Move Guardian reporting and denial accounting into the extension (#44544)

## What changed

- Move assessment event construction, metrics, and analytics tracking into `codex-guardian-reviewer` through `ReviewReport`.
- Store denial accounting in thread extension data through `ReviewDenials`, with core retaining turn interruption and lifecycle cleanup.
- Move failed-review record selection and bounded serialization into the extension, with core supplying captured review context.

## Testing

Add coverage that denial accounting clears on turn completion and interruption. Move the oversized-record test into the extension and exercise the new API.

GitOrigin-RevId: 6e5b3d9d4128b356f99b1d0e74a68111e40261ab
This commit is contained in:
jif
2026-09-10 14:13:57 +00:00
committed by copyberry
parent eca63f0803
commit 4e6d5c0a96
17 changed files with 455 additions and 296 deletions
+10 -92
View File
@@ -1,63 +1,25 @@
//! Captures bounded failed-review evidence for opt-in feedback uploads.
//! Records outlive replaced and short-lived reviewers without writing extra history files.
//! Captures failed-review context for the extension's bounded feedback recorder.
use super::GuardianAssessmentOutcome;
use super::approval_request::format_guardian_action_pretty;
use super::approval_request::guardian_request_target_item_id;
use super::approval_request::guardian_request_turn_id;
use super::prompt::parse_guardian_assessment;
use super::review_session::GuardianReviewSessionOutcome;
use super::review_session::GuardianReviewSessionParams;
use crate::session::session::Session;
use codex_feedback::record_guardian_review_failure;
use codex_protocol::ThreadId;
use codex_protocol::models::ResponseItem;
use serde::Serialize;
use std::io;
use std::io::Write;
const MAX_RECORD_BYTES: usize = 4 * 1024 * 1024;
#[derive(Serialize)]
struct ReviewFeedbackRecord<'a> {
reviewed_thread_id: ThreadId,
reviewed_turn_id: &'a str,
target_item_id: Option<&'a str>,
reviewer_thread_id: ThreadId,
model: &'a str,
status: &'a str,
decision: Option<&'a str>,
action: &'a str,
action_truncated: bool,
instructions: Option<&'a str>,
history: Vec<&'a ResponseItem>,
context_omitted: bool,
}
pub(super) async fn record_failed_review(
reviewer: &Session,
params: &GuardianReviewSessionParams,
outcome: &GuardianReviewSessionOutcome,
) {
if !params.spawn_config.feedback_enabled || params.spawn_config.ephemeral {
let Some(feedback) = codex_guardian_reviewer::FailedReviewFeedback::for_outcome(
outcome,
codex_guardian_reviewer::ReviewFeedbackSettings {
enabled: params.spawn_config.feedback_enabled,
ephemeral: params.spawn_config.ephemeral,
},
) else {
return;
}
let (status, decision) = match outcome {
GuardianReviewSessionOutcome::Completed(Ok(Some(decision))) => {
match parse_guardian_assessment(Some(decision)) {
Ok(assessment) if assessment.outcome == GuardianAssessmentOutcome::Allow => {
return;
}
Ok(_) => ("denied", Some(decision.as_str())),
Err(_) => ("invalid_decision", Some(decision.as_str())),
}
}
GuardianReviewSessionOutcome::Completed(Ok(None)) => ("missing_decision", None),
GuardianReviewSessionOutcome::Completed(Err(_))
| GuardianReviewSessionOutcome::PromptBuildFailed(_)
| GuardianReviewSessionOutcome::SessionFailed { .. } => ("failed", None),
GuardianReviewSessionOutcome::TimedOut => ("timed_out", None),
GuardianReviewSessionOutcome::Aborted => ("aborted", None),
};
let Ok(action) = format_guardian_action_pretty(&params.request) else {
tracing::warn!("Could not serialize Guardian feedback action");
@@ -65,7 +27,7 @@ pub(super) async fn record_failed_review(
};
let instructions = reviewer.get_prompt_base_instructions().await;
let history = reviewer.clone_history().await;
ReviewFeedbackRecord {
feedback.store(codex_guardian_reviewer::ReviewFeedbackContext {
reviewed_thread_id: params.parent_session.thread_id(),
reviewed_turn_id: guardian_request_turn_id(
&params.request,
@@ -74,53 +36,9 @@ pub(super) async fn record_failed_review(
target_item_id: guardian_request_target_item_id(&params.request),
reviewer_thread_id: reviewer.thread_id(),
model: &params.model,
status,
decision,
action: &action.text,
action_truncated: action.truncated,
instructions: Some(&instructions.text),
history: history.raw_items().collect(),
context_omitted: false,
}
.store();
});
}
impl ReviewFeedbackRecord<'_> {
fn store(mut self) {
let mut buffer = BoundedBuffer(Vec::new());
if serde_json::to_writer(&mut buffer, &self).is_err() {
// Preserve the action and decision even when optional reviewer history is too large.
