Replace Guardian tickets with parent response IDs (#43002)

## What changed

Capture the ID from `response.created` and pass it as `parent_response_id` in Guardian review and classifier request metadata. Clear the current response ID before sampling retries so tool reviews cannot inherit a failed response's ID.

Request Guardian credits with `guardian_credits_requested` on eligible parent requests. Remove the ticket protocol, transport attachment, and ticket-specific body redaction. Keep ordinary app-server client `parent_response_id` values inside `x-codex-turn-metadata` rather than promoting them to Guardian request metadata.

## Testing

Update coverage for HTTP fallback, WebSocket review reuse, missing response IDs, classifier transport retries, and code-mode yield/resume attribution. Extend app-server tests to verify client metadata placement over HTTP and WebSocket.

GitOrigin-RevId: c1665e56382ecec8d260fc2ca0ecd5e20b0b2391
This commit is contained in:
jif
2026-09-05 13:02:49 +00:00
committed by copyberry
parent a7a4321593
commit dee21ec1bc
32 changed files with 178 additions and 362 deletions
@@ -73,6 +73,10 @@ async fn turn_start_forwards_client_metadata_to_responses_request_v2() -> Result
timeout(DEFAULT_READ_TIMEOUT, mcp.read_response(thread_req)).await??;
let client_metadata = HashMap::from([
(
"parent_response_id".to_string(),
"client-correlation".to_string(),
),
("fiber_run_id".to_string(), "fiber-start-123".to_string()),
("origin".to_string(), "gaas".to_string()),
(
@@ -109,6 +113,20 @@ async fn turn_start_forwards_client_metadata_to_responses_request_v2() -> Result
.as_deref()
.map(parse_json_header)
.expect("x-codex-turn-metadata header should be present");
let body = request.body_json();
let body_turn_metadata = parse_json_header(
body["client_metadata"]["x-codex-turn-metadata"]
.as_str()
.expect("turn metadata"),
);
assert_eq!(
(
metadata["parent_response_id"].as_str(),
body_turn_metadata["parent_response_id"].as_str(),
body["client_metadata"].get("parent_response_id"),
),
(Some("client-correlation"), Some("client-correlation"), None),
);
assert_eq!(metadata["fiber_run_id"].as_str(), Some("fiber-start-123"));
assert_eq!(metadata["origin"].as_str(), Some("gaas"));
assert_eq!(metadata["thread_source"].as_str(), Some("automation"));
@@ -603,6 +621,10 @@ async fn turn_start_forwards_client_metadata_to_responses_websocket_request_body
timeout(DEFAULT_READ_TIMEOUT, mcp.read_response(thread_req)).await??;
let client_metadata = HashMap::from([
(
"parent_response_id".to_string(),
"client-correlation".to_string(),
),
("fiber_run_id".to_string(), "fiber-start-123".to_string()),
("origin".to_string(), "gaas".to_string()),
]);
@@ -646,6 +668,13 @@ async fn turn_start_forwards_client_metadata_to_responses_websocket_request_body
.as_str()
.map(parse_json_header)
.expect("websocket x-codex-turn-metadata client metadata should be present");
assert_eq!(
(
metadata["parent_response_id"].as_str(),
request["client_metadata"].get("parent_response_id"),
),
(Some("client-correlation"), None),
);
assert_eq!(metadata["fiber_run_id"].as_str(), Some("fiber-start-123"));
assert_eq!(metadata["origin"].as_str(), Some("gaas"));
assert_eq!(metadata["thread_source"].as_str(), Some("automation"));
+6 -2
View File
@@ -94,11 +94,15 @@ pub struct MemorySummarizeOutput {
pub memory_summary: String,
}
/// The server response currently being handled, shared with tool-review extensions.
#[derive(Clone, Debug)]
pub struct ResponseId(pub String);
#[derive(Debug)]
pub enum ResponseEvent {
Created {
/// Runtime-only receipt for this exact server response.
guardian_ticket: Option<codex_protocol::guardian_ticket::GuardianTicket>,
/// Existing server response ID, when supplied by the stream.
response_id: Option<String>,
},
SafetyBuffering(SafetyBuffering),
OutputItemDone(ResponseItem),
+2 -14
View File
@@ -60,8 +60,6 @@ pub struct ResponsesOptions {
pub extra_headers: HeaderMap,
pub compression: Compression,
pub turn_state: Option<Arc<OnceLock<String>>>,
/// Runtime-only parent receipt, attached after inference tracing.
pub guardian_ticket: Option<codex_protocol::guardian_ticket::GuardianTicket>,
}
impl<T: HttpTransport> ResponsesClient<T> {
@@ -103,7 +101,7 @@ impl<T: HttpTransport> ResponsesClient<T> {
)]
pub async fn stream_request(
&self,
mut request: ResponsesApiRequest,
request: ResponsesApiRequest,
options: ResponsesOptions,
) -> Result<ResponseStream, ApiError> {
let ResponsesOptions {
@@ -113,19 +111,9 @@ impl<T: HttpTransport> ResponsesClient<T> {
extra_headers,
compression,
turn_state,
guardian_ticket,
} = options;
crate::guardian_ticket::attach(
&mut request.client_metadata,
guardian_ticket.as_ref(),
self.endpoint,
);
let mut body = EncodedJsonBody::encode(&request)
let body = EncodedJsonBody::encode(&request)
.map_err(|e| ApiError::Stream(format!("failed to encode responses request: {e}")))?;
if guardian_ticket.is_some() {
body = body.without_body_logging();
}
let mut headers = extra_headers;
if let Some(ref thread_id) = thread_id {
@@ -234,10 +234,9 @@ impl ResponsesWebsocketConnection {
)]
pub async fn stream_request(
&self,
mut request: ResponsesWsRequest<'_>,
request: ResponsesWsRequest<'_>,
connection_reused: bool,
turn_state: Option<Arc<OnceLock<String>>>,
guardian_ticket: Option<&codex_protocol::guardian_ticket::GuardianTicket>,
) -> Result<ResponseStream, ApiError> {
let (tx_event, rx_event) =
mpsc::channel::<std::result::Result<ResponseEvent, ApiError>>(1600);
@@ -271,12 +270,6 @@ impl ResponsesWebsocketConnection {
warmup: ws_request.generate == Some(false),
connection_reused,
};
let ResponsesWsRequest::ResponseCreate(ws_request) = &mut request;
crate::guardian_ticket::attach(
&mut ws_request.client_metadata,
guardian_ticket,
self.endpoint,
);
let request_text = serialize_websocket_request(&request)?;
let current_span = Span::current();
-33
View File
@@ -1,33 +0,0 @@
//! Adds opaque receipts only to Guardian requests at the transport boundary.
