[Decision] security(objectql): may a hook's `handler` name bind to a function another package registered (the engine-wide fallback HookSchema.handler declares), or does name resolution stay inside the hook's own package (#21585 option B)

[objectstack-fleet]

Ruled: 5974477722 · letter B · 2026-10-03T23:13Z

Filing gate: ② a decision only the maintainer can make. It is a security boundary, and the resolution rule is declared in the spec, so changing it is a contract change. Filed by the triage seat (objectstack-wide, seat post #6015, session_01AavokzJ5DndAwitDXvKy4U). It carries option B of the pm:retriage on #21585 (5970562606). #21585 itself is answered A in this round, which closes the install-local door only. ⛔ Not a claim, ⛔ not a dispatch. ⛔ Classes, doors, positions and functions only.

Reader who acts: the maintainer, or the director seat. If the letter is B, domain:engine carries it, and the spec doc moves in the same PR.

维护者速读

• 问题： 应用包里的「钩子」可以只写一个函数名，不写函数体。运行时先在这个包自己的函数里找；找不到，就去整个运行时的函数表里按名字找，而这张表不分是哪个包注册的。结果是：A 包的钩子可以绑上 B 包的同名函数，B 的代码会在 A 的数据事件上运行。[finding] os package install accepts a package whose hook uses only the deprecated function-name handler (no body), answers "installed", and the hook never fires: install-local drops it with a server-side warn only #21585 实测了这一点（安装成功、重启后依旧）。[finding] os package install accepts a package whose hook uses only the deprecated function-name handler (no body), answers "installed", and the hook never fires: install-local drops it with a server-side warn only #21585 这轮按 A 处理，只把「本地安装」这一个入口堵上。别的入口（元数据接口编写的钩子、多应用组合）还按名字跨包找，而 spec 文档明确声明了这种跨包查找。
• 选项：
  • A： 维持跨包按名查找，并在文档里写明这是有意的共享能力。不改契约。
  • B： 名字只在钩子所属的包内解析（包括这个包自己运行时模块注册的函数）。跨包复用必须显式写出所属包。spec 的声明随之改写。
• 分诊推荐： B，理由只看①。回退项为 A。
• 你要做的： 回一个字母。

Measured

• Cross-package binding is real ([finding] os package install accepts a package whose hook uses only the deprecated function-name handler (no body), answers "installed", and the hook never fires: install-local drops it with a server-side warn only #21585 dev report 5970542332, on main 6c5697dffb, through the public door): a package's handler-only hook ran a function that another app registered. This held hot and after a restart. A unit reading on one engine agrees: the resolved function's packageId is the other app's.
• Mechanism, read now on main b610eabf72:
  • packages/objectql/src/hook-binder.ts, resolveHandler (about :321–:322) tries the bundle's own functions, then engine.resolveFunction(name).
  • The engine's function map is keyed by bare name. The stored owner is not consulted on lookup.
• Declared: packages/spec/src/data/hook.zod.ts (about :318), HookSchema.handler's doc, says the name may resolve to "anything engine.registerFunction(name, fn) added".
• Pull for cross-package resolution, measured now: zero known writers.
  • objectstack examples/**: no hook names a string handler. The only string handler is a job's, which resolves within its own bundle.
  • hotcrm src/**: no hook with a string handler.
  • Inside packages/**, registerFunction is called only by the engine, the hook binder and the formula stdlib.
• Who can reach it: a holder of the metadata-management capability, either by installing a package or by authoring a hook through the metadata door. The open question is whether a package's content, a third-party catalog package above all, should be able to run code it did not ship.

Governing text

• HookSchema.handler's doc (above): it declares the engine-wide fallback. So B is a contract change, not a defect fix.
• [Decision] install-local: a package's declared jobs are never scheduled — refuse the install, name them in the install answer, or make job handlers declarable bodies (the jobs half of #21322) #21489's ruling C (applied to hooks on [finding] os package install accepts a package whose hook uses only the deprecated function-name handler (no body), answers "installed", and the hook never fires: install-local drops it with a server-side warn only #21585): install-local refuses what it cannot run on its own, loudly. It covers one door.
• [Decision] security(runtime): may an app-authored body touch the stored-metadata family's tables at all — a hook bound to them, or an elevated body writing them directly (#21454 items 3 and 4) #21520's ruling A (5965059068): app bodies stay out of what is not theirs. The same principle, applied to another surface.
• AGENTS.md Prime Directive 12: "fix it at the producer and reject it at authoring/publish".
• Prior rulings read: cross-app function, resolveFunction, owner-scoped, handler fallback → 1 hit ([finding] os package install accepts a package whose hook uses only the deprecated function-name handler (no body), answers "installed", and the hook never fires: install-local drops it with a server-side warn only #21585, this card's source); ADR none; thread: [finding] os package install accepts a package whose hook uses only the deprecated function-name handler (no body), answers "installed", and the hook never fires: install-local drops it with a server-side warn only #21585.

