[codex] Make MCP connection startup fallible (#27261)

## Why

Required MCP server startup was enforced in `Session::new` after
`McpConnectionManager` had already created the clients. That split let
other manager construction paths bypass the same requirement and exposed
manager internals solely so the session could validate them. Keeping
required-server readiness in the constructor gives every caller one
consistent startup contract.

## What changed

- make `McpConnectionManager::new` return `anyhow::Result<Self>` and
fail when an enabled, required server cannot initialize
- pass the startup cancellation token into the constructor so
required-server waits remain cancellable
- propagate constructor failures through resource reads, connector
discovery, and MCP status collection
- preserve the active manager and cancellation token when a refreshed
replacement fails
- keep required-startup failure collection private and cover the
constructor error contract directly

## Validation

- updated the focused connection-manager test to assert the complete
required-server startup error
- local tests not run; relying on CI
This commit is contained in:
Ahmed Ibrahim
2026-06-10 00:17:58 -07:00
committed by GitHub
Unverified
parent 00a25e1e0c
commit a7b6baecc6
8 changed files with 129 additions and 107 deletions
+62 -31
View File
@@ -68,6 +68,7 @@ use serde_json::Value as JsonValue;
use tokio::task::JoinSet;
use tokio_util::sync::CancellationToken;
use tracing::Instrument;
use tracing::info_span;
use tracing::instrument;
use tracing::trace;
use tracing::trace_span;
@@ -105,6 +106,7 @@ pub fn tool_is_model_visible(tool: &ToolInfo) -> bool {
pub struct McpConnectionManager {
clients: HashMap<String, AsyncManagedClient>,
server_metadata: HashMap<String, McpServerMetadata>,
required_servers: Vec<String>,
tool_plugin_provenance: Arc<ToolPluginProvenance>,
host_owned_codex_apps_enabled: bool,
prefix_mcp_tool_names: bool,
@@ -113,7 +115,7 @@ pub struct McpConnectionManager {
}
impl McpConnectionManager {
#[allow(clippy::new_ret_no_self, clippy::too_many_arguments)]
#[allow(clippy::too_many_arguments)]
pub async fn new(
mcp_servers: &HashMap<String, EffectiveMcpServer>,
store_mode: OAuthCredentialsStoreMode,
@@ -121,6 +123,7 @@ impl McpConnectionManager {
approval_policy: &Constrained<AskForApproval>,
submit_id: String,
tx_event: Sender<Event>,
startup_cancellation_token: CancellationToken,
initial_permission_profile: PermissionProfile,
runtime_context: McpRuntimeContext,
codex_home: PathBuf,
@@ -131,8 +134,13 @@ impl McpConnectionManager {
tool_plugin_provenance: ToolPluginProvenance,
auth: Option<&CodexAuth>,
elicitation_reviewer: Option<ElicitationReviewerHandle>,
) -> (Self, CancellationToken) {
let cancel_token = CancellationToken::new();
) -> Self {
let mut required_servers = mcp_servers
.iter()
.filter(|(_, server)| server.enabled() && server.required())
.map(|(name, _)| name.clone())
.collect::<Vec<_>>();
required_servers.sort();
let mut clients = HashMap::new();
let mut server_metadata = HashMap::new();
let mut join_set = JoinSet::new();
@@ -152,7 +160,7 @@ impl McpConnectionManager {
.filter(|(_, server)| server.enabled())
{
server_metadata.insert(server_name.clone(), McpServerMetadata::from(&server));
let cancel_token = cancel_token.child_token();
let cancel_token = startup_cancellation_token.child_token();
let _ = emit_update(
startup_submit_id.as_str(),
&tx_event,
@@ -237,11 +245,12 @@ impl McpConnectionManager {
let manager = Self {
clients,
server_metadata,
required_servers,
tool_plugin_provenance,
host_owned_codex_apps_enabled,
prefix_mcp_tool_names,
elicitation_requests: elicitation_requests.clone(),
startup_cancellation_token: cancel_token.clone(),
startup_cancellation_token: startup_cancellation_token.clone(),
};
tokio::spawn(async move {
let outcomes = join_set.join_all().await;
@@ -265,7 +274,53 @@ impl McpConnectionManager {
})
.await;
});
(manager, cancel_token)
manager
}
/// Waits for every required server and reports their startup failures together.
