The Agent Skills Directory

[EXTERNAL_DOWNLOADS]: The skill instructs the user to install the arize-phoenix-client Python package. This is a legitimate library from the vendor (Arize AI) used for the intended purpose of the skill.
[CREDENTIALS_UNSAFE]: The skill mentions the use of PHOENIX_API_KEY but correctly advises users to set this via environment variables rather than hardcoding it, preventing the key from being stored in shell history or configuration files.
[INDIRECT_PROMPT_INJECTION]: The skill defines an attack surface by ingesting external data from Harbor tasks, ATIF trajectories, and JSON reference files.
Ingestion points: Harbor task metadata (task.toml), ATIF trajectory files, and user-defined JSON reference files.
Boundary markers: The skill does not explicitly define LLM boundary markers as it primarily handles data mapping for an observability plugin.
Capability inventory: The plugin reads files and performs network operations to the Phoenix collector endpoint (PHOENIX_COLLECTOR_ENDPOINT).
Sanitization: The skill explicitly implements path traversal mitigation by rejecting absolute paths, parent directory components (..), and symlinks that escape the task root folder.
[PROMPT_INJECTION]: A static hint regarding concealment was evaluated and found to be a false positive. The skill actually encourages transparency by instructing the agent not to suppress error facts and ensuring users are notified of record conflicts.

phoenix-harbor