The Agent Skills Directory

[DYNAMIC_EXECUTION]: The skill downloads source code snapshots from a remote repository and executes them to verify the update, which involves runtime loading and execution of externally fetched scripts.
[COMMAND_EXECUTION]: The update_bundle.py script uses subprocess.run to call internal verification tools such as show_menu.py and validate_registry.py using computed paths during the update process.
[EXTERNAL_DOWNLOADS]: The skill fetches software updates and manifest data from GitHub's API and archive services (api.github.com and codeload.github.com).
[PROMPT_INJECTION]: Instructions in SKILL.md advise the agent to assume user authorization for the update process if specific keywords like "update" or "upgrade" are detected, instructing the agent not to ask for confirmation again.
[INDIRECT_PROMPT_INJECTION]: The skill ingests new instruction sets (SKILL.md files) from an external repository to update the agent's capabilities, creating a surface where future agent behavior is determined by remote content.
Ingestion points: Fetches release-manifest.json and various SKILL.md files from the vendor's GitHub repository via update_bundle.py.
Boundary markers: The skill does not explicitly define delimiters for the newly ingested instructions, though it relies on platform-level loading.
Capability inventory: The skill has file-write access (atomic_write) and command execution capabilities (subprocess.run).
Sanitization: Includes strict SHA-256 hash verification for all downloaded files and path validation to prevent directory traversal.

yetong-update