The Agent Skills Directory

[PROMPT_INJECTION]: The skill documents a pattern that creates a surface for indirect prompt injection. \n
Ingestion points: Data is read from .claude/*.local.md files located in the project root (SKILL.md, references/parsing-techniques.md). \n
Boundary markers: The implementation pattern does not define or suggest boundary markers or explicit instructions for the agent to ignore instructions embedded within the configuration files. \n
Capability inventory: Extracted content from the files is used as a 'system message' or feedback 'reason', allowing external content to influence the agent's logic and next steps (references/parsing-techniques.md, references/real-world-examples.md). \n
Sanitization: Although JSON construction is safely handled via jq, no semantic validation or filtering of the markdown prompt content is implemented. \n- [COMMAND_EXECUTION]: Risk of command injection through terminal session interaction via tmux. \n
The multi-agent-swarm implementation example in references/real-world-examples.md passes data extracted from local configuration files directly to tmux send-keys. \n
Maliciously modified configuration fields (such as agent_name or coordinator_session) could trigger unintended command execution in target terminal sessions.

Plugin Settings