Skip to content

armgit / codex

Public Imported from GitHub
0

Lightweight coding agent that runs in your terminal

Rust

Record reviewed actions with security risk scores (#40901)

jifd4998d611ad3
diff --git a/codex-rs/app-server-protocol/schema/precomputed/app-server-exports-stable.json.zst b/codex-rs/app-server-protocol/schema/precomputed/app-server-exports-stable.json.zst
index 11383fee3d..71ff2b933e 100644
Binary files a/codex-rs/app-server-protocol/schema/precomputed/app-server-exports-stable.json.zst and b/codex-rs/app-server-protocol/schema/precomputed/app-server-exports-stable.json.zst differ
diff --git a/codex-rs/app-server-protocol/src/protocol/thread_history_projection_tests.rs b/codex-rs/app-server-protocol/src/protocol/thread_history_projection_tests.rs
index 42624fa408..6cd4744d6a 100644
--- a/codex-rs/app-server-protocol/src/protocol/thread_history_projection_tests.rs
+++ b/codex-rs/app-server-protocol/src/protocol/thread_history_projection_tests.rs
@@ -206,6 +206,8 @@ fn ignores_legacy_abort_without_turn_id_and_context_only_records() {
}));
let security_risk = project(RolloutItem::SecurityRiskScore(SecurityRiskScore {
scores: BTreeMap::from([("action_risk".to_string(), 0.92)]),
+ call_id: None,
+ action: None,
sampled_at: None,
}));
diff --git a/codex-rs/core/src/session/rollout_reconstruction_tests.rs b/codex-rs/core/src/session/rollout_reconstruction_tests.rs
index abc875c3ba..4f8c6ceebd 100644
--- a/codex-rs/core/src/session/rollout_reconstruction_tests.rs
+++ b/codex-rs/core/src/session/rollout_reconstruction_tests.rs
@@ -157,6 +157,8 @@ async fn record_initial_history_ignores_security_risk_scores() {
let user_item = user_message("visible user input");
let security_risk = SecurityRiskScore {
scores: BTreeMap::from([("credential_access".to_string(), 0.92)]),
+ call_id: None,
+ action: None,
sampled_at: None,
};
diff --git a/codex-rs/ext/guardian-v2/src/async_scorer/extension.rs b/codex-rs/ext/guardian-v2/src/async_scorer/extension.rs
index 3d47429f0b..550c01ab6d 100644
--- a/codex-rs/ext/guardian-v2/src/async_scorer/extension.rs
+++ b/codex-rs/ext/guardian-v2/src/async_scorer/extension.rs
@@ -453,6 +453,8 @@ 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)]),
+ call_id: None,
+ action: None,
sampled_at: Some(sampled_at.into()),
};
thread_store.insert_if(score.clone(), |previous| {
@@ -772,6 +774,10 @@ impl GuardianV2Extension {
};
let score = SecurityRiskScore {
scores: BTreeMap::from([("action_risk".to_owned(), action_risk)]),
+ call_id: Some(call_id.clone()),
+ action: Some(
+ serde_json::from_str(&planned_action).map_err(|error| error.to_string())?,
+ ),
sampled_at: Some(sampled_at.into()),
};
let accepted =
diff --git a/codex-rs/ext/guardian-v2/src/async_scorer/extension_tests.rs b/codex-rs/ext/guardian-v2/src/async_scorer/extension_tests.rs
index f486ddc016..a4b6dfe048 100644
--- a/codex-rs/ext/guardian-v2/src/async_scorer/extension_tests.rs
+++ b/codex-rs/ext/guardian-v2/src/async_scorer/extension_tests.rs
@@ -259,6 +259,8 @@ fn fail_closed_score_preserves_classification_order() {
let newest_sampled_at = newer_sampled_at + Duration::from_secs(1);
let newer_score = SecurityRiskScore {
scores: BTreeMap::from([("action_risk".to_owned(), 0.25)]),
+ call_id: None,
+ action: None,
sampled_at: Some(newer_sampled_at.into()),
