objectql: a stored field's expression is computed on insert before the autonumber is issued, so an expression that reads the record number stores null (or its has() fallback) until a later update by id

Ruled: 6335312761 · letter A · 2026-10-22T07:31Z

Filing gate ① a product defect with reach measured. The defect was measured by the os-dev on objectstack-ai/hotcrm#2230, which executes #24981 ruling 6308445871 (maintainer 「其他同意」): "a composite title that must round-trip through import is a stored computed field". The run was on HotCRM main c8a7a704 with @objectstack/* 18.0.0 (framework 06cab4849), on the @objectstack/verify boot with the seed replay. Filed by the repo:hotcrm seat 2 (os-support-ai, session session_01KqkTQNLcxVc588eZpJmp1r). ⛔ Not a claim. Grade: triage's.

Reader (who acts): objectstack domain:engine triage. The fix lands here. objectstack-ai/hotcrm#2230 carries Blocked-by: this card and resumes on the @objectstack release that carries the fix.

The contract

FieldSchema.expression (packages/spec/src/data/field.zod.ts:1856 at 06cab4849): "On any other (stored) field: computed by the engine from the row the write stores and stored … on every insert and every update by id".

A row the insert stores carries its issued record number. So an expression that reads an autonumber column should store the value with the number in it.

The finding, as measured

Setup: HotCRM's nine record titles switched locally from Field.formula to a stored Field.text, with the same expression (the shape the ruling prescribes). The change was not committed.
On a seeded boot, 53 of 134 stored titles equal the pre-migration formula value:
- the three objects whose title reads no autonumber are all correct (contact 9/9, lead 21/21, forecast 7/7);
- the six whose title reads one are mostly wrong (product 0/13, knowledge_article 0/24, case 1/38, account 4/9, quote 4/6; campaign 7/7).
- The rows that came out right were rewritten by id later in the seed, which recomputes.
Single inserts as the dev admin through REST (POST /api/v1/data/<object>):
- crm_case answers 201 and stores display_title null beside case_number CASE-00039. The expression is record.case_number + " - " + record.subject.
- crm_product stores "Probe 2230 product" without PRD-0014. Its has(record.product_code) ? … : record.name fallback took the empty branch.
- Control: crm_lead (no autonumber in its title) stores "Pro Be - Probe 2230 Inc", correct.
Reach: every object whose stored-expression field reads an autonumber column, in any app, on every insert. The ruling directs HotCRM's nine titles to exactly this shape, and six of them read a record number.

Seat review (source at `06cab4849`, unchanged on `origin/main` `688052e147`)

packages/objectql/src/engine.ts, insert path:

computeStoredExpressions runs at :15807 (before the hooks) and again at :16410 ("after the hooks and the strips").
applyAutonumbers issues the record number afterwards, at :16466. Its own comment places it after validation on purpose ("no validation rule can depend on the value").
Nothing recomputes the stored-expression plan once the number is issued.
When the driver owns autonumber (driverOwnsAutonumber), the number arrives at the write itself, later still.

Direction (the reader's call):

Recompute the stored-expression plan for the fields whose expression reads an issued autonumber, after applyAutonumbers and before the row is written;
or issue the number before the second stored-expression pass.
Either way, the driver-owned path needs an answer too.
A pin: insert into an object whose stored expression reads its autonumber, and read the stored value back.

Dedupe

Method: every objectstack-ai/objectstack issue title (every state), listed through the repo-scoped issues?state=all endpoint and refreshed with since=2026-10-21. Matched case-insensitively on autonumber combined with expression / formula / stored / computed / title / order, and on stored expression / computeStoredExpressions / #24673 / #24981. The search API is refused for this session.

Results:

Related:
- import: a lookup cell that names the record by the target object's title field answers reference_not_found — resolveRef never reads nameField, and a formula title is not matchable #24981 (the ruling that prescribes this shape);
- spec(data): once PR #24673 lands, a client value for a stored-expression field is overwritten with no report — DroppedFieldsEvent.reason has no value for "written, but recomputed", so onFieldsDropped / droppedFields stay silent #24690 and feat(spec,objectql): report a caller value a stored expression replaced, as DroppedFieldsEvent reason stored_expression #24778 (closed: reporting a client value a stored expression replaces);
- objectql 18.x: under strictReadonlyWrites, refuse a write whose value a stored expression replaces (stored_expression), the deferred half of #24690 #25000 (open: refusing such a write under strictReadonlyWrites);
- ADR-0079 D3, the validate-record-title lint and formulas.mdx still steer a composite record title into a formula field, which ruling 6308445871 says cannot be matched on import (use a stored computed field) #25071 (open: the docs and lint that steer a composite title into a formula).
None is about the insert order against autonumber. New.

Generated by Claude Code

Claim: PM loop round 30 (the lane's p1 and p2 too while seat 2 is vacant, by the maintainer's 「engine 车道有任务为什么不派发」) · 2026-10-22T05:07Z · p1, bug, domain:engine. The lane's only open p1 in pm:queue, after #25144 landed (504d8b3acd). It blocks objectstack-ai/hotcrm#2230 (Blocked-by:), which carries out the #24981 ruling 6308445871.
Session: session_01JtSQQ3EWsbNPzyYMug1tXy (the dispatching seat; the dev runs as its own cloud session, named in the dispatch comment)
Account: os-project-manager (the seat's linked user as GET /user answers it; always the card's assignee)
Branch: claude/issue-25073-stored-expression-after-autonumber
Worktree: objectstack-issue-25073
Domain: domain:engine
Seat: domain:engine#1 (seat post #6367)
Premise, re-read at this claim on origin/main 504d8b3acd (the card is a day old): it holds, and one path is wider than the card says.

The insert's two stored-expression passes still run before the number exists. The second pass, "after the hooks and the strips", is computeStoredExpressions inside the insert body. applyAutonumbers issues the engine's number after it, per row.
The production driver issues the number itself. driver-sql advertises supports.autonumber: true, and sqlite-wasm and turso inherit it. So applyAutonumbers returns [], and SqlDriver.create / bulkCreate fill the number through the protected fillAutoNumberFields right before the INSERT. On that path the engine never holds the number before the write. This is the path HotCRM measured.
A third path the card does not name. createWithAutonumberResync deletes and re-issues the engine's number on a collision, after the expression was computed. That is the fallback drivers' path, memory and mongodb, and its stored value would go stale the same way.
File surface:
Measured first, before any edit: how a stored expression that reads an autonumber field gets the issued number on each of the three paths:
- engine-issued (applyAutonumbers), on single and batch inserts and in partial mode;
- the collision re-issue in createWithAutonumberResync;
- driver-issued (driver-sql create / bulkCreate).
The fix, in packages/objectql/src:
- Recompute the stored-expression fields whose expression reads an issued autonumber once the number exists, before the row is written. Keep applyAutonumbers' placement after validation, so a rejected row still burns no number.
- For the driver-issued path, the route must keep the insert one atomic act. The afterInsert hooks, the response and every reader see the final stored value, and no reader can ever see the row without it.
- A fork, needs_decision, before building it: a route that needs a new IDataDriver member, a new DriverOptions field, or a call into a driver's non-contract method.
- ⛔ No instanceof SqlDriver or driver-name branch.
Pins:
- an insert into an object whose stored expression reads its autonumber, read back from the store. Cover the driver-sql path (SQLite) and the engine-issued path, single and batch, plus the collision re-issue case;
- the card's has() fallback shape;
- controls: an expression that reads no autonumber is unchanged; an update by id recomputes as today.
- Ablation: remove the recompute; the autonumber pins go red and the controls stay green.
The changeset: @objectstack/objectql patch (and any package the measured route moves). Clause-②: no (insert order: a stored expression now reads the record number the insert issues), Metadata-protocol: none.

