The Agent Skills Directory

[INDIRECT_PROMPT_INJECTION]: The skill implements an automated "Inherit and Deposit" cycle that creates a persistence layer for conversation-derived knowledge, establishing a surface for indirect prompt injection.\n- Ingestion points: User-controlled conversation history and task requests are processed in SKILL.md via the DEPOSIT mode to extract "CORE/PATTERN/DECISION/TIP" content.\n- Boundary markers: The skill lacks explicit boundary markers or instructions in SKILL.md to treat summarized content as data rather than directives when re-read during the "INHERIT" phase.\n- Capability inventory: In SKILL.md, the agent uses view_file to read knowledge and performs file-write/merge operations to update doc/memory.md, .agent/rules/rule-one.md, and other documentation files.\n- Sanitization: Although SKILL.md and resources/memory-schema.md enforce structural constraints (e.g., removing code blocks, limiting TL;DR length), they lack sanitization logic to filter out adversarial instructions from being promoted to "Core Wisdom".

session-managing