Skip to content

plugin-security: a controlled_by_parent child's read loads every readable master id with no limit (twice per request), so its CPU grows with the master table and the read answers 500 "too many SQL variables" past ~32.8k masters #24932

Closed

Description

@objectstack-fleet

Ruled: 6342223675 · letter fork 1 A / fork 2 A · Tier H approval in words · 2026-10-22T13:43Z
Ruled: 6329027469 · letter Q1 C / Q2 A · 2026-10-21T23:44Z
Ruled: 6303721609 · letter B · 2026-10-20T23:08Z

Filing gate ①: class (a), measured. A public door answers 500 and its cost grows unbounded with data. Filed by repo:cloud#1 (seat post #6026, session session_01FSg7JRUiTPo8GrgT38tAWP) from the cloud#3053 round 4 dev report (objectstack-ai/cloud#3053, comment 6295337145). ⛔ Not a claim.

Reader who acts: triage grades and routes it (expected domain:services, plugin-security); the lane's execution seat fixes it. Cloud consumes the fix at its next framework pin move.

Dedupe: objectstack issues of both states (444 read) grepped for controlled_by_parent, computeControlledByParent, too many SQL variables, master ids and master-detail … $in. 1 hit: #24291 (closed, a spec docblock about forceCbpMasterDetailRequired, a different subject). No duplicate.

Where (read at objectstack main 74fcec1e; same code at 9d13b5fc)

packages/plugins/plugin-security/src/security-plugin.ts, computeControlledByParentFilter (:8535; the scan at :8615-8626):

const rows = await this.ql.find(rel.master, {
  where: masterFilter ?? {},
  fields: ['id'],
  context: { isSystem: true },
});
masterIds = rows.map((r) => r?.id).filter(…);
return { [rel.fk]: { $in: masterIds } };

There is no limit. When the caller's master scope is unrestricted, masterFilter is null and the scan is the whole master table. The statement log shows select id from MASTER, unbounded, twice per request (once for find, once for count).

Measured (local hosted-shaped rig, cloud main b4a6a777 = framework 9d13b5fc, HotCRM 5.0.0; crm_contact and crm_quote are controlled_by_parent of crm_account; environment owner)

crm_account rows GET /api/v1/data/crm_contact?top=1, main-thread CPU answer
9 ~8 ms 200
1,009 12 ms 200
10,009 45 ms 200
40,009 / 50,009 — 500 DATABASE_ERROR (too many SQL variables, sql-driver log)

crm_account's own read stays at 5–6 ms throughout. Staging's Turso remote was not measured.

Wanted

  1. Push the master scope into the child query as a subquery or semi-join, instead of materialising an id list. When the caller's master scope is unrestricted, short-circuit to no predicate.
  2. Pin it with a statements-per-request test on a controlled_by_parent child, and with a test past the dialect's variable limit (the read answers 200 with the right rows).

Acceptance

  • A CBP child read answers 200 with the rows the master scope allows, at a master count past the dialect's variable limit.
  • Its statement count and CPU do not grow with the master table when the master scope is unrestricted.

Blocking (director seat, ruling Q1 C 6329027469; the two engine cards #25009 and #25010 the earlier lines named closed completed, so those lines left the body)

The plugin-security switch to the ADR-0139 leaf is dispatched after #25144 (the boot-time refusal of a controlled_by_parent master on another datasource or behind an object-scoped read middleware) has merged. Q2 A (service-analytics) rides the switch PR. The empty branch claude/issue-24932-cbp-master-scope is not reused; claude/issue-24932-cbp-leaf-switch is empty at f2b6cf230.

Activity

  1. objectstack-fleet commented on Oct 20, 2026

    @objectstack-fleet
    ContributorAuthor

    Claim: PM loop round 3 · 2026-10-20T18:22Z
    Session: session_01UmqzD97e14iws6hmVhgZ5y
    Account: os-support-ai (the seat's linked user as GET /user answers it; always the card's assignee)
    Branch: claude/issue-24932-cbp-master-scope
    Worktree: objectstack-issue-24932
    Domain: domain:services
    Seat: domain:services#2 (seat post #21118)
    File surface: packages/plugins/plugin-security/src/security-plugin.ts (computeControlledByParentFilter and its two call sites' per-request reuse only), pins beside it in plugin-security, one new .changeset/24932-*.md. ⛔ Not packages/objectql/**, ⛔ not packages/drivers/**, ⛔ not packages/spec/**: if the semi-join needs an engine or driver change, stop and report the design for domain:engine. Stop on breach and explain in the report.
    Container & model: M, mode:subagent, model: opus (TIER_DEFAULT; dispatch-gates --tier on this surface: no path-derived mandate, floor sonnet, default opus, ceiling fable)
    Clause-②: no
    Responsibility: plugin-security materialises every readable master id into an $in list, with no limit, twice per request | none in the filter path today, to be measured: a semi-join spelling the engine already supports | every reader of a controlled_by_parent child; measured at a public door on a hosted-shaped rig, 500 past ~32.8k masters (card)
    Thread-read: none
    Serial constraints cleared: no open PR touches plugin-security/src/security-plugin.ts (all open PRs' files read this act). This seat's #24865 (PR #24930, in the merge queue) touches service-messaging and plugin-auth only.

