生产链: 补证/重写删除业务硬上限，改为授权终态（阶段E第一部分）
This commit is contained in:
zizi
parent
860771a28e
commit
fabcd4ff91
@ -262,7 +262,7 @@ DRAFT/CHECKING/PASSED -> DISCARDED（用户明确决策）
|
|||||||
|
|
||||||
正文、规划、提取、检测和评审分别由单一职责角色执行，主代理不把多个角色合成一次模型调用。
|
正文、规划、提取、检测和评审分别由单一职责角色执行，主代理不把多个角色合成一次模型调用。
|
||||||
|
|
||||||
当前生产编排实现（`run_writer_pipeline`）的事实：机械门 → 语义检测 → 有限补证/重写环、CAS 乐观锁 + 失败收敛终态、生成篇幅与机械接受底线分层合同、评测候选四层强制；代码层的 `MAX_EVIDENCE_REQUESTS=3`、`MAX_REWRITES=2` 是技术保护，业务继续/停止的授权归人（见本节目标合同）。生产写手收到 4000–7000 汉字（目标 7000），可信适配器只拒绝不满 3001 汉字或超过 10000 汉字的异常输出；两层区间由 WriterContext 机械校验保持嵌套。
|
当前生产编排实现（`run_writer_pipeline`）的事实：机械门 → 语义检测 → 单次收敛、证据缺口收敛为授权终态（AUTHORIZATION_REQUIRED，补证/重写经人授权后以新运行继续）、CAS 乐观锁 + 失败收敛终态、生成篇幅与机械接受底线分层合同、评测候选四层强制；业务继续/停止的授权归人，技术保护（超时、预算、取消）在执行策略层。生产写手收到 4000–7000 汉字（目标 7000），可信适配器只拒绝不满 3001 汉字或超过 10000 汉字的异常输出；两层区间由 WriterContext 机械校验保持嵌套。
|
||||||
|
|
||||||
## 6. 运行与落库
|
## 6. 运行与落库
|
||||||
|
|
||||||
@ -276,7 +276,7 @@ DRAFT/CHECKING/PASSED -> DISCARDED（用户明确决策）
|
|||||||
> - 状态机持久化：**已建**——运行级 CAS 链 `example_candidate_cas`，业务态落 `example_candidate`；ARCHIVED 态未启用。
|
> - 状态机持久化：**已建**——运行级 CAS 链 `example_candidate_cas`，业务态落 `example_candidate`；ARCHIVED 态未启用。
|
||||||
> - 结构化事实增量：**已建**——`example_fact_delta`（提案）+ `example_fact_ledger`（正典账本）；模型只提六型闭集增量且必须带正文证据引文，只有用户批准的增量随正文同事务入账本，抽取结果不自动升格。
|
> - 结构化事实增量：**已建**——`example_fact_delta`（提案）+ `example_fact_ledger`（正典账本）；模型只提六型闭集增量且必须带正文证据引文，只有用户批准的增量随正文同事务入账本，抽取结果不自动升格。
|
||||||
> - 待建：生产链上的质量评分环节（盲评仍只接离线评测）；章后抽取的异步执行（接受时只登记 pending 投影）。
|
> - 待建：生产链上的质量评分环节（盲评仍只接离线评测）；章后抽取的异步执行（接受时只登记 pending 投影）。
|
||||||
> - 补证重组装：**已建**——语义 `needs_evidence` 时 `production_evidence_reassemble` 检索本作品 as_of 内正文/实体摘录并追加 factEvidence；命中则有限重写（技术保护上限见 §5；业务继续授权归人）。零命中不禁止写手发明新设定、不重写，缺口升格为 `newSettingCandidates`，当前候选进人闸。新设定是否进入正典由人决定；与既有正典冲突仍失败关闭。
|
> - 补证重组装：**已建**——语义 `needs_evidence` 时 `production_evidence_reassemble` 检索本作品 as_of 内正文/实体摘录并追加 factEvidence；命中则停在授权终态（AUTHORIZATION_REQUIRED），重写经人授权后以新运行继续。零命中不禁止写手发明新设定、不重写，缺口升格为 `newSettingCandidates`，当前候选进人闸。新设定是否进入正典由人决定；与既有正典冲突仍失败关闭。
|
||||||
|
|
||||||
## 7. 开发评测边界
|
## 7. 开发评测边界
|
||||||
|
|
||||||
|
|||||||
@ -34,7 +34,7 @@ Writer 不接收 `runId`、权限信息、manifest、hash、候选版本、验
|
|||||||
2. 伏笔只按细纲动作执行:说埋就埋、说推就推、说收就收;不擅自提前回收,不新开大坑。
|
2. 伏笔只按细纲动作执行:说埋就埋、说推就推、说收就收;不擅自提前回收,不新开大坑。
|
||||||
3. 前情衔接与上一章末场景无缝;章末钩子按文风画像的钩子风格。
|
3. 前情衔接与上一章末场景无缝;章末钩子按文风画像的钩子风格。
|
||||||
4. 缺少细纲字段、`factConstraints` 字段或篇幅合同属于 adapter 输入错误,必须在模型调用前失败。`factConstraints=[]` 在冻结检索确实没有可确认事实时是合法输入，不等于“事实已验证”。写手可以在正文里设计新设定，但不得把新设定冒充已确认事实。
|
4. 缺少细纲字段、`factConstraints` 字段或篇幅合同属于 adapter 输入错误,必须在模型调用前失败。`factConstraints=[]` 在冻结检索确实没有可确认事实时是合法输入，不等于“事实已验证”。写手可以在正文里设计新设定，但不得把新设定冒充已确认事实。
|
||||||
5. detector 只把「对已有正典/前文章节的主张检索不够」标成 `evidenceGaps` 并触发补证重写。写手新写出的设定进 `newSettingCandidates`，不因此重写或禁写；与既有正典冲突才失败关闭。新设定是否进入正典由人决定。Writer 输出不承载补证请求或审查结论。
|
5. detector 只把「对已有正典/前文章节的主张检索不够」标成 `evidenceGaps`，触发授权终态（补证/重写由人授权后以新运行继续）。写手新写出的设定进 `newSettingCandidates`，不因此重写或禁写；与既有正典冲突才失败关闭。新设定是否进入正典由人决定。Writer 输出不承载补证请求或审查结论。
|
||||||
|
|
||||||
## 生产入口
|
## 生产入口
|
||||||
|
|
||||||
@ -42,7 +42,7 @@ Dashboard 与人工生产入口统一调用 `scripts/produce_next_chapter.py`。
|
|||||||
|
|
||||||
## 生产落库
|
## 生产落库
|
||||||
|
|
||||||
- 生产编排走 `run_writer_pipeline`（机械门→语义 detector→补证/重写有限环），状态链用 `scripts/candidate_cas.py` 的 `PostgresCasStateStore` 持久化到 `example_candidate_cas`（一次运行一条链，revision 单调，DB 触发器锁方向闭集）；内存 `InMemoryCasStateStore` 仅供离线测试。
|
- 生产编排走 `run_writer_pipeline`（机械门→语义 detector→单次收敛；证据缺口收敛为 AUTHORIZATION_REQUIRED 授权终态，补证/重写经人授权后以新运行继续），状态链用 `scripts/candidate_cas.py` 的 `PostgresCasStateStore` 持久化到 `example_candidate_cas`（一次运行一条链，revision 单调，DB 触发器锁方向闭集）；内存 `InMemoryCasStateStore` 仅供离线测试。
|
||||||
- `run_writer_with_receipt()` 只负责可信 writer adapter；生产编排在组装后调用 `assemble-context/scripts/persist_context_freeze.py`，审查落库调用 `scripts/persist_writer_run.py`。
|
- `run_writer_with_receipt()` 只负责可信 writer adapter；生产编排在组装后调用 `assemble-context/scripts/persist_context_freeze.py`，审查落库调用 `scripts/persist_writer_run.py`。
|
||||||
- `persist_writer_run.py` 要求本次 `run_id` 已有成功 writer 调用的 raw 指针，随后登记 `example_candidate`（传入 `semantic_report` 时校验绑定并把 `semantic_status`/`semantic_report_sha256` 固化到候选行）、追加 `example_run_receipt` 和机械/语义两行 `example_quality_result`；它不接受正文，Shadow→Canonical 仍只能由 `decide-candidate/scripts/write_canonical.py` 完成（接受通道 DB 级兜底复检 `state=passed` 且 `semantic_status=passed`）。
