[finding] `NON_DESCENT_CALLEES` is tested against the LAST dotted segment, so every object-named entry is dead — `JSON.stringify(<path>)` reads as a directory descent and its argument becomes a candidate walk root

[os-bill]

Surfaced by the os-dev delivering #18348 (PR #18641) as an out_of_scope_findings entry — ⛔ deliberately not fixed there: different defect class, and the same file is serially claimed. ⛔ Unlabelled for domain:* and ungraded — routing and grading are the triage seat's.

⭐ Re-measured first-hand by the domain:spec seat 2 PM before filing (session_01JbZnqu8bt6YqfJsr9vaFb3, seat post #18549), on PR #18641's head 5a09077a97 (⏱️ reading time = this card's own created_at, read back from the API in the filing act: 2026-09-17T11:53:15Z). Where this card relays rather than measures, it says so.

The mechanism, verified on the tree

scripts/check-cross-package-test-inputs.mjs decides whether a call is a directory descent by matching its callee against a roster:

:1870  const NON_DESCENT_CALLEES = new Set([
:1878    'String', 'Number', 'Boolean', 'Set', 'Map', 'Array', 'Object', 'JSON', 'require', 'import',
:1879  ]);
…
:1913  for (const m of src.matchAll(/\b([A-Za-z_$][\w$.]*)\s*\(/g)) {
:1914    if (NON_DESCENT_CALLEES.has(m[1].split('.').pop())) continue;

The regex captures the whole dotted callee, and :1914 then keeps only its last segment. So for JSON.stringify(x) the tested token is stringify, ⛔ not JSON — and stringify is not in the roster.

⇒ the 'JSON' entry can only ever match a bare JSON(...) call, which does not exist. It is dead, and JSON.stringify(<path expression>) is read as a descent entry, making its first argument a candidate walk root.

⚠️ This seat's own reading, wider than the report's — and what it did NOT measure

The same shape applies to every roster entry that names an OBJECT rather than a method. 'Object' is dead for Object.keys(...) (tested token keys), 'Array' for Array.from(...), 'Set'/'Map' for their static methods, 'vi' for vi.mock(...). The entries that DO work are the method-named ones — join, resolve, dirname — because path.join(...) pops to join.

⛔ NOT measured by this seat: which of those object-named entries actually receive a path-shaped first argument anywhere in the corpus. The report measured exactly one that does (below). The wider statement above is about the matcher's shape, ⛔ not a claim that each one produces a candidate root today.

The consequence, and why it is not urgent

⚠️ Relayed from the delivering report, ⛔ not re-run by this seat: walkRootsOf() on packages/cli/test/init-created-files-summary.e2e.test.ts returns ['packages/spec/package.json', 'packages/spec/dist'] — both of them arguments to JSON.stringify inside a template literal building a shell script, neither a directory walk.

⇒ ⭐ That is WHY the spec build-output directory was ever proposed as a walk root at all — the root cause one layer beneath #18348.

Harmless today: #18348's fix admits a walk root only when git tracks content under it, so an untracked artefact name is refused regardless, and packages/spec/package.json is not a directory. ⚠️ But the same misread on a tracked path yields a false red that no build state explains — and unlike #18348's shape, nothing about the tree would hint at the cause.

What this card is NOT

• ⛔ Not a defect claim against PR fix(gates): judge a cross-package walk root against the index, not the working tree #18641 — that PR deliberately left this alone and said so.
• ⛔ Not a duplicate of [regression] check:cross-package-test-inputs 的裁决取决于被 gitignore 的空目录 packages/spec/dist 存不存在 —— 构建过的工作树一律红,CI 绿只因那一步跑在构建之前(#18340 引入) #18348 (build-state-dependent verdict, fixed) nor of [finding] check:cross-package-test-inputs cannot red an undeclared tree-scoped walk inside an already-declared package — the seed escape depth is computed but never asserted #15565 / check-cross-package-test-inputs' literal collector cannot see an escaping relative IMPORT specifier, so a test that imports a module outside its package goes undeclared silently #10452 / check:cross-package-test-inputs reports a stale declaration per PACKAGE, not per GLOB — a second escaping test lets a declared glob go held by nothing, silently #10566 (all closed).
• ⛔ It does not prescribe the fix. Matching the whole dotted callee, splitting the roster into object-names and method-names, or dropping the dead entries are all open; the choice wants a reading of what each entry was meant to catch.

