# 产出示例：hardware_profile.json
{
  "gpu_type": "NVIDIA RTX 3090",
  "cuda_version": "11.8",
  "memory_gb": 24,
  "recommended_package": "torch>=2.0.0"
}

{
  "paper_id": "arxiv:2301.00001",
  "title": "Attention Is All You Need",
  "key_findings": ["Transformer架构", "自注意力机制", "并行计算效率"],
  "method": "encoder-decoder + multi-head attention",
  "limitations": ["计算复杂度O(n²)", "位置信息编码需要额外设计"],
  "research_gap": "长期依赖建模效率问题"
}

# 辩论结构
agents = [
    Agent(role="Critic", perspective="极限性能"),
    Agent(role="Advocate", perspective="实际应用"),
    Agent(role="Devil", perspective="潜在失败模式")
]

# 多轮辩论后生成假设
hypotheses = debate(agents, synthesis_findings)
# → 输出：3个可验证的假设，每个带有novelty_score

code_agent:
  enabled: true
  architecture_planning: true      # 先规划架构，再生成代码
  sequential_generation: true       # 按依赖顺序生成文件
  hard_validation: true             # AST验证，阻止硬编码指标

class SandboxHarness:
    def __init__(self, config):
        self.time_guard = TimeGuard(max_seconds=300)
        self.metric_validator = MetricValidator()
        self.memory_limit = 4 * 1024 * 1024 * 1024  # 4GB
        self.allowed_imports = ['math', 'random', 'numpy', 'torch']
    
    def run(self, code: str) -> ExperimentResult:
        # AST 验证：确保没有硬编码指标
        # 不可变 harness：防止代码修改实验参数
        # NaN/Inf 快速失败：检测到异常值立即终止

class VerifiedRegistry:
    """
    确保论文中的数据来自真实实验
    """
    def __init__(self):
        self.verified_metrics = set()
        self.unverified_warnings = []
    
    def register(self, key: str, value: float, source: str):
        # 所有指标必须来自 experiment_run 或 ITERATIVE_REFINE
        # 手动添加的数据会被标记为 unverified
        if source in ALLOWED_SOURCES:
            self.verified_metrics.add(key)
        else:
            self.unverified_warnings.append(key)
    
    def validate(self, paper_text: str) -> ValidationResult:
        # 论文中的所有数字必须来自 verified_metrics
        # 未验证的数字会被替换为 [VERIFIED: X.X] 或警告

analysis_agents = [
    Agent(perspective="统计显著性"),
    Agent(perspective="实际意义"),
    Agent(perspective="与现有工作对比")
]

analysis_report = multi_agent_debate(results, agents)

review_rubric:
  - 评分维度: ["原创性", "技术深度", "实验完整性", "写作质量"]
  - 评分范围: 1-10
  - 检查项:
      - 所有基线是否包含？
      - 消融实验是否充分？
      - 论文声明与实验证据是否一致？

# 输出目录结构
artifacts/rc-20260310-143200-a1b2c3/
├── paper_final.md              # Markdown 终稿
├── paper.tex                  # LaTeX 源码
├── references.bib             # BibTeX 引用
├── charts/                    # 自动生成图表
│   ├── accuracy_comparison.pdf
│   └── ablation_study.pdf
└── code/                      # 开源代码包
    ├── experiment.py
    └── requirements.txt

class CitationVerifier:
    def verify(self, citation: Citation) -> VerificationResult:
        # Layer 1: arXiv ID check
        arxiv_valid = self.check_arxiv_id(citation.arxiv_id)
        
        # Layer 2: DOI metadata
        doi_valid = self.check_doi(citation.doi)
        
        # Layer 3: Title match
        title_similarity = self.check_title_match(
            citation.title, 
            citation.extracted_from
        )
        
        # Layer 4: LLM relevance
        relevance_score = self.llm_judge_relevance(
            citation.claim,
            citation.full_text
        )
        
        return VerificationResult(
            layers_passed=sum([arxiv_valid, doi_valid, ...]),
            relevance_score=relevance_score
        )

class SmartPause:
    def __init__(self, hitl_config):
        self.confidence_threshold = 0.7
        self.low_confidence_signals = [
            "novelty_score < 0.5",
            "baseline_missing_standard_method",
            "experiment_timeout_rate > 30%"
        ]
    
    def should_pause(self, context: StageContext) -> bool:
        # 基于上下文的置信度决定是否暂停
        confidence = self.calculate_confidence(context)
        return confidence < self.confidence_threshold

# config.arc.yaml
openclaw_bridge:
  use_cron: true              # ⏰ 定时研究运行
  use_message: true           # 💬 进度通知（Discord/Slack/Telegram）
  use_memory: true            # 🧠 跨会话知识持久化
  use_sessions_spawn: true    # 🔀 为并发阶段生成并行子会话
  use_web_fetch: true         # 🌐 文献综述期间实时网络搜索
  use_browser: false          # 🖥️ 基于浏览器的论文收集

# config.yaml — ACP 配置示例
llm:
  provider: "acp"
  acp:
    agent: "claude"   # 任何 ACP 兼容的代理 CLI 命令
    cwd: "."          # 代理的工作目录
  # 无需 base_url 或 api_key —— 代理自行处理认证

# 克隆仓库
git clone https://github.com/aiming-lab/AutoResearchClaw.git
cd AutoResearchClaw

# 创建虚拟环境
python3 -m venv .venv
source .venv/bin/activate

# 安装
pip install -e .

# 交互式配置（推荐）
researchclaw setup
researchclaw init

# 或手动配置
cp config.researchclaw.example.yaml config.arc.yaml

# config.arc.yaml
project:
  name: "my-research"

research:
  topic: "Your research topic here"

llm:
  base_url: "https://api.openai.com/v1"
  api_key_env: "OPENAI_API_KEY"
  primary_model: "gpt-4o"
  fallback_models: ["gpt-4o-mini"]

experiment:
  mode: "sandbox"
  sandbox:
    python_path: ".venv/bin/python"

# 完整自主运行 —— 无需人工干预
researchclaw run --config config.arc.yaml --topic "Your research idea" --auto-approve

# Co-Pilot 模式 —— 在关键决策点协作
researchclaw run --config config.arc.yaml --topic "Your research idea" --mode co-pilot

# 列出所有加载的技能
researchclaw skills list

# 安装自定义技能
researchclaw skills install /path/to/my-skill/

# 禁用技能
researchclaw skills disable scientific-writing

hitl:
  cost_budget_usd: 50.0        # 达到预算时暂停
  notifications:
    channels: ["terminal"]     # terminal | slack | webhook
  timeouts:
    default_human_timeout_sec: 86400  # 24小时超时
    auto_proceed_on_timeout: false