finding(devx): a path appended to a package's test script WIDENS instead of narrowing — pnpm --filter @object-ui/cli test <one file> runs all 16 files and exits 0

Measured by the #7807 dev seat while pinning app-generator's two cwd-dependent cases (PR objectui#7810). Filed by the PM seat because that seat's two dedup channels were both down at the time (REST /search/* → 403; the single MCP search_issues fallback → API rate limit already exceeded for user ID 318158314), and it declined to file blind. The dedup was done here instead, over all 402 open items: no existing card carries this reading.

Measured

$ pnpm --filter @object-ui/cli test packages/cli/src/__tests__/app-generator.test.ts
Test Files  16 passed (16)
Tests       246 passed (246)

The path argument did not narrow the run — it ran the whole package. The reason is in the script itself:

"test": "vitest run --root ../.. packages/cli/"

Vitest unions positional filters. The appended file path does not replace the script's own packages/cli/ filter, it is added beside it, so every file the baked-in filter admits still runs.

Why this is a trap and not a preference

It exits 0 and prints a green summary, so the invocation reads exactly like a successful narrowed run. An agent that appends a path and reports "the file is green, 246 tests" is reporting the package's count as the file's — the number is real, the attribution is not. Nothing on screen distinguishes the two.

The control that this is a property of the script and not of the invocation: the same file narrowed through a form that carries no baked-in filter is 41 tests before the PR / 43 after —

$ pnpm exec vitest run --root ../.. packages/cli/src/__tests__/app-generator.test.ts   # cwd packages/cli
Test Files  1 passed (1)
Tests       43 passed (43)

— against 246 for the same file through pnpm --filter … test <path>. Same file, same tree, one number is 6× the other.

It is a third member of a family this repo already names

scripts/vitest-invocation-guard.mjs documents two invocation traps. This is the third, and it is documented in neither the guard's docstring nor AGENTS.md — both were grepped, by the dev seat and again here.

objectui#7800 — AGENTS.md claims pnpm --filter FIELD test reaches the same conclusion as CI (falsified twice, pm:awaiting-maintainer). A ruling there that removes the baked-in packages/<pkg>/ filter from the package test scripts would remove this hazard as a side effect, so the two should be read together before either is implemented.
objectui#7329 — parseVitestArgv collapses a repeated flag to its last value, the neighbouring "the guard's own argv reading is not what vitest does" defect.
objectui#7524 / objectui#7525 — the other two open cards on the per-package vitest invocation surface.

Scope note

⚠️ No fix is proposed here and none is obvious: the baked-in filter is what makes the bare pnpm --filter <pkg> test mean "this package", and removing it is objectui#7800's question, not this card's. What this card asks for is that the union behaviour stop being invisible — a guard reading, a documented sentence, or a narrowing form that actually narrows.

⛔ Do not fold this into objectui#7800 by editing that card: it is pm:awaiting-maintainer and its sentence is a maintainer's to rule on.

Blocked-by: #7800

⚠️ Added by the domain:devx @ objectui execution seat, 2026-09-08T07:48Z. ⛔ Not a new decision — it makes the existing one machine-readable. This card has been skipped at candidate selection since 2026-09-06 with the reason recorded in comment 5559757487, but it stayed pm:queue, so every dispatcher re-derives the same skip and the card never appears in the unlock scan when objectui#7800 is ruled. pm:blocked + this line is what the state model prescribes for exactly that case: the skip becomes automatic and the return is automatic. The dispatch shape once #7800 is ruled is already written in that comment; ⛔ nothing about the card's substance changes here.

Generated by Claude Code

Deferred at candidate selection, with the known pit recorded here so the next dispatcher does not re-derive it.

domain:devx @ objectui execution seat, PM session session_01FhBNJcLRZLe8M87VcUgpKr, R46, 2026-09-06T14:08Z. Read in candidate order (oldest unassigned pm:queue in the lane); labels untouched — this stays pm:queue.

Why not now. The card names three shapes and rules none: a guard reading, a documented sentence, or a narrowing form that actually narrows. The third is objectui#7800's question (pm:awaiting-maintainer) and the card itself says so. The second lands half in AGENTS.md (governed, domain:skills) and half in the guard's docstring — a sentence in a docstring is invisible at the moment the trap fires, which is the defect being described. The first — the guard refusing or warning when a package script's baked packages/NAME/ filter is unioned with an appended positional — is a gate behaviour change that #7800's ruling may make moot (if the baked filter leaves the scripts, the union has nothing to union with), and parseVitestArgv's own argv reading is already under objectui#7329.

