观察：`ValueDataSource.test.ts` 的 10 处 `$filter: [...] as any` 在 #3909 之后已成多余，且 `as any` 正在抹掉这些 fixture 本可提供的类型面证据

[yinlianghui]

观察类发现（finding，不进 pm:queue），发现于 #3909 / PR #5999 的实施过程。不是缺陷，运行时完全正确 —— 是一处随 #3909 落地而变得多余、且顺带遮蔽了类型面信号的测试写法。

机制

packages/core/src/adapters/__tests__/ValueDataSource.test.ts 有 10 处形如：

$filter: ['role', '=', 'admin'] as any,
$filter: ['and', ['role', 'in', ['admin', 'user']], ['age', '>=', 28]] as any,
const result = await ds.find('users', { $filter: [] as any });

这些 as any 之所以存在，是因为 QueryParams.$filter 当时声明为 Record< string, any >，只描述了 MongoDB 风格的记录形式 —— 作者要写 ObjectQL AST 数组，就得把类型推开。#3909（PR #5999）把该槽位声明成它实际接受的联合（记录形式 | 绑定自 @objectstack/spec/data 的 FilterArray）之后，这些数组字面量已经可以不带任何 cast 直接赋值，as any 成了纯粹的多余物。

有害在哪（很轻，但方向明确）

不是运行时问题 —— 这些用例断言的是 ValueDataSource 的过滤行为，as any 不影响它们跑什么、断言什么，今天它们全绿且断言正确。

值得记一笔的是第二层：as any 把这些 fixture 本可顺带提供的类型面证据抹掉了。这个文件恰好枚举了 AST 数组的九种形态（比较三元组、in / not in、and / or 组合、!=、contains、空数组），是仓内对该形状覆盖最全的一处。若去掉 cast，它们在跑运行时断言的同时，也自动成为「这些形状对 $filter 合法」的编译期证据 —— 一旦有人把该槽位收窄回记录形式，这个文件会转红。带着 as any，同样的收窄对它完全不可见。

即 #3909 已经付过一次的那类代价的小号版本：一个看起来在覆盖某形状、实际对该形状的声明变化免疫的夹具。

落地建议

删掉这 10 处 as any，跑 pnpm --filter @object-ui/core type-check 确认无需替代 cast。不要换成别的断言或加宽任何东西 —— 这些值本来就合法，直接裸写即可。

⚠️ 前置依赖：需要 #3909 的类型改动先落地（PR #5999），否则删掉 cast 会转红。

Blocked-by: #3909

参考位置

• packages/core/src/adapters/__tests__/ValueDataSource.test.ts:168,177,185,194,202,210,219,227,234,241
• packages/types/src/data.ts（$filter 声明，观察:QueryParams.$filter 声明为 Record< string, any >,但仓内下沉口早已往里送 ObjectQL AST 数组 #3909 改动处）

关联：#3909 / PR #5999（发现来源）

[claude]

Concentrated triage batch: finding → pm:queue + domain:ui, Task, S — post-#3909, the 10 $filter: […] as any casts in ValueDataSource.test.ts are redundant and actively erase the type-surface evidence these fixtures would otherwise provide (the file is the repo's fullest enumeration of the AST array forms). Delete the casts; if any assignment fails to compile bare, that is a real narrowing gap in #3909's union — report it rather than re-adding the cast.

---

Generated by Claude Code

[os-sales]

Claim: domain:ui execution seat, session 8ca04858-ea8e-5b85-9182-de59aa49e00c, branch claude/issue-6001-drop-redundant-filter-casts.

Blocker is dead — verified, not assumed

Body carries Blocked-by: #3909. Checked: #3909 is closed/completed, closed by merged PR #5999. Read packages/types/src/data.ts on main directly — the slot now reads:

$filter?: Record<string, any> | FilterArray;

FilterArray is bound from @objectstack/spec/data. The union is on the branch point. Nothing is waiting.

Fence

Exactly one file: packages/core/src/adapters/__tests__/ValueDataSource.test.ts

Do not edit packages/types/, the spec package, or any other file. If the work appears to require touching anything outside that one path, stop and report rather than crossing — that is a signal my order is wrong, not that the fence is.