use std::collections::HashMap;
use crate::endpoint::responses::ResponsesEndpoint;
use codex_protocol::guardian_ticket::GUARDIAN_TICKET_METADATA_KEY;
use codex_protocol::guardian_ticket::GuardianTicket;
pub(crate) fn attach(
metadata: &mut Option<HashMap<String, String>>,
ticket: Option<&GuardianTicket>,
endpoint: ResponsesEndpoint,
) {
if let Some(metadata) = metadata.as_mut() {
metadata.remove(GUARDIAN_TICKET_METADATA_KEY);
if endpoint != ResponsesEndpoint::Responses {
metadata.remove("guardian_ticket_requested");
}
}
if endpoint != ResponsesEndpoint::Responses
&& let Some(ticket) = ticket
{
metadata.get_or_insert_with(HashMap::new).insert(
GUARDIAN_TICKET_METADATA_KEY.to_owned(),
ticket.as_str().to_owned(),
);
}
}
#[cfg(test)]
#[path = "guardian_ticket_tests.rs"]
mod tests;
@@ -1,60 +0,0 @@
use super::*;
use crate::common::ResponseEvent;
use crate::sse::responses::ResponsesStreamEvent;
use crate::sse::responses::process_responses_event;
use pretty_assertions::assert_eq;
use serde_json::json;
#[test]
fn response_created_receipt_is_runtime_only_and_redacted() {
let raw = "a".repeat(43);
let event: ResponsesStreamEvent = serde_json::from_value(json!({
"type": "response.created",
"response": {"id": "resp-1", "headers": {"x-codex-guardian-ticket": raw}}
}))
.unwrap();
let ResponseEvent::Created { guardian_ticket } =
process_responses_event(event).unwrap().unwrap()
else {
panic!("expected response.created");
};
let ticket = guardian_ticket.unwrap();
assert_eq!(ticket.as_str(), raw);
assert!(!format!("{ticket:?}").contains(&raw));
let mut metadata = Some(HashMap::from([("turn_id".into(), "turn-1".into())]));
attach(&mut metadata, Some(&ticket), ResponsesEndpoint::Guardian);
assert_eq!(
metadata,
Some(HashMap::from([
("turn_id".into(), "turn-1".into()),
(GUARDIAN_TICKET_METADATA_KEY.into(), raw),
]))
);
attach(
&mut metadata,
/*ticket*/ None,
ResponsesEndpoint::Guardian,
);
assert_eq!(
metadata,
Some(HashMap::from([("turn_id".into(), "turn-1".into())]))
);
}
#[test]
fn missing_or_malformed_response_receipts_do_not_become_tickets() {
for response in [
json!({}),
json!({"headers": {"x-codex-guardian-ticket": "invalid"}}),
] {
let event =
serde_json::from_value(json!({"type": "response.created", "response": response}))
.unwrap();
assert!(matches!(
process_responses_event(event).unwrap(),
Some(ResponseEvent::Created {
guardian_ticket: None
})
));
}
}
+1 -1
View File
@@ -4,7 +4,6 @@ pub(crate) mod common;
pub(crate) mod endpoint;
pub(crate) mod error;
pub(crate) mod files;
mod guardian_ticket;
pub(crate) mod images;
pub(crate) mod provider;
pub(crate) mod rate_limits;
@@ -40,6 +39,7 @@ pub use crate::common::ReasoningContext;
pub use crate::common::ReasoningSummaryDelivery;
pub use crate::common::ResponseCreateWsRequest;
pub use crate::common::ResponseEvent;
pub use crate::common::ResponseId;
pub use crate::common::ResponseStream;
pub use crate::common::ResponsesApiRequest;
pub use crate::common::ResponsesApiTools;
+6 -16
View File
@@ -9,8 +9,6 @@ use crate::telemetry::SseTelemetry;
use codex_client::ByteStream;
use codex_client::StreamResponse;
use codex_protocol::ResponseUsageMetadata;
use codex_protocol::guardian_ticket::GUARDIAN_TICKET_HEADER;
use codex_protocol::guardian_ticket::GuardianTicket;
use codex_protocol::models::ResponseItem;
use codex_protocol::protocol::MisalignmentErrorDetails;
use codex_protocol::protocol::ModelVerification;
@@ -409,17 +407,11 @@ pub fn process_responses_event(
}
"response.created" => {
if let Some(response) = event.response {
let guardian_ticket = response
.get("headers")
.and_then(Value::as_object)
.and_then(|headers| {
headers
.iter()
.find(|(name, _)| name.eq_ignore_ascii_case(GUARDIAN_TICKET_HEADER))
})
.and_then(|(_, value)| value.as_str())
.and_then(GuardianTicket::from_server);
return Ok(Some(ResponseEvent::Created { guardian_ticket }));
let response_id = response
.get("id")
.and_then(Value::as_str)
.map(str::to_owned);
return Ok(Some(ResponseEvent::Created { response_id }));
}
}
"response.failed" => {
@@ -1630,9 +1622,7 @@ mod tests {
);
assert_matches!(
&events[1],
ResponseEvent::Created {
guardian_ticket: None
}
ResponseEvent::Created { response_id: Some(id) } if id == "resp-1"
);
assert_matches!(
&events[2],
+3 -15
View File
@@ -331,8 +331,7 @@ async fn responses_client_uses_guardian_path() -> Result<()> {
async fn responses_client_stream_request_preserves_item_ids() -> Result<()> {
let state = RecordingState::default();
let transport = RecordingTransport::new(state.clone());
let client = ResponsesClient::new(transport, provider("openai"), Arc::new(NoAuth))
.with_endpoint(ResponsesEndpoint::Guardian);
let client = ResponsesClient::new(transport, provider("openai"), Arc::new(NoAuth));
let request = ResponsesApiRequest {
model: "gpt-test".into(),
instructions: "Say hi".into(),
@@ -357,24 +356,14 @@ async fn responses_client_stream_request_preserves_item_ids() -> Result<()> {
client_metadata: None,
access_programs: None,
};
let raw_ticket = "t".repeat(43);
let ticket = codex_protocol::guardian_ticket::GuardianTicket::from_server(&raw_ticket).unwrap();
let mut expected = serde_json::to_value(&request)?;
expected["client_metadata"] = serde_json::json!({"guardian_ticket": raw_ticket});
let expected = serde_json::to_value(&request)?;
let _stream = client
.stream_request(
request,
ResponsesOptions {
guardian_ticket: Some(ticket),
..Default::default()
},
)
.stream_request(request, ResponsesOptions::default())
.await?;
let requests = state.take_stream_requests();
assert_eq!(requests.len(), 1);
assert!(!format!("{:?}", requests[0]).contains(&raw_ticket));
let prepared = requests[0]
.prepare_body_for_send()
.expect("body should prepare");
@@ -589,7 +578,6 @@ async fn azure_store_sends_ids_and_headers() -> Result<()> {
