Hit your Claude Code limit mid-task? How to switch to Codex without starting over

Developers report losing the thread when Claude Code or Codex hits its usage limit mid-task. Why the next session starts cold, the fixes that work today, and how to keep decisions across tools.

The issue, as reported

Reports on both the Claude Code and Codex trackers describe the same thing: the limit lands mid-task, and whatever picks the work up next, whether the same tool after the reset or the other tool you pay for, starts cold.

A Codex issue from September sets out the cost. After each usage-limit interruption, a large part of the new allowance goes on rereading repository files, rerunning checks and tests, and working out again which steps are finished. The reporter keeps handoff summaries by hand so Codex doesn't redo completed work, and asks for a checkpoint: completed steps, pending steps, modified files, the latest test results, blockers and the exact next action.
A Codex feature request from July, with 12 thumbs-up, asks for the session to pause at the limit and carry on at the reset without having to restart it or re-explain what it was doing.
The same request on the Claude Code tracker puts numbers on it: about 70 of 90 workflow tasks were done when the limit hit, and after resuming the progress showed 46, so roughly 24 finished tasks were thrown away and run again.
Claude Code now continues on its own after the reset, but a September report says that covers the main conversation only: subagents and workflows running at the time fail, and the continuation builds on the half-finished run.
Switching tools has its own version of the cold start. The Hacker News thread on Claude Code reading AGENTS.md (740 points, 285 comments) drew a run of replies from people glad to delete the symlinks and sync scripts they had kept so every tool got the same rules.

Reported in

Recovery after the limit eats the new allowance; the reporter keeps handoff summaries by hand and asks for a checkpoint Codex tracker · September 2026 · 4 thumbs-up, 5 comments
Pause at the limit and carry on at the reset, without restarting or re-explaining Codex tracker · July 2026 · 12 thumbs-up, 2 comments
About 70 of 90 tasks were done at the limit; after resuming, progress showed 46 Claude Code tracker · July 2026 · 2 reactions, 3 comments
Automatic continue picks up the main conversation only; subagents and workflows fail Claude Code tracker · September 2026 · 2 reactions
Claude Code now reads AGENTS.md: people deleting the symlinks and sync scripts they kept Hacker News · September 2026 · 740 points, 285 comments

Why it happens

Every session starts with an empty context window. Claude Code's memory docs say so directly: knowledge crosses sessions only through instruction files (CLAUDE.md or AGENTS.md) and the tool's own auto memory, which is machine-local and belongs to that one tool.
Nothing records the state of the work. A resumed session has the transcript, but no checkpoint of which steps are verified done, what the last test run said or what comes next. So the agent re-reads and re-runs to find out, which is exactly the gap the Codex issue asks the tool to close.
Each tool keeps its own history. Claude Code and Codex save sessions in their own folders and formats. Neither reads the other's conversations unless you import them.
The instruction files didn't match. Codex and OpenCode read AGENTS.md. Claude Code read only CLAUDE.md until version 2.1.277, released in September 2026.
The limit is shared. On Pro and Max plans, Claude and Claude Code draw on the same usage limits, so a busy day in the chat app can end a coding session sooner.

Before the limit the context holds the rules, the steps done, the last test result and the decisions; the next session gets the rules back and empty slots for the rest — The instruction files are re-read at the start of every session. The state of the work is not recorded anywhere, so the next session has to find it out again.

How to fix it today

None of these needs an extra tool.

