codex-cli

Files (1)

• SKILL.md

SKILL.md (first lines)

# Codex CLI

Codex runs non-interactively through `codex exec`. A new run starts a separate Codex session in the selected working directory, streams progress to stderr, and prints its final response to stdout. Redirect stderr to `/tmp/codex.err` so the turn log stays out of the caller context. A run is done when the process exits; stdout is the reply. If the command fails, read `/tmp/codex.err`.

## Run a task

Use `gpt-6.1-sol` with `high` reasoning effort by default:

```bash
codex -a never \
  -C <working-directory> \
  -m gpt-6.1-sol \
  -c model_reasoning_effort=high \
  exec --skip-git-repo-check "<request>" \
  </dev/null 2>/tmp/codex.err
```

`-a never` is a global flag and therefore appears before `exec`. It keeps an unattended run from waiting for interactive approval; failed operations return to Codex instead.

`--skip-git-repo-check` lifts Codex's trusted-directory check; without it an unattended run fails before doing any work.

`</dev/null>` keeps piped stdin from being appended to the prompt. For a long request, write it to a file and pass `- < file` in place of the quoted prompt and `</dev/null` — inline quoting breaks on quotes and newlines.

### Model and reasoning

Available models, in ascending capability and cost: `gpt-6-luna`, `gpt-6.1-sol`. Set the model explicitly with `-m` on every run.

Reasoning effort accepts `medium`, `high`, `xhigh`, or `max` (not `low`). Higher effort gives the model more room to deliberate and generally increases latency. Set it explicitly with `-c model_reasoning_effort=<level>`.

Use `gpt-6.1-sol` with `high` effort for most delegated tasks. Use `gpt-6-luna` with `xhigh` or `max` for simple but tedious tasks to improve reliability at low cost.

## Continue a session

`resume --last` continues the most recent session in that working directory. A thread ID continues a specific one. The thread id is on the `session id:` line in `/tmp/codex.err`. Keep model and effort unless the caller picks different values.

```bash
codex -a never \
  -C <working-directory> \
  -m gpt-6.1-sol \
  -c model_reasoning_effort=high \
  exec resume --skip-git-repo-check --last "<follow-up>" \
  </dev/null 2>/tmp/codex.err
```

```bash
codex -a never \
  -C <working-directory> \
  -m gpt-6.1-sol \
  -c model_reasoning_effort=high \
  exec resume --skip-git-repo-check <thread-id> "<follow-up>" \
  </dev/null 2>/tmp/codex.err
```

## Review code

Pass either one scope flag or a prompt. `--help` lists both; the CLI still rejects that combination (exit 2).

```bash
codex -a never \
  -C <working-directory> \
  -m gpt-6.1-sol \
  -c model_reasoning_effort=high \
  exec review --skip-git-repo-check --uncommitted \
  </dev/null 2>/tmp/codex.err
```

Other shapes: `review --base <branch>`, `review --commit <sha>`, `review "<instructions>"`.

## Command reference

Use `codex exec --help`, `codex exec resume --help`, and `codex exec review --help` for the command surface installed on the machine.

From legibet/agent-skills, MIT.

More from legibet/agent-skills