extra_headers,
compression: Compression::None,
turn_state: None,
guardian_ticket: None,
},
)
.await?;
+17 -7
View File
@@ -1104,14 +1104,24 @@ impl ModelClient {
&& provider.aws.is_none()
}
fn set_guardian_ticket_request(
fn set_guardian_metadata(
&self,
metadata: &mut Option<HashMap<String, String>>,
parent_response_id: Option<&str>,
auth: Option<&CodexAuth>,
endpoint: ResponsesEndpoint,
) {
if let Some(metadata) = metadata.as_mut() {
metadata.remove("guardian_ticket_requested");
metadata.remove("guardian_credits_requested");
metadata.remove("parent_response_id");
}
if endpoint == ResponsesEndpoint::Guardian
&& let Some(parent_response_id) = parent_response_id
{
metadata.get_or_insert_with(HashMap::new).insert(
"parent_response_id".to_owned(),
parent_response_id.to_owned(),
);
}
if self.free_guardian_enabled
&& endpoint == ResponsesEndpoint::Responses
@@ -1125,7 +1135,7 @@ impl ModelClient {
{
metadata
.get_or_insert_with(HashMap::new)
.insert("guardian_ticket_requested".to_owned(), "true".to_owned());
.insert("guardian_credits_requested".to_owned(), "true".to_owned());
}
}
@@ -1351,7 +1361,6 @@ impl ModelClientSession {
},
compression,
turn_state: Some(Arc::clone(&self.turn_state)),
guardian_ticket: responses_metadata.guardian_ticket.clone(),
}
}
@@ -1641,8 +1650,9 @@ impl ModelClientSession {
service_tier.clone(),
responses_metadata,
)?;
self.client.set_guardian_ticket_request(
self.client.set_guardian_metadata(
&mut request.client_metadata,
responses_metadata.parent_response_id.as_deref(),
client_setup.auth.as_ref(),
endpoint,
);
@@ -1900,8 +1910,9 @@ impl ModelClientSession {
),
..ResponseCreateWsRequest::from(&request)
};
self.client.set_guardian_ticket_request(
self.client.set_guardian_metadata(
&mut ws_payload.client_metadata,
responses_metadata.parent_response_id.as_deref(),
client_setup.auth.as_ref(),
endpoint,
);
@@ -1922,7 +1933,6 @@ impl ModelClientSession {
ws_request,
self.websocket_session.connection_reused(),
Some(Arc::clone(&self.turn_state)),
responses_metadata.guardian_ticket.as_ref(),
)
.await;
if let Some(original_item_ids) = original_item_ids {
+2 -6
View File
@@ -844,9 +844,7 @@ async fn response_stream_records_last_model_feedback_ids() {
.set_default();
let api_stream = futures::stream::iter([
Ok(ResponseEvent::Created {
guardian_ticket: None,
}),
Ok(ResponseEvent::Created { response_id: None }),
Ok(ResponseEvent::Completed {
response_id: "resp-123".to_string(),
token_usage: None,
@@ -1062,9 +1060,7 @@ async fn dropped_backpressured_response_stream_traces_cancelled_partial_output()
let backpressured_item_yielded = Arc::new(Notify::new());
let mut events = VecDeque::new();
for _ in 0..super::RESPONSE_STREAM_CHANNEL_CAPACITY {
events.push_back(ResponseEvent::Created {
guardian_ticket: None,
});
events.push_back(ResponseEvent::Created { response_id: None });
}
events.push_back(ResponseEvent::OutputItemDone(output_message(
"1",
+11 -14
View File
@@ -88,8 +88,8 @@ const GUARDIAN_RECENT_ENTRY_LIMIT: usize = 40;
/// step-scoped things past their lifetime (like MCP bindings)
#[derive(Clone)]
pub(crate) struct GuardianReviewContext {
/// Ticket from the response currently handled in this execution context.
pub(crate) guardian_ticket: Option<codex_protocol::guardian_ticket::GuardianTicket>,
/// The response currently handled in this execution context.
pub(crate) parent_response_id: Option<String>,
turn: Arc<TurnContext>,
environments: TurnEnvironmentSnapshot,
// Model and reasoning inputs are carried for the follow-up Guardian and V2 migrations.
@@ -109,11 +109,10 @@ impl GuardianReviewContext {
settings: &ResolvedStepSettings,
) -> Self {
Self {
guardian_ticket: turn
parent_response_id: turn
.extension_data
.get::<codex_protocol::guardian_ticket::GuardianTicket>()
.as_deref()
.cloned(),
.get::<codex_api::ResponseId>()
.map(|id| id.0.clone()),
environments: turn.environments.clone(),
model_info: Arc::clone(&settings.model_info),
reasoning_effort: settings.reasoning_effort().cloned(),
@@ -136,12 +135,11 @@ impl GuardianReviewContext {
impl From<&Arc<StepContext>> for GuardianReviewContext {
fn from(step: &Arc<StepContext>) -> Self {
Self {
guardian_ticket: step
parent_response_id: step
.turn
.extension_data
.get::<codex_protocol::guardian_ticket::GuardianTicket>()
.as_deref()
.cloned(),
.get::<codex_api::ResponseId>()
.map(|id| id.0.clone()),
turn: Arc::clone(&step.turn),
environments: step.environments.clone(),
model_info: Arc::clone(&step.settings.model_info),
@@ -156,11 +154,10 @@ impl From<&Arc<StepContext>> for GuardianReviewContext {
impl From<Arc<TurnContext>> for GuardianReviewContext {
fn from(turn: Arc<TurnContext>) -> Self {
Self {
guardian_ticket: turn
parent_response_id: turn
.extension_data
.get::<codex_protocol::guardian_ticket::GuardianTicket>()
.as_deref()
.cloned(),
.get::<codex_api::ResponseId>()
.map(|id| id.0.clone()),
environments: turn.environments.clone(),
model_info: Arc::clone(turn.model_info()),
reasoning_effort: turn.reasoning_effort().cloned(),
+7 -1
View File
@@ -1300,8 +1300,14 @@ async fn run_review_on_session(
}),
..Default::default()
})
.with_responses_metadata(
params
.parent_context
.parent_response_id
.as_ref()
.map(|id| HashMap::from([("parent_response_id".to_owned(), id.clone())])),
)
.on_start(TurnStartOptions {
guardian_ticket: params.parent_context.guardian_ticket.clone(),
final_output_json_schema: Some(params.schema.clone()),
service_tier: None,
parent_turn_id: Some(parent_turn.sub_id.clone()),
+4 -5
View File
@@ -58,8 +58,7 @@ pub(crate) const WORKSPACES_KEY: &str = "workspaces";
// App-server clients can specify additional metadata in the `responsesapi_client_metadata` param
// when submitting a turn, but they must not override fields owned by core.