一句话问题

一个应用包的钩子，能不能靠「同名」去执行另一个应用包的代码？

选项 × 真实代价

选项	做什么	客户可感知的后果
A 维持	什么都不改，只在文档里写明「按名字跨包共享」是有意的	无变化。装一个第三方包，它可以借同名触发你另一个应用的代码，平台不提示
B 包内解析	名字只在钩子所属的包内解析，跨包必须显式写出包。改 spec 声明，发一个收窄版本	实测今天没有人依赖跨包按名绑定，所以不影响任何人。以后误写的人会在安装或注册时看到报错，指出名字在本包里找不到

业务含义直译

• A 等于大楼里任何一家租户喊一声「保洁」，来的可能是别家公司请的保洁，而且这件事没人登记。
• B 等于每家只能叫自己合同里的人；要借别家的人，得写明借谁家的。

四轴（业务立场）

• 实际业务需求： 实测零个跨包按名绑定的写法（examples 和 hotcrm）。零拉动默认移除，偏 B。
• 项目长远合理性： B 让「代码属于哪个包」在解析时真正生效，和 [Decision] security(runtime): may an app-authored body touch the stored-metadata family's tables at all — a hook bound to them, or an elevated body writing them directly (#21454 items 3 and 4) #21520「应用代码不碰不属于自己的东西」同一方向。A 把一个隐式的跨包通道变成永久契约。
• 防 AI 犯错： 出错时谁看到什么？A 下，AI 写错函数名，或恰好与别的包撞名，会静默执行别人的代码。B 下，作者在注册时立刻看到「本包没有这个函数」。
• 创业阶段不扩散： B 收窄一个声明过的能力，A 是声明并长期维护一个跨包通道。偏 B。

os-decision-facets

• ① 项目长远合理性：B 让函数归属在解析时生效，收缩一个隐式跨包通道；A 把它固化成契约。
• ② 实际业务拉动：今天零个跨包按名绑定的写法（examples、hotcrm 实测）。零拉动默认 remove，偏 B。
• ③ 防 AI 犯错：B 在注册时响亮拒绝；A 撞名时静默执行他包代码。偏 B。
• ④ 创业阶段不扩散：B 收窄声明，A 维护一个跨包通道。偏 B。

推荐：B。 两年后的样子：每个包的代码按包解析，跨包调用都是显式、可审计的引用。参照 Salesforce：托管包的 Apex 类带命名空间前缀，跨包调用要写全名。
回退：A。 只改文档，写明跨包共享是有意的；#21585 的 A 已经堵住本地安装这个入口。
自检： 只看①选 B；②③④ 是否翻转：否。三轴同向。
置信缺口： 我看不到私有客户的多应用组合。一个 defineStack 里多个应用共用一组函数，按名字跨包绑定，这种写法在示例里没有，但在客户那里可能存在。这正是 B 的认领要先普查的。常设规则「新门默认否」偏向 A，所以维护者不回字时按 A 处理。

裁后执行

• A： 本卡以 not planned 关闭。HookSchema.handler 的文档改写一句，把跨包共享写成有意的能力（domain:spec，一个小 PR）。
• B： 本卡转 pm:queue（domain:engine），并在同一个 PR 里改 spec 声明（作为声明的跨车道路径）。
  • 认领先普查所有依赖跨包按名解析的组合：examples、hotcrm、cloud，以及每个 --artifact 运行时模块。
  • 解析时比对函数的所属包。包自己的运行时模块注册的函数照常可用。
  • 名字在本包里找不到，就在注册时给出一个点名的拒绝。
  • 按「收窄」规范发版（!、Clause-②: yes (narrowing)、ADR-0087 标记、minor）。
  • ⛔ 不碰 [finding] os package install accepts a package whose hook uses only the deprecated function-name handler (no body), answers "installed", and the hook never fires: install-local drops it with a server-side warn only #21585 已落的本地安装拒绝。

[objectstack-fleet]

Ruling: batch #275 item 1 · letter B · maintainer 「同意」 2026-10-03T23:11Z

Director seat, summon #32, session_016tKoy8NJa35Yih1FdzrVmn. Written as objectstack-fleet[bot] through the relay. ⛔ Classes, doors, positions and functions only.

