The Agent Skills Directory

[COMMAND_EXECUTION]: The skill executes the acpx CLI tool with the --approve-all flag. This flag is designed to bypass interactive confirmation for changes made by the external implementation persona, which significantly increases the risk of unintended or malicious code being committed if the input task or the tool itself is compromised.
[PROMPT_INJECTION]: The skill exhibits an indirect prompt injection surface by processing external tasks and file contents that are subsequently passed to the acpx tool.
Ingestion points: User-provided task descriptions and local file paths (Phase 0 and Phase 1 in SKILL.md).
Boundary markers: Absent. The skill instructions specify building a prompt file by appending task and context to the role definition without structural isolation (e.g., XML tags or specific delimiters).
Capability inventory: Execution of acpx with write permissions via --approve-all, file reading, and invocation of the code-reviewer task agent.
Sanitization: The skill includes a manual instruction to "Sanitize secrets before including file contents," which mitigates but does not eliminate the risk from malicious data injection.
[DATA_EXFILTRATION]: There is a potential risk of data exposure as the skill reads local files and includes their content in prompt files for the acpx execution environment. While the skill includes a safeguard instruction to "Sanitize secrets before including file contents in any prompt file," the reliability of this process depends on the LLM's ability to identify and redact all sensitive information (credentials, internal paths, etc.) before transmission to the external process.

execute-acpx