The Agent Skills Directory

[SAFE]: The skill is entirely instructional and provides a framework for manual or agent-assisted debugging. It does not include any scripts, binaries, or instructions for remote code execution or network communication.
[PROMPT_INJECTION]: The skill introduces a surface for indirect prompt injection by design, as it requires the agent to read and process external, potentially untrusted data like logs and system outputs to be stored as evidence nodes.
Ingestion points: Case files within the /coe/ directory are read and updated based on external inputs (logs, code, user feedback).
Boundary markers: The skill uses a strict Markdown structure with specific headers (# 问题, ## 假设, ## 证据) to organize information, which provides structural delimitation but not security isolation.
Capability inventory: The agent uses standard file read and write operations to maintain the case files within the project workspace.
Sanitization: There are no instructions provided to sanitize or escape the raw content ('原始内容') captured from external sources before it is written to the case files and subsequently re-read by the agent.

coe-debug