core: surface terminal subagent errors to parent agents (#28375)

## Why

When a subagent exhausts its retries, it emits an `Error`, but the
generic task lifecycle then emits `TurnComplete(None)`. That completion
used to overwrite the subagent's `Errored` status with
`Completed(None)`, so the parent received an empty completion
notification.

This made a failed child look indistinguishable from a child that
completed without an answer. In unattended or long-running multi-agent
work, the root could silently continue without knowing that delegated
work failed or how to restart it.

## Behavior

Before, a terminal stream failure was reduced to an empty completion:

```text
<subagent_notification>
{"agent_path":"/root/worker","status":{"completed":null}}
</subagent_notification>
```

Now the parent receives the actual terminal error, bounded to 1,000
tokens, together with an actionable recovery hint:

```text
<subagent_notification>
{
  "agent_path": "/root/worker",
  "status": {
    "errored": "stream disconnected before completion: stream closed before response.completed"
  },
  "next_action": "This agent's turn failed. If you still need this agent, use `followup_task` to give it another task."
}
</subagent_notification>
```

The notification remains queue-only: it does not wake the root or replay
the failed request. The root sees it at the next sampling boundary and
can use `followup_task` to start a new turn for that agent.

## What changed

- Added terminal-error precedence to the [agent status
reducer](https://github.com/openai/codex/blob/e95fcfe2bb6a02f1a75650afa20048859f556511/codex-rs/core/src/agent/status.rs#L23-L34),
so a closing `TurnComplete` cannot erase an immediately preceding
`Errored` status.
- Made MultiAgentV2 completion forwarding use the retained session
status instead of re-deriving `Completed(None)` from the final event.
- Extended the [subagent notification
fragment](https://github.com/openai/codex/blob/e95fcfe2bb6a02f1a75650afa20048859f556511/codex-rs/core/src/context/subagent_notification.rs#L6-L60)
with a `next_action` for terminal errors and a hard cap on model-visible
error text.
- Kept successful completions and interrupted turns unchanged.