|
- `persist_writer_run.py` 要求本次 `run_id` 已有成功 writer 调用的 raw 指针，随后登记 `example_candidate`（传入 `semantic_report` 时校验绑定并把 `semantic_status`/`semantic_report_sha256` 固化到候选行）、追加 `example_run_receipt` 和机械/语义两行 `example_quality_result`；它不接受正文，Shadow→Canonical 仍只能由 `decide-candidate/scripts/write_canonical.py` 完成（接受通道 DB 级兜底复检 `state=passed` 且 `semantic_status=passed`）。
|
||||||
|
|
||||||
|
|||||||
@ -5,7 +5,7 @@
|
|||||||
读已确认细纲 + 前章正文基线
|
读已确认细纲 + 前章正文基线
|
||||||
→ build_retrieval_plan + retrieve_writer_sources（生产仓储；新书无卡诚实返空）
|
→ build_retrieval_plan + retrieve_writer_sources（生产仓储；新书无卡诚实返空）
|
||||||
→ assemble_context 冻结 WriterContext v1
|
→ assemble_context 冻结 WriterContext v1
|
||||||
→ run_writer_pipeline（持久 CAS 状态链 + 机械门 + 语义 detector，有限补证/重写合同）
|
→ run_writer_pipeline（持久 CAS 状态链 + 机械门 + 语义 detector，单次收敛 + 授权终态合同）
|
||||||
writer 真调经 run_writer_with_receipt：runtime 自动把模型输入/输出原文与调用明细原子落库；
|
writer 真调经 run_writer_with_receipt：runtime 自动把模型输入/输出原文与调用明细原子落库；
|
||||||
篇幅越界在 writer 适配层自动重抽（最多三遍）；语义 detector 走冻结 detector profile 真调；
|
篇幅越界在 writer 适配层自动重抽（最多三遍）；语义 detector 走冻结 detector profile 真调；
|
||||||
语义 needs_evidence 时走 production_evidence_reassemble（检索已有正典摘录）；零命中当新设定交人闸，不禁写不重写
|
语义 needs_evidence 时走 production_evidence_reassemble（检索已有正典摘录）；零命中当新设定交人闸，不禁写不重写
|
||||||
@ -524,6 +524,12 @@ def main():
|
|||||||
print(f"[警告] 被拒候选留库失败: {persist_exc}", file=sys.stderr)
|
print(f"[警告] 被拒候选留库失败: {persist_exc}", file=sys.stderr)
|
||||||
finish_run(run_id, "failed", creator="continuation",
|
finish_run(run_id, "failed", creator="continuation",
|
||||||
trigger_detail={"stage": "writer-production-pipeline", "failureCode": exc.code})
|
trigger_detail={"stage": "writer-production-pipeline", "failureCode": exc.code})
|
||||||
|
if exc.code == "AUTHORIZATION_REQUIRED":
|
||||||
|
gaps = (exc.details or {}).get("evidenceGaps") or []
|
||||||
|
print("[需要授权] 语义检查发现证据缺口，补证或重写需要人授权：")
|
||||||
|
for item in gaps:
|
||||||
|
print(f" - {item.get('gapId')}: {item.get('reason')}（检索：{item.get('query')}）")
|
||||||
|
print("授权后由主代理发起新运行继续补证（本运行已收敛 REJECTED，新运行接续候选版本）。")
|
||||||
print(f"[停止] 生产 pipeline 未通过: code={exc.code}；{exc}")
|
print(f"[停止] 生产 pipeline 未通过: code={exc.code}；{exc}")
|
||||||
print(f"复核 artifacts/{run_id}-pipeline-result.json 后决定下一步。")
|
print(f"复核 artifacts/{run_id}-pipeline-result.json 后决定下一步。")
|
||||||
print(f"\nRUN_ID={run_id}")
|
print(f"\nRUN_ID={run_id}")
|
||||||
|
|||||||
@ -1,5 +1,5 @@
|
|||||||
#!/usr/bin/env python3
|
#!/usr/bin/env python3
|
||||||
"""正文候选的有限补证、重写、机械审查与 CAS 编排。"""
|
"""正文候选的单次收敛编排：机械审查、语义检查、授权终态与 CAS。"""
|
||||||
|
|
||||||
from __future__ import annotations
|
from __future__ import annotations
|
||||||
|
|
||||||
@ -30,8 +30,6 @@ from writer_contract import ( # noqa: E402
|
|||||||
)
|
)
|
||||||
|
|
||||||
|
|
||||||
MAX_EVIDENCE_REQUESTS = 3
|
|
||||||
MAX_REWRITES = 2
|
|
||||||
_HASH_PATTERN = re.compile(r"^sha256:[0-9a-f]{64}$")
|
_HASH_PATTERN = re.compile(r"^sha256:[0-9a-f]{64}$")
|
||||||
|
|
||||||
|
|
||||||
@ -440,7 +438,12 @@ def run_writer_pipeline(
|
|||||||
result_path: str | pathlib.Path | None = None,
|
result_path: str | pathlib.Path | None = None,
|
||||||
initial_candidate_version: int = 1,
|
initial_candidate_version: int = 1,
|
||||||
) -> dict[str, Any]:
|
) -> dict[str, Any]:
|
||||||
"""执行最多 3 次补证、2 次重写的正文候选有限收敛循环。
|
"""正文候选单次收敛：写手 -> 机械门 -> 语义检查，全分支终态化。
|
||||||
|
|
||||||
|
语义检查发现证据缺口时先做确定性检索（evidence_provider，只读）：
|
||||||
|
零命中升格为新设定提案进人闸；命中则收敛为 AUTHORIZATION_REQUIRED 终态，
|
||||||
|
携带缺口报告与重组上下文哈希，由编排方在人授权后发起新运行继续重写。
|
||||||
|
补证/重写业务次数不设硬上限，管控归人；授权门只挡模型重调用，不挡检索。
|
||||||
|
|
||||||
initial_candidate_version 给生产编排接续既有版本号：候选表对
|
initial_candidate_version 给生产编排接续既有版本号：候选表对
|
||||||
(作品, 章, candidate_version) 唯一，同一章重跑必须从「已有最大版本+1」起，
|
(作品, 章, candidate_version) 唯一，同一章重跑必须从「已有最大版本+1」起，
|
||||||
@ -529,7 +532,7 @@ def run_writer_pipeline(
|
|||||||
try:
|
try:
|
||||||
candidate = validate_writer_output(candidate)
|
candidate = validate_writer_output(candidate)
|
||||||
except ContractError as exc:
|
except ContractError as exc:
|
||||||
# 合同错误作为可定位审查失败进入有限重写，而不是绕过状态机。
|
# 合同错误作为可定位审查失败进入机械门定位，而不是绕过状态机。
|
||||||
try:
|
try:
|
||||||
mechanical_report = check_writer_candidate(current_context, candidate, requirements)
