openai/codex
1
0
Fork 0
mirror of https://github.com/openai/codex.git synced 2026-09-09 15:58:47 +00:00
Code Issues Packages Projects Releases Wiki Activity

[codex-analytics] guardian review analytics events emission

Browse Source
This commit is contained in:
rhan-oai
2026-04-17 10:52:31 -07:00
parent a17bef5ee3
commit d4dd4eccc9
5 changed files with 623 additions and 73 deletions
Download Patch File Download Diff File Expand all files Collapse all files

2
codex-rs/core/src/codex_delegate.rs
View File

@@ -3,6 +3,7 @@ use std::sync::Arc;
use async_channel::Receiver;
use async_channel::Sender;
use codex_analytics::GuardianApprovalRequestSource;
use codex_async_utils::OrCancelExt;
use codex_exec_server::EnvironmentManager;
use codex_protocol::protocol::ApplyPatchApprovalRequestEvent;
@@ -753,6 +754,7 @@ fn spawn_guardian_review(
review_id,
request,
retry_reason,
GuardianApprovalRequestSource::DelegatedSubagent,
cancel_token,
));
let _ = tx.send(decision);

2
codex-rs/core/src/guardian/mod.rs
View File

@@ -94,6 +94,8 @@ use prompt::render_guardian_transcript_entries;
#[cfg(test)]
use review::GuardianReviewOutcome;
#[cfg(test)]
use review::GuardianReviewSessionResult;
#[cfg(test)]
use review::run_guardian_review_session as run_guardian_review_session_for_test;
#[cfg(test)]
use review_session::build_guardian_review_session_config as build_guardian_review_session_config_for_test;

483
codex-rs/core/src/guardian/review.rs
View File

@@ -1,5 +1,14 @@
use std::sync::Arc;
use std::time::Instant;
use codex_analytics::GuardianApprovalRequestSource;
use codex_analytics::GuardianReviewDecision;
use codex_analytics::GuardianReviewFailureReason;
use codex_analytics::GuardianReviewSessionKind;
use codex_analytics::GuardianReviewTerminalStatus;
use codex_analytics::GuardianReviewedAction;
use codex_analytics::now_unix_seconds;
use codex_features::Feature;
use codex_protocol::config_types::ApprovalsReviewer;
use codex_protocol::protocol::AskForApproval;
use codex_protocol::protocol::EventMsg;
@@ -10,12 +19,14 @@ use codex_protocol::protocol::GuardianRiskLevel;
use codex_protocol::protocol::GuardianUserAuthorization;
use codex_protocol::protocol::ReviewDecision;
use codex_protocol::protocol::SubAgentSource;
use codex_protocol::protocol::TokenUsage;
use codex_protocol::protocol::WarningEvent;
use tokio_util::sync::CancellationToken;
use crate::session::session::Session;
use crate::session::turn_context::TurnContext;
use super::GUARDIAN_REVIEW_TIMEOUT;
use super::GUARDIAN_REVIEWER_NAME;
use super::GuardianApprovalRequest;
use super::GuardianAssessment;
@@ -26,6 +37,7 @@ use super::approval_request::guardian_request_target_item_id;
use super::approval_request::guardian_request_turn_id;
use super::prompt::guardian_output_schema;
use super::prompt::parse_guardian_assessment;
use super::review_session::GuardianReviewSessionMetadata;
use super::review_session::GuardianReviewSessionOutcome;
use super::review_session::GuardianReviewSessionParams;
use super::review_session::build_guardian_review_session_config;
@@ -73,13 +85,79 @@ pub(crate) fn guardian_timeout_message() -> String {
GUARDIAN_TIMEOUT_INSTRUCTIONS.to_string()
}
#[derive(Debug)]
pub(super) struct GuardianReviewSessionResult {
pub(super) outcome: GuardianReviewOutcome,
pub(super) metadata: Option<GuardianReviewSessionMetadata>,
}
#[derive(Debug)]
pub(super) enum GuardianReviewOutcome {
Completed(anyhow::Result<GuardianAssessment>),
Completed(GuardianAssessment),
Failed(GuardianReviewFailure),
TimedOut,
Aborted,
}
impl GuardianReviewSessionResult {
fn completed(
assessment: GuardianAssessment,
metadata: Option<GuardianReviewSessionMetadata>,
) -> Self {
Self {
outcome: GuardianReviewOutcome::Completed(assessment),
metadata,
}
}
fn failed(
failure: GuardianReviewFailure,
metadata: Option<GuardianReviewSessionMetadata>,
) -> Self {
Self {
outcome: GuardianReviewOutcome::Failed(failure),
metadata,
}
}
fn timed_out(metadata: Option<GuardianReviewSessionMetadata>) -> Self {
Self {
outcome: GuardianReviewOutcome::TimedOut,
metadata,
}
}
fn aborted(metadata: Option<GuardianReviewSessionMetadata>) -> Self {
Self {
outcome: GuardianReviewOutcome::Aborted,
metadata,
}
}
}
#[derive(Debug)]
pub(super) enum GuardianReviewFailure {
PromptBuild(anyhow::Error),
Session(anyhow::Error),
Parse(anyhow::Error),
}
impl GuardianReviewFailure {
fn reason(&self) -> GuardianReviewFailureReason {
match self {
Self::PromptBuild(_) => GuardianReviewFailureReason::PromptBuildError,
Self::Session(_) => GuardianReviewFailureReason::SessionError,
Self::Parse(_) => GuardianReviewFailureReason::ParseError,
}
}
fn error(&self) -> &anyhow::Error {
match self {
Self::PromptBuild(err) | Self::Session(err) | Self::Parse(err) => err,
}
}
}
fn guardian_risk_level_str(level: GuardianRiskLevel) -> &'static str {
match level {
GuardianRiskLevel::Low => "low",
@@ -89,6 +167,176 @@ fn guardian_risk_level_str(level: GuardianRiskLevel) -> &'static str {
}
}
fn guardian_reviewed_action(request: &GuardianApprovalRequest) -> GuardianReviewedAction {
match request {
GuardianApprovalRequest::Shell {
sandbox_permissions,
additional_permissions,
..