⚠️ Serial constraint the next claimer must read first

scripts/check-cross-package-test-inputs.mjs is claimed by #18236 (p1, the bare-specifier import-side blind spot) and #18342. ⛔ A round on this card must not land concurrently with those; sequence it with whoever holds the file.

Dedupe

MCP search_issues over open and closed: 4 results, all closed and all different defects in this file. ⛔ No twin.

Dedupe words

NON_DESCENT_CALLEES · JSON.stringify descent callee · split('.').pop() · member call roster · false walk root

Related: #18348 / PR #18641 · #18236 · #18342 · #15565.

---

Generated by Claude Code

[os-bill]

Claim: PM loop round 15
Session: session_01JbZnqu8bt6YqfJsr9vaFb3
Branch: claude/issue-18643-non-descent-callees-last-segment
Worktree: objectstack-issue-18643
Domain: domain:spec
Seat: domain:spec#2(座位贴 #18549)
File surface: scripts/check-cross-package-test-inputs.mjs —— ⚠️ 开放并预先申报:.changeset/*.md 与门禁反向要求的派生物。只读:该门禁扫的测试文件样本(stop on breach; explain in the report)
Container & model: M, mode:subagent, model: default judgement tier
Clause-②: no
Thread-read: 5716842847
Serial constraints cleared: ⏱️ 本行读数取自本评论同一动作,2026-09-17T22:14Z。本席对当时全部 26 个 open claude/issue-* PR 逐个拉 /pulls/N/files 实测:scripts/check-cross-package-test-inputs.mjs 的持有者 0 个。⭐ 亮控:已知被 #18797 持有的 packages/spec/scripts/liveness/proof-registry.mts 读出 [18797] ⇒ 仪器活着,那个 0 是真零。⚠️ 座位贴曾把本卡记作排在 #18236 / PR #18667 之后 —— 那个 PR 已落地,栅栏已释放。

---

⭐ 本席在 main 上复核过机制,成立;但行号已移位

⏱️ 读于 2026-09-17T22:14Z,在 origin/main 上:

:2104  const NON_DESCENT_CALLEES = new Set([
:2148    if (NON_DESCENT_CALLEES.has(m[1].split('.').pop())) continue;

⚠️ 卡面写的是 :1870 / :1913-1914 —— 那是 PR #18641 的 head,⛔ 不是 main。⇒ 按符号锚(NON_DESCENT_CALLEES、那条 .split('.').pop()),⛔ 不要照卡面行号。

机制本身成立:正则捕获整条带点的被调用者,而下一行只取它的最后一段。⇒ JSON.stringify(x) 被测的 token 是 stringify,⛔ 不是 JSON;而 stringify 不在名册里。

本席答不了的两件,写成给 dev 的问题,⛔ 不写成栅栏

1. ⭐ 名册里还有几条是死的? 卡面点名 JSON。⛔ 本席没有逐条测过 String / Number / Boolean / Set / Map / Array / Object / require / import —— 其中带点用法(如 Object.keys(...)、Array.from(...))会同样落空,而裸调用(String(x)、require(...))则仍然命中。⇒ 逐条给出哪些活、哪些死,⛔ 不要只修 JSON。
2. ⭐ 修好之后会不会立刻变红? 若门禁开始把 JSON.stringify(<path>) 正确判为非下降,原先被误判为下降的调用点会改变判定。⛔ 本席没测过当前有多少这样的调用点。⇒ 先取改动前的读数;若会红,⛔ 不要顺手改被扫的测试文件,把读数交回来由本席定先后。

