← Back to the briefs
Tools

OpenAI Codex 0.169.0 enforces directory trust and isolates plan execution

OpenAI Codex 0.169.0 keeps project configuration and MCP servers disabled in untrusted directories during non-interactive runs, isolates plan handoffs into fresh sessions, and adds tabbed terminal workspaces.

Execution boundaries and session handoffs

OpenAI Codex 0.169.0 tightens non-interactive tool execution by preventing untrusted directories from running project configuration files or Model Context Protocol servers. Under the updated behavior, running codex exec with workspace-write sandbox flags will still leave local project configurations and MCP servers inactive unless an operator passes an explicit trust flag.

The release also improves how planning flows hand work off to implementation sessions. Selecting the context-clearing implementation option transfers the entire approved plan directly into a clean session, while keeping the parent conversation intact if that transfer encounters an error. Turn control logic also isolates cancellations, ensuring that late aborts or abandoned file approvals do not spill into following turns.

Terminal workspaces and gateway setups

For interactive agent management, Codex adds tabbed side chats inside the optional terminal workspace. Operators can cycle across multiple conversations using keyboard shortcuts or mouse clicks and resume local side sessions after application restarts. The built-in agent command center also gains task search across prompt previews, task titles, and working directories.

Engineers deploying Codex across remote infrastructure can now register custom model providers directly through terminal sign-in flows. This includes device-code authentication for supported gateways over SSH connections when local web browsers are unavailable. Platform fixes address Windows sandbox issues with quoted shell paths and working directories located under restricted parent folders.

Our assessment

Automated pipelines that run Codex in CI or headless environments should verify whether their scripts depend on repository-level configurations or MCP servers. If automated runs touch untrusted checkouts, engineers must either supply trust flags deliberately or keep configurations disabled to prevent unreviewed tool loading. Development teams testing complex, multi-stage tasks should also verify whether fresh session handoffs preserve the full required task context without carrying forward memory debt.

Source: OpenAI original publication.

Original source

Read the original material

OpenAI (opens in a new tab)

The article date belongs to Agentic Horizon. The source publication date is listed separately. Vendor claims remain attributed; we have not independently tested the reported capability.

Search the publication

Find a story by title, topic, or keyword.

Press Escape to close

Ask HorizonYour site guide
Your guide to the publication

Ask Horizon.

Ask a question about the publication.

Privacy