} => GuardianReviewedAction::Shell {
sandbox_permissions: *sandbox_permissions,
additional_permissions: additional_permissions.clone(),
},
GuardianApprovalRequest::ExecCommand {
sandbox_permissions,
additional_permissions,
tty,
..
} => GuardianReviewedAction::UnifiedExec {
sandbox_permissions: *sandbox_permissions,
additional_permissions: additional_permissions.clone(),
tty: *tty,
},
#[cfg(unix)]
GuardianApprovalRequest::Execve {
source,
program,
additional_permissions,
..
} => GuardianReviewedAction::Execve {
source: *source,
program: program.clone(),
additional_permissions: additional_permissions.clone(),
},
GuardianApprovalRequest::ApplyPatch { .. } => GuardianReviewedAction::ApplyPatch {},
GuardianApprovalRequest::NetworkAccess { protocol, port, .. } => {
GuardianReviewedAction::NetworkAccess {
protocol: *protocol,
port: *port,
}
}
GuardianApprovalRequest::McpToolCall {
server,
tool_name,
connector_id,
connector_name,
tool_title,
..
} => GuardianReviewedAction::McpToolCall {
server: server.clone(),
tool_name: tool_name.clone(),
connector_id: connector_id.clone(),
connector_name: connector_name.clone(),
tool_title: tool_title.clone(),
},
}
}
struct GuardianReviewAnalyticsContext {
thread_id: String,
turn_id: String,
review_id: String,
target_item_id: Option<String>,
approval_request_source: GuardianApprovalRequestSource,
reviewed_action: GuardianReviewedAction,
started_at: u64,
started_instant: Instant,
}
struct GuardianReviewAnalyticsResult {
decision: GuardianReviewDecision,
terminal_status: GuardianReviewTerminalStatus,
failure_reason: Option<GuardianReviewFailureReason>,
risk_level: Option<GuardianRiskLevel>,
user_authorization: Option<GuardianUserAuthorization>,
outcome: Option<GuardianAssessmentOutcome>,
guardian_thread_id: Option<String>,
guardian_session_kind: Option<GuardianReviewSessionKind>,
guardian_model: Option<String>,
guardian_reasoning_effort: Option<String>,
had_prior_review_context: Option<bool>,
reviewed_action_truncated: bool,
token_usage: Option<TokenUsage>,
time_to_first_token_ms: Option<u64>,
completed_at: u64,
}
impl GuardianReviewAnalyticsResult {
fn from_session_metadata(
metadata: Option<GuardianReviewSessionMetadata>,
completed_at: u64,
) -> Self {
let mut result = Self {
decision: GuardianReviewDecision::Denied,
terminal_status: GuardianReviewTerminalStatus::FailedClosed,
failure_reason: None,
risk_level: None,
user_authorization: None,
outcome: None,
guardian_thread_id: None,
guardian_session_kind: None,
guardian_model: None,
guardian_reasoning_effort: None,
had_prior_review_context: None,
reviewed_action_truncated: false,
token_usage: None,
time_to_first_token_ms: None,
completed_at,
};
if let Some(metadata) = metadata {
result.guardian_thread_id = Some(metadata.guardian_thread_id);
result.guardian_session_kind = Some(metadata.guardian_session_kind);
result.guardian_model = Some(metadata.guardian_model);
result.guardian_reasoning_effort = metadata.guardian_reasoning_effort;
result.had_prior_review_context = Some(metadata.had_prior_review_context);
result.token_usage = metadata.token_usage;
}
result
}
}
impl GuardianReviewAnalyticsContext {
fn track(&self, session: &Session, turn: &TurnContext, result: GuardianReviewAnalyticsResult) {
if !turn.config.features.enabled(Feature::GeneralAnalytics) {
return;
}
let completion_latency_ms = self.started_instant.elapsed().as_millis() as u64;
session
.services
.analytics_events_client
.track_guardian_review(codex_analytics::GuardianReviewEventParams {
thread_id: self.thread_id.clone(),
turn_id: self.turn_id.clone(),
review_id: self.review_id.clone(),
target_item_id: self.target_item_id.clone(),
approval_request_source: self.approval_request_source,
reviewed_action: self.reviewed_action.clone(),
reviewed_action_truncated: result.reviewed_action_truncated,
decision: result.decision,
terminal_status: result.terminal_status,
failure_reason: result.failure_reason,
risk_level: result.risk_level,
user_authorization: result.user_authorization,
