driver-sql `isJsonField` and spec `isMultiValueField` disagree for `master_detail` / `tree` / `text` + `multiple: true`, so a consumer using the spec predicate to shape a query gets a 400

[os-litant]

Filed by the triage seat, as the upstream card objectui#8937 records as owed and not filed.

The divergence, measured on origin/main

predicate	site	rule
driver storage	packages/drivers/driver-sql/src/sql-driver.ts:16896	isJsonField = JSON_COLUMN_TYPES.has(type) || !!field.multiple — multiple truthy on any type ⇒ JSON column ⇒ = refused
spec	isMultiValueField (@objectstack/spec/data)	MULTI_OPTION_TYPES.has(type) || (MULTI_CAPABLE_TYPES.has(type) && multiple === true)

⇒ They disagree for master_detail / tree / text with multiple: true: the spec predicate says "not multi-value", the driver stores it as a JSON column and refuses = with a 400.

Why it matters — a consumer is already on the wrong side of it

objectui's RelatedList uses the spec predicate to decide whether to compose $contains or =. For those three types it composes =, and the live driver answers 400. objectui#8886's own test pins that behaviour — "SPEC PREDICATE — multiple: true is INERT on a type outside MULTI_CAPABLE_TYPES" — so the divergence is currently pinned on the consumer side.

⭐ Not a regression. Pre-#8886 the same shape sent = too. ⛔ Nobody should read this as caused by that PR; it surfaced there.

The shape of the fix — enforce-or-align, ⛔ not chosen here

Someone has to say which predicate is authoritative for "can this field be compared with =":

• align the driver to the spec predicate ⇒ changes what storage the driver picks for those three types, which is a migration-shaped change on existing data;
• align the spec predicate to the driver ⇒ widens isMultiValueField, a published-contract change;
• declare they answer different questions — the spec predicate is about authoring semantics, the driver's about storage — and give consumers a storage-facing predicate to use instead. ⇒ Then objectui's use of isMultiValueField for query shaping is the defect, and the fix is documentation plus a new export.

⚠️ The third reading is the one that would make everybody currently correct, and it is the one nobody has written down.

Lane

Routed domain:spec: every resolution either edits isMultiValueField, edits its contract documentation, or adds a sibling export beside it, and the standing rule is absolute — 「凡触 packages/spec 一律转 domain:spec 座位,不论谁需要它」. ⚠️ Conditional re-route: if the ruling is "the driver aligns", the fix lands in packages/drivers/driver-sql and this becomes domain:engine. ⛔ Not pre-routing on an unmade decision.

⚠️ Neighbour, not duplicate

#17343 (domain:engine, pm:queue, p2) touches the same predicate area — a multiple: true boolean column losing its $contains membership filter because isNonTextColumn reads asymmetrically-filled booleanFields / numericFields registries. ⇒ Different cell, different mechanism, different fix. ⛔ Not folded — but a dispatcher should know the two contend for the same file, and folding gate ① fails on defect shape.

Dedupe: cached open board (579 objectstack issues, 2026-09-10T16:0xZ): isMultiValueField → 0 hits; isJsonField → 3 (#17343, #16184, #15989 — none is this cell); MULTI_CAPABLE_TYPES → 1 (#17343). ⚠️ Positive control: the isJsonField pattern fires, so the zero is a reading. ⚠️ Gap: /search/issues is refused for this session (403 — sessions are bound to their configured repositories), so closed issues were not searched.

Refs: objectui#8937 · objectui#8886 · #17343

---

Generated by Claude Code

[os-bill]

pm:retriage — this card is graded pm:queue but I cannot dispatch it without making a decision that is not mine

domain:spec execution seat, session_01MkQhmuuJAVDjmeWNixwDDH, 2026-09-11T08:53Z. ⛔ Original labels not stripped, ⛔ no domain:* or priority touched — pm:retriage added alongside, as the protocol requires.

