框架: 升格串行铁律——失败当场重试不跳窗+迟到覆写闸+缺键宽容
创始人拍板:窗串行生长,失败不得跳窗后补(断层+覆盖风险)。 - run 窗循环改 while:失败当场撤销半写入同窗重试一次,仍败停书(断点=本窗) - merge_card 加水位闸:win_no<卡水位的覆写只留审计不动卡体(防补跑时间倒流) - 水位更新改 GREATEST 防倒退 - 更新批缺'更新'键宽容为空批(prompt 明确空批输出+代码兜底,6窗实测死因)
This commit is contained in:
zizi
parent
4960dfdf7a
commit
0ac8f7ce04
@ -200,7 +200,7 @@ def update_prompt(contracts, title, a, b, text, cards_with_obs):
|
||||
【相关型字段合同】
|
||||
{render_entity_contracts(contracts, sorted({p.get("type") for _, p, _ in cards_with_obs} & set(ENTITY_TYPES)))}
|
||||
|
||||
【输出规则（只输出一个 JSON 对象）】
|
||||
【输出规则（只输出一个 JSON 对象；即使本批所有卡都无变化，也必须输出 {{"更新": []}}，不得省略"更新"键）】
|
||||
{{"更新": [{{"draft_id": 数字, "变更字段": {{"字段key": "新值或新增条目数组"}}, "别名新增": []}}]}}
|
||||
|
||||
━━━ 本窗材料 ━━━
|
||||
@ -272,6 +272,12 @@ def merge_card(conn, draft_id, win_no, changes, alias_new, valid_keys=None):
|
||||
payload = conn.execute("SELECT draft_payload FROM muse_knowledge_draft WHERE id=%s",
|
||||
(draft_id,)).fetchone()[0]
|
||||
fields = payload.setdefault("字段", {})
|
||||
# 卡水位（该卡最后一次被更新的窗号）：补跑迟到窗（如窗41在窗80后补跑）的覆写类字段
|
||||
# 若直接落卡，会把书末态倒写回中期态（时间倒流污染，2026-07-15 补跑实测坐实）。
|
||||
# 闸门：win_no < 水位 ⇒ 覆写只留审计不动卡体；追加类带窗号标签乱序无害，不拦。
|
||||
wm_row = conn.execute("SELECT watermark_window FROM example_upgrade_card_state WHERE draft_id=%s",
|
||||
(draft_id,)).fetchone()
|
||||
watermark = wm_row[0] if wm_row else 0
|
||||
changes = dict(changes or {})
|
||||
# 模型偶把「别名新增」混进变更字段（窗2实测）：摘出来并入别名流程，不落卡体字段
|
||||
alias_new = list(alias_new or []) + \
|
||||
@ -301,6 +307,15 @@ def merge_card(conn, draft_id, win_no, changes, alias_new, valid_keys=None):
|
||||
old.append(f"[窗{win_no}] {core}")
|
||||
seen.add(core)
|
||||
fields[k] = old
|
||||
elif win_no < watermark:
|
||||
# 迟到覆写弃用：本窗时序早于卡已生长到的窗位，覆写会让卡态倒流——
|
||||
# 只留审计（标记可查），卡体保持高窗态不动
|
||||
conn.execute(
|
||||
"""INSERT INTO example_upgrade_audit
|
||||
(draft_id, window_no, field_name, old_value, new_value, tenant_id)
|
||||
VALUES (%s,%s,%s,NULL,%s,%s)""",
|
||||
(draft_id, win_no, ("迟到覆写弃用:" + str(k))[:100],
|
||||
json.dumps(v, ensure_ascii=False)[:2000], TENANT))
|
||||
else:
|
||||
conn.execute(
|
||||
"""INSERT INTO example_upgrade_audit
|
||||
@ -329,7 +344,9 @@ def merge_card(conn, draft_id, win_no, changes, alias_new, valid_keys=None):
|
||||
conn.execute(
|
||||
"""INSERT INTO example_upgrade_card_state (draft_id, work_id, watermark_window, tenant_id)
|
||||
VALUES (%s,%s,%s,%s)
|
||||
ON CONFLICT (draft_id) DO UPDATE SET watermark_window=EXCLUDED.watermark_window,
|
||||
ON CONFLICT (draft_id) DO UPDATE
|
||||
SET watermark_window=GREATEST(example_upgrade_card_state.watermark_window,
|
||||
EXCLUDED.watermark_window),
|
||||
update_time=now()""",
|
||||
(draft_id, payload.get("_work_id") or 0, win_no, TENANT))
|
||||
|
||||
@ -478,8 +495,14 @@ def run(work_id, max_windows, max_calls, model, redo_window):
|
||||
click.echo(f"[撤销] failed 窗{r0[0]} 旧写入回滚后重跑")
|
||||
undo_window(conn, work_id, r0[0])
|
||||
done_n = 0
|
||||
fail_streak = 0 # 连续失败计数：单窗偶发失败（网络瞬断）continue，连续 2 窗停书（系统性问题）
|
||||
for win_no, a, b in wins:
|
||||
# 串行铁律（创始人 2026-07-15 拍板）：窗与窗是串行生长——后窗的预扫/判重/更新
|
||||
# 全依赖前窗长成的卡。窗失败绝不跳窗（跳窗=知识断层+事后补跑有覆盖风险），
|
||||
# 而是当场撤销半写入→整窗重试一次（挡偶发病）；仍败→停书，断点就在本窗，
|
||||
# 下次启动从本窗续跑（failed 先撤销机制），串行语义天然无损。
|
||||
wi = 0
|
||||
retried = False # 当前窗是否已当场重试过
|
||||
while wi < len(wins):
|
||||
win_no, a, b = wins[wi]
|
||||
if max_windows and done_n >= max_windows:
|
||||
break
|
||||
if max_calls and calls["n"] >= max_calls:
|
||||
@ -552,7 +575,16 @@ def run(work_id, max_windows, max_calls, model, redo_window):
|
||||
p = conn.execute("SELECT draft_payload FROM muse_knowledge_draft WHERE id=%s",
|
||||
(did,)).fetchone()[0]
|
||||
cards.append((did, p, obs_pts))
|
||||
upd, _ = call(update_prompt(contracts, title, a, b, text, cards), ("更新",))
|
||||
# 缺"更新"键宽容为空批：prompt 教"无变化的卡不输出"，某批恰好全无变化时
|
||||
# 模型会顺势连键一起省（批7实测 6 窗全死于此）。语义上缺键≈空批，按空批放行
|
||||
# 并留警告日志可审计；其余格式错误（乱码/解析失败）仍原样抛、窗照 fail。
|
||||
try:
|
||||
upd, _ = call(update_prompt(contracts, title, a, b, text, cards), ("更新",))
|
||||
except RuntimeError as e:
|
||||
if "缺少必需键" not in str(e):
|
||||
raise
|
||||
print(f"[宽容] 窗{win_no} 更新批缺键按空批放行: {str(e)[:80]}", file=sys.stderr)
|
||||
upd = {"更新": []}
|
||||
valid = {d for d, _, _ in cards}
|
||||
# 每卡按其型的合同 key 集校验（+一句话摘要），越合同 key 裁剪留审计
|
||||
did2keys = {d: {f["key"] for f in contracts.get(p.get("type"), {}).get("字段", [])}
|
||||
@ -665,7 +697,8 @@ def run(work_id, max_windows, max_calls, model, redo_window):
|
||||
click.echo(f" 窗{win_no}✓ ({a}-{b}章) 在场{len(onstage)} 新名字{new_n} "
|
||||
f"更新卡{len(to_update)} 调用累计{calls['n']}")
|
||||
done_n += 1
|
||||
fail_streak = 0
|
||||
wi += 1
|
||||
retried = False
|
||||
except SensitiveHardStop as e:
|
||||
with psycopg.connect(DSN) as conn:
|
||||
conn.execute(
|
||||
@ -682,14 +715,17 @@ def run(work_id, max_windows, max_calls, model, redo_window):
|
||||
WHERE tenant_id=%s AND work_id=%s AND window_no=%s""",
|
||||
(str(e)[:500], TENANT, work_id, win_no))
|
||||
conn.commit()
|
||||
fail_streak += 1
|
||||
# 偶发失败（网络瞬断/单窗输出畸形）跳去下一窗，failed 窗留给续跑（自动先撤销）；
|
||||
# 连续 2 窗失败=系统性问题，停书防连环烧额度
|
||||
if fail_streak >= 2:
|
||||
click.echo(f" ⛔ 连续 {fail_streak} 窗失败，本书升格停（末窗{win_no}）: {str(e)[:200]}")
|
||||
return
|
||||
click.echo(f" ✗ 窗{win_no} 失败（已记录，续跑重试；连续第{fail_streak}次）: {str(e)[:200]}")
|
||||
continue
|
||||
if not retried:
|
||||
# 串行铁律第一层：当场撤销半写入→同窗立即重试（挡网络瞬断/偶发格式病）
|
||||
with psycopg.connect(DSN) as conn:
|
||||
undo_window(conn, work_id, win_no)
|
||||
conn.commit()
|
||||
retried = True
|
||||
click.echo(f" ↻ 窗{win_no} 失败，撤销后当场重试（串行铁律不跳窗）: {str(e)[:150]}")
|
||||
continue
|
||||
# 第二层重试仍败：停书，断点=本窗；下次启动 failed 先撤销、从本窗续跑
|
||||
click.echo(f" ⛔ 窗{win_no} 当场重试仍失败，本书升格停（断点=本窗，续跑从此接续）: {str(e)[:200]}")
|
||||
return
|
||||
click.echo(f"《{title}》本次完成 {done_n} 窗，LLM 调用 {calls['n']} 次")
|
||||
|
||||
|
||||
|
||||
Loading…
x
Reference in New Issue
Block a user