Put your rules in one AGENTS.md that all three tools read
Keep build commands, conventions and what not to touch in AGENTS.md at the repository root. Codex reads ~/.codex/AGENTS.md, then every AGENTS.md from the Git root down to your working folder, up to 32 KiB by default (Codex docs). OpenCode reads AGENTS.md, and falls back to CLAUDE.md (OpenCode docs). Claude Code (2.1.277 or later) reads AGENTS.md only when there's no CLAUDE.md, .claude/CLAUDE.md or CLAUDE.local.md in your folder or above it. Check it loaded with /memory in Claude Code, or in Codex with codex --ask-for-approval never "Summarize the current instructions.". Claude Code doesn't read anything under .agents/, so shared skills still need their own arrangement. If you keep both files, set Project instructions to claude-md-and-agents-md in /config, or start your CLAUDE.md with an import:
```
@AGENTS.md

## Claude Code only
Use plan mode for changes under src/billing/.
```
Keep a checkpoint file, and update it at each milestone
Don't wait for the limit warning; it can land mid-reply. Ask the agent to maintain a file both tools can read, using the fields the Codex issue asked for. After the reset, or in the other tool, start with: "Read CHECKPOINT.md first. Continue from the next action and don't redo anything marked done."
```
Keep CHECKPOINT.md in the repo root up to date after each step:
- Goal of this task, in one line
- Steps done (and how each was verified)
- Steps left, in order
- Files changed so far
- Last test or build result
- Decisions made, each with its reason
- Blockers
- The exact next action
```
In Claude Code, let it wait and continue (2.1.234 or later)
For interactive sessions signed in with a claude.ai subscription, automatic continue is on by default: when the limit stops a task, Claude Code waits in the open session and sends a prompt to pick the task up at the reset. Keep the session open. If the menu appears instead, run /rate-limit-options and pick Wait here, then continue automatically. The switch lives in /config as Continue automatically at usage limit.
Claude Code to Codex: use Codex's /import
In the Codex CLI, /import (slash commands) brings in Claude Code setup, projects and chats, finding up to 50 chats from the last 30 days. It arrived in Codex CLI 0.145.0 (21 July 2026), and 0.147.0 keeps imported Claude conversations in sync without duplicates. Point Codex at your checkpoint file as well.
Codex to Claude Code: bring the configuration, carry the rest in the checkpoint
Claude Code's /import codex (add --dry-run to preview) brings Codex configuration: instruction files, MCP servers, commands, subagents and skills. It doesn't bring conversations (command reference), so the checkpoint file does that job. Going the other way to any tool, /export writes a Claude Code conversation out as plain text.

What the fix doesn't cover

A checkpoint is only as good as its last update

If the limit lands between updates, the newest work isn't in it. And automatic continue is narrow: per the docs, it ends if you exit, /resume another session or clear the conversation; it isn't offered in background sessions or -p runs; and it doesn't start on its own for a reset more than 24 hours away. The report above says running subagents and workflows don't come back with it.

What reaches the other tool

RulesBuild commands, conventions and what not to touch, in AGENTS.md, read by all three tools.

ChatsCodex's /import brings in Claude Code chats from the last 30 days. Nothing built in brings Codex chats into Claude Code.

SummariesA checkpoint keeps only what someone thought to write down.

The reasonsWhat was tried, why it failed, which decision replaced which: that stays in the transcript of the tool where it happened.

How Axiom Lift solves it

Axiom Lift is a small helper on your computer that reads the conversations Claude Code, Codex and OpenCode already save, automatically, as you work, so nothing depends on writing a note before the limit hits.

The free fix

Rules move in AGENTS.md; the reasoning stays in the transcript where it happened
A checkpoint keeps only what someone thought to write down, as of its last update
Codex's /import brings Claude Code chats one way, from the last 30 days
Automatic continue waits in the open session, for the main conversation only

With Axiom

Picks out the requirements, decisions, corrections and results as you work, and keeps them in one history
The next session, in the same tool after the reset or in one of the others, asks Axiom over MCP and gets back the few lines that are true now, not the whole transcript
Before each prompt, adds a few lines about what's already decided on that subject
If you changed your mind along the way, the old decision is struck through, not erased, and linked to the moment it was said

What Axiom keeps from the Claude Code session and what it hands the next session in Codex — Axiom keeps the decisions, corrections and results from the session before the limit, and hands the next session, in any of the three tools, the few lines that are true now.

What it doesn't do here:

It doesn't give you more usage, and it doesn't resume a paused run or restart subagents that the limit stopped.
It doesn't move a chat into the other tool's resume list; Codex's /import does that one way.
It covers Claude Code, Codex and OpenCode only. Not Cursor, and not ChatGPT or Claude in a browser.
It doesn't replace AGENTS.md. Keep your house rules there.

It installs with one command in a desktop terminal, on Mac, with Linux and Windows in beta (it needs Python 3), and shows you what it found before anything is sent. The free plan covers one computer.

Switch tools without starting over

One command in the terminal where you use Claude Code, Codex or OpenCode. Free, no card.

Try it free →

Claude, Codex and OpenCode are trademarks of their owners; Axiom works alongside them and is not affiliated.