{
  "grade": "auto_execute",
  "autonomous": true,
  "requires_human": false,
  "threshold": 0.8,
  "margin_threshold": 0.05,
  "decisions": {
    "selected_strategy": {
      "value": "redis_cache",
      "confidence": 0.9231,
      "margin": 0.41,
      "grade": "auto_execute",
      "reason": "confidence 0.92 clears 0.80 and beats the runner-up by 0.41",
      "requires_human": false
    }
  },
  "grade_counts": { "auto_execute": 1, "escalate": 0, "ask_human": 0, "reconsider": 0 },
  "reasons": [],
  "summary": "all 1 answers clear the auto-execute bar; proceed",
  "routing": { "model": "english", "reason": "English Latin text" }
}

result = mcp.call("decide", state=diff, preset="code_change_risk")
grade = result["contract"]["grade"]

if grade == "auto_execute":
    merge_pull_request()                       # 直接干，零推理成本
elif grade == "escalate":
    run_full_llm_review(diff)                  # 花 token 深挖
elif grade == "ask_human":
    notify_reviewer(result["contract"]["reasons"])
else:  # reconsider
    result = mcp.call("decide", state=diff, preset="code_change_risk",
                      threshold="critical")    # 换阈值 / 换问题重试

                        ┌─────────────────────────────┐
   MCP Agent  ────────► │  /mcp  StreamableHTTP       │
 (Claude Code /         │  (9 tools / 4 resources /   │
  Codex / Cursor)       │   2 prompts)                │
                        ├─────────────────────────────┤
   脚本 / 看板 ────────► │  /v1/...  REST (FastAPI)    │
                        └──────────────┬──────────────┘
                                       │  共用同一套实现
                        ┌──────────────▼──────────────┐
                        │        laya_mcp.core        │
                        │  settings / schema / preset │
                        │  engine(常驻+池+缓存)         │
                        │  contract(决策契约) / metrics │
                        └──────────────┬──────────────┘
                                       │
                        ┌──────────────▼──────────────┐
                        │  laya.Router (System-1)     │
                        │  english / multilingual /   │
                        │  typed-decisions            │
                        └─────────────────────────────┘

laya-mcp/
├── src/laya_mcp/
│   ├── core/                 # 与传输方式无关的领域层
│   │   ├── settings.py       #   LAYAMCP_* 配置、阈值等级
│   │   ├── schema.py         #   公开问题 schema ↔ Laya 原生格式
│   │   ├── presets.py        #   5 个内置问题集
│   │   ├── contract.py       #   ★ 概率 → grade 的决策契约
│   │   ├── engine.py         #   单例 Router + 线程池 + LRU 缓存
│   │   ├── errors.py         #   统一错误分类与 HTTP 映射
│   │   └── metrics.py        #   计数 / 延迟分位
│   ├── api/                  # FastAPI 层
│   │   ├── app.py            #   create_app()，挂载 /mcp 与 /v1
│   │   ├── routes.py         #   REST 端点（不含决策逻辑）
│   │   ├── schemas.py        #   请求 / 响应模型
│   │   └── dependencies.py   #   DI：engine / settings / metrics
│   ├── mcp_server/           # MCP 层
│   │   ├── server.py         #   工具 / 资源 / Prompt 注册
│   │   └── questions.py      #   问题解析（inline / preset / 快捷参数）
│   ├── cli.py                # 命令行入口
│   └── logging_config.py     # 结构化日志
├── examples/quickstart_router.py   # 原始 laya-test.py 用法
├── tests/
└── pyproject.toml

git clone <your-repo-url> laya-mcp
cd laya-mcp
uv sync                       # 创建 .venv 并安装依赖

uv run laya-mcp doctor

laya-mcp 0.2.0
python: 3.13.x
settings: models=['english', 'multilingual'] device=auto port=8077
laya: 0.3.5 -> .../laya/__init__.py
torch: 2.x.x cuda=False
HF 缓存: 2.20 GB，包含 laya 检查点: True
fastapi: ok
uvicorn: ok
mcp: ok
pydantic_settings: ok
结论: 环境可用

