codexctllisted

# codexctl One CLI. Agent commands print JSON on stdout (exit 0 on a completed turn, 2 otherwise); plain calls print the model's text on stdout and a stats line on stderr: ``` node ${CLAUDE_SKILL_DIR}/scripts/codexctl.js <command> [options] "prompt" ``` Run with no arguments for full usage. Working directory defaults to the current one; pass `--cwd`. ## Mental model - **Thread** = persistent conversation with a model, cwd, sandbox, approval policy. **Turn** = one prompt→completion cycle producing **items** (agentMessage, commandExecution, fileChange, reasoning). Threads persist on disk under `~/.codex/sessions` and can be resumed, forked, reverted. - The app server asks the client to approve commands (**server→client requests**). codexctl auto-answers with `--decide` (default `accept`). Safety comes from the **sandbox**, not from the approval answer: `workspace-write` (default, writes only inside cwd, no network), `read-only`, `danger-full-access`. - Each invocation normally spawns a private app-server. Run `codexctl serve` once to get a shared WebSocket server; then invocations attach to it, and `steer`/`interrupt` can reach turns started by earlier invocations. - **Two kinds of call.** *Agent* commands (`run`, `resume`, `fork`, `review`) use Codex's own agent prompt and tools in a workspace. *Plain* commands (`ask`, `compare`, `map`) replace that prompt with ours (`assets/plain.md`, or `--system`/`--instructions`), strip skills, plugins and connectors, run read-only in