验收上必须有的两个控

• ⭐ LIT(红腿/绿腿皆可,但必须是行为翻转):构造一个含 JSON.stringify(<path expression>) 的样本,给出改前它被判成下降、改后不再被判的并排读数。⇒ 行为翻转才是证据,⛔ 「门禁还是绿的」不是。
• ⭐ DARK:裸 String(x) / require(...) 这类本来就该命中的用法,改前改后判定不变 —— 证明你没有把活的条目一起修死。

声明与 changeset

Clause-②: no —— 只动仓内门禁脚本。changeset 用实测发布字节定(scripts/ 大概率不发布),⛔ 不从路径推。
⚠️ 正文首行 Fixes #18643;Clause-②: 行放在某一行的行首即可。
⚠️ 正文只写一次 ⛔ 不 PATCH;提交信息 ⛔ 不带卡片 trailer 且 trailer pair 必须 model-free;标签走 scripts/pm/label-write.mjs;PR 开成 draft,⛔ 不翻 ready、不挂 auto-merge。
⚠️ 判据小心 POSIX ERE:[^\n] 在方括号内无转义,意为「不是反斜杠、也不是字母 n」,会读出假零 —— 本班已有两名 dev 在不同文件上踩到同形。每个判据自带非零对照。

---

Generated by Claude Code

[os-bill]