self.instructions = None;
self.history.clear();
self.context_omitted = true;
buffer = BoundedBuffer(Vec::new());
if serde_json::to_writer(&mut buffer, &self).is_err() {
tracing::warn!("Guardian feedback action and decision exceed the record limit");
return;
}
}
record_guardian_review_failure(self.reviewed_thread_id, buffer.0);
}
}
struct BoundedBuffer(Vec<u8>);
impl Write for BoundedBuffer {
fn write(&mut self, bytes: &[u8]) -> io::Result<usize> {
if bytes.len() > MAX_RECORD_BYTES.saturating_sub(self.0.len()) {
return Err(io::Error::other(
"Guardian feedback record exceeds its size limit",
));
}
self.0.extend_from_slice(bytes);
Ok(bytes.len())
}
fn flush(&mut self) -> io::Result<()> {
Ok(())
}
}
#[cfg(test)]
#[path = "feedback_tests.rs"]
mod tests;
@@ -1,48 +0,0 @@
use super::*;
use pretty_assertions::assert_eq;
use serde_json::json;
#[test]
fn oversized_context_keeps_the_action_and_decision() -> anyhow::Result<()> {
let thread_id = ThreadId::new();
let reviewer_thread_id = ThreadId::new();
let decision = r#"{"outcome":"deny","rationale":"Missing approval."}"#;
let action = r#"{"command":"git push"}"#;
ReviewFeedbackRecord {
reviewed_thread_id: thread_id,
reviewed_turn_id: "parent-turn",
target_item_id: Some("push-call"),
reviewer_thread_id,
model: "review-model",
status: "denied",
decision: Some(decision),
action,
action_truncated: false,
instructions: Some(&"x".repeat(MAX_RECORD_BYTES)),
history: Vec::new(),
context_omitted: false,
}
.store();
let contents = codex_feedback::guardian_review_failures(&[thread_id])
.attachment
.expect("failed-review record")
.buffer;
assert_eq!(
serde_json::from_slice::<serde_json::Value>(&contents)?,
json!({
"reviewed_thread_id": thread_id,
"reviewed_turn_id": "parent-turn",
"target_item_id": "push-call",
"reviewer_thread_id": reviewer_thread_id,
"model": "review-model",
"status": "denied",
"decision": decision,
"action": action,
"action_truncated": false,
"instructions": null,
"history": [],
"context_omitted": true,
})
);
Ok(())
}
-490
View File
@@ -1,490 +0,0 @@
use std::time::Duration;
use codex_analytics::GuardianApprovalRequestSource;
use codex_analytics::GuardianReviewAnalyticsResult;
use codex_analytics::GuardianReviewDecision;
use codex_analytics::GuardianReviewFailureReason;
use codex_analytics::GuardianReviewSessionKind;
use codex_analytics::GuardianReviewTerminalStatus;
use codex_analytics::GuardianReviewedAction;
use codex_mcp::CODEX_APPS_MCP_SERVER_NAME;
use codex_otel::GUARDIAN_REVIEW_COUNT_METRIC;
use codex_otel::GUARDIAN_REVIEW_DURATION_METRIC;
use codex_otel::GUARDIAN_REVIEW_TOKEN_USAGE_METRIC;
use codex_otel::GUARDIAN_REVIEW_TTFT_DURATION_METRIC;
use codex_otel::SessionTelemetry;
use codex_otel::sanitize_metric_tag_value;
use codex_protocol::protocol::GuardianAssessmentOutcome;
use codex_protocol::protocol::GuardianRiskLevel;
use codex_protocol::protocol::GuardianUserAuthorization;
use codex_protocol::protocol::TokenUsage;
pub(crate) fn emit_guardian_review_metrics(
session_telemetry: &SessionTelemetry,
result: &GuardianReviewAnalyticsResult,
approval_request_source: GuardianApprovalRequestSource,
reviewed_action: &GuardianReviewedAction,
completion_latency_ms: u64,
) {
let tags = guardian_review_metric_tags(result, approval_request_source, reviewed_action);
let tag_refs: Vec<(&str, &str)> = tags
.iter()
.map(|(key, value)| (*key, value.as_str()))
.collect();
// Custom MCP tool names may contain user data; only attribute tools from the OpenAI Apps server.
let mcp_tags = if let GuardianReviewedAction::McpToolCall {
server, tool_name, ..
} = reviewed_action
&& server == CODEX_APPS_MCP_SERVER_NAME
{
vec![("tool", sanitize_metric_tag_value(tool_name))]
} else {
Vec::new()
};
let mut counter_tag_refs = tag_refs.clone();
counter_tag_refs.extend(mcp_tags.iter().map(|(key, value)| (*key, value.as_str())));
session_telemetry.counter(
GUARDIAN_REVIEW_COUNT_METRIC,
/*inc*/ 1,
&counter_tag_refs,
);
session_telemetry.record_duration(
GUARDIAN_REVIEW_DURATION_METRIC,
Duration::from_millis(completion_latency_ms),
&tag_refs,
);
if let Some(time_to_first_token_ms) = result.time_to_first_token_ms {
session_telemetry.record_duration(
GUARDIAN_REVIEW_TTFT_DURATION_METRIC,
Duration::from_millis(time_to_first_token_ms),
&tag_refs,
);
}
if let Some(token_usage) = result.token_usage.as_ref() {
emit_guardian_token_usage_histograms(session_telemetry, token_usage, tags);
}
}
fn emit_guardian_token_usage_histograms(
session_telemetry: &SessionTelemetry,
token_usage: &TokenUsage,
base_tags: Vec<(&'static str, String)>,
) {
for (token_type, value) in [
