enterprise-agent-designer 0.34.1

package/scripts/verify_v0321_guards.py

@@ -0,0 +1,410 @@
+#!/usr/bin/env python3
+"""Regression checks for v0.32 review, workflow, and evidence guards."""
+
+from __future__ import annotations
+
+import argparse
+import copy
+import importlib.util
+import json
+import subprocess
+import sys
+import tempfile
+from pathlib import Path
+from types import SimpleNamespace
+
+
+ROOT = Path(__file__).resolve().parents[1]
+
+
+def load_module(name: str, path: Path):
+    spec = importlib.util.spec_from_file_location(name, path)
+    if spec is None or spec.loader is None:
+        raise RuntimeError(f"cannot load module: {path}")
+    module = importlib.util.module_from_spec(spec)
+    spec.loader.exec_module(module)
+    return module
+
+
+def validate_receipt(module, receipt: dict[str, object], expected_snapshot: str | None = None) -> list[str]:
+    mode = str(receipt["review_mode"])
+    args = SimpleNamespace(
+        expected_mode=mode,
+        expected_round=receipt["review_round"],
+        expected_snapshot=expected_snapshot if mode == "source-gate" else None,
+    )
+    errors: list[str] = []
+    module.validate_common(receipt, args, errors)
+    if mode == "grounding-gate":
+        module.validate_grounding(receipt, errors)
+    else:
+        module.validate_source(receipt, args, errors)
+    return errors
+
+
+def all_mock_receipts() -> list[dict[str, object]]:
+    payload = json.loads((ROOT / "evaluation/cases.json").read_text(encoding="utf-8"))
+    receipts: list[dict[str, object]] = []
+    for case in payload["cases"]:
+        for mock in case.get("mock_tools", []):
+            result = mock.get("result")
+            if isinstance(result, dict) and result.get("review_mode") in {
+                "grounding-gate",
+                "source-gate",
+            }:
+                receipts.append(result)
+    return receipts
+
+
+def run_controller(*arguments: str) -> subprocess.CompletedProcess[str]:
+    return subprocess.run(
+        [sys.executable, str(ROOT / "scripts/optional/workflow_controller.py"), *arguments],
+        text=True,
+        capture_output=True,
+        encoding="utf-8",
+    )
+
+
+def main() -> int:
+    receipt_module = load_module(
+        "validate_review_receipt",
+        ROOT / "skills/review-enterprise-agent/scripts/validate_review_receipt.py",
+    )
+    finalizer = load_module(
+        "finalize_agent_delivery",
+        ROOT / "skills/design-enterprise-agent/scripts/finalize_agent_delivery.py",
+    )
+
+    receipts = all_mock_receipts()
+    for receipt in receipts:
+        snapshot = (
+            str(receipt["reviewed_snapshot_id"])
+            if receipt["review_mode"] == "source-gate"
+            else None
+        )
+        errors = validate_receipt(receipt_module, receipt, snapshot)
+        if errors:
+            raise AssertionError(f"mock receipt failed: {errors}")
+
+    round_four = next(
+        item
+        for item in receipts
+        if item["review_mode"] == "source-gate" and item["review_round"] == 4
+    )
+    invalid_late_round = copy.deepcopy(round_four)