outcome: result.outcome,
guardian_thread_id: result.guardian_thread_id,
guardian_session_kind: result.guardian_session_kind,
guardian_model: result.guardian_model,
guardian_reasoning_effort: result.guardian_reasoning_effort,
had_prior_review_context: result.had_prior_review_context,
review_timeout_ms: GUARDIAN_REVIEW_TIMEOUT.as_millis() as u64,
// TODO(rhan-oai): plumb nested Guardian review session tool-call counts.
tool_call_count: None,
time_to_first_token_ms: result.time_to_first_token_ms,
completion_latency_ms: Some(completion_latency_ms),
started_at: self.started_at,
completed_at: Some(result.completed_at),
input_tokens: result.token_usage.as_ref().map(|usage| usage.input_tokens),
cached_input_tokens: result
.token_usage
.as_ref()
.map(|usage| usage.cached_input_tokens),
output_tokens: result.token_usage.as_ref().map(|usage| usage.output_tokens),
reasoning_output_tokens: result
.token_usage
.as_ref()
.map(|usage| usage.reasoning_output_tokens),
total_tokens: result.token_usage.as_ref().map(|usage| usage.total_tokens),
});
}
}
/// Whether this turn should route `on-request` approval prompts through the
/// guardian reviewer instead of surfacing them to the user. ARC may still
/// block actions earlier in the flow.
@@ -116,11 +364,24 @@ async fn run_guardian_review(
review_id: String,
request: GuardianApprovalRequest,
retry_reason: Option<String>,
approval_request_source: GuardianApprovalRequestSource,
external_cancel: Option<CancellationToken>,
) -> ReviewDecision {
let started_at = now_unix_seconds();
let started_instant = Instant::now();
let target_item_id = guardian_request_target_item_id(&request).map(str::to_string);
let assessment_turn_id = guardian_request_turn_id(&request, &turn.sub_id).to_string();
let action_summary = guardian_assessment_action(&request);
let analytics_context = GuardianReviewAnalyticsContext {
thread_id: session.conversation_id.to_string(),
turn_id: assessment_turn_id.clone(),
review_id: review_id.clone(),
target_item_id: target_item_id.clone(),
approval_request_source,
reviewed_action: guardian_reviewed_action(&request),
started_at,
started_instant,
};
session
.send_event(
turn.as_ref(),
@@ -142,6 +403,19 @@ async fn run_guardian_review(
.as_ref()
.is_some_and(CancellationToken::is_cancelled)
{
analytics_context.track(
session.as_ref(),
turn.as_ref(),
GuardianReviewAnalyticsResult {
decision: GuardianReviewDecision::Aborted,
terminal_status: GuardianReviewTerminalStatus::Aborted,
failure_reason: Some(GuardianReviewFailureReason::Cancelled),
..GuardianReviewAnalyticsResult::from_session_metadata(
/*metadata*/ None,
now_unix_seconds(),
)
},
);
session
.send_event(
turn.as_ref(),
@@ -163,28 +437,78 @@ async fn run_guardian_review(
let schema = guardian_output_schema();
let terminal_action = action_summary.clone();
let outcome = Box::pin(run_guardian_review_session(
let GuardianReviewSessionResult { outcome, metadata } = Box::pin(run_guardian_review_session(
session.clone(),
turn.clone(),
request,
retry_reason,
retry_reason.clone(),
schema,
external_cancel,
))
.await;
let completed_at = now_unix_seconds();
let result =
|metadata| GuardianReviewAnalyticsResult::from_session_metadata(metadata, completed_at);
let assessment = match outcome {
GuardianReviewOutcome::Completed(Ok(assessment)) => assessment,
GuardianReviewOutcome::Completed(Err(err)) => GuardianAssessment {
risk_level: GuardianRiskLevel::High,
user_authorization: GuardianUserAuthorization::Unknown,
outcome: GuardianAssessmentOutcome::Deny,
rationale: format!("Automatic approval review failed: {err}"),
},
GuardianReviewOutcome::Completed(assessment) => {
let approved = matches!(assessment.outcome, GuardianAssessmentOutcome::Allow);
analytics_context.track(
session.as_ref(),
turn.as_ref(),
GuardianReviewAnalyticsResult {
decision: if approved {
GuardianReviewDecision::Approved
} else {
GuardianReviewDecision::Denied
},
terminal_status: if approved {
GuardianReviewTerminalStatus::Approved
