feat(v2): 纠偏第一批——执行链产出新版预审记录（真实运行验收）
覆盖 AC-46/47/48/49（BR-45/46/47、E-27）。审计确认的 P1 缺口"正常发起评审不产生 新版预审记录"由此关闭：视图由执行链出口产生，不再依赖测试或脚本注入。 后端： - services/prescreen.py：新增 screening_from_run()，把一次真实运行的执行结果装配成 预审视图；维度与阶段按**本次实际执行的组件与核查项**投影 （dimensions_and_evidence：冻结方案声明过、且本次真的执行过的技能/规则才算维度； 经验核查等方案外伪步骤不算；阶段只表达证据成熟度，不看结论好坏）。 - services/assistant_worker.py：运行出口（含失败路径）统一调用 save_prescreen_view()， 运行结束即落库视图；装配失败单独记事件，不让运行一起丢。 - contracts/prescreen.py：视图带 evidence（界面"查看依据"的来源实体）； 新增运行级错误码 RUN_STEP_FAILED（E-27）。 - 维度状态用本次执行的步骤与核查项算（含跑通但无问题的项）， 只看待办 finding 会把通过的维度误判成未核查。 前端：面板新增"核查结果"区（每条可打开依据抽屉，显示来源、定位、原文片段与组件版本） 与降级记录展示；没有证据时明确写"按缺证处理"，不编造。 测试：backend/test_prescreen_from_run.py 12 项，用真实 enqueue + DurableWorker.process_next （临时库）断言视图由执行链产生、失败与模型不可用如实留痕、重跑新增记录。 含反例：执行前不得存在任何记录。 实际验收（浏览器渲染 + 真实运行）：状态 partial、10 维（5 已核查/2 待核实/3 阻断）、 4 条核查结果、14 条证据、1 条 RUN_STEP_FAILED 未完成项；记录见 docs/v2/17-batch-delivery-records.md（含剩余缺口：界面点击未自动化、材料确认未在浏览器走完、 演示角色需 XINGBANG_DEMO_ROLE_LOGIN=1、模型未配置）。 Co-Authored-By: Claude Code <noreply@anthropic.com>
This commit is contained in:
Claude Code
@@ -16,6 +16,8 @@ from __future__ import annotations
|
||||
from contracts.prescreen import (
|
||||
DataDomain,
|
||||
DimensionResult,
|
||||
DimensionSpec,
|
||||
ErrorCode,
|
||||
ProbabilityView,
|
||||
ScreeningRun,
|
||||
)
|
||||
@@ -120,6 +122,8 @@ def assemble(run_body: dict, frozen_plan: dict, decision: dict, judgment_config,
|
||||
"versions": versions_block(frozen_plan, {"trajectory_id": trajectory_id}
|
||||
if trajectory_id else None),
|
||||
"dimension_results": dimension_results(frozen_plan, decision.get("findings")),
|
||||
# 证据是界面"查看依据"的实体来源；超长只截断列表，不改内容
|
||||
"evidence": [dict(item) for item in (run_body.get("evidence") or ())][:200],
|
||||
"run_instruction": run_body.get("requirement") or "",
|
||||
"created_at": run_body.get("created_at") or store.now(),
|
||||
"finished_at": run_body.get("ended_at"),
|
||||
@@ -145,3 +149,161 @@ def assert_versions_frozen(screening: dict) -> None:
|
||||
if config.get("enabled") and (not config.get("profile_id")
|
||||
or config.get("profile_revision") is None):
|
||||
raise ValueError("启用专项的运行必须冻结配置 ID 与修订")
|
||||
|
||||
|
||||
# ---------------------------------------------------------------- 执行链装配（纠偏第一批）
|
||||
# 证据成熟度刻度。**只表达证据成熟度，不看结论好坏**：发现问题不会降阶段，
|
||||
# 通过也不会升到人工复核（见 docs/v2/16-delivery-corrections.md BR-46）。
|
||||
STAGE_SCALE_ID = "evidence-stage"
|
||||
STAGE_SCALE_VERSION = "1.0.0"
|
||||
STAGE_LEVELS = [
|
||||
{"value": 0, "label": "未提供", "criterion": "本次没有该核查项的结果"},
|
||||
{"value": 1, "label": "仅有说明", "criterion": "有结论但未引用任何证据"},
|
||||
{"value": 2, "label": "资料待核", "criterion": "引用了证据；尚无交叉核验记录"},
|
||||
{"value": 3, "label": "交叉核验", "criterion": "有核验记录证明证据之间互相印证"},
|
||||
{"value": 4, "label": "人工复核", "criterion": "有人工复核记录"},
|
||||
]
|
||||
DISABLED_JUDGMENT_CONFIG = {
|
||||
"enabled": False, "provider": "disabled",
|
||||
"disabled_reason": "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",
