SQLite Web: gitea.db

id

Primary key. INTEGER NOT NULL

package_id

INTEGER NOT NULL

creator_id

INTEGER NOT NULL (default 0

version

TEXT NOT NULL

lower_version

TEXT NOT NULL

created_unix

INTEGER NOT NULL

is_internal

INTEGER NOT NULL (default 0

metadata_json

{"scope":"@juhai","name":"client-fact","description":"可信事实受理客户端、事务 Outbox 与幂等 Inbox；Kafka 运输不影响领域契约","license":"UNLICENSED","dependencies":{"@prisma/client":"6.19.3"},"development_dependencies":{"@types/node":"22.19.20","typescript":"5.9.3","vitest":"2.1.9"},"peer_dependencies":{"@juhai/kernel":"0.22.0"},"readme":"# @juhai/client-fact\n\n事实干线（M5）的消费者 SDK：Outbox 投递 + Inbox 幂等/版本/Gap/DLQ 判定 + 可信回查。\n上层只允许 exact pin（`check:pins`），判定语义只在本包，**消费者不得复刻第二份**。\n\n本文记录的是**真实上层接入时踩过的坑**（数字员工基座 `os-employment-projection`，回灌台账 U-29 / U-31 / U-34）。\n平台自己的主线 e2e 用合成宿主表，这些坑在平台侧照不出来。\n\n---\n\n## 1. `baseUrl` 必须包含 `/api` 前缀（U-34）\n\n平台 NestJS 应用 `setGlobalPrefix(\"api\")`，而本包用**相对 URL** 解析端点：\n\n```ts\nnew URL(\"./v1/facts/intents\", baseUrl)   // createFactPublisher 内部\n```\n\n因此：\n\n| baseUrl | 实际请求 | 结果 |\n| --- | --- | --- |\n| `https://host/api` | `https://host/api/v1/facts/intents` | ✅ 正确 |\n| `https://host` | `https://host/v1/facts/intents` | ❌ 404 |\n| `https://host/api/` | `https://host/api/v1/facts/intents` | ✅（结尾斜杠会被补齐） |\n\n**为什么这条必须写在契约里**：少写 `/api` 不会显式报错，而是让**所有**回查返回 404。消费者若把 404\n当成\"事实不存在\"交给 `FactInbox`，SDK 会判 `UNTRUSTED_FACT` → 计入毒消息 → 三次进死信。\n一个配置漏字母会把整条摄入变成全量死信，而日志读起来像是\"平台在篡改事实\"。\n\n**消费者必须区分三种情况，不要混进毒消息计数**：\n\n| 情况 | 应有处置 |\n| --- | --- |\n| 回查端点返回 404 | 该 fact 确实不存在 → 交 SDK 判 `UNTRUSTED_FACT`（正当的毒消息） |\n| 回查不可达 / 5xx / 超时 | **抛错**，保留 broker offset 等待恢复；不要返回 null |\n| 未配置回查地址 | **失败关闭**：每次调用都拒绝，绝不\"跳过校验\" |\n\n平台 e2e 里 `baseUrl = http://127.0.0.1:${port}/api` 曾是这条约定唯一的载体（一行测试赋值）。\n\n## 2. 迁移模板只给表，授权与策略绑定由消费者补齐（U-31）\n\n`migrations/001_outbox_inbox.sql` 建五张表并对每张表 `ENABLE + FORCE ROW LEVEL SECURITY`，\n策略为：\n\n```sql\nCREATE POLICY tenant_isolation ON %I\n  USING (tenant_id = current_setting('app.current_tenant_id', true))\n  WITH CHECK (tenant_id = current_setting('app.current_tenant_id', true));\n```\n\n**模板的两个边界，照抄前必须知道**：\n\n1. **策略没有 `TO \u003crole\u003e` 子句** ⇒ 对 PUBLIC 生效。消费者若有 system / 后台角色需要跨租户扫描\n   （补 gap、重放、运维对账），它同样会被这条���略挡住——除非该角色 `BYPASSRLS`，\n   而 `FORCE ROW LEVEL SECURITY` 下表属主也不豁免。\n2. **模板不含任何 `GRANT`**。消费者库若用独立的低权限运行角色（而不是表属主连接），\n   照抄模板后该角色对这五张表**没有任何权限**，摄入会直接失败。\n\n**契约**：表 DDL 逐字照用（列名与主键是 SDK 的 SQL 依赖，改了就对不上）；\n`GRANT` 与策略的角色绑定**由消费者按自身角色模型补齐**，作为自己的增量迁移。\n参考实现见数字员工基座的 `20260928010000_platform_client_fact_inbox`（tenant / system 两角色 + 逐表显式 GRANT）。\n\n## 3. 处理结果与 rc.3 的行为变更（U-29）\n\n`FactInbox.handle()` 的返回值：\n\n| 返回 | 含义 |\n| --- | --- |\n| `APPLIED` | 首次应用，投影已写 |\n| `DUPLICATE` | 同 `fact_id` 重投，单次效果（digest 不一致则抛 `PROCESSED_FACT_CONFLICT`） |\n| `STALE` | 版本 ≤ 水位，不写投影 |\n| `GAP` | `previous_version ≠ 水位`，落 gap 缓冲等补齐 |\n\n\u003e **`1.0.0-rc.3` 起的行为变更**：`STALE` 之前**不可达**——版本落后时本包抛\n\u003e `FactRejected(\"UNPROCESSED_VERSION_BEHIND_WATERMARK\")`，而 `handle` 把任何 `FactRejected`\n\u003e 计入 attempts，于是至少一次投递下的**正常迟到重投**三次即被打进死信，与\"内容有毒\"共用一套判据。\n\u003e rc.3 改为返回 `\"STALE\"`，并顺手删掉该 fact 可能留在 gap 缓冲里的副本（落后事实永远等不到\n\u003e `previous_version = 水位`，留着只会把缓冲占满到 `GAP_BUFFER_FULL`，drain 循环也会空转）。\n\u003e\n\u003e 消费 rc.2 及更早版本的上层需要在边界做\"错误码 → 结果词\"的翻译；**升到 rc.3 后应删除该翻译**，\n\u003e 判定回到 SDK 单源。\n\n## 4. 写链边界\n\n`appendFactIntent` 必须在**属主自己的业务事务内**调用——它与业务写同 tx，事务回滚则事实意图一并回滚。\n`FactInbox.handle` 同理在消费者事务内完成 `processed_fact` + 投影 + 水位的原子写入。