    ⛔ The visible rows must not change: a child whose foreign key is null or dangling stays outside the read as it is today. A short-circuit that drops the predicate would widen what is visible, which is a security-boundary relaxation and outside this card.


    Generated by Claude Code

  2. objectstack-fleet commented on Oct 20, 2026

    @objectstack-fleet
    ContributorAuthor

    os-dev-report

    {
      "issue": 24932,
      "status": "needs_decision",
      "branch": "claude/issue-24932-cbp-master-scope",
      "pr": null,
      "session": "session_01UmqzD97e14iws6hmVhgZ5y — the dispatching PM's session (mode:subagent shares it)",
      "premise_still_valid": true,
      "summary": "M1 holds on origin/main 223ae939b (unchanged on the measured files through 8cfe1f94c). computeControlledByParentFilter (security-plugin.ts:8535, scan :8613-8626) scans the master with no limit under a system context and returns fk $in ids. It runs once per engine operation, so a findData list request (find, then count, protocol.ts:15134) scans twice. Measured on the real stack (ObjectQL + SqlDriver, better-sqlite3): find+count on a controlled_by_parent child = 4 statements, 2 of them `select id from MASTER`; a 2-level chain = 8 statements; with 33,000 masters both find and count answer 500 DATABASE_ERROR `too many SQL variables` (better-sqlite3 limit measured: 32,766 binds pass, 32,767 fail). M2: exactly one semi-join spelling exists, the ADR-0139 ReadableIds leaf (engine.ts:6120 composeReadableIdsAtBoundary, driver-sql compiles `fk in (select id from MASTER)`), and it does NOT keep the visible rows. Experiment (uncommitted; ablation-replace, blob == HEAD restored and git diff HEAD empty after each run): the leaf gives 1 statement per find, 2 per find+count, and 200 with the right rows past 33k masters. It matches today's rows for restricted, unrestricted, null-FK and dangling-FK children. It differs in 3 measured cases: (1) on-behalf-of read: today c1, leaf 500 INTERNAL_ERROR (engine.ts:6185 refuses a delegated caller, ADR-0139 D5.1); the delegator half cannot be spelled at all, because the engine composes for the outer caller only. (2) A caller holding child read but no object-level read on the master: today every child under an existing master (c1,c2), leaf none (canReadObject gate, engine.ts:6237). (3) A 2-level chain: the leaf nested in the master's scope reaches the driver uncomposed, so 500 INTERNAL_ERROR from driver-sql. From source only: a federated master (engine.ts:6202) and a master narrowed by an object-scoped middleware (engine.ts:6196) are refused 500. getReadFilter (:5943) hands its scope to the analytics raw-SQL door, which refuses the leaf as a non-list comparand. Plugin-security cannot hand the engine its own composed scope either: the boundary refuses a forged composed leaf (filter-comparand-shape.ts:196). ADR-0139 D6 and O5 leave open whether controlled_by_parent becomes its seventh consumer, and ADR-0055 §2 mechanism (b) stands, so switching it is a decision (Prime Directive #13), not a card edit. The interim the PM asked about (one scan per request instead of two) is not sound inside the landing. find and count share only the caller's context object (engine.ts:3722 returns it by reference). That object is not request-bounded for non-REST callers that reuse it across writes, so a memo keyed on it would serve a stale master set, wider or narrower. The one sound precedent (packages/core/src/security/request-grants-memo.ts) needs a scope opened by the request owner plus an engine write observer, and its scope closes before the handler runs. It also leaves the 500 and the unbounded scan in place. Not shipped. The card's Wanted #1 short-circuit was measured as the forbidden widening: with the predicate dropped on an unrestricted scope, c_null and c_dangling become visible. No code, no PR.",
      "tests": "Measurement only, nothing committed. Harness (removed from the worktree, copy kept): /tmp/claude-0/-home-user/baba1a33-5813-5ac5-906a-64b704dc6695/scratchpad/issue-24932/measure-harness.test.ts.txt. Raw outputs are in the same directory: measure-main3.txt (today), measure-leaf.txt and measure-leaf2.txt (leaf), measure-shortcircuit.txt. Command: `bash scripts/pm/os-verify-lock.sh -c 'pnpm --filter @objectstack/plugin-security exec vitest run --maxWorkers=2 src/zz-measure-24932.test.ts'`, every run `VERDICT command-exit 0`, `Tests 3 passed (3)`. TODAY: viewAll c1,c2 · owner c1 · childOnly c1,c2 (direct master read: PERMISSION_DENIED 403) · agentForU1 c1 · chain l1 · `STATEMENTS find 2 find+count 4` (select id from msr_account twice) · `CHAIN STATEMENTS find+count 8` · BIG (33,000 masters): find and count both `THROW DATABASE_ERROR 500 … too many SQL variables`. LEAF: viewAll c1,c2 · owner c1 · childOnly empty · agentForU1 `THROW INTERNAL_ERROR 500 … the engine composes no such leaf for a delegated (on-behalf-of) caller` · chain `THROW INTERNAL_ERROR 500 [driver-sql] … ReadableIds leaf … UNCOMPOSED` · `STATEMENTS find 1 find+count 2` (`account in (select id from msr_account)`) · BIG find bc1,bc2, count 2. SHORT-CIRCUIT: viewAll c1,c2,c_dangling,c_null · childOnly c1,c2,c_dangling,c_null. Mutation and restore lines, each run: `ablation-replace: ok mutation landed: anchor 1 -> 0` then `ok restored: blob == HEAD (25d6a049dc7b) and git diff HEAD is empty`. The first leaf attempt was refused by the tool because the anchor was contained in the replacement (no mutation was measured), then corrected. NOT MEASURED: tenancy-enabled master (the subquery's driver tenant wall vs Layer 0 inside today's scan); a federated master; a master narrowed by an object middleware; Postgres/MySQL/Turso.",
      "gates": "No diff, so no gate families. `node scripts/pm/dispatch-gates.mjs --commands --repo objectstack-ai/objectstack` in the worktree answers exit 2: `this branch changes nothing against 'origin/main' (merge base 223ae939b) — nothing to derive`. Typecheck and dispatch-gates --ran: not applicable (no commit). Built locally: `pnpm --workspace-concurrency=2 --filter '@objectstack/plugin-security^...' build`, VERDICT command-exit 0. No PR, so no check-runs.",
      "mcp_calls": "0",
      "api_writes": "1 — POST /repos/objectstack-ai/objectstack/issues/24932/comments (this report, via scripts/pm/post-stamped.mjs). Also `git push -u origin claude/issue-24932-cbp-master-scope`: the empty-branch write probe, branch head = 223ae939b, not a REST write. pr_create not used.",
      "line_budget": "n/a — no file changed",
      "deviations": [
        "No PR and no code: needs_decision under the dispatch's M2 rule (no spelling inside plugin-security keeps the rows; ADR-0139 O5 is undecided). The remote branch claude/issue-24932-cbp-master-scope stays empty at 223ae939b.",
        "Measurement harness written into the worktree (packages/plugins/plugin-security/src/zz-measure-24932.test.ts), never committed, deleted before cleanup; copy in the scratchpad.",
        "Four experimental mutations of security-plugin.ts through scripts/ablation-replace.mjs: 1 refused by the tool, 3 measured. Each was restored, blob == HEAD 25d6a049dc7b, git diff HEAD empty."
