The Agent Skills Directory

[EXTERNAL_DOWNLOADS]: The skill downloads binary archives and metadata from lightmake.site and Tencent Cloud Object Storage (myqcloud.com).
Evidence: SKILLHUB_INSTALL_KIT_URL, SKILLHUB_SEARCH_URL, and SKILLHUB_PRIMARY_DOWNLOAD_URL_TEMPLATE in scripts/skillhub_tool.py.
[REMOTE_CODE_EXECUTION]: The bootstrap and install commands download zip/tar archives from the internet and extract them into the workspace or user home directory, effectively installing external executable scripts.
Evidence: install_skill_archive and command_bootstrap functions using safe_extract_zip and safe_extract_tar in scripts/skillhub_tool.py.
[COMMAND_EXECUTION]: The script executes shell commands via the subprocess module to configure the local 'openclaw' environment, detect software versions, and manage a local gateway process.
Evidence: Calls to subprocess.run and subprocess.Popen in configure_plugin, disable_plugin_if_present, and restart_gateway_if_needed.
[PROMPT_INJECTION]: The skill has an indirect prompt injection surface as it ingests untrusted metadata (skill names and summaries) from a remote API which are then presented to the agent.
Ingestion points: fetch_search_results in scripts/skillhub_tool.py.
Boundary markers: Absent in the prompt construction.
Capability inventory: Subprocess execution, file system modifications, and network requests are available via the skill's own tooling.
Sanitization: No sanitization is performed on the descriptive strings retrieved from the API before they are output to the agent.

find-skills