# -*- coding: utf-8 -*- """ demo_task.py — Pi Agent 兜底能力 P0 PoC：演示任务驱动脚本 ========================================================== 两种模式： - 默认（mock 模式，无需 LLM）：准备一个真实订单 CSV（demo-data/锐扬APS演示数据/ 05_orders.csv）拷入 runs//inbox/；生成《执行计划草稿》→ 模拟确认 → 用**纯 Python 的 mock executor**（产出与真实 pi 相同结构的事件流，因此走编排器 同一条熔断/日志/stopReason 判定路径）模拟 Pi 的步骤：读 inbox 数据 → 分析 → 写报告到 outbox（报告引用 bridge 签发的 callId 凭证）。随后跑 3 项围墙自检： 熔断（步数闸）/ 凭证（伪造 callId）/ 圈禁（路径越界）。 - `--real`：走真实 Pi headless 路径（Agent-C 用）： `node runtime/.../cli.js -p --mode json --model --tools read,grep,find,ls -e <守卫扩展> "任务"`。无真实 LLM 网关时预期 stopReason=error 显式失败 （这正是本 PoC 要证明的「失败不包装成成功」）。 用法（Git Bash，工作区根目录）： .venv/Scripts/python.exe poc/pi-fallback/demo_task.py # mock 全流程 + 自检 .venv/Scripts/python.exe poc/pi-fallback/demo_task.py --real # 真实 pi（需模型网关） .venv/Scripts/python.exe poc/pi-fallback/demo_task.py --real --model aps-local/qwen3-32b-local --timeout 120 """ from __future__ import annotations import argparse import csv import json import random import shutil import sys import time from pathlib import Path # 允许以脚本方式直接运行（poc/pi-fallback 加入 sys.path） sys.path.insert(0, str(Path(__file__).resolve().parent)) from orchestrator import BreakerConfig, Orchestrator from sandbox import (POC_ROOT, SandboxViolation, create_run_dirs, new_run_id, validate_path, write_guard_extension) from tool_bridge import ToolBridge, ToolBridgeViolation # --------------------------------------------------------------------------- # 集中配置 # --------------------------------------------------------------------------- WORKSPACE_ROOT = POC_ROOT.parent.parent # aps-agent 工作区根 DEMO_CSV = WORKSPACE_ROOT / "demo-data" / "锐扬APS演示数据" / "05_orders.csv" FALLBACK_ROWS = 100 # 找不到素材时生成的模拟工单行数（明确标注模拟数据） REPORT_NAME = "report.md" # 产物唯一出口文件名（outbox/） PLAN_NAME = "执行计划草稿.md" # --------------------------------------------------------------------------- # 素材准备 # --------------------------------------------------------------------------- def prepare_inbox(inbox: Path) -> tuple[Path, str]: """ 把数据分析素材拷入 inbox（L2 可读注入区）。 返回 (素材路径, 来源说明)。找不到真实素材时生成 100 行模拟工单数据并明确标注。 """ if DEMO_CSV.exists(): dst = inbox / DEMO_CSV.name shutil.copy2(DEMO_CSV, dst) return dst, f"真实演示数据（拷贝自 {DEMO_CSV}）" # 兜底：生成模拟工单数据 dst = inbox / "simulated_workorders.csv" rng = random.Random(42) with open(dst, "w", newline="", encoding="utf-8-sig") as f: w = csv.writer(f) w.writerow(["woNo", "orderNo", "productCode", "quantity", "dueDate", "priority", "status"]) for i in range(1, FALLBACK_ROWS + 1): w.writerow([ f"WO-{i:04d}", f"SO-{rng.randint(1, 30):03d}", f"P{rng.randint(1000, 9999)}", rng.randint(5, 200), f"2026-08-{rng.randint(1, 28):02d}", rng.randint(1, 6), rng.choice(["RELEASED", "IN_PROGRESS", "DONE"]), ]) return dst, "【模拟数据】未找到 demo-data 素材，脚本生成 100 行模拟工单" # --------------------------------------------------------------------------- # mock executor：模拟 Pi 会做的步骤（读 inbox → 分析 → 写 outbox 报告） # --------------------------------------------------------------------------- def make_mock_runner(run_id: str, run_dir: Path, inbox_csv: Path, data_source_note: str, bridge: ToolBridge): """ 构造 mock AgentRunner。事件结构与真实 pi headless JSONL 一致 （agent_start → tool_execution_start/end × N → message_end(stopReason) → agent_end）， 因此编排器的熔断/日志/stopReason 判定是同一条代码路径。 