Fail closed on Guardian V2 risk scoring errors (#39307)

## What changed

- Treat configuration, action serialization, thread lookup, and classification errors as elevated risk instead of retaining a prior low-risk result.
- Track asynchronous scoring failures separately from completed scores so approval review falls back to strict review when the latest tool call could not be scored.
- Preserve newer classification results when recording a fail-closed score.

## Testing

- Add coverage for each failure path and for ordering fail-closed scores with concurrent classifications.

GitOrigin-RevId: 7012f078a24031848f2943354a206866286ad9f4
This commit is contained in:
felixxia-oai
2026-08-18 22:38:37 +00:00
committed by copyberry
parent 6ec012668b
commit c97bd2dcb5
2 changed files with 286 additions and 29 deletions

View File

@@ -202,6 +202,7 @@ const CLASSIFICATION_DURATION_METRIC: &str = "codex.guardian_v2.classification.d
struct GuardianV2ScoreProgress {
latest_tool_call: AtomicUsize,
latest_scored_tool_call: AtomicUsize,
latest_failed_tool_call: AtomicUsize,
metrics: Option<Arc<dyn ExtensionMetrics>>,
}
@@ -324,14 +325,13 @@ impl ApprovalReviewContributor for GuardianV2Extension {
thread_store.get::<GuardianV2Enabled>()?;
let guardian_config = thread_store.get::<GuardianV2Config>()?;
let score_progress = thread_store.get::<GuardianV2ScoreProgress>()?;
let latest_scored_tool_call = score_progress
.latest_scored_tool_call
.load(Ordering::Acquire);
let tool_call_lag = score_progress
.latest_tool_call
.load(Ordering::Acquire)
.saturating_sub(
score_progress
.latest_scored_tool_call
.load(Ordering::Acquire),
);
.saturating_sub(latest_scored_tool_call);
if let Some(metrics) = &extension_metrics {
metrics.histogram(
TOOL_CALL_LAG_METRIC,
@@ -350,6 +350,14 @@ impl ApprovalReviewContributor for GuardianV2Extension {
}
return None;
}
if score_progress
.latest_failed_tool_call
.load(Ordering::Acquire)
> latest_scored_tool_call
{
thread_store.insert(StrictReviewReason::ElevatedRisk);
return None;
}
let score = thread_store
.get::<SecurityRiskScore>()
@@ -367,17 +375,34 @@ impl ApprovalReviewContributor for GuardianV2Extension {
impl ToolLifecycleContributor for GuardianV2Extension {
fn on_tool_start<'a>(&'a self, input: ToolStartInput<'a>) -> ToolLifecycleFuture<'a> {
Box::pin(self.score_tool(input))
}
}
impl GuardianV2Extension {
fn record_fail_closed_score(thread_store: &ExtensionData, sampled_at: SystemTime) {
let score = SecurityRiskScore {
scores: BTreeMap::from([("action_risk".to_owned(), 1.0)]),
sampled_at: Some(sampled_at.into()),
};
thread_store.insert_if(score.clone(), |previous| {
previous.is_none_or(|previous| previous.sampled_at <= score.sampled_at)
});
}
async fn score_tool(&self, input: ToolStartInput<'_>) {
let classification_started_at = Instant::now();
let Some(sampler) = input.thread_store.get::<LunaSampler>() else {
return Box::pin(std::future::ready(()));
return;
};
let Some(guardian_config) = input.thread_store.get::<GuardianV2Config>() else {
return Box::pin(std::future::ready(()));
return;
};
let Some(score_progress) = input.thread_store.get::<GuardianV2ScoreProgress>() else {
return Box::pin(std::future::ready(()));
return;
};
let metrics = score_progress.metrics.clone();
let sampled_at = SystemTime::now();
let parent_model = input.thread_store.get::<ModelInfo>();
let model_defaults = parent_model
.as_ref()
@@ -386,6 +411,7 @@ impl ToolLifecycleContributor for GuardianV2Extension {
let guardian_config = match guardian_config.with_model_defaults(model_defaults) {
Ok(config) => config,
Err(error) => {
Self::record_fail_closed_score(input.thread_store, sampled_at);