const RESERVED_METADATA_KEYS: &[&str] = &[
codex_protocol::guardian_ticket::GUARDIAN_TICKET_METADATA_KEY,
"guardian_ticket_requested",
"guardian_credits_requested",
INSTALLATION_ID_KEY,
X_CODEX_INSTALLATION_ID_HEADER,
SESSION_ID_KEY,
@@ -219,8 +218,8 @@ pub(crate) enum TurnToolSource {
/// truth.
#[derive(Clone, Debug)]
pub struct CodexResponsesMetadata {
/// Opaque runtime receipt; deliberately omitted from metadata projections.
pub(crate) guardian_ticket: Option<codex_protocol::guardian_ticket::GuardianTicket>,
/// Guardian parent reference; projected only onto a Guardian request.
pub(crate) parent_response_id: Option<String>,
pub(crate) installation_id: String,
pub(crate) session_id: String,
pub(crate) thread_id: String,
@@ -260,7 +259,7 @@ impl CodexResponsesMetadata {
window_id: String,
) -> Self {
Self {
guardian_ticket: None,
parent_response_id: None,
installation_id,
session_id,
thread_id,
-1
View File
@@ -137,7 +137,6 @@ pub(super) async fn spawn_review_thread(
extension_data.insert(parent_turn_context.skills_snapshot().as_ref().clone());
let review_turn_context = TurnContext {
guardian_ticket: None,
sub_id: review_turn_id.clone(),
trace_id: current_span_trace_id(),
realtime_active: parent_turn_context.realtime_active,
-1
View File
@@ -619,7 +619,6 @@ impl Session {
request_kind,
);
CodexResponsesMetadata {
guardian_ticket: turn_context.guardian_ticket.clone(),
window_number: Some(window_number),
context_window_id: Some(context_window_id),
history_ingest_requested: turn_context
+7 -5
View File
@@ -1437,10 +1437,10 @@ async fn run_sampling_request(
let mut original_input = None;
let mut executed_tool_calls_by_output = HashMap::new();
loop {
// A retry must not lend a previous response's ticket to the next tool call.
// A retry must not attribute the next tool call to the previous response.
turn_context
.extension_data
.remove::<codex_protocol::guardian_ticket::GuardianTicket>();
.remove::<codex_api::ResponseId>();
let prompt_input = if let Some(input) = initial_input.take() {
input
} else {
@@ -2376,9 +2376,11 @@ async fn try_run_sampling_request(
record_turn_ttft_metric(&turn_context, &event).await;
match event {
ResponseEvent::Created { guardian_ticket } => {
if let Some(ticket) = guardian_ticket {
turn_context.extension_data.insert(ticket);
ResponseEvent::Created { response_id } => {
if let Some(response_id) = response_id {
turn_context
.extension_data
.insert(codex_api::ResponseId(response_id));
}
}
ResponseEvent::OutputItemDone(mut item) => {
@@ -185,7 +185,6 @@ impl std::fmt::Debug for TurnEnvironment {
/// Request options for one turn, separate from persistent thread settings.
#[derive(Default)]
pub(crate) struct NewTurnContextOptions {
pub(crate) guardian_ticket: Option<codex_protocol::guardian_ticket::GuardianTicket>,
pub(crate) final_output_json_schema: Option<Value>,
pub(crate) cyber_access_program: Option<CyberAccessProgram>,
}
@@ -193,8 +192,6 @@ pub(crate) struct NewTurnContextOptions {
/// The context needed for a single turn of the thread.
#[derive(Debug)]
pub struct TurnContext {
/// Receipt selected for this Guardian turn, separate from its own responses.
pub(crate) guardian_ticket: Option<codex_protocol::guardian_ticket::GuardianTicket>,
pub(crate) sub_id: String,
pub(crate) trace_id: Option<String>,
pub(crate) realtime_active: bool,
@@ -540,7 +537,6 @@ impl TurnContext {
windows_sandbox_level: self.windows_sandbox_level,
available_models,
unified_exec_shell_mode: self.unified_exec_shell_mode.clone(),
guardian_ticket: self.guardian_ticket.clone(),
final_output_json_schema: self.final_output_json_schema.clone(),
dynamic_tools: self.dynamic_tools.clone(),
turn_metadata_state: self.turn_metadata_state.clone(),
@@ -785,7 +781,6 @@ impl Session {
let extension_data = Arc::new(codex_extension_api::ExtensionData::new(sub_id.clone()));
extension_data.insert(skills_snapshot);
TurnContext {
guardian_ticket: None,
sub_id,
trace_id: current_span_trace_id(),
realtime_active: false,
@@ -1040,7 +1035,6 @@ impl Session {
turn_context.extension_data.insert(trusted_plugin_roots);
turn_context.realtime_active = self.conversation.running_state().await.is_some();
turn_context.guardian_ticket = options.guardian_ticket;
turn_context.final_output_json_schema = options.final_output_json_schema;
if turn_context.config.model_provider_id == codex_model_provider_info::OPENAI_PROVIDER_ID {
turn_context.cyber_access_program = options.cyber_access_program;
-2
View File
@@ -121,7 +121,6 @@ impl PreparedTurnInputSettings {
kind: TurnStartKind,
) -> CodexResult<Option<Arc<TurnContext>>> {
let TurnStartOptions {
guardian_ticket,
turn_trigger,
final_output_json_schema,
service_tier,
@@ -139,7 +138,6 @@ impl PreparedTurnInputSettings {
updates.service_tier_for_turn = service_tier;
let options = NewTurnContextOptions {
guardian_ticket,
final_output_json_schema,
cyber_access_program,
};
-1
View File
@@ -473,7 +473,6 @@ impl Session {
.new_turn_with_default_settings(
sub_id,
NewTurnContextOptions {
guardian_ticket: start_options.guardian_ticket,
final_output_json_schema: start_options.final_output_json_schema,
cyber_access_program: start_options.cyber_access_program,
},
+9
View File
@@ -393,6 +393,14 @@ impl TurnMetadataState {
.read()
.unwrap_or_else(std::sync::PoisonError::into_inner)
.clone();
// Extract Guardian's internal parent before filtering configured metadata keys.