+    invalid_late_round["materiality_audit"][
+        "late_round_blocking_justification"
+    ] = "not_applicable"
+    if not validate_receipt(
+        receipt_module,
+        invalid_late_round,
+        str(invalid_late_round["reviewed_snapshot_id"]),
+    ):
+        raise AssertionError("late-round blocker without justification was accepted")
+    material_same_family = copy.deepcopy(round_four)
+    material_same_family["materiality_audit"][
+        "late_round_blocking_justification"
+    ] = "material-same-family-missed-instance"
+    if validate_receipt(
+        receipt_module,
+        material_same_family,
+        str(material_same_family["reviewed_snapshot_id"]),
+    ):
+        raise AssertionError("material same-family late-round blocker was rejected")
+
+    source_non_pass = next(
+        item
+        for item in receipts
+        if item["review_mode"] == "source-gate" and item["verdict"] != "pass"
+    )
+    if not validate_receipt(receipt_module, source_non_pass, "sha256:different"):
+        raise AssertionError("source non-pass with wrong host snapshot was accepted")
+
+    grounding = next(item for item in receipts if item["review_mode"] == "grounding-gate")
+    invalid_grounding = copy.deepcopy(grounding)
+    invalid_grounding["reviewed_snapshot_id"] = "sha256:not-allowed"
+    if not validate_receipt(receipt_module, invalid_grounding):
+        raise AssertionError("grounding receipt with source snapshot was accepted")
+
+    invalid_enum = copy.deepcopy(grounding)
+    invalid_enum["lowest_failed_layer"] = "unknown-layer"
+    invalid_enum["return_to"] = "unknown-target"
+    if not validate_receipt(receipt_module, invalid_enum):
+        raise AssertionError("receipt with invalid routing enums was accepted")
+
+    files = [ROOT / "SYSTEM_PROMPT.md"]
+    mobile = finalizer.normalize_forbidden_evaluation_terms([" 移动端 ", "移动端"])
+    offline = finalizer.normalize_forbidden_evaluation_terms(["离线"])
+    mobile_manifest = finalizer.build_manifest(ROOT, "SYSTEM_PROMPT.md", files, 2, mobile)
+    mobile_manifest_again = finalizer.build_manifest(
+        ROOT,
+        "SYSTEM_PROMPT.md",
+        files,
+        2,
+        finalizer.normalize_forbidden_evaluation_terms(["移动端"]),
+    )
+    offline_manifest = finalizer.build_manifest(ROOT, "SYSTEM_PROMPT.md", files, 2, offline)
+    if mobile_manifest["source_snapshot_id"] != mobile_manifest_again["source_snapshot_id"]:
+        raise AssertionError("equivalent forbidden terms produced different snapshots")
+    if mobile_manifest["source_snapshot_id"] == offline_manifest["source_snapshot_id"]:
+        raise AssertionError("different forbidden terms produced the same snapshot")
+
+    with tempfile.TemporaryDirectory() as temp_dir:
+        temp_root = Path(temp_dir)
+        evaluation = temp_root / "evaluation.md"
+        evaluation.write_text("移动端", encoding="utf-8")
+        errors: list[str] = []
+        finalizer.validate_forbidden_evaluation_terms(
+            [evaluation], temp_root, mobile, errors
+        )
+        if not errors:
