The Agent Skills Directory

[INDIRECT_PROMPT_INJECTION]: The skill processes data from the knowledge graph and local memory files (e.g., decisions.jsonl, MEMORY.md) which are then formatted into the agent's context. This presents a surface for indirect prompt injection if the stored memories contain adversarial instructions. Ingestion points include the MCP mcp__memory__search_nodes tool and memory files in .claude/memory/**. While results are structured into diagrams or tables, explicit prompt delimiters to isolate ingested content are not defined. The skill grants the agent capabilities like Bash and AskUserQuestion, which could be abused if the agent is successfully injected. Basic sanitization is present in the form of HTML entity escaping and ID sanitization in visualization scripts.- [COMMAND_EXECUTION]: The skill executes local scripts and system commands. SKILL.md defines a cron job that runs a Python script (staleness_cron.py) for maintenance. Additionally, scripts/graph-utils.mjs uses execFileSync to invoke platform-specific browser commands (open, start, or xdg-open) to display generated HTML visualizations.- [EXTERNAL_DOWNLOADS]: The skill references external software registries and CDNs. The staleness_cron.py script refers to pypi.yonyon.ai for the yg-mcp-core package, which is a non-standard PyPI registry associated with the vendor. The HTML visualization template in scripts/playground-template.html loads the vis-network library from the unpkg.com CDN.

memory