The Agent Skills Directory

[COMMAND_EXECUTION]: The skill's orchestration logic involves executing shell commands to manage agent sessions. In references/agents/orchestrate.yaml, the kickoff workflow uses a shell command (bun skills/orchestrate/scripts/cli.ts kickoff "<goal>") that interpolates the user-provided goal directly into the command line, creating a surface for command injection if input is not properly handled by the underlying tool.
[PROMPT_INJECTION]: The skill processes untrusted user input (goals and queries) to generate prompts for sub-agents, as seen in the templates in references/subagent-prompt-templates.md. This represents an indirect prompt injection surface where malicious instructions in a goal could potentially influence the behavior of worker agents.
Ingestion points: User input for <goal> and task requirements enters the context in references/agents/orchestrate.yaml and references/subagent-prompt-templates.md.
Boundary markers: The prompt templates utilize placeholders like <requirements> but do not consistently employ strong boundary markers or "ignore embedded instructions" warnings.
Capability inventory: The skill utilizes subprocess execution (bun, python) and file system operations.
Sanitization: No explicit sanitization or input validation logic is present in the provided skill instructions.

multi-agent