///
/// Callers must make the manager reachable to request handlers before awaiting this method,
/// because server initialization may require client elicitation.
pub async fn validate_required_servers(&self) -> Result<()> {
let failures = async {
let mut failures = Vec::new();
for server_name in &self.required_servers {
let Some(async_managed_client) = self.clients.get(server_name).cloned() else {
failures.push(McpStartupFailure {
server: server_name.clone(),
error: format!("required MCP server `{server_name}` was not initialized"),
});
continue;
};
match async_managed_client.client().await {
Ok(_) => {}
Err(error) => failures.push(McpStartupFailure {
server: server_name.clone(),
error: startup_outcome_error_message(error),
}),
}
}
failures
}
.instrument(info_span!(
"session_init.required_mcp_wait",
otel.name = "session_init.required_mcp_wait",
session_init.required_mcp_server_count = self.required_servers.len(),
))
.await;
if failures.is_empty() {
return Ok(());
}
let details = failures
.iter()
.map(|failure| format!("{}: {}", failure.server, failure.error))
.collect::<Vec<_>>()
.join("; ");
Err(anyhow!(
"required MCP servers failed to initialize: {details}"
))
}
pub fn new_uninitialized_with_permission_profile(
@@ -276,6 +331,7 @@ impl McpConnectionManager {
Self {
clients: HashMap::new(),
server_metadata: HashMap::new(),
required_servers: Vec::new(),
tool_plugin_provenance: Arc::new(ToolPluginProvenance::default()),
host_owned_codex_apps_enabled: false,
prefix_mcp_tool_names,
@@ -374,31 +430,6 @@ impl McpConnectionManager {
}
}
pub async fn required_startup_failures(
&self,
required_servers: &[String],
) -> Vec<McpStartupFailure> {
let mut failures = Vec::new();
for server_name in required_servers {
let Some(async_managed_client) = self.clients.get(server_name).cloned() else {
failures.push(McpStartupFailure {
server: server_name.clone(),
error: format!("required MCP server `{server_name}` was not initialized"),
});
continue;
};
match async_managed_client.client().await {
Ok(_) => {}
Err(error) => failures.push(McpStartupFailure {
server: server_name.clone(),
error: startup_outcome_error_message(error),
}),
}
}
failures
}
/// Returns all tools with model-visible names normalized.
#[instrument(level = "trace", skip_all, fields(mcp_server_count = self.clients.len()))]
pub async fn list_all_tools(&self) -> Vec<ToolInfo> {
@@ -1167,13 +1167,15 @@ async fn no_local_runtime_fails_local_stdio_but_keeps_local_http_server() {
),
]);
let (manager, cancel_token) = McpConnectionManager::new(
let cancel_token = CancellationToken::new();
let manager = McpConnectionManager::new(
&mcp_servers,
OAuthCredentialsStoreMode::default(),
HashMap::new(),
&approval_policy,
String::new(),
tx_event,
cancel_token.clone(),
PermissionProfile::default(),
McpRuntimeContext::new(
Arc::new(EnvironmentManager::without_environments()),
@@ -1201,13 +1203,18 @@ async fn no_local_runtime_fails_local_stdio_but_keeps_local_http_server() {
.wait_for_server_ready("stdio", Duration::from_millis(10))
.await
);
let failures = manager
.required_startup_failures(&["stdio".to_string()])
.await;
assert_eq!(failures.len(), 1);
assert_eq!(failures[0].server, "stdio");
let error = match manager
.clients
.get("stdio")
.expect("stdio client")
.client()
.await
{
Ok(_) => panic!("local stdio MCP startup should fail"),
Err(error) => error,
};
assert_eq!(
failures[0].error,
startup_outcome_error_message(error),
"local stdio MCP server `stdio` requires a local environment"
);
cancel_token.cancel();
+7 -2
View File
@@ -35,6 +35,7 @@ use rmcp::model::ElicitationCapability;
use rmcp::model::ReadResourceRequestParams;
use rmcp::model::ReadResourceResult;
use serde_json::Value;
use tokio_util::sync::CancellationToken;
use crate::codex_apps::codex_apps_tools_cache_key;
use crate::connection_manager::McpConnectionManager;
@@ -293,13 +294,15 @@ pub async fn read_mcp_resource(
.await;
let (tx_event, rx_event) = unbounded();
drop(rx_event);
let (manager, cancel_token) = McpConnectionManager::new(
