Codex 0.140 Makes Claude Code Migration a First-Class Workflow, Not a Copy-Paste Ritual
OpenAI shipped Codex CLI 0.140.0 last week, and buried in the changelog is a feature that deserves more attention than it will probably get: a first-party /import command that pulls Claude Code setup into Codex. The release notes call it a migration workflow. It is actually the beginning of a coding-agent ecosystem war fought on the terrain of accumulated configuration — not model capability.
The mechanics are straightforward. Codex can now import instruction files, settings.json, skills, plugins, project folders, the last 30 days of chat sessions, MCP server configuration, hooks, slash commands, and subagents from an existing Claude Code installation. The import does not delete or overwrite what is already there. It is additive: the user reviews each imported component and decides what to keep. That review step is the entire point.
Because here is what the release notes do not say explicitly: a serious Claude Code environment is not one prompt. It is a whole operating system. It is custom slash commands that encode your team's code review ritual. It is instruction files that define what "done" means for your repos. It is skills that know how to call your internal APIs, handle your monorepo structure, or speak to your MCP servers. It is plugin credentials, auth tokens, hook scripts, project conventions, and sometimes subagent routing that dispatches work to specialized child contexts. That is not a setup. That is engineering capital.
What OpenAI is recognizing is that switching costs for coding agents are no longer about whether the new model can edit your code. They are about whether the new runtime can accommodate the operating system you built around the old one. If migration means rebuilding all of that by hand, the incumbent agent gets switching-cost lock-in even when the competitor becomes technically superior. If migration becomes a guided import with status cards and review prompts, switching becomes a test plan — not a rewrite.
The Import Is Not the End of the Story; It Is the Beginning of the Review
The docs are notably honest about the dangerous part: imported setup is not automatically trusted setup. MCP servers may use custom authentication, headers, environment variables, or transport configurations. Hooks may behave differently in Codex than they did in Claude Code. Prompt templates that rely on shell interpolation or positional arguments can silently change meaning when the execution context shifts. Skills and agents may encode tool restrictions that do not map one-to-one between runtimes. The import surfaces all of this — but it is on the user to actually review it.
This is where the feature will succeed or fail in practice. "Click import and ship" will produce broken environments. "Import, inspect, run parity tests, and remove the copy-paste phase where humans miss half the safety policy" will produce credible migrations. The difference is entirely in whether teams treat the review step as real work or theater.
The release also adds /usage views that show daily, weekly, and cumulative token activity — a feature that belongs in the same story. Once agents can import whole workspaces, run subagents, call MCP tools, and preserve recent chats, the cost surface expands dramatically. A team that has not thought about token budgets suddenly has an observability problem: who is burning how many tokens on what, before the monthly bill arrives. Usage visibility is not a billing nice-to-have for agentic workflows. It is the observability layer for autonomous development work. Without it, you cannot set sane budgets, debug runaway loops, or attribute costs to the right team or project.
What Practitioners Should Actually Do With This
If your team is considering Codex as a Claude Code alternative — or if you are simply evaluating whether the coding-agent market is competitive enough to matter — treat this release as a forcing function for a migration checklist you probably do not have yet.
Before importing anything: snapshot the old agent config. Export, copy, or document the current state of instruction files, skills, plugins, MCP servers, hook scripts, and slash commands. This is not paranoia. It is standard operational practice for any system where configuration has ever mattered, and your Claude Code environment is such a system now.
After importing: run five specific tests. One read-only repo inspection — ask the agent to describe the architecture of a service it has not seen. One edit-and-test task — make a small change and verify tests pass. One MCP-backed task if you use MCP servers — verify the imported MCP configuration connects and responds correctly. One slash-command conversion — test whether a custom command you relied on in Claude Code maps cleanly to Codex or needs adjustment. And if you use subagents, test one subagent dispatch path: verify that child context inherits the right memory, tools, and permissions.
If those five tests pass, Codex migration is credible. If they fail, you have a precise bug report — a specific component that did not survive the import — instead of a vague sense that "the new agent feels wrong." That precision is the actual value of the import feature. It surfaces exactly what does and does not transfer, which is the information you need to make a rational decision about whether the migration is worth finishing.
The Deeper Signal
Codex adding Claude Code import is not a courtesy feature. It is a strategic move in a market where agent switching is becoming real. Developers are no longer evaluating coding agents as isolated chat boxes. They are comparing them as workspaces with config, memory, tools, auth, and team habits attached. The team that wins the next agent war will not be the one with the best model benchmark. It will be the one that makes it cheapest for teams to bring their existing workflows — their skills, their conventions, their safety rails — to the new runtime.
The next agent war, in other words, is not about model IQ. It is about workspace portability. The runtime that treats accumulated agent configuration as transferable engineering capital — not as lock-in ammunition — will be the one that earns long-term adoption from teams that actually ship.
Sources: OpenAI Codex changelog, GitHub release rust-v0.140.0, Import to Codex documentation