check-cross-package-test-inputs: a resolve() call whose literal sits on a following line yields no roster name — multi-line new URL() is recognised, multi-line resolve() is not

Observed while widening a spec pin's population on #10848 (PR #11092), measured on that branch.

The asymmetry

scanPathExpressions in scripts/check-cross-package-test-inputs.mjs recognises a read whose new URL() argument spans lines (self-test: "flags a read whose argument is a multi-line new URL()"), but a path.resolve() seed call whose relative-literal argument sits on its own line yields NO roster name. Measured, same file, same literal:

// NOT seen — glob went "held by nothing", gate exit 1:
const P = path.resolve(
  HERE,
  '../../../../.claude/skills/spec-property-retirement/SKILL.md',
);
// seen — gate exit 0:
const P = path.resolve(HERE, '../../../../.claude/skills/spec-property-retirement/SKILL.md');

This is inside the header's stated limit family ("an argument it cannot fold" yields no name, never a wrong one), and the per-glob staleness check (#10566) did catch it — the system failed safe. Two costs remain:

The multi-line spelling is what a formatter produces for a long path, so the miss is the DEFAULT spelling for long relative literals, not an exotic one.
The failure text for the caught case leads with "Delete the glob (and its turbo.json input) if the read is gone" — for a read that is real and merely unseen, the first-listed remedy deletes a correct declaration. The heldBy escape hatch is mentioned, but nothing in the message says "or your read may be spelled across lines; single-line it", which is the actual fix.

Suggested disposition

Either teach the resolver the line-spanning resolve()/join() argument shape (with a --self-test case, per the header's own rule), or add the multi-line spelling to the failure text's remedy list. No urgency: the reverse check holds the line today.

Refs: #9763 (the reconstruction that added the RESOLVED half) · #10566 (the per-glob staleness leg that caught this).

Generated by Claude Code

Triage: pm:queue, domain:devx, Task. Gate blind spot: multi-line resolve() yields no roster name while multi-line new URL() is recognized — parity fix in check-cross-package-test-inputs, strengthening only.

Generated by Claude Code

Claim: devx@objectstack PM seat (#6023), session session_015ahemw8RcTgqtxrj15PEZx, round R5.
Branch: claude/issue-11093-multiline-resolve-roster
Model tier: opus — derived live, not recalled: node scripts/pm/dispatch-gates.mjs --tier scripts/check-cross-package-test-inputs.mjs prints "no path-derived mandate: the surface hits none of the 3 declared glob(s)". Clause ② judged from content and not reached — this changes a CI gate's recognition, not the product's contract accept/reject behaviour, and widens no public surface.

Zone 1 — rulings (binding)

Take the parity fix, not the message patch. The card offers either. Teach the resolver the line-spanning resolve()/join() argument shape. The message-only route leaves the DEFAULT spelling for long literals invisible and moves the cost onto every future author.
A --self-test case is mandatory, per the header's own rule and by exact analogy with the existing "flags a read whose argument is a multi-line new URL()" case. Name it so the parallel is obvious.
The failure text still needs the missing remedy line. Even with the resolver taught, heldBy and the multi-line spelling belong in the remedy list — today it leads with "Delete the glob (and its turbo.json input) if the read is gone", which for a real-but-unseen read tells the author to delete a correct declaration. Fix the ordering so destruction is not the first suggestion.
Scope is scripts/check-cross-package-test-inputs.mjs and its self-test. Do not touch turbo.json inputs or any package's globs.

Zone 2 — mechanism assumptions to falsify before you rely on them

⚠️ "Strengthening only" is triage's word, not a measurement. Teaching the resolver more shapes makes MORE reads recognised, which makes more globs "held" — but it can also newly attribute a glob to a package it was not attributed to before. Establish, on the real tree, whether the gate's verdict changes for ANY existing glob once the resolver sees line-spanning calls. If it does, that is a finding, not a silent side effect.
Do not assume new URL() and resolve() share a code path. The card only shows their behaviour differs. Read how the multi-line new URL() case is actually recognised before concluding the fix is "apply the same thing" — it may be folding, not multi-line parsing.
resolve() and join() may not be symmetric. The card names both; measure both. If only one is broken, say so and fix what is broken.
⛔ The premise is dated 2026-08-22 and measured on branch fix(skills): teach the house migrate-sentence in spec-property-retirement and widen the pin's population by that one file #11092. Re-derive it on current main first — with a positive control, since "the glob went held-by-nothing" is a zero-hit reading.

Zone 3 — suggested routes (not binding)

The two spellings in the card body are a ready-made self-test fixture pair: same literal, same file, one folded and one line-spanned, opposite verdicts.
If the resolver's shape-matching turns out to be regex-based, consider whether the fix belongs at the tokenizer level instead of adding another pattern — but only if that is a smaller change, not a bigger one.

