The Agent Skills Directory

[PERSISTENCE]: The skill includes a documented mechanism in references/positioning/04_persistence_暂停恢复与隐私.md for saving user progress to the local filesystem. It writes session state, evidence, and candidates into the ./eva-positioning/ directory. This is a functional feature of the skill that includes security controls such as explicit user consent prompts, versioning of state files, and restrictions on writing sensitive data like passwords or cookies.
[INDIRECT_PROMPT_INJECTION]: The skill possesses an attack surface for indirect prompt injection as it is designed to ingest and process untrusted external data provided by the user, specifically platform search results, comments, and account details as described in references/positioning/02_platform-search_平台现实取证.md.
Ingestion points: Untrusted data enters the agent context through user-provided screenshots, text copies, and platform search evidence回传 (回传) in references/positioning/02_platform-search_平台现实取证.md.
Boundary markers: The skill explicitly references an external safety protocol (../eva-shared/references/shared/06_external-material-safety_外部材料安全边界.md) in its loading logic in SKILL.md to handle untrusted materials.
Capability inventory: The skill possesses the capability to write files to the local directory for persistence (references/positioning/04_persistence_暂停恢复与隐私.md).
Sanitization: While the skill requires user confirmation for privacy before saving state, it does not detail specific character-level or instruction-level sanitization for the platform data it processes.

eva-positioning