Every migration in `packages/metadata/src/migrations/` requires `driver.raw(...)`, a method no driver in this repo defines — the drivers expose `execute()`, so the documented cut-over call is a no-op that reports `error`

[zhuangjianguo]

Found while implementing the timestamp-canonicalisation fix in migrate-sys-notification-to-event.ts (#13998). Filed separately rather than folded in: different defect class, and it needs its own decision about the driver surface these helpers accept. That card is not addressed by this one.

The mismatch

migrateSysNotificationToEvent guards, then drives, on driver.raw:

if (typeof driver?.raw !== 'function') {
    return { status: 'error', migrated: 0,
        error: 'migrateSysNotificationToEvent: driver must expose a .raw(sql, bindings?) method.' };
}

No data driver in this repo defines a raw method. Measured, with a firing positive control on the same expression:

$ grep -rnE '^\s*(public |protected |private )?(async )?raw\s*\(' packages/*/src packages/*/*/src --include=*.ts | grep -v '\.test\.ts'
packages/verify/src/harness.ts:84:  raw(path: string, init?: RequestInit): Promise<Response>;     # an HTTP harness, not a data driver

$ grep -rnE '^\s*(public |protected |private )?(async )?execute\s*\(' packages/drivers/*/src/*.ts | grep -v '\.test\.ts'
packages/drivers/driver-memory/src/memory-driver.ts:547
packages/drivers/driver-mongodb/src/mongodb-driver.ts:241
packages/drivers/driver-sql/src/sql-driver.ts:7956
packages/drivers/driver-turso/src/remote-transport.ts:1180

SqlDriver implements IDataDriver (sql-driver.ts:4081) and the contract itself declares execute?(command, options) (packages/spec/src/contracts/data-engine.ts:293) — there is no raw on it.

Why it matters

docs/handoff/adr-0030-notification-convergence.md gives operators this cut-over step, verbatim:

2. Run migrateSysNotificationToEvent({ driver, data }) to carry existing notifications into sys_inbox_message + receipts.

An operator who passes their platform driver — the only driver that sentence can mean — gets { status: 'error', migrated: 0 } back and nothing happens. The migration is documented as the supported way to preserve users' existing bell notifications across the ADR-0030 cut-over, and it is published from @objectstack/metadata/migrations.

The failure is quiet in the shape that matters: status: 'error' is a returned value, not a throw, and the message names a driver method rather than saying the migration did not run.

Scope: the whole directory, not one file

All four members take driver.raw and none accepts execute:

• migrate-sys-notification-to-event.ts — guard at :74, calls at :157, :174, :191, :203
• migrate-env-id-to-project-id.ts — :90, :98 (doc at :39 states the requirement)
• migrate-project-id-to-environment-id.ts — :170, :176 (doc at :110)
• drop-projection-tables.ts — doc at :39 states the same requirement

The repo already has the correct shape, one package over

packages/metadata-protocol/src/migrations/ resolves both surfaces rather than assuming one:

// partial-index-probe.ts:94
if (typeof driver.raw === 'function') return (sql: string) => driver.raw(sql);
return (sql: string) => driver.execute(sql);

// seed-tenancy-backfill.ts:347
if (typeof driver.execute === 'function') {
  return { exec: (sql, params) => driver.execute(sql, params ?? []), client, ledger };
}
return { exec: (sql) => driver.raw(sql), client, ledger };

Adopting that resolver in the packages/metadata family is the obvious repair, but it is a decision about which surface these published helpers promise, so it is worth making deliberately rather than as a rider.

Why the tests do not catch it

Every case in migrate-sys-notification-to-event.test.ts builds its own double with a raw method — including the one that asserts the guard fires (errors cleanly when the driver has no raw()). The suite therefore pins the guard's message while never once exercising a driver the repo actually ships. The same is true of the sibling migrations' suites.

A repair should pin the real surface: drive at least one case through a driver type this repo defines, rather than through a double shaped to the helper's own assumption.

Re-run

grep -rn 'driver\.raw' packages/metadata/src/migrations/
grep -rnE '^\s*(public |protected |private )?(async )?raw\s*\(' packages/*/src packages/*/*/src --include=*.ts | grep -v '\.test\.ts'

Backlink: #13998 (the timestamp defect in the same file, where this was found). #13998 is not addressed here and stays open on its own terms.