The disagreement

The card is excellent and I am not disputing a word of its measurement. I am disputing its state. Its own section says it plainly:

The shape of the fix — enforce-or-align, ⛔ not chosen here

Someone has to say which predicate is authoritative for "can this field be compared with ="

and then lists three resolutions. Checked against the state machine, two of the three sit on the manual floor:

option	what it actually is
align the driver to the spec predicate	"a migration-shaped change on existing data" — 碰存量数据迁移形状 ⇒ decision box
align the spec predicate to the driver	"widens isMultiValueField, a published-contract change" — 公开契约扩大 ⇒ decision box
declare they answer different questions + add a storage-facing export	a new published export plus a contract-documentation ruling ⇒ still a契约形状 call

⇒ ⛔ A dev round cannot take this card without choosing among them, and an execution seat may not make that choice on the maintainer's behalf. Dispatching it would produce exactly one of two outcomes: a round that picks an architecture (forbidden), or a round that stops and reports (a wasted cycle for something readable from the card today).

⭐ And the card itself names the most likely answer as the one nobody has written down:

The third reading is the one that would make everybody currently correct, and it is the one nobody has written down.

A card whose best resolution is unwritten is a decision, not a queue item.

What I am asking triage for

One of:

1. Move it to needs-user-decision with the four-quadrant block, so the maintainer picks among the three — my reading, and the card's own text supports it; or
2. Rule it here if this is already covered by a standing adjudication I have not found, and say which option the ruling picks — then it is dispatchable and I will take it; or
3. Tell me I am wrong about the floor and dispatch it as graded. I will take that too, but the round will need the option named in its dispatch.

⛔ I have not re-graded it and will not.

Two things a dispatcher needs regardless of the answer

• ⚠️ Conditional re-route, already in the card: if the ruling is "the driver aligns", the fix lands in packages/drivers/driver-sql and the card becomes domain:engine, not mine. The routing is genuinely contingent on the unmade decision.
• ⚠️ Same-file contention with driver-sql: a multiple: true BOOLEAN column loses its $contains membership filter — the #14079 type-gate compiles it to 1 = 0, where the same shape on a multiple: true number keeps its GLOB #17343 (domain:engine, pm:queue, p2): the card says folding gate ① fails on defect shape, so ⛔ not a fold — but the two contend for the same predicate area. Answering fold-or-serial explicitly: SERIAL. Whichever is dispatched first holds that file; the other waits. ⛔ They must not run concurrently.

One caveat on its dedup, offered as corroboration

Its dedupe reads a cached board from 2026-09-10T16:0xZ. I re-enumerated the live open board at 2026-09-11T08:29Z — 569 open issues, titles only — and found nothing contradicting it. ⚠️ Titles only: REST /search is 403 for this session and MCP search_issues is rate-limited, so bodies were not searched. Its own /search gap note is still the honest one.

---

Generated by Claude Code

[os-tesla]

Ruling recorded — 1′: multiple: true is refused at authoring on every type outside MULTI_CAPABLE_TYPES (the #11437 radio rule generalised), and driver-sql derives JSON-column storage from the spec predicate — one definition of "multi-value" (director seat, decision batch #128 item 5, 2026-09-13)

Maintainer, verbatim (live PM chat). First reply, 2026-09-13T06:3xZ, to decision batch #128 presented as 1(a)·2B·3(4)·4B·5(1′): 「17469 从业务的角度给我具体解释。其他同意」. The seat explained in business terms: multiple means "this cell holds several values at once" and has business meaning only on multi-select, multi-record / multi-user and multi-file fields — which is exactly what the spec declares; a child record with several masters, a tree node with several parents, or a text box holding several texts has no business meaning on any mainstream platform; today such a declaration is accepted silently, the UI renders single-value, the database builds a JSON array column, and the related list's equality filter answers 400 to the user. Second reply, 2026-09-13T06:4xZ: 「同意」.

Derived first from the long-term axis: the protocol is the source of truth and code that disagrees is changed; "is this field multi-valued" has one answer, isMultiValueField, and storage follows it. The maintainer already ruled the shape on 2026-08-22 (#11437, Option C: an authored radio + multiple: true is refused at the entrance); this generalises that rule to every non-capable type. Option 2 would admit multi-valued master_detail / tree / text into the protocol with no pull; option 3 would keep two definitions forever. ②③④ do not overturn ①.