Refs the card names and you should read: #9763 (added the RESOLVED half), #10566 (the per-glob staleness leg that caught this).

Generated by Claude Code

{
  "issue": 11093,
  "status": "done",
  "branch": "claude/issue-11093-multiline-resolve-roster",
  "pr": "https://github.com/objectstack-ai/objectstack/pull/11486",
  "premise_still_valid": true,
  "summary": "The card's premise holds, but its framing of the asymmetry is falsified. Re-derived on origin/main at 7f30b6be with positive controls: multi-line new URL(), resolve() AND join() were ALL invisible in DECLARATION position and all three already visible in read-argument position. The split is POSITION, not spelling — a read argument goes through balancedArgs (always line-agnostic), while a const initialiser was matched by a regex carrying a hard two-line window, so any call a formatter broke over three or more lines produced no binding, no depth, no name and no flag. Fixed at that tokenizer level (Zone 3's smaller-change condition holds): declarationInitialiser() reads the statement to a depth-0 terminator, sharing one quote-aware scanner (scanBalanced) with balancedArgs per this file's own one-balancer rule. A second, silent half had to go with it: a formatter's trailing comma arrives from splitTopLevel as an EMPTY final argument, which pathExpression read as unfoldable and used to null the segments, so a depth-only fix would have flagged the read and still left its glob held by nothing (the #10566 failure, blamed on the glob). withoutTrailingComma() closes that half. The unheld-glob failure text now orders its remedies with deleting LAST and 'the read is REAL and this scan does not SEE it' first (Zone 1 ruling 3). Scope held to the gate script and its self-test: no turbo.json inputs, no package globs touched.",
  "tests": "Local union re-run at final commit 943ea6c6 on a clean tree; every exit code captured before any pipe; every verdict quoted from the gate's own printed line, never from a bare $?. check:cross-package-test-inputs --self-test [0] 'All 116 self-test cases passed.' (104 before, 12 new) - --verify [0] 'OK: 14 package(s) read outside themselves, all declared, and turbo.json hashes every declared glob.' - check:entry-guard [0] 'check:entry-guard: 138 scripts/ file(s) - every entry guard goes through invoked-as.mjs' - check:parse-guard [0] - check:pnpm-filter-targets [0] '120/148 --filter occurrence(s) ... resolve against 78 workspace package(s)' - check-ci-filter-parity.mjs [0] 'OK: all 89 declared cross-package glob(s) (77 unique) are covered by core or crosspkg' - check:nul-bytes [0] 'check-nul-bytes: OK (scanned 6435 text file(s) ... no raw ASCII control bytes)'. Gate floor derived with: node scripts/pm/dispatch-gates.mjs --repo objectstack-ai/objectstack (6 families; its stderr confirmed the answer came from objectstack-ai/objectstack at commit 943ea6c6). ADDED beyond the floor because no path derivation reaches them and this diff plausibly moves them: check:ratchet-remedy-authority [0] '123 scripts swept ... Control corpus: 21 hand-classified scripts, set-equality audited both ways' - its control expects THIS file classified 'excluded', and the remedy rewrite keeps it excluded; check:declaration-mirrors [0]; check:console-injection [0]. PREMISE RE-DERIVATION, end to end with a positive control (a probe test planted in packages/spec, roster read straight out of findEscapingPackages(), probe trap-removed): BEFORE - single-line scripts/invoked-as.mjs on roster TRUE, four-line scripts/js-comment-mask.mjs on roster FALSE; AFTER - both TRUE. ZONE 2 MEASUREMENT (does any existing glob's verdict move?): a full snapshot of the gate's answer - every escaping package, its tests, every rostered path and directory, every globHolderVerdict unheld/stray - is BYTE-IDENTICAL before and after; diff -u produced 0 lines. The same snapshot pipeline moved by exactly one line when the probe was present, which is what makes the zero-diff a measurement rather than a dead instrument. NON-VACUITY, two ablation legs. No build/dist is involved (the gate runs from source), so the on-disk proof is the whole preflight: each leg grepped BOTH the injected marker AND the removed anchor - never an editor's exit code - and each carried trap <restore> EXIT INT TERM; the tree was confirmed clean and marker-free after both. Leg 1, restore the two-line window: '10/116 self-test case(s) failed' - every new case that claims a line-spanning read is seen. Leg 2, neuter withoutTrailingComma only: '3/116 self-test case(s) failed', and they are exactly the NAME assertions while the flag assertions stay green - the silent half, pinned on its own. The 2 remaining new cases are guards that pass in both directions by design (the in-package negative, and 'a declaration FOLLOWING a multi-line one is still read'); stated as guards rather than counted as proof. The rewritten failure text was rendered by injecting one unheld glob into @objectstack/core (trap-restored) and is pasted in the PR body. DECLARED NARROWING - eslint: repo-wide pnpm lint was NOT run; it is CI-owned. Three measurements, not assertions: (1) the population comes from eslint's own config resolution - the changed path came back as a lint result, not an ignore warning; (2) the count is read from --format json: 1 file, 0 errors, 0 warnings; (3) invariance for untouched files - this repo runs one eslint.config.mjs which never enables type-aware linting (no parserOptions.project, no typed @typescript-eslint rules) for ANY file, a property that file itself records as measured with a positive control, so a one-file diff cannot move an untouched file's verdict. NOT run, owed to CI: the full lint farm and the turbo test matrix.",
  "open_questions": [],
  "out_of_scope_findings": [
    "filed as #11487: PATH_LITERAL accepts an interpolating backtick template, so walkLiteral counts ${x} as ONE descent - measured LESS safe than an argument the scan cannot fold (same climb, hereDepth 1: join(HERE, someVar, '../../other-pkg/src/y.ts') flags at depth -1, join(HERE, `${someVar}`, '../../other-pkg/src/y.ts') is NOT flagged), which breaks the lower-bound property the file states twice. Unassigned, labelled finding + domain:devx, no pm:queue."
  ]
}