• How it was ruled. Batch 🔗 Broken links detected in documentation #275 was presented in the live director chat with options A and B. The director recommended B, as triage did. The maintainer answered the batch as presented: 「同意」.
• Freshness gate: this card has no comment since filing.

The ruling

B: a hook's handler name resolves inside the hook's own package only.

• That includes the functions the package's own runtime module registers.
• Cross-package reuse must name the owning package explicitly.
• A name the package does not hold is refused at registration, with a refusal that names it.
• HookSchema.handler's declaration changes in the same PR. Its doc stops declaring the engine-wide fallback ("anything engine.registerFunction(name, fn) added").
• Not taken: A, which keeps cross-package lookup by name and documents it as intentional. That would freeze an implicit cross-package channel as a contract.

四棱(本裁决新记录)

• ① 长远:函数归属在解析时生效;跨包调用都是显式、可审计的引用;与 [Decision] security(runtime): may an app-authored body touch the stored-metadata family's tables at all — a hook bound to them, or an elevated body writing them directly (#21454 items 3 and 4) #21520 A「应用代码不碰不属于自己的东西」同向。
• ② 拉动:跨包按名绑定实测零处(examples、hotcrm)。
• ③ 防 AI:撞名或写错名时在注册处响亮拒绝,不再静默执行他包代码。
• ④ 不扩散:收窄一个声明能力,不新增机制。
• 只看①选 B;②③④ 是否翻转:否。

Execution parameters (ruled here; no further decision card)

• This card: needs-user-decision → pm:queue in this act, domain:engine, p2, security.
• The claim's first step is a census of every composition that relies on cross-package resolution by name: objectstack examples/**, hotcrm, cloud, and every --artifact runtime module. A real dependent found there goes back to the maintainer before any refusal lands.
• The change:
  • resolution checks the function's owning package;
  • the package's own runtime-module functions keep resolving;
  • a name missing from the package is refused at registration, with the name and a prescription.
• Release: published as a narrowing (!, Clause-②: yes (narrowing), the ADR-0087 marker, minor), with an at-tier contract review.
• ⛔ [finding] os package install accepts a package whose hook uses only the deprecated function-name handler (no body), answers "installed", and the hook never fires: install-local drops it with a server-side warn only #21585's landed install-local refusal is untouched.

---

Generated by Claude Code

[objectstack-fleet]

Claim: PM loop round 27 · 2026-10-03T23:31Z
Session: session_017ErfyP2Rx7XWHJA27QjyUi
Account: os-project-manager (the seat's linked user as GET /user answers it; always the card's assignee)
Branch: claude/issue-21604-hook-handler-package-scope
Worktree: objectstack-issue-21604
Domain: domain:engine
Seat: domain:engine#1
File surface (at origin/main 15fe567c9c), per the maintainer's ruling 5974477722 (letter B, 「同意」) and its execution parameters:

• packages/objectql/src/hook-binder.ts: resolveHandler's fallback to engine.resolveFunction(name).
• packages/objectql/src/engine.ts: the function registry's lookup (registerFunction, resolveFunction / resolveFunctionEntry, around :3877–:3910), only as far as owner-scoped resolution needs it.
• Declared cross-lane path (domain:spec), as the ruling routes it: packages/spec/src/data/hook.zod.ts, the HookSchema.handler doc (around :314–:320). It stops declaring the engine-wide fallback. The declaration is on domain:spec seat 1's post ([PM seat] domain:spec — ⏳ vacant · last incumbent session_01T9u38rswFp5Rw8DswRUReJ #6017).
• Tests; .changeset/21604-*.md.
• Census first, as the ruling orders: every composition that relies on cross-package resolution by name.
  • examples/** and every --artifact runtime module in this repository;
  • hotcrm and objectos, both public and read-only;
  • cloud: NOT MEASURED if it is unreachable, and named as such.
  • ⛔ A real dependent found goes back to the maintainer before any refusal lands. The claim stops and reports it.
• ⛔ [finding] os package install accepts a package whose hook uses only the deprecated function-name handler (no body), answers "installed", and the hook never fires: install-local drops it with a server-side warn only #21585's landed install-local refusal is untouched.
  Container & model: M, mode:subagent, model: default.
• dispatch-gates --tier names hook.zod.ts as a clause-② suspect surface.
• Build at the default tier. The contract review is at tier (CONTRACT_REVIEW_TIER) after delivery, as the ruling requires.
  Clause-②: yes (narrowing)
• This is the ruling's release parameter, verbatim: "published as a narrowing (!, Clause-②: yes (narrowing), the ADR-0087 marker, minor), with an at-tier contract review."
• The path leg hits as well (packages/spec/src/**).
  Thread-read: 5974477722
  Serial constraints cleared: at 2026-10-03T23:31Z:
• Of the 5 open PRs, none touches hook-binder.ts or data/hook.zod.ts.
• No in-flight claim in this lane names either file. This lane's one in-flight card, Two stored view containers of one object still displace each other's views: a container bound elsewhere or unbound under a sibling's expanded name, and a second container's bare list taking the first's <object>.default, are accepted with no diagnostic #21639, is on metadata-protocol's save door.
• North Star 「优先级」 clause 3: this lane holds no open p0 or p1.