Dispatch shape once #7800 is ruled (or if triage rules the guard-reading half independent of it): measurement-first — STEP 1: on origin/main, invoke pnpm --filter @object-ui/cli test packages/cli/src/__tests__/app-generator.test.ts under the guard with argv captured; confirm the guard can see both the baked filter and the appended positional at the point it runs. STEP 2 (conditional): if it can, the smallest LOUD reading (refuse with a message naming the union and the narrowing form that works: pnpm exec vitest run --root ../.. <path>), with a pin whose control is the current silent-green run; if it cannot, report why and stop. ⛔ Not the sentence-only shape; ⛔ not a change to the package test scripts (that is #7800's).

Blocked-by: is not written because the card explicitly declines to be folded into #7800; the dependency is recorded here in prose instead, and this seat re-reads it at the next candidate pass.

Generated by Claude Code

pm:queue → pm:blocked, Blocked-by: #7800. domain:devx @ objectui execution seat, PM session session_01FhBNJcLRZLe8M87VcUgpKr, R47, 2026-09-08T07:48Z. Label and body line written in one pass, read back and diffed.

⛔ Not a new decision. The decision is comment 5559757487 (2026-09-06): this card names three shapes and rules none, two of the three turn on objectui#7800's question, and the card body says so itself. What was wrong was the recording: the skip lived in a comment while the label still said pm:queue.

⇒ Two costs that this fixes, both of them mechanical:

every dispatcher that reached this card in candidate order had to re-derive the same skip by reading a comment — the reason existed but was not in the index that selection reads;
when objectui#7800 is ruled, the unlock scan walks Blocked-by: lines. Without one, this card would not have come back on its own.

objectui#7800 re-read on GitHub at 2026-09-08T07:46Z: open, pm:awaiting-maintainer, unassigned, last updated 2026-09-05T22:12:09Z. ⛔ Still a maintainer's to rule; ⛔ this seat does not answer it and ⛔ does not edit that card.

The dispatch shape for the moment it is ruled is already written out in 5559757487, measurement-first, ⛔ so nothing needs re-deriving then either. ⚠️ Triage may also rule the guard-reading half independent of #7800 — if it does, say so on this card and it returns to pm:queue without waiting.

⚠️ Body-write note, recorded because this seat checks its own writes: the read-back's prefix compared unequal to the pre-write body while the appended block compared equal and the content check passed (all three section headings ×1, the measured block ×1, objectui#7800 ×3, objectui#7329 ×1, a nonsense control ×0). ⇒ Line-ending normalisation by the API, ⛔ not content loss.

Generated by Claude Code

UNBLOCKED — `pm:blocked` → `pm:queue`. Upstream closed and the premise re-verified on today's `main`.

domain:devx @ objectui seat, session_013VGeMu3p6qEFWR6K6GGLaW, R58, 2026-09-14T12:1xZ, unlock scan over this lane's 27 held cards.

Blocked-by: #7800 → objectui#7800 is closed / completed.

⛔ A closed upstream is not by itself an unlock — this lane's 2026-08-27 precedent (objectui#8103) is a card whose upstream had closed while the card's own premise remained unmet, and a naive unlock would have put destructive work back in the queue. ⇒ the premise was re-verified, not assumed:

packages/cli/package.json on main today:

"test": "vitest run --root ../.. packages/cli/"

⭐ The script already names a path. So pnpm --filter @object-ui/cli test <somepath> appends a second positional filter, and vitest unions positional filters rather than intersecting them ⇒ the appended path widens the run instead of narrowing it. That is exactly this card's claim, still visible in the script text, ⛔ not inferred from the card body.

Control on the same read: the package's other scripts (build, dev, lint, type-check) are present, so the reader is parsing a real scripts block and the test value is a reading.

⇒ dispatchable. ⛔ Still unassigned and ⛔ not claimed by this comment — it goes back into the queue, not into flight.

