Repository navigation
Permissions and security
A security tool that reads hostile code should be held to its own standard.
-
One permission: the Home Assistant API. No host network, no open port, no
Supervisor role, and no access to your configuration folder, so
secrets.yamland the auth storage are out of its reach. It never opens the installed copies of your apps. -
What it uses the API for is a fixed list: reading your Home Assistant version;
asking the Marketplace for its installed list and its catalog; subscribing to two
Marketplace signals for auto-sniff; creating the bell notification and, if you picked
one, calling your notify service; listing your AI Task entities and notify services
for the two dropdowns; and calling
ai_task.generate_dataonly when you click an AI button. -
Its own AppArmor profile, enforcing (Home Assistant security rating 8). Every
process in the container can run only the tools the image ships, write only under
/dataand/tmp, and use only ordinary network sockets. Nothing downloaded is executable. - The panel answers only Home Assistant's ingress, which handles the login. It ships no JavaScript: its content security policy allows no scripts.
It is treated as hostile. Barkley never executes, imports, builds or installs it. Archives are extracted into a fresh temporary folder with protections against path tricks, symlinks and oversized files. Dependencies are resolved from prebuilt wheels only, so no build script ever runs.
The release workflow builds the image on GitHub's runners from the source at the release tag. Every tool is pinned by version, the downloaded binaries by checksum, and nothing is compiled during the build. The base image's signature is checked before the build, and a published version is never overwritten. Each image is signed with Cosign; checking that signature is optional, see Installing and updating.
Every push and pull request is scanned before it can reach main: Gitleaks for secrets,
Semgrep's registry rules and CodeQL for the code, Trivy for dependencies and for the
built image (high and critical CVEs that have a fix), Checkov and Hadolint for the
Dockerfile and the workflows. A finding fails the run. Anything accepted is written down
with its reason in the repository.
Barkley was built with AI pair-programming assistance (Claude, via Claude Code) under human direction and review, and most of the code and docs are AI-generated. The checks above run on all of it, and all of it is in the repository to read.