("total", token_usage.total_tokens.max(0)),
("input", token_usage.input_tokens.max(0)),
("cached_input", token_usage.cached_input()),
(
"cache_write_input",
token_usage.cache_write_input_tokens.max(0),
),
("non_cached_input", token_usage.non_cached_input()),
("output", token_usage.output_tokens.max(0)),
(
"reasoning_output",
token_usage.reasoning_output_tokens.max(0),
),
] {
let mut tags = base_tags.clone();
tags.push(("token_type", token_type.to_string()));
let tag_refs: Vec<(&str, &str)> = tags
.iter()
.map(|(key, value)| (*key, value.as_str()))
.collect();
session_telemetry.histogram(GUARDIAN_REVIEW_TOKEN_USAGE_METRIC, value, &tag_refs);
}
}
fn guardian_review_metric_tags(
result: &GuardianReviewAnalyticsResult,
approval_request_source: GuardianApprovalRequestSource,
reviewed_action: &GuardianReviewedAction,
) -> Vec<(&'static str, String)> {
vec![
("decision", decision_tag(result.decision).to_string()),
(
"terminal_status",
terminal_status_tag(result.terminal_status).to_string(),
),
(
"failure_reason",
failure_reason_tag(result.failure_reason).to_string(),
),
(
"approval_request_source",
approval_request_source_tag(approval_request_source).to_string(),
),
("action", reviewed_action_tag(reviewed_action).to_string()),
(
"session_kind",
session_kind_tag(result.guardian_session_kind).to_string(),
),
(
"had_prior_review_context",
optional_bool_tag(result.had_prior_review_context).to_string(),
),
(
"reviewed_action_truncated",
bool_tag(result.reviewed_action_truncated).to_string(),
),
("risk_level", risk_level_tag(result.risk_level).to_string()),
(
"user_authorization",
user_authorization_tag(result.user_authorization).to_string(),
),
("outcome", outcome_tag(result.outcome).to_string()),
(
"guardian_model",
result
.guardian_model
.as_deref()
.map(sanitize_metric_tag_value)
.unwrap_or_else(|| "none".to_string()),
),
(
"guardian_reasoning_effort",
result
.guardian_reasoning_effort
.as_deref()
.map(sanitize_metric_tag_value)
.unwrap_or_else(|| "none".to_string()),
),
]
}
fn decision_tag(decision: GuardianReviewDecision) -> &'static str {
match decision {
GuardianReviewDecision::Approved => "approved",
GuardianReviewDecision::Denied => "denied",
GuardianReviewDecision::Aborted => "aborted",
}
}
fn terminal_status_tag(status: GuardianReviewTerminalStatus) -> &'static str {
match status {
GuardianReviewTerminalStatus::Approved => "approved",
GuardianReviewTerminalStatus::Denied => "denied",
GuardianReviewTerminalStatus::Aborted => "aborted",
GuardianReviewTerminalStatus::TimedOut => "timed_out",
GuardianReviewTerminalStatus::FailedClosed => "failed_closed",
}
}
fn failure_reason_tag(reason: Option<GuardianReviewFailureReason>) -> &'static str {
match reason {
Some(GuardianReviewFailureReason::Timeout) => "timeout",
Some(GuardianReviewFailureReason::Cancelled) => "cancelled",
Some(GuardianReviewFailureReason::PromptBuildError) => "prompt_build_error",
Some(GuardianReviewFailureReason::SessionError) => "session_error",
Some(GuardianReviewFailureReason::ParseError) => "parse_error",
None => "none",
}
}
fn approval_request_source_tag(source: GuardianApprovalRequestSource) -> &'static str {
match source {
GuardianApprovalRequestSource::MainTurn => "main_turn",
GuardianApprovalRequestSource::DelegatedSubagent => "delegated_subagent",
}
}
fn reviewed_action_tag(action: &GuardianReviewedAction) -> &'static str {
match action {
GuardianReviewedAction::Shell { .. } => "shell",
GuardianReviewedAction::UnifiedExec { .. } => "unified_exec",
GuardianReviewedAction::WriteStdin { .. } => "write_stdin",
GuardianReviewedAction::Execve { .. } => "execve",
GuardianReviewedAction::ApplyPatch {} => "apply_patch",
GuardianReviewedAction::NetworkAccess { .. } => "network_access",
GuardianReviewedAction::McpToolCall { .. } => "mcp_tool_call",
GuardianReviewedAction::RequestPermissions {} => "request_permissions",
}
}
fn session_kind_tag(kind: Option<GuardianReviewSessionKind>) -> &'static str {
match kind {
Some(GuardianReviewSessionKind::TrunkNew) => "trunk_new",
Some(GuardianReviewSessionKind::TrunkReused) => "trunk_reused",
Some(GuardianReviewSessionKind::EphemeralForked) => "ephemeral_forked",
None => "none",
}
}
fn optional_bool_tag(value: Option<bool>) -> &'static str {
match value {
Some(true) => "true",
Some(false) => "false",
None => "unknown",
}
}
fn bool_tag(value: bool) -> &'static str {
if value { "true" } else { "false" }
}
fn risk_level_tag(risk_level: Option<GuardianRiskLevel>) -> &'static str {
match risk_level {
Some(GuardianRiskLevel::Low) => "low",
Some(GuardianRiskLevel::Medium) => "medium",
