feat: M4 文风 + M5 生成/多provider/Skill + Kimi Code 订阅接入 + 本地联调修复

M4(文风): style-auditor 双轨(提取指纹/漂移第四审)+ jobs 长任务框架(zombie reaper) + 回炉 refine + GET /style read-back。
M5(生成+扩展): worldbuilder/character-gen(入库 continuity 409 gate + partition_writes 白名单 + schema→JSONB 形变);
  网关多 provider 回退链/熔断/能力降级(Anthropic/Gemini 适配器);Skill registry + 表权限沙箱 + 规则;
  前端 角色生成器/世界观/Codex/规则页/技能库/⌘K 命令面板。
K1(Kimi Code 订阅接入): OAuth device-flow(kimi-code)+ 静态 Console key(kimi-code-key)两路径;
  coding 端点 KimiCLI 伪造头(实测 UA allow-list 门禁,缺则 403)+ JSON 模式结构化(thinking ⊥ tool_choice)。
本地联调修复: CORS 中间件;assemble 注入 premise+「写第N章」指令(修空 prompt 400);
  GET /outline·/draft read-back + 大纲/工作台/审稿页重载;写页 client/server 常量边界 + notFound 健壮化;
  字数 toLocaleString locale 水合;审稿页终稿从已存草稿 seed(修 accept 422)。
门禁: backend ruff/mypy(157)/alembic 无漂移/pytest 451 · frontend lint/tsc/vitest/build。

Co-Authored-By: Claude Opus 4.8 (1M context) <noreply@anthropic.com>
This commit is contained in:
Yaojia Wang
2026-06-20 10:39:58 +02:00
parent 5fb7bfb1de
commit 765dbdfbd4
161 changed files with 17330 additions and 208 deletions

