LIVE|CLI v0.144.0·model GPT-5.6 Sol·verified 2026-07-09
Vol. I · No. 1Independent coverage of OpenAI CodexWednesday, July 15, 2026
Codex Insider
The unofficial wire for OpenAI Codex.
Codex Insider
Features

Codex plan mode

Published Jul 15, 2026 · Checked against the official docs

The short version

Type /plan in a Codex CLI session and it switches to plan mode: Codex reads your repo and writes out an execution plan before it touches any files. You review the plan, then tell it to go. You can pass the task inline, like /plan migrate this service to Postgres. It's the built-in version of the old 'write a plan.md first' trick.

The fastest path is one command. Open a Codex CLI session, type /plan, hit Enter. Codex flips into plan mode, reads your repo, and drafts the steps before it writes anything. You read the plan, tell it what to change or say go, and only then does it start implementing.

Plan mode is the fix for the most expensive kind of failure in an agent: a wrong first move that compounds over an hour of autonomous edits. You lock in the approach before Codex burns tokens acting on it.

Turn it on

  1. Start Codex in your project: run codex.
  2. In the composer, type /plan and press Enter. The official docs describe it as “Switch to plan mode and optionally send a prompt”.
  3. Optionally pass the task inline so it starts planning right away.
  4. Read what it proposes: the approach, the files it intends to touch, the order of operations. Reply with edits, or tell it to proceed. Per the docs, /plan asks Codex to “propose an execution plan before implementation work starts.”
codex
# then, in the composer:
/plan migrate the billing module off Stripe Charges to PaymentIntents

Plan mode vs. read-only

These get conflated. Read-only is an approval and sandbox policy. You set it with /permissions or the sandbox_mode config key, and it physically blocks writes. /plan is the drafting step, where Codex thinks the work through first.

The official docs describe plan mode as a mode switch, not a hard sandbox lock. From the community threads, the planning behavior leans on instructions to the model rather than a runtime block (Discussion #4760 notes “there’s no built-in way to prevent Codex from starting implementation”). So if you want a guarantee that nothing gets written during planning, pair /plan with a read-only sandbox instead of trusting the mode alone.

If it fails

  • /plan not recognized. It’s a recent built-in. Older installs don’t have it — before it shipped, people wired up a custom prompt at ~/.codex/prompts/plan.md (Discussion #4760). Update instead: npm install -g @openai/codex@latest (or brew install --cask codex on Mac). Latest stable was 0.144.4 as of 2026-07-14 per the releases page.
  • Shift+Tab to cycle modes. Third-party write-ups say Shift+Tab cycles the collaboration modes, but the official commands reference I checked doesn’t list that shortcut. Use /plan directly and you don’t have to guess.
  • Sandbox blocks a read you needed for planning. If a command dies asking to retry without the sandbox, that’s the read-only policy, not plan mode. See /errors/sandbox-error-retry-without-sandbox/.

Codex Insider Desk

Sources

More Features guides