ADMINCodex execution modes#
Read-only is the default. The new per-user Full Access preference requires a YAADMIN Web administrator. Core checks the current administrator role when applying the execution policy, and changing the mode interrupts active Codex execution before the preference is saved.
Full Access relaxes the Codex local execution sandbox. YAADMIN-registered tools still execute through their existing RBAC, policy, approval and audit checks. This is not a new Agent desktop permission and does not remove those tool checks.
Context usage and responsive controls#
The Codex HUD uses the latest thread token usage and model context window rather than substituting cumulative session totals. User, authentication mode, thread and conversation scoping protect against stale responses. Unknown telemetry is not shown as a fabricated value.
Reasoning and Codex popovers use viewport-aware positioning, scrolling and keyboard dismissal. Physical mobile keyboards and Safari still require acceptance testing.
Follow-up messages during execution#
Text instructions can be accepted while a chat turn is running, with stable retry IDs and an ordered, bounded persistent queue. Codex steers the current turn where acknowledged; standard providers include pending instructions at a request boundary or continue sequentially.
The interface has an adaptive Send/Stop control. Unknown delivery, unfinished messages after restart and pending approvals require review; they are not silently replayed. Follow-ups are text-only in this implementation.
Embedded Codex console#
The console displays output from the same App Server client and turn that produces the answer. It opens manually, supports ANSI colors, copying, follow-tail scrolling and a full-screen view with Escape. Owner-scoped, bounded history supports reconnects. Other providers retain the normal Live Activity presentation and fallback.
The console is output-only. It does not provide interactive shell input or the complete native CLI TUI/PTY. Output filtering is pattern-based and cannot guarantee recognition of every unlabelled confidential value.
Images and extensionless storage#
Codex uploads are validated against ownership, original extension, MIME, size, hash and decoded bytes. Images stored under an extensionless filename can be delivered as local image input. Generic inline images from authorized tools or MCP results can be delivered as image input alongside text, without putting binary image data into ordinary console output.
No screenshot tool was added by this change. Tool-provided remote URLs are not fetched. This source update does not establish the full-screen Agent reconstruction requirement as complete.
The Excel parser also opens extensionless storage through a binary stream and closes the workbook and stream afterward.
Verification and remaining acceptance#
The repository handoff reports earlier focused development results: 59 Codex/access/context/security tests, 8 follow-up tests, 5 console cases and 5 image cases. Those tests were not rerun during this website documentation update.
The image fixture exercised real image-byte transport through an installed App Server and a localhost provider. It did not use an authenticated live model or establish visual recognition. Earlier browser fixtures do not establish physical mobile/Safari acceptance. No full application suite or Windows infrastructure acceptance is claimed.
Branches remain distinct#
Today's main commit and Agent checkpoint 41ffadf47bf8 diverge from a common baseline. Today's Codex/chat facts do not replace the separately audited Agent 0.3.0 source, its acceptance limitations or the accepted full-screen Vision architecture.