## Verification

- Added a status-reducer test proving that `Errored` survives the
trailing `TurnComplete`.
- Added an integration test that exhausts a subagent's stream retries
and verifies the exact `agent_message` delivered to the parent,
including the error and `followup_task` guidance.
- Re-ran the existing successful-completion and interrupted-turn
notification tests.
This commit is contained in:
jif
2026-06-16 14:34:54 +02:00
committed by GitHub
parent de1f77bfdd
commit 1b24ba912a
7 changed files with 104 additions and 13 deletions
+2 -1
View File
@@ -25,6 +25,7 @@ use codex_protocol::items::TurnItem;
use codex_protocol::items::UserMessageItem;
use codex_protocol::models::ResponseItem;
use codex_protocol::protocol::AskForApproval;
use codex_protocol::protocol::CodexErrorInfo;
use codex_protocol::protocol::EventMsg;
use codex_protocol::protocol::HookCompletedEvent;
use codex_protocol::protocol::HookEventName;
@@ -490,7 +491,7 @@ pub(crate) async fn run_legacy_after_agent_hook(
};
let event = EventMsg::Error(codex_protocol::protocol::ErrorEvent {
message,
codex_error_info: None,
codex_error_info: Some(CodexErrorInfo::Other),
});
sess.send_event(turn_context, event).await;
true
+24 -2
View File
@@ -1654,6 +1654,18 @@ impl Session {
/// 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();
if let EventMsg::Error(error) = &legacy_source
&& error
.codex_error_info
.as_ref()
.is_some_and(CodexErrorInfo::affects_turn_status)
{
turn_context
.terminal_error
.lock()
.await
.replace(error.message.clone());
}
self.services
.rollout_thread_trace
.record_codex_turn_event(&turn_context.sub_id, &legacy_source);
@@ -1705,8 +1717,18 @@ impl Session {
return;
};
let Some(status) = agent_status_from_event(msg) else {
return;
let status = match turn_context.terminal_error.lock().await.take() {
Some(error) => {
let status = AgentStatus::Errored(error);
self.agent_status.send_replace(status.clone());
status
}
None => {
let Some(status) = agent_status_from_event(msg) else {
return;
};
status
}
};
if !is_final(&status) {
return;
+1
View File
@@ -152,6 +152,7 @@ pub(super) async fn spawn_review_thread(
extension_data,
turn_skills: TurnSkillsContext::new(parent_turn_context.turn_skills.outcome.clone()),
turn_timing_state: Arc::new(TurnTimingState::default()),
terminal_error: Arc::new(Mutex::new(None)),
server_model_warning_emitted: AtomicBool::new(false),
model_verification_emitted: AtomicBool::new(false),
};
@@ -167,6 +167,7 @@ pub struct TurnContext {
pub(crate) extension_data: Arc<codex_extension_api::ExtensionData>,
pub(crate) turn_skills: TurnSkillsContext,
pub(crate) turn_timing_state: Arc<TurnTimingState>,
pub(crate) terminal_error: Arc<Mutex<Option<String>>>,
pub(crate) server_model_warning_emitted: AtomicBool,
pub(crate) model_verification_emitted: AtomicBool,
}
@@ -336,6 +337,7 @@ impl TurnContext {
extension_data: Arc::clone(&self.extension_data),
turn_skills: self.turn_skills.clone(),
turn_timing_state: Arc::clone(&self.turn_timing_state),
terminal_error: Arc::clone(&self.terminal_error),
server_model_warning_emitted: AtomicBool::new(
self.server_model_warning_emitted.load(Ordering::Relaxed),
),
@@ -639,6 +641,7 @@ impl Session {
extension_data,
turn_skills: TurnSkillsContext::new(skills_outcome),
turn_timing_state: Arc::new(TurnTimingState::default()),
terminal_error: Arc::new(Mutex::new(None)),
server_model_warning_emitted: AtomicBool::new(false),
model_verification_emitted: AtomicBool::new(false),
}
+16 -1
View File
@@ -1,10 +1,18 @@
use codex_protocol::AgentPath;
use codex_protocol::protocol::AgentStatus;
use codex_utils_output_truncation::TruncationPolicy;
use codex_utils_output_truncation::truncate_text;
use crate::context::ContextualUserFragment;
use crate::context::InterAgentCompletionMessage;
use crate::context::SubagentNotification;
const COMPLETION_MESSAGE_MAX_TOKENS: usize = 1_000;
const COMPLETION_MESSAGE_ENVELOPE_TOKEN_RESERVE: usize = 100;
const ERROR_MAX_TOKENS: usize =
COMPLETION_MESSAGE_MAX_TOKENS - COMPLETION_MESSAGE_ENVELOPE_TOKEN_RESERVE;
const ERROR_NEXT_ACTION: &str = "This agent's turn failed. If you still need this agent, use the available collaboration tools to give it another task.";
// Helpers for model-visible session state markers that are stored in user-role
// messages but are not user intent.
@@ -24,7 +32,10 @@ pub(crate) fn format_inter_agent_completion_message(
let payload = match status {
AgentStatus::Completed(Some(message)) => message.clone(),
AgentStatus::Completed(None) => String::new(),
AgentStatus::Errored(error) => format!("Agent errored: {error}"),
AgentStatus::Errored(error) => {
let error = truncate_text(error, TruncationPolicy::Tokens(ERROR_MAX_TOKENS));
format!("Agent errored: {error}\n\n{ERROR_NEXT_ACTION}")
}
AgentStatus::Shutdown => "Agent shut down.".to_string(),
AgentStatus::NotFound => "Agent was not found.".to_string(),
AgentStatus::PendingInit | AgentStatus::Running | AgentStatus::Interrupted => return None,
@@ -32,6 +43,10 @@ pub(crate) fn format_inter_agent_completion_message(
Some(InterAgentCompletionMessage::new(task_name, sender, payload).render())
}
#[cfg(test)]
#[path = "session_prefix_tests.rs"]
mod tests;
pub(crate) fn format_subagent_context_line(
agent_reference: &str,
agent_nickname: Option<&str>,
+20
View File
@@ -0,0 +1,20 @@
use codex_protocol::AgentPath;
use codex_protocol::protocol::AgentStatus;
use codex_utils_output_truncation::approx_token_count;
use super::COMPLETION_MESSAGE_MAX_TOKENS;
use super::ERROR_NEXT_ACTION;
use super::format_inter_agent_completion_message;
#[test]
fn error_completion_message_stays_below_manual_review_threshold() {
let message = format_inter_agent_completion_message(
AgentPath::root(),
AgentPath::try_from("/root/worker").expect("valid agent path"),
&AgentStatus::Errored("stream disconnected ".repeat(1_000)),
)
.expect("error status should produce a completion message");
assert!(approx_token_count(&message) < COMPLETION_MESSAGE_MAX_TOKENS);
assert!(message.contains(ERROR_NEXT_ACTION));
}