Codex Coding Agent

Scope

Use the selected launcher's --version, --help, exec --help and model catalog (debug models; extract only model/effort fields) to choose supported model/effort settings. Honor explicit choices; keep wrappers opaque because their source and environment may contain credentials. Compare requested settings with observed evidence, leaving absent values unverified.

Fix the workspace, permitted actions and acceptance evidence. Experiments and parallel writers need isolated workspaces. Boundary: before file-only work, strict confinement or permission recovery, read Boundaries. The skill grants no extra scope.

Run

For ordinary bounded headless tasks, use codex_run.py: