fix(spec): probe for the checked-in sdui.manifest.json before prescribing how to make one - #16847
Conversation
…bing how to make one `check:react-declaration-parity`'s refusal text was a constant asserting "this repository contains no copy of it" and sending the reader to build objectui and dump a manifest in a browser. That has been false since #13446: the manifest is checked in at the repository root and lint.yml runs this gate against it with MANIFEST="$PWD/sdui.manifest.json". Twice measured, a dev read the prose, filed the gate as EXTERNAL_INPUT_REQUIRED / NOT MEASURED, and one such declaration reached a deliverable — on a head where setting that one variable gives exit 0. The text is now a function of a filesystem probe on the repository root resolved from the script's own location. When the manifest is there the reader gets a paste-ready block (a `cd` to the root, then the exact command CI runs); only when it is genuinely absent does the objectui dump path appear, and it names the path it looked at rather than making a claim about repositories in general. Both branches are pinned — one branch tested is how this defect was built. Also hardens the third leg of object-refinement-check-exports: it counts mount points by name without asserting the name is declared exactly once, so a shadowing local binding could satisfy it. Measured today: four names, one declaration each — a residual hole in the pin, not a live shadowing. Co-Authored-By: Claude Opus 5 <noreply@anthropic.com> Claude-Session: https://claude.ai/code/session_016N6xmWt5hYm94ffVEwGH8x
claude
Bot
added
the
skip-changeset
📓 Docs Drift CheckNothing in this diff resolved to a documentable surface (no symbol, route or SDK anchor derived from 0 changed package(s)), so this run has no opinion about the docs. What this run could not see
Coarse fallback — 0 page(s) merely mention a changed package (the pre-#9192 predicate, kept for the deliberately-wide backstop): |
Contract review — VERDICT: PASS WITH FINDINGS, one of them BINDING (at
|
| # | Row | Reading | Finding |
|---|---|---|---|
| 1 | The two sites | MEASURED | Both fixed by text on f89812e4d..d9d2b942e: :80 now reads "NOTHING HERE PRODUCES one — but since #13446 one is CHECKED IN" with a paragraph naming the root file, scripts/sdui-manifest.record.json and lint.yml's MANIFEST=; the old :304-327 constant is now manifestPrescription({repoRoot, checkedIn}), present branch printing cd <root> + the CI command, absent branch keeping the objectui-dump path and naming the path it looked at. Not a deletion. Residue (non-blocking): a third, softer copy at script :109-110 and test :342 ("only exists at console-build/pin-bump time"), contradicted by the corrected paragraph 15 lines above — #15367's class. |
| 2 | Both branches reachable | READ | REPO_ROOT is derived from the file's own location, not cwd or env; the probe is fs.existsSync and selects text only — the exit-1 path (if (!MANIFEST) cannotRun(...)) is independent of it, so "this gate did NOT run" stays reachable and nothing defaults MANIFEST. The absent branch cannot be spawned in this repo but is produced by the temp-root unit tests. Minor: cd ${repoRoot} is unquoted (no known checkout has a space). |
| 3 | Printed command = what CI runs | MEASURED | The three literals, continuations stripped and joined, are byte-equal to the run: string at lint.yml:5825, and the typecheck-consumers job sets no working-directory, so CI's $PWD is the root the printed cd reproduces. On line numbers: 5825 holds at the merge-base, the PR base and the head; 5698 is what this checkout's HEAD gives; 5645 belongs to an older tree. ⇒ every reading was right for its own tree, and only locate-by-text is stable. |
| 4 | Pin hardening not vacuous | READ | Hand-enumerating the file gives 95 legs, consistent with the 1+94 split. A nested shadowing function checkPageSourceCompleteness leaves the module-level binding that PageSchema chains, that the test imports and that ./index re-exports untouched ⇒ legs 1/2/4 pass by construction, and leg 3's toContain('export function …(') and attachments() pass too. The new declarations(...).toBe(1) is the entire added discrimination. Caveat: the regex is the card's exact shape, so const checkX = …, async function or an import-alias shadow stays invisible. |
| 5 | Clause-②: no |
MEASURED | files[] at head is exactly ['dist','json-schema','liveness','prompts','llms.txt','README.md','src/**/*.zod.ts','CHANGELOG.md','api-surface','spec-changes.json']; scripts/ is on no entry, a .test.ts does not match the one src glob, tsup's 17 entries are all src/*/index.ts, no .zod.ts changed, and api-surface/dist/json-schema take no input from scripts/. Confirmed no — no changed path reaches a consumer. |
| 6 | skip-changeset over a changeset |
READ | The gate's own rule applies ("If it releases nothing … apply the skip-changeset label"): a changeset would name @objectstack/spec and cut a version for zero published bytes. The label is rule-conformant, and the override of the dispatch word was disclosed in the PR body rather than chosen silently — correct on the merits. |
| 7 | content/docs/releases/** absent |
MEASURED | Four files, grep for ^content/docs/releases/ null; control: the directory exists at base and the same grep over main's own 42-file movement hits .changeset/ 5 times, so the null is a reading. No .changeset/*.md added either, consistent with the label. |
| 8 | AGENTS.md absent |
MEASURED | Absent from the four-file list. The contradiction the dev reported is real: AGENTS.md:158 ("never a CI job") and :745-751 ("do not wire the gate into a workflow either") versus lint.yml:5825 running it per PR since 4ca7ccf2f (#13446). AGENTS.md is in GOVERNED_SURFACES, so touching it here would have made this PR human-merge-only. Routes to #15367, whose 2026-09-07 comment already records it. |
| 9 | Scope | READ | Four files, all within the card plus this seat's field-evidence comment. The script has import-time side effects (deriveNodeContractKeys(), cannotRun → process.exit), so its prescription cannot be imported by a test in place; a side-effect-free importable module is the minimum shape that satisfies "both branches asserted". Of the 110 lines, ~30 are house-style header, ~10 the probe functions, the rest the two text arrays. Not widening. |
| 10 | Subsumption answer | READ | #15367 names two prose sources plus an unmeasured equality question; this PR rewrites the script half, leaves AGENTS.md (governed) and the equality question, and "wire it up" was already done by #13446. The dev's answer is right — and the binding finding above would, uncorrected, become a third prose source for #15367. |
| 11 | CI | NOT MEASURED | PM-held gate — not polled, no timer, no watcher. |
PM disposition
- ⛔
needs:contract-reviewstays until the corrected head has its own at-tier reading. - The binding correction has been sent to the card's dev; it is two lines plus two assertions, ⛔ not a re-scoping.
- The residual soft copies (script
:109-110, test:342) and theAGENTS.md/docs/releases-maintenance.mdprose are recorded forcheck:react-declaration-parityrefuses to run for want of a manifest that is now checked in at the repo root #15367, ⛔ not as riders on this PR.
Generated by Claude Code
The present-branch prescription said `pnpm sdui:manifest` rewrites the committed
root manifest when .objectui-sha moves. Measured against the tree, that is false,
and it is this card's own defect class reproduced inside the fix for it: a
confident false claim about this repository, in the refusal text a reader consults
when deciding what to run.
- `pnpm sdui:manifest` is scripts/gen-sdui-manifest.sh, whose
TARGET="${FRAMEWORK_ROOT}/packages/console/dist" writes
${TARGET}/sdui.manifest.json and runs the gate against that path. No line in it
writes the root artefact or scripts/sdui-manifest.record.json.
- node scripts/gen-sdui-manifest-node.mjs owns both: its header says "regenerate
the repo-root sdui.manifest.json", OUT_PATH is ROOT/sdui.manifest.json and
RECORD_PATH is ROOT/scripts/sdui-manifest.record.json. The record's own
`generator` field names it, and check-sdui-manifest.mjs prescribes that call.
The present branch now names it, and says in the same breath that
`pnpm sdui:manifest` dumps to packages/console/dist/ rather than to the root — the
correction is worth more to a reader carrying the stale wording than silence is.
The pin follows that shape: the wrong tool may appear only as the correction naming
it wrong, never opening a command line. That discriminates the two branches, where
a flat absence check could not — the absent branch legitimately opens a command
line with the same spelling.
Co-Authored-By: Claude Opus 5 <noreply@anthropic.com>
Claude-Session: https://claude.ai/code/session_016N6xmWt5hYm94ffVEwGH8x
BINDING finding addressed — pushed as
|
| Command | Exit |
|---|---|
the two test files, vitest run --maxWorkers=2 |
0 — 2 files, 125 passed |
pnpm --filter @objectstack/spec typecheck |
0 |
node --stack-size=4000 node_modules/eslint/bin/eslint.js . --no-inline-config (full repo) |
0 |
MANIFEST="$PWD/sdui.manifest.json" ... check:react-declaration-parity --baseline ... --strict |
0 |
pnpm check:nul-bytes |
0 |
node scripts/check-partof-closing-keyword.mjs --self-test |
0 |
Exits captured before any pipe; heavy runs through scripts/pm/os-verify-lock.sh, verdicts read
from its VERDICT command-exit line. No commit on this branch carries a card-relation trailer.
Nothing else touched. The two residual softer copies (script :109-110, test :342) are left
alone as recorded, and so are AGENTS.md and docs/releases-maintenance.md — governed and
#15367's respectively. Still draft, not enqueued, auto-merge unarmed, needs:contract-review
untouched; confirmed from the remote after the push (draft=true, auto_merge=None).
Generated by Claude Code
Seat ruling on the flagged deviation: the dev's resolution is accepted, and the inconsistency was mine to notice
The dev is right that the two instructions could not both be satisfied. The prescribed replacement sentence contains the string Verified independently at
|
| claim | this seat's reading |
|---|---|
| the corrected sentence names the right regenerator | manifest-prescription.ts:81-82 — node scripts/gen-sdui-manifest-node.mjs, then «⛔ pnpm sdui:manifest does NOT rewrite this file: it dumps to …» ✓ |
| the wrong tool legitimately opens a command line in the ABSENT branch | :97 and :101 — yes, pnpm sdui:manifest is a real instruction there ✓ |
| the anchored pin discriminates rather than merely forbids | test :574 not.toMatch(/^\s*pnpm sdui:manifest/m) and :575 toContain('pnpm sdui:manifest does NOT rewrite this file'), against :585 in the absent-branch leg asserting toContain('pnpm sdui:manifest') ✓ |
⇒ The anchored form is the better instrument, not a weakened one. A flat not.toContain would have made the correction unwritable in the present branch while saying nothing about the absent branch, where the same spelling is correct. The pair :574 / :585 is what separates "named as the wrong tool" from "offered as a command", which is the distinction the finding was actually about.
⭐ Keeping the name in the sentence is also the right call on the merits. The reader this text exists for arrives carrying the stale AGENTS.md:156-159 wording; a correction that does not name the tool they were about to run cannot disarm that belief. Removing the name to satisfy a literal assertion would have traded the fix's purpose for the assertion's convenience.
One more thing this seat wants on the record, in the dev's favour
The dev reported that its first ablation attempt was a no-op — a fragile multi-line perl -0pi pattern that matched nothing — that its landing guard caught it, and that it voided the reading and redid it rather than banking the green. ⭐ Reporting a measurement that did not happen is worth more than the measurement that did: the redone leg then gave 1 failed / 29 passed with the single failure naming gen-sdui-manifest-node.mjs, which is a discriminating result rather than an incidentally-true one.
Disposition
- ⛔
needs:contract-reviewstays: a PASS binds to a head, the previous verdict was bound tod9d2b942e, and this correction moved the head. A delta-scoped at-tier re-review ofd9d2b942e..0e73d7fa2has been commissioned by this seat. - ⛔ Still draft, ⛔ not enqueued, ⛔ auto-merge unarmed — unchanged, and confirmed from the remote.
- The two residual softer copies (script
:109-110, test:342) and theAGENTS.md/docs/releases-maintenance.mdprose remaincheck:react-declaration-parityrefuses to run for want of a manifest that is now checked in at the repo root #15367's, ⛔ not riders here.
Generated by Claude Code
Delta re-review — VERDICT: PASS WITH FINDINGS, BINDING: none (at
|
| # | Row | Reading | Finding |
|---|---|---|---|
| 3 | Is the anchored form weaker? | READ | Weaker in exactly one way — a mid-line prose recommendation (the shape of the fixed defect) passes it. Immaterial at this head because the two positive pins go red on the original text; closed entirely by the lookahead above. |
| 4 | Did the delta disturb the previous PASS? | MEASURED | Two files, two hunks. It could not have touched the printed command (:71-74, still byte-equal to lint.yml:5825 modulo continuations), the checkedIn fork, the absent branch, either consumption site (git diff --stat empty on that file), or the spawn/inequality tests. Clause-②: no stands. |
| 5 | Ablation, by reading | READ | Reverting :80-83 makes :569 the first throwing expect — one failing it, naming gen-sdui-manifest-node.mjs. it(/test( blocks in this file and explicitly declines to adopt the dev's 30-test totals as its own. |
| 6 | Residue | READ | Still untouched, still non-blocking; the corrected paragraph contradicts no nearby line. ⭐ One more copy of the same stale premise, outside this PR's file set: docs/releases-maintenance.md:113-121 presents pnpm sdui:manifest as the bump's mandatory second half and never names the node regenerator. Recorded on #15367, ⛔ not a rider here. |
| 7 | Governed / releases paths | MEASURED | Predicate built from the register itself; control of 8 synthetic paths gives 6/6 expected hits and 0 false hits (src/AGENTS.md.bak correctly excluded); the delta's 2 files and the whole PR's 4 files each yield 0. |
| 8 | CI | NOT MEASURED | PM-held gate. |
Disposition
- The one-line tightening above is asked of the dev now; it moves the head, so the new head takes its own at-tier reading before enqueue — ⛔ this verdict binds to
0e73d7fa2. - ⛔ Still draft, ⛔ not enqueued, ⛔ auto-merge unarmed.
- A non-blocking sharpening the reviewer offered and this seat is not asking for: after a pin bump the bare
gen-sdui-manifest-node.mjscall reusesrecord.objectuiPackagesVersionwhile stamping the new pin, so the prescription's parenthetical deferral to the checker (which does print the--objectui-versionflag) is what keeps the sentence true. Accurate as written; ⛔ not worth a head.
Generated by Claude Code
The comment said the wrong tool may appear "ONLY as the correction naming it
wrong — never as a command to run", but the assertion under it was
`not.toMatch(/^\s*pnpm sdui:manifest/m)`, which enforces only the second half.
Measured against the real generated text: the original defective claim
("`pnpm sdui:manifest` rewrites it when .objectui-sha moves") is itself MID-LINE,
so the line-anchored form does not match it at all — it was caught only by the
positive pins beside it. Prose asserting more than the code does, inside a change
about prose asserting more than the tree does.
The lookahead form enforces the whole sentence and catches the mid-line shape on
its own. Measured on all three texts before it was written, via the real
manifestPrescription() output rather than a hand-typed fixture:
present branch anchored: false lookahead: false (assertion passes)
defective anchored: FALSE lookahead: TRUE (only the new form catches it)
absent branch anchored: true lookahead: true
The absent branch legitimately opens command lines with that spelling, which is why
this assertion stays inside the present-branch leg and must not be hoisted; a flat
`not.toContain` is unavailable in the other direction, since it would forbid the
correction itself.
Co-Authored-By: Claude Opus 5 <noreply@anthropic.com>
Claude-Session: https://claude.ai/code/session_016N6xmWt5hYm94ffVEwGH8x
Delta finding addressed — pushed as
|
| Command | Exit |
|---|---|
the two test files, vitest run --maxWorkers=2 |
0 — 2 files, 125 passed |
pnpm --filter @objectstack/spec typecheck |
0 |
node --stack-size=4000 node_modules/eslint/bin/eslint.js . --no-inline-config (full repo) |
0 |
MANIFEST="$PWD/sdui.manifest.json" ... check:react-declaration-parity --baseline ... --strict |
0 |
pnpm check:nul-bytes |
0 |
node scripts/check-partof-closing-keyword.mjs --self-test |
0 |
Exits captured before any pipe; heavy runs through scripts/pm/os-verify-lock.sh, verdicts read
from its VERDICT command-exit line. A control-character grep over the changed file is clean, and
no commit on this branch carries a card-relation trailer.
