The Agent Skills Directory

[DYNAMIC_EXECUTION]: The skill spawns a separate, isolated sub-agent (an "adjudicator") using a dedicated system prompt provided in adjudicator.md. This sub-agent is dynamically tasked with verifying repository state and ruling on triage arguments.
[EXTERNAL_DOWNLOADS]: The skill executes npx -y crysk, which fetches and runs code directly from the npm registry. It also downloads remote CI artifacts using gh run download to serve as input for the triage process, creating a dependency on external data sources.
[INDIRECT_PROMPT_INJECTION]: The skill ingests untrusted data from GitHub PR comments (specifically  markers). This creates an attack surface where a malicious finding could influence the agent's decision-making or lead it to apply incorrect code "fixes."
Ingestion points: GitHub PR comments and CI run artifacts.
Boundary markers: Uses specific HTML-style comment tags to identify findings but does not describe sanitization of the finding's body text.
Capability inventory: Includes file system modification, git push, posting GitHub comments, and execution of CLI tools.
Sanitization: No explicit sanitization or escaping of the finding content before it is processed by the agent or adjudicator.
[OBFUSCATION]: The skill utilizes URL encoding for file paths within finding markers and includes shell snippets that decode Base64 data from the GitHub API to inspect workflow configurations.
[COMMAND_EXECUTION]: The skill heavily utilizes shell-based tooling, including git, gh, and npx, to automate the PR review lifecycle and repository updates.

crysk-loop