The Agent Skills Directory

[INDIRECT_PROMPT_INJECTION]: The skill creates a surface for indirect prompt injection by processing untrusted data from GitHub including issues, pull requests, and comments. 1. Ingestion points: Content is ingested from GitHub issue bodies, pull request descriptions, comments, and CI logs as described in references/github-conventions.md. 2. Boundary markers: The skill employs HTML comments to wrap its internal status blocks and includes explicit directives to the agent to treat external content as data rather than instructions. 3. Capability inventory: The agent has permissions to write to the file system, commit/push code via Git, execute shell commands for verification, and spawn additional agent sessions. 4. Sanitization: The skill relies on the agent following explicit instructions to differentiate between human-authored commands and external data, but does not implement automated technical sanitization of the ingested text.
[DYNAMIC_EXECUTION]: The workflow requires executing verification commands such as test suites and linters sourced from repository documentation, which presents a risk if the repository contains malicious configuration. Additionally, through its delegation model, the skill facilitates the spawning of autonomous subagent sessions (workers and investigators) to perform implementation and review tasks, which constitutes a form of dynamic execution across separate agent contexts.

loop-engineering