Project Loop

Run approved long-running project work from file-backed state, continue through self-clearable tasks, pause cleanly at real gates, and recover across sessions.

This skill is coherent for orchestrating long-running, file-backed projects and does not request credentials or install software. Before installing: 1) confirm you only use it for explicitly approved projects and that project folders (agents/<agent-id>/projects/<project-slug>) do not contain sensitive secrets the agent shouldn’t modify; 2) review how your agent identity maps to state.json.owner_agent so the correct agent executes tasks; 3) be aware it can proceed autonomously across multiple turns (it forbids waiting for human confirmation on self-clearable tasks) and will send regular status updates (every ~5 minutes) — decide whether you want that behavior and where updates will be delivered; 4) consider testing it in a sandbox project to verify 'verify reality' and status-update behavior are acceptable before using on critical projects.

Type: OpenClaw Skill Name: project-loop Version: 1.0.2 The skill implements a high-autonomy project management framework that includes several high-risk behaviors. Key indicators include instructions in SKILL.md to bypass security filters by splitting commands if they are 'denied for obfuscation or size', the automated creation of system persistence via 'watchdog cron' jobs, and a directive to prioritize external file-based state over direct user chat, which creates a vector for indirect prompt injection. While the skill includes security guardrails requiring human approval for sensitive actions like credential handling and installs, the combination of automated persistence and filter-bypass logic warrants a suspicious classification.