Some(GuardianRiskLevel::High) => "high",
Some(GuardianRiskLevel::Critical) => "critical",
None => "none",
}
}
fn user_authorization_tag(user_authorization: Option<GuardianUserAuthorization>) -> &'static str {
match user_authorization {
Some(GuardianUserAuthorization::Unknown) => "unknown",
Some(GuardianUserAuthorization::Low) => "low",
Some(GuardianUserAuthorization::Medium) => "medium",
Some(GuardianUserAuthorization::High) => "high",
None => "none",
}
}
fn outcome_tag(outcome: Option<GuardianAssessmentOutcome>) -> &'static str {
match outcome {
Some(GuardianAssessmentOutcome::Allow) => "allow",
Some(GuardianAssessmentOutcome::Deny) => "deny",
None => "none",
}
}
#[cfg(test)]
mod tests {
use super::*;
use codex_otel::MetricsClient;
use codex_otel::MetricsConfig;
use codex_protocol::ThreadId;
use codex_protocol::protocol::SessionSource;
use opentelemetry::KeyValue;
use opentelemetry_sdk::metrics::InMemoryMetricExporter;
use opentelemetry_sdk::metrics::data::AggregatedMetrics;
use opentelemetry_sdk::metrics::data::Metric;
use opentelemetry_sdk::metrics::data::MetricData;
use opentelemetry_sdk::metrics::data::ResourceMetrics;
use pretty_assertions::assert_eq;
use std::collections::BTreeMap;
use test_case::test_case;
fn test_session_telemetry() -> SessionTelemetry {
let exporter = InMemoryMetricExporter::default();
let metrics = MetricsClient::new(
MetricsConfig::in_memory("test", "codex-core", env!("CARGO_PKG_VERSION"), exporter)
.with_runtime_reader(),
)
.expect("in-memory metrics client");
SessionTelemetry::new(
ThreadId::new(),
"gpt-5.4",
"gpt-5.4",
/*account_id*/ None,
/*account_email*/ None,
/*auth_mode*/ None,
"test_originator".to_string(),
/*log_user_prompts*/ false,
"tty".to_string(),
SessionSource::Cli,
)
.with_metrics_without_metadata_tags(metrics)
}
fn find_metric<'a>(resource_metrics: &'a ResourceMetrics, name: &str) -> &'a Metric {
for scope_metrics in resource_metrics.scope_metrics() {
for metric in scope_metrics.metrics() {
if metric.name() == name {
return metric;
}
}
}
panic!("metric {name} missing");
}
fn attributes_to_map<'a>(
attributes: impl Iterator<Item = &'a KeyValue>,
) -> BTreeMap<String, String> {
attributes
.map(|kv| (kv.key.as_str().to_string(), kv.value.as_str().to_string()))
.collect()
}
fn counter_point(
resource_metrics: &ResourceMetrics,
name: &str,
) -> (BTreeMap<String, String>, u64) {
let metric = find_metric(resource_metrics, name);
match metric.data() {
AggregatedMetrics::U64(data) => match data {
MetricData::Sum(sum) => {
let points: Vec<_> = sum.data_points().collect();
assert_eq!(points.len(), 1);
let point = points[0];
(attributes_to_map(point.attributes()), point.value())
}
_ => panic!("unexpected counter aggregation"),
},
_ => panic!("unexpected counter data type"),
}
}
fn histogram_sums(
resource_metrics: &ResourceMetrics,
name: &str,
expected_base_tags: &BTreeMap<String, String>,
) -> BTreeMap<String, u64> {
let metric = find_metric(resource_metrics, name);
match metric.data() {
AggregatedMetrics::F64(data) => match data {
MetricData::Histogram(histogram) => histogram
.data_points()
.map(|point| {
let mut attrs = attributes_to_map(point.attributes());
let token_type = attrs
.remove("token_type")
.unwrap_or_else(|| "sample".to_string());
assert_eq!(&attrs, expected_base_tags);
(token_type, point.sum() as u64)
})
.collect(),
_ => panic!("unexpected histogram aggregation"),
},
_ => panic!("unexpected histogram data type"),
}
}
#[test_case(
GuardianReviewedAction::NetworkAccess {
protocol: codex_protocol::approvals::NetworkApprovalProtocol::Https,
port: 443,
},
"network_access",
BTreeMap::new();
"network access"
)]
#[test_case(
GuardianReviewedAction::McpToolCall {
server: CODEX_APPS_MCP_SERVER_NAME.to_string(),
tool_name: "search docs".to_string(),
connector_id: None,
connector_name: None,
tool_title: None,
},
"mcp_tool_call",
BTreeMap::from([("tool".to_string(), "search_docs".to_string())]);
"openai apps tool call"
)]
#[test_case(
GuardianReviewedAction::McpToolCall {
server: "codex_apps_custom".to_string(),
tool_name: "search_private_docs".to_string(),
connector_id: None,
connector_name: None,
tool_title: None,
},
"mcp_tool_call",
BTreeMap::new();
"custom mcp tool call"
)]
fn guardian_review_metrics_record_counts_durations_and_token_usage(
reviewed_action: GuardianReviewedAction,
expected_action: &str,
expected_mcp_tags: BTreeMap<String, String>,
) {
let session_telemetry = test_session_telemetry();
let result = GuardianReviewAnalyticsResult {
decision: GuardianReviewDecision::Approved,