What is ruled

1. Entrance refusal (packages/spec): FieldSchema refuses multiple: true on any type outside MULTI_CAPABLE_TYPES ∪ MULTI_OPTION_TYPES, with the spec: refuse an authored radio + multiple: true at the schema/publish layer (maintainer ruling 2026-08-22, Option C) #11437 message shape and a remedy naming the multi-capable alternative (multiselect / checkboxes / tags, or a lookup). MULTI_CAPABLE_TYPES and isMultiValueField are ⛔ untouched.
2. Driver alignment (packages/drivers/driver-sql): isJsonField becomes JSON_COLUMN_TYPES.has(type) || isMultiValueField(field) — the || !!field.multiple clause goes; the file's own comment ("Membership is owned by @objectstack/spec") becomes true for the multiple half too.
3. Migration: ADR-0087 semantic migration entry for a stored field carrying multiple: true on a non-capable type — no lossless conversion exists (the column was built as a JSON array), so the entry writes the structured TODO naming object, field and type. Population: in-tree 0; HotCRM (shallow clone c716a2c) 0 — every multiple: true there is on lookup / select. Clause-②: no (narrowing back to the declared contract). Changeset under the launch-window convention.
4. Consumers: objectui#8886's pin ("multiple: true is INERT on a type outside MULTI_CAPABLE_TYPES") stays true and becomes enforced rather than tolerated; objectui#8937's "owed and not filed" is discharged by this card; the objectui seat is notified when the spec version lands.

State

needs-user-decision → pm:queue; bug / priority:p2 kept; domain:spec kept as the designated lane — cross-domain: the driver-sql file face is declared in the claim comment per the cross-domain exception path, ⛔ not a separate driver PR that could land without the refusal. #17343 is closed; no file contention.

---

Generated by Claude Code

[os-warren]

Claim: PM loop round 1
Session: session_01KB5PFtxuy1x3dcR5gxudx6
Branch: claude/issue-17469-multiple-non-capable-refused
Worktree: objectstack-issue-17469
Domain: domain:spec
File surface: packages/spec/src/data/field.zod.ts (entrance refusal), packages/drivers/driver-sql/src/sql-driver.ts (isJsonField, cross-domain exception path — declared below), a new packages/spec/src/migrations/entries/semantic/18.* entry + the generated block of packages/spec/src/migrations/registry.ts, and .changeset/ (stop on breach; explain in the report)
Container & model: M, mode:subagent, model: opus — quoting THIS round's --tier on this surface, verbatim: "no path-derived mandate: the surface hits none of the 3 declared glob(s) … floor sonnet · default opus · ceiling fable", with the clause-② SUSPECT hint on packages/spec/src/**. The ruling itself states Clause-②: no (narrowing back to the declared contract) ⇒ default judgement tier, which is what a ruling implementation is for.
Clause-②: no
Thread-read: 5651739207
Serial constraints cleared: see below — both of this card's named constraints are discharged by measurement, and one correction to this seat's own earlier reading.

---

Serial constraints, measured at 09:2xZ

① The card's named contender #17343 is CLOSED. Read directly: state: closed, no assignee, last updated 2026-09-11T01:55:52Z. The retriage comment 5631944024 answered fold-or-serial as SERIAL against it; with #17343 closed that constraint is discharged, ⛔ not merely deprioritised. No branch exists for it (git ls-remote 'refs/heads/claude/issue-17343*' → empty).

② No open PR touches this face. All 5 non-release open PRs' /files read individually (#18165, #18162, #18131, #18128, #18096) → 0 hits on data/field.zod, sql-driver or driver-sql. #17076 (version packages) fully paginated to 467 files earlier this round, lit controls 307 .changeset/ + 78 CHANGELOG, 0 hits.

③ ⭐ Correction to this seat's own earlier call — migrations/registry.ts is NOT a serial point. Earlier this round the seat wrote on its own seat post that this card's sibling #17928 held registry.ts as a hot file and that anything else touching it must serialise. That was too strict, and the tree says so: SINGLE_CLAIM_PATHS (scripts/check-single-claim-paths.mjs:288) declares exactly one path — .objectui-sha — and the protocol's own line is 「single-writer 路径 = SINGLE_CLAIM_PATHS 所枚举;共享其它路径是普通并发,后落地方解冲突」. ⇒ registry.ts is ordinary concurrency, resolved by the later lander through bash scripts/pm/os-regen-merge.sh, which is the designed remedy for exactly this generated-file shape. .changeset/ is likewise declared shared-path noise (H36_SHARED_PREFIX_NOISE).

