Aident AI
Codex Computer Use node_repl Error? Contain It on Windows
If Codex Computer Use works once on Windows and the next call fails with node_repl exec context not found, treat it as a lost helper context, not as proof that the target app is broken. Start a new task, update Codex, and make each Computer Use step import the helper and find the target again. If an action needs approval but the prompt never appears, stop or open the app yourself. Do not disable approvals or change Windows file ownership to force the action through.
Public reports from August 5 and 6, 2026 describe two useful signatures. In one, a first Computer Use action succeeds and the next state read fails. In another, app listing succeeds but launching an app fails before the approval prompt appears. There is not yet a verified general client-side fix in either report, so the safe response is containment plus a version-scoped retest.
Confirm Which Computer Use Failure You Have
The exact boundary matters because several Windows failures mention the same helper:
What you observe | Likely boundary | Next check |
|---|---|---|
The | Tool injection or task mode | Use the missing Computer Use diagnostic |
One cell can list or control an app, then a later cell says | Helper context was not retained | Start a new task and re-import, re-find, and observe in each step |
| Approval or elicitation path | Open the app manually or stop; do not bypass approval |
The message contains | Different packaging, permission, or helper failure | Follow the Windows sandbox diagnostic |
This guide owns the case where the helper is present and initially works, but its execution context is unavailable on the next action. It does not replace the missing-tool or Windows permission guides.
Reproduce the Boundary Without Changing Anything
Open a new Codex task and ask for a read-only check:
Use Computer Use on Windows. In one JavaScript execution, import
@oai/sky, list available apps, and return only their names. Do not launch, click, type, or change anything.
If that succeeds, make a second read-only request in a new execution:
Import
@oai/skyagain. Re-find the target app and read its current state. Do not reuse asky, window, screenshot, or element reference from the previous execution. Do not take an action.
The second prompt deliberately avoids reused objects. If it still returns node_repl exec context not found, you have a small, supportable reproduction: a fresh task, a successful read-only first execution, and a failed fresh second execution.
Record the Codex version, Windows version, task mode, exact first successful call, exact failing call, and whether an approval prompt appeared. Attach sanitized logs or screenshots to the existing issue instead of opening a duplicate.
Contain the Failure in Fresh, Bounded Steps
The public reports do not establish that restarting, repairing, or reinstalling will fix this signature. Use the narrowest containment while waiting for an update:
Update Codex, then create a new task rather than resuming the failed task.
In every JavaScript execution, import the Computer Use helper again.
Re-enumerate the app or window and obtain fresh state in that execution.
Perform at most the single action the user requested.
Do not carry helper objects, screenshots, element handles, or window references into a later execution.
If the action requires approval and no prompt appears, stop. Open the app manually if that is acceptable, then continue with read-only observation.
Use this prompt to make the boundary explicit:
Use Windows Computer Use for one bounded step. Start a fresh JavaScript execution, import
@oai/sky, re-find the target app, and read fresh state. Perform only [requested action]. Do not reuse any object from an earlier execution. Preserve every approval prompt. If approval cannot be requested or the execution context is missing, stop and report the exact failing call without retrying a destructive action.
This pattern may keep a read-only or already-approved unit of work moving. It does not repair an approval flow that never renders, and it should not be presented as a permanent fix.
Avoid High-Risk Windows Workarounds
Do not respond to this error by taking ownership of WindowsApps, weakening antivirus or sandbox settings, adding broad trusted directories, running the entire app as administrator, or disabling confirmation prompts. Those changes affect a much larger security boundary than the failing helper context.
Also avoid reinstalling the app you wanted Codex to control. A successful first call shows that Codex could see or use that app before its own context disappeared. Reinstalling the target is unlikely to address that evidence and can destroy useful local state.
The official OpenAI Computer Use API guide covers a different product surface, but its safety principles still apply: isolate the environment, restrict allowed actions, and keep a human in the loop for consequential work. A stuck approval path is a reason to stop, not permission to route around the gate.
Remove the Containment After a Verified Fix
Retest after Codex or its Computer Use plugin changes:
Start a new task on the updated version.
Run a read-only app listing in one execution.
Run a fresh state read in a second execution.
Request one harmless action that requires approval.
Confirm that the approval prompt appears and the action runs only after approval.
Remove the extra containment only when both executions succeed and the approval prompt renders reliably. If either test fails, keep the workflow bounded and add the new version and reproduction to the open issue.
Route Online-Service Work Around the Desktop UI
Some Computer Use tasks are really API jobs in disguise: reading a calendar, finding a CRM record, checking an issue, or retrieving analytics. For those jobs, a reviewed read-only integration can remove desktop UI automation from the critical path. It does not fix Codex Computer Use, but it can finish the underlying task without clicking through an app.
Give Codex the canonical setup instruction:
Follow https://aident.ai/SETUP.md
Then use a bounded discovery prompt:
Use Aident Loadout to find a read-only Action for [the online-service job]. Inspect the current schema and show the exact quote before execution. Do not select an Action that creates, edits, publishes, messages, or deletes data. If no reviewed read-only Action exists, stop.
For the full integration pattern, see How to Connect Claude Code and Codex to Real-World Tools. Success is measurable: the intended record or result is returned without a Computer Use action, every input is visible, and no external state changes.
Ready to move one blocked online-service task off the desktop UI? Set up Aident Loadout and route it through a reviewed Action.
Sources
Refresh this guide when either issue changes state, OpenAI documents the desktop helper lifecycle, the Codex or Computer Use version changes, or the linked Aident setup and diagnostic guides change.
The one tool
for every tool
your agent needs.
Give any AI agent real capabilities in seconds. Connect 1,000+ tools once, skip the setup headache, and let your agents execute.