Read the download before installing
The course sample has one manifest and one generated Python entrypoint. It declares no required capabilities or credentials.
Treat a skill as executable capability
A skill adds a tool by installing code that an agent can execute. Source review is part of installation, not optional homework.
Browse and record the current boundary
armaraos skill search <QUERY>
armaraos skill list
Rank source risk
| Source | Risk posture |
|---|---|
| Appstore name | Curated, but still review what it grants. |
| Local path | Read every file before install. |
| Git URL | Arbitrary remote code, highest risk. |
Read the two course files
course-6/downloads/course-safe-skill/skill.toml
course-6/downloads/course-safe-skill/src/main.pyThe entrypoint reads JSON and returns Processed: .... It has no credential access.
Install from the reviewed path
armaraos skill install course-6\downloads\course-safe-skill
armaraos skill listExercise once and read the real evidence
armaraos message <AGENT_UUID> "Use the course_safe_skill tool with input skill check."
armaraos approvals list
The CLI approval queue remained empty. The direct tool result proves use; the security audit recorded the agent message but no separate skill event.
Remove the exact skill
armaraos skill remove course-safe-skill
armaraos skill listRemoval is permanent and has no dry-run. The reviewed download remains available for inspection.
Check it / defend the source choice
Explain why the local sample was acceptable.
Point to its two files, no declared capabilities, simple echo behavior, and the empty final installed list.
Installed skills execute code with agent permissions. Read source first, keep approvals enabled, install one at a time, and never explore third-party code with --yolo.