Make Guardian risk score persistence opt-in (#40911)

## What changed

- Add the `features.guardianv2.persist_scores` configuration option for debugging.
- Default the option to `false`, so Guardian V2 reviewed actions and risk scores are only written to rollout files when explicitly enabled and the session is not ephemeral.

## Testing

- Cover configuration deserialization, default non-persistence, and opt-in persistence for nested code-mode actions.

GitOrigin-RevId: 452af99ef14553d45140340ececc2913389124d0
This commit is contained in:
jif
2026-08-26 16:23:34 +00:00
committed by copyberry
parent bde9db1375
commit 5b92c2d2f5
6 changed files with 24 additions and 22 deletions
+4
View File
@@ -1441,6 +1441,10 @@
"minimum": 0.0,
"type": "integer"
},
"persist_scores": {
"description": "Persist reviewed actions and risk scores to rollout files for debugging.",
"type": "boolean"
},
"reasoning_effort": {
"$ref": "#/definitions/ReasoningEffort"
},
@@ -33,6 +33,7 @@ pub(crate) enum GuardianV2ReviewScope {
#[derive(Clone, Debug, PartialEq)]
pub(crate) struct GuardianV2Config {
local_overrides: GuardianV2ConfigToml,
pub(crate) persist_scores: bool,
pub(crate) classifier_instructions: String,
pub(crate) review_threshold: f64,
pub(crate) max_tool_call_lag: usize,
@@ -229,6 +230,7 @@ impl GuardianV2Config {
Ok(Self {
local_overrides: configured.clone(),
persist_scores: configured.persist_scores.unwrap_or(false),
classifier_instructions: configured
.classifier_instructions
.unwrap_or_else(|| DEFAULT_CLASSIFIER_INSTRUCTIONS.to_owned()),
@@ -804,7 +804,8 @@ impl GuardianV2Extension {
.latest_scored_tool_call
.fetch_max(tool_call_index, Ordering::Release);
classification_finished_at = Some(Instant::now());
if !config.ephemeral
if guardian_config.persist_scores
&& !config.ephemeral
&& let Err(error) = thread
.append_rollout_items(&[RolloutItem::SecurityRiskScore(score)])
.await
@@ -1770,26 +1770,16 @@ async fn contributor_samples_tool_calls_with_the_existing_luna_pool() -> Result<
);
assert!(score.sampled_at.is_some());
test.codex.ensure_rollout_materialized().await;
let persisted_score = tokio::time::timeout(Duration::from_secs(5), async {
loop {
if let Some(persisted_score) = test
.codex
.load_history(/*include_archived*/ false)
.await?
.items
.into_iter()
.find_map(|item| match item {
RolloutItem::SecurityRiskScore(score) => Some(score),
_ => None,
})
{
return Ok::<_, anyhow::Error>(persisted_score);
}
tokio::task::yield_now().await;
}
})
.await??;
assert_eq!(&persisted_score, score.as_ref());
assert!(
!test
.codex
.load_history(/*include_archived*/ false)
.await?
.items
.into_iter()
.any(|item| matches!(item, RolloutItem::SecurityRiskScore(_))),
"risk scores should not be persisted unless explicitly enabled"
);
assert_eq!(
registry
.fast_approval_decision(
@@ -1867,7 +1857,7 @@ async fn contributor_persists_nested_code_mode_action_with_score() -> Result<()>
Vec::new(),
r#"{"path":"README.md"}"#,
Some(TEST_GUARDIAN_POLICY),
"",
"[features.guardianv2]\nenabled = true\npersist_scores = true\n",
/*model_defaults*/ None,
ToolCallSource::CodeMode {
cell_id: "cell-1".to_owned(),
+3
View File
@@ -129,6 +129,9 @@ pub struct GuardianV2ConfigToml {
/// Route Guardian review and classification through the unmetered Codex endpoints.
#[serde(skip_serializing_if = "Option::is_none")]
pub free_guardian: Option<bool>,
/// Persist reviewed actions and risk scores to rollout files for debugging.
#[serde(skip_serializing_if = "Option::is_none")]
pub persist_scores: Option<bool>,
#[serde(skip_serializing_if = "Option::is_none")]
pub classifier_instructions: Option<String>,
#[serde(skip_serializing_if = "Option::is_none")]
+2
View File
@@ -164,6 +164,7 @@ fn guardian_v2_feature_config_deserializes_classifier_and_transcript_settings()
[guardianv2]
enabled = true
free_guardian = true
persist_scores = true
classifier_instructions = "Review this action"
review_threshold = 0.65
max_tool_call_lag = 2
@@ -194,6 +195,7 @@ max_recent_non_user_entries = 12
Some(FeatureToml::Config(crate::GuardianV2ConfigToml {
enabled: Some(true),
free_guardian: Some(true),
persist_scores: Some(true),
classifier_instructions: Some("Review this action".to_owned()),
review_threshold: Some(0.65),
max_tool_call_lag: Some(2),