qiaoqichaoandClaude Code
501a4d9c9b
覆盖 AC-16/17/18/19/20/38/39/40/41（BR-17/18/19/20/37/38/39、E-10/13/14/15/25）。 新增： - backend/contracts/judgment.py：JudgeProfile/TaskValidation 契约（逐任务逐用途 验收登记 + 冻结 checkpoint）。放在独立契约模块，因为规格要求的 tokenizer/预算 字段会撞上 prescreen.py 既有的"契约层不得出现凭据字样"静态守卫；不放宽该守卫。 - backend/services/judgment_adapters.py：统一适配入口。准入（未登记任务/用途拒绝、 runtime 须独立评测）、冻结权重（unresolved 与不一致均拒绝）、上下文预算 （正文与选项分开计，超限返回 E-25 且零发送）、中文否定/条件例外/来源矛盾/未登记 映射一律落 unknown(inconclusive)，不硬选正面结果。 - backend/fixtures/judgment/：中文提炼依据题集、本地模型原始输出、Jev 厂商响应 fixture（虚构内容，离线契约用）。 - backend/test_judgment_adapters.py：30 项测试。 修改（最小面）： - contracts/prescreen.py：新增 E-25 运行级错误码 JUDGMENT_CONTEXT_EXCEEDED； Judgment 增 actual_checkpoint_ref、JudgmentConfig 增 checkpoint_ref。 - services/judgment_provider.py：映射适配层的策略拒绝异常；归一化保留实际权重。 未做（受授权限制）：没有 Jev 凭据与 Laya 权重，因此只交付离线 fixture 路径， live 传输层仅留注入点，未接入硬件与端口探测。fixture 通过不证明中文效果合格。 验证：backend 全量 unittest 440 项 OK（原 410 项无回归）。 Co-Authored-By: Claude Code <noreply@anthropic.com>
217 lines
11 KiB
Python
217 lines
11 KiB
Python
"""T07-b 专项判断的统一入口：disabled 路径与工具策略守卫。
|
||
|
||
依据：docs/v2/11-judgment-providers.md §3–§6、03-detailed-design.md §9.3；
|
||
BR-09/10/17/18/20、E-06/08/10/13/14/15、AC-09/10/21。
|
||
|
||
本批**只做 disabled 路径**（T07 卡片要求"先 disabled 专项路径"）：真正的 Jev /
|
||
本地适配器属 T08，这里不预造。
|
||
|
||
三条不可让渡的规则：
|
||
1. **关闭就明确关闭**：disabled 时已计划题目记 `skipped + disabled`，运行记录关闭
|
||
原因，绝不静默产生判断，也绝不因为"关了"就伪造高分（BR-09 / AC-21）。
|
||
2. **失败不补分**：服务失效/超时/畸形输出记 `failed + error_code` 与降级记录，
|
||
`answer` 必须为空；不生成默认分值（AC-09 / E-06 / E-15）。
|
||
3. **工具白名单来自冻结的 Skill**：材料正文里写什么都**不能**扩展工具集合
|
||
（BR-10 / AC-10）。工具 = Skill 声明 ∩ 注册表白名单，二者取交，不取并。
|
||
"""
|
||
|
||
from __future__ import annotations
|
||
|
||
from contracts.prescreen import (
|
||
ErrorCode,
|
||
Judgment,
|
||
JudgmentAnswer,
|
||
JudgmentConfig,
|
||
NativeMetric,
|
||
)
|
||
from services import assistant_store as store
|
||
from services.component_registry import TOOLS
|
||
# T08：适配层的策略拒绝用异常表达，在这里统一映射成 Judgment 记录。
|
||
# judgment_adapters 只依赖 contracts，不反向依赖本模块，因此没有循环导入。
|
||
from services.judgment_adapters import (
|
||
CheckpointMismatch,
|
||
ContextBudgetExceeded,
|
||
EgressNotApproved,
|
||
TaskNotValidated,
|
||
)
|
||
|
||
DISABLED_REASON_DEFAULT = "本次运行未启用专项判断；主审仍按规则、工具与经验形成意见"
|
||
|
||
# 仅这些错误码表示"提供方不可用"，可按预算重试一次（03 §5）。
|
||
RETRYABLE_PROVIDER_ERRORS = (ErrorCode.JUDGMENT_UNAVAILABLE,)
|
||
|
||
|
||
def resolve_config(enabled: bool, profile_id: str | None = None,
|
||
profile_revision: int | None = None, *, known_profiles=(),
|
||
provider: str = "local") -> JudgmentConfig:
|
||
"""解析并冻结本次运行的专项配置（BR-20：切换只影响新运行）。
|
||
|
||
`known_profiles` 是管理员在后端配置里登记的可用配置；P0 为空，因此在
|
||
enabled=True 且未给 ID 时会明确报错（E-13），而不是悄悄关闭专项。
|
||
"""
|
||
if not enabled:
|
||
if profile_id:
|
||
# 开关 false 同时传 ID 属 E-07，避免静默忽略用户选择（11 §3）
|
||
raise ValueError("专项已关闭却指定了配置 ID；请明确启用或清空选择")
|
||
return JudgmentConfig(enabled=False, provider="disabled",
|
||
disabled_reason=DISABLED_REASON_DEFAULT)
|
||
|
||
if not profile_id:
|
||
default = next((item for item in known_profiles if item.get("is_default")), None)
|
||
if default is None:
|
||
raise ValueError("当前没有可用的专项模型配置，请选择可用配置或关闭专项核查")
|
||
profile_id, profile_revision = default["id"], default.get("revision", 1)
|
||
|
||
profile = next((item for item in known_profiles if item.get("id") == profile_id), None)
|
||
if profile is None:
|
||
raise ValueError("当前专项模型配置不可用，请选择可用配置或关闭专项核查")
|
||
if profile.get("readiness") != "ready":
|
||
raise ValueError("该专项模型配置尚未通过验证，暂不可用")