// Ordinary app-server client metadata stays in `extra`.
let parent_response_id =
if self.subagent_header.as_deref() == Some(crate::guardian::GUARDIAN_REVIEWER_NAME) {
extra.remove("parent_response_id")
} else {
None
};
for key in self
.responses_api_metadata
.read()
@@ -402,6 +410,7 @@ impl TurnMetadataState {
extra.remove(key);
}
let mut metadata = CodexResponsesMetadata {
parent_response_id,
turn_id: Some(self.turn_id.clone()),
agent_name: Some(self.agent_name.clone()),
forked_from_thread_id: self.forked_from_thread_id,
+1 -3
View File
@@ -29,9 +29,7 @@ async fn turn_timing_state_records_ttft_only_once_per_turn() {
state.mark_turn_started(Instant::now()).await;
assert_eq!(
state
.record_ttft_for_response_event(&ResponseEvent::Created {
guardian_ticket: None
})
.record_ttft_for_response_event(&ResponseEvent::Created { response_id: None })
.await,
None
);
+21 -18
View File
@@ -2373,21 +2373,21 @@ async fn code_mode_wait_timeout_reconnects_on_next_exec() -> Result<()> {
}
#[derive(Default)]
struct GuardianTicketObserver {
tickets: Mutex<Vec<(String, Option<String>)>>,
struct ResponseIdObserver {
response_ids: Mutex<Vec<(String, Option<String>)>>,
wait_started: tokio::sync::Notify,
}
impl ToolLifecycleContributor for GuardianTicketObserver {
impl ToolLifecycleContributor for ResponseIdObserver {
fn on_tool_start<'a>(&'a self, input: ToolStartInput<'a>) -> ToolLifecycleFuture<'a> {
Box::pin(async move {
if matches!(input.tool_name.name.as_str(), "exec_command" | "wait") {
self.tickets.lock().unwrap().push((
self.response_ids.lock().unwrap().push((
input.tool_name.name.clone(),
input
.turn_store
.get::<codex_protocol::guardian_ticket::GuardianTicket>()
.map(|ticket| ticket.as_str().to_owned()),
.get::<codex_api::ResponseId>()
.map(|id| id.0.clone()),
));
if input.tool_name.name == "wait" {
self.wait_started.notify_one();
@@ -2403,7 +2403,7 @@ async fn code_mode_can_yield_and_resume_with_wait() -> Result<()> {
skip_if_no_network!(Ok(()));
let server = responses::start_mock_server().await;
let observer = Arc::new(GuardianTicketObserver::default());
let observer = Arc::new(ResponseIdObserver::default());
let mut extensions = ExtensionRegistryBuilder::<Config>::new();
extensions.tool_lifecycle_contributor(observer.clone());
let mut builder = test_codex()
@@ -2431,7 +2431,7 @@ text((await tools.exec_command({{cmd: "printf 'phase 3'"}})).output);
responses::mount_sse_once(
&server,
sse(vec![
serde_json::json!({"type": "response.created", "response": {"id": "resp-1", "headers": {"x-codex-guardian-ticket": "p".repeat(43)}}}),
responses::ev_response_created("resp-1"),
ev_custom_tool_call("call-1", "exec", &code),
ev_completed("resp-1"),
]),
@@ -2464,7 +2464,7 @@ text((await tools.exec_command({{cmd: "printf 'phase 3'"}})).output);
responses::mount_sse_once(
&server,
sse(vec![
serde_json::json!({"type": "response.created", "response": {"id": "resp-3", "headers": {"x-codex-guardian-ticket": "q".repeat(43)}}}),
responses::ev_response_created("resp-3"),
responses::ev_function_call(
"call-2",
"wait",
@@ -2511,7 +2511,7 @@ text((await tools.exec_command({{cmd: "printf 'phase 3'"}})).output);
responses::mount_sse_once(
&server,
sse(vec![
serde_json::json!({"type": "response.created", "response": {"id": "resp-5", "headers": {"x-codex-guardian-ticket": "r".repeat(43)}}}),
responses::ev_response_created("resp-5"),
responses::ev_function_call(
"call-3",
"wait",
@@ -2551,15 +2551,18 @@ text((await tools.exec_command({{cmd: "printf 'phase 3'"}})).output);
);
assert_eq!(text_item(&third_items, /*index*/ 1), "phase 3");
let observed = observer.tickets.lock().unwrap();
let mut nested_tickets = observed
let observed = observer.response_ids.lock().unwrap();
let mut nested_response_ids = observed
.iter()
.filter(|(tool, _)| tool == "exec_command")
.filter_map(|(_, ticket)| ticket.clone())
.skip_while(|ticket| ticket != &"q".repeat(43))
.filter_map(|(_, response_id)| response_id.clone())
.skip_while(|response_id| response_id != &"resp-3".to_owned())
.collect::<Vec<_>>();
nested_tickets.dedup();
assert_eq!(nested_tickets, vec!["q".repeat(43), "r".repeat(43)]);
nested_response_ids.dedup();
assert_eq!(
nested_response_ids,
vec!["resp-3".to_owned(), "resp-5".to_owned()]
);
assert_eq!(
observed
.iter()
@@ -2567,8 +2570,8 @@ text((await tools.exec_command({{cmd: "printf 'phase 3'"}})).output);
.cloned()
.collect::<Vec<_>>(),
vec![
("wait".to_owned(), Some("q".repeat(43))),
("wait".to_owned(), Some("r".repeat(43))),
("wait".to_owned(), Some("resp-3".to_owned())),
("wait".to_owned(), Some("resp-5".to_owned())),
]
);
Ok(())
+26 -26
View File
@@ -135,14 +135,14 @@ impl TimeProvider for RecordingTimeProvider {
#[test_case(CodexAuth::create_dummy_chatgpt_auth_for_testing(), "OpenAI", "/backend-api/codex", true, "/backend-api/codex/guardian", true; "chatgpt_uses_guardian_when_enabled")]
#[test_case(CodexAuth::create_dummy_chatgpt_auth_for_testing(), "OpenAI", "/v1", true, "/v1/responses", true; "custom_openai_url_uses_responses")]
#[test_case(CodexAuth::create_dummy_chatgpt_auth_for_testing(), "Custom", "/backend-api/codex", true, "/backend-api/codex/responses", true; "custom_provider_uses_responses")]
#[test_case(CodexAuth::create_dummy_chatgpt_auth_for_testing(), "OpenAI", "/backend-api/codex", true, "/backend-api/codex/guardian", false; "server_without_tickets")]
#[test_case(CodexAuth::create_dummy_chatgpt_auth_for_testing(), "OpenAI", "/backend-api/codex", true, "/backend-api/codex/guardian", false; "server_without_response_id")]
async fn guardian_session_inherits_parent_http_fallback(
auth: CodexAuth,
provider_name: &str,
base_path: &str,
free_guardian: bool,
expected_guardian_path: &str,
ticket_issued: bool,
response_id_present: bool,
) -> Result<()> {
skip_if_no_network!(Ok(()));
@@ -154,16 +154,16 @@ async fn guardian_session_inherits_parent_http_fallback(