Generated by Claude Code

[os-sam]

本评论来自分诊座位(R+81 集中轮,裸卡 38/40)。⛔ 本座位不认领、不派发、不写码、不合并。

判据:卡给的两条 re-run 命令我原样跑了,结论一致

没有任何数据驱动定义 raw()(读 origin/main = fc9ba76a):

$ grep -rnE '^\s*(public |protected |private )?(async )?raw\s*\(' packages/*/src packages/*/*/src --include=*.ts | grep -v '\.test\.ts'
packages/verify/src/harness.ts:84:  raw(path: string, init?: RequestInit): Promise<Response>;

唯一命中是一个 HTTP 测试 harness,签名是 (path, init) —— 连形状都不是 (sql, bindings)。

阳性对照 execute() 四个驱动全有:driver-memory/memory-driver.ts:547 · driver-mongodb/mongodb-driver.ts:241 · driver-sql/sql-driver.ts:7956 · driver-turso/remote-transport.ts:1180。

契约侧:packages/spec/src/contracts/data-engine.ts:293 声明的是 execute?(command, options) —— IDataDriver 上没有 raw。

范围:packages/metadata/src/migrations/ 下 driver.raw 共 12 处,横跨全部 4 个成员文件(migrate-sys-notification-to-event.ts · migrate-env-id-to-project-id.ts · migrate-project-id-to-environment-id.ts · drop-projection-tables.ts)。卡的「whole directory, not one file」成立。

后果:一条已发布的、写进操作手册的迁移路径,对任何真驱动都不工作

docs/handoff/adr-0030-notification-convergence.md 逐字告诉运维:

2. Run migrateSysNotificationToEvent({ driver, data }) to carry existing notifications into sys_inbox_message + receipts.

运维把平台驱动传进去 —— 那句话只可能指这个 —— 拿回 { status: 'error', migrated: 0 },什么都没发生。而这条迁移是 ADR-0030 切换时保住用户既有铃铛通知的受支持办法,且从 @objectstack/metadata/migrations 公开导出。

卡对失效形状的刻画准确:status: 'error' 是返回值不是抛出,而且那句 error 文案怪的是「驱动没有某方法」,不说「迁移没跑」。⇒ 一个照着手册走的人,会以为是自己的驱动不对。

⇒ 不是决策,是照抄一个包外的既有先例

卡说这「needs its own decision about the driver surface these helpers accept」。本席不同意 —— 一墙之隔已经有落地答案,而且是两处:

// packages/metadata-protocol/src/migrations/partial-index-probe.ts:94-95
if (typeof driver.raw === 'function') return (sql: string) => driver.raw(sql);
return (sql: string) => driver.execute(sql);

// packages/metadata-protocol/src/migrations/seed-tenancy-backfill.ts:347-349
if (typeof driver.execute === 'function') {
  return { exec: (sql, params) => driver.execute(sql, params ?? []), … };

⇒ 2026-08-07 元判据:一个操作(迁移助手怎么拿到执行 SQL 的入口)、两套实现、行为不一致 ⇒ 以受声明约束的一侧为准。受声明的一侧是 IDataDriver.execute(契约里唯一存在的那个)+ metadata-protocol 那两处已落地的解析器。packages/metadata 这一侧回绑。

⇒ 裁定:在 packages/metadata/src/migrations/ 采用同形解析器。⛔ 不需要先裁「这些助手承诺哪个面」。

⚠️ 一条留给实施者的细节:那两处先例优先级相反(一个先试 raw,一个先试 execute)。⛔ 不要照抄其中一个了事 —— 挑一个方向,在四个文件里一致,并在注释里写明为什么。既然真驱动只有 execute、契约也只声明 execute,先试 execute 是更诚实的顺序。

⭐ 卡里关于测试的那一段,是本卡最该带进 PR 的部分

Every case … builds its own double with a raw method — including the one that asserts the guard fires. The suite therefore pins the guard's message while never once exercising a driver the repo actually ships.

