覆盖 AC-50/51/52(BR-48/49、E-28)。审计确认的 P1-3 缺口"历史提炼与卡片 不在实时 API 调用图上"由此关闭:提炼、候选治理、卡片进方案全部有真实调用方。 后端: - contracts:新增 E-28 EXTRACTION_UNAVAILABLE(运行级,不进 HTTP 信封)与 KnowledgeExtraction 记录类型。契约层强制"失败不得带候选", AC-50「模型不可用⇒不产生候选」是类型约束而不是文档约定。 - services/experience_extraction.py(新):真实提炼入口。模型不可用或调用失败 → 记 E-28 且 candidate_id 为空;不产半成品候选。传输纪律同 assistant_model (禁进程内代理、零重试),且 tools=[] 使"越权读材料"在结构上不可能。 自然语言修订同样必须真实调用模型,没有"把中文猜成条件"的回退路径。 - services/knowledge_cards.py:新增 current_draft/revise/list_cards 与 registry_conditions。修订只产草稿、保留 origin 与来源、升版本记 parent_version; test_cases 纳入可改字段——它是 component_evaluation 对自定义 Skill 的发布门禁, 不开放则提炼出的候选永远卡在评测门外。 - services/project_assistant.py:enqueue 改用 review_plan.assemble(), 让 T06-b 的依赖预检(E-19 可定位文案)与维度解析不再被绕过。 - api/prescreen.py:case-timeline、knowledge-extractions、knowledge-candidates (创建/修订/状态机)、review-plans(装配/状态机)。角色门禁与既有 /components 入口一致;id 与数据域由服务端生成;按 id 的路由一律在 run() 内 查归属,否则 KeyError 会变成 500 而不是 404。 - component_registry/review_plan:下沉 current_draft/version_key 与 list_plans。 前端: - 新增 /experience-center 独立入口(BR-50)与 ExperienceCenter 组件; AgentNav 增加「历史经验加工」。 - originBadge 此前零调用方、.origin-* 无任何 CSS——守卫只 grep 库文件字符串, 于是"测试通过但功能不存在"。本批补齐消费方与样式,并用左边框样式区分四种来源, 保证去色仍可区分(AC-42/BR-48)。 - 守卫语义按 AC-55 调整:由"禁止新建路由"改为"禁止重复页面" (面板不得搬到独立路由、已合并页必须保持重定向),并新增 experience-center-ui 守卫(已登记 verify.py,共 12 项检查)。 测试:新增 36 项(提炼 15 / 候选治理 15 / 卡片到运行 6),全量 unittest 无回归; 两次同项目运行对照证明卡片确实改变了核查结果,且 basis_ids 指向 卡片id@版本。 未闭合(记录在 17 的剩余缺口):真实模型提炼尚未实测(需端点); 界面未经浏览器点击验证;例外边界未进入可求值语言;规则类候选不在本批。 Co-Authored-By: Claude Code <noreply@anthropic.com>
364 lines
18 KiB
Python
364 lines
18 KiB
Python
"""T05-b 知识卡片:既有 Skill/Rule 治理的目标投影。
|
||
|
||
**不新建第二套发布库**(docs/v2/03-detailed-design.md §9.2 明文)。本模块只负责三件事:
|
||
|
||
1. **投影字段**:origin、谱系(parent_version)、数据域、卡片范围(project/team)、条件与例外
|
||
2. **硬规则门禁**:hard 必须有正式制度/授权依据与依据证据(BR-26 / E-18)
|
||
3. **有效性过滤**:按数据域 / 状态 / 有效期 / 范围筛选可应用的已发布版本(BR-04)
|
||
|
||
状态机与发布**完全委托** `component_registry.draft()` / `release()`,因此既有保证
|
||
原样生效、不重复实现:
|
||
- 自批拒绝:`release()` 里 `edited_by == actor` → PermissionError(AC-03 / AC-43)
|
||
- 评测门禁:候选哈希必须匹配通过的隔离评测(AC-34)
|
||
- 已发布不可覆盖:`component_version` 以 immutable 写入;改已发布需新版本号
|
||
- 工具白名单、方案预算等既有校验
|
||
|
||
**命名映射(不得互相顶替)**:
|
||
- 本设计的 `card_scope`(project/team)↔ 注册表无此概念,作为附加键保存
|
||
- 本设计的 `data_domain` ↔ 注册表的 `scope`(demo/production)——**同一个键名、不同语义**,
|
||
投影时显式互换,绝不混用
|
||
"""
|
||
|
||
from __future__ import annotations
|
||
|
||
from contracts.prescreen import Condition, KnowledgeCard, Origin, RuleEffect
|
||
from services import assistant_store as store
|
||
from services import component_registry as registry
|
||
|
||
ORIGINS = tuple(item.value for item in Origin)
