Update guardian output schema (#17061)

## Summary
- Update guardian output schema to separate risk, authorization,
outcome, and rationale.
- Feed guardian rationale into rejection messages.
- Split the guardian policy into template and tenant-config sections.

## Validation
- `cargo test -p codex-core mcp_tool_call`
- `env -u CODEX_SANDBOX_NETWORK_DISABLED INSTA_UPDATE=always cargo test
-p codex-core guardian::`

---------

Co-authored-by: Owen Lin <owen@openai.com>
This commit is contained in:
maja-openai
2026-04-08 15:47:29 -07:00
committed by GitHub
Unverified
parent 49677ec71f
commit dcbc91fd39
45 changed files with 673 additions and 312 deletions
+12 -12
View File
@@ -3231,13 +3231,13 @@ fn loads_compact_prompt_from_file() -> std::io::Result<()> {
}
#[test]
fn load_config_uses_requirements_guardian_developer_instructions() -> std::io::Result<()> {
fn load_config_uses_requirements_guardian_policy_config() -> std::io::Result<()> {
let codex_home = TempDir::new()?;
let config_layer_stack = ConfigLayerStack::new(
Vec::new(),
Default::default(),
crate::config_loader::ConfigRequirementsToml {
guardian_developer_instructions: Some(
guardian_policy_config: Some(
" Use the workspace-managed guardian policy. ".to_string(),
),
..Default::default()
@@ -3256,7 +3256,7 @@ fn load_config_uses_requirements_guardian_developer_instructions() -> std::io::R
)?;
assert_eq!(
config.guardian_developer_instructions.as_deref(),
config.guardian_policy_config.as_deref(),
Some("Use the workspace-managed guardian policy.")
);
@@ -3264,13 +3264,13 @@ fn load_config_uses_requirements_guardian_developer_instructions() -> std::io::R
}
#[test]
fn load_config_ignores_empty_requirements_guardian_developer_instructions() -> std::io::Result<()> {
fn load_config_ignores_empty_requirements_guardian_policy_config() -> std::io::Result<()> {
let codex_home = TempDir::new()?;
let config_layer_stack = ConfigLayerStack::new(
Vec::new(),
Default::default(),
crate::config_loader::ConfigRequirementsToml {
guardian_developer_instructions: Some(" ".to_string()),
guardian_policy_config: Some(" ".to_string()),
..Default::default()
},
)
@@ -3286,7 +3286,7 @@ fn load_config_ignores_empty_requirements_guardian_developer_instructions() -> s
config_layer_stack,
)?;
assert_eq!(config.guardian_developer_instructions, None);
assert_eq!(config.guardian_policy_config, None);
Ok(())
}
@@ -4558,7 +4558,7 @@ fn test_precedence_fixture_with_o3_profile() -> std::io::Result<()> {
experimental_realtime_ws_startup_context: None,
base_instructions: None,
developer_instructions: None,
guardian_developer_instructions: None,
guardian_policy_config: None,
include_permissions_instructions: true,
include_apps_instructions: true,
include_environment_context: true,
@@ -4705,7 +4705,7 @@ fn test_precedence_fixture_with_gpt3_profile() -> std::io::Result<()> {
experimental_realtime_ws_startup_context: None,
base_instructions: None,
developer_instructions: None,
guardian_developer_instructions: None,
guardian_policy_config: None,
include_permissions_instructions: true,
include_apps_instructions: true,
include_environment_context: true,
@@ -4850,7 +4850,7 @@ fn test_precedence_fixture_with_zdr_profile() -> std::io::Result<()> {
experimental_realtime_ws_startup_context: None,
base_instructions: None,
developer_instructions: None,
guardian_developer_instructions: None,
guardian_policy_config: None,
include_permissions_instructions: true,
include_apps_instructions: true,
include_environment_context: true,
@@ -4981,7 +4981,7 @@ fn test_precedence_fixture_with_gpt5_profile() -> std::io::Result<()> {
experimental_realtime_ws_startup_context: None,
base_instructions: None,
developer_instructions: None,
guardian_developer_instructions: None,
guardian_policy_config: None,
include_permissions_instructions: true,
include_apps_instructions: true,
include_environment_context: true,
@@ -5041,7 +5041,7 @@ fn test_requirements_web_search_mode_allowlist_does_not_warn_when_unset() -> any
rules: None,
enforce_residency: None,
network: None,
guardian_developer_instructions: None,
guardian_policy_config: None,
};
let requirement_source = crate::config_loader::RequirementSource::Unknown;
let requirement_source_for_error = requirement_source.clone();
@@ -5662,7 +5662,7 @@ async fn explicit_sandbox_mode_falls_back_when_disallowed_by_requirements() -> s
rules: None,
enforce_residency: None,
network: None,
guardian_developer_instructions: None,
guardian_policy_config: None,
};
let config = ConfigBuilder::without_managed_config_for_tests()
+10 -8
View File
@@ -254,8 +254,11 @@ pub struct Config {
/// Developer instructions override injected as a separate message.
pub developer_instructions: Option<String>,
/// Guardian-specific developer instructions override from requirements.toml.
pub guardian_developer_instructions: Option<String>,
/// Guardian-specific tenant policy config override from requirements.toml.
/// This is inserted into the fixed guardian prompt template under the
/// `# Policy Configuration` section rather than replacing the whole
/// guardian developer prompt.
pub guardian_policy_config: Option<String>,
/// Whether to inject the `<permissions instructions>` developer block.
pub include_permissions_instructions: bool,
@@ -1840,9 +1843,8 @@ impl Config {
.include_environment_context
.or(cfg.include_environment_context)
.unwrap_or(true);
let guardian_developer_instructions = guardian_developer_instructions_from_requirements(
config_layer_stack.requirements_toml(),
);
let guardian_policy_config =
guardian_policy_config_from_requirements(config_layer_stack.requirements_toml());
let personality = personality
.or(config_profile.personality)
.or(cfg.personality)
@@ -2065,7 +2067,7 @@ impl Config {
.show_raw_agent_reasoning
.or(show_raw_agent_reasoning)
.unwrap_or(false),
guardian_developer_instructions,
guardian_policy_config,
model_reasoning_effort: config_profile
.model_reasoning_effort
.or(cfg.model_reasoning_effort),
@@ -2273,11 +2275,11 @@ pub(crate) fn uses_deprecated_instructions_file(config_layer_stack: &ConfigLayer
.any(|layer| toml_uses_deprecated_instructions_file(&layer.config))
}
fn guardian_developer_instructions_from_requirements(
fn guardian_policy_config_from_requirements(
requirements_toml: &ConfigRequirementsToml,
) -> Option<String> {
requirements_toml
.guardian_developer_instructions
.guardian_policy_config
.as_deref()
.and_then(|value| {
let trimmed = value.trim();