|
||
if provider not in ("jev", "local"):
|
||
raise ValueError("专项提供方只能是 jev 或 local")
|
||
if profile.get("provider") != provider:
|
||
raise ValueError("所选配置的提供方与请求不一致")
|
||
return JudgmentConfig(
|
||
enabled=True, provider=provider, profile_id=profile_id,
|
||
profile_revision=profile_revision or profile.get("revision", 1),
|
||
model_version=profile.get("model_version") or "unresolved",
|
||
adapter_version=profile.get("adapter_version") or "not_applicable",
|
||
prompt_version=profile.get("prompt_version") or "not_applicable",
|
||
mapping_version=profile.get("mapping_version") or "not_applicable",
|
||
# T08：权重修订随配置一起冻结；缺失记 unresolved 而不是留空猜测（AC-39）
|
||
checkpoint_ref=profile.get("checkpoint_ref") or "unresolved")
|
||
|
||
|
||
def allowed_tools(skill_body: dict, registry_tools=None) -> set[str]:
|
||
"""工具集合 = Skill 声明 ∩ 注册表白名单（取交，不取并）。
|
||
|
||
材料正文、运行补充说明、专项模型输出都**不参与**这个计算，因此无法借它们
|
||
扩展权限（BR-10 / AC-10）。
|
||
"""
|
||
declared = set(skill_body.get("allowed_tools") or ())
|
||
return declared & set(registry_tools or TOOLS)
|
||
|
||
|
||
def assert_tools_within_policy(requested, skill_body, registry_tools=None) -> None:
|
||
"""任何越界工具请求都必须被拒，而不是"尽量满足"。"""
|
||
permitted = allowed_tools(skill_body, registry_tools)
|
||
outside = sorted(set(requested or ()) - permitted)
|
||
if outside:
|
||
raise ValueError(f"工具超出当前 Skill 授权范围，已拒绝：{outside}")
|
||
|
||
|
||
def _skipped(run_id: str, question: dict, config: JudgmentConfig, reason: str) -> dict:
|
||
return {"id": store.uid("judgment"), "run_id": run_id,
|
||
"question_id": question["id"], "question_version": question["version"],
|
||
"type": question["type"], "status": "skipped", "answer": None,
|
||
"evidence_ids": list(question.get("evidence_ids", [])),
|
||
"provider": "disabled", "profile_id": None, "profile_revision": None,
|
||
"model_version": None, "native_metrics": [], "error_code": None,
|
||
"skip_reason": reason}
|
||
|
||
|
||
def judge(config: JudgmentConfig, run_id: str, question: dict,
|
||
provider_call=None) -> dict:
|
||
"""执行一道题的专项判断。
|
||
|
||
`provider_call` 是 T08 才接的适配器；本批只覆盖 disabled 与失败/畸形路径。
|
||
"""
|
||
if config.provider == "disabled":
|
||
return _skipped(run_id, question, config, "disabled")
|
||
|
||
if not question.get("evidence_ids"):
|
||
# 无证据不请求模型：按缺证处理并明确返回 unknown（11 §4）
|
||
return {"id": store.uid("judgment"), "run_id": run_id,
|
||
"question_id": question["id"], "question_version": question["version"],
|
||
"type": question["type"], "status": "completed",
|
||
"answer": {"kind": "unknown", "value": None, "reason_code": "missing_evidence"},
|
||
"evidence_ids": [], "provider": config.provider,
|
||
"profile_id": config.profile_id, "profile_revision": config.profile_revision,
|
||
"model_version": config.model_version, "native_metrics": [],
|
||
"error_code": None, "skip_reason": None}
|
||
|
||
if provider_call is None:
|
||
return _failed(run_id, question, config, ErrorCode.JUDGMENT_UNAVAILABLE)
|
||
|
||
try:
|
||
raw = provider_call(config, question)
|
||
except TimeoutError:
|
||
return _failed(run_id, question, config, ErrorCode.JUDGMENT_UNAVAILABLE)
|
||
except ContextBudgetExceeded:
|
||
# E-25：超预算不外发、不静默截断；evidence_ids 仍保留原引用范围（AC-40）
|
||
return _failed(run_id, question, config, ErrorCode.JUDGMENT_CONTEXT_EXCEEDED)
|
||
except TaskNotValidated:
|
||
# 未获准的任务不给"未完成"以外的结论（AC-38）
|
||
return {**_failed(run_id, question, config, ErrorCode.JUDGMENT_UNAVAILABLE),
|
||
"status": "skipped", "error_code": None, "skip_reason": "unsupported"}