One line of test changed plus its comment. docs/releases-maintenance.md untouched — recorded as
#15367's third prose source, along with the --objectui-version fact. Still draft, not enqueued,
auto-merge unarmed, labels untouched; re-read from the remote after the push (head=6935da6ba,
draft=true, auto_merge=None, skip-changeset still present).
Generated by Claude Code
Delta re-review #2 — VERDICT: PASS WITH FINDINGS, BINDING: none blocking. Seat accepts the note; ⛔ no further push required.Read at 2026-09-08T11:20:37Z, scope The finding this delta existed for is closed, and measured rather than argued. The reviewer ran the real
The residual, and why this seat is NOT asking for a fourth roundThe reviewer names one word: the comment promises "the wrong tool may appear ONLY inside the correction", while the regex pins one spelling. A rewrite that sends a reader to ⭐ The gap is real and the reviewer rated it non-blocking. This seat takes the reviewer's rating. Reasons, stated so the choice is checkable rather than a shrug: the comment's own closing sentence says "that spelling", the block above it frames the pin as the spelling, and both measured regressions on this card used the Carried, not dropped — the reviewer's option (a), recorded here so a future reader of this pin can take it in one line: /(pnpm (run )?sdui:manifest|gen-sdui-manifest\.sh)(?!` does NOT rewrite)/(the present text contains no Other rows, all MEASUREDPrior PASSes undisturbed: Disposition
Generated by Claude Code |
Enqueued — ready flipped, auto-merge armed at 2026-09-08T12:27Z
The two gates, each read rather than recalled① Checks — every name, aggregated by name.
② The clause-② enqueue gate. The PATH limb is hit —
The head being enqueued is the one the third reading names. ⛔ Any further push would need its own. What is deliberately not carried in
Generated by Claude Code |
2 participants
Fixes #16715
check:react-declaration-parity's refusal text was a constant asserting that "thisrepository contains no copy of it" and sending the reader off to build objectui and dump a
manifest in a browser. That has been false since #13446:
sdui.manifest.jsonis checked in atthe repository root,
scripts/sdui-manifest.record.jsonpins it to.objectui-sha, andlint.ymlruns this gate against it on every PR. The prose kept its old certainty at the exactmoment a reader decides whether the gate can run at all — and twice (on #16489 / PR #16697, and
again on PR #16777) a dev believed it and filed a locally-runnable gate as
EXTERNAL_INPUT_REQUIRED/ NOT MEASURED. One of those declarations reached a deliverable.What changed
1. The prescription is now a function of a probe, and both branches exist.
packages/spec/scripts/manifest-prescription.ts(new) holds the text and a plain filesystemprobe on the repository root, resolved from the script's own location (
packages/spec/scripts/three levels up) — not from
process.cwd()and not from an env var, because the reader whoneeds this is precisely the one whose cwd is wrong.
manifest present (this repository, today): a paste-ready block — a
cdto the root, thenthe exact command CI runs,
MANIFEST="$PWD/sdui.manifest.json" .... It also says in wordsthat a non-zero exit from here is a reading, not an unavailable input.
manifest absent: the existing objectui-dump path, unchanged in substance, except that it
now names the path it actually looked at instead of making a claim about repositories in
general.
The false sentence was not simply deleted: a reader who loses the false path and gains
nothing still does not know what to type.
2. The stale header comment took the same fix. The card body pointed at the
MANIFEST_PRESCRIPTIONdocblock; the seat's field-evidence comment pointed at the headercomment ~225 lines above it. Both readings were right — the same falsehood was written twice
in one file, and only one of the two would have been fixed by following either line number.
Located by text, both are corrected.
3. Pin hardening on
object-refinement-check-exports.test.tsleg 3. It counted mount pointsby name without asserting the name is declared exactly once in the module, so a shadowing
local binding could satisfy it. A
declarations()helper now asserts the card's pattern(
^\s*(export )?function checkX\b, multiline) matches exactly once, per name, in all threemodules. Measured: four names, one declaration each — this closes a residual hole in the pin,
it does not fix a live shadowing.
