The Agent Skills Directory

[EXTERNAL_DOWNLOADS]: The skill instructs users to install the arize-phoenix-client package with the harbor extra using pip. This is a standard library from the skill's vendor (Arize AI).
[COMMAND_EXECUTION]: The instructions provide examples for executing the harbor run CLI command to start agent evaluations with the Phoenix plugin enabled.
[INDIRECT_PROMPT_INJECTION]: The skill processes untrusted data originating from Harbor tasks and agent trajectories (ATIF format) for recording and visualization in Phoenix.
Ingestion points: The skill reads Harbor task collections, registry datasets, published packages, local dataset paths, and repository metadata, as well as ATIF trajectory files from agent trials in SKILL.md.
Boundary markers: The instructions define rigid mappings and tables to translate Harbor data into Phoenix datasets, experiments, and traces, which acts as a logical boundary for how data is interpreted.
Capability inventory: The skill environment allows Python package installation, execution of the harbor CLI, and network communication with a Phoenix collector endpoint for data recording.
Sanitization: The skill specifies that LLM inputs are reconstructed from ATIF data and explicitly tagged with metadata.atif.input_source = "reconstructed" to distinguish them from original inputs.

phoenix-harbor