effort,
turn_context.reasoning_summary,
turn_context.config.service_tier,
turn_metadata_header,
)
.instrument(trace_span!("stream_request"))
.or_cancel(&cancellation_token)
.await??;
let tool_runtime = ToolCallRuntime::new(
Arc::clone(&router),
Arc::clone(&sess),
Arc::clone(&turn_context),
Arc::clone(&turn_diff_tracker),
);
let mut in_flight: FuturesOrdered<BoxFuture<'static, CodexResult<ResponseInputItem>>> =
FuturesOrdered::new();
let mut needs_follow_up = false;
let mut last_agent_message: Option<String> = None;
let mut active_item: Option<TurnItem> = None;
let mut should_emit_turn_diff = false;
let plan_mode = turn_context.collaboration_mode.mode == ModeKind::Plan;
let mut assistant_message_stream_parsers = AssistantMessageStreamParsers::new(plan_mode);
let mut plan_mode_state = plan_mode.then(|| PlanModeStreamState::new(&turn_context.sub_id));
let receiving_span = trace_span!("receiving_stream");
let outcome: CodexResult<SamplingRequestResult> = loop {
let handle_responses = trace_span!(
parent: &receiving_span,
"handle_responses",
otel.name = field::Empty,
tool_name = field::Empty,
from = field::Empty,
);
let event = match stream
.next()
.instrument(trace_span!(parent: &handle_responses, "receiving"))
.or_cancel(&cancellation_token)
.await
{
Ok(event) => event,
Err(codex_async_utils::CancelErr::Cancelled) => break Err(CodexErr::TurnAborted),
};
let event = match event {
Some(res) => res?,
None => {
break Err(CodexErr::Stream(
"stream closed before response.completed".into(),
None,
));
}
};
sess.services
.session_telemetry
.record_responses(&handle_responses, &event);
record_turn_ttft_metric(&turn_context, &event).await;
match event {
ResponseEvent::Created => {}
ResponseEvent::OutputItemDone(item) => {
let previously_active_item = active_item.take();
if let Some(previous) = previously_active_item.as_ref()
&& matches!(previous, TurnItem::AgentMessage(_))
{
let item_id = previous.id();
flush_assistant_text_segments_for_item(
&sess,
&turn_context,
plan_mode_state.as_mut(),
&mut assistant_message_stream_parsers,
&item_id,
)
.await;
}
if let Some(state) = plan_mode_state.as_mut()
&& handle_assistant_item_done_in_plan_mode(
&sess,
&turn_context,
&item,
state,
previously_active_item.as_ref(),
&mut last_agent_message,
)
.await
{
continue;
}
let mut ctx = HandleOutputCtx {
sess: sess.clone(),
turn_context: turn_context.clone(),
tool_runtime: tool_runtime.clone(),
cancellation_token: cancellation_token.child_token(),
};
let output_result = handle_output_item_done(&mut ctx, item, previously_active_item)
.instrument(handle_responses)
.await?;
if let Some(tool_future) = output_result.tool_future {
in_flight.push_back(tool_future);
}
if let Some(agent_message) = output_result.last_agent_message {
last_agent_message = Some(agent_message);
}
needs_follow_up |= output_result.needs_follow_up;
}
ResponseEvent::OutputItemAdded(item) => {
if let Some(turn_item) =
handle_non_tool_response_item(&item, plan_mode, Some(&turn_context.cwd)).await
{
let mut turn_item = turn_item;
let mut seeded_parsed: Option<ParsedAssistantTextDelta> = None;
let mut seeded_item_id: Option<String> = None;
if matches!(turn_item, TurnItem::AgentMessage(_))
&& let Some(raw_text) = raw_assistant_output_text_from_item(&item)
{
let item_id = turn_item.id();
let mut seeded =
assistant_message_stream_parsers.seed_item_text(&item_id, &raw_text);
if let TurnItem::AgentMessage(agent_message) = &mut turn_item {
agent_message.content =
vec![codex_protocol::items::AgentMessageContent::Text {
text: if plan_mode {
String::new()
} else {
std::mem::take(&mut seeded.visible_text)
},
}];
}
seeded_parsed = plan_mode.then_some(seeded);
seeded_item_id = Some(item_id);
}
if let Some(state) = plan_mode_state.as_mut()
&& matches!(turn_item, TurnItem::AgentMessage(_))
{
let item_id = turn_item.id();
state
.pending_agent_message_items
.insert(item_id, turn_item.clone());
} else {
sess.emit_turn_item_started(&turn_context, &turn_item).await;
}
if let (Some(state), Some(item_id), Some(parsed)) = (
plan_mode_state.as_mut(),
seeded_item_id.as_deref(),
seeded_parsed,
) {
emit_streamed_assistant_text_delta(
&sess,
&turn_context,
Some(state),
item_id,
parsed,
)
.await;
}
active_item = Some(turn_item);
}
}
ResponseEvent::ServerModel(server_model) => {
if !*server_model_warning_emitted_for_turn
&& sess
.maybe_warn_on_server_model_mismatch(&turn_context, server_model)
.await
{
*server_model_warning_emitted_for_turn = true;
}
}
ResponseEvent::ServerReasoningIncluded(included) => {
sess.set_server_reasoning_included(included).await;
}
ResponseEvent::RateLimits(snapshot) => {
// Update internal state with latest rate limits, but defer sending until
// token usage is available to avoid duplicate TokenCount events.