⛔ Not in scope:

packages/spec, including FieldSchema.expression's text (it already says the insert's row);
a backfill of rows already stored with the null title (HotCRM's seed replay rewrites by id; a migration would be another card);
objectql 18.x: under strictReadonlyWrites, refuse a write whose value a stored expression replaces (stored_expression), the deferred half of #24690 #25000 (strictReadonlyWrites), ADR-0079 D3, the validate-record-title lint and formulas.mdx still steer a composite record title into a formula field, which ruling 6308445871 says cannot be matched on import (use a stored computed field) #25071 (docs and lint), and Store the 9 formula record titles as stored computed fields and backfill them, so import can match them and sort and search can use them (objectstack#24981 ruling A) hotcrm#2230 itself;
driver-memory / driver-mongodb beyond what the engine's own fallback path needs (the [裁决] driver-memory / driver-mongodb 投入冻结 —— 维护者 2026-08-05 口径（跨单锚点） #5499 freeze).

Container & model: M, mode:cloud (three insert paths, one of them owned by the driver, plus a fork check), model: TIER_DEFAULT to build (this run's dispatch-gates.mjs --tier --paths packages/objectql/src/engine.ts: "no path-derived mandate"). The review is TIER_DEFAULT, the seat's own review of record; a fork into the driver contract would move it to CONTRACT_REVIEW_TIER.
Clause-②: no (insert order: a stored expression now reads the record number the insert issues)
Responsibility: an author who writes a stored expression that reads an autonumber field | the insert stores null, or the has() fallback, until a later update by id | every app with such a field on every insert; HotCRM's six autonumber-reading titles (objectstack-ai/hotcrm#2230)
Thread-read: none
Serial constraints cleared: the 12 open PRs' file lists were read at this claim, on origin/main 504d8b3acd. None touches packages/objectql/src. #24876 (changeset-release/main, 294 files) touches none.

Generated by Claude Code

Dispatch · domain:engine seat 1 · session_01JtSQQ3EWsbNPzyYMug1tXy · 2026-10-22T05:09Z: mode:cloud. The dev session is session_017BrhF7dagr9BaqR2ssbhUQ. It builds at the default tier on branch claude/issue-25073-stored-expression-after-autonumber, under claim 6333215775.

Read first: how a stored expression that reads an autonumber field gets the issued number on each of three insert paths. Each reading is shown with a probe, covering atomicity, what afterInsert and the response see, hooks fired, collisions, and batch and partial mode:
- the engine-issued path (applyAutonumbers);
- the collision re-issue path (createWithAutonumberResync);
- the driver-issued path (driver-sql, the production path HotCRM measured).
What lands:
- the recompute of the autonumber-reading stored fields once the number exists and before the row is visible, with applyAutonumbers kept after validation;
- pins on SQLite, on the engine-issued path (single and batch) and on a forced collision, plus the has() fallback;
- controls (no autonumber; update by id) and an ablation;
- a patch changeset: Clause-②: no (insert order: a stored expression now reads the record number the insert issues), Metadata-protocol: none.
The fork: a driver-issued route that needs a new IDataDriver member, a new DriverOptions field or a driver's non-contract method is needs_decision, and so is a route that cannot keep the insert one atomic act. The dev reports the measured routes with a four-axis recommendation before building that path.
Not built: packages/spec; a backfill of stored rows; objectql 18.x: under strictReadonlyWrites, refuse a write whose value a stored expression replaces (stored_expression), the deferred half of #24690 #25000; ADR-0079 D3, the validate-record-title lint and formulas.mdx still steer a composite record title into a formula field, which ruling 6308445871 says cannot be matched on import (use a stored computed field) #25071; Store the 9 formula record titles as stored computed fields and backfill them, so import can match them and sort and search can use them (objectstack#24981 ruling A) hotcrm#2230.
Owed: the seat's TIER_DEFAULT review of record before the queue. A fork into the driver contract would move it to CONTRACT_REVIEW_TIER.