      ],
      "files_changed": [],
      "open_questions": [
        {
          "question": "How should a controlled_by_parent child's read scope stop materialising the master id list (ADR-0139 O5: is controlled_by_parent its seventh consumer), given that the only engine semi-join today, the ReadableIds leaf, changes visible rows in 3 measured cases (delegated caller → 500, no object-level read on the master → no rows, 2-level chain → 500)?",
          "options": [
            "A — O5 = yes, caller half only, chosen up front. Single-level masters, non-delegated callers, same datasource, no object middleware: the caller half writes ReadableIds.of(master). Every other shape keeps today's id list: delegated (both halves), chained, federated or middleware-narrowed master, and the getReadFilter/analytics path. Precondition: the no-master-object-read case must first be ruled and given one answer on the id list, or the switch narrows it. Effort: 1 plugin-security PR (~200-300 lines with real-stack pins) plus a one-line ADR-0139 O5 / ADR-0055 status pointer (Tier H). Business need: fixes the measured HotCRM crm_contact/crm_quote → crm_account case (500 → 200, 4 → 2 statements per list). Leaves HotCRM quote lines (2-level chain) and agent reads at the ceiling. Long-term: one accepted mechanism, but two live spellings of one scope (leaf and id list), the drift getReadFilter's shared helper exists to prevent. AI-safety: neutral, nothing authorable. Startup scope: smallest change that serves the measured case, no new engine form.",
            "B — engine first, then a wholesale switch (domain:engine, ADR-0139 amendment, Tier H). (i) A delegated read composes the leaf with the D10 intersection; this is D5.1's own exit condition, threading the delegator into getReadFilter. (ii) A leaf nested inside a composed scope is composed; measured: driver-sql refuses it today. Then plugin-security moves the caller half, the delegator half and getReadFilter in one PR. Federated and middleware-narrowed masters still need a ruling (keep the list or refuse loudly). Effort: 2-3 engine PRs plus 1 plugin-security PR. Business need: covers chains and agent reads too. Long-term: one mechanism, no fallback; best end state. AI-safety: neutral. Startup scope: largest; widens ADR-0139 beyond what its six consumers needed.",
            "C — stay on ADR-0055 (b) and make the ceiling honest, in this lane: past a bound below the dialect limit, refuse with a located error (the RELATION_FILTER_ID_CAP posture, ADR-0139 letter C) instead of an unattributed 500 DATABASE_ERROR. Effort: small. Business need: serves no rows past the bound, so the card's acceptance is unmet. Long-term: a patch; ADR-0139 already ruled cap moves out for its consumers. AI-safety: a loud refusal beats a 500. Startup scope: smallest, but it does not deliver the card.",
            "D — the driver binds a large $in as one parameter (json_each / = ANY). Not recommended: ADR-0139 ruled out the unbounded $in (its alternative B), and the scan's CPU still grows with the master table, the card's other measured axis."
          ],
          "recommendation": "A now, with B filed as the domain:engine follow-up that retires A's fallback. Business need: A clears the measured HotCRM 500 with one plugin-security PR, and the chain and agent shapes keep exactly today's answer, with no regression. Long-term: B is the end state, and A's fallback is chosen up front in ADR-0139 D5 style, never reached by catching a refusal, so it retires cleanly once B lands. AI-safety: both are non-authorable, a wash. Startup scope: A adds no engine form; B is written only because chains (quote lines) are a measured business shape. Two rulings are needed before A: O5 = yes, and the master object-read question below. Without them, A violates this card's binding row constraint in the no-master-read case."
        },
        {
          "question": "Today a caller with read on a controlled_by_parent child but NO object-level read on its master gets every child row whose master exists (measured: direct master read 403, child read c1,c2). ADR-0055 §3 says 'Reading a detail requires read access to its master (§2)' and §4 says 'a member who cannot read master M can neither read nor by-id-write a detail D under M'. The ADR-0139 leaf answers no rows here (canReadObject). Which answer is the contract?",
          "options": [
            "A — the ADR-0055 §4 reading: no master object read, no detail rows. Fix it on the id-list path first, in its own card (plugin-security: the master's object-level read admission before the scan), so a later leaf switch keeps rows by construction. Effort: small. Business need: none measured relying on today's answer (not searched in HotCRM). Long-term: one answer on both spellings. AI-safety: declared equals enforced. Startup scope: a narrowing fix, nothing new.",
            "B — keep today's answer (the child's object grant plus the master's row scope decide) and record it as the rule. Then the leaf can never be a drop-in for controlled_by_parent without an engine opt-out of canReadObject. That is a second engine path. Effort: an ADR-0055 clarification plus engine work. Long-term: the ADR text and runtime stay apart until edited."
          ],
          "recommendation": "A, because ADR-0055's own §3/§4 text already states it (spec declaration over implementation), and it turns option A of the first question into a row-preserving switch. It needs the maintainer's word because it narrows what some callers see today."
        }
      ],
      "out_of_scope_findings": [
        "class: b · reach: exception: security (may leak data); measured on the real stack at engine.find with a member context (the call GET /api/v1/data/OBJECT makes), not over HTTP · evidence: a caller holding read on controlled_by_parent msr_contact and no object-level read on its private master msr_account gets PERMISSION_DENIED 403 reading the master directly, yet reads c1,c2, every child whose master exists. Contract: ADR-0055 §4 'a member who cannot read master M can neither read nor by-id-write a detail D under M'. Seam: spec:`sharingModel: 'controlled_by_parent'` → runtime:`security-plugin.ts#computeControlledByParentFilter` (no master object admission before the system scan) · dedupe words: controlled_by_parent master object read, detail visible without master grant, canReadObject master, ADR-0055 §4",
        "carrier: #24932 (this card's decision) · noted, not filed · ADR-0139 consumers whose parent object is controlled_by_parent (attachments, comments, activity on a CBP child) compose their subquery scope from getReadFilter(child), which still carries the unbounded master id list. Same ceiling, same family. Source reading, not measured.",
        "carrier: #24932 (this card's decision) · noted, not filed · ADR-0055 Consequences says 'One extra master-id query per controlled_by_parent object per request (cached within the request)'. Measured: no request cache, the master scan runs once per engine operation (find + count = 2). Declared text over a governed record; correct it in the same Tier H edit that answers O5."
      ]
    }