+            raise AssertionError("normalized forbidden term did not reach scanner")
+
+    grounding_pass = next(
+        item
+        for item in receipts
+        if item["review_mode"] == "grounding-gate" and item["verdict"] == "pass"
+    )
+    source_pass = next(
+        item
+        for item in receipts
+        if item["review_mode"] == "source-gate" and item["verdict"] == "pass"
+    )
+
+    with tempfile.TemporaryDirectory() as temp_dir:
+        temp_root = Path(temp_dir)
+        state_path = temp_root / "workflow-state.json"
+        goal_path = temp_root / "DESIGN_CONTEXT.md"
+        goal_path.write_text("overall goal", encoding="utf-8")
+        grounding_receipt = temp_root / "grounding-round-1.json"
+        grounding_receipt.write_text(
+            json.dumps(grounding_pass, ensure_ascii=False), encoding="utf-8"
+        )
+
+        for arguments in (
+            ("init", str(state_path), "--overall-goal-ref", str(goal_path)),
+            ("event", str(state_path), "candidate-grounded"),
+            ("event", str(state_path), "grounding-gate-started"),
+            (
+                "gate",
+                str(state_path),
+                "--gate",
+                "grounding-gate",
+                "--receipt",
+                str(grounding_receipt),
+                "--expected-round",
+                str(grounding_pass["review_round"]),
+            ),
+            ("event", str(state_path), "source-written"),
+        ):
+            result = run_controller(*arguments)
+            if result.returncode != 0:
+                raise AssertionError(f"workflow setup failed: {result.stdout}{result.stderr}")
+
+        state = json.loads(state_path.read_text(encoding="utf-8"))
+        if (
+            state.get("workflow_mode") != "reference_only"
+            or state.get("enforcement_level") != "callable_reference_only"
+        ):
+            raise AssertionError("default workflow mode overstated host integration")
+        if state["phase"] != "source_review" or state["allowed_actions"] != [
+            "prepare-source-review"
+        ]:
+            raise AssertionError("source-written did not force source review preparation")
+        if run_controller("check-terminal", str(state_path)).returncode == 0:
+            raise AssertionError("source-written was incorrectly accepted as terminal")
+
+        snapshot = str(source_pass["reviewed_snapshot_id"])
+        manifest = temp_root / "manifest.json"
+        manifest.write_text(
+            json.dumps({"source_snapshot_id": snapshot}), encoding="utf-8"
+        )
+        for arguments in (
+            (
+                "event",
+                str(state_path),
+                "source-review-prepared",
+                "--manifest",
+                str(manifest),
+            ),
+            ("event", str(state_path), "source-gate-started"),
+        ):
+            result = run_controller(*arguments)
+            if result.returncode != 0:
+                raise AssertionError(f"source review setup failed: {result.stdout}{result.stderr}")
+
+        invalid_receipt = temp_root / "source-attempt-invalid.txt"