terminal_status: GuardianReviewTerminalStatus::Approved,
risk_level: Some(GuardianRiskLevel::Low),
user_authorization: Some(GuardianUserAuthorization::High),
outcome: Some(GuardianAssessmentOutcome::Allow),
guardian_session_kind: Some(GuardianReviewSessionKind::TrunkReused),
guardian_model: Some("gpt-5.4 guardian".to_string()),
guardian_reasoning_effort: Some("low".to_string()),
had_prior_review_context: Some(true),
reviewed_action_truncated: true,
token_usage: Some(TokenUsage {
input_tokens: 10,
cached_input_tokens: 4,
cache_write_input_tokens: 2,
output_tokens: 3,
reasoning_output_tokens: 2,
total_tokens: 15,
codex_rollout_budget_units: None,
}),
time_to_first_token_ms: Some(123),
..GuardianReviewAnalyticsResult::without_session()
};
emit_guardian_review_metrics(
&session_telemetry,
&result,
GuardianApprovalRequestSource::DelegatedSubagent,
&reviewed_action,
/*completion_latency_ms*/ 456,
);
let snapshot = session_telemetry
.snapshot_metrics()
.expect("runtime metrics snapshot");
let (attrs, value) = counter_point(&snapshot, GUARDIAN_REVIEW_COUNT_METRIC);
let expected_base_tags = BTreeMap::from([
("action".to_string(), expected_action.to_string()),
(
"approval_request_source".to_string(),
"delegated_subagent".to_string(),
),
("decision".to_string(), "approved".to_string()),
("failure_reason".to_string(), "none".to_string()),
("guardian_model".to_string(), "gpt-5.4_guardian".to_string()),
("guardian_reasoning_effort".to_string(), "low".to_string()),
("had_prior_review_context".to_string(), "true".to_string()),
("outcome".to_string(), "allow".to_string()),
("reviewed_action_truncated".to_string(), "true".to_string()),
("risk_level".to_string(), "low".to_string()),
("session_kind".to_string(), "trunk_reused".to_string()),
("terminal_status".to_string(), "approved".to_string()),
("user_authorization".to_string(), "high".to_string()),
]);
let mut expected_counter_tags = expected_base_tags.clone();
expected_counter_tags.extend(expected_mcp_tags);
assert_eq!((attrs, value), (expected_counter_tags, 1));
assert_eq!(
histogram_sums(
&snapshot,
GUARDIAN_REVIEW_TOKEN_USAGE_METRIC,
&expected_base_tags
),
BTreeMap::from([
("cached_input".to_string(), 4),
("cache_write_input".to_string(), 2),
("input".to_string(), 10),
("non_cached_input".to_string(), 6),
("output".to_string(), 3),
("reasoning_output".to_string(), 2),
("total".to_string(), 15),
])
);
assert_eq!(
histogram_sums(
&snapshot,
GUARDIAN_REVIEW_DURATION_METRIC,
&expected_base_tags
),
BTreeMap::from([("sample".to_string(), 456)])
);
assert_eq!(
histogram_sums(
&snapshot,
GUARDIAN_REVIEW_TTFT_DURATION_METRIC,
&expected_base_tags
),
BTreeMap::from([("sample".to_string(), 123)])
);
}
}
-6
View File
@@ -7,7 +7,6 @@ mod coverage;
mod decision;
mod feedback;
mod input_budget;
mod metrics;
mod prompt;
pub(crate) use input_budget::PendingReviewContext;
pub(crate) use input_budget::check_pending as check_pending_guardian_input;
@@ -62,7 +61,6 @@ pub(crate) use runtime::ReviewAction;
pub(crate) use codex_guardian_reviewer::REVIEW_TIMEOUT as GUARDIAN_REVIEW_TIMEOUT;
pub(crate) const GUARDIAN_REVIEWER_NAME: &str = "guardian";
pub(crate) use codex_guardian_reviewer::AUTO_REVIEW_DENIAL_WINDOW_SIZE;
pub(crate) const AUTO_REVIEW_DENIED_ACTION_APPROVAL_DEVELOPER_PREFIX: &str =
codex_guardian_context::MANUAL_APPROVAL_DEVELOPER_PREFIX;
const GUARDIAN_MAX_TOOL_ENTRY_TOKENS: usize = codex_guardian_context::ContextProfile::synchronous()
@@ -171,10 +169,6 @@ impl From<&Arc<TurnContext>> for GuardianReviewContext {
#[cfg(test)]
use codex_guardian_reviewer::guardian_output_schema;
pub(crate) use codex_guardian_reviewer::GuardianRejectionCircuitBreaker;
pub(crate) use codex_guardian_reviewer::GuardianRejectionCircuitBreakerAction;
pub(crate) use codex_guardian_reviewer::GuardianRejectionCircuitBreakerPolicy;
pub(crate) use approval_request::format_guardian_action_pretty;
#[cfg(test)]
use approval_request::guardian_assessment_action;
-1
View File
@@ -361,7 +361,6 @@ pub(crate) fn guardian_truncate_text(content: &str, token_cap: usize) -> (String
}
use codex_guardian_reviewer::guardian_output_contract_prompt;
pub(super) use codex_guardian_reviewer::parse_guardian_assessment;
pub(crate) const BUNDLED_GUARDIAN_POLICY: &str = include_str!("../../assets/guardian/policy.md");
pub(crate) const BUNDLED_GUARDIAN_POLICY_TEMPLATE: &str =
+10 -56
View File
@@ -1,4 +1,4 @@
//! Supplies host review preparation, telemetry and runtime configuration.
//! Supplies host review preparation, interruption and runtime configuration.
//! Guardian's extension owns the synchronous review loop and pool.
