The Agent Skills Directory

[INDIRECT_PROMPT_INJECTION]: The skill processes untrusted data from the local repository (source code, ADRs, and configurations) and external web search results. This data is used to generate architectural recommendations and reports, making it a surface for indirect prompt injection attacks where malicious content within the audited project could attempt to manipulate the agent's behavior.
Ingestion points: The skill reads project files such as docs/project-config.json and various reference documents (e.g., project-structure-reference.md) in references/mode-design.md and references/mode-review.md. It also performs external research via web searches in Step 5 of the design mode.
Boundary markers: The skill relies on structured report outputs and adversarial validation steps rather than specific prompt delimiters to isolate ingested data.
Capability inventory: The skill possesses capabilities to read local files, execute project-internal diagnostic scripts (e.g., code_graph), and spawn sub-agents for specialized architecture analysis. No unauthorized network exfiltration or credential access was detected.
Sanitization: The skill contains explicit protocols for redacting secrets (Category G7 in references/mode-scalability.md) and mandatory validation gates using the why-review tool to confirm the accuracy and safety of generated findings before they are finalized.

architecture