    Generated by Claude Code

  3. objectstack-fleet commented on Oct 20, 2026

    @objectstack-fleet
    ContributorAuthor

    [Decision] 主从(controlled_by_parent)子对象读取:是否改用 ADR-0139 的半连接叶子(O5),以及顺序

    domain:services seat 2 (seat post #21118) · os-support-ai · session_01UmqzD97e14iws6hmVhgZ5y · 2026-10-20T18:50Z。依据:本卡开发报告 6298477402(只测量、未提交代码;origin/main 223ae939b,在真实 ObjectQL + SqlDriver 上实测)。

    Release: session_01UmqzD97e14iws6hmVhgZ5y · 因:唯一能消除 ID 列表的引擎形式会改变可见行;是否让 controlled_by_parent 成为 ADR-0139 的第七个消费者是该 ADR 明文留给你的开放问题(O5)· 去向:needs-user-decision,本笔清 assignee。开发分支 claude/issue-24932-cbp-master-scope 为空(未提交任何代码)。

    一句话问题: 子对象(如 HotCRM 的联系人、报价)每次列表读取都把「当前用户能看的全部主记录 ID」查出来拼成 IN (...),一次请求查两遍,没有上限。主表到约 3.3 万行时直接 500。引擎里唯一能改成子查询的现成写法,又会在三种情况下改变用户看到的行。

    背景(实测)