.mount_as_scoped(&server)
.await;
let parent_ticket = "p".repeat(43);
let parent_response_id = "parent-tool";
let responses = mount_sse_sequence(
&server,
vec![
// A failed sampling attempt must not lend its receipt to the retried action.
// A failed sampling attempt must not supply the retried action's parent ID.
sse(vec![json!({"type": "response.created", "response": {
"id": "failed-parent", "headers": {"x-codex-guardian-ticket": "f".repeat(43)}
"id": "failed-parent"
}})]),
sse(vec![
json!({"type": "response.created", "response": {"id": "parent-tool", "headers": ticket_issued.then(|| json!({"x-codex-guardian-ticket": parent_ticket}))}}),
json!({"type": "response.created", "response": {"id": response_id_present.then_some(parent_response_id)}}),
ev_function_call(
"call",
"exec_command",
@@ -241,30 +241,30 @@ async fn guardian_session_inherits_parent_http_fallback(
})
.expect("Guardian reviewer inference request");
assert_eq!(guardian_request.path(), expected_guardian_path);
let tickets_enabled = expected_guardian_path.ends_with("/guardian");
let credits_enabled = expected_guardian_path.ends_with("/guardian");
let body = guardian_request.body_json();
assert_eq!(
(
body["client_metadata"].get("guardian_ticket").cloned(),
body["client_metadata"].get("guardian_ticket_requested"),
body["client_metadata"].get("parent_response_id").cloned(),
body["client_metadata"].get("guardian_credits_requested"),
),
(
(tickets_enabled && ticket_issued).then(|| json!(parent_ticket)),
(credits_enabled && response_id_present).then(|| json!(parent_response_id)),
None
)
);
assert!(!body["input"].to_string().contains(&parent_ticket));
assert!(!body["input"].to_string().contains(parent_response_id));
for request in responses.requests() {
let body = request.body_json();
if body["client_metadata"]["x-openai-subagent"] != "guardian" {
assert_eq!(
(
body["client_metadata"]
.get("guardian_ticket_requested")
.get("guardian_credits_requested")
.cloned(),
body["client_metadata"].get("guardian_ticket"),
body["client_metadata"].get("parent_response_id"),
),
(tickets_enabled.then(|| json!("true")), None)
(credits_enabled.then(|| json!("true")), None)
);
}
}
@@ -409,7 +409,7 @@ async fn guardian_session_prewarms_and_is_reused_for_first_review(
skip_if_no_network!(Ok(()));
let uses_codex_backend = auth.uses_codex_backend();
let tickets_enabled = free_guardian && uses_codex_backend;
let credits_enabled = free_guardian && uses_codex_backend;
let bundled_models = codex_models_manager::bundled_models_response()?.models;
let catalog_auto_review = bundled_models
.iter()
@@ -442,12 +442,12 @@ async fn guardian_session_prewarms_and_is_reused_for_first_review(
"justification": "Exercise Guardian approval routing.",
})
.to_string();
let parent_ticket = "w".repeat(43);
let parent_response_id = "approval-request";
let server = start_websocket_server(vec![
vec![vec![ev_response_created("warm-1"), ev_completed("warm-1")]],
vec![vec![ev_response_created("warm-2"), ev_completed("warm-2")]],
vec![vec![
json!({"type": "response.created", "response": {"id": "approval-request", "headers": {"x-codex-guardian-ticket": parent_ticket}}}),
ev_response_created(parent_response_id),
ev_function_call("approval-call", "exec_command", &tool_args),
ev_completed("approval-request"),
]],
@@ -603,30 +603,30 @@ async fn guardian_session_prewarms_and_is_reused_for_first_review(
assert_eq!(
(
guardian_review["client_metadata"]
.get("guardian_ticket")
.get("parent_response_id")
.cloned(),
guardian_review["client_metadata"].get("guardian_ticket_requested"),
guardian_review["client_metadata"].get("guardian_credits_requested"),
parent_request["client_metadata"]
.get("guardian_ticket_requested")
.get("guardian_credits_requested")
.cloned(),
parent_request["client_metadata"].get("guardian_ticket"),
parent_request["client_metadata"].get("parent_response_id"),
),
(
tickets_enabled.then(|| json!(parent_ticket)),
credits_enabled.then(|| json!(parent_response_id)),
None,
tickets_enabled.then(|| json!("true")),
credits_enabled.then(|| json!("true")),
None,
)
);
assert!(
guardian_prewarm["client_metadata"]
.get("guardian_ticket")
.get("parent_response_id")
.is_none()
);
assert!(
!guardian_review["input"]
.to_string()
.contains(&parent_ticket)
.contains(parent_response_id)
);
for request in [&parent_request, &guardian_review] {
assert_root_turn(request, Some(parent_turn_id))?;
@@ -710,7 +710,7 @@ async fn guardian_session_prewarms_and_is_reused_for_first_review(
assert_eq!(guardian_context_windows, vec![Some(258_400)]);
for handshake in server.handshakes() {
let is_guardian = handshake.header("x-openai-subagent").as_deref() == Some("guardian");
let uses_guardian_endpoint = tickets_enabled && is_guardian;
let uses_guardian_endpoint = credits_enabled && is_guardian;
assert_eq!(
handshake.uri(),
if uses_guardian_endpoint {
@@ -512,11 +512,10 @@ impl GuardianV2Extension {
let thread_id = input.thread_store.level_id().to_owned();
let turn_id = input.turn_id.to_owned();
let root_turn_id = input.root_turn_id.map(str::to_owned);
let guardian_ticket = input
let parent_response_id = input
.turn_store
.get::<codex_protocol::guardian_ticket::GuardianTicket>()
.as_deref()
.cloned();
.get::<codex_api::ResponseId>()
.map(|id| id.0.clone());
let thread_context: Result<_, String> = async {
let parsed_thread_id =
ThreadId::from_string(&thread_id).map_err(|error| error.to_string())?;
@@ -839,7 +838,7 @@ impl GuardianV2Extension {
let instructions = guardian_config.render_classifier_instructions(policy);
let output = match sampler
.sample(LunaSamplingRequest {
guardian_ticket,
parent_response_id,
instructions,
trusted_review_evidence,
trusted_tool_context,
@@ -90,8 +90,8 @@ pub struct LunaSamplerConfig {
/// One tool-less Luna classification request over an already-open connection.
