Commit Graph

3 Commits

Author SHA1 Message Date
1445043649 8bdd308ea0 docs(multi-tenant): 据审计对齐 ADR-002/003/004/005 与 phase-0 计划
非 spike 驱动的对齐改动:
- ADR-002/004/005:头部加现状提示 + 链接审计报告(沙箱出网/资源缺位、
  token_version 已存在 MembershipCache 全新建、ObjectStorage/7 表/KMS
  全部从 0 起)
- ADR-003 LLM 计费:修正 TokenUsageMiddleware 当前只 log 不持久化的描述;
  补充 create_chat_model sync→async 改造的连带影响说明
- phase-0 计划:新增 §3.5 底座先行(Postgres 测试夹具 / ObjectStorage
  Protocol / KMS 抽象 3 件并行做),时间盒 2 → 3 周;ADR-006/007 摘要
  对齐;DoD 加底座骨架检查项

Co-Authored-By: Claude Opus 4.7 (1M context) <noreply@anthropic.com>
2026-05-09 21:03:32 +08:00
1445043649 27c4f14233 docs(multi-tenant): 加入 ADR 审计 + spike,并据其修订 ADR-001/006/007
新增两份评审产出物:
- adr-vs-code-audit:7 份 ADR 与现状代码的差异核对,标注每条假设是否成立
- adr-spike-langgraph-postgres:实测 langgraph-checkpoint-postgres==3.0.5
  注入能力,确认不存在 connection_factory 参数,且 psycopg_pool 自带的
  configure callback 不是 per-acquire hook

据 spike 与审计修订三份 ADR:
- ADR-001 数据隔离:LangGraph 表改为应用层强校验 + threads_meta unique
  约束兜底(不再挂 RLS、不 ALTER 表);hook 点从 AssistantsCompat 修正
  为 threads.py + thread_runs.py
- ADR-006 运行时与渠道:§2.1 完全重写为应用层强校验;MCP OAuth token
  从"无持久化进程内存"直接做加密 DB;channel store binding 改为新建
  channel_bindings 表
- ADR-007 路由与前端:删除 Better Auth 假设(前端实际无此依赖),改为
  扩展现有 auth/jwt.py TokenPayload 加 tid/role 字段

Co-Authored-By: Claude Opus 4.7 (1M context) <noreply@anthropic.com>
2026-05-09 21:03:16 +08:00
1445043649 dce5e9598b docs(multi-tenant): add ADRs and phase-0 plan for multi-tenant redesign
Add 7 ADRs and a phase-0 plan covering the multi-tenant redesign of
DeerFlow, plus an architecture-overview snapshot of the current state.

ADRs:
- 001 data isolation: row-level tenant_id + Postgres RLS, including
  LangGraph-owned checkpoint tables (subquery RLS or column upgrade path).
- 002 sandbox isolation: K8s namespace + gVisor + NetworkPolicy default-
  deny, threat model and pod spec defaults.
- 003 LLM key & billing: hybrid platform/BYO with pessimistic reservation
  to handle the "ghost token" overflow on the last call, plus a usage
  category split for memory/title/summarization charges.
- 004 RBAC: two-level (owner/admin/member), JWT-with-role + 30s LRU
  cache for reads, strict DB lookup for sensitive writes, token_version
  bump as the single revocation path.
- 005 storage topology: Postgres (structured) + S3-compatible object
  store (large objects) + emptyDir (ephemeral); explicit treatment of
  the extensions_config.json migration's downstream effects.
- 006 runtime & channel tenancy: per-tenant MCP cache, dual-track skills
  loader, sandbox provider routing by namespace, internal LLM call
  billing, IM channel-to-tenant binding model.
- 007 routing & frontend: path-slug URL form, JWT-only API auth,
  TenantProvider, hard-reload tenant switch, Better Auth integration.

These docs are decision records; no code changes are included.
2026-05-08 23:45:33 +08:00