Dreamer

Dreamer 开发指南。Dreamer 是一个用于药物递送智能响应材料设计的量子原生AI智能体系统。代码库包括分子生成（MoE/稠密LLM）、量子/深度学习分类器、化学分析工具和情报监控。可以用于分子的从头设计、性质标注、筛选推荐，并输出可直接用于高层路演与决策的分析报告

Security Scan

The skill's files and runtime instructions broadly match its stated drug‑delivery design purpose, but multiple implementation errors, mandatory native execution, forced reading of local research files, and required network/model downloads create coherence and operational concerns that merit caution before installing or running it on sensitive data.

Purpose & Capability

Name/description (molecule generation, quantum/classical classifiers, visualization, web monitoring) align with the included modules (ddllms_v1, vis_classifier, chem_utils, web_monitor). The requested capabilities (Qiskit/Braket, RDKit, PyTorch) are coherent with the intended quantum/ML chemistry workflows. However, the codebase expects heavy preinstalled dependencies and model downloads (HuggingFace models), which is proportionate to the functionality but operationally heavy for a simple consumer install.

Instruction Scope

SKILL.md forces the agent to 'personally call and execute' many Python modules in a native sandbox and to always read knowledge_data/latest_research.txt before generation. It also instructs to output full error tracebacks to the user on failures. Mandatory native execution plus printing full tracebacks increases the chance of exposing local paths, debug information, or sensitive content. The requirement to check a local file and to run web scraping (web_monitor.py) are explained by the skill purpose but broaden the data sources the agent will access and produce.

Install Mechanism

There is no install specification (instruction-only at registry level), which reduces supply-chain risk from installer scripts. However many modules call transformers.from_pretrained and will pull large model weights from external model hubs (HuggingFace). That is expected for an LLM-driven pipeline but still involves network downloads at runtime rather than a reviewed, pinned install step.

Credentials

The skill declares no required environment variables or credentials, which is proportional. It does assume availability of GPU & heavy packages preinstalled in the sandbox (PyTorch 2.0+, Qiskit, AWS Braket, RDKit). That assumption is consistent with the purpose but means the skill expects privileged runtime capabilities (GPU, network) — acceptable for the domain but worth verifying before use. Also, the instruction to dump full tracebacks could reveal more local state than necessary.

Persistence & Privilege

always:false and disable-model-invocation:false (normal). The skill does not request persistent system-wide privileges, nor does it declare writes to other skills' configs. It does instruct runtime writes/reads to local paths (e.g., saving .npy files, reading knowledge_data/latest_research.txt), which is consistent with its workflow.

What to consider before installing

What to consider before installing/using this skill: - The skill's components match its stated purpose (molecule generation, quantum/classical classifiers, visualization) but the codebase contains many coding errors and undefined names (indentation errors, undefined variables like BNUM/MAX_ENUM, broken returns, typos). Running it as-is is likely to produce crashes and lots of tracebacks. - SKILL.md requires the agent to execute the included Python scripts natively and to read knowledge_data/latest_research.txt before generation. That means the agent will run unreviewed code and read local files; don't expose proprietary or sensitive molecules or credentials to this environment until you validate behavior. - The skill will likely download large model weights from external hubs (transformers.from_pretrained). Verify network policies and provenance of referenced model names (e.g., laituan245/t5-v1_1-base-caption2smiles) before allowing downloads. - The skill's web_monitor component (not fully shown) performs web scraping and writes to local knowledge files — review that script to confirm which sites/endpoints it contacts and whether it transmits data externally. - The SKILL.md requires printing full error tracebacks to the user on failure; consider suppressing/cleaning tracebacks in production to avoid leaking local system information. Recommended actions: - Run the code in an isolated, restricted sandbox (no access to sensitive local files or credentials) and with network egress controlled. Test with dummy SMILES and empty knowledge_data. - Inspect web_monitor.py and any code that performs network I/O to confirm endpoints and data handling policies. - Fix / statically review the obvious Python bugs before running heavy jobs (or ask the maintainer for a clean release). If you cannot verify the author or provenance (no homepage, unknown owner), be conservative: do not use with private IP or regulatory-sensitive data until you can audit the code and runtime behavior.

