The Agent Skills Directory

[INDIRECT_PROMPT_INJECTION]: The skill represents an attack surface for indirect prompt injection because its primary purpose is to ingest, process, and evaluate traces and spans from LLM applications, which frequently contain untrusted user inputs.
Ingestion points: Data enters the agent's context through client.get_spans_dataframe and dataset creation methods in SKILL.md and references/advanced-usage.md.
Boundary markers: No explicit instruction-ignoring delimiters or boundary markers are suggested for use in evaluation templates (e.g., CUSTOM_EVAL_TEMPLATE in references/advanced-usage.md).
Capability inventory: The skill utilizes LLM-as-judge capabilities via llm_classify, makes network calls to the Phoenix server, and logs evaluation results back to a central platform.
Sanitization: There are no mentions of sanitizing or escaping the input or output data before it is interpolated into evaluation prompts.
[EXTERNAL_DOWNLOADS]: The skill documentation guides the user to download several external resources.
Python packages: Instructions include installing arize-phoenix and its various sub-packages from PyPI.
Docker images: Deployment guides reference the arizephoenix/phoenix:latest image from Docker Hub.
[COMMAND_EXECUTION]: The skill provides numerous shell commands for environment setup and server management.
Server operations: Use of phoenix serve to launch the observability backend.
Database configuration: Instructions for exporting sensitive environment variables like PHOENIX_SQL_DATABASE_URL for PostgreSQL connections.

phoenix-observability