Automatically discovers and loads CLAUDE.md files in the project root with support for @include directives and reverse order loading up to 40,000 characters.

What to check before installing: - Confirm the trigger model: SKILL.md says it auto-runs on each user message, but the registry does not set always:true. Ask the author how and when the skill is invoked by the agent. - Understand what will be injected: the code will read any CLAUDE.md files in the current directory and up to 3 parents and print their contents (combined and truncated). If your repo stores sensitive info in such files, that content could be added to the agent's system prompt/context. - Ask about @include handling: the README promises @include parsing but neither Python nor Node files implement it — this is a functional mismatch you should resolve. - Note there are no network calls or credential requests in the code, which lowers exfiltration risk, but the main risk is sensitive local content being loaded into the agent context. Consider testing in a sandboxed project or asking for code updates to explicitly implement and document include handling, invocation semantics, and any safety filters (e.g., skip files in .git or CI folders) before using in sensitive environments.

Type: OpenClaw Skill Name: claude-md-auto-discovery Version: 0.1.0 The skill implements a discovery mechanism for 'CLAUDE.md' files to inject project-specific instructions into the agent's system context. It is classified as suspicious because it instructs the agent to automatically execute and inject unvalidated file content into the system prompt, creating a high-risk prompt injection vector. Furthermore, the code specifically searches for a file with a leading space (' CLAUDE.md') in `claude-md.py` and `discover.mjs`, which is a common technique for hiding files, and the documentation in `SKILL.md` claims to support features like '@include' that are entirely absent from the actual implementation.