Keep coding when Claude Code or Codex cuts you off: hand the task to the other tool

Paste this into the tool you are leaving. If it has already run out, use it as the first message when the window comes back, or skip to step 2 and let the other tool read the work as it stands.

Stop here. Before you stop, write HANDOFF.md in the project root for another coding agent that has not seen this conversation. Include the goal, what is done, what is left, any decisions made and why, the files you touched, and the exact command to check the work. Keep it under 40 lines.

Open Claude Code or Codex in the same folder and paste:

Read HANDOFF.md and continue the work from where it stops. Check it with the command it gives.

When your first tool's window comes back, the same two steps hand the work back the other way.

Two small tasks, each a Node.js module with three functions and tests, run from scratch folders on one machine:

• Claude Code to Codex. Claude Code wrote the first function with 8 tests, then the note: the goal, what was left, why money was held as whole cents, which inputs it rejected, and npm test as the check. Codex read only the note, wrote the other two functions with their tests, and all 14 tests passed. Codex used about 12,000 tokens to finish.
• Codex to Claude Code. Codex wrote the first function with its tests, using about 18,500 tokens, and a note listing its decisions, such as removing accents. Claude Code read only the note, finished the other two functions, and all 16 tests passed, in seven turns and about half a minute. It then updated the note itself to say the work was done.

What made the notes work was the "decisions and why" part: the second tool kept the choices the note explained, such as holding money as whole cents. Where a note left a question open, the second tool made its own call. Claude Code's note asked whether negative amounts should be allowed when formatting, and Codex decided to allow them. So read the second tool's summary before you carry on.

On subscriptions, handing over costs only allowance in the second tool, which you already pay for. In money, the first step of our test, writing one function with 8 tests on Opus 5.5, came to about 27 cents at Anthropic's API prices, which is what it would cost if you paid by the token:

Claude Code's own figure for the same run was $0.27. The reason to hand over with a note rather than drag a long session along is the size of what gets re-sent. On the machine we tested, the largest requests of the week each re-read close to a million tokens of history. At $0.20 per million for a cache read, that is about 19 cents a turn; if the cache has expired and the history is written again at the five-minute rate of $5 per million, it is about $4.85 for one turn. A fresh session with a forty-line note starts near zero.

• Hand over between steps, not in the middle of one. A half-written file is harder to describe than a finished one.
• Commit, or at least note, what is done before you switch, so you can see exactly what the second tool changed.
• The other tool has its own limits. Codex on ChatGPT Plus has its own 5-hour window, and what you can do now has both tools' reset rules.
• Codex reads project instructions from AGENTS.md. Claude Code reads CLAUDE.md, and can also read AGENTS.md. If your instructions live only in CLAUDE.md, Codex will not see them, so mention them in the note or keep them in AGENTS.md.
• Some people keep several accounts with the same vendor and move between them instead. That may break the vendor's terms of service, and this page does not explain how.

Once you work across both tools, neither vendor's screen shows the whole picture. OpenReason Intel, a dashboard for Claude Code and Codex usage, keeps both tools' 5-hour and weekly windows side by side over time, from every machine you use, with each turn's estimated share of the window.

It shows you what happened; it does not hand work over for you, it does not stop a session as you work, and it cannot change either vendor's limits.