record_classification(
metrics.as_deref(),
classification_started_at.elapsed(),
@@ -396,7 +422,7 @@ impl ToolLifecycleContributor for GuardianV2Extension {
turn_id: Some(input.turn_id.to_owned()),
message: error,
});
return Box::pin(std::future::ready(()));
return;
}
};
input.thread_store.insert(guardian_config.clone());
@@ -404,7 +430,6 @@ impl ToolLifecycleContributor for GuardianV2Extension {
.latest_tool_call
.fetch_add(/*val*/ 1, Ordering::Relaxed)
.saturating_add(1);
let sampled_at = SystemTime::now();
let latest_parent_compaction = input
.conversation_history
.items()
@@ -436,19 +461,13 @@ impl ToolLifecycleContributor for GuardianV2Extension {
_ => false,
})
{
let score = SecurityRiskScore {
scores: BTreeMap::from([("action_risk".to_owned(), 1.0)]),
sampled_at: Some(sampled_at.into()),
};
input.thread_store.insert_if(score.clone(), |previous| {
previous.is_none_or(|previous| previous.sampled_at <= score.sampled_at)
});
Self::record_fail_closed_score(input.thread_store, sampled_at);
record_classification(
metrics.as_deref(),
classification_started_at.elapsed(),
"failure",
);
return Box::pin(std::future::ready(()));
return;
}
let event_sink = Arc::clone(&self.event_sink);
let thread_manager = self.thread_manager.clone();
@@ -476,6 +495,38 @@ impl ToolLifecycleContributor for GuardianV2Extension {
};
tokio::spawn(async move {
let thread_context: Result<_, String> = async {
let parsed_thread_id =
ThreadId::from_string(&thread_id).map_err(|error| error.to_string())?;
let manager = thread_manager
.upgrade()
.ok_or_else(|| "thread manager is unavailable".to_string())?;
let thread = manager
.get_thread(parsed_thread_id)
.await
.map_err(|error| error.to_string())?;
Ok((manager, thread))
}
.await;
let (manager, thread) = match thread_context {
Ok(context) => context,
Err(error) => {
score_progress
.latest_failed_tool_call
.fetch_max(tool_call_index, Ordering::Release);
record_classification(
metrics.as_deref(),
classification_started_at.elapsed(),
"failure",
);
event_sink.emit_warning(ExtensionWarning {
thread_id,
turn_id: Some(turn_id),
message: format!("Guardian V2 risk scoring failed: {error}"),
});
return;
}
};
let transcript = guardian_config
.transcript
.build(conversation_history.items());
@@ -486,6 +537,7 @@ impl ToolLifecycleContributor for GuardianV2Extension {
let planned_action = match action.render(guardian_config.max_action_tokens) {
Ok(planned_action) => planned_action,
Err(error) => {
Self::record_fail_closed_score(thread.thread_extension_data(), sampled_at);
record_classification(
metrics.as_deref(),
classification_started_at.elapsed(),
@@ -511,15 +563,6 @@ impl ToolLifecycleContributor for GuardianV2Extension {
]);
let mut classification_finished_at = None;
let result: Result<&str, String> = async {
let parsed_thread_id =
ThreadId::from_string(&thread_id).map_err(|error| error.to_string())?;
let manager = thread_manager
.upgrade()
.ok_or_else(|| "thread manager is unavailable".to_string())?;
let thread = manager
.get_thread(parsed_thread_id)
.await
.map_err(|error| error.to_string())?;
let config = thread.config().await;
let review_model_messages = if config.guardian_policy_config.is_none() {
let review_model_id = review_model_override.as_deref().unwrap_or_else(|| {
@@ -616,6 +659,9 @@ impl ToolLifecycleContributor for GuardianV2Extension {
Ok("success")
}
.await;
if result.is_err() {
Self::record_fail_closed_score(thread.thread_extension_data(), sampled_at);
}
record_classification(
metrics.as_deref(),
classification_finished_at
@@ -633,8 +679,6 @@ impl ToolLifecycleContributor for GuardianV2Extension {
});
}
});
Box::pin(std::future::ready(()))
}
}

