The Agent Skills Directory

[METADATA_POISONING]: The skill contains fabricated bibliographic data, including a 2026 publication date for a research survey and an invalid social media status ID in the discussion links. There is also a discrepancy between the attribution to 'ara.so' in the document and the provided author context of 'reason-machines'.
[INDIRECT_PROMPT_INJECTION]: The skill instructs the agent to clone and parse a remote GitHub repository (RUC-NLPIR/Awesome-Long-Horizon-Agents). This creates an attack surface where instructions embedded in the repository's markdown files could influence agent behavior.
Ingestion points: git clone and grep operations on the downloaded repository content in SKILL.md.
Boundary markers: None; the agent is not provided with delimiters or instructions to ignore content within the downloaded files.
Capability inventory: Shell execution (git, grep, cd) and potentially arbitrary tool usage via the MCPAgent pattern described in the examples.
Sanitization: Absent; no filtering or validation is performed on the ingested content before the agent processes it.
[EXTERNAL_DOWNLOADS]: The skill initiates a download from a remote GitHub repository (https://github.com/RUC-NLPIR/Awesome-Long-Horizon-Agents.git). While the target organization is known, the retrieval of external data for processing introduces risks associated with untrusted content.
[COMMAND_EXECUTION]: The skill utilizes shell commands (git clone, cd, grep) to manage and search external data, which is an expected but powerful capability that must be monitored.

awesome-long-horizon-agents-survey