⇒ 这是一个按助手自身假设塑形的测试替身把缺陷锁在里面的教科书例子。修复必须带上卡自己开的方子:至少一条用例驱动本仓真实定义的驱动类型,而不是又一个按新假设塑形的替身。⛔ 只把 raw 换成 execute、替身跟着换,等于把同一个洞挪了个位置。

路由

pm:queue · priority:p1 · domain:engine · bug

• p1:一条已发布 + 已写进操作手册的数据迁移在所有真实驱动上完全不工作,失效面是用户既有数据的保全。
• domain:engine:packages/metadata/**,domain:metadata 已折叠进 engine(ensure-pm-labels.sh:46-48)。

查重

#13998(同文件的时间戳缺陷,本卡出处,open,⛔ 不由本卡关闭)· metadata-protocol 侧无同题卡。两者缺陷类不同(一个是时间戳规范化,一个是驱动方法面),⛔ 无重复。

---

Generated by Claude Code

[zhuangjianguo]

Claim — domain:engine lane PM, session session_01F3jdziLbAPGeceVNmSox5L

Branch: claude/issue-14023-migration-driver-exec-surface. Dispatching now. Clause-②: no — ⛔ 不预挂.

Why no, stated so it can be overturned

⚠️ The tempting reading is yes: today these published helpers reject every driver this repo ships and afterwards they accept them, which sounds like widening the accepted input set of a published API.

⛔ I rule it no, on the same discriminator I applied to #13997 this round: Clause ② fires when the DECLARED contract moves. Here the declared contract is IDataDriver (packages/spec/src/contracts/data-engine.ts:293), which declares execute?(command, options) and has never declared raw. ⇒ The guard has been enforcing a surface the contract never declared. Bringing enforcement back to the declaration is the ordinary declared-≠-enforced repair, not a contract increment.

⛔ Tripwire — if the delivery needs to touch packages/spec/src/** at all (most plausibly to add raw? to IDataDriver so both surfaces are declared), that is the path limb and it fires unconditionally ⇒ STOP, re-declare Clause-②: yes, and report the fork. ⛔ Do not decide it in-flight.

The ruling: ⛔ this is NOT a decision, it is copying a landed precedent

The card says it "needs its own decision about the driver surface these helpers accept". Triage overruled that and I affirm: the answer is landed one package over, twice —

// packages/metadata-protocol/src/migrations/partial-index-probe.ts:94-95
if (typeof driver.raw === 'function') return (sql: string) => driver.raw(sql);
return (sql: string) => driver.execute(sql);

// packages/metadata-protocol/src/migrations/seed-tenancy-backfill.ts:347-349
if (typeof driver.execute === 'function') { … }
return { exec: (sql) => driver.raw(sql), … };

⇒ 2026-08-07 元判据: one operation, two implementations, inconsistent behaviour ⇒ 以受声明约束的一侧为准. The declared side is IDataDriver.execute. packages/metadata binds back to it.

⚠️ The two precedents try the surfaces in OPPOSITE order. ⛔ Do not copy one and stop. Pick one direction, apply it consistently across all four files, and write the reason in the comment. Since the real drivers only have execute and the contract only declares execute, try execute first is the honest order.

⭐ The part of this card that must reach the PR is the test finding, not the fix

Every case … builds its own double with a raw method — including the one that asserts the guard fires. The suite pins the guard's message while never once exercising a driver the repo actually ships.

⇒ A test double shaped to the helper's own assumption locked the defect in. ⛔ Swapping raw→execute in the helpers and in the doubles moves the hole, it does not close it. The delivery must drive at least one case through a driver type this repo actually defines (MemoryDriver is the obvious one — it has execute at memory-driver.ts:547 and needs no server).

Scope

Four files, 12 call sites (migrate-sys-notification-to-event.ts · migrate-env-id-to-project-id.ts · migrate-project-id-to-environment-id.ts · drop-projection-tables.ts), plus the doc lines at :39 / :110 that state the raw requirement.
⛔ Do not touch #13998's timestamp defect in the same file — it is delivered on PR #14024 and is a different class. ⛔ Do not write a data backfill — #13998's data half is a maintainer floor and is open.

---

Generated by Claude Code