Merge PR #8533: fix(pi-natives): guard DeviceCheck token generation on GUI session (@roboomp)

This commit is contained in:
can1357
2026-08-14 14:11:51 +02:00
2 changed files with 44 additions and 0 deletions
+41
View File
@@ -99,6 +99,43 @@ mod platform {
static _NSConcreteStackBlock: *const c_void;
}
// `SessionGetInfo` reports the caller's login-session attributes; consulted
// to detect a GUI/graphic session before touching the GUI-only DeviceCheck
// daemon.
#[link(name = "Security", kind = "framework")]
unsafe extern "C" {
fn SessionGetInfo(session: u32, session_id: *mut u32, attributes: *mut u32) -> i32;
}
/// `callerSecuritySession` — query the session hosting the current process.
const CALLER_SECURITY_SESSION: u32 = u32::MAX;
/// `sessionHasGraphicAccess` attribute bit from `Security/AuthSession.h`.
const SESSION_HAS_GRAPHIC_ACCESS: u32 = 0x0010;
/// Whether the caller's security session has GUI/graphic access.
///
/// `DCDevice.isSupported` synchronously opens an XPC connection to the
/// per-user DeviceCheck metadata daemon, which exists only in an
/// interactive GUI login session. From a session without graphic access
/// (SSH, a launchd `LaunchDaemon`, a CI runner, a service account, a
/// sandbox) the connection setup hits `_xpc_api_misuse` and aborts the
/// whole process with `SIGTRAP` before the completion handler can run — so
/// there is no error to return, only a dead process. Gating on the
/// documented session attribute keeps the call out of that trapping path.
///
/// Returns `false` when graphic access is absent *or* the session cannot be
/// queried: degrading to "unsupported" merely drops the attestation header
/// (exactly as on every non-macOS host), whereas optimistically assuming
/// "supported" risks the process-killing trap.
fn session_has_graphic_access() -> bool {
let mut attributes: u32 = 0;
// SAFETY: `SessionGetInfo` writes the caller session's attribute bits
// through the out-pointer; the session-id slot is unused, so it is null.
let status =
unsafe { SessionGetInfo(CALLER_SECURITY_SESSION, ptr::null_mut(), &mut attributes) };
status == 0 && attributes & SESSION_HAS_GRAPHIC_ACCESS != 0
}
/// Outcome delivered once from the completion block to the waiting worker.
enum Completion {
Token(String),
@@ -289,6 +326,10 @@ mod platform {
error: None,
latency_ms: 0.0,
};
if !session_has_graphic_access() {
result.error = Some("DeviceCheck unavailable without a GUI login session".to_owned());
return result;
}
// SAFETY: `c"DCDevice"` is a valid null-terminated class name.
let class = unsafe { objc_getClass(c"DCDevice".as_ptr()) };
if class.is_null() {
+3
View File
@@ -26,6 +26,9 @@
### Fixed
- Fixed an issue where shell-internal background jobs (such as `yes >/dev/null &`) could survive a one-shot shell session and consume CPU indefinitely after the command returned.
### Fixed
- Fixed `deviceCheckGenerateToken` aborting the whole process with `SIGTRAP` when called from a macOS session without GUI/graphic access (SSH, a launchd `LaunchDaemon`, a CI runner, a service account, a sandbox), which made every `openai-codex/*` OAuth model unusable for such accounts. `-[DCDevice isSupported]` synchronously opens an XPC connection to the per-user DeviceCheck metadata daemon, which exists only in an interactive GUI login session; without one the connection setup hits `_xpc_api_misuse` and traps before any completion handler runs, so the promise never rejects. The binding now checks the caller's security session for the `sessionHasGraphicAccess` attribute first and resolves `{ supported: false, error: … }` instead of touching DeviceCheck when it is absent ([#8353](https://github.com/can1357/oh-my-pi/issues/8353)).
## [17.2.12] - 2026-08-08