};
thread_store.insert(newer_score.clone());
@@ -272,6 +274,8 @@ fn fail_closed_score_preserves_classification_order() {
GuardianV2Extension::record_fail_closed_score(&thread_store, newest_sampled_at);
let fail_closed_score = SecurityRiskScore {
scores: BTreeMap::from([("action_risk".to_owned(), 1.0)]),
+ call_id: None,
+ action: None,
sampled_at: Some(newest_sampled_at.into()),
};
assert!(!thread_store.insert_if(newer_score.clone(), |previous| {
@@ -412,6 +416,8 @@ async fn computer_use_only_scores_cannot_approve_other_actions() -> Result<()> {
thread_store.insert(config);
thread_store.insert(SecurityRiskScore {
scores: BTreeMap::from([("action_risk".to_owned(), 0.25)]),
+ call_id: None,
+ action: None,
sampled_at: None,
});
let progress = thread_store
@@ -673,6 +679,25 @@ async fn sample_configured_conversation_history(
guardian_policy: Option<&str>,
guardian_config: &str,
model_defaults: Option<GuardianV2ModelConfig>,
+) -> Result<(serde_json::Value, TestCodex, ExtensionRegistry<Config>)> {
+ sample_configured_conversation_history_with_source(
+ conversation_history,
+ arguments,
+ guardian_policy,
+ guardian_config,
+ model_defaults,
+ ToolCallSource::Direct,
+ )
+ .await
+}
+
+async fn sample_configured_conversation_history_with_source(
+ conversation_history: Vec<ResponseItem>,
+ arguments: &str,
+ guardian_policy: Option<&str>,
+ guardian_config: &str,
+ model_defaults: Option<GuardianV2ModelConfig>,
+ source: ToolCallSource,
) -> Result<(serde_json::Value, TestCodex, ExtensionRegistry<Config>)> {
let thread_server = responses::start_mock_server().await;
let guardian_policy = guardian_policy.map(str::to_owned);
@@ -789,7 +814,7 @@ async fn sample_configured_conversation_history(
tool_name: &tool_name,
payload: &tool_payload,
conversation_history: Arc::new(conversation_history),
- source: ToolCallSource::Direct,
+ source,
})
.await;
@@ -837,6 +862,8 @@ impl GuardianFailureFixture {
let thread_store = self.test.codex.thread_extension_data();
thread_store.insert(SecurityRiskScore {
scores: BTreeMap::from([("action_risk".to_owned(), 0.25)]),
+ call_id: None,
+ action: None,
sampled_at: None,
});
let turn_store = ExtensionData::new("turn-1");
@@ -1196,6 +1223,8 @@ max_recent_non_user_entries = 8
assert_eq!(thread_store.get::<StrictReviewReason>(), None);
thread_store.insert(SecurityRiskScore {
scores: BTreeMap::from([("action_risk".to_owned(), 0.65)]),
+ call_id: None,
+ action: None,
sampled_at: None,
});
assert_eq!(
@@ -1217,6 +1246,8 @@ max_recent_non_user_entries = 8
);
thread_store.insert(SecurityRiskScore {
scores: BTreeMap::from([("action_risk".to_owned(), 0.55)]),
+ call_id: None,
+ action: None,
sampled_at: None,
});
assert_eq!(
@@ -1561,14 +1592,23 @@ async fn contributor_uses_model_defaults_and_preserves_local_overrides() -> Resu
(2, false, 384, true)
);
assert!(thread_store.get::<NodeReplReviewEvidence>().is_some());
- tokio::time::timeout(Duration::from_secs(5), async {
- while thread_store.get::<SecurityRiskScore>().is_none() {
+ let score = tokio::time::timeout(Duration::from_secs(5), async {
+ loop {
+ if let Some(score) = thread_store.get::<SecurityRiskScore>() {
+ return score;
+ }
tokio::task::yield_now().await;
}
})
.await?;
+ assert_eq!(
+ score.action,
+ Some(serde_json::from_str::<serde_json::Value>(action)?)