+        invalid_receipt.write_text("not json", encoding="utf-8")
+        first_invalid = run_controller(
+            "gate",
+            str(state_path),
+            "--gate",
+            "source-gate",
+            "--receipt",
+            str(invalid_receipt),
+            "--expected-round",
+            str(source_pass["review_round"]),
+        )
+        if first_invalid.returncode != 2:
+            raise AssertionError("first invalid receipt did not enter retryable state")
+        state = json.loads(state_path.read_text(encoding="utf-8"))
+        if state["phase"] != "source_review" or state["allowed_actions"] != [
+            "retry-source-gate"
+        ]:
+            raise AssertionError("invalid receipt did not preserve source review retry")
+
+        retry_start = run_controller("event", str(state_path), "source-gate-started")
+        if retry_start.returncode != 0:
+            raise AssertionError("source gate retry was not allowed")
+        second_invalid = run_controller(
+            "gate",
+            str(state_path),
+            "--gate",
+            "source-gate",
+            "--receipt",
+            str(invalid_receipt),
+            "--expected-round",
+            str(source_pass["review_round"]),
+        )
+        if second_invalid.returncode != 2:
+            raise AssertionError("second invalid receipt did not stop normally")
+        state = json.loads(state_path.read_text(encoding="utf-8"))
+        if state["phase"] != "blocked" or state["package_status"] != "not_started":
+            raise AssertionError("invalid receipt exhaustion unlocked packaging")
+
+    with tempfile.TemporaryDirectory() as temp_dir:
+        temp_root = Path(temp_dir)
+        state_path = temp_root / "workflow-state.json"
+        goal_path = temp_root / "DESIGN_CONTEXT.md"
+        goal_path.write_text("overall goal", encoding="utf-8")
+        grounding_receipt = temp_root / "grounding-round.json"
+        grounding_receipt.write_text(
+            json.dumps(grounding_pass, ensure_ascii=False), encoding="utf-8"
+        )
+        source_receipt = temp_root / "source-round.json"
+        source_receipt.write_text(
+            json.dumps(source_pass, ensure_ascii=False), encoding="utf-8"
+        )
+        snapshot = str(source_pass["reviewed_snapshot_id"])
+        manifest = temp_root / "manifest.json"
+        manifest.write_text(
+            json.dumps({"source_snapshot_id": snapshot}), encoding="utf-8"
+        )
+
+        successful_path = (
+            ("init", str(state_path), "--overall-goal-ref", str(goal_path)),
+            ("event", str(state_path), "candidate-grounded"),
+            ("event", str(state_path), "grounding-gate-started"),
+            (
+                "gate",
+                str(state_path),
+                "--gate",
+                "grounding-gate",
+                "--receipt",
+                str(grounding_receipt),
+                "--expected-round",
+                str(grounding_pass["review_round"]),
+            ),
+            ("event", str(state_path), "source-written"),
+            (