pub struct LunaSamplingRequest {
/// Runtime receipt of the response handling the classified tool.
pub guardian_ticket: Option<codex_protocol::guardian_ticket::GuardianTicket>,
/// ID of the response handling the classified tool.
pub parent_response_id: Option<String>,
/// Trusted instructions describing the requested classification.
pub instructions: String,
/// Host-supplied Guardian reviews isolated from untrusted transcript entries.
@@ -454,7 +454,7 @@ impl LunaSampler {
}
// A classification is its own inference turn; retries keep that identity.
let turn_id = Uuid::now_v7().to_string();
let guardian_ticket = request.guardian_ticket;
let parent_response_id = request.parent_response_id;
let parent_turn_id = request.parent_turn_id;
let root_turn_id = request.root_turn_id;
let mut input = vec![
@@ -629,6 +629,11 @@ impl LunaSampler {
turn_metadata["root_turn_id"] = json!(root_turn_id);
}
client_metadata.insert(TURN_METADATA_KEY.to_owned(), turn_metadata.to_string());
if lease.connection.endpoint == ResponsesEndpoint::GuardianClassifier
&& let Some(parent_response_id) = &parent_response_id
{
client_metadata.insert("parent_response_id".to_owned(), parent_response_id.clone());
}
request.client_metadata = Some(client_metadata);
let mut stream = match lease
.connection
@@ -637,7 +642,6 @@ impl LunaSampler {
ResponsesWsRequest::ResponseCreate((&request).into()),
/*connection_reused*/ true,
/*turn_state*/ None,
guardian_ticket.as_ref(),
)
.await
{
@@ -192,7 +192,7 @@ async fn connect_sampler(config: LunaSamplerConfig) -> Result<LunaSampler> {
fn sample_request(parent_turn_id: &str) -> LunaSamplingRequest {
LunaSamplingRequest {
guardian_ticket: None,
parent_response_id: None,
instructions: "Return high for high risk or low for low risk.".to_owned(),
trusted_review_evidence: Vec::new(),
trusted_tool_context: None,
@@ -447,7 +447,7 @@ async fn preconnected_sampler_reuses_authenticated_websocket_for_classifications
let first = sampler
.sample(LunaSamplingRequest {
guardian_ticket: None,
parent_response_id: None,
instructions: "Return high for high risk or low for low risk.".to_owned(),
trusted_review_evidence: Vec::new(),
trusted_tool_context: None,
@@ -479,7 +479,7 @@ async fn preconnected_sampler_reuses_authenticated_websocket_for_classifications
manager.refresh_token_from_authority().await?;
let second = sampler
.sample(LunaSamplingRequest {
guardian_ticket: None,
parent_response_id: None,
instructions: "Return high for high risk or low for low risk.".to_owned(),
trusted_review_evidence: Vec::new(),
trusted_tool_context: None,
@@ -655,7 +655,7 @@ async fn sampler_returns_classification_token_before_terminal_response_events()
let output = tokio::time::timeout(
Duration::from_secs(2),
sampler.sample(LunaSamplingRequest {
guardian_ticket: None,
parent_response_id: None,
instructions: "Return high for high risk or low for low risk.".to_owned(),
trusted_review_evidence: Vec::new(),
trusted_tool_context: None,
@@ -1119,7 +1119,7 @@ async fn sampler_limits_transient_recovery_attempts() -> Result<()> {
}
#[tokio::test(flavor = "multi_thread", worker_threads = 2)]
async fn guardian_ticket_survives_classifier_transport_retry() -> Result<()> {
async fn parent_response_id_survives_classifier_transport_retry() -> Result<()> {
skip_if_no_network!(Ok(()));
for free_guardian in [false, true] {
let healthy = responses::start_websocket_server(vec![vec![vec![
@@ -1146,10 +1146,9 @@ async fn guardian_ticket_survives_classifier_transport_retry() -> Result<()> {
);
}
let sampler = connect_sampler(config).await?;
let raw_ticket = "t".repeat(43);
let parent_response_id = "resp-parent";
let mut request = sample_request("turn-1");
request.guardian_ticket =
codex_protocol::guardian_ticket::GuardianTicket::from_server(&raw_ticket);
request.parent_response_id = Some(parent_response_id.to_owned());
assert_eq!(sampler.sample(request).await?, "low");
for server in [&expired, &healthy] {
let requests = server.single_connection();
@@ -1157,12 +1156,12 @@ async fn guardian_ticket_survives_classifier_transport_retry() -> Result<()> {
let body = requests[0].body_json();
assert_eq!(
(
body["client_metadata"].get("guardian_ticket").cloned(),
body["client_metadata"].get("guardian_ticket_requested"),
body["client_metadata"].get("parent_response_id").cloned(),
body["client_metadata"].get("guardian_credits_requested"),
),
(free_guardian.then(|| json!(raw_ticket)), None)
(free_guardian.then(|| json!(parent_response_id)), None)
);
assert!(!body["input"].to_string().contains(&raw_ticket));
assert!(!body["input"].to_string().contains(parent_response_id));
}
}
+2 -24
View File
@@ -11,7 +11,7 @@ use std::time::Duration;
/// Clones share the encoded allocation. Internally, the body can also hold the
/// final compressed wire bytes while retaining the original JSON only when
/// request-body trace logging is enabled.
