You signed in with another tab or window. Reload to refresh your session.You signed out in another tab or window. Reload to refresh your session.You switched accounts on another tab or window. Reload to refresh your session.Dismiss alert
{{ message }}
Repository navigation
Non-jailed brokered Codex legs may overstate seat_filesystem_unconfined after agent-harness#1916 #1930
Follow-up from the independent re-check of agent-harness#1916 (non-blocking, pre-existing).
Non-jailed Codex legs still receive seat_filesystem_unconfined (panel_invoker.py ~16597) and the plan-time mode (~2247), because codex is not in _SANDBOX_INCAPABLE_BROKERED_LEGS (~6378). After agent-harness#1916 the retained brokered non-jailed Codex route (_brokered_codex_command) is sealed: read-only, shell_tool disabled, --cd set to the output dir. It is reached with a staged tree only when route resolution isn't eligible (an injected seam or a research seat).
To do:
Decide whether the notice is accurate for that sealed route.
If it isn't, add codex to _SANDBOX_INCAPABLE_BROKERED_LEGS or emit a narrower notice.
Related, from the 0.7.31 advisor-board review (agent-harness#1939, board comment 6733612917): the board confirmed that a non-jailed Codex leg can still be counted as sandbox-capable in capability accounting. Fail-closed for containment; accounting only. Targeting 0.7.32 alongside the seat-prompt alignment follow-up.
Activity
ViperJuice commentedon Nov 9, 2026
Related, from the 0.7.31 advisor-board review (agent-harness#1939, board comment 6733612917): the board confirmed that a non-jailed Codex leg can still be counted as sandbox-capable in capability accounting. Fail-closed for containment; accounting only. Targeting 0.7.32 alongside the seat-prompt alignment follow-up.