Skip to content

armgit / codex

Public Imported from GitHub
0

Lightweight coding agent that runs in your terminal

Rust

Honor Guardian runtime settings from model defaults (#39738)

felixxia-oai88da5520d41c
diff --git a/codex-rs/ext/guardian-v2/src/async_scorer/config.rs b/codex-rs/ext/guardian-v2/src/async_scorer/config.rs
index 9fc6d949e5..76f8eeb5cd 100644
--- a/codex-rs/ext/guardian-v2/src/async_scorer/config.rs
+++ b/codex-rs/ext/guardian-v2/src/async_scorer/config.rs
@@ -75,6 +75,9 @@ impl GuardianV2Config {
.review_threshold_basis_points
.map(|basis_points| f64::from(basis_points) / 10_000.0)
});
+ configured.max_tool_call_lag = configured
+ .max_tool_call_lag
+ .or(model_defaults.max_tool_call_lag);
configured.reasoning_effort = configured
.reasoning_effort
.or_else(|| model_defaults.reasoning_effort.clone());
@@ -84,6 +87,9 @@ impl GuardianV2Config {
configured.max_classifier_instruction_tokens = configured
.max_classifier_instruction_tokens
.or(model_defaults.max_classifier_instruction_tokens);
+ configured.reuse_parent_compaction = configured
+ .reuse_parent_compaction
+ .or(model_defaults.reuse_parent_compaction);
configured.max_parent_compaction_tokens = configured
.max_parent_compaction_tokens
.or(model_defaults.max_parent_compaction_tokens);
@@ -109,6 +115,9 @@ impl GuardianV2Config {
.collect::<Result<Vec<_>, _>>()?,
);
}
+ transcript.include_images = transcript
+ .include_images
+ .or(model_transcript.include_images);
transcript.max_message_entry_tokens = transcript
.max_message_entry_tokens
.or(model_transcript.max_message_entry_tokens);
diff --git a/codex-rs/ext/guardian-v2/src/async_scorer/config_tests.rs b/codex-rs/ext/guardian-v2/src/async_scorer/config_tests.rs
index 065eec33ac..9836693e81 100644
--- a/codex-rs/ext/guardian-v2/src/async_scorer/config_tests.rs
+++ b/codex-rs/ext/guardian-v2/src/async_scorer/config_tests.rs
@@ -1,6 +1,8 @@
use codex_features::GuardianV2ConfigToml;
use codex_features::GuardianV2ReviewScopeConfigToml;
+use codex_features::GuardianV2TranscriptConfigToml;
use codex_protocol::openai_models::GuardianV2ModelConfig;
+use codex_protocol::openai_models::GuardianV2TranscriptModelConfig;
use codex_protocol::openai_models::ReasoningEffort;
use codex_protocol::protocol::TruncationPolicy;
use pretty_assertions::assert_eq;
@@ -145,3 +147,49 @@ fn model_prompt_and_explicit_threshold_precedence_are_preserved() {
0.5,
);
}
+
+#[test]
+fn model_runtime_settings_preserve_local_overrides() {
+ let defaults = GuardianV2ModelConfig {
+ max_tool_call_lag: Some(1),
+ reuse_parent_compaction: Some(false),
+ transcript: Some(GuardianV2TranscriptModelConfig {
+ include_images: Some(true),
+ ..Default::default()
+ }),
+ ..Default::default()
+ };
+ let inherited = GuardianV2Config::from_overrides(GuardianV2ConfigToml::default())
+ .unwrap()
+ .with_model_defaults(Some(&defaults))
+ .unwrap();
+ assert_eq!(
+ (
+ inherited.max_tool_call_lag,
+ inherited.reuse_parent_compaction,
+ inherited.transcript.include_images,
+ ),
+ (1, false, true)
+ );
+
+ let overridden = GuardianV2Config::from_overrides(GuardianV2ConfigToml {
+ max_tool_call_lag: Some(4),
+ reuse_parent_compaction: Some(true),
+ transcript: Some(GuardianV2TranscriptConfigToml {
+ include_images: Some(false),
+ ..Default::default()
+ }),
+ ..Default::default()
+ })
+ .unwrap()
+ .with_model_defaults(Some(&defaults))
+ .unwrap();
+ assert_eq!(
+ (
+ overridden.max_tool_call_lag,
+ overridden.reuse_parent_compaction,
+ overridden.transcript.include_images,
+ ),
+ (4, true, false)
+ );
+}
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 5f57d76d34..8bf8c64bec 100644
--- a/codex-rs/ext/guardian-v2/src/async_scorer/extension.rs
+++ b/codex-rs/ext/guardian-v2/src/async_scorer/extension.rs
@@ -461,6 +461,12 @@ impl GuardianV2Extension {
return;
}
};
+ if guardian_config.transcript.include_images {