+ );
thread_store.insert(SecurityRiskScore {
scores: BTreeMap::from([("action_risk".to_owned(), 0.65)]),
+ call_id: None,
+ action: None,
sampled_at: None,
});
assert_eq!(
@@ -1723,6 +1763,8 @@ async fn contributor_samples_tool_calls_with_the_existing_luna_pool() -> Result<
score.as_ref(),
&SecurityRiskScore {
scores: BTreeMap::from([("action_risk".to_string(), 1.0)]),
+ call_id: Some("call-1".to_owned()),
+ action: Some(json!({"path": "README.md", "tool": "read_file"})),
sampled_at: score.sampled_at,
}
);
@@ -1761,6 +1803,8 @@ async fn contributor_samples_tool_calls_with_the_existing_luna_pool() -> Result<
);
thread_store.insert(SecurityRiskScore {
scores: BTreeMap::from([("action_risk".to_string(), 0.5)]),
+ call_id: None,
+ action: None,
sampled_at: None,
});
assert_eq!(
@@ -1777,6 +1821,8 @@ async fn contributor_samples_tool_calls_with_the_existing_luna_pool() -> Result<
thread_store.insert(SecurityRiskScore {
scores: BTreeMap::from([("action_risk".to_string(), 0.49)]),
+ call_id: None,
+ action: None,
sampled_at: None,
});
assert_eq!(
@@ -1794,6 +1840,8 @@ async fn contributor_samples_tool_calls_with_the_existing_luna_pool() -> Result<
let disabled_thread_store = ExtensionData::new("disabled-thread");
disabled_thread_store.insert(SecurityRiskScore {
scores: BTreeMap::from([("action_risk".to_string(), 0.25)]),
+ call_id: None,
+ action: None,
sampled_at: None,
});
assert_eq!(
@@ -1811,6 +1859,57 @@ async fn contributor_samples_tool_calls_with_the_existing_luna_pool() -> Result<
Ok(())
}
+#[tokio::test(flavor = "multi_thread", worker_threads = 2)]
+async fn contributor_persists_nested_code_mode_action_with_score() -> Result<()> {
+ skip_if_no_network!(Ok(()));
+
+ let (_request, test, _registry) = sample_configured_conversation_history_with_source(
+ Vec::new(),
+ r#"{"path":"README.md"}"#,
+ Some(TEST_GUARDIAN_POLICY),
+ "",
+ /*model_defaults*/ None,
+ ToolCallSource::CodeMode {
+ cell_id: "cell-1".to_owned(),
+ runtime_tool_call_id: "nested-1".to_owned(),
+ },
+ )
+ .await?;
+ test.codex.ensure_rollout_materialized().await;
+
+ let score = tokio::time::timeout(Duration::from_secs(5), async {
+ loop {
+ if let Some(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>(score);
+ }
+ tokio::task::yield_now().await;
+ }
+ })
+ .await??;
+
+ assert_eq!(
+ score,
+ SecurityRiskScore {
+ scores: BTreeMap::from([("action_risk".to_owned(), 1.0)]),
+ call_id: Some("call-1".to_owned()),
+ action: Some(json!({"path": "README.md", "tool": "read_file"})),
+ sampled_at: score.sampled_at,
+ }
+ );
+
+ Ok(())
+}
+
#[tokio::test(flavor = "multi_thread", worker_threads = 2)]
async fn contributor_skips_models_requiring_managed_guardian_review() -> Result<()> {
skip_if_no_network!(Ok(()));
@@ -1881,6 +1980,8 @@ async fn contributor_skips_models_requiring_managed_guardian_review() -> Result<
thread_store.insert(model_info);
thread_store.insert(SecurityRiskScore {
scores: BTreeMap::from([("action_risk".to_owned(), 0.25)]),
+ call_id: None,
+ action: None,
sampled_at: None,
});
assert_eq!(
@@ -1961,6 +2062,8 @@ async fn contributor_counts_failed_thread_lookups_toward_score_lag() -> Result<(
.await?;