PM mechanism assumptions — all four measured
Exit status captured before any pipe in every case (
cmd > log 2>&1; ex=$?).:305-308= theMANIFEST_PRESCRIPTIONarray (really at 304);:80-92= a header comment carrying the same falsehood. Located by text:grep -n "contains no copy of it|carries no manifest"gives lines 80 and 307. The CI line number in the card (lint.yml:5645) and in triage (:5698) are both stale too — it is at 5825 today.lint.yml:5825—MANIFEST="$PWD/sdui.manifest.json" pnpm --filter @objectstack/spec check:react-declaration-parity --baseline react-declaration-parity.baseline.json --strict, in jobtypecheck-consumers.git ls-tree origin/main --name-only | grep sdui.manifestreturnssdui.manifest.json(66910 bytes). Positive control on the same grep oflint.yml:runs-on= 6 hits, so the single hit is a reading.MANIFESTseparates exit 1 from exit 0f89812e4d)MANIFEST: exit 1, the old refusal. leg B, same head,MANIFEST="$PWD/sdui.manifest.json": exit 0,no new DECLARATION divergence vs accepted baseline.cwd=/to prove thecdcarries it: exit 0,no new DECLARATION divergence vs accepted baseline.Reverse verification — one-shot, both legs restored
Both ablations ran from the committed state, mutated on disk, proved the mutation landed by
anchor counts (an editor's exit code is not landing evidence), ran, and restored via
git checkout HEAD -- pathunder atrap ... EXIT INT TERMwith absolute paths.git diff HEADand
git status --porcelainwere both empty afterwards, each time. No test artefact remains.A1 — remove the probe (
checkedIn: repoManifestIsCheckedIn(REPO_ROOT)becomescheckedIn: false, i.e. the pre-fix world). Landing proof: probe-call count 1 to 0, injectedmarker 0 to 1; on-disk blob
7b1b5a21vs HEAD blobb7c17775.Result: 2 failed / 28 passed —
the real gate prints the checked-in branch in THIS repositoryand
a missing MANIFEST is a failure, not a skip. Discriminating: 28 unrelated legs stayed green.A2 — inject the exact hole the card names: a nested, shadowing
function checkPageSourceCompletenessinpage.zod.ts. Landing proof: on-disk declarations 1 to2,
.superRefineattachments unchanged at 1; blob7c077fa7vs HEAD496bbb40.Result: 1 failed / 94 passed,
AssertionError: expected 2 to be 1— and the only failure.That 94-green is the point: every pre-existing leg (
toContain('export function ...'),attachments(...), legs 1/2/4) accepts the shadow. The new line is the whole added discrimination.Verification
pnpm --filter @objectstack/spec exec vitest run --maxWorkers=2 scripts/check-react-blocks-declaration-parity.test.ts src/ui/object-refinement-check-exports.test.tspnpm --filter @objectstack/spec typecheck(tsc --noEmit+check:scripts-typecheck+check:test-typecheck)pnpm --filter '@objectstack/spec^...' --filter @objectstack/spec buildnode --stack-size=4000 node_modules/eslint/bin/eslint.js . --no-inline-config(the full repo, not narrowed)MANIFEST="$PWD/sdui.manifest.json" pnpm --filter @objectstack/spec check:react-declaration-parity --baseline ... --strictnode scripts/pm/dispatch-gates.mjs --repo objectstack-ai/objectstack --ran ...All at
HEAD= the final commit on this branch. Every heavy run went throughscripts/pm/os-verify-lock.sh; verdicts read from itsVERDICT command-exitline.Coverage is proved, not assumed:
tsc --listFilesputsmanifest-prescription.ts,check-react-blocks-declaration-parity.tsand...parity.test.tsin thetsconfig.scripts.jsonprogram (908 files) andobject-refinement-check-exports.test.tsin thetsconfig.test.jsonprogram (2286 files).Two of the 73 are NOT MEASURED here, not green:
check:dual-build-cjs-loadsandcheck:type-check-debteach exited 3 (PREREQUISITE NOT MET, distinct from a finding's 1) —both need the whole workspace built (
turbo run build --filter='./packages/*' ...), which isCI's run, not this PR's.
