codex2claude

Import any Codex session into Claude Code — quit, then claude --resume to continue.

把任意 Codex 会话导入 Claude Code——退出后 claude --resume 接着聊。

Codex imports from Claude; this is the missing reverse, as a one-command Claude Code plugin.

Codex 能从 Claude 导入，这条补上反方向——一个命令的 Claude Code 插件。

Install

Requires the transession engine for transcoding:

cargo install transession

Install the plugin (inside Claude Code):

/plugin marketplace add MisterBrookT/codex2claude
/plugin install codex2claude

Usage

Inside Claude Code:

/import-codex            # default: import ALL Codex sessions (the reverse of Codex's)
/import-codex all        # same, explicit
/import-codex <id>       # import a specific session id only

Each session is written into its matching project directory by its own cwd, with deduplication. Quit after importing, then run claude -r in any project — the picker lists the just-imported Codex sessions; pick one to continue. No manual transcoding commands.

Zero-token option: the slash command still passes through the model once per call. To burn zero tokens, run bash ~/.claude/bin/import-codex.sh directly in the terminal (or wrap it in an alias). Claude Code's custom slash commands must pass through one model round — this is a platform limit.

How it works

~/.codex/sessions/*.jsonl  ──transession──▶  ~/.claude/projects/<cwd>/<id>.jsonl  ──▶  claude -r <id>

• detect: by default grabs every Codex rollout; can also take a session id
• transcode: reuses transession's universal IR to interconvert between the two formats
• normalize: rewrites the output into a native-faithful transcript (fixes Codex content tags, image sources, picker title, missing native fields) so it loads and resumes without errors
• dedup: ~/.claude/codex-import-ledger.tsv tracks codex_sid → claude_sid so the same session isn't re-imported

Limitations

Converts messages / tool calls / timestamps / metadata; drops the reasoning payload and runtime cache — enough context to resume, not a token-perfect replica.

License

MIT