Codex computer use

Operate the desktop through a local Codex CLI process using OpenAI's Computer Use runtime. Use this route whenever this skill applies and the host's tool-routing instructions permit it. An explicit request for Codex computer use selects this route, including for Chrome. Ordinary file edits and shell work stay with their own tools.

Launch

Write a narrow brief to a file: target app/site and account, requested outcome, permitted changes, and an observable completion check. Carry forward the user's authorization and limits. Say whether the task is inspection, navigation, or editing. Specify screenshot paths only when useful.
Check for another desktop-control task before starting. The desktop is shared across T3 conversations. If another task holds it, wait or coordinate when authorized. A subprocess running in the background can still bring apps forward and affect focus.
From this skill's directory, launch the bundled runner through the host's terminal tool:

python3 scripts/run.py \
  --brief /absolute/path/brief.txt \
  --output-dir /absolute/path/new-run-directory

Use a fresh output directory per attempt. The runner uses ~/.codex, rather than T3's inherited CODEX_HOME proxy directory, so the local computer-use plugins and normal Codex authentication are available. --codex-home selects another deliberately configured local installation. It removes ELECTRON_RUN_AS_NODE only in the child environment.

The default shell sandbox is workspace-write. Use --sandbox read-only under a read-only host policy. --sandbox danger-full-access is available only when the current host policy and authorized task permit it; never widen permissions to evade a refusal. The runner does not use approval-bypass flags, change login, copy credentials, or rewrite Codex configuration. It retains the local model selection; pass --model only when the user requested one.