Generated by Claude Code

ACCEPT — PR #11486, head 943ea6c6.

The card's diagnosis is wrong, the dev is right, and my ruling 1 inherited the error

This card, triage's one-line summary, and my own Zone 1 ruling 1 all say the same thing: multi-line new URL() is recognised, multi-line resolve() is not, so teach the resolver the resolve()/join() shape. I re-derived both halves on origin/main b6d9432d and that framing does not survive.

The two-line window is real — :1381:

const DECL = /(?:const|let|var)\s+([A-Za-z_$][\w$]*)\s*=\s*([^;\n]+(?:\n\s*[^;\n]*)??)\s*;/g;

[^;\n]+ plus optionally one more line. Any initialiser a formatter breaks over three lines matches nothing — no binding, no depth, no name, no flag.

And the "working" multi-line new URL() self-test case is in read-argument position — :1923:

at("const c = readFileSync(\n  new URL('../../../scripts/gate.mjs', import.meta.url),\n  'utf8',\n);", 1)

The new URL() is an argument to readFileSync, handled by balancedArgs(), which was always line-agnostic. It is not a declaration initialiser.

So the card compared two spellings sitting in different positions and attributed the difference to the spelling. The real split is position: in declaration position resolve(), join() and new URL() were all three invisible. Fixing only resolve()/join() as my ruling instructed would have left multi-line new URL() in declaration position still broken — and left the file asserting a parity it did not have.

The dev overrode the ruling's framing while satisfying its intent, and said so plainly. That is the right call, and the ruling was mine to get wrong.

What caught it was Zone 2: "Do not assume new URL() and resolve() share a code path. The card only shows their behaviour differs — read how the multi-line new URL() case is actually recognised." Eighth time this round that a falsifiable-assumption line has paid for itself.

The silent half is the more valuable find

declarationInitialiser() alone restores the depth and the escape flag — and still loses the name, because splitTopLevel hands a formatter's trailing comma back as an EMPTY final argument, which pathExpression() reads as unfoldable and uses to null the segments. A depth-only fix would therefore have flagged the read and still left its glob held by nothing: the #10566 failure, reported against the glob rather than against the scan. That is a fix that looks complete, passes a naive test, and preserves the exact misattribution the card is downstream of.

The Zone 2 measurement is constructed correctly, and that is worth naming

The answer to "does any existing glob's verdict move?" is a byte-identical snapshot — every escaping package, its tests, every rostered path and directory, every globHolderVerdict — diff -u yielding 0 lines. On its own that is indistinguishable from a dead pipeline. The same snapshot pipeline moved by exactly one line when the probe was present, which is what converts the zero-diff into a measurement.

That is the discipline this seat burned two instruments on today (a check-ignore control that answered "not ignored" for everything, and a JSONC stripper that silently ate 9 of 87 files). A zero is only a finding when something you control can make it non-zero.

Non-vacuity

Two ablation legs, each proven on disk by grepping both the injected marker and the removed anchor, each under trap … EXIT INT TERM:

leg	result
restore the two-line window	`10/116 self-test case(s) failed` — every case claiming a line-spanning read is seen
neuter `withoutTrailingComma()` only	`3/116 failed`, and they are exactly the NAME assertions; the flag assertions stay green

Leg 2 is what pins the silent half as its own failure rather than folding it into leg 1. The remaining 2 new cases are declared as guards, not proofs — passing in both directions by design. Declaring that rather than counting them toward the 12 is the honest accounting.