+                "event",
+                str(state_path),
+                "source-review-prepared",
+                "--manifest",
+                str(manifest),
+            ),
+            ("event", str(state_path), "source-gate-started"),
+            (
+                "gate",
+                str(state_path),
+                "--gate",
+                "source-gate",
+                "--receipt",
+                str(source_receipt),
+                "--expected-round",
+                str(source_pass["review_round"]),
+            ),
+        )
+        for arguments in successful_path:
+            result = run_controller(*arguments)
+            if result.returncode != 0:
+                raise AssertionError(f"workflow pass path failed: {result.stdout}{result.stderr}")
+
+        mutation = run_controller("event", str(state_path), "source-changed")
+        if mutation.returncode != 0:
+            raise AssertionError("source mutation event failed")
+        state = json.loads(state_path.read_text(encoding="utf-8"))
+        if (
+            state["phase"] != "source_review"
+            or state["source_gate_status"] != "not_attempted"
+            or state["source_snapshot_id"] is not None
+        ):
+            raise AssertionError("source mutation did not invalidate the review grant")
+
+    performance_markers = {
+        "skills/design-enterprise-agent/SKILL.md": [
+            "tool_evidence_status",
+            "contract_only",
+            "callable_or_trace",
+            "runtime_design.performance",
+        ],
+        "skills/design-enterprise-agent/references/runtime-and-integration.md": [
+            "性能证据与 Tool 接口",
+            "not_exposed",
+        ],
+        "skills/design-enterprise-agent/references/41-performance-worked-example.md": [
+            "necessary_path_events",
+            "first_divergence",
+            "使用本案例时的停止条件",
+        ],
+        "skills/review-enterprise-agent/SKILL.md": [
+            "条件核验性能设计与声明",
+            "contract_only",
+            "callable_or_trace",
+        ],
+        "RUNTIME_ASSEMBLY.md": [
+            "41-performance-worked-example.md",
+            "tool_evidence_status = contract_only",
+        ],
+    }
+    for relative, markers in performance_markers.items():
+        text = (ROOT / relative).read_text(encoding="utf-8")
+        for marker in markers:
+            if marker not in text:
+                raise AssertionError(f"performance marker missing: {relative} -> {marker}")
+
+    categories = {
+        case["category"]
+        for case in json.loads((ROOT / "evaluation/cases.json").read_text(encoding="utf-8"))["cases"]
+    }
+    required_performance_categories = {
+        "performance_no_tool_prospective_design",
+        "performance_contract_only_static_analysis",
+        "performance_trace_bound_measured_claim",
+    }
+    if not required_performance_categories.issubset(categories):
+        raise AssertionError("performance evidence-tier cases are incomplete")
+
+    contract_only_case = next(
+        case
+        for case in json.loads((ROOT / "evaluation/cases.json").read_text(encoding="utf-8"))["cases"]