View File

@@ -3,6 +3,7 @@ use std::sync::Arc;
use std::sync::Mutex;
use std::sync::atomic::Ordering;
use std::time::Duration;
use std::time::SystemTime;
use anyhow::Result;
use codex_core::config::Config;
@@ -50,6 +51,7 @@ use serde_json::json;
use super::CLASSIFICATION_DURATION_METRIC;
use super::CLASSIFICATION_METRIC;
use super::GuardianV2Extension;
use super::GuardianV2ScoreProgress;
use super::REVIEW_FALLBACK_METRIC;
use super::StrictReviewReason;
@@ -223,6 +225,37 @@ impl ConversationHistorySnapshot for TestConversationHistory {
}
}
#[test]
fn fail_closed_score_preserves_classification_order() {
let thread_store = ExtensionData::new("thread-1");
let newer_sampled_at = SystemTime::UNIX_EPOCH + Duration::from_secs(1);
let newest_sampled_at = newer_sampled_at + Duration::from_secs(1);
let newer_score = SecurityRiskScore {
scores: BTreeMap::from([("action_risk".to_owned(), 0.25)]),
sampled_at: Some(newer_sampled_at.into()),
};
thread_store.insert(newer_score.clone());
GuardianV2Extension::record_fail_closed_score(&thread_store, SystemTime::UNIX_EPOCH);
assert_eq!(
thread_store.get::<SecurityRiskScore>().as_deref(),
Some(&newer_score)
);
GuardianV2Extension::record_fail_closed_score(&thread_store, newest_sampled_at);
let fail_closed_score = SecurityRiskScore {
scores: BTreeMap::from([("action_risk".to_owned(), 1.0)]),
sampled_at: Some(newest_sampled_at.into()),
};
assert!(!thread_store.insert_if(newer_score.clone(), |previous| {
previous.is_none_or(|previous| previous.sampled_at < newer_score.sampled_at)
}));
assert_eq!(
thread_store.get::<SecurityRiskScore>().as_deref(),
Some(&fail_closed_score)
);
}
#[test]
fn encrypted_parent_compaction_preserves_the_latest_valid_item() {
let older = ResponseItem::Compaction {
@@ -533,6 +566,186 @@ async fn sample_configured_conversation_history(
Ok((request.body_json(), test, registry))
}
struct GuardianFailureFixture {
test: TestCodex,
registry: ExtensionRegistry<Config>,
session_store: ExtensionData,
}
impl GuardianFailureFixture {
async fn new() -> Result<Self> {
let (_, test, registry) = sample_conversation_history(
Vec::new(),
r#"{"path":"README.md"}"#,
Some(TEST_GUARDIAN_POLICY),
)
.await?;
let thread_store = test.codex.thread_extension_data();
tokio::time::timeout(Duration::from_secs(5), async {
while thread_store.get::<SecurityRiskScore>().is_none() {
tokio::task::yield_now().await;
}
})
.await?;
Ok(Self {
test,
registry,
session_store: ExtensionData::new("session-1"),
})
}
async fn score_tool(&self, tool_name: ToolName) {
let thread_store = self.test.codex.thread_extension_data();
thread_store.insert(SecurityRiskScore {
scores: BTreeMap::from([("action_risk".to_owned(), 0.25)]),
sampled_at: None,
});
let turn_store = ExtensionData::new("turn-1");
let payload = ToolPayload::Function {
arguments: r#"{"path":"README.md"}"#.to_owned(),
};
self.registry.tool_lifecycle_contributors()[0]
.on_tool_start(ToolStartInput {
session_store: &self.session_store,
thread_store,
turn_store: &turn_store,
turn_id: "turn-1",
call_id: "call-1",
tool_name: &tool_name,
payload: &payload,
conversation_history: Arc::new(TestConversationHistory(Vec::new())),
source: ToolCallSource::Direct,
})
.await;
}
async fn assert_fails_closed(&self) -> Result<()> {
let thread_store = self.test.codex.thread_extension_data();
let score_progress = thread_store
.get::<GuardianV2ScoreProgress>()
.expect("Guardian v2 should track score progress per thread");
tokio::time::timeout(Duration::from_secs(5), async {