let cancel_token = CancellationToken::new();
let manager = McpConnectionManager::new(
&mcp_servers,
config.mcp_oauth_credentials_store_mode,
auth_statuses,
&config.approval_policy,
String::new(),
tx_event,
cancel_token.clone(),
PermissionProfile::default(),
runtime_context,
config.codex_home.clone(),
@@ -363,13 +366,15 @@ pub async fn collect_mcp_server_status_snapshot_with_detail(
let (tx_event, rx_event) = unbounded();
drop(rx_event);
let (mcp_connection_manager, cancel_token) = McpConnectionManager::new(
let cancel_token = CancellationToken::new();
let mcp_connection_manager = McpConnectionManager::new(
&mcp_servers,
config.mcp_oauth_credentials_store_mode,
auth_status_entries.clone(),
&config.approval_policy,
submit_id,
tx_event,
cancel_token.clone(),
PermissionProfile::default(),
runtime_context,
config.codex_home.clone(),
+4 -1
View File
@@ -17,6 +17,7 @@ use codex_protocol::models::PermissionProfile;
use codex_tools::DiscoverableTool;
use rmcp::model::ToolAnnotations;
use serde::Deserialize;
use tokio_util::sync::CancellationToken;
use tracing::warn;
use crate::config::Config;
@@ -285,13 +286,15 @@ pub async fn list_accessible_connectors_from_mcp_tools_with_mcp_manager(
let (tx_event, rx_event) = unbounded();
drop(rx_event);
let (mut mcp_connection_manager, cancel_token) = McpConnectionManager::new(
let cancel_token = CancellationToken::new();
let mut mcp_connection_manager = McpConnectionManager::new(
&mcp_servers,
config.mcp_oauth_credentials_store_mode,
auth_status_entries,
&config.permissions.approval_policy,
INITIAL_SUBMIT_ID.to_owned(),
tx_event,
cancel_token.clone(),
PermissionProfile::default(),
// Connector discovery is threadless. Use an actually configured env if
// one exists, but do not reintroduce the old hidden-local fallback.
+3 -1
View File
@@ -46,6 +46,7 @@ use std::path::Path;
use std::path::PathBuf;
use std::sync::Arc;
use tempfile::tempdir;
use tokio_util::sync::CancellationToken;
use tracing::Instrument;
use tracing::Level;
use tracing_subscriber::fmt::format::FmtSpan;
@@ -1256,13 +1257,14 @@ fn codex_apps_auth_failure_metadata() -> McpToolApprovalMetadata {
async fn install_host_owned_codex_apps_manager(session: &Session, turn_context: &TurnContext) {
let auth = session.services.auth_manager.auth().await;
let (manager, _cancel_token) = codex_mcp::McpConnectionManager::new(
let manager = codex_mcp::McpConnectionManager::new(
&HashMap::new(),
turn_context.config.mcp_oauth_credentials_store_mode,
HashMap::new(),
&turn_context.approval_policy,
turn_context.sub_id.clone(),
session.get_tx_event(),
CancellationToken::new(),
turn_context.permission_profile(),
codex_mcp::McpRuntimeContext::new(Arc::clone(&session.services.environment_manager), {
#[allow(deprecated)]
+7 -12
View File
@@ -338,18 +338,21 @@ impl Session {
turn_context.cwd.to_path_buf(),
),
};
{
let mcp_startup_cancellation_token = {
let mut guard = self.services.mcp_startup_cancellation_token.lock().await;
guard.cancel();
*guard = CancellationToken::new();
}
let (refreshed_manager, cancel_token) = McpConnectionManager::new(
let cancellation_token = CancellationToken::new();
*guard = cancellation_token.clone();
cancellation_token
};
let refreshed_manager = McpConnectionManager::new(
&mcp_servers,
store_mode,
auth_statuses,
&turn_context.approval_policy,
turn_context.sub_id.clone(),
self.get_tx_event(),
mcp_startup_cancellation_token,
turn_context.permission_profile(),
mcp_runtime_context,
config.codex_home.to_path_buf(),
@@ -366,14 +369,6 @@ impl Session {
let current_manager = self.services.mcp_connection_manager.read().await;
refreshed_manager.set_elicitations_auto_deny(current_manager.elicitations_auto_deny());
}
{
let mut guard = self.services.mcp_startup_cancellation_token.lock().await;
if guard.is_cancelled() {
cancel_token.cancel();
}
*guard = cancel_token;
}
let mut old_manager = {
let mut manager = self.services.mcp_connection_manager.write().await;
std::mem::replace(&mut *manager, refreshed_manager)
