The Agent Skills Directory

[INDIRECT_PROMPT_INJECTION]: The utility script scripts/context_manager.py creates a surface for indirect prompt injection by assembling context from potentially untrusted sources.
Ingestion points: The build_agent_context function in scripts/context_manager.py accepts task and documents parameters, and ProgressiveDisclosureManager reads from external file paths.
Boundary markers: The ContextBuilder.build method concatenates context sections using simple newline separators without enforcing structured delimiters or instructions to the model to ignore embedded commands.
Capability inventory: The skill contains scripts that perform local filesystem reads and string manipulation to construct agent prompts.
Sanitization: The implementation lacks sanitization, escaping, or schema validation for the ingested strings before they are incorporated into the final context.

context-fundamentals