Generated by Claude Code

Claim: PM loop round R58
Session: session_013VGeMu3p6qEFWR6K6GGLaW
Branch: claude/issue-7814-vitest-argv-union-guard
Worktree: objectui-issue-7814
Domain: domain:devx
File surface: scripts/vitest-invocation-guard.mjs (+ its pin under scripts/__tests__/). ⛔ NOT any package's test script, ⛔ NOT AGENTS.md. Stop on breach and explain in the report.
Container & model: M, mode:subagent, model: default judgment tier (opus)
Clause-②: no
Thread-read: 5559757487, 5581247919, 5663668009 — all three; comments field 3 == enumerable 3 ✓

⭐ The dispatch shape was written 8 days ago and I am NOT re-deriving it

Comment 5559757487 (2026-09-06) already laid out the measurement-first shape and ruled out the wrong ones. ⭐ That seat's foresight is why this dispatch is cheap, and the shape below is its work, carried forward deliberately rather than reinvented.

The blocker is discharged — and the half that mooted it did NOT happen

Blocked-by: #7800 → objectui#7800 is closed / completed, its PR objectui#8952 merged by a maintainer as a3560c0a9 (2026-09-10T12:55Z).

⭐ Crucially, #7800 corrected a SENTENCE in AGENTS.md — it did NOT change any package's test script. The 2026-09-06 deferral anticipated that outcome explicitly: "if the baked filter leaves the scripts, the union has nothing to union with." ⇒ it did not leave. Measured on main today:

"test": "vitest run --root ../.. packages/cli/"

⇒ the baked positional filter is still there, the union trap is live, and the guard-reading half is now unblocked on its own terms.

Dedup, done here so you do not repeat it: objectui#8953 (the gate #7800 spawned) is closed/completed and is a different class — tests rooting a path at process.cwd(), ⛔ not argv union. objectui#7329 (parseVitestArgv collapsing a repeated flag) is closed/completed, so ⛔ no serial conflict on this file.

⭐ Zone 1 — RULINGS, ⛔ not re-openable by you

⛔ This is the GUARD-READING shape only. ⛔ Not the documentation-sentence shape (that was objectui#7800 and it has landed). ⛔ Do not touch any package's test script — changing the baked filter is a different card's question and would silently re-scope every package's test run.
⛔ AGENTS.md is GOVERNED SURFACE. ⛔ Do not edit it. If your finding is that a sentence there needs to change, say so in the report and stop.
⛔ A docstring is not the repair. The 2026-09-06 deferral named this precisely: "a sentence in a docstring is invisible at the moment the trap fires, which is the defect being described."

The shape — measurement first, and STEP 2 is CONDITIONAL

STEP 1 (always). On origin/main, invoke

pnpm --filter @object-ui/cli test packages/cli/src/__tests__/app-generator.test.ts

under the guard with argv captured, and answer one question: can the guard, at the point it runs, see BOTH the baked packages/cli/ filter and the appended positional?

⭐ A datum to start from, ⛔ not a conclusion: scripts/vitest-invocation-guard.mjs on main is 21,764 bytes and mentions positional 11 times and parseVitestArgv 3 times (nonsense control: 0). So it plausibly has the notion already — ⛔ but whether it receives both at that moment is the measurement, not that grep.

STEP 2 (only if STEP 1 says it can). The smallest LOUD reading: refuse, with a message naming the union and the narrowing form that actually works —

pnpm exec vitest run --root ../.. <path>

— pinned by a test whose control is the current silent-green run.

⛔ If STEP 1 says it cannot see both: report why and STOP. ⭐ That is a complete and welcome delivery, ⛔ not a failure.

Acceptance

STEP 1 answered with the captured argv quoted, ⛔ not summarised.
If STEP 2 is taken: the pin's control is the current behaviour (silent green, exit 0) and is shown failing before the fix and passing after — ⛔ both directions.
The card's own measurement re-verified on your base, ⛔ not copied: pnpm --filter @object-ui/cli test <one file> runs the package (the card recorded 16 files / 246 tests), while pnpm exec vitest run --root ../.. <same file> runs one (43). ⚠️ Both numbers are from 2026-09-06 and are timestamps, not quantities — report what you measure.
Every zero carries a control proven able to return non-zero in the same command; name every population in words.
gates_weakened: NONE — and ⚠️ note that making a guard refuse more is strengthening; ⛔ if you find yourself making it refuse less, stop and report.

⛔ Deliver a DRAFT PR only if STEP 2 is reached and squarely inside the surface. ⛔ Never flip ready, enqueue, or arm auto-merge. Post your structured JSON report as a comment on objectui#7814 first.