thread_store.insert(SecurityRiskScore {
scores: BTreeMap::from([("action_risk".to_owned(), 0.25)]),
+ call_id: None,
+ action: None,
sampled_at: None,
});
assert_eq!(
@@ -2451,6 +2554,8 @@ async fn contributor_reuses_the_latest_compatible_parent_compaction() -> Result<
fail_closed_score.as_ref(),
&SecurityRiskScore {
scores: BTreeMap::from([("action_risk".to_owned(), 1.0)]),
+ call_id: None,
+ action: None,
sampled_at: fail_closed_score.sampled_at,
}
);
diff --git a/codex-rs/external-agent-migration/src/sessions/append_tests.rs b/codex-rs/external-agent-migration/src/sessions/append_tests.rs
index 17d1c48b05..81f3973f79 100644
--- a/codex-rs/external-agent-migration/src/sessions/append_tests.rs
+++ b/codex-rs/external-agent-migration/src/sessions/append_tests.rs
@@ -66,6 +66,8 @@ fn requires_a_strict_nonempty_model_prefix() {
}
let security_risk = RolloutItem::SecurityRiskScore(SecurityRiskScore {
scores: BTreeMap::from([("action_risk".to_string(), 0.92)]),
+ call_id: None,
+ action: None,
sampled_at: None,
});
metadata_changed.push(security_risk.clone());
diff --git a/codex-rs/memories/write/src/phase1.rs b/codex-rs/memories/write/src/phase1.rs
index 4d98599b93..89d87e39ee 100644
--- a/codex-rs/memories/write/src/phase1.rs
+++ b/codex-rs/memories/write/src/phase1.rs
@@ -728,6 +728,8 @@ mod tests {
RolloutItem::ResponseItem(skill_message.into()),
RolloutItem::SecurityRiskScore(SecurityRiskScore {
scores: BTreeMap::from([("action_risk".to_string(), 0.92)]),
+ call_id: None,
+ action: None,
sampled_at: None,
}),
RolloutItem::ResponseItem(subagent_message.clone().into()),
diff --git a/codex-rs/protocol/src/security_risk.rs b/codex-rs/protocol/src/security_risk.rs
index ca16fae4f5..151a0a98d8 100644
--- a/codex-rs/protocol/src/security_risk.rs
+++ b/codex-rs/protocol/src/security_risk.rs
@@ -12,6 +12,12 @@ use std::collections::BTreeMap;
#[derive(Clone, Debug, PartialEq, Serialize, Deserialize, JsonSchema)]
pub struct SecurityRiskScore {
pub scores: BTreeMap<String, f64>,
+ /// The tool call whose action produced this score, when available.
+ #[serde(default, skip_serializing_if = "Option::is_none")]
+ pub call_id: Option<String>,
+ /// The bounded tool action that was reviewed to produce this score.
+ #[serde(default, skip_serializing_if = "Option::is_none")]
+ pub action: Option<serde_json::Value>,
/// When sampling started, if this snapshot was written by a timestamp-aware client.
#[serde(default, skip_serializing_if = "Option::is_none")]
#[schemars(with = "Option<String>")]
diff --git a/codex-rs/rollout/src/recorder_tests.rs b/codex-rs/rollout/src/recorder_tests.rs
index 42cf907028..8ccad419b4 100644
--- a/codex-rs/rollout/src/recorder_tests.rs
+++ b/codex-rs/rollout/src/recorder_tests.rs
@@ -445,6 +445,8 @@ async fn load_rollout_items_preserves_security_risk_scores() -> std::io::Result<
("action_risk".to_string(), 0.76),
("data_exfiltration".to_string(), 0.31),
]),
+ call_id: Some("call-1".to_owned()),
+ action: Some(serde_json::json!({"path": "README.md", "tool": "read_file"})),
sampled_at: None,
};
let security_risk_item = RolloutItem::SecurityRiskScore(security_risk.clone());