Generated by Claude Code

os-dev-report

{
  "issue": 25073,
  "status": "needs_decision",
  "branch": "claude/issue-25073-stored-expression-after-autonumber",
  "pr": null,
  "session": "session_017BrhF7dagr9BaqR2ssbhUQ (dispatched by seat session_01JtSQQ3EWsbNPzyYMug1tXy, claim 6333215775 names this branch)",
  "premise_still_valid": true,
  "summary": "Measured before any edit, on origin/main 504d8b3acd. Probe committed as 7dadfd07 on the branch, at packages/runtime/src/zz-probe-25073.integration.test.ts. No product code was changed and no PR was opened. The defect reproduces on all three paths. P1, engine-issued, InMemoryDriver: single, batch and insertMany partial mode all store display_title null beside CASE-00001..00005, and the product has() shape stores 'Probe product' without PRD-0001. P2, collision re-issue, memory: CASE-00003 is issued after the re-seed and the title is still null. P3, driver-issued, SqlDriver better-sqlite3: the single response and the stored row carry null; a batch stores null too; the product takes the fallback branch. The control plain_title is correct everywhere, and an update by id heals the value ('CASE-00001 - one!'). P6: an afterInsert hook sees null and so does the response. Path 3 has exactly one route inside packages/objectql/src: insert, then a by-id driver.update of the dependent fields, inside a unit the engine opens. P4 measures what that unit costs. With a stale SQL counter (an out-of-band row CASE-00002), an insert outside any unit self-heals to CASE-00003 through SqlDriver.create's own re-seed. The same insert inside engine.transaction answers DUPLICATE_RECORD twice in a row and never converges, because SqlDriver.create sets mayRetry false when options.transaction is set. So the route turns a wrong title into a permanent insert outage for that object whenever its counter falls behind the store. The route also cannot be atomic on driver-turso's remote arm, which advertises autonumber: true and transactionsUnsupported: true. Under the fork rule this is needs_decision. I am not landing paths 1 and 2 alone; the reasons are in open_questions.",
  "tests": "Probe only. Command: bash scripts/pm/os-verify-lock.sh -c 'pnpm --filter @objectstack/runtime exec vitest run --project local --maxWorkers=2 src/zz-probe-25073.integration.test.ts'. VERDICT command-exit 0, Tests 6 passed (6), run on the tree that became 7dadfd07. Readings, verbatim from the run: 'P1 single response {n:CASE-00001,t:null}'; 'P1 partial outcomes [{n:CASE-00004,t:null},dead,{n:CASE-00005,t:null}]'; 'P1 product response {n:PRD-0001,t:Probe product}'; 'P2 response {n:CASE-00003,t:null}'; 'P3 single response {n:CASE-00001,t:null,p:S:one}'; 'P3 product response {n:PRD-0001,t:Probe product}'; 'P3 after update by id {n:CASE-00001,t:CASE-00001 - one!}'; 'P6 afterInsert saw [null] response null'; 'P4 outside-unit insert {ok:true,n:CASE-00003}'; 'P4 inside-unit insert 0 {ok:false,code:DUPLICATE_RECORD}' and the same again for insert 1; 'P5 stored after unit {n:CASE-00001,t:CASE-00001 - proto}', with updated_at 1 ms after created_at. Built first: turbo build of the objectql, driver-sql and runtime dependency closures, Tasks 14/14 and 32/32, VERDICT command-exit 0 both times. No ablation: no tested behaviour changed.",
  "gates": "NOT MEASURED: dispatch-gates --commands families, reason: the branch carries no product diff, only the measurement probe, and no PR is opened on a needs_decision. When the build resumes, the derived families are owed on the real diff.",
  "dogfood_battery": "not owed",
  "line_budget": "none set; branch diff vs 504d8b3acd: packages/runtime/src/zz-probe-25073.integration.test.ts +139/-0 (measurement probe, to be dropped on resume)",
  "files_changed": ["packages/runtime/src/zz-probe-25073.integration.test.ts (measurement probe only; another lane's package, runtime, named for the seat)"],
  "deviations": [
    "The branch carries one measurement commit (7dadfd07) instead of being empty. This is the measurement-branch push the needs_decision budget allows. Whoever resumes deletes the probe file in the first commit.",
    "Turso remote arm: read from code, not probed (turso-driver.ts sets supports.transactionsUnsupported to this.transportMode === 'remote', and the driver inherits autonumber: true). MongoDB: not probed. It takes the same engine fallback code as InMemoryDriver, so P1 and P2 stand for it."
  ],
  "measurements": {
    "path_1_engine_issued": "applyAutonumbers runs after the second computeStoredExpressions pass, so the plan is computed over a row without the number: null on single, batch and partial (P1). Route: recompute the dependent fields right after applyAutonumbers, per live row, before the driver write. Atomic by construction (one create or bulkCreate). afterInsert and the response see the value. No update hooks. Validation stays before issue, so a rejected row burns no number. Partial mode: dead rows are skipped already.",
    "path_2_collision_reissue": "createWithAutonumberResync deletes the burned slot and calls applyAutonumbers again; nothing recomputes. Route: recompute after the re-issue, inside the same loop, before the retried create. One write, atomic. Same costs as path 1.",
    "path_3_driver_issued": "applyAutonumbers returns [] and SqlDriver fills the number inside create or bulkCreate through the protected fillAutoNumberFields. The engine first holds the number in the row the driver returns. Route R1 (objectql only): open a unit when none is ambient (the opensAfterAbortUnit pattern), insert, compute the dependent fields over the sent row plus the returned number, then driver.update or bulkUpdate by id in the same unit, then patch the result rows before formula hydration and afterInsert. Measured costs: (a) P4: inside a unit the driver's collision re-seed is off, so a stale counter means permanent DUPLICATE_RECORD for that object. Objects with an aborting metadata afterInsert hook already carry this cost today; R1 extends it to every autonumber-title object. (b) Turso remote has no unit, so the atomicity bar is unmet there. (c) An object on a non-default datasource cannot get a unit from engine.transaction. (d) One extra UPDATE per row and one transaction per insert; on Postgres the sequence row stays locked until commit. (e) updated_at is restamped 1 ms after created_at (P5). (f) No update hooks or audit events, by construction (a driver-level update). (g) The pinyin __search companion is stamped in beforeInsert from the pre-number title. Cost (g) is shared by every route that recomputes after the hooks, paths 1 and 2 included."
  },
  "mcp_calls": "2 — mcp__github__issue_read (get, get_comments), read-only; no MCP write tool",
  "api_writes": "1 — POST /repos/objectstack-ai/objectstack/issues/25073/comments (this report) through the fleet-write relay; git pushes: 2 to the branch (the empty branch, then the probe commit 7dadfd07)",
  "open_questions": [
    {
      "question": "Path 3 (driver-issued, driver-sql, sqlite-wasm and turso, which is HotCRM's production path): how does a stored expression get the record number the driver issues?",
      "options": [
        "A — contract route: add an optional IDataDriver member that issues a row's autonumbers on request (the public form of fillAutoNumberFields). The engine calls it where applyAutonumbers runs today, after validation, recomputes the dependent fields, and hands create a row whose number is already filled. One INSERT, atomic by construction, no transaction needed, so it also works on Turso remote. The collision retry moves to the engine's existing createWithAutonumberResync, which re-issues through the same member: one retry loop for every driver. Cost: a contract member (CONTRACT_REVIEW_TIER), driver-sql implementation that sqlite-wasm and turso inherit, and a changeset on spec, driver-sql and objectql. Size: M.",
        "B — R1, objectql only: insert, then a same-unit by-id driver update. Measured costs: the driver's collision self-heal is lost (P4, a permanent 409 on a stale counter); Turso remote is not atomic; a non-default datasource gets no unit; one extra UPDATE and one transaction per insert; updated_at drift. Size: S–M.",
        "C — B, plus a fallback when no unit is available or a collision happens inside the unit: the non-atomic two-statement write, warned once. It breaks the dispatch's atomicity bar on purpose.",
        "D — refuse the shape at publish (a stored expression reading an autonumber). This contradicts the #24981 ruling 6308445871, which prescribes exactly this shape for a composite title that must round-trip through import. Not recommended."
      ],
      "recommendation": "A, on the four axes. Business need: the measured producer is HotCRM's six autonumber titles on driver-sql (hotcrm#2230 is Blocked-by this card), and A is the only route that serves every SQL arm, Turso remote included. Long-term soundness: A keeps the driver as the single owner of the counter (the #6664 A ruling cited in createWithAutonumberResync) and gives every driver one ordering: issue, then compute, then write once. B is a workaround that buys a correct title with a rare but permanent insert outage. Preventing AI mistakes: under A the declared contract text ('computed from the row the write stores ... on every insert') is delivered by construction on every driver; under B or C it holds only on some datasources and transports, which is the declared-but-not-delivered shape. Startup scope: A adds one optional member with three in-repo implementers, all of them inheritance of one method; no new gate. If a contract member is ruled out, B is the fallback, and its P4 regression must be accepted in writing."
    },
    {
      "question": "Should paths 1 and 2 (engine-issued: memory and mongodb fallback, plus the collision re-issue) land now without path 3?",
      "options": ["A — no: land them with the path-3 route in one PR", "B — yes: land the engine recompute now as its own PR"],
      "recommendation": "A. No measured producer runs this shape on memory or mongodb, and that investment is frozen (#5499). Landing them alone would make a memory-backed test store the right title while driver-sql production still stores null for the same metadata. A test environment that disagrees with production hides the defect from an AI author, the worst outcome on the third axis. The recompute helper (dependent-field selection through analyzeRelationshipTraversals; the probe confirms it sees record.x, has(record.x), record.?x and record['x']) is exactly the code route A's engine half calls after its issue step, so nothing is lost by landing it once, with A."
    },
    {
      "question": "Client-value reporting under any route: should the replaced-value report (#24690, #24778, strictReadonlyWrites) judge a client-sent value against the final, number-bearing value?",
      "options": ["A — keep the report where it is (before validation, before any number is issued) and report once, judged on the pre-number value", "B — move the judgement after the issue step"],
      "recommendation": "A. The strict refusal must land before a number is burned (applyAutonumbers stays after validation). A client cannot know an unissued number, so judging a field that reads one before the issue loses no real case."
    }
  ],
  "out_of_scope_findings": [
    "class: a · reach: an object with an aborting metadata afterInsert hook (opensAfterAbortUnit opens a unit) on driver-sql, after any out-of-band row lands ahead of the sequence: every insert answers DUPLICATE_RECORD and never converges, because SqlDriver.create skips its #5495 re-seed when options.transaction is set · evidence: probe P4, inside-unit inserts 0 and 1 both DUPLICATE_RECORD, the same setup outside a unit heals to CASE-00003 · dedupe words: autonumber collision transaction mayRetry, DUPLICATE_RECORD sequence stale unit, opensAfterAbortUnit autonumber"
  ]
}

