feat(v2): T10 轨迹回放与会议结果复盘——只读轨迹、unmatched 单列、更正追加
覆盖 AC-22/32/33（BR-13/21/31/32、E-05/23/26）。
后端：
- contracts/trajectory.py：TrajectoryStep 与 ReviewOutcome。
TrajectoryStep 失败必有错误码、成本未知时币种留空、observable() 给出
可读字段白名单（机器可读地排除任何推理过程字段）；
ReviewOutcome 只在 matched 且有明确映射口径时才有二元标签，附条件/撤回/
未送审一律留空，unmatched 不得挂快照。
- services/trajectory.py：
· project_steps 把既有 run 的步骤投影成轨迹（顺序即 sequence，部分完成
如实标注，失败保留，成本缺省留空不填 0）；
· replay 只读回放并按 sequence 稳定分页，不存在或跨项目抛 KeyError；
· record 追加会议结果：未知快照记 unmatched，更正要指向本项目既有记录，
链必须能终止；denominator 只数 matched 且有通过/否决口径的记录。
- api/prescreen.py：新增轨迹回放、会议结果列表与追加三个端点。
未做：轨迹写入仍由既有 DurableWorker.save_step 负责（本批不重做运行执行），
因此目前没有 judgment/rule 类步骤的产出方；会议标签映射（E-26）仍属 T11。
验证：backend 全量 unittest 494 项 OK（T09 时 469 项，无回归）。
Co-Authored-By: Claude Code <noreply@anthropic.com>
This commit is contained in:
Claude Code
@@ -0,0 +1,174 @@
|
||||
"""T10 轨迹回放与会议结果复盘。
|
||||
|
||||
依据：docs/v2/03-detailed-design.md §9.1/§9.4、14-experience-loop.md §7/§8；
|
||||
BR-13/31/32、E-05/23/26、AC-22/32/33。
|
||||
|
||||
三条不可让渡的规则：
|
||||
1. **只读回放**：轨迹从既有 run 的步骤投影出来，本模块不写 run，也不改历史结果。
|
||||
2. **失败保留**：失败的步骤照样出现并带错误码，不因为"看起来不好"被过滤。
|
||||
3. **结果追加**：会议结果只能新增；更正生成新记录并指向 `supersedes_id`，
|
||||
找不到材料版本的一律记 unmatched 并单列，不当成否决、也不进通过率分母。
|
||||
"""
|
||||
|
||||
from __future__ import annotations
|
||||
|
||||
from contracts.prescreen import EventOutcome
|
||||
from contracts.trajectory import ReviewOutcome, TrajectoryStep
|
||||
from services import assistant_store as store
|
||||
|
||||
KIND = "review_outcome"
|
||||
DEFAULT_PAGE = 50
|
||||
|
||||
# 既有步骤状态 → 轨迹状态。partial 不是完成，保留原样说明，避免看起来"全过"。
|
||||
_STATUS_MAP = {"running": "running", "completed": "completed", "partial": "completed",
|
||||
"failed": "failed", "skipped": "skipped", "cancelled": "cancelled"}
|
||||
_PARTIAL_NOTE = "部分完成（原始状态 partial）；未完成的部分不计入已核查"
|
||||
_FAILURE_CODE_FALLBACK = "RUN_STEP_FAILED"
|
||||
|
||||
RAW_ALIASES = {
|
||||
"通过": EventOutcome.approved, "同意": EventOutcome.approved,
|
||||
"否决": EventOutcome.rejected, "未通过": EventOutcome.rejected,
|
||||
"附条件": EventOutcome.conditional, "附条件通过": EventOutcome.conditional,
|
||||
"暂缓": EventOutcome.deferred, "撤回": EventOutcome.withdrawn,
|
||||
"未送审": EventOutcome.not_submitted,
|
||||
}
|
||||
|
||||
|
||||
def _step(run_body: dict, raw: dict, index: int) -> TrajectoryStep:
|
||||
status = _STATUS_MAP.get(str(raw.get("status")), "skipped")
|
||||
errors = raw.get("errors") or ()
|
||||
notes = []
|
||||
if raw.get("status") == "partial":
|
||||
notes.append(_PARTIAL_NOTE)
|
||||
if raw.get("missing_inputs"):
|
||||
notes.append("缺证：" + "、".join(str(item) for item in raw["missing_inputs"]))
|
||||
skill = raw.get("skill") or {}
|
||||
component = {"id": skill["id"], "version": str(skill.get("version", ""))} if skill.get("id") else None
|
||||
return TrajectoryStep(
|
||||
id=str(raw.get("step_id")), run_id=str(raw.get("run_id") or run_body.get("id")),
|
||||