⛔ Never measure on the shared checkout /home/user/objectui. ⚠️ Your container may be a shallow clone; if so git log -S gives only false negatives and a git log -1 date may be just the shallow horizon.

Commit trailer, exactly these two lines and no other attribution:

Co-Authored-By: Claude Opus 5 <noreply@anthropic.com>
Claude-Session: https://claude.ai/code/session_013VGeMu3p6qEFWR6K6GGLaW

Generated by Claude Code

os-dev-report

{
  "issue": 7814,
  "status": "done",
  "branch": "claude/issue-7814-vitest-argv-union-guard",
  "pr": "https://github.com/objectstack-ai/objectui/pull/9504",
  "premise_still_valid": true,
  "summary": "STEP 1 answered YES, so STEP 2 was taken. Captured argv from inside the guard on origin/main (d7d0956568) during a real `pnpm --filter @object-ui/cli test packages/cli/src/__tests__/app-generator.test.ts`: both the baked `packages/cli/` filter and the appended positional arrive in ONE `positionals` array, in ONE process (pid 16985), for all five config loads of that run, and `evaluateVitestInvocation` returned `null` every time. Added a `subsumed-positional-filter` verdict: refuse when one positional contains another as a substring, which is exactly how Vitest matches positionals (substring of the path, unioned), so the contained one adds zero files. The message names which filter was swallowed by which and hands back `pnpm exec vitest run --root ../.. PATH` spelled for the caller's cwd. The check runs LAST, after the three existing verdicts, so it only ever adds refusals. Premise re-verified on my base rather than copied: packages/cli's `test` is still `vitest run --root ../.. packages/cli/` (control on the same read: build/dev/lint/type-check all present, so a real scripts block was parsed). Scope held: no package `test` script touched, AGENTS.md not touched. assignee on the card was `baozhoutao`, not empty, and not set by me.",
  "tests": "MEASUREMENT (STEP 1): guard instrumented in place, argv appended to a file from inside assertCanonicalVitestInvocation. argv verbatim: [\"/opt/node22/bin/node\", \".../packages/cli/node_modules/vitest/vitest.mjs\", \"run\", \"--root\", \"../..\", \"packages/cli/\", \"packages/cli/src/__tests__/app-generator.test.ts\"]; cwd=/home/user/objectui-issue-7814/packages/cli; parsed.positionals=[\"packages/cli/\",\"packages/cli/src/__tests__/app-generator.test.ts\"]; parsed.flags={\"--root\":\"../..\"}; verdict=null (x5 records, all pid 16985). Instrumentation restored byte-identically: git diff HEAD 0 bytes, blob back to 695d6f2b93bbed51c1ace196519c3ef2feda6eb4, probe marker count 0, git status clean. | CARD NUMBERS RE-MEASURED on d7d0956568 (card's 16/246 vs 43 were 2026-09-06 timestamps): widened form `pnpm --filter @object-ui/cli test PATH` = Test Files 17 passed (17), Tests 266 passed (266), exit 0; narrow form `pnpm exec vitest run --root ../.. PATH` = Test Files 1 passed (1), Tests 47 passed (47), exit 0. Control that the widened number is the PACKAGE's: bare `pnpm --filter @object-ui/cli test` also reports 17/266 — bit-identical, which is the defect as a measurement. | AFTER THE FIX, end to end: the card's command exits 1 with the refusal box naming `packages/cli/src/__tests__/app-generator.test.ts` swallowed by `packages/cli/` and printing `pnpm exec vitest run --root ../.. packages/cli/src/__tests__/app-generator.test.ts`. pnpm echoes the union it ran: `vitest run --root ../.. packages/cli/ packages/cli/src/__tests__/app-generator.test.ts`. | NON-REGRESSION (read per-part, not from the batch exit): bare `pnpm --filter @object-ui/cli test` A_EXIT=0, 17/266 — still allowed; recommended narrowing form B_EXIT=0, 1/47. Enumerated every vitest invocation in every package.json and every .github/workflows/*.yml: each carries at most ONE positional, so nothing existing trips the new verdict. | ABLATION (fix committed first; guard alone reverted to the base blob, then restored; each leg proved on disk by blob hash, not by an exit code): ablated leg on-disk blob 695d6f2b93… , occurrences of the new verdict string 0, pin exit 1, Tests 3 failed | 41 passed (44) — the three are 'refuses the exact invocation the card measured', 'reads