Phoenix Loop

Auto-diagnose agent failures, extract reusable recovery patterns, and create local skills to fix recurring blockers while keeping all data private and local.

This skill appears to do what it claims and keeps data local, but review these before installing: 1) Platform compatibility — the SKILL.md uses PowerShell commands; ensure your agent runtime supports PowerShell or adapt the commands. 2) Review and approve generated files — the skill will create/update files in skills/local/ that the agent can later run; you should inspect new recovery skills before trusting them for autonomous fixes. 3) Privacy filtering is regex-based and imperfect — run the provided privacy-checklist and consider stronger secret-detection if you have high-sensitivity data. 4) Filename/verification inconsistencies — the skill has a small mismatch in example filenames; test the workflow in a sandbox copy of your repository. 5) Back up skills/local/ and memory/ before first run, and consider restricting autonomous invocation or enabling a manual approval step for new skills until you are comfortable with the loop's behavior.

Type: OpenClaw Skill Name: phoenix-loop Version: 1.0.0 The 'phoenix-loop' skill aims to improve agent performance by diagnosing failures and creating reusable recovery skills. However, it instructs the agent to extract 'solution steps' and 'fallback actions' from local failure logs (`memory/blocked-items.md`, `memory/tasks.md`) and then execute these steps. If these logs contain attacker-controlled input (e.g., a crafted error message or task description), the agent could be prompted to generate and execute arbitrary commands as part of its self-healing process. While the skill includes a 'Sensitive Data Filter' in `SKILL.md` and `references/privacy-checklist.md` to prevent sensitive data leakage, it does not explicitly sanitize commands or instructions extracted from logs, creating a significant prompt injection vulnerability against the agent.