Generated by Claude Code

[待裁] 新建记录时，引用编号的存储字段存成空值：编号由 SQL 驱动在写库前一刻才生成，引擎要怎样拿到它？

domain:engine seat 1（seat post #6367）· os-project-manager · session_01JtSQQ3EWsbNPzyYMug1tXy · 2026-10-22T05:42Z。认领 6333215775 写明：驱动生成编号这条路，如果要动驱动契约（新增 IDataDriver 成员、新增 DriverOptions 字段、调用驱动的非契约方法），或者做不到一次原子写入，就报分叉，不许先做。dev 先实测、没改产品代码，报了 needs_decision（报告 6333405453）。分支上只有一个测量探针提交 7dadfd07，没有开 PR。本席复核了实测，选项落在公开契约的形状上，现有规范定不了，按 SKILL〈升级与决策〉交维护者裁决。认领 6333215775 同笔释放。

一句话问题： 标题写成「工单号 + 主题」的对象，新建工单时标题存成空值，或者只剩主题；要等下一次按 id 修改这条记录才会补上。HotCRM 有六个对象的标题是这种写法。

背景（实测）

dev 在 origin/main 504d8b3acd 上用真实引擎跑了探针（packages/runtime/src/zz-probe-25073.integration.test.ts，6/6 通过，只读测量）。

三条写入路径都复现：
- 引擎发号（内存、MongoDB 驱动）：单条、批量、部分成功模式都存成 null。旁边的编号是 CASE-00001…00005。产品对象的 has() 兜底写法存成「Probe product」，没有 PRD-0001。
- 编号冲突后重发：重新发出 CASE-00003，标题仍是 null。
- 驱动发号（driver-sql，SQLite）：单条的响应和落库的行都是 null，批量也是 null。这是 HotCRM 生产走的路径。afterInsert 钩子和响应看到的也都是 null。
- 对照： 不读编号的标题全部正确；按 id 修改一次即补上（CASE-00001 - one!）。
为什么第三条路引擎拿不到号：
- driver-sql 声明 supports.autonumber: true，driver-sqlite-wasm 和 driver-turso 继承这一声明。
- 编号在 SqlDriver.create / bulkCreate 里，由 protected 方法 fillAutoNumberFields 在 INSERT 前一刻填入。
- 引擎第一次见到这个号，是驱动写完后返回的行。
只改引擎的唯一做法（下文 B）实测有新害：
- 做法是在引擎开的事务里先插入，再按 id 回写标题。
- 事务内 SqlDriver.create 会关掉自己的编号冲突自愈（const mayRetry = options?.transaction === undefined）。
- 实测：计数器落后于库里已有的行时，事务外插入会自愈成 CASE-00003；事务内两次插入都回 DUPLICATE_RECORD，永远不收敛。
- driver-turso 远程模式不支持事务（transactionsUnsupported: this.transportMode === 'remote'），做不到原子。

Governing text

packages/spec/src/data/field.zod.ts（FieldSchema.expression）：「On any other (stored) field: computed by the engine from the row the write stores and stored … on every insert and every update by id」。插入落库的那一行带着它发出的编号，所以契约要求标题里有号。
docs/adr/0034-transactional-writes-and-ambient-transaction.md §3「Autonumber (DB-native autonumber sequence (multi-instance correctness) #1603)」：「have the engine defer autonumber generation to the driver's sequence」。计数器归驱动。
driver-memory census in undeclared-field-write-driver-split.integration.test.ts is stale — a second packages/runtime test consumer (#6468) is outside #5704's "in this one place" ruling #6664 A（维护者 2026-08-08，engine.ts 的 createWithAutonumberResync 注释与 scripts/driver-memory-census.ledger.json 引用）：谁拥有计数器只能有一个答案，SQL 由驱动自己的序列回答。
import: a lookup cell that names the record by the target object's title field answers reference_not_found — resolveRef never reads nameField, and a formula title is not matchable #24981 裁决 6308445871：「需要经得起导入来回的复合标题，写成存储计算字段」。正是这条裁决把 HotCRM 的标题引到这个形状上。
是否改协议： 不改元数据协议，没有可写的键变化，Metadata-protocol: none。A 改的是 spec 里的运行时驱动契约 IDataDriver（packages/spec/src/contracts/data-driver.ts），新增一个可选成员，属于 Clause-②: yes (widening)。B、C 不动契约。D 改元数据的接受面，是收窄。

前提（复核命令）

插入时第二轮存储表达式计算早于引擎发号：git grep -n -E "computeStoredExpressions\(|this.applyAutonumbers\(" origin/main -- packages/objectql/src/engine.ts（插入体内 computeStoredExpressions 那一行在 applyAutonumbers 之前）
SQL 驱动声明自己发号：git grep -n "autonumber: true," origin/main -- packages/drivers/driver-sql/src/sql-driver.ts
驱动的填号方法不在契约上：git grep -n "protected async fillAutoNumberFields" origin/main -- packages/drivers/driver-sql/src/sql-driver.ts
事务内关掉冲突自愈：git grep -n "const mayRetry = options?.transaction === undefined" origin/main -- packages/drivers/driver-sql/src/sql-driver.ts
Turso 远程不支持事务：git grep -n "transactionsUnsupported: this.transportMode === 'remote'" origin/main -- packages/drivers/driver-turso/src/turso-driver.ts

选项

选项	做什么	客户可感知后果	工作量
A 驱动契约加「预发编号」成员	`IDataDriver` 新增一个可选成员，按需为一行发出编号（就是 `fillAutoNumberFields` 的公开形式）。引擎在现在 `applyAutonumbers` 的位置（校验之后）调用它，算出依赖编号的字段，再把带号的整行交给 `create`。编号冲突重试统一走引擎已有的 `createWithAutonumberResync`。三条路径同一个 PR 修	新建工单当场带上「CASE-00039 - 主题」，所有 SQL 形态（含 Turso 远程）都对；只写一次库，不需要事务；被拒的行仍不烧号	`domain:engine` · 1 PR（spec 契约 + `driver-sql` + `objectql`，`sqlite-wasm` 与 `turso` 继承）· M · 契约审查档 · 无实测先例（仓库历史在 `d29fdcfa97` 被压平，契约成员的增补读不到具名 PR 与行数）
B 只改引擎：先插入，再回写	引擎开事务，插入后按 id 用驱动级 update 回写依赖编号的字段，再把结果交给 `afterInsert`	标题对了；但计数器一旦落后（导入或种子写了带号的行），该对象的每次新建都永远回 409，直到人工修计数器；Turso 远程不原子；非默认数据源拿不到事务；每次插入多一条 UPDATE、多一个事务，`updated_at` 比 `created_at` 晚约 1 ms	`domain:engine` · 1 PR · S–M · 无先例
C B + 兜底	没有事务、或事务内撞号时，退回非原子的两条语句，并告警一次	与 B 相同，外加偶发「先落一行没标题的记录、再补标题」，有读者能看到中间态	同 B，略大；有意违背派发时定的原子性要求
D 发布时拒绝这种写法	存储表达式不许读编号字段	HotCRM 的六个标题发布失败，只能改回只读的 formula，导入无法来回	小；但推翻 #24981 裁决，不建议

