The Agent Skills Directory

[INDIRECT_PROMPT_INJECTION]: The skill is designed to ingest and store data from potentially untrusted sources such as external API error messages and user feedback.
Ingestion points: Data enters the system when the agent logs details from '命令/工具失败' (Tool failures), '用户纠正' (User corrections), or '外部 API 失败' (External API failures) into the .learnings/ directory.
Boundary markers: The skill uses Markdown headers (e.g., ### Details, ### Summary) to structure the logs, but does not implement specific boundary tokens or 'ignore instructions' warnings to prevent the agent from accidentally executing commands embedded within the logged data during future reviews.
Capability inventory: The skill utilizes bash via grep for searching logs and has the capability to write and modify files within the project directory.
Sanitization: No explicit sanitization or escaping of external inputs (like API error strings) is mentioned before they are written to the Markdown logs.
[COMMAND_EXECUTION]: The skill uses local shell commands for maintenance and search tasks.
Evidence: The skill instructs the agent to use grep -r "keyword" .learnings/ to search for duplicate patterns and grep -h "Status\*\*: pending" .learnings/*.md | wc -l for periodic reviews of logged items.

self-improvement