} else {
GuardianReviewTerminalStatus::Denied
},
failure_reason: None,
risk_level: Some(assessment.risk_level),
user_authorization: Some(assessment.user_authorization),
outcome: Some(assessment.outcome),
..result(metadata)
},
);
assessment
}
GuardianReviewOutcome::Failed(failure) => {
let rationale = format!("Automatic approval review failed: {}", failure.error());
analytics_context.track(
session.as_ref(),
turn.as_ref(),
GuardianReviewAnalyticsResult {
decision: GuardianReviewDecision::Denied,
terminal_status: GuardianReviewTerminalStatus::FailedClosed,
failure_reason: Some(failure.reason()),
..result(metadata)
},
);
GuardianAssessment {
risk_level: GuardianRiskLevel::High,
user_authorization: GuardianUserAuthorization::Unknown,
outcome: GuardianAssessmentOutcome::Deny,
rationale,
}
}
GuardianReviewOutcome::TimedOut => {
let rationale =
"Automatic approval review timed out while evaluating the requested approval."
.to_string();
analytics_context.track(
session.as_ref(),
turn.as_ref(),
GuardianReviewAnalyticsResult {
decision: GuardianReviewDecision::Denied,
terminal_status: GuardianReviewTerminalStatus::TimedOut,
failure_reason: Some(GuardianReviewFailureReason::Timeout),
..result(metadata)
},
);
session
.send_event(
turn.as_ref(),
@@ -212,6 +536,16 @@ async fn run_guardian_review(
return ReviewDecision::TimedOut;
}
GuardianReviewOutcome::Aborted => {
analytics_context.track(
session.as_ref(),
turn.as_ref(),
GuardianReviewAnalyticsResult {
decision: GuardianReviewDecision::Aborted,
terminal_status: GuardianReviewTerminalStatus::Aborted,
failure_reason: Some(GuardianReviewFailureReason::Cancelled),
..result(metadata)
},
);
session
.send_event(
turn.as_ref(),
@@ -311,6 +645,7 @@ pub(crate) async fn review_approval_request(
review_id,
request,
retry_reason,
GuardianApprovalRequestSource::MainTurn,
/*external_cancel*/ None,
))
.await
@@ -322,16 +657,18 @@ pub(crate) async fn review_approval_request_with_cancel(
review_id: String,
request: GuardianApprovalRequest,
retry_reason: Option<String>,
approval_request_source: GuardianApprovalRequestSource,
cancel_token: CancellationToken,
) -> ReviewDecision {
Box::pin(run_guardian_review(
run_guardian_review(
Arc::clone(session),
Arc::clone(turn),
review_id,
request,
retry_reason,
approval_request_source,
Some(cancel_token),
))
)
.await
}
@@ -356,11 +693,16 @@ pub(super) async fn run_guardian_review_session(
retry_reason: Option<String>,
schema: serde_json::Value,
external_cancel: Option<CancellationToken>,
) -> GuardianReviewOutcome {
) -> GuardianReviewSessionResult {
let live_network_config = match session.services.network_proxy.as_ref() {
Some(network_proxy) => match network_proxy.proxy().current_cfg().await {
Ok(config) => Some(config),
Err(err) => return GuardianReviewOutcome::Completed(Err(err)),
Err(err) => {
return GuardianReviewSessionResult::failed(
GuardianReviewFailure::PromptBuild(err),
/*metadata*/ None,
);
}
},
None => None,
};
@@ -410,37 +752,94 @@ pub(super) async fn run_guardian_review_session(
);
let guardian_config = match guardian_config {
Ok(config) => config,
Err(err) => return GuardianReviewOutcome::Completed(Err(err)),
Err(err) => {
return GuardianReviewSessionResult::failed(
GuardianReviewFailure::PromptBuild(err),
/*metadata*/ None,
);
}
};
match Box::pin(
session
.guardian_review_session
.run_review(GuardianReviewSessionParams {
parent_session: Arc::clone(&session),
parent_turn: turn.clone(),
spawn_config: guardian_config,
request,
retry_reason,
schema,
model: guardian_model,
reasoning_effort: guardian_reasoning_effort,
reasoning_summary: turn.reasoning_summary,
personality: turn.personality,
external_cancel,
}),
)
.await
{
GuardianReviewSessionOutcome::Completed(Ok(last_agent_message)) => {
GuardianReviewOutcome::Completed(parse_guardian_assessment(
last_agent_message.as_deref(),
))
let (session_outcome, session_metadata) = Box::pin(session.guardian_review_session.run_review(
GuardianReviewSessionParams {