The change

Delete the 10 redundant as any casts at lines 168, 177, 185, 194, 202, 210, 219, 227, 234, 241. The card's own instruction, quoted verbatim and not translated:

删掉这 10 处 as any，跑 pnpm --filter @object-ui/core type-check 确认无需替代 cast。不要换成别的断言或加宽任何东西 —— 这些值本来就合法，直接裸写即可。

⛔ Do not substitute a different assertion, do not widen any type, do not add a helper, do not touch the runtime assertions. The values are already legal; write them bare.

If any one cast will not come off bare, the standing triage instruction on this card governs — stop and report it as a real narrowing gap in the union, do not re-add the cast and do not work around it:

if any assignment fails to compile bare, that is a real narrowing gap in #3909's union — report it rather than re-adding the cast.

Line 241 ($filter: [] as any) is the likeliest candidate, being the degenerate empty-array form. Report it; don't patch around it.

A premise in the card that I believe is false — measure it, don't take my word

The card justifies itself with a second claim: that once the casts are gone the fixtures become compile-time evidence, because "一旦有人把该槽位收窄回记录形式，这个文件会转红".

I think that is wrong, and the source is #3909's own landed text. The doc comment PR #5999 wrote into data.ts states, as a measurement:

Note the declaration does not narrow anything: Record<string, any> already structurally accepts arrays (they satisfy its string index), so the union documents the shapes that were always accepted rather than admitting new ones.

#3909's body says the same and cites tsc --strict exit 0. If that holds, then narrowing the slot back to the bare record form would not turn this file red, and the casts were removable before #3909 ever landed.

So: do not assert the "turns red" claim in the PR body. Settle it by measurement instead, as a local uncommitted probe:

1. Commit the cast deletion first.
2. Predict the direction out loud before looking.
3. Locally narrow the slot in packages/types/src/data.ts back to Record<string, any> — prove the mutation landed on disk (anchored grep plus a git hash-object blob-hash difference, not an exit code).
4. Run the type-check and observe whether this test file goes red.
5. Restore, and prove the restore by observation — git diff HEAD empty and the blob hash back to its original value. Never by exit code.

Report the observed direction either way. If it stays green, the deletion is still correct (the casts are genuinely redundant) but its value is "removes a misleading no-op cast", not "adds a type-surface witness" — and the PR body must say the true one. A zero here needs a positive control: if nothing goes red, show something that does go red under the same probe shape, or the zero is NOT MEASURED rather than green.

The probe touches a file outside the fence. That is authorised only as a local, uncommitted, proven-restored measurement — it must never appear in a commit or the diff.

Verification before pushing

• pnpm --filter @object-ui/core type-check — green, with no replacement casts.
• Run the test file itself; the runtime assertions must be untouched and still passing.
• Anchored grep: 10 occurrences of the cast before, 0 after, in that one file.
• Changeset: test-only change. Follow the repo's changeset policy and let the Changeset Declaration gate rule whether one is owed — do not pre-empt it in either direction.

Draft PR, Fixes #6001.

---

Generated by Claude Code

[os-sales]

⚠️ Correction to the dispatch order above — three fragments were damaged in transit. Read these in place of the corresponding lines in that comment. The order is otherwise unchanged and still governs.

The generic type argument was stripped from Record< string, any > in three places, leaving a bare Record that reads as complete but is not. Restating each with spaces inside the brackets so it survives:

1. The slot on main (the fenced block under "Blocker is dead") should read:

$filter?: Record< string, any > | FilterArray;

not $filter?: Record | FilterArray;, which is not a type.

2. The quotation from the doc comment PR #5999 landed in packages/types/src/data.ts should read:

Note the declaration does not narrow anything: Record< string, any > already structurally accepts arrays (they satisfy its string index), so the union documents the shapes that were always accepted rather than admitting new ones.

3. Probe step 3 should read: locally narrow the slot in packages/types/src/data.ts back to Record< string, any > — i.e. drop the | FilterArray arm, leaving the record form alone. As written it said "back to Record", which is not executable.

Nothing else changes: same one-file fence, same 10 line numbers, same ⛔ on substituting assertions or widening, same stop-and-report rule if a cast will not come off bare, same requirement that the probe stay local, uncommitted and proven-restored.