uv run laya-mcp serve

INFO  laya_mcp.engine  loading laya router  models=['english','multilingual'] device=cpu
INFO  laya_mcp.engine  laya router ready  loaded=['english','multilingual']
INFO  laya_mcp.api     engine ready in 67.8s
INFO  mcp.server...    StreamableHTTP session manager started
Uvicorn running on http://127.0.0.1:8077

curl -s http://127.0.0.1:8077/v1/decide \
  -H 'content-type: application/json' \
  -d '{
    "state": "Fix a typo in README, one file, no interface change",
    "preset": "code_change_risk"
  }' | python -m json.tool

uv run laya-mcp ask \
  --state "Fix a typo in README, one file, no interface change" \
  --preset code_change_risk

MCP Streamable HTTP 端点：http://127.0.0.1:8077/mcp

claude mcp add --transport http laya http://127.0.0.1:8077/mcp

{
  "mcpServers": {
    "laya": {
      "type": "streamable-http",
      "url": "http://127.0.0.1:8077/mcp"
    }
  }
}

{
  "mcpServers": {
    "laya": {
      "command": "uv",
      "args": ["run", "--directory", "/abs/path/to/laya-mcp", "laya-mcp", "serve", "--transport", "stdio"]
    }
  }
}

route_request  <  list_presets / describe_questions  <  decide_yes_no  <  decide  <  decide_shortlist
   (<1ms)                  (<1ms)                      (~10ms)         (10~50ms)      (20~80ms)

{
  "contract": { /* 见上文「决策契约」 */ },
  "grade": "auto_execute",          // 整体 grade，Agent 只需看这个
  "autonomous": true,               // 是否可完全自主执行
  "requires_human": false,
  "summary": "all 5 answers clear the auto-execute bar; proceed",
  "answers": {
    "blast_radius": {
      "kind": "choice",
      "value": "single_file",
      "confidence": 0.9642,
      "margin": 0.8713,
      "grade": "auto_execute",
      "reason": "confidence 0.96 clears 0.80 and beats the runner-up by 0.87",
      "requires_human": false,
      "probabilities": { "single_file": 0.9642, "module": 0.0929, "...": 0.0 },
      "raw": { /* Laya 原始返回，未加工 */ }
    }
  },
  "routing": { "model": "english", "reason": "English Latin text", "repo": "convaiinnovations/laya" },
  "latency_ms": 31.4,
  "cached": false,
  "usage": {}
}

# 查看全部 preset
uv run laya-mcp presets
# 查看单个 preset 的完整问题定义
curl -s http://127.0.0.1:8077/v1/presets/code_change_risk | python -m json.tool

{
  "state": "<diff 或工单>",
  "preset": "code_change_risk",
  "questions": {
    "target_release": {
      "kind": "choice",
      "instructions": "这个改动更适合进哪个版本？",
      "criteria": {
        "hotfix": "紧急修复，必须立刻发",
        "next_minor": "下个小版本即可",
        "backlog": "不急，排入待办"
      }
    }
  }
}

{
  "question_id": {
    "kind": "choice",                  // choice | score | yes_no
    "instructions": "选哪个方案？",       // 给模型的指令
    "criteria": {                      // 见下方「三种 kind」
      "redis_cache": "适合读多写少、容忍秒级延迟",
      "db_index": "适合查询条件固定、数据量中等"
    },
    "weight": 1.0,                     // 可选，建议性权重
    "description": "..."               // 可选，仅在罗列时展示
  }
}

curl -s http://127.0.0.1:8077/v1/questions/describe \
  -H 'content-type: application/json' \
  -d '{"preset": "input_guard"}' | python -m json.tool

