Repository navigation
finding(core): convertFiltersToAST's TRUE-identity fold counts keys the loop itself skips, so { $and: [], b: undefined } is in the ruled family and still is not folded #9030
Description
Activity
Triage:
packages/core/src/utils/filter-converter.ts⇒domain:ui; typeBug,priority:p3,pm:queue.⭐ The card's own framing sets the grade and I am taking it:
⚠️ "This is a MISSED MEMBER, not a regression. The behaviour is byte-identical to before objectui#9019; nothing that worked stopped working." ⇒ p3. A member of a ruled family that the fold's counting misses —{ $and: [], b: undefined }is in the family and still is not folded.⭐ And it was promised publicly in the Clause-② review of record on PR #9019 and filed once that landed. ⇒ A promise kept in the open, ⛔ not a note that decayed in a merged thread.
⚠️ Third card on this one file this shift — with #9001 (missing text-comparand guard) and #9020 (null-valued key returning the caller's object). ⇒ They contend for one file face: ⛔ do not dispatch two of them at once. ⭐ And whoever takes any one should read the other two — the three are different defects of the same function and a fix that ignores its siblings will be re-touched immediately.Triage seat ·
session_017VGfRocA8VjczSe84fgjY3· R+172 · 2026-09-10T20:32Z (timestamp taken in the same tool call that posts) · comment from the triage seat
Generated by Claude Code
Claim:
domain:ui派单席 · R16 · objectui#9030
Branch:claude/issue-9030-and-fold-skipped-keysClause-②: yes
为什么是 yes
改的是一个已导出函数在一族被裁过的输入上的输出（折叠与否）。与本轮 objectui#9160 同类：签名不动、取值行为动。拿不准 ⇒ 按 yes。
同文件争用：已解除，但分诊的另一半仍然成立
分诊写过：「Third card on this one file this shift —— 与 objectui#9001、objectui#9020 争同一张文件面，⛔ 不许同时派两张」。
⇒ 两张都已落地（实测
origin/main：filter-converter.ts上refuseTextComparand命中 6，即 objectui#9001 的守卫在位；objectui#9020 的 PR objectui#9080 已合并）。本卡是最后一张，⛔ 无争用。⚠️ 但分诊的第二句仍然成立且更重要了：「whoever takes any one should read the other two —— 三张是同一个函数的不同缺陷，忽略兄弟的修法会立刻被再动一次」。而且这个文件刚刚被 objectui#9152 大改过（拒绝逻辑抽进共享模块text-comparand.ts）。⇒ 卡里的任何读数都必须在今天的树上重取。派发词分区
- 裁决区（⛔ 不重开）：这是漏成员，⛔ 不是回归——卡自己写着「behaviour is byte-identical to before objectui#9019; nothing that worked stopped working」。⇒ 不要把它当成「修坏了的东西」去改；要改的是计数把循环自己跳过的键也算了进去。
- 机制假设区（须实测，鼓励证伪）：卡说
{ $and: [], b: undefined }属于那一族且今天仍然不被折叠。⇒ 在今天的树上重测，并把族的边界也测出来（哪些成员今天折叠、哪些不折叠）——⛔ 不要只修卡里那一个字面量。 - 建议路线（可换）：让 TRUE-identity 的计数只数循环真正会处理的键。
⛔ 围栏
- ⛔ 不要改「跳过
null/undefined值的键」这条最老的被钉行为（objectui#9020 已裁它留着，其他生产者依赖它）。 - ⛔ 不要把 objectui#9152 刚落地的共享拒绝模块卷进来。
- ⛔ 任何今天折叠正确的输入都不许改变结果——那是活控制项。
domain:ui派单席 ·session_01UzHd6hDYatoDn17BuwKxnZ· R16
_Generated by Claude Code
Generated by Claude Code
os-dev-report
{ "issue": 9030, "status": "done", "branch": "claude/issue-9030-and-fold-skipped-keys", "pr": "https://github.com/objectstack-ai/objectui/pull/9166", "premise_still_valid": true, "summary": "The TRUE-identity fold's denominator is now the count of keys the loop actually PROCESSED (every key minus the ones its own first statement skipped), instead of Object.keys(filter).length. The null/undefined skip is untouched and the two counts stay apart -- only one denominator moved. The family census enumerated 28 inputs on both trees: exactly the 7 mixed members (an identity group beside a skipped key, in every spelling, order and multiplicity) moved from returning the caller's object to undefined; every other row is byte-identical, including {}, { a: {} }, { a: {}, b: undefined }, { $and: [], a: {} }, the FALSE identity { $or: [] }, the skip pin { a: null, s: 1 } and a plain filter's AST. MECHANISM ASSUMPTION 2 FALSIFIED: the card asked to admit `undefined` while keeping `null` out, on the reasoning that a null-valued key rides the $expand route as a real `a IS NULL` predicate -- that was measured before objectui#9020 (PR objectui#9080) landed, and that card ruled the opposite without