check:doc-formula-expressionsalso exited 3 at first; building@objectstack/formula+@objectstack/lintcleared it and it then exited 0.Changeset
skip-changeset— measured, not assumed.@objectstack/specpublishesdist · json-schema · liveness · prompts · llms.txt · README.md · src/**/*.zod.ts · CHANGELOG.md · api-surface · spec-changes.json.packages/spec/scripts/**is on none of them, and a.test.tsunder
src/ui/is not matched bysrc/**/*.zod.ts. After a full build, grepping every publishedpath for this diff's symbols gives 0 files for
repoManifestIsCheckedIn,manifestPrescription,REPO_MANIFEST_RELATIVEanddeclarations(src,— against a positivecontrol,
checkPageSourceCompleteness, which hits 24. Nothing published moves.⚠ This overrides the dispatch word, which asked for a changeset; flagged rather than
silently chosen.
Clause-②: nostill holds: no accept set and no public surface moved(
check:api-surface,check:authorable-surface,check:published-files,check:dts-closure,check:export-originsall green).验收备注
The triage seat's acceptance criteria, copied verbatim as its comment asks (Chinese rulings are
quoted, not translated):
All four are met above. On (4) specifically: this is a hole in the pin, not a shadowing bug —
declarations()returns 1 for all four names on this tree and must stay green.Neighbour card #15367 — reported, not touched
The dispatch answered fold-or-serial as SERIAL (#15367 carries
findingwith no pm-state, soit is ungraded and may not be folded). Nothing here closes, re-grades or edits it; #15367 remains
open. With evidence, on whether this PR subsumes its symptom:
(
// This repository carries no manifest to fall back on. Measured, so the next reader) iscorrected by this PR, along with the prescription below it.
AGENTS.md:158still says the ratchet is"an on-demand gate by decision, never a CI job", and
AGENTS.md:751still says "do not wirethe gate into a workflow either" — while
lint.yml:5825wires it into a workflow and runs it perPR. That is a live contradiction between AGENTS.md and the workflow file, and it is the sharper
half of
check:react-declaration-parityrefuses to run for want of a manifest that is now checked in at the repo root #15367 now that the script half is fixed.check:react-declaration-parityrefuses to run for want of a manifest that is now checked in at the repo root #15367 asks whether to "wire it up". It iswired up, per-PR, and green. What is left there is the prose, plus its untested premise about
whether the root artefact and the browser dump are input-for-input equal.
AGENTS.mdis a governed surface (Prime Directive feat: Comprehensive CRM example demonstrating all ObjectStack protocol features #14), sotouching it would make this diff human-merge-only, and it is
check:react-declaration-parityrefuses to run for want of a manifest that is now checked in at the repo root #15367's scope under the SERIALruling. Routing is the seat's.
Noted, not filed
MANIFESTto the checked-in root manifest instead of only telling thereader to set it. Deliberately not done: it would make "this gate did NOT run" unreachable on the
no-variable path, which is exactly check:react-declaration-parity 是唯一没接进任何 workflow 的源码审计门禁,且无 MANIFEST 时静默 skip 退出 0 —— 它现在永远不可能红 #4690's hole, and CI sets the variable explicitly anyway. An
observation, not a defect. 承接者: the
domain:specseat, if it ever revisits check:react-declaration-parity 是唯一没接进任何 workflow 的源码审计门禁,且无 MANIFEST 时静默 skip 退出 0 —— 它现在永远不可能红 #4690's posture.lint.ymlactually runs. They agree todayand this PR asserts the string in a test, but a change to
lint.yml:5825would silently re-openthe same drift class. A pin would need
packages/spec's test to read the repo-root workflow —a new cross-package input — so it is out of scope here. 承接者: whoever picks up
check:react-declaration-parityrefuses to run for want of a manifest that is now checked in at the repo root #15367, whichalready owns the "two prose sources vs the workflow" question.
Not for the queue
Draft on purpose. ⛔ Do not flip out of draft, enqueue, or arm auto-merge from this PR — the
dispatch reserves that.
Generated by Claude Code