parent_session: Arc::clone(&session),
parent_turn: turn.clone(),
spawn_config: guardian_config,
request,
retry_reason,
schema,
model: guardian_model,
reasoning_effort: guardian_reasoning_effort,
reasoning_summary: turn.reasoning_summary,
personality: turn.personality,
external_cancel,
},
))
.await;
match session_outcome {
GuardianReviewSessionOutcome::Completed(Ok(last_agent_message)) => match last_agent_message
{
Some(last_agent_message) => {
match parse_guardian_assessment(Some(&last_agent_message)) {
Ok(assessment) => {
GuardianReviewSessionResult::completed(assessment, session_metadata)
}
Err(err) => GuardianReviewSessionResult::failed(
GuardianReviewFailure::Parse(err),
session_metadata,
),
}
}
None => GuardianReviewSessionResult::failed(
GuardianReviewFailure::Session(anyhow::anyhow!(
"guardian review completed without an assessment payload"
)),
session_metadata,
),
},
GuardianReviewSessionOutcome::Completed(Err(err)) => GuardianReviewSessionResult::failed(
GuardianReviewFailure::Session(err),
session_metadata,
),
GuardianReviewSessionOutcome::PromptBuildFailed(err) => {
GuardianReviewSessionResult::failed(
GuardianReviewFailure::PromptBuild(err),
session_metadata,
)
}
GuardianReviewSessionOutcome::Completed(Err(err)) => {
GuardianReviewOutcome::Completed(Err(err))
GuardianReviewSessionOutcome::TimedOut => {
GuardianReviewSessionResult::timed_out(session_metadata)
}
GuardianReviewSessionOutcome::Aborted => {
GuardianReviewSessionResult::aborted(session_metadata)
}
GuardianReviewSessionOutcome::TimedOut => GuardianReviewOutcome::TimedOut,
GuardianReviewSessionOutcome::Aborted => GuardianReviewOutcome::Aborted,
}
}
#[cfg(test)]
mod review_tests {
use super::*;
#[test]
fn guardian_review_failure_reason_distinguishes_failure_kinds() {
let parse_failure = GuardianReviewFailure::Parse(anyhow::anyhow!("bad guardian JSON"));
let prompt_failure =
GuardianReviewFailure::PromptBuild(anyhow::anyhow!("bad prompt/config"));
let session_failure =
GuardianReviewFailure::Session(anyhow::anyhow!("guardian runtime failed"));
assert!(matches!(
parse_failure.reason(),
GuardianReviewFailureReason::ParseError
));
assert!(matches!(
prompt_failure.reason(),
GuardianReviewFailureReason::PromptBuildError
));
assert!(matches!(
session_failure.reason(),
GuardianReviewFailureReason::SessionError
));
}
}

183
codex-rs/core/src/guardian/review_session.rs
View File

@@ -5,6 +5,7 @@ use std::sync::Arc;
use std::time::Duration;
use anyhow::anyhow;
use codex_analytics::GuardianReviewSessionKind;
use codex_protocol::config_types::Personality;
use codex_protocol::config_types::ReasoningSummary as ReasoningSummaryConfig;
use codex_protocol::models::DeveloperInstructions;
@@ -17,6 +18,7 @@ use codex_protocol::protocol::Op;
use codex_protocol::protocol::RolloutItem;
use codex_protocol::protocol::SandboxPolicy;
use codex_protocol::protocol::SubAgentSource;
use codex_protocol::protocol::TokenUsage;
use serde_json::Value;
use tokio::sync::Mutex;
use tokio_util::sync::CancellationToken;
@@ -58,10 +60,21 @@ const GUARDIAN_FOLLOWUP_REVIEW_REMINDER: &str = concat!(
#[derive(Debug)]
pub(crate) enum GuardianReviewSessionOutcome {
Completed(anyhow::Result<Option<String>>),
PromptBuildFailed(anyhow::Error),
TimedOut,
Aborted,
}
#[derive(Debug, Clone)]
pub(crate) struct GuardianReviewSessionMetadata {
pub(crate) guardian_thread_id: String,
pub(crate) guardian_session_kind: GuardianReviewSessionKind,
pub(crate) guardian_model: String,
pub(crate) guardian_reasoning_effort: Option<String>,
pub(crate) had_prior_review_context: bool,
pub(crate) token_usage: Option<TokenUsage>,
}
pub(crate) struct GuardianReviewSessionParams {
pub(crate) parent_session: Arc<Session>,
pub(crate) parent_turn: Arc<TurnContext>,
@@ -101,6 +114,21 @@ struct GuardianReviewState {
last_committed_fork_snapshot: Option<GuardianReviewForkSnapshot>,
}
fn had_prior_review_context(prompt_mode: &GuardianPromptMode) -> bool {
matches!(prompt_mode, GuardianPromptMode::Delta { .. })