|
||
EFFECTS = tuple(item.value for item in RuleEffect)
|
||
|
||
HARD_BASIS_REQUIRED = "硬性规则需要正式制度或授权依据,仅凭历史频次不足;当前仅可保存为建议草稿"
|
||
|
||
|
||
def _check_hard_basis(body: dict) -> None:
|
||
"""hard 规则的资格校验(BR-26 / E-18)。advisory 不设门槛。"""
|
||
if body.get("kind") != "rule_set":
|
||
return
|
||
effect = body.get("effect", "advisory")
|
||
if effect not in EFFECTS:
|
||
raise ValueError(f"规则效果取值须为 {' 或 '.join(EFFECTS)}")
|
||
if effect != "hard":
|
||
return
|
||
if not (body.get("policy_ref") or body.get("authorization_ref")):
|
||
raise ValueError(HARD_BASIS_REQUIRED)
|
||
if not body.get("basis_evidence_ids"):
|
||
raise ValueError("硬性规则必须附依据证据,不能只有结论")
|
||
|
||
|
||
# 注册表既有的条件形状与本设计的 Condition **不是同一套**:
|
||
# 注册表:{code, field, operator, threshold, title, suggestion}
|
||
# 本设计:{field, operator, value}
|
||
# 因此投影层做**翻译**,而不是把一方强加给另一方——注册表体按它自己的形状保存,
|
||
# 卡片侧只暴露可安全求值的字段,展示性字段(title/suggestion)不进契约。
|
||
# 只映射注册表**实际可能产出**的算子:component_registry.validate() 要求
|
||
# operator ∈ (gt, lt, eq, gte, lte) 且 threshold 可解析为数值。
|
||
# 本设计的 Condition 还允许 in/present/absent,但它们不可能来自注册表条件,
|
||
# 故不在此表中——出现即报错,避免把不可能的取值当成已支持。
|
||
_OPERATORS = {"eq": "eq", "lt": "lt", "lte": "lte", "gt": "gt", "gte": "gte"}
|
||
|
||
|
||
def _translate_conditions(conditions, label: str) -> list[dict]:
|
||
"""把注册表条件翻译为设计的安全条件;无法表达即报错,不静默丢弃。"""
|
||
parsed = []
|
||
for index, item in enumerate(conditions or ()):
|
||
if not isinstance(item, dict):
|
||
raise ValueError(f"{label}第 {index + 1} 条不是对象")
|
||
operator = _OPERATORS.get(str(item.get("operator", "")).strip().lower())
|
||
if operator is None:
|
||
raise ValueError(f"{label}第 {index + 1} 条算子不受支持:{item.get('operator')!r}")
|
||
raw = item.get("threshold", item.get("value"))
|
||
value = raw if isinstance(raw, (list, bool)) else (None if raw is None else str(raw))
|
||
try:
|
||
parsed.append(Condition(field=str(item.get("field", "")), operator=operator,
|
||
value=value).model_dump(mode="json"))
|
||
except Exception as exc:
|
||
raise ValueError(f"{label}第 {index + 1} 条不合法:{exc}") from exc
|
||
if not parsed and conditions:
|
||
raise ValueError(f"{label}无法表达为安全条件")
|
||
return parsed
|
||
|
||
|
||
def propose(db, body: dict, actor: str, project_id: str, origin: str = "manual",
|
||
source_event_ids=(), source_evidence_ids=(), parent_version: dict | None = None,
|
||
card_scope: str = "project", team_id: str | None = None):
|
||
"""提议一张知识卡片草稿。**只生成草稿,绝不直接生效**(BR-03 / AC-26)。"""
|
||
if origin not in ORIGINS:
|
||
raise ValueError(f"来源取值须为 {' 或 '.join(ORIGINS)}")
|
||
if card_scope not in ("project", "team"):
|
||
raise ValueError("卡片范围须为 project 或 team")
|
||
