Auto Research Pipeline

OpenClaw 原生的自动化研究 pipeline。从一个研究 topic 出发，经过 23 个 stage 产出完整论文。 每个 Phase 由独立 sub-agent 执行（context 隔离），Phase 间通过文件系统传递产出。 触发词：Research X、跑研究、文献调研、写论文、研究 pipel...

This skill is coherent for automating a research workflow, but proceed cautiously. Key things to check before installing or running: - Notifications: SKILL.md mentions pushing Feishu (飞书) messages but provides no Feishu token/config. Decide where notifications will go and supply credentials only if you trust that endpoint. - Generated-code execution: The pipeline asks the LLM to generate experiment code and then executes it. Ensure your execution environment actually enforces the promised sandbox (no network, restricted file writes, timeouts). If the platform cannot guarantee sandboxing, do not run the experiment-execution stages. - Network access: literature_search.py performs HTTP requests (arXiv, Semantic Scholar). Confirm you are comfortable with those outbound requests (rate limits, data leaving your environment). Semantic Scholar API keys are optional in code but not declared; if you supply a key, provide it securely and only if needed. - Data residency & secrets: artifacts are stored under ~/.openclaw/workspace/auto-research/. If you have sensitive files or tokens on the same filesystem, verify file permissions and isolation. - Unspecified tools: SKILL.md expects platform tools (memory_search, web_search/web_fetch, sessions_spawn). Understand what those tools send/receive and whether they transmit your prompts or files externally. If you decide to use it: run initial tests in an isolated environment (throwaway account or VM), disable network at the runtime layer if possible, and inspect any generated experiment.py before allowing execution.

Type: OpenClaw Skill Name: openclaw-auto-research Version: 1.0.0 The skill bundle implements a complex 23-stage research pipeline that inherently possesses a high risk of Remote Code Execution (RCE) by design. Specifically, Phase D (S10) and Phase E (S12) involve generating Python code via an LLM and then executing it using shell commands (`python3 experiment.py`) in a loosely defined sandbox (`/tmp/researchclaw_sandbox`). While the instructions in SKILL.md and the prompts in references/phase-d-design.md attempt to enforce 'real' computation and self-containment, the system is highly vulnerable to prompt injection; a malicious research topic could trick the agent into generating and executing code that exfiltrates data or performs other harmful actions. Additionally, the 'Evolution Overlay' feature in SKILL.md uses memory_search to load 'lessons' from previous runs, which could allow for persistent indirect prompt injection if the agent's memory is poisoned with malicious instructions.