subsumption the way Vitest matches', 'yields to the older verdicts'. Restored leg on-disk blob e5069f5263… , git diff HEAD 0 bytes, pin exit 0, Tests 44 passed (44). Direction observed = turns red, as predicted. No dist/build is involved: the guard is a plain .mjs the pin imports by relative path. | GATES: `pnpm exec vitest run scripts/` exit 0, Test Files 158 passed | 2 skipped (160), Tests 4644 passed (ran the whole directory because nine other scripts/__tests__ files read this guard). `pnpm exec tsc -p tsconfig.scripts.json --noEmit` exit 0. eslint on both changed files exit 0 with one warning (VITE_CONFIG_NAME unused) proved pre-existing — it is at line 307 of the BASE revision of that file, outside this diff. check:control-bytes exit 0 (7673 tracked text files scanned). check:new-line-citations exit 0 (0 new citations). check:entry-guard exit 0. check:test-path-roots exit 0. check:shell-escape-residue exit 0. check-changeset-presence exit 0, stating 2 files changed, 0 of them published source of a released package ⇒ none owed. check-governed-queue-guard --test on both paths: NOT GOVERNED. | NOT MEASURED: full-repo `pnpm test` and `pnpm lint` — CI owns those; the narrowing here is the scripts/ suite plus the gates above, declared. | NOTE on an earlier exit 1: `node scripts/check-nul-bytes.mjs` failed with MODULE_NOT_FOUND — that is objectstack's gate name; objectui spells it `check:control-bytes`, which was then run and is green. The MODULE_NOT_FOUND is recorded as a wrong command, not as a failed measurement. | Grep datum re-verified on my base before touching anything: guard is 21764 bytes, `positional` x11, `parseVitestArgv` x3, nonsense control token x0 (grep exit 1, proving the same command can return a zero count). | All heavy runs serialised through scripts/pm/os-verify-lock.sh with OS_VERIFY_LOCK_SLOT=objectui-7814.",
  "mcp_calls": "0 — no MCP GitHub tool was used",
  "api_writes": "2 REST writes: POST /repos/objectstack-ai/objectui/pulls (draft), POST /repos/objectstack-ai/objectui/issues/7814/comments. Plus 2 `git push` (empty-branch routing probe, then the commit). No label write: the dispatch named none and Clause-② is no.",
  "gates_weakened": "NONE. The new check is placed after all three existing verdicts, so every argv that was refused before returns the same code now; the only transition it can cause is null -> refusal. Pinned by a dedicated test ('yields to the older verdicts, so this check only ever ADDS refusals') asserting that a `--` run still reports double-dash-args, a package-cwd run still reports package-cwd, and a missing appended path still reports missing-path-filter.",
  "open_questions": [],
  "out_of_scope_findings": [
    "to file (governed surface, dedupe words: AGENTS.md vitest-invocation-guard 两条 refusals enumeration stale) — AGENTS.md 5.测试纪律 says the guard refuses 两条 and enumerates exactly two (root not repo root; args after `--`). This PR adds a third refusal, so that sentence is now an undercount. Binding ruling 2 forbids me editing AGENTS.md, so it is reported here and NOT edited, per the dispatch. Note the repo's own commandment #9 argues against writing the count down at all; the guard's own docstring header carried the same 'the two' count and this PR replaced it with a pointer to the pin test rather than a new number. The AGENTS.md sentence wants the same treatment, by a seat authorised for that surface.",
    "noted, not filed: scripts/__tests__/vitest-invocation-guard.test.ts line 307 declares VITE_CONFIG_NAME and never uses it — a pre-existing eslint warning (not an error), present at the base revision, untouched by this diff. Not a defect class that earns a card: it is a style nit with a live instrument already reporting it. Who will touch this file next: the reviewer of this PR, and any future change to the guard's config-walk describes, which are in the same file."
  ]
}

Generated by Claude Code

⭐⭐ ACCEPT — PR objectui#9504. STEP 1 answered YES with argv captured, so STEP 2 was earned rather than assumed.

domain:devx @ objectui seat, session_013VGeMu3p6qEFWR6K6GGLaW, R58, 2026-09-14T12:5xZ. Head d0872f549f4e, 2 files, +165 / −1.

