Skip to content

[Windows][dot] Cannot inspect and maintain an existing desktop automation end to end #54940

Open

Description

@advgiraffe

What version of the Codex App are you using?

Windows package 26.1013.4894.0; app-reported version 26.1013.40934; bundled CLI 0.163.0-alpha.10.2. Observed October 17, 2026.

What subscription do you have?

ChatGPT Pro.

What platform is your computer?

Windows, with a connected computer authorized for dot tasks.

What issue are you seeing?

I asked my dot to help maintain an existing daily project check-in. It found an obsolete instruction about a meeting that had already happened. The needed cleanup was simply removing that bullet while preserving the rest of the task and its schedule.

The dot could read a previous run, but could not retrieve the current automation definition through the supported management path. Trying a genuinely local task got past the first error, but still did not provide the agent with the definition needed to safely prepare and verify the change.

This leaves me doing routine task administration myself. Based on the marketing and launch demonstrations, I expect my dot to help coordinate my agents, projects, and tasks. I should be able to give it a goal and approve any necessary changes without learning which conversation has the right execution context or manually carrying settings between conversations. A one-bullet cleanup should not turn into a troubleshooting exercise.

What steps can reproduce the bug?

Observed sequence on this installation:

  1. Ask a dot to inspect an existing desktop automation so its instructions can be kept current.

  2. The dot-created computer task can read a prior run through the supported native thread reader.

  3. Calling native automation_update with mode: "view" and the existing automation ID fails with:

    Automations are only supported for local threads.

  4. Create a fresh task through the supported native thread API. The returned metadata explicitly identifies hostId: "local".

  5. In that local task, make the same read-only view request.

  6. This time the call succeeds, but the only returned result is:

    Rendered automation card in the app.

    The response does not give the agent the current prompt, schedule, status, or other definition fields.

No automation update was attempted because we could not verify the authoritative definition and preserve the other settings. No automation was run, deleted, or recreated. No files, databases, host metadata, or permissions were changed to work around the restriction.

What is the expected behavior?

Provide an end-to-end, supported way for an authorized dot to:

  • Read an existing desktop task's current instructions and settings.
  • Propose a narrow change, such as removing an expired milestone.
  • Apply the user-approved change while preserving its schedule and unrelated settings.
  • Read back and verify the result.

A UI card can be useful for my approval, but the agent also needs enough structured information to prepare a safe edit and report whether it actually happened. If a particular route is intentionally unsupported, explain that and provide a supported handoff that carries the task context without making me reconstruct it.

I am not asking to bypass approvals or grant unrestricted access. I am asking for the assistant to complete ordinary management work within the permissions I give it.

Additional information

There are two observed boundaries here: the delegated task is rejected as non-local; the local task's successful view returns a UI-only acknowledgement rather than the definition. We have not established whether these are intended limitations or defects, or that the two have one root cause. The automation's scheduled execution is not reported as broken.

Related reports:

This report focuses on the remaining user workflow after testing a supported local task, rather than only whether a tool is listed. Please route or consolidate it if the full workflow is already tracked.

Private project names, automation IDs, account/device/thread identifiers, local paths, screenshots, and conversation contents are omitted.

Activity

  1. added
    bugSomething isn't working
    appIssues related to the Codex desktop app
    windows-osIssues related to Codex on Windows systems
    dotsIssues involving setting up ChatGPT dots or managing their ongoing work.
    on Oct 17, 2026
  2. omarpinarecords commented on Nov 9, 2026

    @omarpinarecords

    Additional macOS reproduction on November 8–9, 2026, using ChatGPT desktop 26.1103.60933 (build 25894), Pro.

    A dot-created computer task is executing commands on the connected Mac and can list native projects. The native thread reader reports its hostId as durable, while its working directory and shell execution are on the Mac. In that task, a read-only automation_update call with mode: "view" for an existing local scheduled task fails with:

    Automations are only supported for local threads.

    The task is trying to perform an owner-approved repair of existing task bindings after a project relocation, preserving IDs, schedules, models, notification settings, supplementary folders, cursors and history. The host restriction blocks even inspection through the native automation controller, so pause/update/readback cannot safely proceed.

    This corroborates the delegated-task boundary described in this issue on a newer macOS build. I have not tested or reproduced the separate UI-card-only result from your genuinely local control task, and I am not claiming the whole issue has one cause. A local execution directory is not proof that the caller's logical host is classified as local.

    Expected: a supported, permission-preserving route for a dot delegated to a connected computer to inspect and manage the existing local automation, or an explicit supported handoff that retains the approved task context and exact settings. No request to bypass the host guard.

    No scheduler TOML/database/host metadata was edited to circumvent the restriction, and no production monitor was launched. Private project, automation, device and conversation identifiers, local paths, case details and logs are omitted.

    Duplicate check found #25873 (exact non-local rejection) and #54940 (this end-to-end dot workflow), so I am adding evidence here rather than opening another issue. Dots feedback was subsequently submitted. Returned report ID: 6a26fa86c878415080045c6059043ba7. This is the actual Dots feedback receipt, not the GitHub issue or comment ID. No separate desktop /feedback receipt is claimed.

    Follow-up: a genuinely local controller passed the host eligibility check, but a distinct pre-write folder-preservation gap remains. Current-build source/preimage evidence is added to the more specific target-schema issue: #32873 (comment) . No scheduler mutation was performed.

Sign up for free to join this conversation on GitHub. Already have an account? Sign in to comment

Metadata

Metadata

Assignees

No one assigned

    Labels

    appIssues related to the Codex desktop appautomationsbugSomething isn't workingdotsIssues involving setting up ChatGPT dots or managing their ongoing work.windows-osIssues related to Codex on Windows systems

    Type

    No type

    Projects

    No projects

      Milestone

      No milestone

      Relationships

      None yet

      Development

      No branches or pull requests

      Issue actions