}
fn token_usage_delta(start: &TokenUsage, end: &TokenUsage) -> TokenUsage {
TokenUsage {
input_tokens: (end.input_tokens - start.input_tokens).max(0),
cached_input_tokens: (end.cached_input_tokens - start.cached_input_tokens).max(0),
output_tokens: (end.output_tokens - start.output_tokens).max(0),
reasoning_output_tokens: (end.reasoning_output_tokens - start.reasoning_output_tokens)
.max(0),
total_tokens: (end.total_tokens - start.total_tokens).max(0),
}
}
struct EphemeralReviewCleanup {
state: Arc<Mutex<GuardianReviewSessionState>>,
review_session: Option<Arc<GuardianReviewSession>>,
@@ -267,10 +295,14 @@ impl GuardianReviewSessionManager {
pub(crate) async fn run_review(
&self,
params: GuardianReviewSessionParams,
) -> GuardianReviewSessionOutcome {
) -> (
GuardianReviewSessionOutcome,
Option<GuardianReviewSessionMetadata>,
) {
let deadline = tokio::time::Instant::now() + GUARDIAN_REVIEW_TIMEOUT;
let next_reuse_key = GuardianReviewSessionReuseKey::from_spawn_config(&params.spawn_config);
let mut stale_trunk_to_shutdown = None;
let mut spawned_trunk = false;
let trunk_candidate = match run_before_review_deadline(
deadline,
params.external_cancel.as_ref(),
@@ -304,16 +336,17 @@ impl GuardianReviewSessionManager {
{
Ok(Ok(review_session)) => Arc::new(review_session),
Ok(Err(err)) => {
return GuardianReviewSessionOutcome::Completed(Err(err));
return (GuardianReviewSessionOutcome::PromptBuildFailed(err), None);
}
Err(outcome) => return outcome,
Err(outcome) => return (outcome, None),
};
state.trunk = Some(Arc::clone(&review_session));
spawned_trunk = true;
}
state.trunk.as_ref().cloned()
}
Err(outcome) => return outcome,
Err(outcome) => return (outcome, None),
};
if let Some(review_session) = stale_trunk_to_shutdown {
@@ -321,9 +354,12 @@ impl GuardianReviewSessionManager {
}
let Some(trunk) = trunk_candidate else {
return GuardianReviewSessionOutcome::Completed(Err(anyhow!(
"guardian review session was not available after spawn"
)));
return (
GuardianReviewSessionOutcome::Completed(Err(anyhow!(
"guardian review session was not available after spawn"
))),
None,
);
};
if trunk.reuse_key != next_reuse_key {
@@ -349,20 +385,30 @@ impl GuardianReviewSessionManager {
}
};
let (outcome, keep_review_session) =
Box::pin(run_review_on_session(trunk.as_ref(), &params, deadline)).await;
let guardian_session_kind = if spawned_trunk {
GuardianReviewSessionKind::TrunkNew
} else {
GuardianReviewSessionKind::TrunkReused
};
let (outcome, keep_review_session, metadata) = Box::pin(run_review_on_session(
trunk.as_ref(),
&params,
guardian_session_kind,
deadline,
))
.await;
if keep_review_session && matches!(outcome, GuardianReviewSessionOutcome::Completed(_)) {
trunk.refresh_last_committed_fork_snapshot().await;
}
drop(trunk_guard);
if keep_review_session {
outcome
(outcome, Some(metadata))
} else {
if let Some(review_session) = self.remove_trunk_if_current(&trunk).await {
review_session.shutdown_in_background();
}
outcome
(outcome, Some(metadata))
}
}
@@ -459,7 +505,10 @@ impl GuardianReviewSessionManager {
reuse_key: GuardianReviewSessionReuseKey,
deadline: tokio::time::Instant,
fork_snapshot: Option<GuardianReviewForkSnapshot>,
) -> GuardianReviewSessionOutcome {
) -> (
GuardianReviewSessionOutcome,
Option<GuardianReviewSessionMetadata>,
) {
let spawn_cancel_token = CancellationToken::new();
let mut fork_config = params.spawn_config.clone();
fork_config.ephemeral = true;
@@ -478,17 +527,18 @@ impl GuardianReviewSessionManager {
.await
{
Ok(Ok(review_session)) => Arc::new(review_session),
Ok(Err(err)) => return GuardianReviewSessionOutcome::Completed(Err(err)),
Err(outcome) => return outcome,
Ok(Err(err)) => return (GuardianReviewSessionOutcome::PromptBuildFailed(err), None),
Err(outcome) => return (outcome, None),
};
self.register_active_ephemeral(Arc::clone(&review_session))
.await;
let mut cleanup =
EphemeralReviewCleanup::new(Arc::clone(&self.state), Arc::clone(&review_session));
