The Agent Skills Directory

[COMMAND_EXECUTION]: The skill executes Git and package management commands (npm install, git worktree add) to manage isolated environments for subagents. These operations are scoped to the project's worktree directory.
[EXTERNAL_DOWNLOADS]: Fetches project dependencies from official package registries (NPM) during the environment setup phase for delegated tasks.
[PROMPT_INJECTION]: The skill ingest implementation plans and review reports which are used to generate prompts for subagents, creating a surface for indirect prompt injection.
Ingestion points: Task details are extracted from implementation plans in SKILL.md and failure contexts are read from state files in references/fix-mode.md.
Boundary markers: Templates in references/implementer-prompt.md and references/fixer-prompt.md use markdown headers (e.g., '## Task Description', '## Issue to Fix') to delimit untrusted content.
Capability inventory: Spawned subagents have the ability to modify files, execute tests, and perform git operations (commit/push).
Sanitization: No explicit sanitization or escaping of the plan/review content was identified before interpolation into subagent prompts.

delegation