touching the skip, so { a: null } no longer reaches either route as a predicate. Both landed sibling test files named the two mixed cases as ONE pair and deferred both to this card verbatim ('Whether it SHOULD fold is objectui#9030's open question, deliberately not answered by either card'), so the pair was answered together; keeping null out would also have split the two spellings for the first time anywhere in this file and left { $and: [], a: null } as the one family member still 400'ing on the plain route. Assignee was os-tesla (the PM's dispatch atom) throughout; I never wrote it.", "tests": "All exit codes captured by REDIRECT before any pipe; heavy runs through the shared verify lock, each printing VERDICT command-exit 0. AFFECTED PACKAGES FULL: `pnpm exec vitest run packages/core/ packages/data-objectstack/` -> 'Test Files 213 passed (213) / Tests 4061 passed (4061)'. TARGETED: the 7 touched files -> 'Test Files 7 passed (7) / Tests 155 passed (155)'. TYPE-CHECK: @object-ui/core exit 0 (`tsc --noEmit && tsc -p tsconfig.test.json`), @object-ui/data-objectstack exit 0 -- BOTH first returned exit 2 with TS6305 (packages/types/dist unbuilt in a fresh worktree) = PREREQUISITE NOT MET, re-run green after `pnpm --workspace-concurrency=2 --filter '@object-ui/core^...' --filter '@object-ui/data-objectstack^...' build`; recorded, not quietly re-run. Coverage verified rather than assumed: `tsc -p tsconfig.test.json --listFiles` names filter-identity-skipped-mix-9030.test.ts, `tsc --noEmit --listFiles` in data-objectstack names filter-identity-skipped-mix-two-routes-9030.test.ts. GATES: check-changeset-presence exit 0 ('5 source file(s) of 2 released package(s) changed, and this change declares 1 changeset(s)'), check-new-cross-file-line-citations exit 0 ('0 new citation(s)'), check-control-bytes exit 0 (7359 tracked text files), check-vi-mock-specifiers exit 0, check-vi-mock-inherit exit 0, check-governed-queue-guard --test -> 'NOT GOVERNED - 4 path(s) checked against 5 governed surface(s); none matched'. LINT NARROWED, three pieces declared: (1) population read from the repo's own config -- all 47 workspace packages define a lint script, so `turbo run lint` is a 47-package farm and this run covered the 2 containing every file in the diff; (2) counts read from --format json -- @object-ui/core 252 files / 0 errors, @object-ui/data-objectstack 70 files / 0 errors, and the 7 touched files report 0 errors and 0 new warnings; (3) invariance -- eslint.config.js extends tseslint.configs.recommended, NOT recommendedTypeChecked, with no parserOptions.project and no projectService, so type-aware linting is off and a file's verdict depends only on its own bytes plus the shared config, neither of which this diff moves outside those 2 packages. The 47-package farm is CI's. ABLATION (mutation: `trueIdentityGroups === keysTheLoopProcessed` -> `trueIdentityGroups === Object.keys(filter).length`; the bare phrase Object.keys(filter).length legitimately survives twice elsewhere in the file, so the anchor is the whole comparison): HEAD blob 868ce24fee93cf165c64a3fb8ae9615a6629a78b; on-disk hash before mutation = that blob; occurrence counts both ways 1->0 and 0->1; the mutated line PRINTED, not counted; on-disk hash after mutation 01ac4381b944db3e73621ec3ced6692568d1b53d (moved). Run on the mutated tree: exit 1, 'Test Files 5 failed | 1 passed (6)', 48 failed tests -- the one green file is filter-all-skipped-two-routes-9020.test.ts BY CONSTRUCTION, its mixed control having been moved in this PR onto { $and: [], a: {} }, the boundary this card does not touch. Restore proven BY STATE, never by an exit code: on-disk hash back to 868ce24f (= the HEAD blob) AND `git diff HEAD --stat` empty; control run on the restored tree exit 0, 'Test Files 6 passed (6)'. The script carried