    • 现状: 列表读取(查询加计数)共 4 条 SQL,其中 2 条是 select id from 主表;两级主从链是 8 条;3.3 万主记录时查询和计数都返回 500 too many SQL variables(SQLite 上限实测为 32,766)。
    • 引擎现成的半连接(ADR-0139 的 ReadableIds 叶子): 每次查询 1 条 SQL,超过 3.3 万主记录仍返回 200 且行正确。受限、不受限、外键为空、外键悬空这几种情况结果与今天一致。
    • 但它在 3 种实测情况下改变结果:
      1. 代他人读取(代理身份):今天能读到,叶子直接 500(ADR-0139 D5.1 明确拒绝代理调用方);
      2. 有子对象读权限、但对主对象没有对象级读权限:今天能读到全部子记录,叶子一条都不给。这一条席位已作为安全缺陷另立 plugin-security: a controlled_by_parent detail is readable by a caller with no read on its master object, where ADR-0055 refuses it (detail withheld under the disclosure ruling) #24962,因为 ADR-0055 §3 写明「读子记录需要对主记录有读权限」,今天的行为本身不符合协议;
      3. 两级主从链:叶子嵌在主对象的范围里未被组合,500。
    • 按源码推断(未实测): 联邦主对象、被对象级中间件收窄的主对象也会被拒绝;分析接口不接受叶子。
    • 席位内能做的临时办法都不成立:
      • 「主范围不受限就去掉条件」会让外键为空或悬空的子记录变得可见(实测),这是放宽安全边界;
      • 「一次请求只扫一遍」找不到以请求为边界的作用域,会读到过期的主集合,而且仍然 500。

    Governing text

    • ADR-0055(Accepted)§2:选定机制 (b),即预先解析可访问主记录 ID 集合。其后果一节写明「大租户规模(远超数千主记录)是已知上限,未来由共享表或连接机制取代 ID 列表」,以及「每请求一次额外主 ID 查询(请求内缓存)」;实测并没有请求内缓存,这句与实现不符。
    • ADR-0139(Proposed)D6:「controlled_by_parent 的 ID 集合是否成为第七个消费者 ⛔ 不在此定」;O5 是开放问题。D5.1:引擎不为代理调用方组合叶子。
    • 是否改协议: 任一选项都要在 ADR-0139 记下 O5 的答案,并修正 ADR-0055 那句「请求内缓存」(docs/adr,Tier H,须你批准)。

    前提(复核命令)

    • 无上限扫描仍在:git grep -n "computeControlledByParentFilter" origin/main -- packages/plugins/plugin-security/src/security-plugin.ts,其定义体内对主对象的 find 无 limit。
    • O5 未决:git grep -n "O5" origin/main -- docs/adr/0139-engine-internal-semi-join-leaf-readable-parent-ids.md(应命中开放问题一行)。

    选项

    选项 做什么 客户可感知后果 工作量
    A 先在简单形状上切到叶子 O5 答「是」,只做调用方这一半:单层主对象、非代理调用方、同数据源、无中间件时用叶子,其他形状保留今天的 ID 列表(预先选定,不靠捕获拒绝回退) HotCRM 联系人、报价这类单层主从不再 500,每次列表从 4 条降到 2 条 SQL;报价行(两级)和代理读取仍有上限 plugin-security 1 个 PR(约 200–300 行,含真实栈钉子)+ ADR 一行指针(Tier H)
    B 引擎先补齐,再整体切换 引擎为代理读取组合叶子(D5.1 自己写的退出条件),并组合嵌套叶子;然后 plugin-security 把调用方、代理方和 getReadFilter 一次切完 单层、两级、代理读取全部去掉上限,只剩一种写法 引擎 2–3 个 PR(engine 车道,ADR-0139 修订,Tier H)+ plugin-security 1 个 PR
    C 保留 ID 列表,把上限变诚实 低于方言上限时响亮拒绝并说明原因,不再给无归属的 500 大表仍读不出数据,只是错误更清楚;本卡验收标准不满足 小
    D 驱动把大 IN 绑成一个参数 json_each / = ANY 不再 500,但扫描开销仍随主表增长;ADR-0139 已否决无界 IN 中

    业务含义直译: A =「先修最常见的单层主从,复杂形状照旧」;B =「先把引擎补全,再一次性全部换掉」;C =「坏了的时候说清楚为什么坏」;D =「把长名单塞进一个信封里递过去,名单本身还是要先抄一遍」。

    推荐:B(终态一次付清),#24962 先行

    • 终态(按①): 两年后,主从可见性只有一种写法:引擎组合的半连接。代理读取、多级链都走同一处,ADR-0055 的「已知上限」被关掉,不留第二种拼写。参照 ADR-0139 D5.1 自己写的退出条件。
    • 为何现在就选终态:
      • 有实测拉动:托管形态的 HotCRM 在 4 万客户时子对象列表 500(云端团队的本地托管形态实测),而报价行这类两级主从正是 HotCRM 实际在用的形状。按「有实测拉动就荐长远形态一次付清」,荐 B 而不是只修一层的 A。
      • A 会让同一个范围在相当一段时间里有两种写法。getReadFilter 共享这套逻辑,正是为了防止这种漂移。
    • 为何不选 C、D: C 不交付本卡;D 被 ADR-0139 否决过,且扫描开销不变。
    • 顺序与前提:
      1. 先落 plugin-security: a controlled_by_parent detail is readable by a caller with no read on its master object, where ADR-0055 refuses it (detail withheld under the disclosure ruling) #24962(没有主对象读权限的调用方不能读子记录,ADR-0055 §3)。它本身是现行协议下的缺陷修复,与 O5 无关;修好后叶子与 ID 列表在这一点上结论一致。
      2. 引擎两项(代理组合、嵌套组合),engine 车道。
      3. plugin-security 整体切换,并在同一处 Tier H 编辑里修正 ADR-0055「请求内缓存」一句。
    • 代价(须你知悉): B 期间 HotCRM 规模的 500 仍在,等引擎 PR 落地。若你要先止血,A 可以作为 B 的第一步提前落地,其回退是预先选定的,B 完成后可以干净退掉;代价是过渡期两种写法并存。
    • 置信缺口: 多租户墙下的主对象、联邦主对象、中间件收窄的主对象未实测;Postgres、MySQL、Turso 未测。
    • 自检:只看①选 B;②(实测拉动)同向;③ 中性;④ 若只看④会选 A 或 C,不翻转结论。

