CODEX EXECUTION ENVIRONMENTS · REVIEWED OCT 20, 2026
Codex works in Cloud but fails in CLI: what to compare
Start by identifying where the failing command runs. A working Cloud task and a failing terminal session can involve different files, installed tools, access and saved state. Compare one small failing operation across those environments before concluding that Codex itself is unavailable.
The Cloud overview describes remote coding tasks with separate workspaces. The CLI documentation describes terminal-based work and Cloud handoff. A request started in a terminal can therefore be Cloud work: the place you type is not sufficient evidence of the place commands execute.
Compare the evidence in this order
| Check | Record in each environment | Useful next step |
|---|---|---|
| Execution location | Local working directory, or Cloud task and environment name | Confirm both tests are actually running where expected. |
| Repository state | Commit, branch and relevant uncommitted files | Compare equivalent input; do not overwrite work to make it match. |
| Toolchain | Required runtime and package versions, exact failing command | Separate a missing executable from a failing application test. |
| Service access | Destination hostname and a redacted connection error | Distinguish name resolution, connection, authentication and authorization. |
| Account and service | Selected account, Usage message and relevant incident time | Follow the quota or incident route only when its evidence matches. |
For example, “tests fail in Cloud” is too broad. “The same test command cannot find the package manager in this task” points to setup. A command that reaches a registry but receives an authorization error calls for a different check. Keep the original output and change one condition at a time, so a later success has an explainable cause.
Check which Cloud setup the task used
OpenAI’s Cloud environments guide distinguishes published setup from an existing task’s saved state. New tasks start from the published filesystem; existing tasks keep their own state. Editing a reusable setup does not prove an already-running task received the change.
For current Cloud service connections, review the required hostnames and credentials separately. The guide directs users to test setup, publish or republish, and verify in a new task. An allowed destination does not grant permission inside that service. Use the specific host needed for the failing operation rather than treating unrestricted access as a diagnostic default.
Older advice needs a scope check. The Legacy internet-access page explicitly applies to Codex Cloud (Legacy). Do not transfer its agent-network defaults to the current Cloud experience without checking which environment you use.
When to investigate quota or an outage
If Codex reports a usage limit, use the usage-limit checklist and read the account’s actual notice. Successful work elsewhere does not establish that every account, model or execution surface has identical available capacity. A public reset announcement also does not confirm an individual account’s eligibility.
If several unrelated operations fail around the same time, compare the relevant component and incident timestamps on the Codex status view and its official sources. A resolved incident may explain an earlier error; a green summary does not diagnose a local installation. For a general stalled-tool problem, start with the slow or stuck tools guide.
Prepare a useful report
Save the time and timezone, execution location, client version, selected model, minimal command, expected result and actual error. Include whether the same input failed in a fresh task or only an existing one. Redact tokens, private repository contents and service credentials. Do not replace a failed result with a successful retry when describing the sequence.
The goal is a narrow conclusion: which operation failed, in which environment, and which difference explains it. If the evidence still cannot separate setup, access and service behavior, say that it remains unresolved rather than calling it a general model regression.