|
mechanical_report = check_writer_candidate(current_context, candidate, requirements)
|
||||||
candidate_failures = _validate_mechanical_report(mechanical_report, candidate)
|
candidate_failures = _validate_mechanical_report(mechanical_report, candidate)
|
||||||
@ -659,36 +662,11 @@ def run_writer_pipeline(
|
|||||||
result_path=result_path,
|
result_path=result_path,
|
||||||
) from exc
|
) from exc
|
||||||
if evidence_gaps:
|
if evidence_gaps:
|
||||||
if evidence_request_count + len(evidence_gaps) > MAX_EVIDENCE_REQUESTS:
|
# 缺口先做确定性检索（只读、无模型开销）；结果决定收敛方向：
|
||||||
raise _terminal_failure(
|
# 零命中 = 新设定提案，升格 passed 进人闸；命中 = 需要模型重写，
|
||||||
code="EVIDENCE_REQUEST_LIMIT_REACHED",
|
# 停在授权终态等人授权，由编排方以新运行继续，不自动重写。
|
||||||
message="detector 证据缺口累计超过 3 次",
|
|
||||||
token=checking,
|
|
||||||
state_store=state_store,
|
|
||||||
run_id=run_id,
|
|
||||||
evidence_request_count=evidence_request_count,
|
|
||||||
rewrite_count=rewrite_count,
|
|
||||||
trace=trace,
|
|
||||||
candidate=candidate,
|
|
||||||
result_path=result_path,
|
|
||||||
)
|
|
||||||
if rewrite_count >= MAX_REWRITES:
|
|
||||||
raise _terminal_failure(
|
|
||||||
code="REWRITE_LIMIT_REACHED",
|
|
||||||
message="补证后重写已达到 2 次",
|
|
||||||
token=checking,
|
|
||||||
state_store=state_store,
|
|
||||||
run_id=run_id,
|
|
||||||
evidence_request_count=evidence_request_count,
|
|
||||||
rewrite_count=rewrite_count,
|
|
||||||
trace=trace,
|
|
||||||
candidate=candidate,
|
|
||||||
result_path=result_path,
|
|
||||||
)
|
|
||||||
evidence_request_count += len(evidence_gaps)
|
|
||||||
next_attempt = current_context["attempt"] + 1
|
next_attempt = current_context["attempt"] + 1
|
||||||
previous_snapshot = current_context["contextSnapshot"]["contextSha256"]
|
previous_snapshot = current_context["contextSnapshot"]["contextSha256"]
|
||||||
previous_creative_input = build_writer_creative_input(current_context)
|
|
||||||
previous_fact_ids = {
|
previous_fact_ids = {
|
||||||
str(item.get("evidenceId"))
|
str(item.get("evidenceId"))
|
||||||
for item in (current_context.get("factEvidence") or [])
|
for item in (current_context.get("factEvidence") or [])
|
||||||
@ -719,9 +697,9 @@ def run_writer_pipeline(
|
|||||||
for item in (next_context.get("factEvidence") or [])
|
for item in (next_context.get("factEvidence") or [])
|
||||||
if isinstance(item, Mapping)
|
if isinstance(item, Mapping)
|
||||||
}
|
}
|
||||||
# 正典检索零命中 = 缺口是新设定提案，不是可补的检索债。
|
|
||||||
# 无语义冲突时升格为 passed 交人闸；有冲突则不重写、走失败关闭。
|
|
||||||
if not (next_fact_ids - previous_fact_ids):
|
if not (next_fact_ids - previous_fact_ids):
|
||||||
|
# 正典检索零命中 = 缺口是新设定提案，不是可补的检索债。
|
||||||
|
# 无语义冲突时升格为 passed 交人闸；有冲突则不升格、走失败关闭。
|
||||||
if not candidate_failures and semantic_report is not None:
|
if not candidate_failures and semantic_report is not None:
|
||||||
semantic_report = _promote_unfillable_gaps_to_settings(semantic_report)
|
semantic_report = _promote_unfillable_gaps_to_settings(semantic_report)
|
||||||
evidence_gaps = []
|
evidence_gaps = []
|
||||||
@ -730,7 +708,6 @@ def run_writer_pipeline(
|
|||||||
next_context["runId"] != run_id
|
next_context["runId"] != run_id
|
||||||
or next_context["attempt"] != next_attempt
|
or next_context["attempt"] != next_attempt
|
||||||
or next_context["contextSnapshot"]["contextSha256"] == previous_snapshot
|
or next_context["contextSnapshot"]["contextSha256"] == previous_snapshot
|
||||||
or build_writer_creative_input(next_context) == previous_creative_input
|
|
||||||
):
|
):
|
||||||
raise _terminal_failure(
|
raise _terminal_failure(
|
||||||
code="REASSEMBLED_CONTEXT_INVALID",
|
code="REASSEMBLED_CONTEXT_INVALID",
|
||||||
@ -744,28 +721,35 @@ def run_writer_pipeline(
|
|||||||
candidate=candidate,
|
candidate=candidate,
|
||||||
result_path=result_path,
|
result_path=result_path,
|
||||||
)
|
)
|
||||||
rejected = _cas_or_fail(
|
|
||||||
state_store.transition(checking, "REJECTED"), "CHECKING -> REJECTED"
|
|
||||||
)
|
|
||||||
rewrite_count += 1
|
|
||||||
candidate_version += 1
|
|
||||||
# 补证环的审计痕迹同样携带机械门报告，被拒候选留库不因走补证路径而丢失证据。
|
|
||||||
trace.append({
|
trace.append({
|
||||||
"attempt": checking.attempt,
|
"attempt": checking.attempt,
|
||||||
"candidateVersion": checking.candidate_version,
|
"candidateVersion": checking.candidate_version,
|
||||||
"candidateSha256": candidate["candidateSha256"],
|
"candidateSha256": candidate["candidateSha256"],
|
||||||
"status": "evidence_gap",
|
"status": "needs_authorization",
|
||||||
"gapIds": [item["gapId"] for item in evidence_gaps],
|
"gapIds": [item["gapId"] for item in evidence_gaps],
|
||||||
"mechanicalPassed": mechanical_report.get("passed", False),
|
"mechanicalPassed": mechanical_report.get("passed", False),
|
||||||
"mechanicalReport": dict(mechanical_report),
|
"mechanicalReport": dict(mechanical_report),
|
||||||
"semanticReport": dict(semantic_report),
|
"semanticReport": dict(semantic_report) if semantic_report else None,
|
||||||
})
|
})
|
||||||
current_context = next_context
|
raise _terminal_failure(
|
||||||
draft = _cas_or_fail(
|
code="AUTHORIZATION_REQUIRED",
|
||||||
state_store.start_next(rejected, attempt=next_attempt, candidate_version=candidate_version),
|
message="补证检索命中，重写需要人授权（以新运行继续）",
|
||||||
"REJECTED -> next DRAFT",
|
token=checking,
|
||||||
|
state_store=state_store,
|
||||||
|
run_id=run_id,
|
||||||
|
evidence_request_count=evidence_request_count,
|
||||||
|
rewrite_count=rewrite_count,
|
||||||
|
trace=trace,
|
||||||
|
candidate=candidate,
|
||||||
|
result_path=result_path,
|
||||||
|
details={
|
||||||
|
"evidenceGaps": [dict(item) for item in evidence_gaps],
|
||||||
|
"mechanicalPassed": mechanical_report.get("passed", False),
|
||||||
|
"semanticStatus": semantic_report.get("status") if semantic_report else None,
|
||||||
|
"nextAttempt": next_attempt,
|
||||||
|
"reassembledContextSha256": next_context["contextSnapshot"]["contextSha256"],
|
||||||
|
},
|
||||||
)
|
)
|
||||||
continue
|
|
||||||
trace.append(
|
trace.append(
|
||||||
{
|
{
|
||||||
"attempt": checking.attempt,
|
"attempt": checking.attempt,
|
||||||
@ -824,8 +808,6 @@ def run_writer_pipeline(
|
|||||||
|
|
||||||
|
|
||||||
__all__ = [
|
__all__ = [
|
||||||
"MAX_EVIDENCE_REQUESTS",
|
|
||||||
"MAX_REWRITES",
|
|
||||||
"CasToken",
|
"CasToken",
|
||||||
"CasStateStore",
|
"CasStateStore",
|
||||||
"InMemoryCasStateStore",
|
"InMemoryCasStateStore",
|
||||||
|
|||||||
@ -2520,7 +2520,7 @@ def _run_decision_menu_panel(run_id: str, candidates: list, terminal_state, pipe
|
|||||||
f"写库决策请用 <a href='http://127.0.0.1:8767/candidates/{esc(cid)}'>决策通道 :8767</a>"
|
f"写库决策请用 <a href='http://127.0.0.1:8767/candidates/{esc(cid)}'>决策通道 :8767</a>"
|
||||||
f"（看板只读）</p></div>"
|
f"（看板只读）</p></div>"
|
||||||
)
|
)
|
||||||
# 无候选行：本 run 失败关闭常见于 REASSEMBLED_CONTEXT_INVALID
|
# 无候选行：本 run 失败关闭常见于 AUTHORIZATION_REQUIRED / 语义或机械拒绝
|
||||||
return (
|
return (
|
||||||
"<div class='card' style='border-color:var(--serious,#a40)'><div class='hd'>决策菜单</div>"
|
"<div class='card' style='border-color:var(--serious,#a40)'><div class='hd'>决策菜单</div>"
|
||||||
"<div style='padding:12px 16px'>"
|
"<div style='padding:12px 16px'>"
|
||||||
|
|||||||
50
docs/plans/2026-08-22-阶段E-生产链切换-第一部分-授权终态.md
Normal file
50
docs/plans/2026-08-22-阶段E-生产链切换-第一部分-授权终态.md
Normal file
@ -0,0 +1,50 @@
|
|||||||
|
# 阶段 E：生产链切换（第一部分：授权终态合同）
|
||||||
|
|
||||||
|
日期：2026-08-22
|
||||||
|
总 plan：[2026-08-22-agent-example整体收敛总plan.md](2026-08-22-agent-example整体收敛总plan.md)
|
||||||
|
状态：第一部分完成并提交；第二部分（写作/检测智能体接入框架派发、授权后继续的接线、真实生产烟测）另行启动，真实模型调用需人工授权。
|
||||||
|
|
||||||
|
## 1. 意图
|
||||||
|
|
||||||
|
删除补证/重写的业务硬上限（补证≤3 / 重写≤2），改为授权终态：检查发现缺口后系统停在结构化报告，由主代理转述给人，人授权后以新运行继续。技术保护（超时、预算、取消）保留在执行策略层。
|
||||||
|
|
||||||
|
## 2. 合同设计（本部分核心决策）
|
||||||
|
|
||||||
|
授权门只挡昂贵的模型重调用，不挡廉价的确定性检索：
|
||||||
|
|
||||||
|
```text
|
||||||
|
语义检查发现证据缺口
|
||||||
|
→ 确定性检索（evidence_provider，只读、无模型开销）
|
||||||
|
├─ 零命中 = 新设定提案 → 升格 passed 进人闸（原行为保留）
|
||||||
|
└─ 命中 = 需要模型重写 → AUTHORIZATION_REQUIRED 授权终态
|
||||||
|
携带：缺口清单、机械/语义检查快照、下一 attempt、重组上下文哈希
|
||||||
|
人授权后：编排方以新运行继续（新候选版本，会话可复用）
|
||||||
|
```
|
||||||
|
|
||||||
|
依据：CAS 链按运行唯一（`ON CONFLICT (run_id) DO NOTHING`），同运行不能开新轮；授权后的继续天然是新运行，与"一章复用同一写作智能体会话"通过框架会话续接（阶段 D 能力）组合。
|
||||||
|
|
||||||
|
## 3. 文件台账
|
||||||
|
|
||||||
|
| 处置 | 文件 | 原因 |
|
||||||
|
|---|---|---|
|
||||||
|
| 修改 | `run_writer_pipeline.py` | 删除 `MAX_EVIDENCE_REQUESTS/MAX_REWRITES` 常量、自动补证/重写循环与两个上限错误码；缺口分支改为"检索→零命中升格/命中授权终态"；单次收敛文档化。CAS 基元（含 `start_next`）保留，状态机闭集不动 |
|
||||||
|
| 修改 | `produce_next_chapter.py` | AUTHORIZATION_REQUIRED 显式分支：输出结构化缺口报告与授权指引；模块头口径同步 |
|
||||||
|
| 修改 | `test_run_writer_pipeline.py` | 两个自动补证测试改写为授权终态合同（缺口停终态、授权报告落结果文件）；provider 改为真实检索推进 |
|
||||||
|
| 修改 | `test_production_evidence_reassemble.py` | 命中重写测试改为授权终态（写手单次调用、details 带 nextAttempt 与重组上下文哈希）；零命中升格测试不变 |
|
||||||
|
| 修改 | `test_candidate_cas.py` | CAS 上的补证环测试改为授权终态链断言（create→CHECKING→REJECTED，revision=3） |
|
||||||
|
| 修改 | `test_writer_acceptance.py` | 失败码 fixture 对齐（REWRITE_LIMIT_REACHED → AUTHORIZATION_REQUIRED） |
|
||||||
|
| 修改 | `05-创作流程领域`、`write-next-chapter/SKILL.md`、`dashboard/server.py` 注释 | 补证口径全仓同步（检索语义不变、命中停授权终态） |
|
||||||
|
| 保留 | `production_evidence_reassemble.py` | 确定性检索是新合同的组成部分（授权前置检索），职责不变 |
|
||||||
|
| 保留 | `start_next` 等 CAS 基元 | 总 plan 不重造状态机；续跑走新运行的 `create` 链 |
|
||||||
|
| 遗留 | 授权后继续的编排接线（新运行 + 重组上下文 + 会话续接） | 阶段 E 第二部分 |
|
||||||
|
| 遗留 | 写作/检测智能体框架派发接入 | 阶段 E 第二部分；真实烟测需人工授权 |
|
||||||
|
|
||||||
|
## 4. 验证
|
||||||
|
|
||||||
|
- 管线测试：10 项通过（含授权终态 2 项新合同测试）。
|
||||||
|
- 补证重组装测试：5 项通过（命中授权终态 + 零命中升格）。
|
||||||
|
- CAS 测试：离线全通过；真实库集成测试（PostgresCasStateStore）全部通过。
|
||||||
|
- 候选接受测试：通过（失败码对齐后）。
|
||||||
|
- 全量离线清单：102 通过、8 项外部依赖阻断、0 失败。
|
||||||
|
- 技能严格审计：58 个 Skill，阻断 0；索引一致性、架构门禁、`git diff --check` 通过。
|
||||||
|
- 旧口径零残留：`补证≤3 / 重写≤2 / EVIDENCE_REQUEST_LIMIT / REWRITE_LIMIT` 全仓检索无生产代码引用（写手篇幅适配器的"≤3 遍重抽"是独立技术重试合同，不属于补证/重写，保留）。
|
||||||
@ -327,7 +327,7 @@ class WriterAcceptanceTest(unittest.TestCase):
|
|||||||
|
|
||||||
rejected_detector = copy.deepcopy(detector)
|
rejected_detector = copy.deepcopy(detector)
|
||||||
rejected_detector["status"] = "REJECTED"
|
rejected_detector["status"] = "REJECTED"
|
||||||
rejected_detector["failureCode"] = "REWRITE_LIMIT_REACHED"
|
rejected_detector["failureCode"] = "AUTHORIZATION_REQUIRED"
|
||||||
cases.append(("DETECTOR_NOT_PASSED", candidate, rejected_detector, live))
|
cases.append(("DETECTOR_NOT_PASSED", candidate, rejected_detector, live))
|
||||||
|
|
||||||
wrong_detector_schema = copy.deepcopy(detector)
|
wrong_detector_schema = copy.deepcopy(detector)
|
||||||
|
|||||||
@ -207,45 +207,46 @@ class PipelineOnPostgresStoreTest(unittest.TestCase):
|
|||||||
self.assertEqual(chain["state"], "PASSED")
|
self.assertEqual(chain["state"], "PASSED")
|
||||||
self.assertEqual(chain["revision"], 3) # create→CHECKING→PASSED
|
self.assertEqual(chain["revision"], 3) # create→CHECKING→PASSED
|
||||||
|
|
||||||
def test_pipeline_evidence_loop_uses_start_next(self) -> None:
|
def test_pipeline_evidence_hit_stops_at_authorization(self) -> None:
|
||||||
|
"""新合同：补证检索命中后停在授权终态，CAS 收敛 REJECTED，不开新轮。"""
|
||||||
context, _ = _valid_pair()
|
context, _ = _valid_pair()
|
||||||
rows: dict = {}
|
rows: dict = {}
|
||||||
|
|
||||||
def detector(current: dict, candidate: dict, _mechanical: dict) -> dict:
|
def detector(current: dict, candidate: dict, _mechanical: dict) -> dict:
|
||||||
if candidate["candidateVersion"] == 1:
|
report = {
|
||||||
report = {
|
"schemaVersion": "semantic-detection-v3",
|
||||||
"schemaVersion": "semantic-detection-v3",
|
"runId": current["runId"], "sampleId": "cas-sample",
|
||||||
"runId": current["runId"], "sampleId": "cas-sample",
|
"opaqueArmId": "cas-candidate", "inputSha256": "sha256:" + "1" * 64,
|
||||||
"opaqueArmId": "cas-candidate", "inputSha256": "sha256:" + "1" * 64,
|
"candidateVersion": candidate["candidateVersion"],
|
||||||
"candidateVersion": candidate["candidateVersion"],
|
"candidateSha256": candidate["candidateSha256"],
|
||||||
|
"contextSnapshotSha256": current["contextSnapshot"]["contextSha256"],
|
||||||
|
"modelReceiptSha256": "sha256:" + "2" * 64, "status": "needs_evidence",
|
||||||
|
"claims": [], "findings": [], "assertionVerdicts": [],
|
||||||
|
"hardConstraintVerdicts": [], "newSettingCandidates": [],
|
||||||
|
"evidenceGaps": [{
|
||||||
|
"gapId": "gap-1", "query": "补证", "reason": "缺口", "priority": "high",
|
||||||
"candidateSha256": candidate["candidateSha256"],
|
"candidateSha256": candidate["candidateSha256"],
|
||||||
"contextSnapshotSha256": current["contextSnapshot"]["contextSha256"],
|
"candidateQuote": "旧徽章", "startCodePoint": 11, "endCodePoint": 14,
|
||||||
"modelReceiptSha256": "sha256:" + "2" * 64, "status": "needs_evidence",
|
}],
|
||||||
"claims": [], "findings": [], "assertionVerdicts": [],
|
}
|
||||||
