The Agent Skills Directory

[DATA_EXFILTRATION]: The skill accesses sensitive platform-specific directories such as ~/.claude/projects/ to read session transcripts (via scripts/context-usage.py) and ~/.claude/memory/ to retrieve user-specific context and decisions. While this is performed to manage sub-agent context windows and provide personalized orchestration, it involves reading sensitive interaction history and private memory files.
[INDIRECT_PROMPT_INJECTION]: The orchestrator ingests untrusted project data from various sources, including planning documents (epics.md), project conventions (CLAUDE.md), and sub-agent outputs. It employs a 'document-first' handoff strategy where instructions and findings are often passed verbatim between agents. This creates a potential surface for instructions embedded in the project files to influence the behavior of the orchestrator or its sub-agents.
[COMMAND_EXECUTION]: The skill utilizes a broad range of local system commands to build, run, and validate code across multiple technology stacks (e.g., Node.js, Python, Go, Rust, Embedded). It also uses git tools to manage commits and project state.
[EXTERNAL_DOWNLOADS]: Validation playbooks within the skill recommend the use of various developer tools and Docker containers. These include services and organizations such as Cloudflare, Vercel Labs, and official language registries (NPM, PyPI), which are used for auditing, testing, and environment simulation.
[DYNAMIC_EXECUTION]: The skill includes a local Python script (scripts/context-usage.py) that is executed to parse JSONL transcripts and calculate token usage. Additionally, several validation guides involve the generation and execution of temporary scripts (e.g., Dockerfiles or Python snippets) to verify application behavior.

bmad-auto