Problem or use case
When an agent uses Orca's computer-use skill to operate the Orca desktop UI itself, it can navigate away from the active agent conversation—for example, into project settings—complete and verify the requested change, and then finish while leaving the user on that intermediate screen.
From the user's perspective, this makes completion ambiguous: the conversation that contains the agent's progress and final response is no longer visible, so the user cannot easily tell whether the operation is still running or already finished.
This appears to be a completion-contract gap in the shipped skill. The current snapshot → act → snapshot loop verifies the target action, but does not require preserving and restoring the user's originating UI context.
Steps to reproduce
- Start an agent conversation in an Orca workspace.
- Ask the agent to use Computer Use to change an Orca project setting.
- The agent navigates to Settings, changes the value, and verifies that it was saved.
- The agent completes the task without navigating back to the originating workspace and conversation tab.
Actual behavior
Orca remains on the settings or intermediate UI used by Computer Use. The user must manually find and reopen the original workspace/conversation to see whether the agent completed the task.
Expected behavior
The Computer Use skill should treat UI restoration as part of task completion:
- Before the first UI-changing action, record the originating app, window, Orca workspace, and conversation/tab when they can be identified.
- Perform the requested UI operation and verify its durable result.
- Unless the user explicitly asked to leave the destination screen open, return to the originating workspace and conversation/tab.
- Verify that restoration succeeded before reporting completion.
- If restoration is unsafe, unavailable, or cannot be verified, leave the current UI intact and explicitly report that restoration did not occur.
Suggested skill rule
Before changing desktop UI state, capture the user's originating app/window/workspace/tab when possible. After verifying the requested result, restore that context unless the user asked to keep the destination open or restoration would be unsafe. Verify the restored state before completing; otherwise report that restoration could not be confirmed.
This could likely be added to the completion/action rules in skills/computer-use/SKILL.md without requiring Computer Use itself to undo the requested operation.
Environment
- Orca: 1.4.141
- Platform: Windows
- Agent: Codex
- Target UI: Orca project settings
Problem or use case
When an agent uses Orca's
computer-useskill to operate the Orca desktop UI itself, it can navigate away from the active agent conversation—for example, into project settings—complete and verify the requested change, and then finish while leaving the user on that intermediate screen.From the user's perspective, this makes completion ambiguous: the conversation that contains the agent's progress and final response is no longer visible, so the user cannot easily tell whether the operation is still running or already finished.
This appears to be a completion-contract gap in the shipped skill. The current snapshot → act → snapshot loop verifies the target action, but does not require preserving and restoring the user's originating UI context.
Steps to reproduce
Actual behavior
Orca remains on the settings or intermediate UI used by Computer Use. The user must manually find and reopen the original workspace/conversation to see whether the agent completed the task.
Expected behavior
The Computer Use skill should treat UI restoration as part of task completion:
Suggested skill rule
This could likely be added to the completion/action rules in
skills/computer-use/SKILL.mdwithout requiring Computer Use itself to undo the requested operation.Environment