"hardConstraintVerdicts": [], "newSettingCandidates": [],
|
report["reportSha256"] = canonical_sha256(report)
|
||||||
"evidenceGaps": [{
|
return report
|
||||||
"gapId": "gap-1", "query": "补证", "reason": "缺口", "priority": "high",
|
|
||||||
"candidateSha256": candidate["candidateSha256"],
|
|
||||||
"candidateQuote": "旧徽章", "startCodePoint": 11, "endCodePoint": 14,
|
|
||||||
}],
|
|
||||||
}
|
|
||||||
report["reportSha256"] = canonical_sha256(report)
|
|
||||||
return report
|
|
||||||
return _semantic_pass(current, candidate, _mechanical)
|
|
||||||
|
|
||||||
result = run_writer_pipeline(
|
with self.assertRaises(PipelineError) as caught:
|
||||||
context=context,
|
run_writer_pipeline(
|
||||||
requirements=_requirements(),
|
context=context,
|
||||||
writer=lambda current, version: _writer_output(current, version),
|
requirements=_requirements(),
|
||||||
evidence_provider=lambda current, _gaps, attempt: _advance_context(current, attempt),
|
writer=lambda current, version: _writer_output(current, version),
|
||||||
semantic_detector=detector,
|
evidence_provider=lambda current, _gaps, attempt: _advance_context(current, attempt),
|
||||||
state_store=_store(rows),
|
semantic_detector=detector,
|
||||||
)
|
state_store=_store(rows),
|
||||||
self.assertEqual(result["status"], "PASSED")
|
)
|
||||||
|
self.assertEqual(caught.exception.code, "AUTHORIZATION_REQUIRED")
|
||||||
|
self.assertEqual(caught.exception.details["nextAttempt"], 2)
|
||||||
chain = rows[context["runId"]]
|
chain = rows[context["runId"]]
|
||||||
# create→CHECKING→REJECTED→新DRAFT→CHECKING→PASSED
|
# create→CHECKING→REJECTED：授权终态不开同运行新轮，续跑是新运行的事。
|
||||||
self.assertEqual((chain["state"], chain["revision"], chain["attempt"],
|
self.assertEqual((chain["state"], chain["revision"], chain["attempt"],
|
||||||
chain["candidate_version"]), ("PASSED", 6, 2, 2))
|
chain["candidate_version"]), ("REJECTED", 3, 1, 1))
|
||||||
|
|
||||||
|
|
||||||
if __name__ == "__main__":
|
if __name__ == "__main__":
|
||||||
|
|||||||
@ -21,7 +21,7 @@ from production_evidence_reassemble import ( # noqa: E402
|
|||||||
EvidenceReassembleError,
|
EvidenceReassembleError,
|
||||||
reassemble_writer_context_for_gaps,
|
reassemble_writer_context_for_gaps,
|
||||||
)
|
)
|
||||||
from run_writer_pipeline import InMemoryCasStateStore, run_writer_pipeline # noqa: E402
|
from run_writer_pipeline import InMemoryCasStateStore, PipelineError, run_writer_pipeline # noqa: E402
|
||||||
from writer_contract import build_writer_creative_input, retrieval_identity # noqa: E402
|
from writer_contract import build_writer_creative_input, retrieval_identity # noqa: E402
|
||||||
|
|
||||||
from test_check_writer_candidate import _requirements, _valid_pair # noqa: E402
|
from test_check_writer_candidate import _requirements, _valid_pair # noqa: E402
|
||||||
@ -98,8 +98,8 @@ class ProductionEvidenceReassembleTests(unittest.TestCase):
|
|||||||
work_id=12,
|
work_id=12,
|
||||||
)
|
)
|
||||||
|
|
||||||
def test_pipeline_needs_evidence_then_pass_with_reassemble(self) -> None:
|
def test_pipeline_evidence_hit_stops_at_authorization(self) -> None:
|
||||||
"""needs_evidence + 检索命中 → 重写 → 再检 passed。"""
|
"""needs_evidence + 检索命中 -> 授权终态（重写需人授权，以新运行继续）。"""
|
||||||
|
|
||||||
context, _ = _valid_pair()
|
context, _ = _valid_pair()
|
||||||
calls: list[int] = []
|
calls: list[int] = []
|
||||||
@ -121,25 +121,23 @@ class ProductionEvidenceReassembleTests(unittest.TestCase):
|
|||||||
return _writer_output(current, version)
|
return _writer_output(current, version)
|
||||||
|
|
||||||
def detector(current: dict, candidate: dict, _mechanical: dict) -> dict:
|
def detector(current: dict, candidate: dict, _mechanical: dict) -> dict:
|
||||||
if current["attempt"] == 1:
|
return _semantic_report(
|
||||||
return _semantic_report(
|
current,
|
||||||
current,
|
candidate,
|
||||||
candidate,
|
status="needs_evidence",
|
||||||
status="needs_evidence",
|
gaps=[
|
||||||
gaps=[
|
{
|
||||||
{
|
"gapId": "gap-1",
|
||||||
"gapId": "gap-1",
|
"query": "旧徽章归属",
|
||||||
"query": "旧徽章归属",
|
"reason": "缺口",
|
||||||
"reason": "缺口",
|
"priority": "high",
|
||||||
"priority": "high",
|
"candidateSha256": candidate["candidateSha256"],
|
||||||
"candidateSha256": candidate["candidateSha256"],
|
"candidateQuote": "旧徽章",
|
||||||