+        if case["category"] == "task_rubric_contract_only_not_blocked"
+    )
+    grounding_receipt = contract_only_case["mock_tools"][0]["result"]
+    if (
+        grounding_receipt["verdict"] != "pass"
+        or grounding_receipt["grounding_audit"]["standard_mode"]
+        != "runtime-input-only"
+    ):
+        raise AssertionError("runtime-input-only rubric was still treated as a universal blocker")
+
+    print(
+        f"V0321_GUARDS_PASS receipts={len(receipts)} "
+        "workflow_cases=3 performance_cases=3 standard_routes=3"
+    )
+    return 0
+
+
+if __name__ == "__main__":
+    raise SystemExit(main())
+

package/skills/design-enterprise-agent/SKILL.md

@@ -0,0 +1,131 @@
+---
+name: design-enterprise-agent
+description: 设计、新建、整体优化、重构、修改、生成或测试企业领域 Agent；从已经通过 Grounding Gate 的首版专业任务建立职业内核、七层联合行为和最小责任拓扑，写出自然高质量的 System Prompt、真正需要的 Skills、诚实的依赖契约与行为评测，并在完整源码任务中完成冻结、Source Gate 和可验证交付。用户只要分析或只读裁决时使用 review-enterprise-agent。
+---
+
+# 设计企业 Agent
+
+Skill-Version: 0.34.1
+
+## 入口与交付责任
+
+本 Skill 把已经充分理解的专业任务编译成真实 Agent。完整源码任务先核对调研交接、原始材料和 Grounding Gate 收据；限定局部修改只核对受影响事实与责任，不因入口检查要求重新取得完整调研收据。当前平台没有已核证的生命周期 Hook；完整交付由 Agent 主动走完，不能把内部自检称为独立通过，也不能把主动执行声称为宿主强制。
+
+交付形态由你裁决，不询问业务用户：
+
+- `full_agent_source`：新建、整体优化、重设计或用户要求完整 Agent，默认交付；
+- `local_asset_edit`：用户只授权修改指定 Prompt、Skill 或相邻资产；
+- `design_only`：用户明确只要方案或禁止写文件。
+
+完整源码至少有独立 System Prompt、每个真实 Skill 的源码、确有必要的依赖契约、正常任务和最高风险任务评测。文件数量不预设。
+
+若当前交接仍存在无法通过安全限制、暂定基线、平台待接入或延期扩展处理的高影响业务决定，返回相应调研 Skill。不要在本 Skill 里另建问卷；也不要因为真实 rubric、API 或校准样本尚未接入而自动退回，只要首版能够诚实按 `contract_only` 或安全限制设计。
+
+## 形成总体设计判断
+
+先独立回答，不沿用旧目录：
+
+- 用户真正需要改善什么业务结果，主要消费者依据结果做什么；
+- 为什么需要 Agent，而不是 Prompt、单一 Skill 或确定性程序；
+- Agent 代表谁观察，对哪一种判断质量负责，误判的实际后果是什么；
+- 正常和最高风险任务中，专家先看什么、怎样比较、哪些证据会改变策略；
+- 旧设计哪些行为已经有效，哪些只是文件或方法假设；
+- 当前理解来自哪些工作经历、产物和例外，用户的表面诉求与实际需要是否存在差别；材料不足以定义专业任务时返回访谈，不用预设方案选择代替需求探索；
+- 推荐的首版边界、主要代价和什么事实会改变方案。
+
+给用户的解释使用业务语言。设计师可以挑战旧稿和用户提出的错误架构，但用户对业务方向和最终取舍有最终决定权。
+
+## 设计职业内核和七层联合行为
+
+选择一个正常任务或联合压力事件，先写可观察行为，再让七层共同控制：
+
+- 底线保护什么，越权后怎样合法继续；
+- 岗位处理什么、转交什么，代表谁判断；
+- 原则怎样比较合法方案，任务目标怎样随证据和授权变化；
+- 专业方法怎样观察、比较、反证和形成结论；
+- Tool 与知识提供什么事实，失败怎样返回 Agent 重判；
+- Output 怎样让消费者行动，责任怎样交接；
+- Trace 怎样支持争议复审和真实能力范围内的恢复。
+
+目标 Agent 必须有职业差异。把岗位名换成相邻岗位后仍成立，说明专业观察和判断方法尚未建立。写出至少一个“表面合格但实质不可用”和一个“表达粗糙但专业上仍可用”的正反比较，让判断标准可以被理解和复审。
+
+## 推导最小责任拓扑
+
+先决定谁形成最终专业结论，再决定是否需要 Skill。Agent 保留总体意图、跨 Skill 综合和最终责任；Skill 承担可独立触发、可复用、可验收的专业判断闭环；Tool 负责确定读取、计算和动作；Output 投影同一事实；运行环境承担真实存在的状态、调度、审批与恢复；人承担正式政策、授权和最终业务责任。
+
+满足以下全部条件才拆 Skill：
+
+1. 有独立业务意图，而不只是流水线步骤；
+2. 需要自己的专业证据和判断；
+3. 有独立成功、停止和恢复；
+4. 可以单独测试；
+5. 不与 Agent 或其他 Skill 重复形成同一总评、风险或建议。
+
+解析、格式化、消费者投影、纯路由、写入、去重、统计聚合和固定状态转换通常不是 Skill。相似能力优先合并；“职责需要被说明”不等于“职责需要成为 Skill”。
+
+## 设计知识、Tool 和运行接入
+
+逐项说明哪种专业判断依赖哪种知识、现场数据或动作。记录证据状态：
+
+- `connected`：当前环境可调用并有真实返回；
+- `contract_only`：只有有权接口定义，可设计调用语义和模拟行为；
+- `required_not_connected`：能力必要但尚无接口证据；
+- `not_required`：首版不依赖。
+
+没有真实平台文档时，契约只写业务目的、必要语义、权限与副作用、成功／部分成功／失败类别、平台待回答问题和安全降级。不得补端点、字段、错误码、重试次数、阈值、哈希和幂等默认。契约存在只能证明设计路径，不得标成 Tool 已接入或数据已取得。
+
