The Agent Skills Directory

[COMMAND_EXECUTION]: The scripts scripts/loop-file.sh and scripts/loop-task.sh execute the ralph command with the --allow-all flag. This flag is explicitly designed to auto-approve all tool permissions, effectively removing human-in-the-loop (HITL) oversight and allowing the agent to execute shell commands, modify files, and access the network without requesting user confirmation for each tool use.
[EXTERNAL_DOWNLOADS]: The skill documentation (references/ralph-options.md) and installation checks in the scripts refer to and encourage the installation of the @th0rgal/ralph-wiggum package from the npm registry and its associated GitHub repository (Th0rgal/opencode-ralph-wiggum).
[PROMPT_INJECTION]: The skill implements an iterative loop that is susceptible to indirect prompt injection. If the agent is instructed to process untrusted data (such as external source code or documentation) as part of a task, malicious instructions hidden in that data could hijack the agent's behavior. Ingestion points: The loop-file.sh script reads instructions from markdown files (PROMPT_FILE) which may contain untrusted content. Boundary markers: There are no markers or system-level instructions in the scripts to prevent the AI from following commands embedded within the data it reads. Capability inventory: The underlying agent possesses comprehensive capabilities (shell access, file manipulation, etc.) which are granted full autonomy via the --allow-all flag. Sanitization: The skill lacks any mechanism to sanitize or validate the input data before it is processed by the AI in the loop.

loop