while thread_store
.get::<SecurityRiskScore>()
.is_none_or(|score| score.scores.get("action_risk") != Some(&1.0))
&& score_progress
.latest_failed_tool_call
.load(Ordering::Acquire)
<= score_progress
.latest_scored_tool_call
.load(Ordering::Acquire)
{
tokio::task::yield_now().await;
}
})
.await?;
assert_eq!(
self.registry
.approval_review(
&self.session_store,
thread_store,
"review action",
/*extension_metrics*/ None,
)
.await,
None
);
Ok(())
}
}
#[tokio::test(flavor = "multi_thread", worker_threads = 2)]
async fn contributor_fails_closed_when_thread_lookup_fails() -> Result<()> {
skip_if_no_network!(Ok(()));
let fixture = GuardianFailureFixture::new().await?;
fixture
.test
.thread_manager
.remove_thread(&fixture.test.session_configured.thread_id)
.await
.expect("the test thread should exist before simulating a failed lookup");
fixture.score_tool(ToolName::plain("read_file")).await;
fixture.assert_fails_closed().await
}
#[tokio::test(flavor = "multi_thread", worker_threads = 2)]
async fn contributor_fails_closed_when_model_configuration_is_invalid() -> Result<()> {
skip_if_no_network!(Ok(()));
let fixture = GuardianFailureFixture::new().await?;
let mut parent_model = fixture
.test
.thread_manager
.get_models_manager()
.get_model_info("gpt-5.5", &fixture.test.config.to_models_manager_config())
.await;
parent_model
.model_messages
.as_mut()
.expect("test model should expose model messages")
.guardian_v2 = Some(GuardianV2ModelConfig {
max_action_tokens: Some(1),
..Default::default()
});
fixture
.test
.codex
.thread_extension_data()
.insert(parent_model);
fixture.score_tool(ToolName::plain("read_file")).await;
fixture.assert_fails_closed().await
}
#[tokio::test(flavor = "multi_thread", worker_threads = 2)]
async fn contributor_fails_closed_when_action_serialization_fails() -> Result<()> {
skip_if_no_network!(Ok(()));
let fixture = GuardianFailureFixture::new().await?;
let oversized_tool_name = ToolName::plain(
"x".repeat(TruncationPolicy::Tokens(DEFAULT_MODEL_CONTEXT_ITEM_TOKENS).byte_budget()),
);
fixture.score_tool(oversized_tool_name).await;
fixture.assert_fails_closed().await
}
#[tokio::test(flavor = "multi_thread", worker_threads = 2)]
async fn contributor_fails_closed_when_luna_classification_fails() -> Result<()> {
skip_if_no_network!(Ok(()));
let fixture = GuardianFailureFixture::new().await?;
let invalid_score = vec![
ev_assistant_message("sample", r#"{"scores":{"action_risk":"invalid"}}"#),
ev_completed("response-invalid"),
];
let server = responses::start_websocket_server(vec![Vec::new(), vec![invalid_score]]).await;
let mut config = fixture.test.config.clone();
config.model_provider = ModelProviderInfo::create_openai_provider(Some(format!(
"http://{}/v1",
server.uri().trim_start_matches("ws://")
)));
config.features.enable(Feature::GuardianV2)?;
fixture.registry.thread_lifecycle_contributors()[0]
.on_thread_start(ThreadStartInput {
config: &config,
session_source: &SessionSource::Exec,
persistent_thread_state_available: false,
environments: &[],
mcp_resource_client: None,
extension_metrics: None,
session_store: &fixture.session_store,
thread_store: fixture.test.codex.thread_extension_data(),
})
.await;
fixture.score_tool(ToolName::plain("read_file")).await;
fixture.assert_fails_closed().await
}
#[tokio::test(flavor = "multi_thread", worker_threads = 2)]
async fn contributor_renders_policy_inside_a_configured_prompt() -> Result<()> {
skip_if_no_network!(Ok(()));