WisePick | 智选

The Deterministic Scaffold for Agentic Capability Routing | 面向 Agent 能力路由的确定性脚手架

🚀 WisePick Decision API (WPDA) v0.1.5

WisePick does not recommend apps to humans. It routes executable capabilities to agents at 0.0s latency.
智选不向人类推荐应用；它为 Agent 提供 0.0s 延迟的确定性决策路由。

🛡️ Decision Infrastructure | 决策基础设施

From intent → one executable capability unit (ECU).

输入意图 → 单一可执行能力单元（ECU）。

Executable Capability Unit (ECU)

A standardized executable capability an agent can route, invoke, and learn from.

可执行能力单元（ECU）：可被路由、调用并通过反馈学习的标准化能力抽象。

❓ What Problem It Solves | 解决什么问题

Most agents fail from poor capability routing, not weak models.

多数失败来自能力路由失准，而非模型能力不足。

• Blind capability search · 盲目遍历能力
• Trial-and-error execution · 反复试错执行
• No execution feedback loop · 缺少反馈闭环

WisePick replaces guessing with learned routing.

智选用数据驱动的能力路由替代猜测。

🚀 Quick Start | 快速启动

Configure DATABASE_URL in .env (see .env.example). Full API and ops detail: README_API.md.

在 .env 中配置 DATABASE_URL（参考 .env.example）。完整接口与运维说明见 README_API.md。

pip install -r requirements.txt

uvicorn app.main:app --reload --host 0.0.0.0 --port 8000

Smoke test:

curl -s http://localhost:8000/health

⚡ Why Integrate WisePick | 为什么接入智选

• Lower cost and latency by cutting trial-and-error.
  更低成本与延迟：减少无效试错与 Token 浪费。

• Deterministic selection: one best ECU per decision.
  确定性输出：每次决策对应单一最优 ECU。

• Self-evolving loop: routing improves from execution feedback.
  自进化闭环：执行反馈持续修正路由统计。

📜 Performance | 性能报告

Production-oriented routing: under high load, sub-millisecond average latency on the isolated scaffold. Benchmarks: STRESS_TEST_RESULTS.md.

面向生产的确定性路由；隔离脚手架下平均亚毫秒级延迟。基准数据见 STRESS_TEST_RESULTS.md。

🔌 Integration | 集成接入

WisePick is a stateless decision layer: /v1/decide maps intent → ECU only; it does not retain execution state. Run tools in your runtime, then POST /v1/feedback to close the loop.

无状态决策层： /v1/decide 仅完成意图 → ECU 映射；不保存执行状态。能力在你侧执行后，通过 /v1/feedback 回传结果。

Prerequisite: Deploy and run the API on your infrastructure (deployment guide).

前提条件： 自托管部署并完成启动（README_API.md — 部署与运行）。

import requests

BASE_URL = "http://localhost:8000"  # replace with your host

decision = requests.post(
    f"{BASE_URL}/v1/decide",
    json={"task": "Generate a technical summary"},
).json()

# Execute locally: your_agent.run(decision["capability_id"], decision["provider"])

requests.post(
    f"{BASE_URL}/v1/feedback",
    json={"decision_id": decision["decision_id"], "success": True},
)

🤖 Agent Runtime Integration | 运行时集成

Machine-readable contract and runtime loop: AGENTS.md.

机器可读的集成语义与运行时闭环见 AGENTS.md。

🧠 How It Works | 工作原理

Capability Matching | 能力匹配

Task text → capability labels derived from bootstrap rules.

任务文本 → 由引导规则得到能力标签。

task → capabilities

Capability Scoring | 能力评分

score =
capability_match       * 0.70  (语义匹配度)
execution_success_rate * 0.20  (历史成功率)
bootstrap_weight       * 0.10  (初始权重 / 冷启动偏好)

Optional YantrikDB | 可选 YantrikDB

Enterprise cluster awareness · 企业级集群感知

Optional integration via YANTRIK_DB_URL (and optional YANTRIK_DB_API_KEY): reads YantrikDB /v1/health, may scale ECU scores under high replication lag—no primary schema change.

可选接入：读取 YantrikDB /v1/health，复制滞后过高时可缩放 ECU 分数；不修改主库 Schema。

Feedback Loop | 反馈闭环

decision → execution → feedback → capability_stats → next decision

Routing updates from real execution outcomes.

路由统计随真实执行结果更新。

Components | 核心组件

• Routing core (decision_engine) — Task → ECU scoring and selection. · 路由核心 — 任务评分与 ECU 选择。
• Capability registry (api_tool_specs) — Enabled providers, capability tags, bootstrap weights. · 能力注册表 — 可用 provider、标签与冷启动权重。
• Execution memory (tool_stats, feedback) — Success rates and outcomes for closed-loop learning. · 执行记忆 — 成功率与反馈闭环。

🦜 Semantic Upgrade | 语义升级

WisePick evolved from tool selection → executable capability routing.

演进路径：从「选工具」到「可执行能力路由」。

🔬 Example ECU Response | ECU 响应示例

{
  "decision_id": "xxx",
  "capability_id": "audio_transcription",
  "execution_type": "api",
  "provider": "feishu_minutes",
  "tool_key": "feishu_minutes",
  "confidence": 0.87,
  "callable": true
}

Program against capabilities, not product names.

对「能力」编程，不对「产品名」编程。

🧪 Agent Workflow | Agent 工作流

Ask WisePick for routing        → 请求路由
Receive ECU                     → 获取 ECU
Map ECU → local API / MCP       → 映射到本地实现
Execute                         → 执行
Send feedback                   → 回传反馈

WisePick provides decision, routing, and execution learning—not task execution.

智选提供决策、路由与执行侧学习信号；不替代任务执行本身。

🔮 Vision | 愿景

Today: Local execution learning.
Tomorrow: Shared decision memory.

当下： 本地执行反馈驱动学习。
下一步： 共享决策记忆。

Execution outcomes become portable capability experience—not repeated trial and error.

执行结果沉淀为可迁移的能力经验，而非重复试错。

🗺️ Roadmap | 路线图

• ✅v0.1: Core Capability Routing · 核心路由层实现
  Sub-millisecond isolated core latency.
  ECU protocol & feedback loop logic.
• 🔄v0.2: Agentic Workflow Routing · 复杂 Agent 流转路由支持（从单点路由向多步协同演进）
• 🔄v0.3: Collective Decision Memory · 集体决策记忆（让真实的执行结果沉淀为可复用的路由经验）
• 🔄Ongoing: ECU Ecology · 持续扩展主流 MCP 与 API 能力库，构建最全的可执行能力索引

🤗 Feedback & Integration | 反馈与集成

Share use cases, routing results, or failure reports.

欢迎反馈接入场景、路由结果或失败案例。

• Issues: GitHub Issues
• Email: [email protected]

Every routing decision is observable, feedback-driven, and reproducible.
每一次路由决策可观测、可反馈、可复现。

Every decision sharpens the path to perfect agency.
每一次决策，都在打磨通往完美能动性的路径。˗ˋˏ( ´͈ ᗜ `͈ )ˎˊ˗