#[derive(Clone, PartialEq, Eq)]
#[derive(Debug, Clone, PartialEq, Eq)]
pub struct EncodedJsonBody {
bytes: Bytes,
trace_bytes: Option<Bytes>,
@@ -33,27 +33,11 @@ impl EncodedJsonBody {
&self.bytes
}
/// Omits credential-bearing bodies from tracing and debug output, including retries.
pub fn without_body_logging(mut self) -> Self {
self.trace_bytes = Some(Bytes::from_static(b"<redacted>"));
self
}
pub(crate) fn trace_bytes(&self) -> &[u8] {
self.trace_bytes.as_ref().unwrap_or(&self.bytes)
}
}
impl std::fmt::Debug for EncodedJsonBody {
fn fmt(&self, formatter: &mut std::fmt::Formatter<'_>) -> std::fmt::Result {
formatter
.debug_struct("EncodedJsonBody")
.field("body", &String::from_utf8_lossy(self.trace_bytes()))
.field("prepared", &self.prepared)
.finish()
}
}
#[derive(Debug, Clone, Copy, Default, PartialEq, Eq)]
pub enum RequestCompression {
#[default]
@@ -136,13 +120,7 @@ impl Request {
self.body,
Some(RequestBody::Json(_) | RequestBody::EncodedJson(_))
);
let retained_trace_bytes = match self.body.as_ref() {
Some(RequestBody::EncodedJson(body)) => body.trace_bytes.clone(),
Some(RequestBody::Json(_) | RequestBody::Raw(_)) | None => None,
};
let trace_bytes = if retained_trace_bytes.is_some() {
retained_trace_bytes
} else if self.compression != RequestCompression::None
let trace_bytes = if self.compression != RequestCompression::None
&& tracing::enabled!(target: "codex_http_client::transport", tracing::Level::TRACE)
{
match self.body.as_ref() {
@@ -1,7 +1,4 @@
use super::*;
use crate::request::EncodedJsonBody;
use crate::request::RequestCompression;
use pretty_assertions::assert_eq;
use serde_json::json;
use std::io::Write;
use std::sync::Arc;
@@ -31,35 +28,6 @@ async fn disabled_request_logging_suppresses_transport_url_and_body() {
assert!(!logs.contains("body-secret"));
}
#[test]
fn sensitive_encoded_body_stays_redacted_after_preparation_and_retries() {
let payload = json!({"client_metadata": {"guardian_ticket": "secret-receipt"}});
for compression in [RequestCompression::None, RequestCompression::Zstd] {
let mut request = Request::new(Method::POST, "https://example.com/responses".into())
.with_compression(compression);
request.body = Some(RequestBody::EncodedJson(
EncodedJsonBody::encode(&payload)
.unwrap()
.without_body_logging(),
));
assert_eq!(request_body_for_trace(&request), "<redacted>");
for _ in 0..2 {
request = request.clone().into_prepared().unwrap();
assert_eq!(request_body_for_trace(&request), "<redacted>");
assert!(!format!("{request:?}").contains("secret-receipt"));
let body = request.prepare_body_for_send().unwrap().body.unwrap();
let decoded = match compression {
RequestCompression::None => body.to_vec(),
RequestCompression::Zstd => zstd::stream::decode_all(body.as_ref()).unwrap(),
};
assert_eq!(
serde_json::from_slice::<serde_json::Value>(&decoded).unwrap(),
payload
);
}
}
}
#[tokio::test]
async fn connection_failures_are_classified_without_exposing_request_urls() {
let unavailable_server =
-37
View File
@@ -1,37 +0,0 @@
//! Opaque parent-inference receipts kept out of persisted and model-visible context.
use std::fmt;
/// Header carried on the parent response's metadata, for both SSE and WebSocket.
pub const GUARDIAN_TICKET_HEADER: &str = "x-codex-guardian-ticket";
/// Request metadata consumed by Codex backend before forwarding reviewer inference.
pub const GUARDIAN_TICKET_METADATA_KEY: &str = "guardian_ticket";
/// A server-issued receipt. Parsing checks its shape, not its authenticity.
///
/// Only the backend can validate ownership, expiry, and remaining allowance. This
/// type intentionally has no serialization implementation or unredacted Debug.
#[derive(Clone, PartialEq, Eq)]
pub struct GuardianTicket(String);
impl GuardianTicket {
/// Accept the bounded URL-safe representation used by the receipt issuer.
pub fn from_server(value: &str) -> Option<Self> {
(value.len() == 43
&& value
.bytes()
.all(|byte| byte.is_ascii_alphanumeric() || matches!(byte, b'-' | b'_')))
.then(|| Self(value.to_owned()))
}
/// Expose the receipt only when writing transport metadata.
pub fn as_str(&self) -> &str {
&self.0
}
}
impl fmt::Debug for GuardianTicket {
fn fmt(&self, formatter: &mut fmt::Formatter<'_>) -> fmt::Result {
formatter.write_str("GuardianTicket([REDACTED])")
}
}
-1
View File
@@ -25,7 +25,6 @@ pub mod dynamic_tools;
mod environment;
pub mod error;
pub mod exec_output;
pub mod guardian_ticket;
pub mod items;
mod legacy_events;
pub mod local_media;
-2
View File
@@ -158,8 +158,6 @@ pub enum CyberAccessProgram {
/// child input, Core also compares root lineage to detect ambiguity.
#[derive(Clone, Debug, Default)]
pub struct TurnStartOptions {
/// Parent inference receipt for an internal Guardian turn. Never persisted.
pub guardian_ticket: Option<crate::guardian_ticket::GuardianTicket>,
/// Source classification for the caller that starts a new turn.
/// Ignored when the submitted input steers an active turn.
pub turn_trigger: Option<String>,