mirror of
https://github.com/openai/codex.git
synced 2026-10-20 00:18:49 +00:00
Constrain Guardian reviews to parent filesystem permissions (#38377)
## Why Guardian review sessions must not gain access to paths that the parent turn is not allowed to read. ## What changed - Derive Guardian permissions by intersecting managed parent filesystem rules with read-only access, preserving denied paths and restricting network access. - Offer Guardian execution tools only when a managed sandbox can enforce those rules. - Include the selected environment IDs in the review-session reuse key so a session is not reused across different environment sets. ## Testing Update the Guardian reuse integration test to verify that a review cannot read a parent-denied file or write a local file while consecutive reviews still reuse the same session. GitOrigin-RevId: 20f17a6c379f1eda651e8508459d642a51e4ce94
This commit is contained in:
@@ -44,6 +44,7 @@ use crate::config::NetworkProxySpec;
|
||||
use crate::config::Permissions;
|
||||
use crate::context::ContextualUserFragment;
|
||||
use crate::context::GuardianFollowupReviewReminder;
|
||||
use crate::environment_selection::TurnEnvironmentSnapshot;
|
||||
use crate::session::GitEnrichmentPolicy;
|
||||
use crate::session::SessionIo;
|
||||
use crate::session::session::Session;
|
||||
@@ -184,6 +185,7 @@ struct GuardianReviewSessionReuseKey {
|
||||
zsh_path: Option<PathBuf>,
|
||||
features: ManagedFeatures,
|
||||
use_experimental_unified_exec_tool: bool,
|
||||
environment_ids: Vec<String>,
|
||||
}
|
||||
|
||||
impl GuardianReviewSessionReuseKey {
|
||||
@@ -221,8 +223,18 @@ impl GuardianReviewSessionReuseKey {
|
||||
zsh_path: spawn_config.zsh_path.clone(),
|
||||
features: spawn_config.features.clone(),
|
||||
use_experimental_unified_exec_tool: spawn_config.use_experimental_unified_exec_tool,
|
||||
environment_ids: Vec::new(),
|
||||
}
|
||||
}
|
||||
|
||||
fn with_environments(mut self, environments: &TurnEnvironmentSnapshot) -> Self {
|
||||
self.environment_ids = environments
|
||||
.captured_environments()
|
||||
.into_keys()
|
||||
.collect::<Vec<_>>();
|
||||
self.environment_ids.sort_unstable();
|
||||
self
|
||||
}
|
||||
}
|
||||
|
||||
fn encrypted_parent_compaction<'a, I>(items: I) -> Option<ResponseItem>
|
||||
@@ -359,16 +371,18 @@ impl GuardianReviewSessionManager {
|
||||
.enabled(Feature::GuardianReuseParentCompaction)
|
||||
.then(|| encrypted_parent_compaction(parent_history.raw_items()))
|
||||
.flatten();
|
||||
let parent_context = GuardianReviewContext::from(parent_turn);
|
||||
let reuse_key = GuardianReviewSessionReuseKey::from_spawn_config(
|
||||
&spawn_config,
|
||||
parent_session.user_instructions().await,
|
||||
parent_history.history_version(),
|
||||
);
|
||||
)
|
||||
.with_environments(parent_context.environments());
|
||||
let spawn_cancel_token = self.cancellation_token.child_token();
|
||||
let spawn_cancel_guard = spawn_cancel_token.clone().drop_guard();
|
||||
let review_session = spawn_guardian_review_session(
|
||||
&parent_session,
|
||||
&GuardianReviewContext::from(parent_turn),
|
||||
&parent_context,
|
||||
spawn_config,
|
||||
reuse_key,
|
||||
spawn_cancel_token.clone(),
|
||||
@@ -439,8 +453,8 @@ impl GuardianReviewSessionManager {
|
||||
¶ms.spawn_config,
|
||||
params.parent_session.user_instructions().await,
|
||||
parent_history.history_version(),
|
||||
);
|
||||
let mut stale_trunk_to_shutdown = None;
|
||||
)
|
||||
.with_environments(params.parent_context.environments());
|
||||
let mut spawned_trunk = false;
|
||||
let trunk_candidate = match run_before_review_deadline(
|
||||
deadline,
|
||||
@@ -460,8 +474,9 @@ impl GuardianReviewSessionManager {
|
||||
if let Some(trunk) = state.trunk.as_ref()
|
||||
&& trunk.reuse_key != next_reuse_key
|
||||
&& trunk.review_lock.try_acquire().is_ok()
|
||||
&& let Some(stale_trunk) = state.trunk.take()
|
||||
{
|
||||
stale_trunk_to_shutdown = state.trunk.take();
|
||||
stale_trunk.shutdown_in_background();
|
||||
}
|
||||
|
||||
if state.trunk.is_none() {
|
||||
@@ -504,10 +519,6 @@ impl GuardianReviewSessionManager {
|
||||
}
|
||||
};
|
||||
|
||||
if let Some(review_session) = stale_trunk_to_shutdown {
|
||||
review_session.shutdown_in_background();
|
||||
}
|
||||
|
||||
let Some(trunk) = trunk_candidate else {
|
||||
return (
|
||||
GuardianReviewSessionOutcome::Completed(Err(anyhow!(
|
||||
@@ -880,7 +891,7 @@ async fn run_review_on_session(
|
||||
.total_token_usage()
|
||||
.await
|
||||
.unwrap_or_default();
|
||||
let guardian_permission_profile = PermissionProfile::read_only();
|
||||
let guardian_permission_profile = params.spawn_config.permissions.permission_profile().clone();
|
||||
let parent_turn_environments = params.parent_context.environments().to_selections();
|
||||
// TODO(anp): Migrate guardian review thread settings to a PathUri fallback cwd so foreign
|
||||
// parent environments do not fall back to the host-native config cwd.
