The Agent Skills Directory

[COMMAND_EXECUTION]: The skill dynamically constructs shell commands in SKILL.md by directly interpolating the ${BRANCH} and ${CUSTOM_DIR} variables into the Bash tool invocation string. If a user inputs shell metacharacters (such as ;, &&, or $()) via a custom branch name or directory path, it can result in arbitrary code execution within the host environment.
[INDIRECT_PROMPT_INJECTION]: The skill exposes a vulnerability surface by accepting untrusted data via user input or prompt arguments without validation.
Ingestion points: SKILL.md handles unvalidated input strings for custom branches and directory paths.
Boundary markers: Absent.
Capability inventory: Executes shell scripts using the platform's Bash tool.
Sanitization: There is no filtering or escaping of shell metacharacters before executing the concatenated command string.
[EXTERNAL_DOWNLOADS]: The scripts/download_openharmony.sh script automatically downloads an external executable script from a GitCode repository (https://raw.gitcode.com/gitcode-dev/repo/raw/main/repo-py3) and runs pip3 install using a Huawei Cloud PyPI mirror. While these are recognized mirrors for the OpenHarmony ecosystem, dynamically downloading and running scripts at runtime introduces supply chain risks.

openharmony-download