Codex Upgrades Build Reliable AI Coding Workbench
OpenAI's Codex evolves from CLI tool to full workbench via desktop browser/computer use, CLI v0.122-0.125 reliability fixes, plugin ecosystems, enterprise permissions, Bedrock support, and GPT-5.5 as default model.
Desktop App Enables Visual Testing and Background Monitoring
Use Codex's in-app browser to preview local sites or public pages, provide feedback on renders, and have the agent fix issues automatically—this closes the loop on UI verification beyond file edits. On macOS, computer use lets Codex see, click, and type in native apps for GUI bugs, simulator flows, or settings without terminal commands. Start chats without project folders for research, planning, or analysis; set thread automations to resume on schedules with full context. Task sidebar offers context-aware suggestions and better PR workflows; artifact viewer handles PDFs, docs, spreadsheets. Multi-window/terminal support, Intel Mac/Windows tray, and memory aid long sessions.
Codex Pets provide a floating overlay showing active thread status (running, waiting, ready), progress prompts, and agent state while using other apps—toggle via /pet, settings, or command menu. Create custom pets with 'hatch pet' skill for project-inspired companions, solving oversight without reopening threads.
CLI Versions 0.122-0.125 Fix Workflows for Production Use
In v0.122.0, queue / commands or ! shell prompts during agent work to avoid rigidity; use /side for quick questions without derailing main threads (e.g., "What does this file do?"). Plan mode starts implementation in fresh context, previewing usage to avoid messy discussions bloating tokens. Plugins gain tabbed browsing, inline toggles, remote/local marketplaces—install 'hatch pet' skill and reload for custom pets.
Standalone installs self-contain; app command opens/installs reliably on Windows/Intel Macs. Tool discovery/image generation default-on improves UI debugging with high-detail handling.
v0.123.0 adds Amazon Bedrock provider (AWS profiles/SigV4); /mcp verbose for diagnostics/templates. v0.124.0 introduces Alt+, (lower reasoning) / Alt+. (raise) for quick terminal tweaks; multi-env app servers switch directories per turn. Hooks stabilize for MCP observation, patches, bash. v0.125.0 enhances app-server plumbing (Unix sockets, pagination, sticky envs), remote plugin installs/upgrades, consistent permissions across CLI/app/MCP/shell.
Fixes prevent stale approvals, stuck states, Unicode issues, ensuring reliable resumes/forks.
Permissions and Sandboxing Build Enterprise Trust
Deny-read glob policies, managed requirements, platform sandbox enforcement, and isolated exec runs ignore user configs—protect private keys, env files, client code. Trusted workspaces required for hooks/exec; automatic approval reviews route risky actions through reviewer agent, showing risk/status (approved/denied/timed out) for safer delegation.
Permission profiles sync across sessions, user turns, MCP sandbox, shell escalation—keeps CLI/app/server aligned on access.
GPT-5.5 and Integrations Unlock Broader Capabilities
GPT-5.5 recommends for implementation/refactors/debugging/testing/validation/artifacts (GPT-5.4 fallback during rollout)—update CLI/app/IDE to access. Browser use lets Codex operate in-app browser for clicking UIs, reproducing visual bugs. Bedrock expands beyond OpenAI models; multi-env/remotes suit AWS-heavy teams. ChatGPT plans default to fast tier, boosting value for heavy users.