if card_scope == "team" and not team_id:
|
||
raise ValueError("team 范围的卡片必须有 team_id")
|
||
if origin == "history_derived" and not (source_event_ids or source_evidence_ids):
|
||
raise ValueError("历史提炼卡片必须保留来源(事件或证据),不得成为无源知识")
|
||
if "data_domain" in body:
|
||
raise ValueError("请用 scope 声明数据域(注册表既有约定),不要另传 data_domain")
|
||
|
||
_check_hard_basis(body)
|
||
enriched = dict(
|
||
body,
|
||
origin=origin,
|
||
project_id=project_id,
|
||
card_scope=card_scope,
|
||
team_id=team_id,
|
||
parent_version=parent_version,
|
||
source_event_ids=list(source_event_ids),
|
||
source_evidence_ids=list(source_evidence_ids),
|
||
)
|
||
# 只校验"能否表达为安全条件",不改写注册表既有的条件形状
|
||
_translate_conditions(body.get("conditions"), "适用条件")
|
||
_translate_conditions(body.get("exceptions"), "例外")
|
||
# 委托既有治理:normalize + validate + 写入 component_draft
|
||
return registry.draft(db, enriched, actor)
|
||
|
||
|
||
def publish(db, draft_id: str, actor: str, preview: bool = False):
|
||
"""发布草稿。自批与评测门禁由 `component_registry.release()` 保证。"""
|
||
body = store.get(db, "component_draft", draft_id)
|
||
# 门禁在发布时**再查一次**:草稿可能在提议之后被改动,不能只信提议时的校验
|
||
_check_hard_basis(body)
|
||
return registry.release(db, draft_id, actor, preview)
|
||
|
||
|
||
def _project(body: dict) -> dict:
|
||
"""把注册表对象投影为 KnowledgeCard。scope↔data_domain 在此完成互换。"""
|
||
domain = body.get("scope")
|
||
if domain not in ("demo", "production"):
|
||
raise ValueError("组件的数据域不合法,无法投影为卡片")
|
||
# 内置种子组件是**全局**的(无 project_id / edited_by),不是项目知识卡片。
|
||
# 明确报错而不是让 pydantic 抛出难懂的信息——否则容易把内置 Skill 误当项目卡片。
|
||
if not body.get("project_id") or not body.get("edited_by"):
|
||
raise ValueError(
|
||
"该组件不是项目知识卡片(缺少 project_id 或 edited_by);"
|
||
"内置 Skill/规则集属全局配置,请勿投影为卡片"
|
||
)
|
||
card = {
|
||
"id": body.get("component_id") or body["id"],
|
||
"project_id": body.get("project_id") or "",
|
||
"data_domain": domain, # ← 注册表 scope
|
||
"revision": body.get("revision", 1),
|
||
"version": body.get("version", "1.0.0"),
|
||
"kind": body.get("kind", ""),
|
||
"title": body.get("name") or body.get("title") or "",
|
||
"origin": body.get("origin", "manual"),
|
||
"status": body.get("status", "draft"),
|
||
"source_event_ids": body.get("source_event_ids", []),
|
||
"source_evidence_ids": body.get("source_evidence_ids", []),
|
||
"parent_version": body.get("parent_version"),
|
||
"created_by": body.get("owner") or body.get("edited_by", ""),
|
||
"edited_by": body.get("edited_by", ""),
|
||
"card_scope": body.get("card_scope", "project"),
|
||
"team_id": body.get("team_id"),
|
||
"conditions": _translate_conditions(body.get("conditions"), "适用条件"),
|
||
"exceptions": _translate_conditions(body.get("exceptions"), "例外"),