|
||||
@@ -1127,9 +1138,16 @@ pub(crate) fn build_guardian_review_session_config(
|
||||
guardian_config.notify = None;
|
||||
guardian_config.developer_instructions = None;
|
||||
guardian_config.permissions.approval_policy = Constrained::allow_only(AskForApproval::Never);
|
||||
let guardian_permission_profile = parent_config
|
||||
.permissions
|
||||
.permission_profile()
|
||||
.intersect_with_read_only()
|
||||
.unwrap_or(PermissionProfile::External {
|
||||
network: codex_protocol::permissions::NetworkSandboxPolicy::Restricted,
|
||||
});
|
||||
guardian_config
|
||||
.permissions
|
||||
.set_permission_profile(PermissionProfile::read_only())
|
||||
.set_permission_profile(guardian_permission_profile)
|
||||
.map_err(|err| {
|
||||
anyhow::anyhow!("guardian review session could not set permission profile: {err}")
|
||||
})?;
|
||||
@@ -1931,7 +1949,8 @@ mod tests {
|
||||
.clone_history()
|
||||
.await
|
||||
.history_version(),
|
||||
);
|
||||
)
|
||||
.with_environments(params.parent_context.environments());
|
||||
let manager = Arc::new(GuardianReviewSessionManager {
|
||||
state: Arc::new(Mutex::new(GuardianReviewSessionState {
|
||||
trunk: Some(Arc::new(review_session)),
|
||||
|
||||
@@ -70,6 +70,7 @@ use codex_protocol::dynamic_tools::DynamicToolNamespaceTool;
|
||||
use codex_protocol::dynamic_tools::DynamicToolSpec;
|
||||
use codex_protocol::error::CodexErrorDetails;
|
||||
use codex_protocol::error::Result as CodexResult;
|
||||
use codex_protocol::models::PermissionProfile;
|
||||
use codex_protocol::openai_models::ConfigShellToolType;
|
||||
use codex_protocol::openai_models::InputModality;
|
||||
use codex_protocol::openai_models::ToolMode;
|
||||
@@ -890,9 +891,16 @@ fn code_mode_namespace_descriptions(
|
||||
#[instrument(level = "trace", skip_all)]
|
||||
fn add_core_tool_sources(context: &CoreToolPlanContext<'_>, registry: &mut ToolRegistry) {
|
||||
// Guardian reviewers receive only `exec_command`, `write_stdin`, and `view_image`
|
||||
// when an environment is available; all general tool sources stay excluded.
|
||||
// when a managed sandbox can enforce the parent's filesystem restrictions;
|
||||
// all general tool sources stay excluded.