#[path = "review_request.rs"]
@@ -7,11 +7,6 @@ mod request;
use crate::context::GuardianContextMode;
use codex_analytics::GuardianApprovalRequestSource;
use codex_analytics::GuardianReviewAnalyticsResult;
use codex_analytics::GuardianReviewDecision;
use codex_analytics::GuardianReviewFailureReason;
use codex_analytics::GuardianReviewTerminalStatus;
use codex_analytics::GuardianReviewTrackContext;
use codex_analytics::GuardianReviewedAction;
use codex_core_plugins::PluginCommandAttribution;
use codex_extension_api::ThreadIdleCause;
use codex_features::Feature;
@@ -22,9 +17,6 @@ use codex_guardian_reviewer::GuardianReviewSessionLimits;
use codex_protocol::config_types::ApprovalsReviewer;
use codex_protocol::protocol::AskForApproval;
use codex_protocol::protocol::EventMsg;
use codex_protocol::protocol::GuardianAssessmentDecisionSource;
use codex_protocol::protocol::GuardianAssessmentEvent;
use codex_protocol::protocol::GuardianAssessmentStatus;
use codex_protocol::protocol::InternalSessionSource;
use codex_protocol::protocol::ReviewDecision;
use codex_protocol::protocol::SessionSource;
@@ -42,21 +34,18 @@ use crate::session::session::Session;
use crate::session::turn_context::TurnContext;
use crate::turn_timing::now_unix_timestamp_ms;
use super::AUTO_REVIEW_DENIAL_WINDOW_SIZE;
use super::ApprovalRequestReasons;
#[cfg(test)]
use super::GUARDIAN_REVIEW_TIMEOUT;
use super::GUARDIAN_REVIEWER_NAME;
use super::GuardianApprovalRequest;
use super::GuardianAssessmentOutcome;
use super::GuardianRejectionCircuitBreakerAction;
use super::GuardianRejectionCircuitBreakerPolicy;
use super::GuardianReviewContext;
use super::approval_request::format_guardian_action_pretty;
use super::approval_request::guardian_assessment_action;
use super::approval_request::guardian_request_target_item_id;
use super::approval_request::guardian_request_turn_id;
use super::approval_request::guardian_reviewed_action;
use super::metrics::emit_guardian_review_metrics;
use super::review_session::GuardianReviewSessionParams;
use super::review_session::build_guardian_review_session_config;
use codex_guardian_reviewer::guardian_output_schema;
@@ -147,48 +136,17 @@ pub(crate) fn is_basic_session_source(session_source: &SessionSource) -> bool {
}
}
fn track_guardian_review(
session: &Session,
tracking: &GuardianReviewTrackContext,
approval_request_source: GuardianApprovalRequestSource,
reviewed_action: &GuardianReviewedAction,
result: GuardianReviewAnalyticsResult,
completed_at_ms: u64,
) {
emit_guardian_review_metrics(
&session.services.session_telemetry,
&result,
approval_request_source,
reviewed_action,
completed_at_ms.saturating_sub(tracking.started_at_ms),
);
session
.services
.analytics_events_client
.track_guardian_review(tracking, result, completed_at_ms);
}
pub(super) async fn record_guardian_non_denial(session: &Arc<Session>, turn_id: &str) {
session
.services
.guardian_rejection_circuit_breaker
.lock()
.await
.record_non_denial(turn_id);
codex_guardian_reviewer::ReviewDenials::for_thread(&session.services.thread_extension_data)
.record_non_denial(turn_id)
.await;
}
async fn record_guardian_denial(session: &Arc<Session>, turn: &Arc<TurnContext>, turn_id: &str) {
let policy = GuardianRejectionCircuitBreakerPolicy::from(turn.model_info().as_ref());
let action = session
.services
.guardian_rejection_circuit_breaker
.lock()
.await
.record_denial(turn_id, policy);
let GuardianRejectionCircuitBreakerAction::InterruptTurn {
consecutive_denials,
recent_denials,
} = action
let Some(message) =
codex_guardian_reviewer::ReviewDenials::for_thread(&session.services.thread_extension_data)
.record_denial(turn_id, turn.model_info())
.await
else {
return;
};
@@ -200,11 +158,7 @@ async fn record_guardian_denial(session: &Arc<Session>, turn: &Arc<TurnContext>,
session
.send_event(
turn.as_ref(),
EventMsg::GuardianWarning(WarningEvent {
message: format!(
"Automatic approval review rejected too many approval requests for this turn ({consecutive_denials} consecutive, {recent_denials} in the last {AUTO_REVIEW_DENIAL_WINDOW_SIZE} reviews); interrupting the turn."
),
}),
EventMsg::GuardianWarning(WarningEvent { message }),
)
.await;
+36 -115
View File
@@ -4,21 +4,12 @@
use super::*;
use crate::codex_thread::GuardianAuthorizationVersion;
use codex_guardian_reviewer::ReviewHost;
use codex_protocol::approvals::GuardianAssessmentAction;
use codex_protocol::approvals::GuardianReviewReason;
pub(in crate::guardian) struct PreparedApproval {
request: GuardianApprovalRequest,
turn: Arc<TurnContext>,
review_reason: GuardianReviewReason,
assessment_turn_id: String,
target_item_id: Option<String>,
plugin_id: Option<String>,
script_path: Option<String>,
action_summary: GuardianAssessmentAction,
reviewed_action: GuardianReviewedAction,
review_tracking: GuardianReviewTrackContext,
started_at_ms: i64,
report: codex_guardian_reviewer::ReviewReport,
root_authorization_version: Option<GuardianAuthorizationVersion>,
user_message_revision: u64,
review_evidence: Option<(
@@ -98,80 +89,46 @@ impl ReviewHost for super::super::runtime::ReviewRuntime {
.as_ref()
.map(PluginCommandAttribution::serialized_fields)
.unzip();
