Wei Devils Advocate

Stress-test ideas using multiple LLMs in adversarial roles to generate counterarguments, cross-check reasoning, and expose hidden risks and failure modes. 易找...

What to check before installing: - Metadata mismatch: The registry metadata shows no required environment variables, but SKILL.md and code require OPENROUTER_API_KEY and/or DASHSCOPE_API_KEY. Don’t assume there are no credentials needed — the skill will ask for provider API keys. - Review client code: The included scripts make network calls to the configured api_base endpoints (openrouter.ai and dashscope.aliyuncs.com). If you will supply API keys, ensure you trust those provider endpoints and understand that queries (and any user data you pass) will be sent there. - Inspect prompt files: The pre-scan detected a prompt‑injection string in SKILL.md. Open prompts/debater.txt and prompts/judge.txt (and SKILL.md) and confirm the suspicious phrases are only documented as examples or defenses, not embedded in ways that could cause unintended model behavior. - File writes: The tool will write reports and intermediate outputs to reports/ and intermediate/ in the skill directory. Don’t run it from a location containing sensitive files you don’t want mixed or leaked. - Install steps: SKILL.md instructs installing Bun via the official bun.sh installer (downloaded by curl). That is common but you should only run it if you trust the source. The package dependencies are standard (axios, dotenv). - If you are unsure: request the author to update registry metadata to declare required env vars, provide a formal install spec, and confirm why any prompt phrases that resemble injection attacks are present. If you will provide API keys, avoid sending secrets or extremely sensitive data through the skill until you have audited the client implementations and prompts.

Type: OpenClaw Skill Name: wei-devils-advocate Version: 1.0.0 The skill bundle is a legitimate multi-LLM adversarial analysis tool designed to stress-test ideas using various language models. The code implementation in `scripts/agent.ts` and the HTTP clients in `scripts/clients/` are well-structured, using standard libraries like Axios and Dotenv. Notably, the agent includes a proactive `sanitizeInput` function to mitigate prompt injection attacks by redacting common injection phrases. Sensitive data such as API keys are correctly handled via environment variables, and file operations are restricted to local report generation within the skill's directory. No evidence of data exfiltration, malicious execution, or unauthorized persistence was found.