|
||
"effective_at": body.get("effective_from") or store.now(),
|
||
"expires_at": body.get("expires_at"),
|
||
"reviewer_id": body.get("approved_by"),
|
||
"rule": None,
|
||
}
|
||
if body.get("kind") == "rule_set":
|
||
card["rule"] = {
|
||
"effect": body.get("effect", "advisory"),
|
||
"predicate": _translate_conditions(
|
||
body.get("predicate") or body.get("conditions"), "规则条件") or [
|
||
{"field": "unknown", "operator": "present"}],
|
||
"on_match": body.get("on_match", "flag"),
|
||
"basis_evidence_ids": body.get("basis_evidence_ids", []),
|
||
"policy_ref": body.get("policy_ref"),
|
||
"authorization_ref": body.get("authorization_ref"),
|
||
}
|
||
return KnowledgeCard(**card).model_dump(mode="json")
|
||
|
||
|
||
def registry_conditions(mapping: dict) -> list[dict]:
|
||
"""{标准字段: 触发值} → 注册表既有的等值条件形状。
|
||
|
||
注册表条件只能是 `{code, field, operator, threshold}` 且算子限于数值阈值,
|
||
因此这里固定用 `eq`——**不试图表达"字段存在/缺失"**,那种语义在注册表条件语言里
|
||
没有对应算子(`present` 被 test_knowledge_cards 的既有断言锁定拒绝)。
|
||
历史提炼与人工修订共用本函数,避免两处各写一份形状。
|
||
"""
|
||
return [{"code": "history.condition." + str(index + 1), "field": str(field),
|
||
"operator": "eq", "threshold": value}
|
||
for index, (field, value) in enumerate(sorted((mapping or {}).items()))]
|
||
|
||
|
||
def draft_card(db, draft_id: str) -> dict:
|
||
return _project(store.get(db, "component_draft", draft_id))
|
||
|
||
|
||
# ---------------------------------------------------------------- 候选读取与修订(AC-51)
|
||
def current_draft(db, card_ref: str) -> dict:
|
||
"""按草稿 id(`组件id@版本`)或组件 id 找当前候选草稿。
|
||
|
||
判据与方案草稿完全相同,因此直接复用注册表实现,不在这里再写一份排序规则。
|
||
"""
|
||
return registry.current_draft(db, card_ref)
|
||
|
||
|
||
def list_cards(db, project_id: str, domain: str | None = None) -> list[dict]:
|
||
"""项目下的知识卡片(草稿与已发布)。内置全局组件不是项目卡片,按 `_project` 的
|
||
既有判据排除,不在这里另写一套过滤。
|
||
"""
|
||
cards = []
|
||
for kind in ("component_draft", "component_version"):
|
||
for row in store.listing(db, kind, project_id):
|
||
if row.get("project_id") != project_id:
|
||
continue
|
||
if domain and row.get("scope") != domain:
|
||
continue
|
||
try:
|
||
cards.append(_project(row))
|
||
except ValueError:
|
||
continue
|
||
cards.sort(key=lambda card: (card["id"], card["status"] != "published", card["revision"]))
|
||
return cards
|
||
|
||
|
||
# 可改字段:适用范围、例外说明,以及**独立测试样本**。
|
||
# test_cases 在这里不是可选项——`component_evaluation` 对自定义 Skill 的发布门禁就是
|
||
# "必须有独立测试样本与期望输出"(14 §4 的"专家修改 → 测试"步骤)。不开放它,
|
||
# 提炼出来的候选就永远卡在评测门外,闭环无法走通。
|
||
EDITABLE_CHANGES = ("conditions", "exception_note", "test_cases")
|
||
|
||
|
||
def _bump(version: str) -> str:
|
||
parts = str(version).split(".")