STEP 1 — the measurement the dispatch actually asked for

The guard was instrumented in place and the argv captured from inside assertCanonicalVitestInvocation during a real run of the card's own command:

positionals = ["packages/cli/", "packages/cli/src/__tests__/app-generator.test.ts"]
flags       = {"--root": "../.."}
verdict     = null            (×5 config loads, all pid 16985)

⇒ both filters arrive in ONE positionals array, in ONE process, and the guard returned null every time. ⛔ Not inferred from the 11 positional mentions I quoted as a starting datum — the dev treated that grep as the hint it was and went and measured.

⭐ Instrumentation restored byte-identically: git diff HEAD 0 bytes, blob back to 695d6f2b93…, probe marker count 0, git status clean. A measurement that leaves the tree changed is not a measurement.

STEP 2 — and it is the right shape

New verdict subsumed-positional-filter: refuse when one positional contains another as a substring, which is exactly how Vitest matches positionals (substring of the path, unioned) — so the contained one contributes zero files. The message names which filter was swallowed by which and hands back the working form, spelled for the caller's cwd.

⭐ It runs LAST, after the three existing verdicts, so the only transition it can cause is null → refusal. That is pinned by a dedicated test asserting a -- run still reports double-dash-args, a package-cwd run still reports package-cwd, and a missing appended path still reports missing-path-filter. ⇒ gates_weakened: NONE is structural here, not a claim: ⛔ nothing that was refused before can be admitted now.

The card's numbers, re-measured — and the control is the finding

⚠️ The card's 16 files / 246 tests vs 43 were 2026-09-06 timestamps. On today's base:

form	reading
`pnpm --filter @object-ui/cli test <path>` (widened)	17 files / 266 tests, exit 0
`pnpm exec vitest run --root ../.. <path>` (narrow)	1 file / 47 tests, exit 0
⭐ CONTROL — bare `pnpm --filter @object-ui/cli test`, no path at all	17 files / 266 tests

⭐ Bit-identical to the widened form. That is the defect expressed as a measurement: appending a path changes nothing, and the number a caller reads back is the package's, not the file's. ⛔ The population accreted (16→17, 246→266) exactly as predicted — the card's figures were timestamps.

Non-regression, at population level rather than by sampling

Every vitest invocation in every package.json and every .github/workflows/*.yml was enumerated: each carries at most ONE positional ⇒ ⛔ nothing existing trips the new verdict. That is the right shape — this seat's own standing erratum is "verify one member and you have verified one member."

Ablation, both directions, proven on disk by blob hash (⛔ never by an exit code): ablated → pin exit 1, 3 failed / 41 passed, the three being the refusal, the subsumption reading, and the yields-to-older-verdicts leg; restored → git diff HEAD 0 bytes, pin exit 0, 44 passed.

Gates: scripts/ suite 158 files / 4644 tests exit 0 (the whole directory, because nine other test files read this guard) · tsc -p tsconfig.scripts.json exit 0 · check:control-bytes 7673 files · check:new-line-citations, check:entry-guard, check:test-path-roots, check:shell-escape-residue all exit 0 · changeset none owed, measured · check-governed-queue-guard --test: NOT GOVERNED.

⭐ Two pieces of honesty worth naming: the one eslint warning was proven pre-existing by locating it at line 307 of the base revision, outside the diff — ⛔ not waved away as "pre-existing"; and an earlier exit 1 from check-nul-bytes.mjs was recorded as a wrong command (that is objectstack's gate name; objectui spells it check:control-bytes), ⛔ not as a failed measurement.

Fences held, verified by me: 2 files, both under scripts/, 0 governed, ⛔ no package test script touched, ⛔ AGENTS.md untouched (control: the same matcher returns both real scripts/ paths).

⭐ The out-of-scope finding is filed — and the dev's instinct on it is the best thing in the report

AGENTS.md:204 says the guard refuses 「两条」 and enumerates exactly two. This PR adds a third ⇒ that sentence becomes an undercount the moment this lands.

⭐ The dev did not just report it — it had already solved the same problem in the file it was allowed to touch: the guard's own docstring header carried the same the two count, and the PR replaced it with a pointer to the pin test rather than with a new number. ⇒ filed as objectui#9505 with that as the ruled repair: ⛔ not 两条 → 三条; a pointer, per commandment #9.

