Cloudflare Codex: two paths for commerce developers

Cloudflare’s experimental Codex approaches have different recovery limits. Check saved work and repeatable actions before connecting live commerce operations.

SSophie Harrington2 min read

Cloudflare has published an experimental way to run the Codex agent loop inside a Durable Object, giving coding-agent work persistent state across interruptions. For teams building commerce apps, the useful question is which work survives a restart—and whether retrying a task could repeat a live-store change.

The 9 November announcement describes a WebAssembly adaptation of Codex’s Rust turn-processing code. Cloudflare promoted it again on 11 November. This is developer infrastructure, not a ready-made Shopify automation app.

Two approaches, different recovery limits

In the announced example, the agent loop runs inside the Durable Object. Cloudflare’s Tasks capability records model and tool work: completed steps can reuse their saved results after a restart, while interrupted steps may execute again. Other capabilities retain conversations, replay events and provide persistent files.

The current Codex harness guide, updated on 10 November, describes another approach: the Codex CLI runs in a Linux container, coordinated by a Durable Object. That preserves the CLI’s own tools and configuration, but its recovery behaviour differs.

For the container approach, the guide says the filesystem is snapshotted after five idle minutes. A container failure during a turn can leave it unanswered; retrying is an option when effects are safe to repeat. File changes made after the last snapshot can be lost. Cloudflare labels the interface experimental and warns that it may change.

Cloudflare’s original post includes the WebAssembly example’s code graphic. View the announcement on X.

What commerce developers should test

A sensible first trial is a development task with a reviewable output: analyse an app repository, propose a fix or generate tests. Interrupt it deliberately and check what was saved, what ran again and whether the result can still be reviewed. Treat those as acceptance tests for your application, rather than assuming persistence makes every action safe.

For a future workflow that changes products, refunds or fulfilment records, deduplicate the business action itself. Replaying a model step and preventing a second refund are different requirements. Keep deployment and store-write approval separate from generating a proposed change.

The container guide also requires authentication on routes reaching the agent and warns that prompts can execute commands and consume model credit. Its model-credential mechanism keeps those credentials outside the container; that does not make arbitrary secrets placed in its filesystem safe. Budget for the services and model usage your implementation actually uses rather than treating this example as a free hosted service.

Before choosing either design, confirm the current example, recovery guarantees and tools against your workload. The announcement establishes a new experimentation path; it does not establish production readiness for an unattended commerce operation.

ℹ️
Cloudflare’s experimental Codex approaches have different recovery limits. Check saved work and repeatable actions before connecting live commerce operations.
Filed underNews

More from News

See all →