|
||
if len(parts) == 3 and all(part.isdigit() for part in parts):
|
||
return f"{parts[0]}.{parts[1]}.{int(parts[2]) + 1}"
|
||
return f"{version}.1"
|
||
|
||
|
||
def _redraft_body(row: dict) -> dict:
|
||
"""把已存在的草稿行还原成可再次提交的注册表体。
|
||
|
||
必须去掉 `registry.draft()` 自己维护或校验用的键:`revision`(否则会被当成
|
||
expected 与 `id` 拼接重复计版本)、`status`/`edited_by`/`approved_by` 等由服务端
|
||
决定的字段、以及 `content_hash`(会重算)。
|
||
"""
|
||
body = {key: value for key, value in row.items()
|
||
if key not in ("revision", "content_hash", "status", "edited_by", "approved_by",
|
||
"published_at", "evaluation_id", "schema_version", "dependencies")}
|
||
body["id"] = row.get("component_id") or row["id"]
|
||
return body
|
||
|
||
|
||
def revise(db, card_ref: str, changes: dict, actor: str, note: str, expected_revision=None):
|
||
"""按人工修改形成**草稿修订**(AC-51 / BR-42)。
|
||
|
||
三条不变量,全部由既有治理保证而不是本函数自己承诺:
|
||
|
||
- **只产草稿**:走 `propose()` → `registry.draft()`,发布仍需独立审核与隔离评测;
|
||
- **来源不可洗**:origin 与来源事件原样继承,历史卡片修订后仍是 `history_derived`;
|
||
- **不自批**:本函数不发布;发布时 `registry.release()` 才检查 `edited_by == actor`。
|
||
|
||
已发布的当前版本会被**升版本**并记 `parent_version`;未发布时原地产生新 revision
|
||
(与既有"草稿可反复修订"一致)。
|
||
"""
|
||
if not str(note or "").strip():
|
||
raise ValueError("修订必须填写变更说明,否则无法追溯改了什么")
|
||
unknown = sorted(set(changes or ()) - set(EDITABLE_CHANGES))
|
||
if unknown:
|
||
raise ValueError(f"不支持的修订字段:{unknown};本版只允许改适用范围、例外说明与独立测试样本")
|
||
if not changes:
|
||
raise ValueError("修订必须至少改动一个字段")
|
||
|
||
draft = current_draft(db, card_ref)
|
||
component_id = draft["component_id"]
|
||
version_id = f"{component_id}@{draft['version']}"
|
||
try:
|
||
pointer = store.get(db, "release", component_id)
|
||
except KeyError:
|
||
pointer = None
|
||
if pointer and pointer.get("version_id") == version_id:
|
||
version = _bump(draft["version"])
|
||
parent = {"id": component_id, "version": draft["version"]}
|
||
else:
|
||
version = draft["version"]
|
||
parent = draft.get("parent_version")
|
||
|
||
body = _redraft_body(draft)
|
||
body["version"] = version
|
||
body["change_note"] = str(note)
|
||
if "conditions" in changes:
|
||
body["conditions"] = registry_conditions(changes["conditions"])
|
||
if "exception_note" in changes:
|
||
body["exception_note"] = str(changes["exception_note"])
|
||
if "test_cases" in changes:
|
||
cases = changes["test_cases"]
|
||
if not isinstance(cases, list) or any(
|
||
not isinstance(case, dict) or not str(case.get("name") or "").strip()
|
||
for case in cases):
|
||
raise ValueError("测试样本必须是带 name 的对象列表")
|
||
body["test_cases"] = cases
|
||
if expected_revision is not None:
|