|
||||
}
|
||||
# 阶段 3/4 需要"核验记录"或"人工复核记录"这类依据，P0 的执行链还产不出，
|
||||
# 因此自动映射只会落在 0–2；3/4 保留给后续判别路径，不在这里臆造。
|
||||
|
||||
|
||||
def stage_from_check(item) -> int:
|
||||
"""核查项 → 证据成熟度阶段（BR-46）。"""
|
||||
if item.get("result") in (None, "not_applicable"):
|
||||
return 0
|
||||
return 2 if item.get("evidence_ids") else 1
|
||||
|
||||
|
||||
def _component_of(step, skills, rules):
|
||||
"""步骤归属的组件：**必须是冻结方案里声明过的技能或规则**。
|
||||
|
||||
既有执行器把技能、规则、经验核查都落成 step，其中经验核查这类不在方案里的
|
||||
步骤不是核查维度；反过来方案声明的组件若本次没执行，也不会出现在结果里
|
||||
（维度的定义是"跑了什么"，不是"方案里写了什么"）。
|
||||
"""
|
||||
step_id = str(step.get("step_id") or "")
|
||||
skill_id = str((step.get("skill") or {}).get("id") or "")
|
||||
lookup = skill_id or step_id
|
||||
if not lookup or lookup != step_id:
|
||||
return None
|
||||
item = next((s for s in skills or () if s.get("id") == lookup), None)
|
||||
if item is None:
|
||||
item = next((r for r in rules or () if r.get("id") == lookup), None)
|
||||
if item is None:
|
||||
return None
|
||||
return lookup, str(item.get("name") or lookup), str(item.get("version") or "1.0.0")
|
||||
|
||||
|
||||
def dimensions_and_evidence(steps, skills, rules):
|
||||
"""本次实际执行的组件 → 维度定义 + 用于判定状态与阶段的核查输入（BR-46）。
|
||||
|
||||
只投影真实执行记录：步骤状态决定该维度是否算已核查，核查项决定是否阻断；
|
||||
技能/规则跑通但没有产出核查项时，维度按"已核查"成立——这不等于通过审批。
|
||||
方案没有执行记录就不产生维度，绝不补占位轴。
|
||||
"""
|
||||
dimensions, items, order = [], [], 0
|
||||
for step in steps or ():
|
||||
component = _component_of(step, skills, rules)
|
||||
if not component:
|
||||
continue
|
||||
component_id, label, version = component
|
||||
findings = [dict(item) for item in step.get("findings") or ()]
|
||||
codes = [str(item.get("code")) for item in findings if item.get("code")]
|
||||
status = str(step.get("status") or "")
|
||||
if status == "failed":
|
||||
step_result = "error"
|
||||
elif status in ("partial", "running", "queued"):
|
||||
step_result = "unknown"
|
||||
elif any(item.get("result") not in ("pass", "not_applicable") for item in findings):
|
||||
step_result = "issue"
|
||||
else:
|
||||
step_result = "pass"
|
||||
evidence_ids = sorted({str(e) for item in findings for e in (item.get("evidence_ids") or ())})
|
||||
refs = [{"id": component_id, "version": version}] + [
|
||||
{"id": code, "version": version} for code in codes]
|
||||
dimensions.append(DimensionSpec(
|
||||
id=component_id, label=label, order=order, scale_id=STAGE_SCALE_ID,
|
||||
scale_version=STAGE_SCALE_VERSION, levels=STAGE_LEVELS,
|
||||
component_refs=refs).model_dump(mode="json"))
|
||||
order += 1
|
||||
step_item = {"code": component_id, "result": step_result,
|
||||
"evidence_ids": evidence_ids, "stage": None}
|
||||
step_item["stage"] = stage_from_check(step_item)
|
||||
items.append(step_item)
|
||||
for finding, code in zip(findings, codes):
|
||||
row = {"code": code, "result": finding.get("result"),
|
||||
"evidence_ids": list(finding.get("evidence_ids") or ())}
|
||||
row["stage"] = stage_from_check(row)