View File

@@ -0,0 +1,225 @@
"""文风端点:学文风(异步长任务)+ 回炉 + 最新指纹读取
C3 扩 / ARCH §5.4 style-auditor 提取轨 / §7.2 / §7.4 jobs / UX §6.9 / §8.3)。
独立 router仿 outline.py挂 `app.include_router`。三端点:
- `POST /projects/{id}/style` → 202 `{job_id}`:学文风走 jobsM4-c。立即写一行 job
返 202提取经 BackgroundTask `run_job` 跑(**自建独立 session**,请求 session 已关闭)。
无凭据 → 流前 `LLM_UNAVAILABLE`503在调度 job 之前拦下,避免凭空写注定失败的 job
- `POST /projects/{id}/chapters/{no}/refine` → 200 `{original, refined}`同步回炉writer
网关纯文本重写选中段;不写库(不变量 #3作者采纳经既有 draft 自动保存合入);末尾
`commit()` 让 usage_ledger 落库(网关 ledger add-only同 draft/review 纪律)。
- `GET /projects/{id}/style` → 200 最新指纹(完整 16 维 + 证据 + 版本UX §6.9)。
不变量 #2经 `get_*_gateway`analyst/writer 档注入spec 只声明档位不传 model。
不变量 #3提取/回炉 agent 只读不写其它表;文风指纹落库经此作者发起的学文风路径。
"""
from __future__ import annotations
import uuid
from typing import Annotated, Any
from fastapi import APIRouter, BackgroundTasks, Depends, Request, Response
from sqlalchemy.ext.asyncio import AsyncSession
from ww_agents import refiner_spec, style_extract_spec
from ww_core.domain import JobRepo, ProjectRepo, StyleFingerprintWriteRepo
from ww_core.domain.style_repo import SqlStyleFingerprintWriteRepo
from ww_core.orchestrator import run_style_extraction
from ww_db import get_session
from ww_llm_gateway import Gateway
from ww_llm_gateway.types import Block, LlmRequest, Scope
from ww_shared import AppError, ErrorCode
from ww_api.logging_config import get_logger
from ww_api.schemas.style import (
RefineRequest,
RefineResponse,
StyleFingerprintResponse,
StyleLearnRequest,
StyleLearnResponse,
)
from ww_api.services.credentials import STUB_OWNER_ID, SqlCredentialStore
from ww_api.services.foreshadow_scan import SessionFactory
from ww_api.services.job_runner import run_job
from ww_api.services.project_deps import (
build_gateway_for_tier,
get_job_repo,
get_project_repo,
get_refine_gateway,
get_session_factory,
get_style_extract_gateway,
get_style_write_repo,
)
log = get_logger("ww.api.style")
router = APIRouter(prefix="/projects", tags=["style"])
ProjectRepoDep = Annotated[ProjectRepo, Depends(get_project_repo)]
JobRepoDep = Annotated[JobRepo, Depends(get_job_repo)]
StyleWriteRepoDep = Annotated[StyleFingerprintWriteRepo, Depends(get_style_write_repo)]
StyleExtractGatewayDep = Annotated[Gateway, Depends(get_style_extract_gateway)]
RefineGatewayDep = Annotated[Gateway, Depends(get_refine_gateway)]
SessionDep = Annotated[AsyncSession, Depends(get_session)]
SessionFactoryDep = Annotated[SessionFactory, Depends(get_session_factory)]
_JOB_KIND_STYLE_LEARN = "style_learn"
def _split_fingerprint(result: Any) -> tuple[dict[str, str], dict[str, list[str]]]:
"""把 `StyleFingerprintResult` 拆成 DB 两列:`{name:value}` + `{name:[evidence]}`。"""
dimensions = {dim.name: dim.value for dim in result.dimensions}
evidence = {dim.name: list(dim.evidence) for dim in result.dimensions}
return dimensions, evidence
def _make_style_learn_work(project_id: uuid.UUID, samples_text: str) -> Any:
"""构造学文风后台工作闭包:在 `run_job` 自建的独立 session 上跑提取 + 落库。
`work(session)` 自建 analyst 网关(从凭据)+ 写侧 repo用 `run_job` 传入的独立
session→ `run_style_extraction` → 拆指纹 → `append`(版本化)→ 返回 result 摘要
`{version, dims_count}`。提交归 `run_job`(业务写 + job done 同一事务一次 commit
"""
async def work(session: AsyncSession) -> dict[str, Any]:
store = SqlCredentialStore(session)
gateway = await build_gateway_for_tier(session, store, "analyst")
result = await run_style_extraction(
style_extract_spec,
samples_text=samples_text,
gateway=gateway,
user_id=STUB_OWNER_ID,
project_id=project_id,
)
dimensions, evidence = _split_fingerprint(result)
repo = SqlStyleFingerprintWriteRepo(session)
version = await repo.append(project_id, dimensions_json=dimensions, evidence_json=evidence)
return {"version": version, "dims_count": len(dimensions)}
return work
@router.post("/{project_id}/style", status_code=202)
async def learn_style(
project_id: uuid.UUID,
body: StyleLearnRequest,
request: Request,
response: Response,
background_tasks: BackgroundTasks,
project_repo: ProjectRepoDep,
job_repo: JobRepoDep,
gateway: StyleExtractGatewayDep, # 凭据探测(无凭据 → dep 解析阶段 503
session: SessionDep,
session_factory: SessionFactoryDep,
) -> StyleLearnResponse:
"""学文风:写一行 job 返 202提取经 BackgroundTask 异步跑。
项目不存在 → 404无凭据 → 503在调度 job 前拦下)。`mode` 不影响落库逻辑
(写侧始终 append 新版本version+1 自动),仅供前端区分首学/更新语义。
"""
request_id = getattr(request.state, "request_id", None)
project = await project_repo.get(STUB_OWNER_ID, project_id)
if project is None:
raise AppError(ErrorCode.NOT_FOUND, f"project not found: {project_id}")
# `gateway` 已在 dep 解析阶段验过凭据(无凭据 → 503提取本体在 BackgroundTask
# 里用独立 session 重新构网关跑,此处不复用它(请求 session 即将关闭)。
samples_text = "\n\n".join(body.samples)
job = await job_repo.create(project_id, _JOB_KIND_STYLE_LEARN)
await session.commit() # job 行需在 202 返回前持久化(供前端立即轮询)。
background_tasks.add_task(
run_job,
session_factory,
job.id,
_make_style_learn_work(project_id, samples_text),
request_id=request_id,
)
log.info(
"style_learn_scheduled",
project_id=str(project_id),
request_id=request_id,
job_id=str(job.id),
sample_count=len(body.samples),
mode=body.mode,
)
response.status_code = 202
return StyleLearnResponse(job_id=job.id)
@router.get("/{project_id}/style")
async def get_style(
project_id: uuid.UUID,
project_repo: ProjectRepoDep,
style_repo: StyleWriteRepoDep,
) -> StyleFingerprintResponse:
"""取最新文风指纹(完整 16 维 + 证据 + 版本)。项目不存在 → 404无指纹 → 404。"""
project = await project_repo.get(STUB_OWNER_ID, project_id)
if project is None:
raise AppError(ErrorCode.NOT_FOUND, f"project not found: {project_id}")
latest = await style_repo.latest(project_id)
if latest is None:
raise AppError(ErrorCode.NOT_FOUND, f"no style fingerprint for project: {project_id}")
return StyleFingerprintResponse(
dimensions=latest.dimensions,
evidence=latest.evidence,
version=latest.version,
)
def _build_refine_input(segment: str, instruction: str | None) -> str:
"""组回炉输入:待重写段 + 可选指令(保留上下文语气,只重写该段)。"""
parts = [f"【待重写段落】\n{segment}"]
if instruction:
parts.append(f"【改写指令】\n{instruction}")
return "\n\n".join(parts)
@router.post("/{project_id}/chapters/{chapter_no}/refine")
async def refine_segment(
project_id: uuid.UUID,
chapter_no: int,
body: RefineRequest,
request: Request,
project_repo: ProjectRepoDep,
gateway: RefineGatewayDep,
session: SessionDep,
) -> RefineResponse:
"""同步回炉writer 网关纯文本重写选中段,返回 {original, refined}。
项目不存在 → 404无凭据 → 503。不写库不变量 #3末尾 `commit()` 让网关
usage_ledger 落库add-only否则记账静默丢失同 draft/review 纪律)。
"""
request_id = getattr(request.state, "request_id", None)
project = await project_repo.get(STUB_OWNER_ID, project_id)
if project is None:
raise AppError(ErrorCode.NOT_FOUND, f"project not found: {project_id}")
req = LlmRequest(
tier=refiner_spec.tier, # 不变量 #2writer 档,不传 model
system=[Block(text=refiner_spec.system_prompt, cache=True)],
input=_build_refine_input(body.segment, body.instruction),
output_schema=None, # refiner 纯文本(无结构化 schema
scope=Scope(user_id=STUB_OWNER_ID, project_id=project_id),
)
resp = await gateway.run(req)
# 提交边界:网关 ledger add-only → 端点末尾 commit否则记账静默丢失见 gotcha
await session.commit()
log.info(
"refine_done",
project_id=str(project_id),
chapter_no=chapter_no,
request_id=request_id,
segment_len=len(body.segment),
refined_len=len(resp.text),
has_instruction=body.instruction is not None,
)
return RefineResponse(original=body.segment, refined=resp.text)