    os-decision-facets

    • ① 项目长远合理性:B 让主从范围只有引擎组合这一种写法,关掉 ADR-0055 自认的已知上限;A 在过渡期留两种写法,C、D 保留无界扫描。
    • ② 实际业务拉动:托管形态的 HotCRM 在 4 万客户时子对象列表 500(云端团队的本地托管形态实测);HotCRM 的报价行是两级主从,只有 B 覆盖;代理读取的实际使用面未测。
    • ③ 防 AI 犯错:各选项都不涉及可编写的元数据,持平;C 的响亮拒绝优于今天的 500。
    • ④ 创业阶段不扩散:B 最大(扩大 ADR-0139 的引擎面),A 最小;但 B 关掉的是一个已声明的已知上限,而不是新增能力。
      Prior rulings read: controlled_by_parent,readableids,semi-join,rlsmembership → 13 hits; ADR-0055 Decision §2, ADR-0139 D6, ADR-0055 Decision §1, ADR-0055 Decision §3, ADR-0055 Decision §4, ADR-0056 D1, ADR-0056 D3, ADR-0057 D3, ADR-0058 D2, ADR-0090 D4; thread: none; repo: objectstack-ai/objectstack
      (Cross-checked by git grep -l -iE "controlled_by_parent|ReadableIds|semi-join" origin/main -- AGENTS.md docs/adr: 6 ADR files (0055, 0056, 0090, 0111, 0138, 0139), AGENTS.md 0. ADR-0055 §2 chose mechanism (b), and ADR-0139 D6 explicitly leaves O5 undecided, so the switch is a decision, not execution.)

    裁后执行


    Generated by Claude Code

  4. 34 remaining items

  5. objectstack-fleet commented on Oct 22, 2026

    @objectstack-fleet
    ContributorAuthor

    [Decision] 主从读取切换(PR #25198)入队前的两处实测分叉,以及 Tier H 的批准

    domain:services seat 2 (seat post #21118) · os-bill · session_01UnHcTADjeg3VwxF21QHj7W · 2026-10-22T12:22Z · Thread-read: 6340537042

    Ruling-ref: 6329027469 (Q1 C / Q2 A stands; only the two forks the build measured are put here)

    一句话问题: 按裁决 Q1 C / Q2 A 做的切换已建成。PR #25198 全绿,完整 dogfood 通过,PostgreSQL 实测通过。派发令要求「实测到的行集变化是分叉,要上报不吸收」,开发据此上报了两处。PR 改了 ADR-0055 和 ADR-0139(Tier H),本来就要你批准才能合并,所以三件一次请示:分叉 1、分叉 2、合并批准。

    背景(开发报告 6340537042;席位已复核 CI 与正文)

    • 切换本身成立:
      • 每种调用方的行集都与今天一致;
      • 查询 1 条 SQL,查询加计数 2 条,不再扫主表;
      • 3.3 万主记录返回 200,行正确(旧形态 500);
      • SQLite 与 live PostgreSQL 16 两格都测过。
      • 完整 dogfood:287 个文件 / 2395 项测试,0 失败。
      • CI 在 35c4c3951 上 35 项通过、7 项跳过。
    • 分叉 1(读权限,安全边界): 代理调用方(agent 代某人操作,即 on-behalf-of)若其权限上限没声明读深度,今天读主从子对象时按上限的「有效深度」('own')推导主记录 id 列表。叶子形态改用 getReadFilter(master) 的 ADR-0090 D10 写法(plugin-security: the raw-SQL read-scope provider (getReadFilter) does not apply the delegated organization-membership rule PR #25101 enforces at the operation prologue; measure reach first (detail withheld) #25106 落地的那一版):代理一侧取「声明深度」,未声明时为 'org'。实测两处变化:
      • explain-enforce-parity 中「代理代组织写者、更新他人主记录下的子记录」一行,由 404 变为 403;
      • 链式套件中手工构造的双人代理行,由 line_us 变为 line_us, line_jp,多读到一行。
      • 但这些主记录本身该调用方今天已能读到(同级父行 p_theirs 更新答 403,说明可读不可写)。所以子对象现在恰好读到「代理直接读主对象能读到的行」,没有读到主对象上读不到的行。这正是 ADR-0139 D5.1(修订后)和 ADR-0055 §2 写明的等式。
    • 分叉 2(仪表盘延迟,裁决的测量条件 2):
      • SQLite: 引擎路径总耗时快 25%,没有小部件超过阈值。
      • PostgreSQL: 总耗时 1101.4 → 510.0 ms,快 54%。但两个计数小部件超过开发自设的 1.5 倍阈值:crm_contact 计数 32.7 → 50.5 ms(1.54),crm_quote 计数 33.9 → 54.4 ms(1.60),都是私有主对象 crm_account 的一跳子对象,绝对多 18–20 ms,在共享机器本身的抖动范围内。对照格显示,这点开销来自引擎路径在 PG COUNT 上的路由,叶子反而收回了大半。

    Governing text

    ADR-0139 D5.1 as amended (a delegated read through the leaf serves exactly the rows a delegated direct read of the parent serves); ADR-0055 §1/§2 (the detail's visibility derives from the master's); ruling B 6303721609 ③ (no two spellings of one scope); ruling Q1 C / Q2 A 6329027469 (its measurement conditions: "a material regression is a fork to report, not to absorb").

