Evidence scope: Desktop packaging and the
ohmyagentengine are documented in the pinned official desktop README. Local-directory capability is listed on the product site. This guide is a reproducible procedure, not a completed desktop-task recording.
1. Prepare a disposable folder
Download the starter ZIP and extract it into a new folder such as monkeycode-first-task. Run python3 -m unittest discover -v there: the touching-interval case should fail among eight tests. Save the original ZIP so you can start again without deleting unrelated files.
2. Open only that folder
Install the desktop client from the official product download entry, then select the extracted folder as the working directory. Record the app version shown by your installation and the selected model route. The local working directory and the model provider are separate choices; local file access does not mean all model processing stays on your machine.
3. Run and inspect the bounded task
Paste the exact bug-fix prompt from the first-task guide. Confirm any required file or command permissions for that folder. Check the changed file in your own editor and run the tests independently in a terminal. Accept only the one-file change and eight passing tests before continuing to the test and documentation tasks.
If Python is unavailable to the app, compare its command environment with your terminal. If files are missing, inspect the selected folder rather than expanding access to your entire home directory. If the model request fails, inspect the account or configured endpoint separately from folder permissions.
4. Keep or discard the result
Save a diff and the exact check output. To retry, open a newly extracted copy. Stop the task before changing the working directory or disconnecting the account. Removing a folder from the app does not necessarily delete its local files, task history or provider-side records; verify each action in the installed version.
The public reference outputs show what the checks should establish. They are not a MonkeyCode run, and the public reference must not be scored as a blind benchmark.