每个"工具调用"都经 ToolBridge 签发 callId 并落 calls.jsonl， 报告里的成果声明逐条附 [callId: ...] 凭证。 """ outbox = run_dir / "outbox" def runner(task: str, work_dir: Path): yield {"type": "session", "id": run_id, "mode": "mock"} yield {"type": "agent_start"} yield {"type": "turn_start"} evidence: list[tuple[str, str]] = [] # (成果声明, callId) def step(tool: str, params: dict, do): """一次模拟工具调用：签发 callId → 发 start 事件 → 执行 → 发 end 事件 → 落账。""" call_id = bridge.issue_call(tool, params) yield {"type": "tool_execution_start", "toolName": tool, "toolCallId": call_id, "input": params} try: result = do() bridge.complete_call(call_id, result, ok=True) except Exception as exc: bridge.complete_call(call_id, exc, ok=False) raise yield {"type": "tool_execution_end", "toolName": tool, "toolCallId": call_id, "isError": False} return call_id, result # -- 步骤 1：fs_read 读 inbox 数据 ---------------------------------- def do_read(): # 读路径也过 L2 校验（mock 侧自觉走围墙） safe = validate_path(inbox_csv, run_dir, zone="inbox", must_exist=True) with open(safe, encoding="utf-8-sig", newline="") as f: return list(csv.DictReader(f)) cid_read, rows = yield from _wrap(step("fs_read", {"path": str(inbox_csv)}, do_read)) evidence.append((f"读取注入区数据 {inbox_csv.name}，共 {len(rows)} 行", cid_read)) # -- 步骤 2：aps_query 做只读统计 ----------------------------------- def do_stats(): by_status: dict = {} by_level: dict = {} qty_total = 0 high_pri = 0 for r in rows: by_status[r.get("status", "?")] = by_status.get(r.get("status", "?"), 0) + 1 lvl = r.get("customerLevel", "?") if lvl: by_level[lvl] = by_level.get(lvl, 0) + 1 try: qty_total += int(r.get("quantity") or 0) except ValueError: pass try: if int(r.get("priority") or 0) >= 5: high_pri += 1 except ValueError: pass return {"rows": len(rows), "by_status": by_status, "by_customer_level": by_level, "qty_total": qty_total, "high_priority": high_pri} cid_stats, stats = yield from _wrap(step("aps_query", {"sql_like": "select status,count(*) group by status"}, do_stats)) evidence.append((f"订单状态分布 {stats['by_status']}", cid_stats)) evidence.append((f"订单总量 {stats['qty_total']} 件，优先级>=5 的 {stats['high_priority']} 单", cid_stats)) # -- 步骤 3：fs_write + report_emit 写报告到 outbox ------------------ def do_write(): lines = [ "# Pi 兜底分析报告（mock 模式产物）", "", f"- runId: `{run_id}`", f"- 任务: {task}", f"- 数据来源: {data_source_note}", "", "## 成果声明（每条附 callId 凭证）", "", ] for claim, cid in evidence: lines.append(f"- {claim} [callId: {cid}]") lines += [ "", "## 结论建议", "", "优先级 >= 5 的订单建议优先排产；B 级客户订单量大可关注交付风险。", "（本报告由 mock executor 生成，用于验证编排/围墙/凭证链路，非真实 LLM 产出）", ] target = validate_path(outbox / REPORT_NAME, run_dir, zone="outbox") target.write_text("\n".join(lines) + "\n", encoding="utf-8") return str(target) cid_write, report_path = yield from _wrap(step("fs_write", {"path": str(outbox / REPORT_NAME)}, do_write)) yield from _wrap(step("report_emit", {"md": report_path}, lambda: report_path)) report_text = Path(report_path).read_text(encoding="utf-8") # -- 收尾事件：stopReason=stop（编排器据此判成功） -------------------- assistant_msg = { "role": "assistant", "stopReason": "stop", "content": [{"type": "text", "text": report_text}], } yield {"type": "message_end", "message": assistant_msg} yield {"type": "turn_end"} yield {"type": "agent_end", "messages": [assistant_msg]} # step() 是生成器函数包装：_wrap 让 "执行 + yield 事件" 能 return 值 return runner def _wrap(gen): """驱动 step 生成器，透传 yield 的事件，取回其 return 值。""" value = yield from gen return value # --------------------------------------------------------------------------- # 围墙自检（mock 模式下验证熔断/凭证/圈禁各一条） # --------------------------------------------------------------------------- def self_checks() -> dict: """ 3 项自检，对应 GOAL.md 的 3 个越狱测试判据： 1) 熔断：步数闸触发 → 运行显式 failed； 2) 凭证：报告引用伪造 callId → validate 判 invalid； 3) 圈禁：越界路径（../../server）→ validate_path 拒绝。 """ results = {} # 1) 熔断：步数上限 def runaway_runner(task, work_dir): yield {"type": "agent_start"} for i in range(20): # 远超 max_steps=3 yield {"type": "tool_execution_start", "toolName": "fs_read", "toolCallId": f"runaway-{i}", "input": {}} yield {"type": "tool_execution_end", "toolName": "fs_read", "toolCallId": f"runaway-{i}", "isError": False} yield {"type": "agent_end", "messages": []} orch = Orchestrator(BreakerConfig(max_steps=3, timeout_sec=60)) r = orch.run("runaway self-check", runaway_runner, run_id=new_run_id("selfcheck-breaker")) results["breaker_step_limit"] = { "pass": (not r.success) and "max_steps" in r.stop_reason, "detail": f"success={r.success} stop_reason={r.stop_reason} steps={r.steps}", } # 2) 凭证：伪造 callId 必须判 invalid run_id = new_run_id("selfcheck-credential") dirs = create_run_dirs(run_id) bridge = ToolBridge(run_id, dirs["root"]) real_cid = bridge.issue_call("fs_read", {"path": "inbox/x.csv"}) bridge.complete_call(real_cid, "ok") good = bridge.validate_report_citations(f"成果 A [callId: {real_cid}]") forged = "call-" + "deadbeef-0000-0000-0000-000000000000" bad = bridge.validate_report_citations(f"成果 B [callId: {forged}]") results["credential_forgery"] = { "pass": good["valid"] and (not bad["valid"]) and bad["missing"] == [forged], "detail": f"真凭证 valid={good['valid']}；伪造凭证 valid={bad['valid']} missing={bad['missing']}", } # 3) 圈禁：越界路径必须拒绝 run_id = new_run_id("selfcheck-sandbox") dirs = create_run_dirs(run_id) checks = [] try: validate_path("../../server/evil.py", dirs["root"]) checks.append("越界未拦截!") except SandboxViolation: checks.append("越界已拦截") try: validate_path(dirs["outbox"] / "ok.md", dirs["root"], zone="outbox") checks.append("界内放行") except SandboxViolation as e: checks.append(f"界内误杀: {e}") results["sandbox_escape"] = { "pass": checks == ["越界已拦截", "界内放行"], "detail": "; ".join(checks), } return results # --------------------------------------------------------------------------- # 主流程 # --------------------------------------------------------------------------- def run_mock(task: str) -> int: """mock 模式完整流程：建目录 → 备料 → 计划草稿 → 模拟确认 → 编排执行 → 凭证校验 → 自检。""" run_id = new_run_id() dirs = create_run_dirs(run_id) run_dir = dirs["root"] print(f"[demo] run_id={run_id}") inbox_csv, source_note = prepare_inbox(dirs["inbox"]) print(f"[demo] inbox 素材: {inbox_csv.name}（{source_note}）") # —— 产出《执行计划草稿》（方案 §4.4 agent.fallback.propose，沙盒语义不写主干）—— plan = f"""# 执行计划草稿（propose 阶段产物） - runId: `{run_id}` - 任务: {task} - 数据来源: {source_note} ## 计划步骤（卡片上的承诺即执行上限） | 步 | 工具 | 权力 | 动作 | |----|------|------|------| | 1 | fs_read | P0 | 读 inbox/{inbox_csv.name} | | 2 | aps_query | P0 | 状态/优先级/客户等级只读统计 | | 3 | fs_write | P1 | 写报告到 outbox/{REPORT_NAME} | | 4 | report_emit | P1 | 报告走凭证校验出口 | ## 预计影响面 只读 inbox，仅写 outbox/{REPORT_NAME}；不触主干数据；无需 checkpoint 回滚。 """ (run_dir / PLAN_NAME).write_text(plan, encoding="utf-8") # —— 模拟确认（产品态为 P2 确认卡；PoC 落一个确认记录文件）—— (run_dir / "确认记录.md").write_text( f"# 模拟确认\n\nrunId: {run_id}\n计划已（模拟）人工确认，进入 execute 阶段。\n", encoding="utf-8") print(f"[demo] 《{PLAN_NAME}》已生成并模拟确认") # —— 编排执行（mock executor 走编排器同一条熔断/日志/判定路径）—— bridge = ToolBridge(run_id, run_dir) runner = make_mock_runner(run_id, run_dir, inbox_csv, source_note, bridge) orch = Orchestrator(BreakerConfig()) # 默认三闸：10min / 50 步 / 10MiB t0 = time.time() res = orch.run(task, runner, run_id=run_id) print(f"[demo] 编排完成: success={res.success} stop_reason={res.stop_reason} " f"steps={res.steps} output={res.output_bytes}B elapsed={res.elapsed_sec:.2f}s") # —— 成果校验：报告引用的 callId 必须全部真实存在（方案 §4.5）—— report_path = dirs["outbox"] / REPORT_NAME if res.success and report_path.exists(): v = bridge.validate_report_citations(report_path.read_text(encoding="utf-8")) print(f"[demo] 凭证校验: valid={v['valid']} cited={len(v['cited'])} " f"issued={len(v['issued'])} missing={v['missing']}") if not v["valid"]: print("[demo] ⚠️ 报告含无凭证的成果声明 → 判 invalid", file=sys.stderr) return 2 else: print("[demo] ⚠️ 编排失败或无报告产物", file=sys.stderr) return 2 # —— 围墙自检 —— checks = self_checks() all_pass = True print("[demo] 围墙自检:") for name, r in checks.items(): print(f" [{'PASS' if r['pass'] else 'FAIL'}] {name}: {r['detail']}") all_pass = all_pass and r["pass"] print(f"[demo] 运行目录: {run_dir}") print(f"[demo] 报告产物: {report_path}") print(f"[demo] 总耗时 {time.time() - t0:.1f}s，自检 {'全部通过' if all_pass else '存在失败'}") return 0 if all_pass else 3 def run_real(task: str, model: str, timeout: float) -> int: """--real 模式：真实 pi headless。无 LLM 网关时预期 stopReason=error 显式失败。""" run_id = new_run_id("real") dirs = create_run_dirs(run_id) inbox_csv, source_note = prepare_inbox(dirs["inbox"]) guard = write_guard_extension(dirs["root"]) print(f"[real] run_id={run_id} 素材={inbox_csv.name}（{source_note}）") print(f"[real] 守卫扩展: {guard}") orch = Orchestrator(BreakerConfig(timeout_sec=timeout)) runner = orch.pi_headless_runner(model=model, tools="read,grep,find,ls", guard_ext=guard) full_task = (f"{task}\n数据文件在 inbox/{inbox_csv.name}（相对当前目录上一级），" f"分析报告写到 outbox/{REPORT_NAME}。") res = orch.run(full_task, runner, run_id=run_id) print(f"[real] 结果: success={res.success} stop_reason={res.stop_reason} " f"steps={res.steps} elapsed={res.elapsed_sec:.1f}s") if res.error_message: print(f"[real] error: {res.error_message[:300]}") print(f"[real] 运行目录: {dirs['root']}（events.jsonl / orchestrator.log / result.json）") # 显式失败也是成功交付：返回码只在 harness 自身错误时非零 return 0 def main() -> int: ap = argparse.ArgumentParser(description="Pi 兜底 PoC 演示任务驱动") ap.add_argument("--real", action="store_true", help="走真实 Pi headless（需模型网关）") ap.add_argument("--model", default="aps-local/qwen3-32b-local", help="--real 用模型") ap.add_argument("--timeout", type=float, default=600.0, help="超时秒数（默认 600）") ap.add_argument("--task", default="分析注入区的订单数据：状态分布、优先级分布、客户等级结构，给出排产建议", help="任务简报") args = ap.parse_args() if args.real: return run_real(args.task, args.model, args.timeout) return run_mock(args.task) if __name__ == "__main__": sys.exit(main())