Codex Attention

Know when Codex needs you.

By Kohei Elimi Lab · 日本語 · More add-ons

A Windows VS Code extension that flashes taskbar buttons and brings the requesting editor window forward when the OpenAI Codex extension asks a question or requests approval.

• Questions and command/file/permission approvals.
• Multiple windows matching the same workspace name? All matching windows flash. Only the requesting window is focused.
• Answering one overlapping request does not clear another window's pending attention.
• Separate switches for flashing, focus, approvals and opening the Codex sidebar.
• No telemetry, no API key, no automatic answers.

Install

Download the VSIX from Releases. In VS Code, choose Extensions → … → Install from VSIX, then run Developer: Reload Window in every window you want monitored. This is a VS Code extension (also called an add-on), not a Codex skill or a standalone app.

Requires Windows desktop VS Code 1.96+ and the official OpenAI Codex extension running a native Windows codex.exe app-server in the same extension host. Reload once after installation to load the updated extension. Both already-running and newly started app-server children are attached, so Codex may activate first. Requests emitted before attachment cannot be replayed. Tested locally against Codex extension 26.1007.21434 and VS Code 1.142.0.

Use

The status bar explicitly shows monitoring, disconnected, disabled or pending state. The Codex Attention output channel records each attached server and whether it was already running. Run Codex Attention: 点滅・前面化をテスト from the command palette for a manual flash/focus test. Settings are under codexAttention.*.

Limits and privacy

Codex only. Other assistants, terminals, the Codex desktop app, WSL/SSH/remote app-servers, macOS, Linux and browser VS Code are not supported.

Uses an observer on the shared Node ChildProcess spawn implementation and discovers live children through the extension host's active handles for native Codex app-server JSON-RPC. This also catches callers that cached their spawn function before Attention loaded. It leaves official extension files and transmitted bytes unchanged and never submits an answer. Active-handle discovery and protocol monitoring are private integration points rather than a supported OpenAI event API; future Node, VS Code or Codex changes can break detection.

Question bodies, shell commands and credentials are not stored or transmitted by this extension. Only request metadata is retained in memory. Window handles/process IDs and heartbeats are shared through temporary per-user files to coordinate overlapping requests. Those are local coordination data, not telemetry.

Windows can refuse foreground activation. Local testing confirmed activation, but repeated foreground checks were not consistently successful under concurrent desktop activity. Flashing is the fallback; activation is best effort. Bring each newly background-launched VS Code window forward once to bind its native window. Flashing also follows Windows taskbar grouping behavior.

Disable or uninstall, then reload VS Code to remove the observer. This independent extension is not affiliated with or endorsed by OpenAI or Microsoft.

Build and test

npm test runs protocol tests. node --test test/duplicates.test.js runs a Windows-only native two-window test. Compile native/WindowAttention.cs with the .NET Framework C# compiler and System.Web.Extensions.dll; source and compiled helper are included. python tools/package_vsix.py builds a dependency-free VSIX. The packaging allowlist excludes local integration logs and fixture executables.

MIT · © 2026 Kohei Elimi Lab