"candidateQuote": "旧徽章",
|
"startCodePoint": 11,
|
||||||
"startCodePoint": 11,
|
"endCodePoint": 14,
|
||||||
"endCodePoint": 14,
|
}
|
||||||
}
|
],
|
||||||
],
|
)
|
||||||
)
|
|
||||||
return _semantic_report(current, candidate, status="passed")
|
|
||||||
|
|
||||||
def provider(current, gaps, attempt):
|
def provider(current, gaps, attempt):
|
||||||
with patch(
|
with patch(
|
||||||
@ -150,17 +148,20 @@ class ProductionEvidenceReassembleTests(unittest.TestCase):
|
|||||||
current, gaps, attempt, work_id=int(current["workId"])
|
current, gaps, attempt, work_id=int(current["workId"])
|
||||||
)
|
)
|
||||||
|
|
||||||
result = run_writer_pipeline(
|
with self.assertRaises(PipelineError) as caught:
|
||||||
context=context,
|
run_writer_pipeline(
|
||||||
requirements=_requirements(),
|
context=context,
|
||||||
writer=writer,
|
requirements=_requirements(),
|
||||||
evidence_provider=provider,
|
writer=writer,
|
||||||
semantic_detector=detector,
|
evidence_provider=provider,
|
||||||
state_store=InMemoryCasStateStore(),
|
semantic_detector=detector,
|
||||||
)
|
state_store=InMemoryCasStateStore(),
|
||||||
self.assertEqual(result["status"], "PASSED")
|
)
|
||||||
self.assertEqual(calls, [1, 2])
|
self.assertEqual(caught.exception.code, "AUTHORIZATION_REQUIRED")
|
||||||
self.assertGreaterEqual(result.get("rewriteCount", 0), 1)
|
# 只跑单次收敛：写手只被调用一次，重写等人授权。
|
||||||
|
self.assertEqual(calls, [1])
|
||||||
|
self.assertEqual(caught.exception.details["nextAttempt"], 2)
|
||||||
|
self.assertTrue(caught.exception.details["reassembledContextSha256"].startswith("sha256:"))
|
||||||
|
|
||||||
def test_unfillable_gaps_pass_without_rewrite(self) -> None:
|
def test_unfillable_gaps_pass_without_rewrite(self) -> None:
|
||||||
"""正典零命中的缺口是新设定提案：不重写，报告升格 passed，进人闸。"""
|
"""正典零命中的缺口是新设定提案：不重写，报告升格 passed，进人闸。"""
|
||||||
|
|||||||
@ -109,45 +109,45 @@ class RunWriterPipelineV3Test(unittest.TestCase):
|
|||||||
for forbidden in ("claimLedger", "evidenceRequests", "newSettingDeclarations"):
|
for forbidden in ("claimLedger", "evidenceRequests", "newSettingDeclarations"):
|
||||||
self.assertNotIn(forbidden, result["candidateArtifact"])
|
self.assertNotIn(forbidden, result["candidateArtifact"])
|
||||||
|
|
||||||
def test_only_detector_evidence_gaps_trigger_fresh_attempt(self) -> None:
|
def test_evidence_gaps_stop_at_authorization_terminal(self) -> None:
|
||||||
|
"""新合同：补证检索命中后收敛为 AUTHORIZATION_REQUIRED，重写等人授权。"""
|
||||||
context, _ = _valid_pair()
|
context, _ = _valid_pair()
|
||||||
writer_calls: list[tuple[int, int, str]] = []
|
writer_calls: list[tuple[int, int]] = []
|
||||||
detector_calls: list[int] = []
|
|
||||||
provider_calls: list[list[str]] = []
|
|
||||||
|
|
||||||
def writer(current: dict, version: int) -> dict:
|
def writer(current: dict, version: int) -> dict:
|
||||||
writer_calls.append((current["attempt"], version, current["contextSnapshot"]["contextSha256"]))
|
writer_calls.append((current["attempt"], version))
|
||||||
return _writer_output(current, version)
|
return _writer_output(current, version)
|
||||||
|
|
||||||
def detector(current: dict, candidate: dict, _mechanical: dict) -> dict:
|
def detector(current: dict, candidate: dict, _mechanical: dict) -> dict:
|
||||||
detector_calls.append(candidate["candidateVersion"])
|
return _semantic_report(current, candidate, status="needs_evidence", gaps=[{
|
||||||
if candidate["candidateVersion"] == 1:
|
"gapId": "gap-1", "query": "旧徽章来源", "reason": "证据不足", "priority": "high",
|
||||||
return _semantic_report(current, candidate, status="needs_evidence", gaps=[{
|
"candidateSha256": candidate["candidateSha256"], "candidateQuote": "旧徽章",
|
||||||
"gapId": "gap-1", "query": "旧徽章来源", "reason": "证据不足", "priority": "high",
|
"startCodePoint": 11, "endCodePoint": 14,
|
||||||
"candidateSha256": candidate["candidateSha256"], "candidateQuote": "旧徽章",
|
}])
|
||||||
"startCodePoint": 11, "endCodePoint": 14,
|
|
||||||
}])
|
|
||||||
return _semantic_report(current, candidate)
|
|
||||||
|
|
||||||
def provider(current: dict, gaps: list[dict], attempt: int) -> dict:
|
store = InMemoryCasStateStore()
|
||||||
provider_calls.append([item["gapId"] for item in gaps])
|
with tempfile.TemporaryDirectory() as directory:
|
||||||
return _advance_context(current, attempt, add_evidence=True)