|
||||
items.append(row)
|
||||
return dimensions, items
|
||||
|
||||
|
||||
def _degradations(steps, run) -> list[dict]:
|
||||
"""失败步骤与运行级失败都留痕，并给出可定位的原因（AC-48）。"""
|
||||
rows = []
|
||||
for step in steps or ():
|
||||
if step.get("status") != "failed":
|
||||
continue
|
||||
error = (step.get("errors") or [{}])[0]
|
||||
kind = str(error.get("type") or "")
|
||||
code = (ErrorCode.RUN_BUDGET_EXCEEDED if kind == "TimeoutError"
|
||||
else ErrorCode.RUN_STEP_FAILED)
|
||||
label = str(step.get("name") or step.get("step_id") or "未命名步骤")
|
||||
detail = str(error.get("message") or "步骤未完成")
|
||||
rows.append({"code": code, "message": f"{label}：{detail[:400]}"})
|
||||
if run.get("status") == "failed" and not rows:
|
||||
rows.append({"code": ErrorCode.RUN_STEP_FAILED,
|
||||
"message": str(run.get("error") or "运行未完成")[:400]})
|
||||
return rows
|
||||
|
||||
|
||||
def _missing_inputs(steps, run) -> list[str]:
|
||||
"""缺证项汇总：来自各步骤与数据错误，去重后保持顺序。"""
|
||||
raw = [str(item) for step in steps or () for item in (step.get("missing_inputs") or ())]
|
||||
for item in run.get("data_errors") or ():
|
||||
raw.append(str(item.get("message") if isinstance(item, dict) else item))
|
||||
seen, rows = set(), []
|
||||
for item in raw:
|
||||
if item and item not in seen:
|
||||
seen.add(item)
|
||||
rows.append(item[:200])
|
||||
return rows
|
||||
|
||||
|
||||
def screening_from_run(run: dict, data_domain: str = "demo",
|
||||
judgment_config=None) -> dict:
|
||||
"""把**一次真实运行的执行结果**装配成新版预审视图（AC-46/47/48/49）。
|
||||
|
||||
这是执行链的出口：由 worker 在运行结束时调用，因此正常发起评审就会产生记录，
|
||||
不需要测试或演示脚本注入（BR-45）。方案取自运行创建时冻结的 components，
|
||||
不在装配时重读目录，避免"跑的是旧版本、显示的是新版本"。
|
||||
"""
|
||||
components = run.get("components") or {}
|
||||
checks = [dict(item) for item in run.get("checks") or ()]
|
||||
steps = list(run.get("steps") or ())
|
||||
skills = list(components.get("skills") or ())
|
||||
rules = list(components.get("rules") or ())
|
||||
dimensions, dimension_inputs = dimensions_and_evidence(steps, skills, rules)
|
||||
frozen_plan = {
|
||||
"plan": components.get("plan") or {},
|
||||
"skills": skills,
|
||||
"rules": rules,
|
||||
"template": components.get("template") or {},
|
||||
"content_hash": components.get("content_hash"),
|
||||
"dimensions": dimensions,
|
||||
}
|
||||
decision = policy.decide(checks, _missing_inputs(steps, run), _degradations(steps, run))
|
||||
decision["findings"] = [dict(item) for item in run.get("findings") or ()]
|
||||
view = assemble(run, frozen_plan, decision,
|
||||
judgment_config or DISABLED_JUDGMENT_CONFIG,
|
||||
data_domain=data_domain,
|
||||
status=str(run.get("status") or "completed"),
|
||||
trajectory_id=run.get("id"))
|
||||
# 维度状态用**本次执行的步骤与核查项**算（含跑通但无问题的项）：
|
||||
# 只看待办 finding 会把通过的维度误判成未核查。
|
||||
view["dimension_results"] = dimension_results(frozen_plan, dimension_inputs)
|
||||
return ScreeningRun(**view).model_dump(mode="json")
|
||||
|
||||
Reference in New Issue
Block a user