let (outcome, _) = Box::pin(run_review_on_session(
let (outcome, _, metadata) = Box::pin(run_review_on_session(
review_session.as_ref(),
&params,
GuardianReviewSessionKind::EphemeralForked,
deadline,
))
.await;
@@ -496,7 +546,7 @@ impl GuardianReviewSessionManager {
cleanup.disarm();
review_session.shutdown_in_background();
}
outcome
(outcome, Some(metadata))
}
}
@@ -543,8 +593,13 @@ async fn spawn_guardian_review_session(
async fn run_review_on_session(
review_session: &GuardianReviewSession,
params: &GuardianReviewSessionParams,
guardian_session_kind: GuardianReviewSessionKind,
deadline: tokio::time::Instant,
) -> (GuardianReviewSessionOutcome, bool) {
) -> (
GuardianReviewSessionOutcome,
bool,
GuardianReviewSessionMetadata,
) {
let (send_followup_reminder, prompt_mode) = {
let state = review_session.state.lock().await;
@@ -559,6 +614,14 @@ async fn run_review_on_session(
(send_followup_reminder, prompt_mode)
};
let mut guardian_metadata = GuardianReviewSessionMetadata {
guardian_thread_id: review_session.codex.session.conversation_id.to_string(),
guardian_session_kind,
guardian_model: params.model.clone(),
guardian_reasoning_effort: params.reasoning_effort.map(|effort| effort.to_string()),
had_prior_review_context: had_prior_review_context(&prompt_mode),
token_usage: None,
};
if send_followup_reminder {
append_guardian_followup_reminder(review_session).await;
}
@@ -583,6 +646,14 @@ async fn run_review_on_session(
prompt_mode,
)
.await?;
// A fresh Guardian review session may not have token info yet. Treat that
// as a zero baseline so the first review can still report usage.
let token_usage_at_review_start = review_session
.codex
.session
.total_token_usage()
.await
.unwrap_or_default();
review_session
.codex
@@ -602,29 +673,44 @@ async fn run_review_on_session(
})
.await?;
Ok::<GuardianTranscriptCursor, anyhow::Error>(prompt_items.transcript_cursor)
Ok::<(GuardianTranscriptCursor, TokenUsage), anyhow::Error>((
prompt_items.transcript_cursor,
token_usage_at_review_start,
))
}),
)
.await;
let submit_result = match submit_result {
Ok(submit_result) => submit_result,
Err(outcome) => return (outcome, false),
Err(outcome) => return (outcome, false, guardian_metadata),
};
let transcript_cursor = match submit_result {
Ok(transcript_cursor) => transcript_cursor,
let (transcript_cursor, token_usage_at_review_start) = match submit_result {
Ok(submit_result) => submit_result,
Err(err) => {
return (GuardianReviewSessionOutcome::Completed(Err(err)), false);
return (
GuardianReviewSessionOutcome::PromptBuildFailed(err),
false,
guardian_metadata,
);
}
};
let outcome =
wait_for_guardian_review(review_session, deadline, params.external_cancel.as_ref()).await;
if matches!(outcome.0, GuardianReviewSessionOutcome::Completed(_)) {
if outcome.2
&& let Some(total_token_usage) = review_session.codex.session.total_token_usage().await
{
guardian_metadata.token_usage = Some(token_usage_delta(
&token_usage_at_review_start,
&total_token_usage,
));
}
let mut state = review_session.state.lock().await;
state.prior_review_count = state.prior_review_count.saturating_add(1);
state.last_reviewed_transcript_cursor = Some(transcript_cursor);
}
outcome
(outcome.0, outcome.1, guardian_metadata)
}
async fn append_guardian_followup_reminder(review_session: &GuardianReviewSession) {
@@ -653,7 +739,7 @@ async fn wait_for_guardian_review(
review_session: &GuardianReviewSession,
deadline: tokio::time::Instant,
external_cancel: Option<&CancellationToken>,
) -> (GuardianReviewSessionOutcome, bool) {
) -> (GuardianReviewSessionOutcome, bool, bool) {
let timeout = tokio::time::sleep_until(deadline);
tokio::pin!(timeout);
let mut last_error_message: Option<String> = None;
@@ -662,7 +748,7 @@ async fn wait_for_guardian_review(
tokio::select! {
_ = &mut timeout => {
let keep_review_session = interrupt_and_drain_turn(&review_session.codex).await.is_ok();
return (GuardianReviewSessionOutcome::TimedOut, keep_review_session);