Note for the lane: this is the sanitizer failure the protocol warns about — a fenced code block and backticks are both not protection, and what survived was a plausible-looking shorter string rather than an obvious hole. The spelling with spaces inside the angle brackets is what the surrounding issue bodies use, and it is what survives.

---

Generated by Claude Code

[os-sales]

os-dev-report

{
 "issue": 6001,
 "status": "done",
 "branch": "claude/issue-6001-drop-redundant-filter-casts",
 "pr": "https://github.com/objectstack-ai/objectui/pull/6630",
 "premise_still_valid": true,
 "summary": "The actionable premise held completely: all 10 `as any` casts on the AST-array `$filter` fixtures in packages/core/src/adapters/__tests__/ValueDataSource.test.ts (lines 168,177,185,194,202,210,219,227,234,241) came off BARE, including line 241's degenerate empty array, which the order flagged as the likeliest to resist. No replacement assertion, no widening, no helper, no runtime assertion touched. TWO SECONDARY PREMISES OF THE CARD ARE FALSIFIED BY MEASUREMENT, and the PR body states the true thing instead of the card's claim. (1) The 'this file turns red if the slot is narrowed' claim is FALSE - measured, see `tests`. (2) Therefore `Blocked-by: #3909` was never real: since the record arm always accepted arrays structurally, these casts were removable BEFORE #3909/PR #5999 ever landed. The merge value is deleting a misleading no-op cast, not adding a type-surface witness. FENCE NOTE, needs your eye: the order fenced the commit to exactly one file, but also ordered me to follow the repo's changeset policy and let the gate decide. The gate decided a declaration IS owed (check-changeset-presence.mjs guards packages/core/src/** with no carve-out for tests and answers a test-only change with its empty-frontmatter exemption). I resolved the tension by adding .changeset/6001-drop-redundant-filter-casts.md with EMPTY frontmatter, so the branch carries 2 files, not 1. I did NOT apply objectui's phantom `skip-changeset` label. Nothing else was crossed: packages/types/ appears in ZERO commits on this branch.",
 "tests": "All results are the tool's own printed verdict with the exit status captured before any pipe; union re-run on the FINAL pushed head 234415d6.\n1) TYPE-CHECK at 234415d6: `pnpm --filter @object-ui/core type-check` (= `tsc --noEmit && tsc -p tsconfig.test.json`) -> EXIT 0, no replacement cast anywhere.\n2) TESTS: `pnpm exec vitest run --maxWorkers=2 packages/core/src/adapters/__tests__/ValueDataSource.test.ts` from the REPO ROOT -> 'Test Files 1 passed (1) / Tests 56 passed (56)'. NOTE: the package-scoped form `pnpm --filter @object-ui/core exec vitest` was REFUSED by this repo's own vitest guard as a known false-green (it silently runs apps/console's 22 files and reports them as passed); the root-relative form is the sanctioned one.\n3) ANCHORED GREP in the one file: `$filter` lines carrying `as any` = 10 BEFORE, 0 AFTER. The 11 other `as any` in the file are unrelated record-shape assertions, untouched. File blob hash d03fa188 -> 4430019e; diff is exactly 10 insertions / 10 deletions.\n4) APPARATUS PROOF (before trusting any zero): `tsc -p tsconfig.test.json --listFiles` shows ValueDataSource.test.ts IS in the checked program (1 hit) and that the program reads packages/types/dist/data.d.ts - NOT the source. packages/core/tsconfig.test.json sets empty `paths` on purpose, so a probe that edits only packages/types/src would have changed nothing the checker sees and returned a meaningless green.\n5) PROBE (local, uncommitted, proven-restored; trap on EXIT INT TERM; absolute paths). MUTATION = drop the `| FilterArray` arm, leaving the record form alone. CONFIRMED ON DISK, not by exit code: anchored grep union-form 1->0 and narrowed-form 0->1 in the source; source blob hash 21899561 -> f9f82421. REBUILT `pnpm --filter @object-ui/types build` and CONFIRMED THE MUTATION REACHED dist: dist/data.d.ts blob hash eeaf49dc -> 3536a047 with the same 1->0 / 0->1 flip. RESULT: `pnpm --filter @object-ui/core type-check` EXIT 0 - zero errors in the program, zero mentioning ValueDataSource.test.ts. OBSERVED DIRECTION: NOT RED. Prediction was stated before looking and matched. RESTORE PROVEN BY OBSERVATION: source hash back to 21899561, dist hash back to eeaf49dc, `git diff HEAD` empty, `git status --short` empty.