The Agent Skills Directory

[INDIRECT_PROMPT_INJECTION]: The skill processes untrusted text from X (Twitter) posts, which could contain instructions intended to manipulate the agent. The current implementation does not use boundary markers or delimiters when formatting this content into Markdown. Mandatory evidence: 1. Ingestion points: The fetch_via_fxtwitter and fetch_via_syndication functions in fetch_x.py retrieve data from external web APIs. 2. Boundary markers: The generate_markdown function in fetch_x.py lacks delimiters or warnings to prevent an agent from interpreting embedded instructions within the fetched tweet text. 3. Capability inventory: The skill can write files to the local file system (save_markdown_to_dir in scripts/main.py) and perform network requests to download media (download_media_files in scripts/main.py). 4. Sanitization: While the skill uses yaml.safe_load for its own configuration, it does not sanitize or escape the external content retrieved from the web before formatting it into Markdown.
[EXTERNAL_DOWNLOADS]: The skill connects to api.fxtwitter.com and cdn.syndication.twimg.com to retrieve tweet data, and downloads media files from various hosts found within tweet data (e.g., pbs.twimg.com) when the media download feature is enabled.
[COMMAND_EXECUTION]: The agent instructions in SKILL.md involve using shell commands such as test -f to verify the existence of configuration files in project and user directories.

x-fetcher