return (GuardianReviewSessionOutcome::TimedOut, keep_review_session, false);
}
_ = async {
if let Some(cancel_token) = external_cancel {
@@ -672,7 +758,7 @@ async fn wait_for_guardian_review(
}
} => {
let keep_review_session = interrupt_and_drain_turn(&review_session.codex).await.is_ok();
return (GuardianReviewSessionOutcome::Aborted, keep_review_session);
return (GuardianReviewSessionOutcome::Aborted, keep_review_session, false);
}
event = review_session.codex.next_event() => {
match event {
@@ -684,18 +770,20 @@ async fn wait_for_guardian_review(
return (
GuardianReviewSessionOutcome::Completed(Err(anyhow!(error_message))),
true,
true,
);
}
return (
GuardianReviewSessionOutcome::Completed(Ok(turn_complete.last_agent_message)),
true,
true,
);
}
EventMsg::Error(error) => {
last_error_message = Some(error.message);
}
EventMsg::TurnAborted(_) => {
return (GuardianReviewSessionOutcome::Aborted, true);
return (GuardianReviewSessionOutcome::Aborted, true, false);
}
_ => {}
},
@@ -703,6 +791,7 @@ async fn wait_for_guardian_review(
return (
GuardianReviewSessionOutcome::Completed(Err(err.into())),
false,
false,
);
}
}
@@ -974,4 +1063,44 @@ mod tests {
assert_eq!(outcome.unwrap(), 42);
assert!(!cancel_token.is_cancelled());
}
#[test]
fn had_prior_review_context_tracks_prompt_mode() {
assert!(!had_prior_review_context(&GuardianPromptMode::Full));
assert!(had_prior_review_context(&GuardianPromptMode::Delta {
cursor: GuardianTranscriptCursor {
parent_history_version: 7,
transcript_entry_count: 42,
}
}));
}
#[test]
fn token_usage_delta_never_reports_negative_usage() {
let start = TokenUsage {
input_tokens: 10,
cached_input_tokens: 8,
output_tokens: 6,
reasoning_output_tokens: 4,
total_tokens: 28,
};
let end = TokenUsage {
input_tokens: 15,
cached_input_tokens: 7,
output_tokens: 10,
reasoning_output_tokens: 2,
total_tokens: 34,
};
assert_eq!(
token_usage_delta(&start, &end),
TokenUsage {
input_tokens: 5,
cached_input_tokens: 0,
output_tokens: 4,
reasoning_output_tokens: 0,
total_tokens: 6,
}
);
}
}

26
codex-rs/core/src/guardian/tests.rs
View File

@@ -15,6 +15,7 @@ use crate::config_loader::Sourced;
use crate::session::session::Session;
use crate::session::turn_context::TurnContext;
use crate::test_support;
use codex_analytics::GuardianApprovalRequestSource;
use codex_config::config_toml::ConfigToml;
use codex_exec_server::LOCAL_FS;
use codex_model_provider::create_model_provider;
@@ -703,6 +704,7 @@ async fn cancelled_guardian_review_emits_terminal_abort_without_warning() {
.to_string(),
},
/*retry_reason*/ None,
GuardianApprovalRequestSource::MainTurn,
cancel_token,
)
.await;
@@ -918,7 +920,11 @@ async fn guardian_review_request_layout_matches_model_visible_request_snapshot()
/*external_cancel*/ None,
)
.await;
let GuardianReviewOutcome::Completed(Ok(assessment)) = outcome else {
let GuardianReviewSessionResult {
outcome: GuardianReviewOutcome::Completed(assessment),
..
} = outcome
else {
panic!("expected guardian assessment");
};
assert_eq!(assessment.outcome, GuardianAssessmentOutcome::Allow);
@@ -1127,13 +1133,25 @@ async fn guardian_reuses_prompt_cache_key_and_appends_prior_reviews() -> anyhow:
)
.await;
let GuardianReviewOutcome::Completed(Ok(first_assessment)) = first_outcome else {
let GuardianReviewSessionResult {
outcome: GuardianReviewOutcome::Completed(first_assessment),
..
} = first_outcome
else {
panic!("expected first guardian assessment");
};
let GuardianReviewOutcome::Completed(Ok(second_assessment)) = second_outcome else {
let GuardianReviewSessionResult {
outcome: GuardianReviewOutcome::Completed(second_assessment),
..
} = second_outcome
else {
panic!("expected second guardian assessment");
};
let GuardianReviewOutcome::Completed(Ok(third_assessment)) = third_outcome else {
let GuardianReviewSessionResult {
outcome: GuardianReviewOutcome::Completed(third_assessment),
..
} = third_outcome
else {
panic!("expected third guardian assessment");
};
assert_eq!(first_assessment.outcome, GuardianAssessmentOutcome::Allow);