The Agent Skills Directory

[PROMPT_INJECTION]: The skill identifies an attack surface for indirect prompt injection because it is designed to ingest and interpret external, potentially untrusted architecture descriptions (such as Mermaid, PlantUML, or C4 diagrams) and baseline threat reports.
Ingestion points: Untrusted data enters the agent's context through architecture files during format detection (Phase 1) and through existing baseline threats.md files during correlation (Phase 1a+).
Boundary markers: The provided instructions do not define specific delimiters or "ignore" directives to prevent embedded instructions within diagrams or baseline files from influencing the orchestrator's logic.
Capability inventory: The orchestrator is documented to use the Read tool to load reference materials and dispatches specialized sub-agents (e.g., tachi-spoofing, tachi-agent-autonomy) to perform analysis on identified components.
Sanitization: While a preprocessing pipeline is described (lowercasing, tokenization, sorting), it is intended for finding similarity comparison and does not include logic for filtering or sanitizing potential prompt injection payloads.

tachi-orchestration