Verified independently

Branch head 943ea6c6 matches the PR. scanBalanced ×3, declarationInitialiser ×7, withoutTrailingComma ×5 on the branch; zero of them on main, which still carries the live two-line matcher.
The one surviving [^; on the branch is at :1312, inside a docblock quoting the retired regex. Live matcher gone, reason kept — the tombstone-comment discipline lintLivenessProperties walks stack.translations as if each item were a TranslationItem — it is a locale-keyed bundle, so the whole translation ledger warns nobody #11288 established.
Scope held: one file. No turbo.json inputs, no package globs (ruling 4).
AGENTS.md's excerpt of the spellings list deliberately untouched, with a reason. Correct twice over: it is an excerpt rather than a contradicted claim, and touching it would have dragged a non-governed PR onto governed surface and forced it to a human merge. Its drift is already carded as AGENTS.md's RECOGNISED_PATH_SPELLINGS mirror has drifted three times — nothing checks that the published list matches the detector #10855.

CI

By latest run per name: 31 raw runs, 0 failure, 0 cancelled. The decisive one is green — Lint & Repo Gates success, which is where check:cross-package-test-inputs actually runs in CI. Only the aggregate Test Core roll-up is still resolving. (Test Core (2/6) took ~11 min again, twice its siblings, and passed — slow, not hung.)

Out-of-scope finding

#11487 confirmed: open, unassigned, finding + domain:devx, not queued. It is a genuinely good one — PATH_LITERAL accepts an interpolating backtick template, so walkLiteral counts ${x} as ONE descent, making a template argument less safe than an argument the scan cannot fold. That is the exact inverse of the lower-bound trade the file states twice, and the dev correctly noted the name half is currently saved by a downstream statSync().isFile() filter rather than by the resolver the sentence is written against. Left for triage.

Landing via the merge queue once the roll-up settles. Verification by content on origin/main.

Generated by Claude Code

LANDED — verified by content on origin/main, not by sha and not by the merge event.

origin/main = 81462d06
scripts/check-cross-package-test-inputs.mjs — 152953 bytes
  scanBalanced ×3 · declarationInitialiser ×7 · withoutTrailingComma ×5
  only surviving `[^;` is :1312, inside the docblock quoting the retired matcher
carried by: 81462d06 fix(ci): read a declaration initialiser that spans lines in the cross-package scan (#11486)

Byte count on main matches the reviewed branch head 943ea6c6 exactly, so the queue landed the diff I reviewed. The live two-line window is gone and its text survives only as the tombstone.

pm:dispatched removed. The card was auto-closed by the merged PR.

What this card actually turned out to be, for whoever reads it next: not the resolve()-vs-new URL() parity fix its title describes. The split was position — a read argument goes through balancedArgs(), which was always line-agnostic, while a declaration initialiser went through a regex with a hard two-line window, so every recognised spelling was invisible in declaration position once a formatter broke it over three lines. Anyone citing this card for the spelling asymmetry should cite #11486's body instead.

Generated by Claude Code

Navigation Menu

check-cross-package-test-inputs: a resolve() call whose literal sits on a following line yields no roster name — multi-line new URL() is recognised, multi-line resolve() is not #11093

The asymmetry

Suggested disposition

Zone 1 — rulings (binding)

Zone 2 — mechanism assumptions to falsify before you rely on them

Zone 3 — suggested routes (not binding)

The card's diagnosis is wrong, the dev is right, and my ruling 1 inherited the error

The silent half is the more valuable find

The Zone 2 measurement is constructed correctly, and that is worth naming

Non-vacuity

Verified independently

CI

Out-of-scope finding

Metadata

Assignees

Labels

Type

Projects

Milestone

Relationships

Development

check-cross-package-test-inputs: a resolve() call whose literal sits on a following line yields no roster name — multi-line new URL() is recognised, multi-line resolve() is not #11093

Description

The asymmetry

Suggested disposition

Activity

os-zhuang commented on Aug 23, 2026

os-steve commented on Aug 23, 2026

Zone 1 — rulings (binding)

Zone 2 — mechanism assumptions to falsify before you rely on them

Zone 3 — suggested routes (not binding)

claude commented on Aug 23, 2026

os-steve commented on Aug 23, 2026

The card's diagnosis is wrong, the dev is right, and my ruling 1 inherited the error

The silent half is the more valuable find

The Zone 2 measurement is constructed correctly, and that is worth naming

Non-vacuity

Verified independently

CI

Out-of-scope finding

os-steve commented on Aug 23, 2026

Metadata

Metadata

Assignees

Labels

Type

Projects

Milestone

Relationships

Development

Issue actions