The Agent Skills Directory

[SAFE]: The skill instructions define a benign workflow for tracking project milestones and session progress using local memory tools (mcp__claude-flow__memory_*). No external network requests or malicious command patterns were identified.
[COMMAND_EXECUTION]: The skill manifest lists Bash, Read, and Write in its allowed-tools frontmatter, granting the agent these capabilities. However, the skill instructions themselves do not invoke these tools, focusing instead on structured data storage and retrieval via the memory interface.
[PROMPT_INJECTION]: The skill implements a data ingestion surface by storing user-defined objective names and milestone descriptions in persistent memory. While these fields are retrieved in future sessions and could theoretically contain instructions targeting the agent (Indirect Prompt Injection), the skill uses structured JSON formatting for storage and retrieval, and the behavior is essential to its primary purpose of cross-session tracking.
Ingestion points: User-provided objective names and milestone criteria stored via mcp__claude-flow__memory_store in SKILL.md.
Boundary markers: The instructions mandate the use of JSON structures for storing state, providing basic technical boundaries.
Capability inventory: The skill has access to Bash, Read, and Write tools as defined in the frontmatter.
Sanitization: No explicit sanitization or instruction-filtering steps are defined in the workflow.

horizon-track