feat(xingbang): 兴邦金租 AI 评审平台模块(构建于 mini-OpenClaw)

新增兴邦金租评审项目模块, 作为 mini-OpenClaw 之上的完整交付: - 五智能体流水线(data_curator/risk_scout/case_finder/approval_advisor/risk_monitor) - evolution 自进化子系统(轨迹采集+奖励计算+模式归纳prompt增强) - /api/xingbang/* 端点(config/workflow/stream/evolution/trajectories) - 运行时产物(客户产出情报卡 + 进化轨迹数据)已 gitignore, 不含隐私遵循兴邦三档分级审批制度, 数据源指向兴邦实验数据包
2026-09-01 16:58:44 +08:00
parent 42ce992a0f
commit c4ac5aa82e
11 changed files with 1111 additions and 0 deletions
@@ -0,0 +1,229 @@
+"""XingbangWorkflow — 五智能体审批流水线 + 自进化闭环。
+
+执行流程:
+  data_curator → risk_scout → case_finder → approval_advisor → risk_monitor
+  (case_finder 与 risk_scout 可并行, 此处先串行保证可追溯)
+
+每个 agent 的产出写入输出目录, 供下游读取; 同时全程记录
+轨迹到 evolution 模块, 跑完后计算奖励并更新进化洞察。
+
+API 风格: run() 返回完整结果; 也可用 run_stream() 逐步 yield 进度。
+"""
+
+from __future__ import annotations
+
+import time
+import uuid
+from pathlib import Path
+from typing import Any, AsyncGenerator
+
+from xingbang.config import agent_specs
+from xingbang.agents import build_agents, XingbangAgent
+from xingbang.evolution.trajectory import TrajectoryRecorder
+from xingbang.evolution.reward import compute_reward_v1
+from xingbang.evolution.optimizer import EvolutionEngine
+
+
+class XingbangWorkflow:
+    """兴邦五智能体审批工作流。"""
+
+    def __init__(
+        self,
+        llm,
+        base_dir: Path,
+        output_dir: Path | None = None,
+        evolution_enabled: bool = True,
+    ):
+        self._base_dir = Path(base_dir)
+        self._output_dir = output_dir or (self._base_dir / "xingbang_output")
+        self._output_dir.mkdir(parents=True, exist_ok=True)
+
+        # 自进化模块
+        self.evolution_enabled = evolution_enabled
+        self._recorder = TrajectoryRecorder(self._base_dir, enabled=evolution_enabled)
+        self._evolution = EvolutionEngine(self._base_dir, self._recorder, enabled=evolution_enabled)
+
+        # 五智能体(注入证据回调 → 轨迹)
+        self._agents: dict[str, XingbangAgent] = build_agents(
+            llm=llm,
+            base_dir=self._output_dir,
+            output_dir=self._output_dir,
+            on_evidence=self._on_evidence,
+        )
+
+        # 流水线顺序
+        self._pipeline = ["data_curator", "risk_scout", "case_finder",
+                          "approval_advisor", "risk_monitor"]
+        self._results: dict[str, dict] = {}
+
+    def _on_evidence(self, kind: str, data: dict):
+        # 轨迹采集(保持简单: 在 workflow 层显式 add_step)
+        pass
+
+    @property
+    def recorder(self) -> TrajectoryRecorder:
+        return self._recorder
+
+    @property
+    def evolution(self) -> EvolutionEngine:
+        return self._evolution
+
+    # ── 单次评审 ────────────────────────────────────────
+    async def run(self, task: str, customer: str = "") -> dict[str, Any]:
+        """运行完整五智能体审批流水线,返回各步产出 + 轨迹 + 奖励。"""
+        workflow_id = f"wf_{int(time.time())}_{uuid.uuid4().hex[:4]}"
+        traj_id = self._recorder.begin(workflow_id=workflow_id, customer=customer)
+
+        # 第一步任务描述
+        tasks = self._build_tasks(task)
+        final_output = ""
+        summary_parts = []
+
+        for key in self._pipeline:
+            agent = self._agents[key]
+            agent_task = tasks.get(key, task)
+
+            # 收集上游上下文(该 agent 依赖的产出)
+            ctx = self._collect_context(key, agent_task)
+
+            t0 = time.time()
+            try:
+                result = await agent.run(agent_task, context=ctx)
+                success = True
+                output = result["output"]
+                file = result.get("file")
+            except Exception as e:  # noqa
+                success = False
+                output = f"[执行失败] {e}"
+                file = None
+
+            elapsed_ms = int((time.time() - t0) * 1000)
+
+            # 记录这一步到轨迹
+            step = {
+                "agent": key,
+                "name": agent.spec.name,
+                "input": agent_task[:300],
+                "tools": ["read_xingbang_data", "read_output"],
+                "output_file": str(file) if file else None,
+                "output_preview": output[:200],
+                "success": success,
+                "time_ms": elapsed_ms,
+            }
+            self._recorder.add_step(step)
+            self._results[key] = {"output": output, "file": file, "success": success}
+            summary_parts.append(f"[{key}] {'✅成功' if success else '❌失败'}")
+            final_output = output  # 最后一环(风险哨兵)作为总结论
+
+        # 收尾: 计算奖励 + 保存进化洞察
+        trajectory = self._recorder.finalize(
+            success=all(self._results[k]["success"] for k in self._pipeline),
+            summary=" | ".join(summary_parts),
