The Agent Skills Directory

[COMMAND_EXECUTION]: The skill is designed to run shell commands ('verification commands' and 'gates') that are retrieved from local project files such as docs/verification-plan.xml and docs/operational-packets.xml.
[REMOTE_CODE_EXECUTION]: By executing commands sourced from data files, the skill creates a vector for arbitrary code execution if the configuration files are compromised or provided by an untrusted source.
[DATA_EXFILTRATION]: The ability to execute arbitrary shell commands can be leveraged to exfiltrate sensitive information, such as environment variables, local credentials, or source code, to external servers.
[PROMPT_INJECTION]: The skill is susceptible to indirect prompt injection as it processes and follows instructions embedded in external XML files without adequate sanitization.
Ingestion points: docs/development-plan.xml, docs/knowledge-graph.xml, docs/verification-plan.xml, and docs/operational-packets.xml.
Boundary markers: None present; the agent processes the XML content as authoritative instructions.
Capability inventory: Shell command execution (via verification steps and vendor-provided CLI tools like grace), file system writes (during code generation), and git commit operations.
Sanitization: No evidence of command validation or input sanitization before execution.

grace-execute