The Agent Skills Directory

[PRIVILEGE_ESCALATION]: The skill documents and utilizes a 'bypassPermissions' mode for subagents via the Task tool. This mode allows subagents to perform sensitive operations, such as shell command execution and file system modifications, without requiring manual user approval for each specific action. A transparency risk is present in the 'Show plan and confirm' step (Step 4), as the provided markdown template for the user confirmation table omits the permission mode, which may lead to users unknowingly approving subagents that operate with bypassed safety prompts.- [INDIRECT_PROMPT_INJECTION]: The skill is susceptible to indirect prompt injection because it processes untrusted user-provided task descriptions and interpolates them into the instructions for dynamically created subagents. While the skill employs a human-in-the-loop confirmation mechanism, it lacks explicit sanitization or boundary enforcement for the user input before it is used to define subagent behaviors.
Ingestion points: The main task description provided by the user in SKILL.md.
Boundary markers: The skill uses the AskUserQuestion tool to seek user confirmation of the team structure and task allocation before starting work.
Capability inventory: Depending on the subagent type, the spawned agents have access to tools including Bash, Edit, Write, and WebFetch (as defined in references/agent-types.md).
Sanitization: No specific sanitization, filtering, or escaping of the user-provided task description is specified in the skill's logic.- [COMMAND_EXECUTION]: The skill facilitates a broad command execution surface by creating subagents of the 'Bash' or 'general-purpose' types. These subagents are explicitly granted access to shell command tools for tasks such as environment setup, CI/CD operations, and build processes, all orchestrated by the primary agent.- [DYNAMIC_EXECUTION]: The skill performs dynamic execution by generating role-specific prompts and task definitions at runtime based on the user's initial request. These generated instructions are then used to initialize and execute new agent contexts via the platform's Task tool.

swarm