Roundtable Adaptive

Adaptive multi-model AI roundtable. Runs up to 4 AI models (configurable) in 2 debate rounds with cross-critique and formal consensus scoring. Requires a con...

This skill appears to do what it says, but check these before installing: 1) The skill requires an Anthropic API key for the Claude panel — only provide it if you trust the skill and understand provider billing. 2) Full multi-model operation is optional: OpenAI and Blockrun integrations are supported but require extra credentials/config (OPENAI_API_KEY, BLOCKRUN_PROXY_URL) and, for Blockrun, installing a plugin and funding a wallet — expect small per-run costs if you enable them. 3) The skill will post to Discord channels/threads and needs permission to list/create threads and post messages; set the output channel intentionally (ROUNDTABLE_OUTPUT_CHANNEL/panels.json). 4) Results and logs are written to your workspace memory directory ({workspace}/memory/roundtables or ~/clawd/memory/roundtables) — if you need to prevent local persistence, do not enable the skill or modify the paths. 5) The package source/homepage is missing and the registry lists an owner ID but no homepage — if provenance matters, ask the publisher for a source repo or audit the files. 6) If you won't use Blockrun/OpenAI features, you can run in Claude-only degraded mode; otherwise, verify you trust any third-party proxies (Blockrun) before pointing local services at them. If you want greater assurance, request the source repo or a signed release and check the exact message/action APIs the agent will call in your environment.

Type: OpenClaw Skill Name: roundtable-adaptive Version: 2.9.1 The skill is classified as suspicious due to critical prompt injection vulnerabilities. The user's input `[PROMPT]` is directly inserted into multiple AI prompt templates (`prompts/meta-panel.md`, `prompts/round1.md`, `prompts/round2-cross-critique.md`, `prompts/round3-validation.md`) without apparent sanitization. This allows a malicious user to inject arbitrary instructions into the AI models, potentially leading to unauthorized actions, data exfiltration (e.g., API keys via `env()` or `message()` actions), or remote code execution if the agent has access to shell commands. Additionally, the `[CURRENT_CONTEXT]` derived from `web_search(query = prompt)` also presents a potential vector for injecting malicious instructions via manipulated search results.