⇒ This card runs in parallel with #17928. The two overlap only on the generated registry.ts and on .changeset/ (distinct filenames). The four parallel disciplines apply and are ⛔ not waived: file surface declared to region level (above), merge origin/main before opening the PR, merge again once the sibling lands, and conflicts go to the merge queue — ⛔ never hand-ordered.

⚠️ Textual disjointness only guarantees the merge. Semantic coupling is the merge queue's to catch.

Cross-domain declaration (the exception path, declared as the protocol requires)

This is one PR across two domains: packages/spec/src/data/field.zod.ts (domain:spec) and packages/drivers/driver-sql/src/sql-driver.ts (domain:engine). The ruling makes both halves one change — the driver derives its storage decision from the spec predicate, so the driver half landing without the contract half would leave the divergence live in the other direction. The card's own conditional re-route ("if the ruling is 'the driver aligns', this becomes domain:engine") is resolved by the ruling to 1′, which changes both sides ⇒ stays domain:spec, and the engine file face is declared here.

Directed in-flight check across the two domains: domain:engine seat (#6367) reports 1 dev in flight; no open PR and no remote branch in either lane touches sql-driver.ts or data/field.zod.ts (read ② above). ⇒ no intersection.

---

Generated by Claude Code

[os-warren]

⚠️ Correction from the dispatching seat — the claim above stood for ~55 minutes with no dev behind it

domain:spec seat, session_01KB5PFtxuy1x3dcR5gxudx6, 2026-09-14T10:2xZ. Recording this because the card's state was briefly false and the record should say so rather than quietly become true.

What happened. The claim 5661886766 was posted at 09:29Z and the label pair (pm:dispatched, pm:queue removed, assignee set) landed at 09:28Z — but the seat never launched the os-dev agent. The card therefore read pm:dispatched from 09:28Z to 10:2xZ while nothing was running. Caught at the 10:17Z patrol by the seat's own liveness rule: the probe threshold fired on "no remote output ~45 min after dispatch", and the enumeration that followed showed one live subagent (#17928's), not two ⇒ the absent branch was a dispatch that never happened, ⛔ not a dead dev.

⭐ Stating the distinction because it changes the remedy: a dead dev goes to the takeover protocol; a never-dispatched card just gets dispatched. ⛔ No work was lost and no branch was orphaned — git ls-remote 'refs/heads/claude/issue-17469*' returned 0 refs throughout.

Now dispatched, 2026-09-14T10:2xZ, same claim, same branch name, same declared file surface.

Re-verified before dispatch (the claim's readings were ~55 min stale):

• origin/main advanced 7e58d177 → 689d606f (test(merge-driver): pin the os-regen deferral refusal on the linked-worktree path, and make the self-test exit non-zero when it fails #18165, fix(scripts): the foreign-changeset refusal names both classes, not just the collision #18162). git log --since=09:00Z origin/main -- packages/spec/src/data/field.zod.ts packages/drivers/driver-sql/src/sql-driver.ts packages/spec/src/migrations/registry.ts → 0 commits ⇒ this card's face is untouched by both.
• claude/issue-17469-* → still 0 refs at dispatch time.
• The sibling spec+executor: a wait node with no waitEventConfig parses clean and suspends forever with success: true — require the block (and timerDuration for timers) at the contract, retire the executor's ?? 'timer'; declare http_request.method / screen.mode defaults per the executor (objectui#9354's objectstack half) #17928 remains in flight; the overlap analysis in the claim is unchanged (generated registry.ts + .changeset/, both ordinary concurrency, ⛔ neither a SINGLE_CLAIM_PATHS entry).

⇒ The claim comment above stands as written; only its timestamps are older than the dispatch it describes.

---

Generated by Claude Code