|
||||
if crate::guardian::is_guardian_reviewer_source(&context.turn_context.session_source) {
|
||||
let turn_context = context.turn_context;
|
||||
if !matches!(
|
||||
turn_context.permission_profile(),
|
||||
PermissionProfile::Managed { .. }
|
||||
) {
|
||||
return;
|
||||
}
|
||||
let environment_mode = tool_environment_mode(context.environments);
|
||||
if environment_mode.has_environment() {
|
||||
let include_environment_id = matches!(environment_mode, ToolEnvironmentMode::Multiple);
|
||||
|
||||
@@ -19,8 +19,14 @@ use codex_history::RolloutLine;
|
||||
use codex_login::CodexAuth;
|
||||
use codex_protocol::ThreadId;
|
||||
use codex_protocol::config_types::ApprovalsReviewer;
|
||||
use codex_protocol::models::PermissionProfile;
|
||||
use codex_protocol::openai_models::MODEL_SPECIALTY_CYBER;
|
||||
use codex_protocol::openai_models::ModelsResponse;
|
||||
use codex_protocol::permissions::FileSystemAccessMode;
|
||||
use codex_protocol::permissions::FileSystemPath;
|
||||
use codex_protocol::permissions::FileSystemSandboxEntry;
|
||||
use codex_protocol::permissions::FileSystemSandboxPolicy;
|
||||
use codex_protocol::permissions::NetworkSandboxPolicy;
|
||||
use codex_protocol::protocol::AskForApproval;
|
||||
use codex_protocol::protocol::EventMsg;
|
||||
use codex_protocol::protocol::Op;
|
||||
@@ -33,7 +39,6 @@ use core_test_support::responses::assert_parent_turn;
|
||||
use core_test_support::responses::assert_root_turn;
|
||||
use core_test_support::responses::ev_assistant_message;
|
||||
use core_test_support::responses::ev_completed;
|
||||
use core_test_support::responses::ev_custom_tool_call;
|
||||
use core_test_support::responses::ev_function_call;
|
||||
use core_test_support::responses::ev_response_created;
|
||||
use core_test_support::responses::mount_response_once_match;
|
||||
@@ -353,25 +358,44 @@ async fn guardian_session_is_reused_for_consecutive_tool_reviews_without_prewarm
|
||||
"Guardian approval actions require host-native paths"
|
||||
);
|
||||
|
||||
const SECRET: &str = "guardian-parent-policy-test-secret";
|
||||
let server = start_mock_server().await;
|
||||
let approval_policy = AskForApproval::OnRequest;
|
||||
let sandbox_policy = SandboxPolicy::WorkspaceWrite {
|
||||
writable_roots: vec![],
|
||||
network_access: false,
|
||||
exclude_tmpdir_env_var: true,
|
||||
exclude_slash_tmp: true,
|
||||
};
|
||||
let sandbox_policy_for_config = sandbox_policy.clone();
|
||||
|
||||
let mut builder = test_codex().with_config(move |config| {
|
||||
config.permissions.approval_policy = Constrained::allow_any(approval_policy);
|
||||
config.approvals_reviewer = ApprovalsReviewer::AutoReview;
|
||||
config
|
||||
.set_legacy_sandbox_policy(sandbox_policy_for_config)
|
||||
.expect("set sandbox policy");
|
||||
});
|
||||
let mut builder = test_codex()
|
||||
.with_config(move |config| {
|
||||
let secret_file = config.cwd.join("guardian-secret.txt");
|
||||
config.permissions.approval_policy = Constrained::allow_any(approval_policy);
|
||||
config.approvals_reviewer = ApprovalsReviewer::AutoReview;
|
||||
let mut file_system_policy = FileSystemSandboxPolicy::workspace_write(
|
||||
&[],
|
||||
/*exclude_tmpdir_env_var*/ true,
|
||||
/*exclude_slash_tmp*/ true,
|
||||
);
|
||||
file_system_policy.entries.push(FileSystemSandboxEntry::new(
|
||||
FileSystemPath::Path { path: secret_file },
|
||||
FileSystemAccessMode::Deny,
|
||||
));
|
||||
config
|
||||
.permissions
|
||||
.set_permission_profile(PermissionProfile::from_runtime_permissions(
|
||||
&file_system_policy,
|
||||
NetworkSandboxPolicy::Restricted,
|
||||
))
|
||||
.expect("set parent permission profile");
|
||||
})
|
||||
.with_workspace_setup(|cwd, fs| async move {
|
||||
fs.write_file(
|
||||
&codex_utils_path_uri::PathUri::from_abs_path(&cwd.join("guardian-secret.txt")),
|
||||
SECRET.as_bytes().to_vec(),
|
||||
/*sandbox*/ None,
|
||||
)
|
||||