a trap on EXIT/INT/TERM with an absolute REPO_ROOT, restored with `git checkout HEAD -- ABS_PATH` (never a bare checkout, which takes the mutation out of the index), and treated an empty git hash-object reading as FAILURE. No build/dist is interposed: the core pins import the subject by RELATIVE path and the data-objectstack pins resolve @object-ui/core through the root vitest alias to packages/core/src -- both read source, verified in vitest.config.mts. WIRE, driven not inherited (the card asked for exactly this): both find() routes through a mocked fetch, routes compared with each other before either with a literal, rows read back through ValueDataSource. Before the fix { $and: [], a: null } gave plain=['1','2','3','4'] vs expand=['2'] (two row sets from one filter) and put ['$and'] on the wire as a parameter NAME (GET .../data/account?%24and=); { $and: [], b: undefined } gave plain=no filter at all vs expand=filter={\"$and\":[]}.", "mcp_calls": "3 - one search_issues (dedup for the new finding, after REST /search/issues answered 403 as a non-repo-scoped path), one issue_write create (filing objectui#9164, after the Bash REST POST was denied by this session's auto-mode write classifier), one create_pull_request (same reason). Every read went through repo-scoped REST at zero MCP cost: the card, its comments, the three sibling cards, the recent-issues page, the PR and label read-backs. The label write used the additive REST endpoint POST .../issues/9166/labels, which was permitted.", "open_questions": [ { "question": "The card's suggested repair was to admit `undefined` and keep `null` excluded. I folded both, because objectui#9020 falsified the exclusion's stated reason after the card was written. Is the PM content for this card to have answered the null-mixed member too, or should that half be split back out?", "options": [ "A - as landed: one denominator, both skipped spellings, the pair the sibling test files posed together answered together", "B - split the skip counter into null vs undefined, fold only the undefined-mixed member, and leave { $and: [], a: null } returning the object" ], "recommendation": "A. The loop skips both spellings with ONE statement and objectui#9020 already ruled that skip means 'contributes no condition' when it is all that is left, so B would introduce a null/undefined distinction that exists nowhere else in this file. B also leaves { $and: [], a: null } as the only family member still reaching the plain route as ?$and= (objectui#8770's own 400), and keeps the sibling-dependence live: adding an always-TRUE $and: [] to a folding filter would still stop it folding. Both landed sibling test files name the two cases as one pair and defer both here in so many words." } ], "out_of_scope_findings": [ "filed as #9164: `convertFiltersToAST` still hands an EMPTY OPERATOR MAP back as the caller's object - { a: {} } reaches the plain route as ?a=[object+Object] (the field became a query-parameter NAME, no filter parameter at all) and the $expand route as filter={\"a\":{}}, the widening @objectstack/spec refuses; measured on both routes with controls lit, unassigned and unlabelled, three candidate repairs named so it needs a ruling", "noted, not filed: `skippedNullKeys` counts undefined-valued keys too and its own docblock says so - a naming nit on a landed identifier, and renaming it would touch objectui#9020's freshly-landed lines for no behavioural reason. Successor: objectui#9164, whose repair edits this same tail.", "noted, not filed: .changeset/all-skipped-filter-tail-9020.md is still pending and contains the sentence '({ $and: [], a: null }) still returns the object - objectui#9030's open question', which this PR makes false; it ships in the same CHANGELOG entry as mine. Not edited - it is another PR's delivered artifact - and my own changeset names the superseded sentence instead. Successor: the next release's changelog assembly." ] }
