Hello everyone,
I have no background in IT or programming, and I do not have technical knowledge about computers. Because of my health problems, I have been trying to build a personal assistant, with the help of AI tools, to support me in my everyday work and activities.
I would like to make clear from the outset that my goal is a local AI assistant running on my own Mac, including voice conversations, with my data and conversations staying on the computer. I am looking for help building this local solution, not recommendations for cloud chatbots or cloud voice services. Using online AI tools to help with development does not change that goal.
That is the purpose of this project: to have assistance I can rely on, especially with things I do not know how to do on my own. The first step is to be able to have stable conversations with it in Romanian and German.
However, I have become stuck while preparing to verify the app before installation, and I do not know how to continue safely without losing existing conversations or data.
If someone with macOS experience could guide me or provide practical help to get past this step, it would mean a great deal to me. The technical appendix below was prepared with the help of AI tools, based on recorded technical evidence.
Technical appendix
Verification goal
Before replacing the existing app, a separate copy needs to be verified in an isolated environment: without access to personal files unrelated to the test, without network access, and without affecting the existing app or its data.
Complete saving and restoration of the actual conversations remain requirements that must be demonstrated before replacement; we do not consider them already verified.
Environment and recorded error
When this question was being prepared, system information reported macOS 27.0.1, build 26A434. However, the error below comes from a saved earlier attempt, not a new reproduction. The record of that attempt does not include the macOS or Ruby versions.
The preparation used a custom sandbox-exec/SBPL profile and a /usr/bin/ruby supervisor, which was intended to launch /usr/bin/codesign to ad-hoc sign a verification copy.
After removing the personal path prefix, the error was:
bounded_child.rb: Operation not permitted @ realpath_rec - /Users (Errno::EPERM)
The Ruby supervisor exited with code 1. No PID or outcome for the codesign process was recorded, and the copy remained unchanged and unsigned. We therefore do not present this result as a demonstrated failure of the signing tool or an Apple restriction on voice apps.
Findings and corrections
Source inspection showed that the preparation mechanism enumerated the ancestor directories beneath /Users but omitted /Users itself. The profile restricted reads in this area and did not have an explicit exception for metadata access to that directory. This finding is consistent with the observed denial, but it is not a diagnosis confirmed by a new access test.
In a separate, inert copy, the directory enumeration and preflight check were corrected: if the necessary permissions are missing, launch is stopped. Tests of this correction verify only the preparation logic, not actual access within the sandbox.
The denied attempt has not been repeated, the profile has not been broadened, and the candidate has not subsequently been signed, installed, or launched.
Specific question
While researching the project, we found this Apple DTS explanation of the difference between custom SBPL profiles and App Sandbox configured through entitlements. I am not looking for undocumented rules or a way to bypass the access denial.
What is the documented, Apple-supported approach for isolating this verification workflow? Is an App Sandbox/XPC architecture appropriate when the workflow needs to invoke an Apple signing tool, or should app preparation, signing, and verification be separated into other isolation environments, such as a dedicated macOS virtual machine?
Guidance on the appropriate mechanisms and their limitations would help us define the next safe step for the local app.
Help in Romanian, German, or English is welcome.
My heartfelt thanks to everyone who takes the time to read this and help.
I would do this verification in a VM.
The immediate issue you’re having is that you’re trying to re-initialise the sandbox. That’s not allowed. Once a process is in a sandbox, it can neither get out of that sandbox nor apply a different sandbox.
Furthermore, it looks like the third-party infrastructure you’re using is playing games with a custom SBPL sandbox. That’s not something we support, as I explained in that other thread you found.
Finally, the nice thing about using a VM in this context is that you start fresh each time. That guarantees that state from build N of your product doesn’t affect build N+1.
Share and Enjoy
—
Quinn “The Eskimo!” @ Developer Technical Support @ Apple
let myEmail = "eskimo" + "1" + "@" + "apple.com"