The Agent Skills Directory

[REMOTE_CODE_EXECUTION]: Fetches and executes the uv installation script from astral.sh, a well-known service for Python tooling.
[COMMAND_EXECUTION]: Automatically executes a shell script named prepare.sh found within a cloned task directory, allowing for arbitrary code execution from a remote source.
Evidence: bash prepare.sh in SKILL.md (Step 5).
[COMMAND_EXECUTION]: Requests elevated privileges using sudo to install system-level Python packages on Linux environments.
Evidence: sudo apt-get install -y python3 python3-pip in SKILL.md (Step 1).
[EXTERNAL_DOWNLOADS]: Installs Python dependencies defined in a requirements.txt file from a cloned repository, which may lead to the installation of untrusted packages.
Evidence: uv pip install -r requirements.txt in SKILL.md (Step 5).
[CREDENTIALS_UNSAFE]: Accesses and manages sensitive SSH private keys and their filesystem permissions in the user's home directory.
Evidence: ~/.hive/keys/ in SKILL.md (Step 4).
[PROMPT_INJECTION]: The skill is susceptible to indirect prompt injection as it ingests and processes untrusted data from a remote hive server.
Ingestion points: hive task list and hive task clone commands in SKILL.md.
Boundary markers: No delimiters or warnings are used to separate task descriptions or metadata from instructions.
Capability inventory: The agent has the ability to run bash prepare.sh, perform pip installations, and execute sudo commands.
Sanitization: No validation or sanitization of the server-provided task data is performed before execution.

hive-setup