The Agent Skills Directory

[INDIRECT_PROMPT_INJECTION]: The skill aggregates terminal results from multiple parallel subagents to generate a consolidated report. This creates a surface for indirect prompt injection if the subagents process untrusted external data, as malicious instructions could be embedded in their outputs.
Ingestion points: Terminal results from N subagents (Phase C).
Boundary markers: The instructions mandate a "compact result table" and "one-line evidenced issues," which provides a structured format that helps mitigate accidental obedience to embedded instructions.
Capability inventory: The skill uses a subagent tool and git worktree commands to manage parallel tasks.
Sanitization: It explicitly instructs the agent to "Do not paste raw worker dumps," providing a level of content filtering.
[COMMAND_EXECUTION]: The skill instructs the agent to use git worktree add and git worktree remove to manage disjoint file system environments for parallel subagents that need to write to the repository.
[DYNAMIC_EXECUTION]: The skill utilizes platform-specific tools (subagent tool or t3_thread_launch) to dynamically spawn multiple concurrent processes based on task-specific briefs and roles defined at runtime.

swarm