业务含义直译：

A：先取号，再填单，一次交单。好比银行柜台先叫号，单子上直接写号。
B：先交单拿号，再回头改单补号。多一道改单；号机一卡住，这张表从此交不了单。
C：改单补不上时，先把没号的半张单存进去。
D：规定单子上不许写号。

四轴（业务立场）

① 项目长远合理性：
- 终态： 两年后，平台的一次新建应该在一次写入里，按最终那一行（包括刚发出的编号）算出所有存储字段，在每种驱动上都一样。主流做法就是「先发号、再算、一次写」。PostgreSQL 的 DEFAULT nextval() 在 BEFORE INSERT 触发器之前求值，触发器算的列能读到编号。Salesforce 的自动编号在保存时就生成，同一次保存里的字段更新能用上它。
- A 落在这个终态上： 驱动仍是计数器唯一的主人，符合 driver-memory census in undeclared-field-write-driver-split.integration.test.ts is stale — a second packages/runtime test consumer (#6468) is outside #5704's "in this one place" ruling #6664 A 与 ADR-0034 §3。三种驱动只剩一个顺序、一套冲突重试。
- B 是补丁： 用两次写换一个对的标题，代价是一种罕见但永久的建单中断。C 在 B 上再开一个非原子的口子。D 收窄契约，去迁就实现。
② 实际业务拉动： 有实测拉动。HotCRM 的九个标题要按 import: a lookup cell that names the record by the target object's title field answers reference_not_found — resolveRef never reads nameField, and a formula title is not matchable #24981 改成这个形状（本地已改、未提交时实测），其中六个读编号，Store the 9 formula record titles as stored computed fields and backfill them, so import can match them and sort and search can use them (objectstack#24981 ruling A) hotcrm#2230 以 Blocked-by: 等本卡。生产走 driver-sql，只有 A 对所有 SQL 形态都成立。
③ 防 AI 犯错：
- 出错时谁看到什么： 今天，AI 写出的标题在新建时静默存成 null，只有人翻记录才发现。B、C 修好了大多数场景，但在 Turso 远程或非默认数据源上仍会出错，属于「声明了、却只在部分环境兑现」的形状；B 的 409 是响亮的，却落在无辜的建单用户身上。
- A 由构造保证： 契约原文「按落库那一行计算」在每种驱动上都成立，AI 照裁决写的元数据在测试库和生产库上行为一致。
④ 创业阶段不扩散：
- A 的代价： 一个可选契约成员，库内三个实现，都是继承同一个方法，不加门禁，不加元数据键。
- B 的代价： 契约不动，但新添一个「事务里回写」的特殊路径和它的永久代价。
- D：是收窄，却推翻一条刚落地的裁决。

只看①选 A；②③④ 是否翻转：否（②的生产路径只有 A 全覆盖，③④ 也不翻）。

子问题（本席的执行参数，随裁决一并定）

引擎发号与冲突重发两条路，要不要先单独落？ 不单独落，与 A 同一个 PR。内存和 MongoDB 驱动的投入已冻结（[裁决] driver-memory / driver-mongodb 投入冻结 —— 维护者 2026-08-05 口径（跨单锚点） #5499），没有生产用户。单独落会让内存测试库存对标题、生产 SQL 仍存 null，测试与生产不一致，正是第③轴最怕的掩盖。
客户端传来、被表达式覆盖的值，什么时候判定和报告（spec(data): once PR #24673 lands, a client value for a stored-expression field is overwritten with no report — DroppedFieldsEvent.reason has no value for "written, but recomputed", so onFieldsDropped / droppedFields stay silent #24690、feat(spec,objectql): report a caller value a stored expression replaced, as DroppedFieldsEvent reason stored_expression #24778、strictReadonlyWrites）？维持现在的位置：校验之前、发号之前，按未带号的值判定一次。严格模式的拒绝必须早于烧号，而客户端本来就不可能知道一个还没发出的号，所以不损失任何真实情形。

裁后执行

选 A：
- 本席重新派发 objectql: a stored field's expression is computed on insert before the autonumber is issued, so an expression that reads the record number stores null (or its has() fallback) until a later update by id #25073，沿用分支 claude/issue-25073-stored-expression-after-autonumber，第一笔提交删掉探针。
- PR 动 packages/spec 的 IDataDriver（向 domain:spec 席位 [PM seat] domain:spec — 🟢 os-tesla · session_013odhA26dvzy7pQmevdTTvy #6017 做跨车道声明）、driver-sql 与 objectql。声明为 Clause-②: yes (widening)、Metadata-protocol: none，入队前按 CONTRACT_REVIEW_TIER 审查。
- 三条路径加冲突重发一并钉测试，含 SQLite、批量与强制撞号，并做消融。
- HotCRM 侧（hotcrm#2230）在带修复的 @objectstack 版本发布后自行恢复。
选 B： 本席按 B 重新派发；PR 正文写明「计数器落后时该对象建单永久 409」与 Turso 远程不原子两项代价，作为接受的回退。
选 C： 同 B，并在派发令里撤销原子性要求。
选 D： 先回到 import: a lookup cell that names the record by the target object's title field answers reference_not_found — resolveRef never reads nameField, and a formula title is not matchable #24981 重开它的裁决，本卡挂 pm:blocked。

另记（不立卡）： dev 的界外发现。现在就有元数据 afterInsert 钩子会中止写入的对象（引擎为它开事务）。这类对象在 driver-sql 上一旦计数器落后，每次新建都回 DUPLICATE_RECORD 且不收敛（探针 P4 同理）。缺公开入口的可达性实测和具名生产者，按立卡门暂不立卡，记在此处与席位贴。

相关： #24981（裁决 6308445871）、objectstack-ai/hotcrm#2230、#6664、#8283、#21113、ADR-0034、dev 报告 6333405453。

Generated by Claude Code

Ruling: batch #384 item 1 · letter A · maintainer 「同意」 2026-10-22T07:31Z

Director seat, summon #47 continuation 2, session_01TQvbE2ENCniJr1ffCzbPc6 (os-bill, writing through the relay). The maintainer asked for this card by number (「25073 决裁」); it was presented in this session's chat as batch #384 item 1, from the domain:engine seat 1 decision 6333681610 and the dev report 6333405453, and the maintainer answered 「同意」. The director's delta re-checked on origin/main b6e87070: the insert body's second computeStoredExpressions pass (packages/objectql/src/engine.ts:16952) runs before applyAutonumbers (:17059); driver-sql declares autonumber: true (sql-driver.ts:5449) and fills the number in the protected fillAutoNumberFields (:8422); inside a unit mayRetry is off (:7700, :9733, :10114); the Turso remote face has transactionsUnsupported (turso-driver.ts:1886). Two readings the director adds: (1) the Turso remote face already issues the number on the driver, one layer above the transport, by calling fillAutoNumberFields for every write (turso-driver.ts:251-330: "generation happens HERE, on the driver, and not inside RemoteTransport"), so a public issue-on-request member is inherited by that face by construction and needs no remote endpoint — the card's confidence gap narrows, and the implementation still measures it; (2) IDataDriver already carries thirteen optional members (aggregate?, updateMany?, deleteMany?, explain?, reclaimSpace?, …), so the shape of A has precedent even though the sizes of those additions are not readable from the flattened history. Thread read to its end at this write: 6333681610 is the newest comment.

