The Agent Skills Directory

[INDIRECT_PROMPT_INJECTION]: The skill is designed to ingest and act upon data from external, untrusted sources which creates an attack surface for indirect prompt injection.
Ingestion points: The researcher agent reads comments from GitHub issues using the gh CLI, parses provider changelogs and documentation via web searches, and references external benchmarks from the aiewf-eval repository.
Boundary markers: External content fetched from GitHub is formatted into a local file (digest-comments.md) without explicit delimiters or instructions for the agent to ignore any embedded directives within the comments.
Capability inventory: The agent has the ability to execute shell commands, manage git branches and worktrees, use the gh tool for repository interaction, and run local Python scripts via uv that can interact with external APIs.
Sanitization: While the instructions explicitly warn against including credentials in output, there is no mechanism to sanitize or validate the natural language content fetched from external sources to ensure it does not contain malicious instructions aimed at the agent.
[COMMAND_EXECUTION]: The skill relies heavily on executing shell commands to perform its maintenance and research tasks.
Evidence: Extensive use of git, gh, uv, jq, and python scripts is documented in SKILL.md and RESEARCH_GUIDE.md for operations such as cloning repositories, viewing issues, and probing API endpoints.
[EXTERNAL_DOWNLOADS]: The skill performs network operations to clone and update repositories.
Evidence: The orchestrator runs gh repo clone pipecat-ai/provider-watch-reports _reports to manage local report files.
Context: These downloads target the official repository of the skill's authoring organization and are necessary for the skill's stated purpose of tracking provider updates.

provider-research