Ralph Loops

Runs autonomous iterative AI loops for requirements, planning, or building phases using structured prompts and fresh context per iteration.

Install only if you are comfortable running an autonomous coding loop with broad local authority. Use a disposable branch or isolated VM/container, provide only task-specific credentials, disable or restrict the dashboard to loopback with authentication, review diffs before any push or tag, and do not run it in repositories or workspaces containing sensitive files until the dashboard command-injection/XSS and unauthenticated control issues are fixed.

Type: OpenClaw Skill Name: ralph-loops Version: 1.0.2 This skill is classified as suspicious due to its reliance on `--dangerously-skip-permissions` for the Claude CLI, which grants the AI agent full system access, bypassing all built-in safeguards. Additionally, the `scripts/ralph-loop.mjs` script includes a `checkCmd` argument that allows arbitrary command execution via `execSync`, creating a direct command injection vector if the agent's arguments can be manipulated. The skill also instructs the agent to perform `git push` operations and use the `browser` tool to open URLs (e.g., `browser action=open targetUrl="http://localhost:PORT"`), which are powerful capabilities that, while intended for development, significantly increase the attack surface for prompt injection or misuse. These risky capabilities are present in `SKILL.md`, `scripts/ralph-loop.mjs`, `templates/loop.sh`, and `templates/PROMPT_build.md`.