||
body["revision"] = expected_revision # registry.draft 用它做乐观并发校验
|
||
|
||
return propose(db, body, actor, draft["project_id"], origin=draft["origin"],
|
||
source_event_ids=draft.get("source_event_ids"),
|
||
source_evidence_ids=draft.get("source_evidence_ids"),
|
||
parent_version=parent,
|
||
card_scope=draft.get("card_scope", "project"),
|
||
team_id=draft.get("team_id"))
|
||
|
||
|
||
def published_cards(db, domain: str) -> list[dict]:
|
||
return [_project(row) for row in store.listing(db, "component_version")
|
||
if row.get("status") == "published" and row.get("scope") == domain]
|
||
|
||
|
||
def effective(db, domain: str, project_id: str | None = None, at: str | None = None) -> list[dict]:
|
||
"""可应用的已发布卡片(BR-04:先按数据域/状态/有效期/范围过滤)。
|
||
|
||
用户级权限(OQ-02)尚未建设,故此处只做域、有效期与范围过滤;
|
||
**不得**据此宣称已完成授权过滤。
|
||
"""
|
||
moment = at or store.now()
|
||
result = []
|
||
for card in published_cards(db, domain):
|
||
if card["expires_at"] and card["expires_at"] <= moment:
|
||
continue
|
||
if card["card_scope"] == "project" and project_id and card["project_id"] != project_id:
|
||
continue
|
||
if card["card_scope"] == "team" and not card["team_id"]:
|
||
continue
|
||
result.append(card)
|
||
return result
|
||
|
||
|
||
# ---------------------------------------------------------------- AC-26
|
||
# 运行时的"本次预审补充说明"只影响**本次运行**(run_instruction 落在运行对象上,
|
||
# 见 03 §9.3)。要把它变成持久知识,唯一途径是**显式提议为草稿**,且草稿走正常
|
||
# 审核流程;指令本身不得改变工具权限。
|
||
RUN_SCOPE_FIELDS = ("run_instruction", "tools", "allowed_tools", "permissions", "engine")
|
||
|
||
|
||
def draft_from_run_instruction(db, project_id: str, run_id: str, body: dict, actor: str):
|
||
"""把一次运行的补充说明转成**待审草稿**(AC-26)。
|
||
|
||
三条约束:
|
||
1. 只产草稿,绝不因"运行中说过"而直接生效
|
||
2. 指令不得携带工具/权限/引擎字段——扩大工具授权必须走独立提议与审核
|
||
3. 保留来源运行,便于回溯"这条知识从哪次运行来"
|
||
"""
|
||
forbidden = sorted(field for field in RUN_SCOPE_FIELDS if field in body)
|
||
if forbidden:
|
||
raise ValueError(
|
||
"本次预审补充说明不能持久化工具权限或执行引擎;"
|
||
f"请移除这些字段并通过独立提议与审核:{forbidden}"
|
||
)
|
||
if not body.get("name") or not body.get("kind"):
|
||
raise ValueError("转为持久知识需要明确的名称与类型")
|
||
|
||
origin = body.get("origin", "manual")
|
||
if origin not in ORIGINS:
|
||
raise ValueError(f"来源取值须为 {' 或 '.join(ORIGINS)}")
|
||
# 运行指令天然没有历史来源,历史来源必须另行提供证据
|
||
if origin == "history_derived":
|
||
raise ValueError("运行补充说明不能声明为历史提炼;历史提炼需提供来源事件或证据")
|
||
|
||
enriched = dict(body, derived_from_run=run_id, proposed_via="run_instruction")
|
||
return propose(db, enriched, actor, project_id, origin=origin,
|
||
card_scope=body.get("card_scope", "project"),
|
||
team_id=body.get("team_id"))
|