|
||
except EgressNotApproved:
|
||
# 未批准出口：不发送，记录策略拒绝（AC-16/19）
|
||
return {**_failed(run_id, question, config, ErrorCode.JUDGMENT_UNAVAILABLE),
|
||
"status": "skipped", "error_code": None, "skip_reason": "not_authorized"}
|
||
except CheckpointMismatch:
|
||
# 实际权重与冻结修订不符：拒绝并记失败，不得静默换模型（AC-39）
|
||
return _failed(run_id, question, config, ErrorCode.JUDGMENT_OUTPUT_INVALID)
|
||
except Exception:
|
||
return _failed(run_id, question, config, ErrorCode.JUDGMENT_UNAVAILABLE)
|
||
|
||
return _normalise(run_id, question, config, raw)
|
||
|
||
|
||
def _failed(run_id: str, question: dict, config: JudgmentConfig, code: ErrorCode) -> dict:
|
||
"""失败必须留痕且**不补分**：answer 为空，不生成任何默认取值。"""
|
||
return {"id": store.uid("judgment"), "run_id": run_id,
|
||
"question_id": question["id"], "question_version": question["version"],
|
||
"type": question["type"], "status": "failed", "answer": None,
|
||
"evidence_ids": list(question.get("evidence_ids", [])),
|
||
"provider": config.provider, "profile_id": config.profile_id,
|
||
"profile_revision": config.profile_revision,
|
||
"model_version": config.model_version, "native_metrics": [],
|
||
"error_code": code, "skip_reason": None}
|
||
|
||
|
||
def _normalise(run_id: str, question: dict, config: JudgmentConfig, raw: dict) -> dict:
|
||
"""校验并归一化提供方输出；畸形即 E-15，不猜一个分值补齐。"""
|
||
if not isinstance(raw, dict):
|
||
return _failed(run_id, question, config, ErrorCode.JUDGMENT_OUTPUT_INVALID)
|
||
answer_kind = raw.get("kind")
|
||
if answer_kind == "unknown":
|
||
if raw.get("reason_code") not in ("missing_evidence", "inconclusive"):
|
||
return _failed(run_id, question, config, ErrorCode.JUDGMENT_OUTPUT_INVALID)
|
||
answer = JudgmentAnswer(kind="unknown", reason_code=raw["reason_code"])
|
||
else:
|
||
if answer_kind != question["type"]:
|
||
# 题型不符（含"把连续输出硬当 binary"）属能力不支持，不是失败
|
||
return {**_failed(run_id, question, config, ErrorCode.JUDGMENT_CAPABILITY_UNSUPPORTED),
|
||
"status": "skipped", "error_code": None, "skip_reason": "unsupported"}
|
||
try:
|
||
answer = JudgmentAnswer(kind=answer_kind, value=raw.get("value"))
|
||
except Exception:
|
||
return _failed(run_id, question, config, ErrorCode.JUDGMENT_OUTPUT_INVALID)
|
||
|
||
metrics = []
|
||
for item in raw.get("native_metrics") or ():
|
||
try:
|
||
metrics.append(NativeMetric(**item).model_dump(mode="json"))
|
||
except Exception:
|
||
return _failed(run_id, question, config, ErrorCode.JUDGMENT_OUTPUT_INVALID)
|
||
|
||
body = {"id": store.uid("judgment"), "run_id": run_id,
|
||
"question_id": question["id"], "question_version": question["version"],
|
||
"type": question["type"], "status": "completed",
|
||
"answer": answer.model_dump(mode="json"),
|
||
"evidence_ids": list(question.get("evidence_ids", [])),
|
||
"provider": config.provider, "profile_id": config.profile_id,
|
||
"profile_revision": config.profile_revision,
|
||
"model_version": config.model_version, "native_metrics": metrics,
|
||
"error_code": None, "skip_reason": None,
|
||
"actual_checkpoint_ref": raw.get("actual_checkpoint_ref")}
|
||
return Judgment(**body).model_dump(mode="json")
|
||
|
||
|
||
def plan_questions(config: JudgmentConfig, questions) -> dict:
|
||
"""关闭专项时把已计划题目统一记 skipped，并给出运行级关闭原因。"""
|
||
if config.provider != "disabled":
|
||
return {"judgments": [], "disabled_reason": None}
|
||
return {"judgments": [_skipped("", question, config, "disabled") for question in questions or ()],
|
||
"disabled_reason": config.disabled_reason}
|