# 1) 是非问题：这个动作能直接执行吗？
curl -s http://127.0.0.1:8077/v1/decide/yes-no \
  -H 'content-type: application/json' \
  -d '{
    "state": {"action": "DROP TABLE users", "env": "production"},
    "statement": "this action is safe to run without human review",
    "threshold": "critical"
  }' | python -m json.tool

# 2) 中文工单分诊（自动路由到 multilingual checkpoint）
curl -s http://127.0.0.1:8077/v1/decide \
  -H 'content-type: application/json' \
  -d '{"state": "客户反馈发票被重复扣款，语气很急", "preset": "support_triage"}' \
  | python -m json.tool

# 3) 只路由，不推理
curl -s http://127.0.0.1:8077/v1/route \
  -H 'content-type: application/json' \
  -d '{"state": "给我翻译一下这段中文"}' | python -m json.tool
# {"model":"multilingual","reason":"non-Latin script (han, 100% of letters); ..."}

{
  "error": "invalid_question",
  "message": "at most 32 questions per call",
  "details": { "count": 40, "limit": 32 },
  "request_id": "0f3a9c1b2d4e5f60"
}

uv run laya-mcp --help

uv run laya-mcp serve \
  --host 0.0.0.0 \
  --port 8077 \
  --models english,multilingual,typed-decisions \
  --device auto \
  --log-level INFO

# N 选 1（--choices 是快捷写法）
uv run laya-mcp ask \
  --state "高并发下数据库查询变慢，Query 全表扫描，QPS 5000/s" \
  --instructions "根据性能瓶颈选择最合适的优化方案" \
  --choices "redis_cache,db_index,read_write_split,search_engine"

# 读取文件作为 state（@ 前缀）
uv run laya-mcp ask --state @./diff.patch --preset code_change_risk

# 输出完整 JSON（含契约）
uv run laya-mcp ask --state "..." --preset input_guard --json

# 是非问题
uv run laya-mcp ask --state "<发布计划>" \
  --statement "this can ship without human approval"

5/5 answers fail the auto-execute bar (...): ask a human before acting

blast_radius: single_file  (choice, confidence 0.64, margin 0.82, ask_human)
needs_human_review: True   (yes_no, confidence 0.60, margin 0.21, ask_human)

checkpoint: english - English Latin text

# 参考 Dockerfile
FROM ghcr.io/astral-sh/uv:python3.13-bookworm-slim

WORKDIR /app
ENV UV_COMPILE_BYTECODE=1 \
    UV_LINK_MODE=copy \
    LAYAMCP_HOST=0.0.0.0 \
    LAYAMCP_PORT=8077 \
    HF_HOME=/models

COPY pyproject.toml uv.lock README.md ./
RUN uv sync --frozen --no-dev --no-install-project

COPY src ./src
RUN uv sync --frozen --no-dev

EXPOSE 8077
CMD ["uv", "run", "--no-dev", "laya-mcp", "serve"]

docker build -t laya-mcp .
docker run --rm -p 8077:8077 -v laya-models:/models \
  -e LAYAMCP_MODELS=english,multilingual \
  laya-mcp

uv sync
uv run laya-mcp doctor          # 环境自检
uv run laya-mcp tools           # 列出工具（不加载模型，适合快速验证）
uv run pytest -q                # 运行测试

# 仅启动 app，不预热（秒级返回）
LAYAMCP_WARMUP=false uv run python -c "
from fastapi.testclient import TestClient
from laya_mcp.api.app import create_app
with TestClient(create_app()) as c:
    print(c.get('/').json())
    print(c.get('/v1/healthz').json())
"

laya: this checkpoint ships temperatures outside [0.5, 5] ... Treat confidence
from the affected buckets as uncalibrated.

{
  "state": "...",
  "questions": {
    "priority": {
      "kind": "choice",
      "instructions": "这条需求应该排什么优先级？",
      "criteria": {
        "p0": "线上故障，立刻处理",
        "p1": "本周内必须完成",
        "p2": "可以排入下个迭代"
      }
    }
  }
}