parent_step_id=None, sequence=index,
|
||||
kind="skill" if raw.get("skill") else "tool" if raw.get("tool") else "plan",
|
||||
component_ref=component,
|
||||
input_refs=[str(raw.get("input_snapshot_id"))] if raw.get("input_snapshot_id") else [],
|
||||
output_refs=sorted(str(key) for key in (raw.get("outputs") or {})),
|
||||
evidence_ids=sorted({str(item) for finding in raw.get("findings") or ()
|
||||
for item in (finding.get("evidence_ids") or ())}),
|
||||
status=status,
|
||||
started_at=raw.get("started_at"), finished_at=raw.get("finished_at"),
|
||||
reason_summary="；".join(notes) or None,
|
||||
error_code=(str(errors[0].get("code")) if errors and errors[0].get("code")
|
||||
else _FAILURE_CODE_FALLBACK) if status == "failed" else None,
|
||||
latency_ms=raw.get("latency_ms"), cost=None, cost_currency=None)
|
||||
# 成本缺省一律留空：既有运行没有成本口径时不得填 0 或估计值（06 §5）。
|
||||
|
||||
|
||||
def project_steps(run_body: dict) -> list[dict]:
|
||||
"""把既有 run 的步骤投影成轨迹步骤，顺序即 sequence，不改写原 run。"""
|
||||
return [_step(run_body, raw, index).model_dump(mode="json")
|
||||
for index, raw in enumerate(run_body.get("steps") or ())]
|
||||
|
||||
|
||||
def replay(db, project_id: str, run_id: str, cursor: str | None = None,
|
||||
limit: int = DEFAULT_PAGE) -> dict:
|
||||
"""只读回放：按 sequence 稳定分页。不存在或跨项目 → KeyError（接口层转 404）。"""
|
||||
run = store.get(db, "run", run_id)
|
||||
if run.get("project_id") != project_id:
|
||||
raise KeyError("运行不存在或不属于该项目")
|
||||
steps = project_steps(run)
|
||||
start = max(0, int(cursor or 0))
|
||||
window = steps[start:start + max(1, min(limit, 200))]
|
||||
end = start + len(window)
|
||||
return {
|
||||
"run_id": run_id, "project_id": project_id, "status": run.get("status"),
|
||||
"snapshot_id": run.get("input_snapshot_id"),
|
||||
"steps": window,
|
||||
"next_cursor": str(end) if end < len(steps) else None,
|
||||
"read_only": True,
|
||||
"note": "轨迹只记录可观察动作与版本；不包含模型内部推理过程",
|
||||
}
|
||||
|
||||
|
||||
def _outcomes(db, project_id: str) -> list[dict]:
|
||||
return list(store.listing(db, KIND, project_id))
|
||||
|
||||
|
||||
def _chain_has_cycle(rows: list[dict], start: str) -> bool:
|
||||
"""防御性检查：追加写入本身造不出环，但迁移/导入的旧行可能有（03 §9.1）。"""
|
||||
seen, current = set(), start
|
||||
while current:
|
||||
if current in seen:
|
||||
return True
|
||||
seen.add(current)
|
||||
current = next((item.get("supersedes_id") for item in rows
|
||||
if item["id"] == current), None)
|
||||
return False
|
||||
|
||||
|
||||
def record(db, project_id: str, body: dict, actor: str,
|
||||
known_snapshots=()) -> dict:
|
||||
"""追加一条会议结果。更正只能新增，且必须指向同项目的旧记录。"""
|
||||
rows = _outcomes(db, project_id)
|
||||
raw_code = str(body.get("raw_code") or "").strip()
|
||||
if not raw_code:
|
||||
raise ValueError("会议结果必须给出原始结果码，未提供时不得推断")
|
||||
|
||||
normalized = body.get("normalized") or RAW_ALIASES.get(raw_code, EventOutcome.unknown)
|
||||
if isinstance(normalized, str):
|