+11 -53
View File
@@ -469,10 +469,6 @@ impl Session {
#[instrument(name = "session_init", level = "info", skip_all)]
#[allow(clippy::too_many_arguments)]
#[expect(
clippy::await_holding_invalid_type,
reason = "session initialization must serialize access through session-owned manager guards"
)]
pub(crate) async fn new(
mut session_configuration: SessionConfiguration,
config: Arc<Config>,
@@ -1116,15 +1112,6 @@ impl Session {
sess.send_event_raw(event).await;
}
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 enabled_mcp_server_count =
mcp_servers.values().filter(|server| server.enabled()).count();
let required_mcp_server_count = required_mcp_servers.len();
let tool_plugin_provenance = mcp_manager.tool_plugin_provenance(config.as_ref()).await;
let host_owned_codex_apps_enabled = config
.features
@@ -1137,11 +1124,13 @@ impl Session {
} else {
ElicitationCapability::default()
};
{
let mcp_startup_cancellation_token = {
let mut cancel_guard = sess.services.mcp_startup_cancellation_token.lock().await;
cancel_guard.cancel();
*cancel_guard = CancellationToken::new();
}
let cancel_token = CancellationToken::new();
*cancel_guard = cancel_token.clone();
cancel_token
};
let turn_environment = crate::environment_selection::resolve_environment_selections(
sess.services.environment_manager.as_ref(),
session_configuration.environment_selections(),
@@ -1164,13 +1153,14 @@ impl Session {
session_configuration.cwd().to_path_buf(),
),
};
let (mcp_connection_manager, cancel_token) = McpConnectionManager::new(
let mcp_connection_manager = McpConnectionManager::new(
&mcp_servers,
config.mcp_oauth_credentials_store_mode,
auth_statuses.clone(),
auth_statuses,
&session_configuration.approval_policy,
INITIAL_SUBMIT_ID.to_owned(),
tx_event.clone(),
mcp_startup_cancellation_token,
session_configuration.permission_profile(),
mcp_runtime_context,
config.codex_home.to_path_buf(),
@@ -1185,43 +1175,11 @@ impl Session {
.instrument(info_span!(
"session_init.mcp_manager_init",
otel.name = "session_init.mcp_manager_init",
session_init.enabled_mcp_server_count = enabled_mcp_server_count,
session_init.required_mcp_server_count = required_mcp_server_count,
))
.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)
.instrument(info_span!(
"session_init.required_mcp_wait",
otel.name = "session_init.required_mcp_wait",
session_init.required_mcp_server_count = required_mcp_server_count,
))
.await;
if !failures.is_empty() {
let details = failures
.iter()
.map(|failure| format!("{}: {}", failure.server, failure.error))
.collect::<Vec<_>>()
.join("; ");
anyhow::bail!("required MCP servers failed to initialize: {details}");
}
}
sess.services
.install_mcp_connection_manager(mcp_connection_manager)
.await?;
sess.schedule_startup_prewarm(session_configuration.base_instructions.clone())
.await;
let session_start_source = match &initial_history {
+21
View File
@@ -15,6 +15,7 @@ use crate::tools::code_mode::CodeModeService;
use crate::tools::network_approval::NetworkApprovalService;
use crate::tools::sandboxing::ApprovalStore;
use crate::unified_exec::UnifiedExecProcessManager;
use anyhow::Result;
use arc_swap::ArcSwap;
use arc_swap::ArcSwapOption;
use codex_analytics::AnalyticsEventsClient;
@@ -82,3 +83,23 @@ pub(crate) struct SessionServices {
/// the same manager through child-thread spawn paths without reconstructing it.
pub(crate) environment_manager: Arc<EnvironmentManager>,
}
impl SessionServices {
/// Installs the manager before validating required servers so startup-time elicitation can
/// resolve through the session's manager while validation waits.
#[expect(
clippy::await_holding_invalid_type,
reason = "required MCP validation keeps the installed manager reachable for startup-time elicitation"
)]
pub(crate) async fn install_mcp_connection_manager(
&self,
manager: McpConnectionManager,
) -> Result<()> {
*self.mcp_connection_manager.write().await = manager;
self.mcp_connection_manager
.read()
.await
.validate_required_servers()
.await
}
}