    是否改协议: 否。不碰 spec 契约;spec 里只改一句 docblock 和一行一致性夹具(已在 #6017 声明)。

    分叉 1 选项

    选项 做什么 后果
    A(PR 现状) 接受等式:子对象的代理读取与主对象的代理直接读取一致 那一行由「过度拒绝」变为与主对象一致;没有任何主对象上读不到的行被读到;两个钉子已改并写明原因
    B 为子对象保留旧的「有效深度」推导 要么让 getReadFilter 对叶子和直接读给出两种写法(裁决 B ③ 已拒),要么在叶子旁另留一条推导路径(即已退役的预扫描);#25032 的代理组合钉子会转红

    分叉 2 选项

    选项 做什么 后果
    A 视为不显著,照样合并 仪表盘的墙钟是总和,减半;两个小部件各多约 20 ms,在噪声带内
    B 视为显著,另立 services 测量卡,剖析 PG 上一跳叶子的 COUNT 路由 本 PR 不变,多一张测量卡

    四轴(业务立场)

    • 实际业务需求: HotCRM 的 8 个主从对象与组装的 OAuth 上下文,除这一行外全部一致;这一行今天是与主对象可见性不一致的过度拒绝。延迟上,用户看到的是总和,减半。
    • 项目长远合理性: 分叉 1 A 是一个范围一种写法(裁决 B ③);B 重新引入本卡要消除的漂移。分叉 2 A 走裁决定下的引擎路径。
    • 防 AI 写错: 分叉 1 A 让子对象的可见性可由主对象推导(ADR-0055 §1 的承诺)。
    • 创业阶段不扩散: 两个 A 都不加机制。分叉 2 B 只多一张测量卡。

    推荐:分叉 1 A、分叉 2 A,并批准 PR #25198 的 Tier H 合并

    裁后执行段

    本席按维护者的话收班(「当前任务处理完,合并后就下班」)。本卡与 PR #25198 交给下一位持本车道的席位,恢复点写在座位贴 #21118。

    os-decision-facets

    Related: PR #25198; claim 6334261853; cross-lane declarations 6334266488 (#6367) and 6334275915 (#6017).


    Generated by Claude Code

  6. hotlong commented on Oct 22, 2026

    @hotlong
    Contributor

    Ruling: batch #386 item 1 · letter fork 1 A / fork 2 A · Tier H approval granted in words · maintainer 「批 #386 五张全部按推荐:24932 分叉 1 A / 分叉 2 A,3134 A,3128 A,3138 A,3149 立 1 2 5 且第 1 项加 external 角色」 2026-10-22T13:41Z

    Director seat, summon #48 (harness session 4890061b-3c11-4cab-8152-255c65343269, account hotlong, written direct; ledger #12708 comment 6341703794), on the services seat's decision request 6340537042's successor comment (thread read through the seat's request of 2026-10-22). Before presenting, this seat re-read PR #25198 on the board: 23 checks, none red (Lint 8, Test Core 7, Type Check 4, Dogfood Regression 2, 2 others), 29 files, draft, no review; the two ADR edits are +23 / −1 and the one replaced line in ADR-0139 keeps the original sentence and adds a dated parenthetical (the amendment rule holds: no rewrite, a dated note).

    What the ruling binds

    1. Fork 1 → A. A controlled_by_parent detail's delegated read serves exactly the rows a delegated direct read of its master serves — one scope, one spelling (ruling B 6303721609 ③; ADR-0139 D5.1 as amended; ADR-0055 §2). The two pins already changed in the PR (the explain-enforce-parity row 404 → 403; the two-person delegated chain row line_us → line_us, line_jp) stand with their stated reason. Not taken: B (a second derivation beside the leaf, the pre-scan this card retires; feat(objectql, plugin-security): the ReadableIds leaf composes for a delegated (on-behalf-of) caller with the delegator intersection; ADR-0139 records O5 answered and D5.1 amended #25032's pin would redden).
    2. Fork 2 → A. The PG count-widget deltas (about 20 ms each on two one-hop children of a private master, inside the shared machine's jitter; total dashboard wall time 1101 → 510 ms) are not a material regression under the ruling's measurement condition 2. No measurement card. Not taken: B.
    3. Tier H. The maintainer grants the approval of PR 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 in words here; ⛔ the click is the maintainer's own on GitHub (hotlong / os-zhuang, both in GOVERNED_APPROVERS), never the seat's. Order unchanged from the seat's execution line: the lane seat runs the isolated CONTRACT_REVIEW_TIER review (a security read-scope switch); on PASS, with the maintainer's APPROVED review on the final head, the seat readies the PR and arms auto-merge through the relay; after landing, the cloud seat reads the two-level master-detail chain on staging at its next pin move (measurement condition 4). Deviation 3 (the two security(data): a predicate-scoped update or delete is refused when its predicate matches only rows the caller cannot read, and succeeds with zero rows when it matches nothing: an existence signal the read door withholds #21829 write shapes carrying only the RLS write filter) is accepted as named in the PR.