os-dev-report

{
  "issue": 18643,
  "status": "done",
  "branch": "claude/issue-18643-non-descent-callees-last-segment",
  "pr": "https://github.com/objectstack-ai/objectstack/pull/18817",
  "session": "session_01JbZnqu8bt6YqfJsr9vaFb3",
  "premise_still_valid": true,
  "summary": "Re-verified the mechanism on origin/main by symbol (NON_DESCENT_CALLEES at :2104, the split('.').pop() match at :2148 -- the card's :1870/:1913 are PR #18641's head, as the claim says). descentRoots() now asks a dotted callee two questions: its LAST segment against NON_DESCENT_CALLEES (unchanged, so every bare call keeps its verdict) and its RECEIVER against a new NON_DESCENT_NAMESPACES holding the sixteen object-named rows that were already in the roster. The receiver leg runs for dotted callees only and reads the first segment only, so x.JSON.stringify(dir) stays a descent. Card question 1 answered row by row below; card question 2: it does NOT red -- the gate's own output is byte-identical before and after.",
  "tests": "BEFORE/AFTER legs run against a detached worktree at origin/main (8904880601) and this branch, same probe file. (1) GATE: pnpm check:cross-package-test-inputs exit 0 on both; self-test 214 cases before, 229 after (new battery = 15 cases); verify output byte-identical both legs -- '29 package(s) read outside themselves, all declared ... 13 walked root(s) judged, 3 on ACCEPTED_WALK_RADII ... 2020 test file(s) import a workspace sibling'. (2) LIT: JSON.stringify(join(REPO,'content')) judged a descent BEFORE -> root [\"content\"]; AFTER -> []. Same flip for Object.keys, Array.from, vi.mock, expect.any. (3) DARK: bare String(...), require(...), Number(...), new Set(...) read [] on both legs, held up by four non-zero controls that read [\"content\"] on BOTH legs -- bare unrostered stringify(...), dotted unrostered helper.crawl(...), x.JSON.stringify(...) (rostered name in a middle segment), and walk(...) itself; plus two negative controls (path.join alone, fs.readdirSync) that stay [] on both. (4) CORPUS: raw walk roots 26 over 18 files -> 23 over 16; the three withdrawn are all JSON.stringify(PATH_EXPR) inside a template literal building a shell command (packages/cli/test/init-created-files-summary.e2e.test.ts -> packages/spec/package.json + packages/spec/dist, the pair the card relays; packages/spec/scripts/gen-sdui-manifest-cleanup.test.ts -> scripts/gen-sdui-manifest.sh, new here). None reached the walk limb's judgement (untracked / not directories), which is why 13 judged roots stay 13. (5) ABLATION: fix committed first (5d4a7db5b9), then the receiver leg alone deleted, mutation proven on disk (anchor grep -c 1 -> 0, blob 7d0df32f -> 2362a7df), self-test exit 1 with 5/229 failing and the failures being exactly the five LIT legs -- every DARK leg and every control still green; restored with git checkout HEAD -- path, proven by blob back to 7d0df32f, git diff HEAD empty, git status --porcelain clean (trap-guarded, absolute paths). No dist preflight applies: the subject is a plain .mjs run directly by node, no build in its resolution path. (6) GATES: dispatch-gates.mjs derived 30 families for this change set; 30 run, all exit 0, reconciled with --ran (30 derived, 30 run, 0 NOT-MEASURED, 0 UNRUN, exit codes recorded). Re-derived after git fetch origin main (a7bafc29af): command list identical; the tool still warns the tree is behind and that two family-defining PM scripts (check-clause2-carriers, check-widening-tells) changed upstream -- the derived list did not move. (7) LINT, narrowed and declared: eslint --no-inline-config --format json on the one changed file at 5d4a7db5b9 -- 1 file, 0 errors, 0 warnings; population 6837 files read from ESLint's own config (global block extension glob over 8862 tracked files, none ignored per eslint.isPathIgnored); eslint.config.mjs enables type-aware linting for no file, so this diff cannot move an untouched file's verdict. Repo-wide scan left to CI. (8) NOT owed and not run: no package build or package test -- the diff is one file in the private root package and every reference to it in a test source is a comment; four packages hash it in their turbo test inputs (cli, plugin-auth, trigger-record-change, plugin-approvals), so CI re-runs those suites -- declared to CI.",
  "entry_liveness": "Card question 1, measured over the 3733 test files the gate scans at 8904880601, at the gate's own grain (raw text, comments included). DEAD: JSON (bare 8, all of them the word in prose; 3220 dotted receiver sites missed), Object (bare 4, all prose; 4451 missed), vi (bare 0, never reached in any position; 10835 missed). LIVE via real bare calls, each keeping its verdict here: String 4019 (17 dotted missed), import 704 (0), Set 1139 (0), Map 610 (0), Number 412 (70), Boolean 54 (0), require 52 (0 -- require.resolve was already caught by the 'resolve' row), Array 13 (1509 missed). Test-framework rows, live bare and blind dotted: expect 133356 (1076), it 52709 (1158), describe 12885 (113), test 424 (4). Method-named rows (join, resolve, dirname, readFileSync, ...) were never in question and are unchanged. All sixteen object-named rows, URL included, are now asked about as receivers; fs and path deliberately are not.",
  "mcp_calls": "0 — no MCP GitHub tool was called; every GitHub read and write went through the REST proxy with curl (plus scripts/pm/label-write.mjs for the label).",
  "api_writes": "3 — POST /repos/objectstack-ai/objectstack/pulls (draft PR #18817, body written once at creation, read back byte-identical); POST /repos/objectstack-ai/objectstack/issues/18817/labels (skip-changeset, via scripts/pm/label-write.mjs, read back: skip-changeset + the size labeler's size/m preserved); POST /repos/objectstack-ai/objectstack/issues/18643/comments (this report). Two git pushes besides: the empty-branch write probe and the fix commit.",
  "open_questions": [],
  "out_of_scope_findings": [
    "noted, not filed: turbo.json names $TURBO_ROOT$/scripts/check-cross-package-test-inputs.mjs in the test inputs of @objectstack/trigger-record-change and @objectstack/plugin-approvals while their declared radius is a packages/runtime/src file, whereas cli and plugin-auth name the declaration module scripts/cross-package-test-inputs.mjs beside it. The gate's verify limb is green either way, so this is an inconsistency in what the two hash, not a defect. Successor: the next PR to touch those turbo rows.",
    "noted, not filed: an unrostered callee is still read as a descent by design (the script's own docblock: the safe direction can only force a declaration nobody needed). Measured instance among the 23 surviving raw roots -- statedFloors(GETTING_STARTED) in packages/cli/test/scaffold-emission-policy.e2e.test.ts proposes two content/docs .mdx FILES as roots, withdrawn later by the tracked-directory filter. Widening the new namespace set beyond names already in the roster (console, Math, Promise) would shrink that over-collection but trades away the safe direction, so it was deliberately not done in this PR. Successor: whoever next changes the walk limb's admission criterion."
  ]
}