\n本包不提供事务，也不管理连接：宿主传入符合 `FactTransaction` 结构的 tx 即可（不要求宿主生成 Prisma schema）。\n\n## 5. 可信回查读取器 `createFactReader`（候选）\n\n\u003e 本节为 `3074b75` 引入的读取面文档，原文保留（英文），与上面三节的接入契约互补：\n\u003e 上面讲「接进来会踩什么」，本节讲「读取器本身怎么用、边界在哪」。\n\n```ts\nconst reader = createFactReader({\n  baseUrl: \"https://platform.example/api\",\n  tenantId,\n  token: getConsumerBearer,\n});\nconst inbox = new FactInbox({\n  tenantId, consumerId, tenant: consumerTenantTransaction,\n  trustedLookup: reader.lookup,\n  apply: applyOwnerFact,\n});\n// On shutdown: stop consumer delivery, then reader.close().\n```\n\nThe server binds the verified JWT tenant/application/subject to one enrolled consumer. It checks the persisted consumer's activation and source-domain/fact-type scope on every lookup. `replayAllowed` is not required for point lookups; replay remains a separate operation. The candidate server is enabled only with `PLATFORM_FACT_READ_ENABLED=1` in development/test with JWT, `DATABASE_URL_FACT`, and an exact `FACT_READERS_JSON` list of `{tenantId,appId,subject,consumerId}`. It can run independently of intake and relay. Production/default assembly remains unavailable pending CHG-004 and live service identity integration.\n\nThe reader uses HTTPS (HTTP only on literal loopback/localhost), denies redirects and URL credentials, and bounds token/fetch/body processing to 5 seconds and 128 KiB by default. It verifies tenant/id/ordering/receipt and the canonical submission digest. It does not cache authorization or facts. A typed M5 `FACT_NOT_FOUND` returns null; all other transport/authentication/integrity errors throw `FactReadUnavailable`, which is deliberately distinct from `FactRejected`. Dependency failures must retain the Kafka offset and must not increment poison counters. `close()` aborts active lookups and rejects future ones; callbacks that ignore cancellation may continue internally but cannot produce an accepted lookup.\n\nA deactivated or deleted consumer registration answers 403, which the reader turns into a retryable `FactReadUnavailable`: the Kafka offset is kept and nothing is counted as poison. A registration that is still active but no longer covers the fact's source domain or type answers the same 404 as an unknown id, which is a null lookup and therefore `UNTRUSTED_FACT` — a dead letter. Narrowing `sourceDomain`/`factTypes` on a live consumer will dead-letter facts that are already in flight, while deactivating the same consumer only pauses it. Deactivate first, change scope, then reactivate; treat a scope edit as a migration, not as a switch.\n\nThe server refuses a bearer carried in the query string (`FACT_READ_URL_CREDENTIAL_REFUSED`, 401) even when a valid `Authorization` header is also present, because the token has already reached the proxy access log by then. The reader never produces such a URL; the refusal exists for anything else that calls the surface.\n\nAn authenticated M5 response plus the receipt digest is the trust boundary here, not a cryptographic signature proving the Owner. A compromised M5 authority or whole-row rewrite with a recomputed digest is outside this reader's detection. Kafka self-asserted hashes are never sufficient. Consumers still need the Owner's domain validation, bootstrap watermark reconciliation and appropriate action enforcement.\n","repository":{"type":"","url":""}}

TEXT

download_count

INTEGER NOT NULL (default 0

Update row 26 in package_version