let action_summary = guardian_assessment_action(&request);
let reviewed_action = guardian_reviewed_action(&request);
let review_tracking = GuardianReviewTrackContext::new(
session.thread_id.to_string(),
assessment_turn_id.clone(),
review_id.clone(),
target_item_id.clone(),
approval_request_source,
reviewed_action.clone(),
GUARDIAN_REVIEW_TIMEOUT.as_millis() as u64,
);
let started_at_ms = review_tracking.started_at_ms.try_into().unwrap_or_default();
let report =
codex_guardian_reviewer::ReviewReport::new(codex_guardian_reviewer::ReviewMetadata {
thread_id: session.thread_id.to_string(),
turn_id: assessment_turn_id,
review_id,
target_item_id,
plugin_id,
script_path,
approval_request_source,
reviewed_action: guardian_reviewed_action(&request),
action: guardian_assessment_action(&request),
review_reason,
});
session
.send_event(
turn.as_ref(),
EventMsg::GuardianAssessment(GuardianAssessmentEvent {
review_reason: Some(review_reason),
id: review_id.clone(),
target_item_id: target_item_id.clone(),
plugin_id: plugin_id.clone(),
script_path: script_path.clone(),
turn_id: assessment_turn_id.clone(),
started_at_ms,
completed_at_ms: None,
status: GuardianAssessmentStatus::InProgress,
risk_level: None,
user_authorization: None,
rationale: None,
decision_source: None,
action: action_summary.clone(),
}),
EventMsg::GuardianAssessment(report.started_event()),
)
.await;
if external_cancel
.as_ref()
.is_some_and(CancellationToken::is_cancelled)
{
let completed_at_ms = now_unix_timestamp_ms();
track_guardian_review(
session.as_ref(),
&review_tracking,
approval_request_source,
&reviewed_action,
GuardianReviewAnalyticsResult {
decision: GuardianReviewDecision::Aborted,
terminal_status: GuardianReviewTerminalStatus::Aborted,
failure_reason: Some(GuardianReviewFailureReason::Cancelled),
..GuardianReviewAnalyticsResult::without_session()
},
let completed = report.complete(
GuardianReviewOutcome::Error(GuardianReviewError::Cancelled),
turn.model_info(),
GuardianReviewAnalyticsResult::without_session(),
completed_at_ms,
);
report.track(
&session.services.session_telemetry,
&session.services.analytics_events_client,
completed.analytics,
completed_at_ms.try_into().unwrap_or_default(),
);
session
.send_event(
turn.as_ref(),
EventMsg::GuardianAssessment(GuardianAssessmentEvent {
review_reason: Some(review_reason),
id: review_id,
target_item_id,
plugin_id: plugin_id.clone(),
script_path: script_path.clone(),
turn_id: assessment_turn_id.clone(),
started_at_ms,
completed_at_ms: Some(completed_at_ms),
status: GuardianAssessmentStatus::Aborted,
risk_level: None,
user_authorization: None,
rationale: None,
decision_source: Some(GuardianAssessmentDecisionSource::Agent),
action: action_summary,
}),
)
.send_event(turn.as_ref(), EventMsg::GuardianAssessment(completed.event))
.await;
record_guardian_non_denial(&session, &assessment_turn_id).await;
record_guardian_non_denial(&session, report.turn_id()).await;
return Err(ReviewDecision::Abort);
}
@@ -205,15 +162,7 @@ impl ReviewHost for super::super::runtime::ReviewRuntime {
Ok(PreparedApproval {
request,
turn,
review_reason,
assessment_turn_id,
target_item_id,
plugin_id,
script_path,
action_summary,
reviewed_action,
review_tracking,
started_at_ms,
report,
root_authorization_version,
user_message_revision,
review_evidence,
@@ -246,23 +195,12 @@ impl ReviewHost for super::super::runtime::ReviewRuntime {
let PreparedApproval {
request: _,
turn,
review_reason,
assessment_turn_id,
target_item_id,
plugin_id,
script_path,
action_summary,
reviewed_action,
review_tracking,
started_at_ms,
report,
root_authorization_version,
user_message_revision,
review_evidence,
} = prepared;
let session = Arc::clone(&self.session);
let review_id = self.review_id.clone();
let approval_request_source = self.options.approval_request_source;
let terminal_action = action_summary.clone();
if session.guardian_context_mode == GuardianContextMode::ThreadOwned
&& matches!(&outcome, GuardianReviewOutcome::Completed(assessment) if assessment.outcome == GuardianAssessmentOutcome::Allow)
&& (root_authorization_version
@@ -284,32 +222,15 @@ impl ReviewHost for super::super::runtime::ReviewRuntime {
}
let completed_at_ms = now_unix_timestamp_ms();
let completed = codex_guardian_reviewer::complete_review(
let completed = report.complete(
outcome,
turn.model_info(),
GuardianAssessmentEvent {
review_reason: Some(review_reason),
id: review_id,
target_item_id,
plugin_id,
script_path,
turn_id: assessment_turn_id.clone(),
started_at_ms,
completed_at_ms: Some(completed_at_ms),
status: GuardianAssessmentStatus::InProgress,
risk_level: None,
user_authorization: None,
rationale: None,
decision_source: Some(GuardianAssessmentDecisionSource::Agent),
action: terminal_action,
},
analytics_result,
completed_at_ms,
);
track_guardian_review(
session.as_ref(),
&review_tracking,
approval_request_source,
&reviewed_action,
report.track(
&session.services.session_telemetry,
&session.services.analytics_events_client,
completed.analytics,
completed_at_ms.try_into().unwrap_or_default(),
);
@@ -336,9 +257,9 @@ impl ReviewHost for super::super::runtime::ReviewRuntime {