.await?;
|
||||
Ok(())
|
||||
});
|
||||
let test = builder.build_with_auto_env(&server).await?;
|
||||
|
||||
let secret_file = test.config.cwd.join("guardian-secret.txt");
|
||||
let guardian_output_file = test.cwd.path().join("guardian-write.txt");
|
||||
let first_output_file = test.cwd.path().join("guardian-first.txt");
|
||||
let second_output_file = test.cwd.path().join("guardian-second.txt");
|
||||
let first_command = format!("printf first > {}", first_output_file.display());
|
||||
@@ -388,6 +412,14 @@ async fn guardian_session_is_reused_for_consecutive_tool_reviews_without_prewarm
|
||||
"sandbox_permissions": SandboxPermissions::RequireEscalated,
|
||||
"justification": "Exercise the second Guardian approval.",
|
||||
});
|
||||
let guardian_tool_args = json!({
|
||||
"cmd": format!(
|
||||
"cat {}; printf hostile > {}",
|
||||
secret_file.display(),
|
||||
guardian_output_file.display()
|
||||
),
|
||||
"sandbox_permissions": SandboxPermissions::UseDefault,
|
||||
});
|
||||
let responses = mount_sse_sequence(
|
||||
&server,
|
||||
vec![
|
||||
@@ -402,6 +434,15 @@ async fn guardian_session_is_reused_for_consecutive_tool_reviews_without_prewarm
|
||||
]),
|
||||
sse(vec![
|
||||
ev_response_created("resp-guardian-first-review"),
|
||||
ev_function_call(
|
||||
"exec-guardian-denied-read",
|
||||
"exec_command",
|
||||
&serde_json::to_string(&guardian_tool_args)?,
|
||||
),
|
||||
ev_completed("resp-guardian-first-review"),
|
||||
]),
|
||||
sse(vec![
|
||||
ev_response_created("resp-guardian-first-assessment"),
|
||||
ev_assistant_message(
|
||||
"msg-guardian-first-review",
|
||||
&json!({
|
||||
@@ -412,7 +453,7 @@ async fn guardian_session_is_reused_for_consecutive_tool_reviews_without_prewarm
|
||||
})
|
||||
.to_string(),
|
||||
),
|
||||
ev_completed("resp-guardian-first-review"),
|
||||
ev_completed("resp-guardian-first-assessment"),
|
||||
]),
|
||||
sse(vec![
|
||||
ev_response_created("resp-parent-second-tool"),
|
||||
@@ -456,7 +497,6 @@ async fn guardian_session_is_reused_for_consecutive_tool_reviews_without_prewarm
|
||||
environments: Some(local_selections(test.config.cwd.clone())),
|
||||
approval_policy: Some(approval_policy),
|
||||
approvals_reviewer: Some(ApprovalsReviewer::AutoReview),
|
||||
sandbox_policy: Some(sandbox_policy),
|
||||
..Default::default()
|
||||
}),
|
||||
)
|
||||
@@ -465,25 +505,22 @@ async fn guardian_session_is_reused_for_consecutive_tool_reviews_without_prewarm
|
||||
matches!(event, EventMsg::TurnComplete(_))
|
||||
})
|
||||
.await;
|
||||
|
||||
let guardian_requests = responses
|
||||
.requests()
|
||||
.into_iter()
|
||||
let requests = responses.requests();
|
||||
assert!(
|
||||
requests
|
||||
.iter()
|
||||
.all(|request| !request.body_contains_text(SECRET)),
|
||||
"Guardian disclosed a file denied by the parent task"
|
||||
);
|
||||
let guardian_requests = requests
|
||||
.iter()
|
||||
.filter(|request| {
|
||||
request.body_json()["client_metadata"]["x-openai-subagent"].as_str() == Some("guardian")
|
||||
})
|
||||
.collect::<Vec<_>>();
|
||||
assert_eq!(guardian_requests.len(), 2);
|
||||
assert_eq!(guardian_requests.len(), 3);
|
||||
let first_guardian_request = guardian_requests[0].body_json();
|
||||
let second_guardian_request = guardian_requests[1].body_json();
|
||||
let parent_request = responses.requests()[0].body_json();
|
||||
let parent_turn_id = parent_request["client_metadata"]["turn_id"]
|
||||
.as_str()
|
||||
.expect("reviewed parent turn id");
|
||||
for request in [&first_guardian_request, &second_guardian_request] {
|
||||
assert_parent_turn(request, Some(parent_turn_id))?;
|
||||
assert_root_turn(request, Some(parent_turn_id))?;
|
||||
}
|
||||
let second_guardian_request = guardian_requests[2].body_json();
|
||||
let first_guardian_thread_id = first_guardian_request["client_metadata"]["thread_id"]
|
||||
.as_str()
|
||||
.expect("first Guardian review should have a thread id");