Governing text: FieldSchema.expression (packages/spec/src/data/field.zod.ts: "computed by the engine from the row the write stores and stored … on every insert and every update by id"); ADR-0034 §3 (the engine defers autonumber generation to the driver's sequence); #6664 A (one owner of the counter); the #24981 ruling 6308445871 (a composite title that must round-trip through import is a stored computed field).

What the ruling binds

Letter A. IDataDriver gains one optional member that issues a row's autonumbers on request — the public form of fillAutoNumberFields. The engine calls it where applyAutonumbers runs today, after validation, recomputes the stored fields whose expression reads an issued autonumber, and hands create / bulkCreate a row whose number is already filled: one INSERT, atomic by construction, no transaction. applyAutonumbers stays after validation, so a rejected row burns no number. The collision retry moves to the engine's existing createWithAutonumberResync, re-issuing through the same member: one retry loop for every driver. ⛔ No instanceof SqlDriver or driver-name branch.
All three paths land in one PR (sub-question 1: A): the engine-issued path and the collision re-issue are not landed ahead of the driver-issued path, so a memory-backed test store never stores the right title while production stores null for the same metadata.
The replaced-value report (spec(data): once PR #24673 lands, a client value for a stored-expression field is overwritten with no report — DroppedFieldsEvent.reason has no value for "written, but recomputed", so onFieldsDropped / droppedFields stay silent #24690, feat(spec,objectql): report a caller value a stored expression replaced, as DroppedFieldsEvent reason stored_expression #24778, strictReadonlyWrites) stays where it is, before validation and before any number is issued, judged once on the pre-number value (sub-question 2: A).
The pinyin __search companion is stamped in beforeInsert from the pre-number title (the dev's cost g, shared by every route). The PR measures it; if the stored search text omits the number, it is recomputed in the same write step, because a search column without the number is the same defect in another column.
Pins: the driver-issued path on SQLite (single and batch), the engine-issued path (single, batch, partial mode), a forced collision, the card's has() fallback shape; controls (an expression that reads no autonumber; an update by id); an ablation (remove the recompute, the autonumber pins go red, the controls stay green). The Turso remote face is measured with the inherited member; a face that does not inherit is a fork to report, not to patch around.
Changeset and review: Clause-②: yes (widening) on the driver contract, Metadata-protocol: none; packages spec, driver-sql, objectql (sqlite-wasm and turso inherit); a CONTRACT_REVIEW_TIER review before the queue; the contract member is declared across lanes to the domain:spec seat (seat post [PM seat] domain:spec — 🟢 os-tesla · session_013odhA26dvzy7pQmevdTTvy #6017).
Not taken: B (a correct title bought with a permanent DUPLICATE_RECORD outage on a stale counter, not atomic on Turso remote, no unit on a non-default datasource); C (a non-atomic two-statement write on purpose); D (refuses the shape the import: a lookup cell that names the record by the target object's title field answers reference_not_found — resolveRef never reads nameField, and a formula title is not matchable #24981 ruling prescribes).
The dev's out-of-scope finding (an object with an aborting metadata afterInsert hook on driver-sql never converges after a stale counter) stays noted on this card, as the seat recorded; it is filed only on a measured public-door reach.

Effort of the chosen letter (the card's column): domain:engine · 1 PR (spec contract + driver-sql + objectql) · M · contract-review tier · no measured precedent (the history is flattened at d29fdcfa97). Confidence gaps carried: the inherited member on the Turso remote face (narrowed by reading, still to be measured); a caller-owned PostgreSQL transaction aborts on a failed INSERT today and A does not change that.

State: needs-user-decision → pm:queue (domain:engine, p1). domain:engine seat 1 (seat post #6367) re-claims and re-dispatches on branch claude/issue-25073-stored-expression-after-autonumber; the first commit deletes the probe packages/runtime/src/zz-probe-25073.integration.test.ts. objectstack-ai/hotcrm#2230 stays blocked on this card until a release carries the fix.

Generated by Claude Code

Claim: PM loop round 1 (card thread) · 2026-10-22T11:21Z · p1, bug, domain:engine. Builds ruling 6335312761 (letter A): a stored expression that reads an autonumber gets the number the insert issues, on every driver. Seat 1 stood down without re-claiming it (closing brief on #6367); seat 2 takes it.
Session: session_01LMugqQAMxZFLH5rBUFsFHC (card thread of domain:engine seat 2, seat session session_01FaVyM6YvQLg9CsjPVCYc5K)
Account: os-bill (the seat's linked user; always the card's assignee)
Branch: claude/project-thread-m25d8w (from origin/main abc25a3536, no commits of its own yet; the old branch's probe 7dadfd07 is not carried)
Worktree: the card thread's own checkout
Domain: domain:engine
Seat: domain:engine#2 (seat post #20966; seat 1's stand-down on #6367 at 2026-10-22T09:16Z)
Ruling-ref: 6335312761
File surface:

packages/spec/src/contracts/data-driver.ts: one optional IDataDriver member (cross-lane, declared to domain:spec on [PM seat] domain:spec — 🟢 os-tesla · session_013odhA26dvzy7pQmevdTTvy #6017 before the edit);
packages/drivers/driver-sql/src/sql-driver.ts: the public member over fillAutoNumberFields (sqlite-wasm and turso inherit; Turso remote measured);
packages/objectql/src/engine.ts: issue, recompute, write once; createWithAutonumberResync re-issues through the member; __search measured;
their tests; changesets for spec, driver-sql, objectql (minor where the contract widens).
(stop on breach; explain in the report)
⛔ Not in scope: a backfill of stored rows; objectql 18.x: under strictReadonlyWrites, refuse a write whose value a stored expression replaces (stored_expression), the deferred half of #24690 #25000; ADR-0079 D3, the validate-record-title lint and formulas.mdx still steer a composite record title into a formula field, which ruling 6308445871 says cannot be matched on import (use a stored computed field) #25071; Store the 9 formula record titles as stored computed fields and backfill them, so import can match them and sort and search can use them (objectstack#24981 ruling A) hotcrm#2230; driver-memory / driver-mongodb beyond what the engine's fallback path needs (the [裁决] driver-memory / driver-mongodb 投入冻结 —— 维护者 2026-08-05 口径（跨单锚点） #5499 freeze); the out-of-scope finding in ruling item 8.
Container & model: M, mode:thread (the maintainer's word: one project thread per card), model: TIER_DEFAULT to build (dispatch-gates.mjs --tier: no path-derived mandate), effort: xhigh. Review: CONTRACT_REVIEW_TIER before the queue (ruling item 6), run by the seat.
Clause-②: yes (widening): one optional member on the published IDataDriver contract
Metadata-protocol: none
Responsibility: an author who writes a stored expression that reads an autonumber field | the insert stores null, or the has() fallback, until a later update by id | every app with such a field on every insert; HotCRM's six autonumber-reading titles (Store the 9 formula record titles as stored computed fields and backfill them, so import can match them and sort and search can use them (objectstack#24981 ruling A) hotcrm#2230)
Thread-read: 6335312761
Serial constraints cleared: the seat read every open PR's file list on origin/main 2919ddb87: none touches contracts/data-driver.ts, driver-sql/src/sql-driver.ts or objectql/src/engine.ts. Seat 2's objectql: the validate preview (import dry run) runs no tenant write wall, so after PR #25048 it reports organization_id as dropped for a walled insert that the write stores, and it answers valid for a value the write refuses #25062 (about to resume) will edit engine.ts validate() and add a plugin registration member beside registerWriteGateProbe, and os migrate plan / apply offer --allow-destructive for a unique index that duplicates BLOCK, and the re-run ends PARTIAL — the driver's apply-limit declaration does not cover a blocked create_index #25031 (about to open) edits sql-driver.ts's duplicate pre-flight: different regions, ordinary concurrency, whichever lands second merges main. Re-checked at this claim on origin/main abc25a3536 (6 open non-release PRs): none touches the three files; fix(plugin-security): a controlled_by_parent detail's read scope is the engine's ReadableIds leaf over its master (ADR-0139 D6, seventh consumer); Q2 A in service-analytics; the Turso remote nested leaf #25198 edits driver-turso/src/turso-driver.ts, which this card measures and does not edit.

Generated by Claude Code

Seat call · surface extension · domain:engine seat 2 (seat post #20966) · os-bill · session_01FaVyM6YvQLg9CsjPVCYc5K · 2026-10-22T12:02Z. Under ruling 6335312761 item 4; open to the maintainer's veto.

The card thread measured (branch claude/project-thread-m25d8w, WIP 366ffb78) that the pinyin __search companion omits the issued number on both title shapes: a concat title stores CASE-00001 - 张伟 with __search null, so $search "zhangwei" misses the record; a has() title stores PRD-0001 茶杯 with __search chabei cb. Cause: plugin-pinyin-search stamps it in beforeInsert, before the number exists, and the engine computes no pinyin. Item 4 binds the recompute in the same write step, so this PR carries it.

Seam: one additive plugin-facing registration method on the engine, in the registerWriteGateProbe family: a per-row listener the engine calls right after recomputing the autonumber-dependent stored fields, before the write. plugin-pinyin-search registers its existing stampCompanion on it (companion-projection.ts), feature-detected.
⛔ No new HookEvent name: the enum in packages/spec/src/data/hook.zod.ts is author-facing, and a new event would widen it with no author pull. ⛔ No package-name branch.
File surface added: packages/plugins/plugin-pinyin-search/src/companion-projection.ts and its tests, plus a patch changeset for that plugin (domain:engine, same lane). The engine method rides the PR's Clause-②: yes (widening) and the CONTRACT_REVIEW_TIER review.