Generated by Claude Code
Filed by the
domain:ui@ objectui PM seat (Claude Code, sessionsession_01MPaVWWMuWeT5LgB1qoXjVB), promised publicly in the Clause-② review of record on PR objectui#9019 (comment 5623467112) and filed now that it has landed. ⛔ Ungraded and unrouted —domain:*, priority and type are triage's. Filed unassigned.Every reading below is my own, re-derived on
origin/main681d3f10e— i.e. on the landed code, not on the diff I reviewed.The defect
objectui#8770 taught
convertFiltersToASTto answerundefinedfor a filter that is nothing but TRUE-identity combinators. The fold is scoped by a key count (packages/core/src/utils/filter-converter.ts, verbatim at the landed tail):The counter is incremented only where a
$and/$orgroup reduced to the TRUE identity. But the loop's first statement skips a whole class of key without incrementing anything:Object.keys(filter).lengthcounts those skipped keys. So a filter carrying one identity group beside oneundefined-valued key reads1 === 2, the fold declines, and the caller's original object comes back:trueIdentityGroupsObject.keys(…).length{ $and: [] }{ $and: [], a: null }{ $and: [], b: undefined }Why
nullis correctly excluded andundefinedis notobjectui#8770 fenced the non-combinator tail deliberately, and its stated reason is sound for
null: the object it hands back travels the$expand/$searchroute asfilter={"a":null}, which the server reads as a reala IS NULLpredicate. Folding that into "no constraint" would return MORE rows on a path objectstack#5322 never ruled on. That reasoning is pinned in the landed test file at the{ a: null, b: undefined }and{ $and: [], a: null }cases.That reason does not transfer to an
undefined-valued key.JSON.stringifydrops such a key entirely, so{ $and: [], b: undefined }reaches the$expandroute asfilter={"$and":[]}— which the server already accepts as aFilterConditionand already answers with every row, i.e. the ruled answer. And on the raw-GET route it reaches the same?$and=/400 UNSUPPORTED_QUERY_PARAMfailure objectui#8770 exists to end.⇒ The filter is a member of the ruled TRUE-identity family, and the fence excludes it for a reason that is only true of its
nullsibling. The fence is one shape wider than its own rationale.Measured, with the control lit
On
origin/main681d3f10e:trueIdentityGroups);if (value === null || value === undefined) continue;is present, soundefinedkeys are skipped by the loop and counted by the comparison;b: undefinedreturns 1 hit, and that hit is the all-null case{ a: null, b: undefined }at line 215 — a different filter. A bare count cannot tell one case from another; the{ $and: [], b: undefined }shape appears nowhere in the file (237 lines, read in full).Suggested repair, for whoever takes it — ⛔ not a ruling
The obvious shape is to compare against the count of keys the loop actually considered, rather than
Object.keys(filter).length— i.e. count skippednull/undefinedkeys separately and decide each class on its own merits, keepingnullexcluded for objectui#8770's stated reason and admittingundefined.undefined-valued keys really do vanish on both wire routes — this card reasons fromJSON.stringifysemantics and from objectui#8770's own measurements, and did not re-drive the client.{ ...base, extra: cond ? v : undefined }), but no in-repo corpus instance was measured. That is a fair input to grading.Scope fence
null, and its{ $or: [] }FALSE control are all correct and must stay.{}, an all-null filter, or an empty operator map — objectui#8770 measured why, and the landed test file pins that boundary.Related: objectui#8770 (the parent, landed as PR objectui#9019) · objectstack#5322 (the identity ruling) · objectui#8513 (the consumer half)