||||
normalized = EventOutcome(normalized)
|
||||
snapshot = body.get("reviewed_snapshot_id")
|
||||
matched = bool(snapshot) and (not known_snapshots or snapshot in set(known_snapshots))
|
||||
|
||||
supersedes = body.get("supersedes_id")
|
||||
if supersedes:
|
||||
previous = next((item for item in rows if item["id"] == supersedes), None)
|
||||
if previous is None:
|
||||
raise ValueError("更正要指向本项目既有的会议结果记录")
|
||||
if _chain_has_cycle(rows, supersedes):
|
||||
raise ValueError("更正链出现循环，已拒绝")
|
||||
mapping_ref = body.get("mapping_ref")
|
||||
label = None
|
||||
if matched and mapping_ref and normalized in (EventOutcome.approved, EventOutcome.rejected):
|
||||
label = normalized == EventOutcome.approved
|
||||
|
||||
revision = 1 + max([item.get("revision", 0) for item in rows
|
||||
if item.get("meeting_id") == body.get("meeting_id")] or [0])
|
||||
outcome = ReviewOutcome(
|
||||
id=store.uid("outcome"), project_id=project_id,
|
||||
meeting_id=str(body.get("meeting_id") or ""), revision=revision,
|
||||
occurred_at=body.get("occurred_at") or store.now(), recorded_at=store.now(),
|
||||
reviewed_snapshot_id=snapshot if matched else None,
|
||||
raw_code=raw_code, normalized=normalized, mapping_ref=mapping_ref, label=label,
|
||||
conditions=list(body.get("conditions") or ()),
|
||||
evidence_ids=list(body.get("evidence_ids") or ()),
|
||||
status="matched" if matched else "unmatched",
|
||||
supersedes_id=supersedes, recorded_by=actor)
|
||||
saved = outcome.model_dump(mode="json")
|
||||
store.put(db, KIND, saved)
|
||||
store.event(db, project_id, "review_outcome.recorded", actor,
|
||||
{"outcome_id": saved["id"], "status": saved["status"],
|
||||
"raw_code": saved["raw_code"]}, None)
|
||||
return saved
|
||||
|
||||
|
||||
def listing(db, project_id: str, status: str | None = None) -> dict:
|
||||
"""按匹配状态分组返回。unmatched 单列，不混进通过率分母。"""
|
||||
rows = sorted(_outcomes(db, project_id),
|
||||
key=lambda item: item.get("recorded_at") or "")
|
||||
observed = [item for item in rows if item["status"] == "matched"]
|
||||
unmatched = [item for item in rows if item["status"] == "unmatched"]
|
||||
if status == "matched":
|
||||
return {"items": observed}
|
||||
if status == "unmatched":
|
||||
return {"items": unmatched}
|
||||
return {"items": rows, "matched": observed, "unmatched": unmatched,
|
||||
"counts_toward_denominator": denominator(rows)}
|
||||
|
||||
|
||||
def denominator(rows) -> dict:
|
||||
"""已观察分母只数"匹配到材料版本且有通过/否决口径"的记录（AC-22/37）。"""
|
||||
labelled = [item for item in rows if item.get("status") == "matched"
|
||||
and item.get("normalized") in ("approved", "rejected")]
|
||||
excluded = [(item["id"], item.get("raw_code"), item.get("normalized"))
|
||||
for item in rows if item not in labelled]
|
||||
return {"observed": len(labelled), "excluded": excluded,
|
||||
"note": "未送审、撤回、附条件与未知结果不计入已观察通过率；不把未知当否决"}
|
||||
Reference in New Issue
Block a user