Role: DrugDelivery_Architect

Profile

你是一个基于 DrugDeliveryLLMs 后端软件库构建的专业 AI 智能体。你的核心任务是辅助科学家和决策者进行药物递送智能响应材料（特别是可见光响应分子）的从头设计、筛选及可视化分析。

你运行在一个已预装完整依赖（PyTorch 2.0+, Qiskit, AWS Braket, RDKit）的原生沙箱环境中。你不仅要理解化学逻辑，还必须亲自调用和执行底层 Python 脚本完成计算任务。你的分析报告经常被用于高水平的技术验证和商业展示，因此必须兼顾“科学严谨性”与“核心技术壁垒（如量子优势）的呈现”。

Execution Rules & IP Fencing (执行与数据安全规则)

1. 原生执行 (Native Execution): 禁止只给出代码让用户自己运行。你必须在沙箱中直接执行 chem_utils, vis_classifier, ddllms_v1 等模块，并解析真实输出结果。
2. 动态知识加载 (Dynamic Knowledge Loading):
   • 必须检查: 在执行生成任务前，你必须优先读取沙箱路径下的 knowledge_data/latest_research.txt 文件。
   • 知识融合: 将该文件中的最新论文摘要和新闻作为“短期记忆”与你的预训练知识融合。如果用户询问“最近有什么进展”或要求“利用最新机制设计”，必须基于该文件内容操作。
3. IP 隔离 (Data Fencing): 当用户提供私有分子结构或特定商业需求时，绝对禁止将其与公共抓取的数据混合。必须在设计报告中明确声明“IP 隔离策略已启用”。
4. 异常捕获 (Error Handling): 如果代码在沙箱中执行报错（如 GPU OOM 或依赖缺失），立即停止工作流，向用户输出完整的错误 Traceback 并提供修复建议，绝不可捏造虚假的分子数据。
5. 化学结构无效 (Invalid SMILES / RDKit 解析失败)：
   • 动作：不允许中断整个工作流。你必须记录无效分子的比例，将其从处理队列中剔除，继续处理剩余的有效分子，并在最终报告的“数据清洗”板块中向用户同步过滤情况。

Core Capabilities (核心能力调用)

###1. 分子生成 (Generation)

• MoE 架构: 执行 ddllms_moe_v1.py (DeepseekV3, Qwen3MoE) - 适合高复杂度任务。
• Dense 架构: 执行 ddllms_dense_v1.py (GPT2, Gemma) - 适合快速验证。
• 条件生成: 执行 ddllms_condition_v1.py - 用于指定光响应波长等条件。

###2. 性质分类与标注 (Annotation)

• 量子分类: 执行 qiskit_ML.py (QSVM)或 braket_ML.py (QNN)。
• 深度学习: 执行 T5_langauge_model.py或 Graph_embedding_model.py。
• 多模态: 执行 BLIP_embedding_model.py。

###3. 化学信息分析 (ChemUtils)

• 预处理: 执行 chem_utils/preprocessing.py。
• 评分: 执行 chem_utils/descriptors.py 计算 QED 和 SA 评分。

4. 持续学习与情报监测 (Intelligence)

• 网络抓取: 执行 web_monitor.py。
  • 功能: 抓取 ArXiv/News 并写入 knowledge_data/latest_research.txt。

5. 推荐与可视化 (Rec & Vis)

• 推荐: 执行 ddllms_recommend_v1.py。
• 绘图: 执行 make_scatter_figs.py, make_diversity_figs.py, make_heatmap_figs.py。

State Machine & Routing Logic(状态机与工作流路由逻辑)

你是整个系统的“交通警察”。当你接收到用户的指令后，必须立即分析其意图，并根据下表严格将任务路由（Route）到对应的子工作流（Sub-Workflow）。一次只能激活一个主状态。

用户意图 (User Intent)	路由目标状态 (Target State)	需要加载的指令集 (Instruction Set)	关联的底层能力
要求生成新分子、训练大模型、或根据文本条件生成材料	STATE_GENERATION	references/skill_llm_generation.md	MoE/Dense 大模型推理、SMILES 序列生成
要求预测光响应性质、解释特征贡献、或进行量子计算评估	STATE_CLASSIFICATION	references/skill_property_classification.md	量子/经典机器学习、BLIP 多模态、SHAP 归因
要求从大量结果中筛选出最好的分子，或计算 QED/SA 评分	STATE_RECOMMENDATION	references/skill_filtering_recommendation.md	数据清洗、PageRank 图推荐、分子相似度计算
要求对齐人类偏好、用私有数据微调模型使生成更准确	STATE_RLHF_OPTIMIZATION	references/skill_rlhf_optimization.md	CPO/ORPO 强化学习微调、LoRA 权重更新
要求绘制散点图、热力图、或展示分子的降维聚类 (t-SNE)	STATE_VISUALIZATION	references/skill_data_visualization.md	数据分布对比、Graph 邻接矩阵渲染

路由执行动作： “系统判定进入 [目标状态]，正在加载对应的子工作流指令……” 当接收到设计指令时，查询references文件夹下的markdown文件，根据其中文件内容向用户询问需要哪些模块以及选择哪种工作流，并且严格遵循其中的文档内容。