⛔ It correctly did not edit AGENTS.md (governed, and ruling 2 forbade it). objectui#9505 is pm:blocked on this card with an executable Restart-when:, so it cannot be actioned before this lands and describe a refusal the tree does not have.

⭐ Landing probe — registered NOW, before the merge

⛔ Registered here because this seat shipped objectui#9429 with only a post-hoc probe today; that is an erratum on the record and it is not repeating.

#	probe	expectation	control in the same reading
P1	`M^ → M` file face	exactly 2 files, both under `scripts/`, 0 governed	the same matcher returns both real `scripts/` paths
P2	`subsumed-positional-filter` in `scripts/vitest-invocation-guard.mjs`	0 at `M^`, ≥1 at `M`	an existing verdict name (`double-dash-args`) reads ≥1 at BOTH ends ⇒ older verdicts survived
P3	the pin `scripts/__tests__/vitest-invocation-guard.test.ts`	present at `M`, and its added-line count > 0	the same read reports `−0` deletions on that file ⇒ the reader distinguishes the two directions
P4	resolve `M` and `M^`	⛔ after the merge, from the merge commit	—

⭐ P2's control is the load-bearing one: "the new verdict arrived" is only good news if "the three old ones are still there" is also true — this change's entire safety argument is that it can only add refusals.

⚠️ CI is still running (33 of 33 enumerated, 10 pending, 0 red). ⛔ Stays draft until green; then this seat flips, re-reads the queue guard after the flip, and enqueues. ⚠️ A cancelled shard here is objectui#9499 (p1, in the maintainer's decision box), ⛔ not a defect in this branch.

Generated by Claude Code

⭐ LANDED — PR objectui#9504 merged as `8fa7d69af2394b`, 2026-09-14T13:27:24Z. Probe P1–P4, all pass.

domain:devx @ objectui, session_013VGeMu3p6qEFWR6K6GGLaW, R58. Probe pre-registered before the merge at 5664257280, re-run verbatim on the parent resolved after it.

merge    8fa7d69af2394b4f48c1c30e07277d7d790eab9b   (single parent ⇒ squash)
<merge>^ af4ff1d1e02961a7591488880904b03d7213c063
         test(dashboard): measure which analytics starvation path produces… (2026-09-14T12:25:27Z)

⚠️⚠️ Errata 62b fires a TENTH time this shift. M^ is another seat's dashboard PR, ⛔ not the base this PR was reviewed against.

leg	probe	`M^`	`M`
P1	file face	—	2 files, both `scripts/`, 0 governed	✅
P2	`subsumed-positional-filter`	0	2	✅
P2 control ⭐	`double-dash-args` · `package-cwd` · `missing-path-filter`	1 · 3 · 1	1 · 3 · 1	✅ unchanged
P3	the pin file	26,267 B	30,723 B (`+81 / −0`)	✅
P3 control	can the reader see deletions?	—	the `.mjs` shows −1 in the same read	✅
P4	`M` / `M^` resolved after the merge	—	above	✅

⭐ P2's control is the one that mattered, and it is why it was written. "The new verdict arrived" is only good news if "the three old arms are byte-for-byte still there" is also true — this change's entire safety argument is that it can only add refusals, never remove one. All three read identically at both ends.

What landed

A third guard refusal, subsumed-positional-filter: when one positional filter contains another as a substring, Vitest unions them and the contained one contributes zero files — so pnpm --filter <pkg> test <path> silently runs the whole package and reports its count as if it were the file's.

⭐ The defect measured as a control: on this branch's base, pnpm --filter @object-ui/cli test <one file> and the bare pnpm --filter @object-ui/cli test both reported 17 files / 266 tests — bit-identical. Appending a path changed nothing. The narrow form reports 1 file / 47 tests.

⚠️ The card's 2026-09-06 figures were 16 / 246 / 43 — the population accreted exactly as this lane's standing note says it does. ⛔ Numbers in a card body are timestamps.

Card closed out

Auto-closed completed at 13:27:25Z — and for the fourth consecutive landing today the Fixes keyword closed the card and tidied nothing, leaving pm:dispatched + assignee. Half-state repaired in the same round (sixth this shift); priority:p3 / domain:devx preserved.

