The Agent Skills Directory

[DYNAMIC_EXECUTION]: The script scripts/invoke.py dynamically identifies the path to an execution engine (kntnt.py) by searching through a pre-defined list of potential global skill directories associated with various AI agent harnesses (e.g., ~/.claude/skills, ~/.cursor/skills). It then executes this engine using uv run, which constitutes dynamic code execution from computed paths.
[INDIRECT_PROMPT_INJECTION]: The skill retrieves ticket bodies and comments from an external tracker (GitHub) and injects them into subagent prompts via the references/review.md template. This untrusted content is not sanitized, creating a vulnerability where a malicious ticket could influence subagent behavior. 1. Ingestion points: GitHub ticket content (body and comments) read in Step 2 of SKILL.md. 2. Boundary markers: The prompt uses Markdown headers but lacks explicit 'ignore embedded instructions' warnings for the subagent. 3. Capability inventory: The skill uses gh and uv, and subagents are tasked with repository analysis; scripts/invoke.py uses subprocess.run. 4. Sanitization: No escaping or validation is performed on the ingested ticket data.
[COMMAND_EXECUTION]: The scripts/invoke.py file uses the subprocess.run function to execute system commands (uv run). While it uses an argument list rather than a shell string, it enables the execution of the found manager tool and passes user-provided input through stdin.

ready-for-agent-check