sess.update_rate_limits(&turn_context, snapshot).await;
}
ResponseEvent::ModelsEtag(etag) => {
// Update internal state with latest models etag
sess.services.models_manager.refresh_if_new_etag(etag).await;
}
ResponseEvent::Completed {
response_id: _,
token_usage,
} => {
flush_assistant_text_segments_all(
&sess,
&turn_context,
plan_mode_state.as_mut(),
&mut assistant_message_stream_parsers,
)
.await;
sess.update_token_usage_info(&turn_context, token_usage.as_ref())
.await;
should_emit_turn_diff = true;
needs_follow_up |= sess.has_pending_input().await;
break Ok(SamplingRequestResult {
needs_follow_up,
last_agent_message,
});
}
ResponseEvent::OutputTextDelta(delta) => {
// In review child threads, suppress assistant text deltas; the
// UI will show a selection popup from the final ReviewOutput.
if let Some(active) = active_item.as_ref() {
let item_id = active.id();
if matches!(active, TurnItem::AgentMessage(_)) {
let parsed = assistant_message_stream_parsers.parse_delta(&item_id, &delta);
emit_streamed_assistant_text_delta(
&sess,
&turn_context,
plan_mode_state.as_mut(),
&item_id,
parsed,
)
.await;
} else {
let event = AgentMessageContentDeltaEvent {
thread_id: sess.conversation_id.to_string(),
turn_id: turn_context.sub_id.clone(),
item_id,
delta,
};
sess.send_event(&turn_context, EventMsg::AgentMessageContentDelta(event))
.await;
}
} else {
error_or_panic("OutputTextDelta without active item".to_string());
}
}
ResponseEvent::ReasoningSummaryDelta {
delta,
summary_index,
} => {
if let Some(active) = active_item.as_ref() {
let event = ReasoningContentDeltaEvent {
thread_id: sess.conversation_id.to_string(),
turn_id: turn_context.sub_id.clone(),
item_id: active.id(),
delta,
summary_index,
};
sess.send_event(&turn_context, EventMsg::ReasoningContentDelta(event))
.await;
} else {
error_or_panic("ReasoningSummaryDelta without active item".to_string());
}
}
ResponseEvent::ReasoningSummaryPartAdded { summary_index } => {
if let Some(active) = active_item.as_ref() {
let event =
EventMsg::AgentReasoningSectionBreak(AgentReasoningSectionBreakEvent {
item_id: active.id(),
summary_index,
});
sess.send_event(&turn_context, event).await;
} else {
error_or_panic("ReasoningSummaryPartAdded without active item".to_string());
}
}
ResponseEvent::ReasoningContentDelta {
delta,
content_index,
} => {
if let Some(active) = active_item.as_ref() {
let event = ReasoningRawContentDeltaEvent {
thread_id: sess.conversation_id.to_string(),
turn_id: turn_context.sub_id.clone(),
item_id: active.id(),
delta,
content_index,
};
sess.send_event(&turn_context, EventMsg::ReasoningRawContentDelta(event))
.await;
} else {
error_or_panic("ReasoningRawContentDelta without active item".to_string());
}
}
}
};
flush_assistant_text_segments_all(
&sess,
&turn_context,
plan_mode_state.as_mut(),
&mut assistant_message_stream_parsers,
)
.await;
drain_in_flight(&mut in_flight, sess.clone(), turn_context.clone()).await?;
if cancellation_token.is_cancelled() {
return Err(CodexErr::TurnAborted);
}
if should_emit_turn_diff {
let unified_diff = {
let mut tracker = turn_diff_tracker.lock().await;
tracker.get_unified_diff()
};
if let Ok(Some(unified_diff)) = unified_diff {
let msg = EventMsg::TurnDiff(TurnDiffEvent { unified_diff });
sess.clone().send_event(&turn_context, msg).await;
}
}
outcome
}
pub(super) fn get_last_assistant_message_from_turn(responses: &[ResponseItem]) -> Option<String> {
for item in responses.iter().rev() {
if let Some(message) = last_assistant_message_from_item(item, false) {
return Some(message);
}
}
None
}
use crate::memories::prompts::build_memory_tool_developer_instructions;
#[cfg(test)]
pub(crate) use tests::make_session_and_context;
#[cfg(test)]
pub(crate) use tests::make_session_and_context_with_dynamic_tools_and_rx;
#[cfg(test)]
pub(crate) use tests::make_session_and_context_with_rx;
#[cfg(test)]
pub(crate) use tests::make_session_configuration_for_tests;
#[cfg(test)]
#[path = "codex_tests.rs"]
mod tests;