+ input
+ .thread_store
+ .get_or_init(NodeReplReviewEvidence::default)
+ .enable_image_capture();
+ }
input.thread_store.insert(guardian_config.clone());
let tool_call_index = score_progress
.latest_tool_call
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 eb46697ebf..4af0c635a5 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
@@ -7,6 +7,7 @@ use std::time::SystemTime;
use anyhow::Result;
use codex_core::config::Config;
+use codex_core::context::NodeReplReviewEvidence;
use codex_extension_api::ConversationHistorySnapshot;
use codex_extension_api::ExtensionData;
use codex_extension_api::ExtensionMetrics;
@@ -1231,15 +1232,18 @@ async fn contributor_uses_model_defaults_and_preserves_local_overrides() -> Resu
let model_defaults = GuardianV2ModelConfig {
classifier_instructions: Some("Use the experimental model-owned prompt.".to_owned()),
review_threshold_basis_points: Some(6_000),
+ max_tool_call_lag: Some(2),
reasoning_effort: Some(ReasoningEffort::Minimal),
transcript: Some(GuardianV2TranscriptModelConfig {
sources: Some(vec!["reasoning".to_owned()]),
+ include_images: Some(true),
max_message_entry_tokens: Some(128),
max_message_transcript_tokens: Some(256),
..Default::default()
}),
max_action_tokens: Some(128),
max_classifier_instruction_tokens: Some(256),
+ reuse_parent_compaction: Some(false),
max_parent_compaction_tokens: Some(384),
};
let conversation_history = vec![
@@ -1312,13 +1316,19 @@ async fn contributor_uses_model_defaults_and_preserves_local_overrides() -> Resu
let session_store = ExtensionData::new("session-1");
let thread_store = test.codex.thread_extension_data();
+ let guardian_config = thread_store
+ .get::<crate::async_scorer::config::GuardianV2Config>()
+ .expect("Guardian v2 configuration should be installed");
assert_eq!(
- thread_store
- .get::<crate::async_scorer::config::GuardianV2Config>()
- .expect("Guardian v2 configuration should be installed")
- .max_parent_compaction_tokens,
- 384
+ (
+ guardian_config.max_tool_call_lag,
+ guardian_config.reuse_parent_compaction,
+ guardian_config.max_parent_compaction_tokens,
+ guardian_config.transcript.include_images,
+ ),
+ (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() {
tokio::task::yield_now().await;
diff --git a/codex-rs/protocol/src/openai_models/guardian_v2.rs b/codex-rs/protocol/src/openai_models/guardian_v2.rs
index 814bf3ad23..953a89a4e7 100644
--- a/codex-rs/protocol/src/openai_models/guardian_v2.rs
+++ b/codex-rs/protocol/src/openai_models/guardian_v2.rs
@@ -13,6 +13,8 @@ pub struct GuardianV2ModelConfig {
#[serde(default, skip_serializing_if = "Option::is_none")]
pub review_threshold_basis_points: Option<u16>,
#[serde(default, skip_serializing_if = "Option::is_none")]
+ pub max_tool_call_lag: Option<usize>,
+ #[serde(default, skip_serializing_if = "Option::is_none")]
pub reasoning_effort: Option<ReasoningEffort>,
#[serde(default, skip_serializing_if = "Option::is_none")]
pub transcript: Option<GuardianV2TranscriptModelConfig>,
@@ -21,6 +23,8 @@ pub struct GuardianV2ModelConfig {
#[serde(default, skip_serializing_if = "Option::is_none")]
pub max_classifier_instruction_tokens: Option<usize>,
#[serde(default, skip_serializing_if = "Option::is_none")]
+ pub reuse_parent_compaction: Option<bool>,
+ #[serde(default, skip_serializing_if = "Option::is_none")]
pub max_parent_compaction_tokens: Option<usize>,
}
@@ -30,6 +34,8 @@ pub struct GuardianV2TranscriptModelConfig {
#[serde(default, skip_serializing_if = "Option::is_none")]
pub sources: Option<Vec<String>>,
#[serde(default, skip_serializing_if = "Option::is_none")]
+ pub include_images: Option<bool>,
+ #[serde(default, skip_serializing_if = "Option::is_none")]
pub max_message_entry_tokens: Option<usize>,
#[serde(default, skip_serializing_if = "Option::is_none")]
pub max_tool_entry_tokens: Option<usize>,