|
result_path = pathlib.Path(directory) / "result.json"
|
||||||
|
with self.assertRaises(PipelineError) as caught:
|
||||||
result = run_writer_pipeline(
|
run_writer_pipeline(
|
||||||
context=context,
|
context=context,
|
||||||
requirements=_requirements(),
|
requirements=_requirements(),
|
||||||
writer=writer,
|
writer=writer,
|
||||||
evidence_provider=provider,
|
evidence_provider=lambda current, _gaps, attempt: _advance_context(current, attempt, add_evidence=True),
|
||||||
semantic_detector=detector,
|
semantic_detector=detector,
|
||||||
state_store=InMemoryCasStateStore(),
|
state_store=store,
|
||||||
)
|
result_path=result_path,
|
||||||
self.assertEqual(result["status"], "PASSED")
|
)
|
||||||
self.assertEqual(writer_calls[0][:2], (1, 1))
|
result = json.loads(result_path.read_text(encoding="utf-8"))
|
||||||
self.assertEqual(writer_calls[1][:2], (2, 2))
|
self.assertEqual(caught.exception.code, "AUTHORIZATION_REQUIRED")
|
||||||
self.assertNotEqual(writer_calls[0][2], writer_calls[1][2])
|
# 只跑单次收敛：写手只被调用一次，无自动第二轮。
|
||||||
self.assertEqual(detector_calls, [1, 2])
|
self.assertEqual(writer_calls, [(1, 1)])
|
||||||
self.assertEqual(provider_calls, [["gap-1"]])
|
self.assertEqual(store.latest(context["runId"]).state, "REJECTED")
|
||||||
self.assertEqual(result["evidenceRequestCount"], 1)
|
# 授权终态携带结构化缺口报告，供主代理转述给人。
|
||||||
|
gaps = caught.exception.details["evidenceGaps"]
|
||||||
|
self.assertEqual([item["gapId"] for item in gaps], ["gap-1"])
|
||||||
|
self.assertEqual(result["failureCode"], "AUTHORIZATION_REQUIRED")
|
||||||
|
self.assertEqual(result["status"], "REJECTED")
|
||||||
|
|
||||||
def test_writer_cannot_trigger_evidence_provider(self) -> None:
|
def test_writer_cannot_trigger_evidence_provider(self) -> None:
|
||||||
context, _ = _valid_pair()
|
context, _ = _valid_pair()
|
||||||
@ -165,7 +165,8 @@ class RunWriterPipelineV3Test(unittest.TestCase):
|
|||||||
code="MECHANICAL_DETECTION_REJECTED",
|
code="MECHANICAL_DETECTION_REJECTED",
|
||||||
)
|
)
|
||||||
|
|
||||||
def test_reassembled_context_must_have_new_snapshot_and_attempt(self) -> None:
|
def test_authorization_terminal_publishes_gap_report(self) -> None:
|
||||||
|
"""授权终态结果文件必须带缺口与检查快照，编排方据此向人请求授权。"""
|
||||||
context, _ = _valid_pair()
|
context, _ = _valid_pair()
|
||||||
|
|
||||||
def detector(current: dict, candidate: dict, _mechanical: dict) -> dict:
|
def detector(current: dict, candidate: dict, _mechanical: dict) -> dict:
|
||||||
@ -175,15 +176,24 @@ class RunWriterPipelineV3Test(unittest.TestCase):
|
|||||||
"startCodePoint": 11, "endCodePoint": 14,
|
"startCodePoint": 11, "endCodePoint": 14,
|
||||||
}])
|
}])
|
||||||
|
|
||||||
same_snapshot = copy.deepcopy(context)
|
store = InMemoryCasStateStore()
|
||||||
same_snapshot["attempt"] = 2
|
with tempfile.TemporaryDirectory() as directory:
|
||||||
self._assert_failure(
|
result_path = pathlib.Path(directory) / "result.json"
|
||||||
context=context,
|
with self.assertRaises(PipelineError) as caught:
|
||||||
writer=lambda current, version: _writer_output(current, version),
|
run_writer_pipeline(
|
||||||
detector=detector,
|
context=context,
|
||||||
provider=lambda *_args: same_snapshot,
|
requirements=_requirements(),
|
||||||
code="REASSEMBLED_CONTEXT_INVALID",
|
writer=lambda current, version: _writer_output(current, version),
|
||||||
)
|
evidence_provider=lambda current, _gaps, attempt: _advance_context(current, attempt, add_evidence=True),
|
||||||
|
semantic_detector=detector,
|
||||||
|
state_store=store,
|
||||||
|
result_path=result_path,
|
||||||
|
)
|
||||||
|
result = json.loads(result_path.read_text(encoding="utf-8"))
|
||||||
|
self.assertEqual(caught.exception.code, "AUTHORIZATION_REQUIRED")
|
||||||
|
self.assertEqual(result["candidateSha256"], caught.exception.result["candidateSha256"])
|
||||||
|
self.assertEqual(caught.exception.details["mechanicalPassed"], True)
|
||||||
|
self.assertEqual(caught.exception.details["semanticStatus"], "needs_evidence")
|
||||||
|
|
||||||
def test_old_semantic_v2_report_fails_closed(self) -> None:
|
def test_old_semantic_v2_report_fails_closed(self) -> None:
|
||||||
context, _ = _valid_pair()
|
context, _ = _valid_pair()
|
||||||
|
|||||||
Loading…
x
Reference in New Issue
Block a user