\n6) POSITIVE CONTROL for that zero, identical probe shape (same slot, same file, rebuild, same program): narrowed the slot to `Record< string, any > & { length?: undefined }`, which genuinely excludes arrays while leaving the record-form fixtures alone. `tsc -p tsconfig.test.json` EXIT 2 with 10 errors at EXACTLY lines 168,177,185,194,202,210,219,227,234,241, e.g. \"ValueDataSource.test.ts(234,45): error TS2322: Type 'never[]' is not assignable\". Run through the test project DIRECTLY so no `&&` short-circuit could mask it. So the checker does see these ten expressions: the zero in (5) is a MEASUREMENT, not blindness. Restore proven the same way (hashes back, git diff HEAD empty). HONEST DETAIL: in the control leg my single-line dist grep returned 0 because tsc re-prints an intersection across multiple lines; I re-ran the leg to confirm the emitted text (`$filter?: Record< string, any > & {` at dist line 64) and the hash difference. The mutation landed; the grep ANCHOR was too narrow, not the mutation.\n7) GATES: check-changeset-presence.mjs EXIT 1 before (naming this file), EXIT 0 after with the empty-frontmatter exemption; check-changeset-no-major EXIT 0; check-changeset-overwrite EXIT 0; check-control-bytes EXIT 0 (5485 files) plus a manual control-byte scan of my two files: clean.\n8) DECLARED NARROWING (lint): targeted, not the repo-wide farm. Population 3873 files read from ESLint's OWN ignore resolution over tracked sources; the run linted 1 file counted from `--format json`; 0 errors and 36 pre-existing no-explicit-any warnings (10 fewer than before). eslint.config.js sets no `parserOptions.project` and no `projectService`, so type-aware linting is OFF and this diff cannot move any untouched file's verdict. CI runs the full farm.\n9) LOCK DISCLOSURE: the heavy half (types build + full core type-check + --listFiles) ran UNDER scripts/pm/os-verify-lock.sh, 'VERDICT command-exit 0 - held the lock 56s - waited 294s'. The remaining single-file vitest and single-project control ran OFF the lock as a DECLARED narrowing, after three consecutive `exit 99` queue-timeouts (~30 min) behind holder pid 18808/25261/32636 = objectui-issue-6366's multi-shard `pnpm test`. Those 99s are recorded as NOT MEASURED, never as reds and never inside a green count.",
 "mcp_calls": "2 - create_pull_request, add_issue_comment. All issue/PR reads went through the zero-quota unauthenticated REST channel (probed 200 first).",
 "open_questions": [
 {
 "question": "Your fence said exactly one file may appear in my commits, but you also ordered me to follow the repo's changeset policy and let the Changeset Declaration gate decide. The gate ruled a declaration IS owed for packages/core/src/**, with no carve-out for tests. Was the one-file fence meant to exclude the gate-mandated changeset too?",
 "options": [
 "A - Intended as I resolved it: the fence targets SOURCE files (packages/types, spec, renderer code), and .changeset/6001-drop-redundant-filter-casts.md with empty frontmatter is the gate's own sanctioned answer, so the branch legitimately carries 2 files.",
 "B - The fence was literal: I should have stopped and reported instead of adding the changeset, and you would have decided how to satisfy the gate.",
 "C - Drop the changeset and let CI's Changeset Declaration gate fail, then decide."
 ],
 "recommendation": "A. The two clauses of the order conflict only if 'file' means any file rather than any source file, and you explicitly delegated this decision to the gate with 'do not pre-empt it in either direction'. The gate pre-empted nothing: it ran red naming my file, then green once declared. C is a knowingly red CI. If you meant B, the changeset is one isolated commit (234415d6) and drops cleanly."
 }
 ],
 "out_of_scope_findings": []
}

Generated by Claude Code

---

Generated by Claude Code