+输出效力也是依赖。若 Agent 要形成总评、等级、评级草案、准入建议或流程状态，必须逐项追溯其名称、维度关系、权重、阈值、映射和批准效力由谁提供；只有分项锚点而没有有权聚合／评级规则时，只能交付分项结论、缺口与冲突，不能因为“草案”二字就自行生成正式等级。
+
+需要设计 Tool 路径、交互前沿、恢复或性能时，读取 [references/runtime-and-integration.md](references/runtime-and-integration.md)。只有用户明确要求性能设计，或真实轨迹显示重复理解、发现、构造、执行时，再读取 [references/41-performance-worked-example.md](references/41-performance-worked-example.md)。记录 `tool_evidence_status = none / contract_only / callable_or_trace` 并据此选择声明强度，没有轨迹不填写真实性能收益；需要时把路径分析写入 `runtime_design.performance`，不把案例数字复制到目标 Agent。
+
+同一岗位有显著不同的任务结构、长任务上下文／恢复问题、模型或 Tool 变化，或用户要求 Harness 适配时，先读取 [references/task-adaptive-runtime.md](references/task-adaptive-runtime.md)。先判断现有路径是否足够，再在稳定责任与当前权限内调整运行方法；四个运行视角不拆成四个 Skill。普通一次性任务不触发该路线。所需依据复用已有运行契约，不额外制造运行档案。
+
+## 编译高质量生产 Prompt 与 Skills
+
+写 Prompt 前读取 [references/cold-start-and-writing.md](references/cold-start-and-writing.md)。生产 Prompt 默认用七个自然中文业务标题承载七层，并在短段落中说明联合控制，但不能露出设计协议章节、L1—L4、内部门禁编号或开发期状态。
+
+完成三遍编译：
+
+1. **职业内核**：身份张力、消费者决定、专业观察、比较方法、误判后果、表面与实质；
+2. **条件运行**：证据、风险、授权和 Tool 返回变化时怎样直接推进、询问、拒绝、转交、降级和恢复；
+3. **管理可信**：质量优先级、边界保护对象、责任交接、自然解释和可复审输出。
+
+再去规格化：删除方法术语、设计痕迹、空洞口号、字段堆叠和对业务员工无意义的内部状态。关键策略先写给模型使用的业务目的、因果依据和边界，再让它结合当前事实推导动作；这些依据必须进入目标 Prompt／Skill，不能只留在设计说明，也不能退化成“完成后解释理由”。只保留真实依赖、安全与易错操作必需的固定顺序。向用户解释的是决定性依据与代价，不是隐藏思维链，也不要求每个动作都写理由。
+
+每个 Skill 写成“业务意图 × 当前状态 → 专业路径”：意图包含结果用途与结论强度，状态来自当前有效证据、成果、冲突和授权。先说明什么决定证据是否充分、动作是否有用，以及误判会损害什么，再给必要的方法与例子；模型可在边界内推导未列出的合法路径，不把例子编译成封闭菜单。相同意图在不同证据下、相同事实用于不同决定时可能需要不同做法；不穷举矩阵，简单确定任务可保持单一路径。
+
+把协作能力编译到目标岗位自身：需求尚未成形时用经历回放、专业追问或对照帮助用户理解问题；事实明确时直接完成；缺事实时最小澄清；有真实取舍时给有依据的建议；实际行动另核授权。不能只要求“少问问题”，也不把所有岗位变成访谈员。用户纠正后识别影响范围，保留有效成果，改正受影响结论并继续原任务；跨会话保存只由真实授权载体承担。
+
+同一决策事实只由有权来源或工具提供，交接保留终局意图、当前有效结果、未决事项及必要入口。不要让 Agent 重新拼装可由确定性工具给出的状态，也不能把工具建议当成行动授权。这些内容进入实际 Prompt／Skill／必要契约，不仅写在设计说明。
+
+## 写入、回读与内部反证
+
+写入前冻结最小资产计划：每个文件的独立责任、为什么不能与其他文件合并、由什么行为验收。`full_agent_source` 写入独立源码根目录；汇总设计稿只能额外生成。
+
+写入顺序以质量为中心：System Prompt → 核心 Skills → 必要契约 → 评测 → README。逐份回读并回答：
+
+- Prompt 是否比设计说明更聪明、更自然；
+- Skill 是否真有独立专业闭环；
+- 正常和最高风险任务是否可以走通；
+- 无权政策、猜测接口和未验证能力是否进入生产资产；
+- 同一结论是否被多处重复形成；
+- 依赖状态是否区分契约、接入和真实返回；
+- 评测是否检查行为，而不是标题和关键词。
+
+行为评测必须可独立回放核心判断：正常任务至少给出足够具体的对象片段或结构化事实、适用锚点／规则、模拟 Tool 返回（如有）和逐项预期证据关系；“有一份文档、有三个锚点、应正确评审”不构成可复演 fixture。
+
+对确有策略分叉或交互承诺的岗位，在上述正常／风险案例内选择最能证伪设计的对照，不额外穷举测试集：同意图不同状态、同状态不同意图，或中途纠正／新证据后的接续。必要时实际运行生成的 Prompt 和 Skill，观察取证、提问、工具动作与终止，而不只让设计师解释应该怎样做。开放访谈看是否沿回答深化理解并能收敛，明确任务看是否免于无效访谈；不按问题数量少、回答短或出现关键词判通过。模拟与真实运行分别记录，未执行保持 not_run。
+
+内部反证发现问题直接修订，但不能签发 Grounding Gate 或 Source Gate 通过。
+
+## Source Gate 与交付
+
+完整源码内部回读后，调用终结器的 `--prepare-source-review` 在源码根目录之外生成冻结清单。只有清单成功并取得 `source_snapshot_id` 后，才能启动新的隔离 `source-gate`；未冻结源码上的评审只能算内部意见，不能保存为正式 Source Gate 轮次。
+
+Source Gate 每轮都重新检查总体目标、职业质量、最小拓扑、权威与证据、生产行文和跨资产一致性。有效 `revision_required` 返回本 Skill，只修订裁决指出的决定性失败族；修订后重新冻结。无效返回由宿主保存为 invalid attempt，不修改源码、不占用正式轮次。
+
+只有绑定当前快照的有效 `pass` 才调用 `finalize_agent_delivery.py`。终结器核对入口、Skill 身份、评测、源码未变化、收据绑定和 ZIP 回读一致性；它不评价职业设计、模型行为或平台可用。
+
+当前默认是 `agent_driven`：源码写入后不要结束，继续冻结、调用新的隔离 Source Gate、按裁决修订、重新冻结，并在有效 `pass` 后运行终结器和回读 ZIP。若真实失败使下一动作不可执行，保存依据并报告阻塞。可选宿主控制参考不进入默认装配；本轮主动完成不能宣称为平台自动闭环。
+
+最终先说明岗位与主设计判断，再列真实源码和 ZIP。分别报告设计质量、源码／包、独立评审、行为证据、工作流装配和平台接入。新增检查已经不会改变这些判断时停止。