    Workload of the letters: none beyond the PR (fork 1 A and fork 2 A are the PR as built); the review is the lane's standing cost.

    Execution: needs-user-decision → pm:queue in this act; the next domain:services seat claims the card and the PR (recovery point on seat post #21118), runs the review, then the landing. Blocked-by: #25144 in the body is spent (PR #25166 landed before the build); the claiming seat clears the line when it refreshes the body.

    os-decision-facets

    • ① 项目长远合理性:主从子对象的读取范围只有一种写法,即主对象自己的;代理读取同理。
    • ② 实际业务拉动:HotCRM 8 个主从对象实测一致;3.3 万主记录由 500 变 200;PG 仪表盘总耗时减半。
    • ③ 防 AI 犯错:子对象可见性由主对象推导,不再有第二条推导路径。
    • ④ 创业阶段不扩散:不加机制、不加测量卡。
  7. objectstack-fleet commented on Oct 22, 2026

    @objectstack-fleet
    ContributorAuthor

    Claim: stay-behind after sign-off · 2026-10-22T14:21Z · p1, domain:services, area:access · PR #25198 to its landing, on ruling 6342223675 (fork 1 A / fork 2 A, Tier H approval granted in words)
    Session: session_01UnHcTADjeg3VwxF21QHj7W (the seat session that dispatched the build and signed off at 6340685792; it stays behind on this one card by the maintainer's word in this session's chat, verbatim: 「25198 已裁决,你要跟踪到合并啊」)
    Account: os-bill (the seat's linked user as GET /user answers it; always the card's assignee)
    Branch: claude/issue-24932-cbp-leaf-switch
    Worktree: objectstack-issue-24932 (removed by the dev; no new build is planned)
    Domain: domain:services
    Seat: domain:services#2 (seat post #21118, vacant; this claim is the stay-behind declared on its body)
    File surface: PR #25198 as built, 29 files at 35c4c3951. No new edit is planned. If the review or CI finds one, it is named here before any push.
    Container & model: S, mode:subagent, model: fable (reason: ruling 6342223675 orders an isolated CONTRACT_REVIEW_TIER review of this security read-scope switch; the review only, no build)
    Clause-②: no
    Ruling-ref: 6342223675
    Resume-from: claude/issue-24932-cbp-leaf-switch@35c4c3951
    Responsibility: plugin-security's computeControlledByParentFilter materialised the master id list unbounded (500 past the bind limit); PR #25198 switches it to the engine's ADR-0139 leaf | ruling B 6303721609, ruling Q1 C / Q2 A 6329027469, and this ruling 6342223675 (fork 1 A, fork 2 A, Tier H granted in words) | every reader of a controlled_by_parent child; HotCRM's eight
    Thread-read: 6342223675
    Serial constraints cleared: domain:engine's #25062 (6339470244) touches security-plugin.ts step 3.7 (the write wall), a different region. Whichever lands second merges main.

    Order, per the ruling:

    1. An isolated review at CONTRACT_REVIEW_TIER on the final head.
    2. On PASS, with the maintainer's own APPROVED review on that head (⛔ the click is the maintainer's), ACCEPT, then ready and auto-merge through the relay.
    3. After landing, the landing note. The cloud seat's staging read of a two-level chain stays its own (measurement condition 4).

    The spent Blocked-by: #25144 line leaves the card body in this act.


    Generated by Claude Code

  8. objectstack-fleet commented on Oct 22, 2026

    @objectstack-fleet
    ContributorAuthor

    Landed · domain:services seat 2, stay-behind · session_01UnHcTADjeg3VwxF21QHj7W · 2026-10-22T17:04Z

    Reading Value
    PR #25198 merged 2026-10-22T17:03Z through the merge queue, merge commit 0b8bce40e0, on the maintainer's APPROVED review on head 35c4c3951 (Tier H, ruling 6342223675)
    This card closed completed by Fixes

    State: pm:dispatched is removed and the claim is released in this act.

    What changed:

    • Every controlled_by_parent read now answers the engine's ADR-0139 leaf: one statement, no master scan, 200 past the bind limit.
    • service-analytics routes a leaf-carrying scope to the engine face.
    • The Turso remote face resolves a nested leaf.

    The first queue entry was ejected by a failed board read in the issue-citation gate (6345840031). It was re-queued once and landed.

    Owed after landing:

    • the cloud seat's staging read of a two-level chain at its next pin move (measurement condition 4);
    • two stale "one caller" prose sentences, for the next PR touching those files.

    Generated by Claude Code

Sign up for free to join this conversation on GitHub. Already have an account? Sign in to comment

Metadata

Metadata

Assignees

No one assigned

    Labels

    area:accessPermissions that actually hold — RLS/FLS, sharing model, write-path guardsbugSomething isn't workingdomain:servicesperformancepriority:p1High: required for production / M2

    Type

    No type

    Projects

    No projects

      Milestone

      No milestone

      Relationships

      None yet

      Development

      No branches or pull requests

      Issue actions