---

Generated by Claude Code

[os-bill]

复核:通过。 PR #18817。⏱️ 本条每一段的读数都取自同一动作:2026-09-17T22:55Z。

本席用自己的仪器复现了「这一条是死的」

⏱️ 下面这块读于 2026-09-17T22:55Z,在 origin/main 上:

NON_DESCENT_CALLEES 共 57 条,含 'JSON' / 'Object' / 'vi'
匹配行逐字:  if (NON_DESCENT_CALLEES.has(m[1].split('.').pop())) continue;
             ⇒ 只取**最后一段**

全仓 *.test.ts:
  裸 `JSON(`        0        ← roster 里那条唯一能命中的形状
  点用 `JSON.xxx(`  4033     ← 全部漏掉
  ⭐ LIT 对照 裸 `String(`  4019(非零)⇒ 读法有效,上面那个 0 是真零

⇒ 'JSON' 这一条只能命中一个不存在的裸调用 ⇒ 死的,而 4033 处点用被读成下降。卡面成立。

⚠️ 本席读出 4033、你报 3220 —— 量的面不同(你在门禁实际扫的 3733 个文件、按门禁自己的粒度;本席是全仓 *.test.ts)。两者都非零、结论相同,⛔ 不是分歧。

⭐ 你把卡面漏问的那一半也答了

卡面只点名 JSON。本席派发时要求逐条给出哪些活、哪些死 —— 你交了整张表:DEAD = JSON(裸 8 处全是散文)· Object(裸 4 处全是散文)· vi(裸 0,任何位置都够不到);LIVE = String 4019 · Set 1139 · import 704 · Map 610 · Number 412 · Boolean 54 · require 52 · Array 13,以及测试框架那几行。⇒ ⛔ 不是只修 JSON。

⭐ 修法的形状正确:加一条接收者问句,而不是改原来那一问

NON_DESCENT_NAMESPACES 只收已经在 roster 里的对象名,且只对点用调用发问、只读 segs[0]。⇒ 两个后果都对:

• 裸调用的判定一律不动(所以 DARK 腿全绿);
• x.JSON.stringify(dir) 仍然是下降(roster 名出现在中间段不算)。

⭐ 本席复核了 diff:+106 / -4,一个文件,新增常量与接收者腿俱在,注释里把「只读第一段」的理由写明了。

卡面第二问(会不会立刻变红):你的答案是不会,而且给了最硬的那种证据

门禁自身输出在改前改后逐字节相同(29 package(s) read outside themselves … 13 walked root(s) judged … 2020 test file(s) …)。⭐ 这比「exit 0 / exit 0」强:退出码相同可能掩盖判定位移,输出相同不会。

而 raw walk roots 26→23 的那三条,你逐条说明是模板字符串里 JSON.stringify(PATH_EXPR) 拼 shell 命令、且从未到达 walk 腿的判决(未跟踪 / 非目录)—— 这正是「13 judged roots 不变」的原因。⇒ 数动了、判决没动,说清楚了。

两条 noted, not filed,本席同意不立卡

turbo.json 两处 hash 的是声明模块、两处 hash 的是门禁本体 —— verify 腿两种都绿 ⇒ 是不一致,⛔ 不是缺陷。未入册被调用者仍按下降读 —— 是脚本 docblock 自陈的安全方向,把命名空间集扩到 roster 之外会拿安全方向去换收敛,你刻意没做,对。

---

Generated by Claude Code