|
||||
@@ -491,67 +528,16 @@ async fn guardian_session_is_reused_for_consecutive_tool_reviews_without_prewarm
|
||||
.as_str()
|
||||
.expect("second Guardian review should have a thread id");
|
||||
assert_eq!(first_guardian_thread_id, second_guardian_thread_id);
|
||||
assert!(
|
||||
!guardian_output_file.exists(),
|
||||
"Guardian wrote a local file"
|
||||
);
|
||||
assert_eq!(fs::read_to_string(first_output_file)?, "first");
|
||||
assert_eq!(fs::read_to_string(second_output_file)?, "second");
|
||||
|
||||
Ok(())
|
||||
}
|
||||
|
||||
#[tokio::test(flavor = "multi_thread", worker_threads = 2)]
|
||||
async fn guardian_review_of_code_mode_exec_command_includes_outer_source() -> Result<()> {
|
||||
skip_if_no_network!(Ok(()));
|
||||
skip_if_sandbox!(Ok(()));
|
||||
skip_if_wine_exec!(
|
||||
Ok(()),
|
||||
"Guardian approval actions require host-native paths"
|
||||
);
|
||||
|
||||
let server = start_mock_server().await;
|
||||
let mut builder = test_codex()
|
||||
.with_model("test-gpt-5.1-codex")
|
||||
.with_config(|config| {
|
||||
config.permissions.approval_policy = Constrained::allow_any(AskForApproval::OnRequest);
|
||||
config.approvals_reviewer = ApprovalsReviewer::AutoReview;
|
||||
let _ = config.features.enable(Feature::CodeMode);
|
||||
});
|
||||
let test = builder.build_with_auto_env(&server).await?;
|
||||
|
||||
let code = r#"const outerContext = "guardian-code-mode-outer-context";
|
||||
await tools.exec_command({cmd: "true", sandbox_permissions: "require_escalated", justification: "test"});"#;
|
||||
let responses = mount_sse_sequence(
|
||||
&server,
|
||||
vec![
|
||||
sse(vec![
|
||||
ev_custom_tool_call("code-mode-call", "exec", code),
|
||||
ev_completed("resp-parent"),
|
||||
]),
|
||||
sse(vec![
|
||||
ev_assistant_message("guardian", r#"{"outcome":"allow"}"#),
|
||||
ev_completed("resp-guardian"),
|
||||
]),
|
||||
sse(vec![ev_completed("resp-done")]),
|
||||
],
|
||||
)
|
||||
.await;
|
||||
|
||||
let prompt = "run a nested command that requires Guardian review";
|
||||
test.submit_text_turn(prompt).await?;
|
||||
|
||||
let transcript = responses.requests()[1].message_input_texts("user")[2..6].concat();
|
||||
assert_eq!(
|
||||
transcript,
|
||||
format!(
|
||||
">>> TRANSCRIPT START\n\
|
||||
[1] user: {prompt}\n\
|
||||
\n\
|
||||
[2] tool exec call: {code}\n\
|
||||
>>> TRANSCRIPT END\n"
|
||||
)
|
||||
);
|
||||
|
||||
Ok(())
|
||||
}
|
||||
|
||||
#[tokio::test(flavor = "multi_thread", worker_threads = 2)]
|
||||
async fn interrupted_guardian_tool_review_aborts_without_executing_the_command() -> Result<()> {
|
||||
skip_if_no_network!(Ok(()));
|
||||
|
||||
@@ -387,6 +387,33 @@ impl PermissionProfile {
|
||||
}
|
||||
}
|
||||
|
||||
/// Intersects managed filesystem permissions with read-only access and restricts network.
|
||||
///
|
||||
/// Returns `None` when filesystem enforcement belongs to an external caller.
|
||||
pub fn intersect_with_read_only(&self) -> Option<Self> {
|
||||
let mut file_system = self.file_system_sandbox_policy();
|
||||
match file_system.kind {
|
||||
FileSystemSandboxKind::Restricted => {
|
||||
for entry in &mut file_system.entries {
|
||||
entry.access = match entry.access {
|
||||
FileSystemAccessMode::Read | FileSystemAccessMode::Write => {
|
||||
FileSystemAccessMode::Read
|
||||
}
|
||||
FileSystemAccessMode::Deny => FileSystemAccessMode::Deny,
|
||||
};
|
||||
}
|
||||
}
|
||||
FileSystemSandboxKind::Unrestricted => {
|
||||
file_system = FileSystemSandboxPolicy::read_only();
|
||||
}
|
||||
FileSystemSandboxKind::ExternalSandbox => return None,
|
||||
}
|
||||
Some(Self::from_runtime_permissions(
|
||||
&file_system,
|
||||
NetworkSandboxPolicy::Restricted,
|
||||
))
|
||||
}
|
||||
|
||||
/// Managed workspace-write filesystem access with restricted network
|
||||
/// access.
|
||||
///
|
||||
|
||||
Reference in New Issue
Block a user