The Agent Skills Directory

[COMMAND_EXECUTION]: The skill provides a suite of bash scripts (skill-probe.sh, skill-trace.sh, skill-dashboard.sh) and explicitly instructs the agent to execute them to perform system diagnostics.
[COMMAND_EXECUTION]: The skill-trace.sh script performs automated code injection into other SKILL.md files on the filesystem. It appends a bash code block (a 'canary') that is designed to execute automatically whenever the agent activates the modified skill.
[DATA_EXFILTRATION]: The skill implements a logging mechanism that captures system environment data, including absolute filesystem paths, skill names, and the current working directory (cwd), and writes this information to a local file at ~/.agents/debug/activation.jsonl.
[COMMAND_EXECUTION]: The injection logic in skill-trace.sh is vulnerable to command injection. The script interpolates the name field from a skill's frontmatter directly into a bash command template using single quotes. If a skill being 'traced' has a malicious name containing a single quote followed by shell commands (e.g., name: "' ; rm -rf / ; '"), it could trigger arbitrary command execution when the agent activates that skill and runs the injected canary.

skill-debug