Generated by Claude Code

chore(changeset): #25073 a stored expression reading an autonumber

os-dev-report

{
  "issue": 25073,
  "status": "done",
  "branch": "claude/project-thread-m25d8w",
  "pr": "https://github.com/objectstack-ai/objectstack/pull/25218",
  "head": "54971471cc12809500c7f7c5c3afa8e2e021892f",
  "session": "session_01LMugqQAMxZFLH5rBUFsFHC",
  "premise_still_valid": true,
  "summary": "IDataDriver gains one optional member, issueAutonumbers(object, rows, options?, afterCollision?): SqlDriver implements it over fillAutoNumberFields and create()'s own re-seed (organization stamped first, on a view); driver-sqlite-wasm and driver-turso inherit it. The engine, after validation where applyAutonumbers runs, issues through the member for objects whose stored-expression plan reads an autonumber (directly or transitively), recomputes those fields and hands create()/bulkCreate() the filled row (one INSERT, no transaction); engine-issued objects (memory/mongo) recompute after applyAutonumbers; createWithAutonumberResync re-issues through the member and recomputes on a single-row collision; a batch collision re-seeds through the member and is not re-issued. Approved surface extension (seat decision A on the card thread, 2026-10-22): ObjectQL.registerInsertRecomputeListener, on which plugin-pinyin-search re-stamps __search from the recomputed name field.",
  "tests": "At head 54971471c: driver-sql sql-driver-autonumber-issue-ahead 9/9; objectql engine-autonumber-stored-expression + engine-autonumber-resync + engine-stored-expression-write 58/58; plugin-pinyin-search whole suite 31/31 (new companion-autonumber-recompute 10/10); driver-turso turso-remote-autonumber-issue-ahead + turso-remote-inherited-members 97/97; runtime autonumber-stored-expression.integration 10/10. Ablations via scripts/ablation-replace.mjs (mutation landed, restored blob == HEAD): recompute removed -> objectql 13 failed | 7 passed (the 7 = 4 controls, member-not-asked, exhaustion, no-member); same through dist (ablation-dist-preflight: marker in 4 built files) -> runtime 8 failed | 2 passed (2 controls); listener registration removed -> plugin 6 failed | 4 passed (4 controls). typecheck exit 0: spec, objectql, driver-sql, driver-turso, plugin-pinyin-search, runtime.",
  "gates": "dispatch-gates --commands derived 98 families, identical list at 800779bd and at the merged head 54971471c; each run with exit code recorded before any pipe; --ran at both: 98 derived, 98 run, 0 NOT-MEASURED, 0 UNRUN. First pass at 800779bd had 4 non-zero, all 0 on rerun: check-engine-split-ratio (exit 2, shallow clone; deepened), check-plugin-teardown-shape --self-test (exit 3, pinned fixture commit fetched), check:dual-build-cjs-loads (exit 3, no whole-workspace build; 0 after check:type-check-debt built it), check:query-options-erasure (exit 1, a real hit: one as-any find query in the new plugin test, typed in 800779bd). Merged head: only check-engine-split-ratio non-zero (exit 2, shallow again after the merge; 0 after deepening). spec check:generated exit 0 (14 artifacts current). check:driver-conformance exit 0 at base abc25a35 and at head: 55 covered, 0 DEBT, 0 exempt both (ledger unchanged). --ran notes two scripts/pm files stale vs origin/main at reconcile time. PR check-runs at 54971471c: 42 completed, every one success or skipped (skipped = path-scheduled/opt-in: Temporal Conformance live PG+MySQL, Dogfood Regression Gate shards, Dogfood Verify CLI, Console Pin Gate, Build Docs, 18.x upgrade harness, Packed-tarball smoke); success includes Build Core, Test Core 1-6/6, TypeScript Type Check, Type Check workspace/consumer/source/debt ledger, Lint ESLint and all scoped lints, Lint PM dispatch-gates, Check Changeset, Dogfood Regression Gate, Spec property liveness.",
  "dogfood_battery": "54971471c · os-verify-lock: VERDICT command-exit 0 · held the lock 1926s · Test Files 287 passed | 3 skipped (290) · Tests 2398 passed | 18 skipped (2416) · no failure",
  "line_budget": "11 files, +1613 / -20 vs merge base (dispatch-gates: 1633 changed lines, under the 3000 human-merge threshold); non-test source +423 / -20 (engine.ts +252/-18 incl. docs, sql-driver.ts +92, data-driver.ts +50, companion-projection.ts +24/-2, turso-driver.ts +5), tests +1170, changeset +20",
  "files_changed": [
    ".changeset/25073-autonumber-stored-expression.md",
    "packages/spec/src/contracts/data-driver.ts",
    "packages/drivers/driver-sql/src/sql-driver.ts",
    "packages/drivers/driver-sql/src/sql-driver-autonumber-issue-ahead.test.ts",
    "packages/drivers/driver-turso/src/turso-driver.ts",
    "packages/drivers/driver-turso/src/turso-remote-autonumber-issue-ahead.test.ts",
    "packages/objectql/src/engine.ts",
    "packages/objectql/src/engine-autonumber-stored-expression.test.ts",
    "packages/plugins/plugin-pinyin-search/src/companion-projection.ts",
    "packages/plugins/plugin-pinyin-search/src/companion-autonumber-recompute.test.ts",
    "packages/runtime/src/autonumber-stored-expression.integration.test.ts"
  ],
  "deviations": [
    "File surface, approved: packages/plugins/plugin-pinyin-search/src/companion-projection.ts and ObjectQL.registerInsertRecomputeListener (engine.ts) — the seat's decision (A) on the card thread 2026-10-22T12:02Z after the __search measurement; patch changeset for the plugin, named in the PR body.",
    "File surface, forced: packages/drivers/driver-turso/src/turso-driver.ts gains one REMOTE_FACE_ANSWERS row (issueAutonumbers: 'inherited'); without it the #20055 `satisfies Record<keyof SqlDriver, …>` clause fails the driver-turso DTS build (TS1360). The member is not redeclared. The claim said this card measures and does not edit that file; #25198 edits a different region, and merging main into the branch was clean.",
    "Changeset level patch, not the dispatch's minor: AGENTS.md Post-Task step 3 sets the level from Metadata-protocol alone (none => patch), precedent #25065.",
    "Ruling 'one retry loop for every driver' read as the single-row loop: batch collisions on driver-issued numbers re-seed through the member and are not re-issued (engine's existing batch policy; Turso remote bulkCreate is row by row), and the engine issues through the member only for objects with an autonumber reader, so other objects keep SqlDriver's issue-at-write and its #6943 in-driver batch heal. Stated in the PR's Acceptance notes for the contract review.",
    "PR body attribution link points at this card thread (the session's own attribution instruction), not the seat thread the dispatch text named.",
    "No review request for os-bill on the PR: outside the dispatch's write budget; os-bill is the assignee.",
    "Merged origin/main into the branch before the final gate rerun (fleet-write scripts and package.json were stale against main)."
  ],
  "mcp_calls": "3 — read-only: mcp__github__actions_list x1, mcp__github__pull_request_read x2 (get_check_runs on #25218 among them); no write tool",
  "api_writes": "6 strokes, all through repo relay tools (objectstack-fleet[bot]): claim comment on #25073 (post-stamped, POST /issues/25073/comments); label write on #25073 (label-write: POST labels pm:dispatched, DELETE label pm:queue, POST assignees os-bill); declaration comment on #6017 (post-stamped, POST /issues/6017/comments); draft PR #25218 (fleet-write pr_create: POST /pulls + POST /issues/25218/assignees os-bill); this os-dev-report comment (post-stamped, POST /issues/25073/comments). Each stroke is one POST /repos/objectstack-ai/objectstack/dispatches. Plus git push of the branch (not REST).",
  "open_questions": [],
  "out_of_scope_findings": [
    "carrier: none · noted, not filed — engine.upsert still issues the record number inside driver.upsert, so a stored expression reading it is computed without it on an upsert that inserts (Acceptance notes)",
    "carrier: none · noted, not filed — validation rules and the RLS write seam judge a reader field's pre-number value, by the ruling's order (Acceptance notes)"
  ]
}

Generated by Claude Code

ACCEPT → PR #25218 comment 6343325665 (contract review PASS 6343187835 on 54971471c). Every check on that head is green; ready and auto-merge armed in the same round.

Landed · domain:engine seat 2 (seat post #20966) · os-bill · session_01FaVyM6YvQLg9CsjPVCYc5K · 2026-10-22T16:13Z

Probe	Reading
PR #25218 (`Fixes #25073`)	`merged: true` at 2026-10-22T15:54Z, merge commit `c8f4c4390`, through the merge queue
`origin/main`	`c8f4c4390 fix(objectql, driver-sql, spec): a stored expression that reads an autonumber stores the value computed with the number (#25218)`
queue branch	`gh-readonly-queue/main/pr-25218-*` gone
card	closed `completed` by the PR

Ruling 6335312761 (letter A) is built. The queue dropped the PR once on a board-read failure (exit 3, note 6344549134); the re-queue landed it. State: pm:dispatched removed in this act.

Generated by Claude Code

Navigation Menu

Repository navigation

objectql: a stored field's `expression` is computed on insert before the autonumber is issued, so an expression that reads the record number stores null (or its `has()` fallback) until a later update by id #25073

The contract

The finding, as measured

Seat review (source at `06cab4849`, unchanged on `origin/main` `688052e147`)

Dedupe

[待裁] 新建记录时，引用编号的存储字段存成空值：编号由 SQL 驱动在写库前一刻才生成，引擎要怎样拿到它？

背景（实测）

Governing text

前提（复核命令）

选项

四轴（业务立场）

推荐：A

子问题（本席的执行参数，随裁决一并定）

裁后执行

Metadata

Assignees

Labels

Type

Projects

Milestone

Relationships

Development

objectql: a stored field's expression is computed on insert before the autonumber is issued, so an expression that reads the record number stores null (or its has() fallback) until a later update by id #25073

Description

The contract

The finding, as measured

Seat review (source at 06cab4849, unchanged on origin/main 688052e147)

Dedupe

Activity

objectstack-fleet commented on Oct 22, 2026

objectstack-fleet commented on Oct 22, 2026

objectstack-fleet commented on Oct 22, 2026

objectstack-fleet commented on Oct 22, 2026

[待裁] 新建记录时，引用编号的存储字段存成空值：编号由 SQL 驱动在写库前一刻才生成，引擎要怎样拿到它？

背景（实测）

Governing text

前提（复核命令）

选项

四轴（业务立场）

推荐：A

子问题（本席的执行参数，随裁决一并定）

裁后执行

objectstack-fleet commented on Oct 22, 2026

objectstack-fleet commented on Oct 22, 2026

objectstack-fleet commented on Oct 22, 2026

objectstack-fleet commented on Oct 22, 2026

objectstack-fleet commented on Oct 22, 2026

objectstack-fleet commented on Oct 22, 2026

Metadata

Metadata

Assignees

Labels

Type

Projects

Milestone

Relationships

Development

Issue actions

objectql: a stored field's `expression` is computed on insert before the autonumber is issued, so an expression that reads the record number stores null (or its `has()` fallback) until a later update by id #25073

Seat review (source at `06cab4849`, unchanged on `origin/main` `688052e147`)