Codex

Codex is OpenAI’s coding agent for the terminal, and the second engine in Cole’s fleet. Codex sessions run in tmux right next to Claude Code sessions, with the same skills, the same hooks and the same memory, but a different model family underneath. That difference is the point: Codex gives the independent second review, takes lanes in a parallel fan-out, and keeps working when Anthropic’s API is having a bad hour.

What it is

The CLI is a terminal UI (codex), a non-interactive runner for scripts (codex exec), and codex resume to reopen a recent session. It reads project instructions from AGENTS.md files, the way Claude Code reads CLAUDE.md, and its settings from ~/.codex/config.toml.

Two settings decide how much it may do on its own:

Sandbox	What it allows
read-only	Read files and run commands inside a read-only sandbox
workspace-write	Edit and run commands in the working directory; network off unless enabled; approval needed outside it
danger-full-access	No sandbox at all

Approval policies (on-request, never) decide when Codex stops to ask, and --dangerously-bypass-approvals-and-sandbox turns both off, the equivalent of Claude Code’s bypass mode. The model and its reasoning effort are per invocation: -m <model> and -c model_reasoning_effort="high", or /model inside a session.

Installing and signing in

npm install -g @openai/codex      # or: brew install --cask codex
codex                             # then choose "Sign in with ChatGPT"

Cole signs in with a ChatGPT Pro subscription (from $100 per month ($100, $200 or $500 tiers) (as of Oct 28, 2026, learn.chatgpt.com)), the top individual plan, because a fleet runs a lot of Codex; ChatGPT Plus ($20 per month (as of Oct 28, 2026, learn.chatgpt.com)) is the place to start. Codex can also run OpenAI models through Amazon Bedrock, which Cole uses for automated reviews on AWS.

How Cole uses it

Second reviews. Cole’s rule for any non-trivial change: two independent reviews of the same frozen commit before it ships, one on Codex and one on Claude, gated on severity. Every finding is checked against the real code before anyone fixes it, because a plausible finding is not a real one.

Lanes in a fan-out. The parallel worktrees launcher runs any lane on Codex, either for the whole run (--engine=codex) or per manifest row (codex:sol:xhigh picks a model alias and an effort). Each Codex window unsets any API key so it uses the ChatGPT login, and starts with approvals and sandbox bypassed, because an unattended lane in its own worktree needs git and gh with network access.

Subagents through tmux. The codex-subagents skill lets any agent fan a task out to Codex: one tmux window split into four panes, one interactive codex per pane (read-only, -s read-only, for reviewers), each briefing written to a file and pasted in as a tmux buffer, then a check that every pane really started. For a single bounded job, codex exec --output-last-message <file> runs headless and leaves its answer in a file.

Outage diagnosis. When a Claude turn dies on an API error that retrying will not fix, Cole’s recovery hook starts a diagnostic agent, and it picks Codex when Codex is installed: a different provider still works during an Anthropic outage.

From Drover. Drover runs one read-only codex app-server per machine and reads every thread through it, so Codex sessions show up on the same board as Claude’s, and a message to a Codex session in a terminal goes in through codex queue. A launch can start Codex at any reasoning effort the model supports.

What it shares with Claude Code

Codex is not a separate world on Cole’s machines. The Dotfiles give it the same pieces:

• Rules. A global AGENTS.md carries the same working style, todo-list habit, definition of done and safety rules as Claude Code’s CLAUDE.base.md, plus a section on reading instructions written for Claude: a skill’s knowledge (conventions, checklists, commands) applies as written, while Claude-only mechanics (the Skill and Agent tools, slash commands) are translated to what Codex has.
• Skills. The installer links every skill in ~/.claude/skills/ into ~/.codex/skills/, never touching Codex’s own bundled ones. Codex reads the same SKILL.md format, so one copy serves both. See Skills and Plugins.
• Hooks. ~/.codex/hooks.json points at the same scripts Claude Code runs: tmux window names, turn timestamps, the stop-too-early nudge, the background reflector before compaction and at session end, and a Codex mode of the context advisor. See Hooks.
• Memory. A SessionStart hook hands each Codex session the repo’s open items from the reflection store. A bridge copies Cole’s curated Claude memory into Codex’s memory folder as reference material (never as instructions), and copies Codex’s own memory summary back for Claude to see. Both agents also read and write the shared brain over MCP.

Reading what Codex did

The Codex TUI collapses tool output just like Claude Code’s does, so a pane is a poor witness. Every session writes a rollout file under ~/.codex/sessions/<year>/<month>/<day>/, and Codex keeps thread state in a local SQLite database. The codex-transcripts skill reads both: the real commands, outputs and reasoning, and the history a compacted session no longer shows. Cole’s rule is the same as for Claude: read the transcript, or ask the session; do not trust the pane.

Gotchas

• Multi-line prompts can submit early. Typing a long prompt into the TUI with tmux send-keys sends it at the first newline. Paste it as a buffer (tmux load-buffer, then paste-buffer -p), then press Enter once.
• An Enter after a paste can be lost. Check the pane for Codex’s working status line, and send one more Enter if the text is still sitting in the composer.
• workspace-write has no network. It is a good default by hand, but an unattended lane that runs gh stalls there.
• Hooks need trust. Codex skips a new or changed hook until someone reviews it in /hooks, so every machine needs one visit after a hook changes.
• Codex compacts on its own. Its window is smaller than a 1M Claude session’s, so the context advisor’s Codex mode only reflects and adds a one-line note; the PreCompact hook saves a reflection first.
• Keep subagent defaults on the command line. Pass model, effort and sandbox per invocation instead of editing ~/.codex/config.toml, so one run never changes the next.