When Does Codex Cloud Need a Persistent Host?
Use Codex Cloud for independent repository tasks with a reusable setup. Test Codex CLI on an always-on host when a task requires a specific device or mutable state, and size the host for backlog recovery.
Separate durable handoffs from machine-local state
A published Codex Cloud environment gives independent tasks a common starting setup, and those tasks can continue while your computer sleeps. To see whether that fits a real workflow, take a job that consumes a previous run’s output and give it a versioned input manifest containing the base revision, fixture checksum, and expected output path. Run the same manifest in two fresh tasks. Compare the fixture checksum and the normalized result, ignoring timestamps that have no bearing on correctness. Then remove the fixture from the manifest for a control run. If the control succeeds by reading stale data, the workflow still has a hidden dependency. If both fresh runs agree and the control fails clearly, the dependency can travel explicitly. This matters when a failed job must be replayed against last week’s input after today’s checkout has moved on.
Codex Cloud | ChatGPT LearnMeasure the queue that forms when the host disappears
Cursor’s Self-Hosted Machines illustrate a distinction worth testing on any persistent-host setup. Tool calls run on your hardware while the agent loop remains in Cursor’s cloud, and the selected machine must stay online during sessions. With Codex CLI on a dedicated host, use a harmless recurring fixture to measure the local executor and scheduler together. In a disposable test queue, suspend dispatch for a planned interval while the scheduler keeps recording due jobs. Resume dispatch and capture each job’s due time, first start, and accepted output ID. In an illustrative workload with one job every five minutes, a 30-minute outage creates six overdue jobs. Compare your clearance time with the deadline for the oldest job. A host can have ample average throughput and still miss that deadline after a single restart.
Choose where Cloud Agents run | Cursor DocsPublish the setup and test two tasks
Create a Codex Cloud environment with the repositories and tools your job needs. Configure any required network access in the environment settings and publish the setup. Start two cloud tasks against it. Ask both to run the same existing project check, and ask the first to create an uncommitted marker file in its workspace. After it does so, ask the second task to look for the marker.
Pass if both tasks can run the project check and the marker appears only in the first workspace. If both lack a dependency, inspect the shared setup and verify the environment is published before retrying. If the second sees the marker, check whether the first task committed it or wrote it to an external location that both tasks can reach.
Codex Cloud | ChatGPT LearnTest the dependency that drives the choice
In an illustrative example, two developers need to run independent documentation checks while their laptops are asleep. Both jobs need the same formatter and repository setup. A third job rewrites a generated index across several runs and expects the previous run’s uncommitted index to remain available. The first two jobs fit the published-environment model in Codex Cloud. The third calls for a persistent host if that index must stay in the workspace.
Run the index job twice against a disposable repository. Give the first run seven records and the second three more. Verify that the final index contains ten unique records and that retrying the second batch leaves the count at ten. If the index can be stored and retrieved explicitly between runs, test that design too. It may let the job use isolated workspaces.
Codex Cloud | ChatGPT LearnCheck what happens when the computer sleeps
If you are considering Codex Remote for a checkout or tool available only on a particular computer, confirm its current device, wake, and sign-in requirements in OpenAI’s product help and your account. Use a test repository and a harmless local marker. Try reading it from the second device while the host is awake, then repeat with the host asleep. Record whether the request completes, waits, or becomes unavailable.
Use Codex Cloud as a control with an equivalent repository task that has no dependency on the local marker. OpenAI documents that a cloud task can continue while your computer sleeps. If a cloud task fails because it needs something on that specific computer, decide whether you can provision the dependency in a published environment or need to keep the computer reachable.
Codex Cloud | ChatGPT Learn