The Agent Skills Directory

[PROMPT_INJECTION]: The skill's SKILL.md file contains a 'CRITICAL CONSTRAINTS' section that uses high-priority markers ('override all other instructions', 'NEVER VIOLATE') to enforce fully automated execution. This is explicitly designed to bypass human-in-the-loop confirmations during a long-running orchestration process (up to 2 hours), increasing the risk if the agent is misled by external data.
[REMOTE_CODE_EXECUTION]: The orchestration scripts rely on an external executable named codeagent-wrapper. The skill searches for this tool in various locations, including the system PATH and folders associated with Claude tools (e.g., ~/.claude/bin). This utility is the primary agent for executing code and managing state during orchestration tasks.
[COMMAND_EXECUTION]: Several scripts, including orchestration_loop.py and dispatch_batch.py, use the subprocess module to launch local Python scripts and the wrapper binary. The system also leverages tmux to manage parallel execution environments.
[PROMPT_INJECTION]: The skill exhibits a surface for indirect prompt injection (Category 8) by processing task descriptions from a tasks.md file and including them in prompts for other agents.
Ingestion points: scripts/spec_parser.py reads and parses tasks.md into JSON data structures.
Boundary markers: Absent; descriptions and details from the tasks are interpolated into prompt templates without protective delimiters.
Capability inventory: The system can execute arbitrary code via codeagent-wrapper, write to the project file system, and manage terminal sessions.
Sanitization: No validation or escaping is applied to the content extracted from tasks.md before it is sent to worker or review agents.

multi-agent-orchestration