.send_event(turn.as_ref(), EventMsg::GuardianAssessment(completed.event))
.await;
if completed.assessment_outcome == Some(GuardianAssessmentOutcome::Deny) {
record_guardian_denial(&session, &turn, &assessment_turn_id).await;
record_guardian_denial(&session, &turn, report.turn_id()).await;
} else {
record_guardian_non_denial(&session, &assessment_turn_id).await;
record_guardian_non_denial(&session, report.turn_id()).await;
}
completed.decision
}
-1
View File
@@ -1514,7 +1514,6 @@ impl Session {
models_manager: Arc::clone(&models_manager),
git_root_discovery,
tool_approvals: Mutex::new(ApprovalStore::default()),
guardian_rejection_circuit_breaker: Mutex::new(Default::default()),
runtime_handle: tokio::runtime::Handle::current(),
skills_service,
agents_md_manager,
-2
View File
@@ -6635,7 +6635,6 @@ pub(crate) async fn make_session_and_context() -> (Session, TurnContext) {
models_manager: Arc::clone(&models_manager),
git_root_discovery: Arc::default(),
tool_approvals: Mutex::new(ApprovalStore::default()),
guardian_rejection_circuit_breaker: Mutex::new(Default::default()),
runtime_handle: tokio::runtime::Handle::current(),
skills_service,
agents_md_manager: Arc::new(AgentsMdManager::new(/*user_instructions*/ None)),
@@ -8809,7 +8808,6 @@ where
models_manager: Arc::clone(&models_manager),
git_root_discovery: Arc::default(),
tool_approvals: Mutex::new(ApprovalStore::default()),
guardian_rejection_circuit_breaker: Mutex::new(Default::default()),
runtime_handle: tokio::runtime::Handle::current(),
skills_service,
agents_md_manager: Arc::new(AgentsMdManager::new(/*user_instructions*/ None)),
@@ -1324,3 +1324,48 @@ async fn guardian_subagent_does_not_inherit_parent_exec_policy_rules() {
);
drop(io);
}
#[test_case(TerminalEventKind::TurnComplete; "completion")]
#[test_case(TerminalEventKind::TurnAborted; "interruption")]
#[tokio::test]
async fn terminal_turn_clears_extension_owned_denials(terminal: TerminalEventKind) {
let (session, turn, events) = make_session_and_context_with_rx().await;
let finish = Arc::new(tokio::sync::Notify::new());
session
.spawn_task(
Arc::clone(&turn),
Vec::new(),
HeldStepTask {
kind: TaskKind::Regular,
finish: Arc::clone(&finish),
},
)
.await;
let denials =
codex_guardian_reviewer::ReviewDenials::for_thread(&session.services.thread_extension_data);
for _ in 0..2 {
assert_eq!(
denials.record_denial(&turn.sub_id, turn.model_info()).await,
None
);
}
match terminal {
TerminalEventKind::TurnComplete => finish.notify_one(),
TerminalEventKind::TurnAborted => {
session.abort_all_tasks(TurnAbortReason::Interrupted).await
}
}
recv_terminal_event(&events, terminal).await;
// Delivery precedes accounting cleanup. Wait for the runtime to finish the turn.
timeout(Duration::from_secs(/*secs*/ 5), async {
while session.active_turn.lock().await.is_some() {
tokio::task::yield_now().await;
}
})
.await
.expect("turn becomes idle");
assert_eq!(
denials.record_denial(&turn.sub_id, turn.model_info()).await,
None
);
}
-2
View File
@@ -11,7 +11,6 @@ use crate::current_time::TimeProvider;
use crate::elicitation::ElicitationService;
use crate::environment_selection::ThreadEnvironments;
use crate::exec_policy::ExecPolicyManager;
use crate::guardian::GuardianRejectionCircuitBreaker;
use crate::mcp::McpManager;
use crate::mcp_tool_exposure::McpHandlerCache;
use crate::tools::ExecutedToolCalls;
@@ -68,7 +67,6 @@ pub(crate) struct SessionServices {
pub(crate) git_root_discovery: Arc<GitRootDiscovery>,
pub(crate) session_telemetry: SessionTelemetry,
pub(crate) tool_approvals: Mutex<ApprovalStore>,
pub(crate) guardian_rejection_circuit_breaker: Mutex<GuardianRejectionCircuitBreaker>,
pub(crate) runtime_handle: Handle,
pub(crate) skills_service: Arc<HostSkillsService>,
pub(crate) agents_md_manager: Arc<AgentsMdManager>,
+15 -15
View File
@@ -306,11 +306,11 @@ impl Session {
let cancellation_token = CancellationToken::new();
let done = Arc::new(Notify::new());
self.services
.guardian_rejection_circuit_breaker
.lock()
.await
.clear_turn(&turn_context.sub_id);
codex_guardian_reviewer::ReviewDenials::clear_turn(
&self.services.thread_extension_data,
&turn_context.sub_id,
)
.await;
let (pending_items, _) = self.input_queue.drain_mailbox_input_items().await;
let turn_state = {
@@ -835,11 +835,11 @@ impl Session {
})
};
self.send_event(turn_context.as_ref(), event).await;
self.services
.guardian_rejection_circuit_breaker
.lock()
.await
.clear_turn(&turn_context.sub_id);
codex_guardian_reviewer::ReviewDenials::clear_turn(
&self.services.thread_extension_data,
&turn_context.sub_id,
)
.await;
let cleared_active_turn = {
let mut active = self.active_turn.lock().await;
@@ -991,11 +991,11 @@ impl Session {
duration_ms,
});
self.send_event(task.turn_context.as_ref(), event).await;
self.services
.guardian_rejection_circuit_breaker
.lock()
.await
.clear_turn(&task.turn_context.sub_id);
codex_guardian_reviewer::ReviewDenials::clear_turn(
&self.services.thread_extension_data,
&task.turn_context.sub_id,
)
.await;
// Regular items were flushed before this terminal event was appended; buffering
// thread writers may not flush it without another explicit barrier.
if let Err(err) = self.flush_rollout().await {