The Agent Skills Directory

[INDIRECT_PROMPT_INJECTION]: The skill takes user-supplied goals (e.g., via the /go command) and interpolates them directly into the instructions for specialized subagents (Explore, Plan, Workers) without sanitization or protective boundary markers.
Ingestion points: The userGoal variable in processGoal and analyzeGoal functions within SKILL.md.
Boundary markers: Absent; the input is interpolated directly into prompt strings.
Capability inventory: The skill can read/write files, manage a message bus, spawn subagents via the Task tool, and execute CLI commands.
Sanitization: None detected; the skill relies on subagents to interpret potentially adversarial user input.
[PROMPT_INJECTION]: The skill contains instructions that explicitly mandate the avoidance of user oversight. In agentic mode, the agent is told to "NEVER ask the user" and to resolve ambiguities or blockers autonomously. This reduces transparency and could allow a compromised subagent to perform actions without the user's knowledge.
[REMOTE_CODE_EXECUTION]: The skill references an external Node.js dependency parallel-dispatch via a require() statement. This package is not a standard library and its source is not defined within the skill, representing an unverifiable dependency risk.
[COMMAND_EXECUTION]: The orchestration logic involves spawning external processes (e.g., executing the claude CLI for Board of Directors deliberation) which includes interpolated variables like {question}, potentially creating a path for command injection if the input is not strictly validated.

conductor-orchestrator