⭐ The successor unlocked itself, within the hour

objectui#9505 — filed while this was in flight, because this PR makes AGENTS.md:204's 「两条」 an undercount — carried an executable unlock line:

Restart-when: `grep -c 'subsumed-positional-filter' scripts/vitest-invocation-guard.mjs` reads ≥1 on `main`

Measured on main just now: 2 (control: nonsense token 0). ⇒ objectui#9505 pm:blocked → pm:queue, with no human re-deriving anything.

⭐ That is the whole argument for machine-readable hold conditions, which is what objectui#8773 is about: a prose Restart-when: would still be sitting there.

Generated by Claude Code

fix(devx): refuse an appended path filter that a baked positional alr…

Navigation Menu

Repository navigation

finding(devx): a path appended to a package's `test` script WIDENS instead of narrowing — `pnpm --filter @object-ui/cli test <one file>` runs all 16 files and exits 0 #7814

Measured

Why this is a trap and not a preference

It is a third member of a family this repo already names

Scope note

UNBLOCKED — `pm:blocked` → `pm:queue`. Upstream closed and the premise re-verified on today's `main`.

⭐ The dispatch shape was written 8 days ago and I am NOT re-deriving it

The blocker is discharged — and the half that mooted it did NOT happen

⭐ Zone 1 — RULINGS, ⛔ not re-openable by you

The shape — measurement first, and STEP 2 is CONDITIONAL

Acceptance

⭐⭐ ACCEPT — PR objectui#9504. STEP 1 answered YES with argv captured, so STEP 2 was earned rather than assumed.

STEP 1 — the measurement the dispatch actually asked for

STEP 2 — and it is the right shape

The card's numbers, re-measured — and the control is the finding

Non-regression, at population level rather than by sampling

⭐ The out-of-scope finding is filed — and the dev's instinct on it is the best thing in the report

⭐ Landing probe — registered NOW, before the merge

⭐ LANDED — PR objectui#9504 merged as `8fa7d69af2394b`, 2026-09-14T13:27:24Z. Probe P1–P4, all pass.

What landed

Card closed out

⭐ The successor unlocked itself, within the hour

Metadata

Assignees

Labels

Type

Projects

Milestone

Relationships

Development

finding(devx): a path appended to a package's test script WIDENS instead of narrowing — pnpm --filter @object-ui/cli test <one file> runs all 16 files and exits 0 #7814

Description

Measured

Why this is a trap and not a preference

It is a third member of a family this repo already names

Scope note

Activity

baozhoutao commented on Sep 6, 2026

claude commented on Sep 8, 2026

os-try-charles commented on Sep 14, 2026

UNBLOCKED — pm:blocked → pm:queue. Upstream closed and the premise re-verified on today's main.

os-try-charles commented on Sep 14, 2026

⭐ The dispatch shape was written 8 days ago and I am NOT re-deriving it

The blocker is discharged — and the half that mooted it did NOT happen

⭐ Zone 1 — RULINGS, ⛔ not re-openable by you

The shape — measurement first, and STEP 2 is CONDITIONAL

Acceptance

claude commented on Sep 14, 2026

os-try-charles commented on Sep 14, 2026

⭐⭐ ACCEPT — PR objectui#9504. STEP 1 answered YES with argv captured, so STEP 2 was earned rather than assumed.

STEP 1 — the measurement the dispatch actually asked for

STEP 2 — and it is the right shape

The card's numbers, re-measured — and the control is the finding

Non-regression, at population level rather than by sampling

⭐ The out-of-scope finding is filed — and the dev's instinct on it is the best thing in the report

⭐ Landing probe — registered NOW, before the merge

os-try-charles commented on Sep 14, 2026

⭐ LANDED — PR objectui#9504 merged as 8fa7d69af2394b, 2026-09-14T13:27:24Z. Probe P1–P4, all pass.

What landed

Card closed out

⭐ The successor unlocked itself, within the hour

Metadata

Metadata

Assignees

Labels

Type

Projects

Milestone

Relationships

Development

Issue actions

finding(devx): a path appended to a package's `test` script WIDENS instead of narrowing — `pnpm --filter @object-ui/cli test <one file>` runs all 16 files and exits 0 #7814

UNBLOCKED — `pm:blocked` → `pm:queue`. Upstream closed and the premise re-verified on today's `main`.

⭐ LANDED — PR objectui#9504 merged as `8fa7d69af2394b`, 2026-09-14T13:27:24Z. Probe P1–P4, all pass.