The Agent Skills Directory

[COMMAND_EXECUTION]: The skill uses the mode: "bypassPermissions" flag when calling the Task tool in SKILL.md. This bypasses standard security gates and permission prompts for sub-agents, granting them elevated autonomy which can be exploited if they receive malicious instructions.
[PROMPT_INJECTION]: The skill is vulnerable to Indirect Prompt Injection through its agent-to-agent data flow.
Ingestion points: Results from earlier tasks (e.g., {result_1} in SKILL.md and {architect_result} in references/prompt-templates.md) are directly interpolated into the prompts for subsequent agents.
Boundary markers: The templates do not use any delimiters (like triple quotes or XML tags) or safety instructions to distinguish between the system's instructions and the untrusted data being processed.
Capability inventory: Sub-agents have the ability to execute code, modify files, and update task statuses via the Task and TaskUpdate tools, often under the bypassPermissions flag.
Sanitization: There is no evidence of sanitization, filtering, or validation of the content produced by one agent before it is used to direct the behavior of the next agent.

team-assemble