+        )
+        reward = compute_reward_v1(trajectory)
+        trajectory["cumulative_reward"] = reward
+        insights_path = self._evolution.save_insights()
+
+        return {
+            "workflow_id": workflow_id,
+            "trajectory_id": traj_id,
+            "customer": customer,
+            "reward": reward,
+            "results": {k: {kk: vv for kk, vv in v.items() if kk in ("output", "file", "success")}
+                        for k, v in self._results.items()},
+            "final_summary": summary_parts,
+            "evolution_insights_file": str(insights_path),
+        }
+
+    # ── 逐步流式(供前端展示进度) ───────────────────────
+    async def run_stream(self, task: str, customer: str = "") -> AsyncGenerator[dict, None]:
+        """逐步 yield 每步进度与产出,便于前端 SSE 展示。"""
+        workflow_id = f"wf_{int(time.time())}_{uuid.uuid4().hex[:4]}"
+        traj_id = self._recorder.begin(workflow_id=workflow_id, customer=customer)
+        tasks = self._build_tasks(task)
+
+        for key in self._pipeline:
+            agent = self._agents[key]
+            agent_task = tasks.get(key, task)
+            ctx = self._collect_context(key, agent_task)
+            yield {"event": "agent_start", "agent": key, "name": agent.spec.name}
+
+            t0 = time.time()
+            try:
+                result = await agent.run(agent_task, context=ctx)
+                success = True
+                output = result["output"]
+                file = result.get("file")
+            except Exception as e:  # noqa
+                success = False
+                output = f"[执行失败] {e}"
+                file = None
+            elapsed_ms = int((time.time() - t0) * 1000)
+
+            step = {
+                "agent": key, "name": agent.spec.name, "input": agent_task[:300],
+                "tools": ["read_xingbang_data", "read_output"],
+                "output_file": str(file) if file else None,
+                "output_preview": output[:200], "success": success, "time_ms": elapsed_ms,
+            }
+            self._recorder.add_step(step)
+            self._results[key] = {"output": output, "file": file, "success": success}
+            yield {"event": "agent_end", "agent": key, "name": agent.spec.name,
+                   "success": success, "output": output, "file": str(file) if file else None}
+
+        trajectory = self._recorder.finalize(success=True)
+        reward = compute_reward_v1(trajectory)
+        trajectory["cumulative_reward"] = reward
+        insights_path = self._evolution.save_insights()
+        yield {"event": "done", "workflow_id": workflow_id,
+               "trajectory_id": traj_id, "reward": reward,
+               "insights_file": str(insights_path)}
+
+    # ── 辅助 ────────────────────────────────────────────
+    def _build_tasks(self, user_task: str) -> dict[str, str]:
+        """为每个智能体构造针对性的任务描述。"""
+        return {
+            "data_curator": (
+                f"请读取兴邦数据包中的多源数据, 为客户『{user_task}』(申请评审) "
+                f"生成结构化客户情报卡, 保存到 customer_intel.md。"
+            ),
+            "risk_scout": (
+                f"基于上游 customer_intel.md, 对客户『{user_task}』做尽调分析, "
+                f"生成尽调报告初稿, 保存到 due_diligence.md。"
+            ),
+            "case_finder": (
+                f"检索历史案例库与制度规则, 为『{user_task}』这类项目找出最相关案例, "
+                f"生成历史案例参考, 保存到 case_reference.md。"
+            ),
+            "approval_advisor": (
+                f"汇总情报卡/尽调/历史案例, 按三档分级制度为『{user_task}』"
+                f"输出审批参考意见, 保存到 approval_opinion.md。"
+            ),
+            "risk_monitor": (
+                f"审批前动态复核最新风控数据, 为『{user_task}』输出风控复核结论, "
+                f"保存到 risk_recheck.md。"
+            ),
+        }
+
+    def _collect_context(self, agent_key: str, task: str) -> str:
+        """收集该 agent 依赖的上游产出作为上下文。"""
+        spec = agent_specs()[agent_key]
+        parts = []
+        for dep in spec.inputs_from:
+            res = self._results.get(self._dep_owner(dep))
+            if res and res.get("output"):
+                parts.append(f"[{dep}]\n{res['output'][:2000]}")
+        if parts:
+            return "\n\n".join(parts)
+        return ""
+
+    def _dep_owner(self, output_file: str) -> str:
+        """根据产出文件名反查产生它的 agent key。"""
+        mapping = {
+            "customer_intel.md": "data_curator",
+            "due_diligence.md": "risk_scout",
+            "case_reference.md": "case_finder",
+            "approval_opinion.md": "approval_advisor",
+            "risk_recheck.md": "risk_monitor",
+        }
+        return mapping.get(output_file, "data_curator")