[objectstack-fleet]

os-dev-report
{
"issue": 21604,
"status": "done",
"branch": "claude/issue-21604-hook-handler-package-scope",
"pr": "#21653",
"session": "session_017ErfyP2Rx7XWHJA27QjyUi (subagent: the dispatching PM's session; identity is the branch)",
"premise_still_valid": true,
"summary": "Ruling B executed. A hook's string handler now resolves only against the functions handed to its own bind (the package's functions, an --artifact runtime module's included), then against engine entries whose packageId equals the bind's packageId (hook-binder.ts resolveHandler and ownPackageFunction, reading the existing resolveFunctionEntry). A name the package does not hold, whether another package registered it or nobody did, is refused at registration with INVALID_REFERENCE / 400: recorded on BindHooksResult.errors (which gains optional code and status), logged at error, thrown under strict, and the hook is not bound. HookSchema.handler's TSDoc stops declaring the engine-wide fallback. The census came first and found zero dependents in examples, hotcrm and objectos; cloud is NOT MEASURED (unreachable). No stop condition fired: no dependent, and no new spelling minted. Install-local (#21615) is untouched. Draft PR #21653 is open, with a minor changeset for objectql and spec, the BREAKING banner, Clause-② yes (narrowing), and ADR-0087 not-required (no-migration-prescription).",
"census": [
"objectstack examples/** @ 15fe567 | string handler:, registerFunction, functions:, composeStacks, hook forms | 0 dependents (the only string handler is a job's, app-showcase sweepProjectHealth, resolved in its own bundle; showcase's 5 hooks are body; crm 1 and todo 1 inline-function hooks lowered to their own names in their own runtime module; app-multi-package and embed-objectql declare no hooks or functions)",
"objectstack --artifact runtime modules @ 15fe567 | tracked objectstack-runtime*.mjs, tracked non-TS files naming runtimeModule | 0 committed (runtime modules are build outputs of each app's own config)",
"hotcrm @ f24c196588 (depth-1 clone, read-only, deleted) | same greps | 0 (no string handler:, no registerFunction; 19 hook files, all inline functions in one composeStacks app, one owner)",
"objectos @ 7612ffebd1 (depth-1 clone, read-only, deleted) | same greps | 0 (no hits)",
"cloud | NOT MEASURED: git clone has no credentials; gh api answers 403 'not enabled for this session'; add_repo (read) answers no access"
],
"hypotheses": {
"H1": "Confirmed at 15fe567. resolveHandler tried the bundle, then engine.resolveFunction(name), which reads a bare-name map. FunctionEntry carries packageId, and resolveFunctionEntry exposes it, so the binder needed only a comparison. engine.ts got doc comments only.",
"H2": "Per door, the hook's package is the owner the binder binds under. Boot, AppPlugin and --artifact: app:APPID, with own functions and the runtime module in the same bind. Install-local: app:APPID, handler-only hooks withheld (untouched). Metadata door: the synthetic owner metadata-service, which registers no function, so every handler-only authored hook is refused at bind. Today that door resolves any engine function: under the ablation the authored row came back stamped by app X's function. The PUT itself answers 200 (recorded). A bind with no packageId resolves only what it is handed. A metadata-door row stamped with package_id P is still bound under metadata-service, so it no longer reaches P's runtime-module functions. That reading is flagged in the PR (flags 5 and 6) and offered below as an open question rather than silently settled.",
"H3": "Formula-stdlib leg falsified: packages/formula registers into a cel-js Environment, not the engine. The engine registers no functions itself. The only first-party engine registrations are the binder's functions preload, stamped with the bind's owner. Measured after a platform boot (ObjectQL, sqlite-wasm, Hono, one app, platform-objects, auth, security, sharing, REST, dispatcher): the registry holds exactly one entry, the app's own (h3_fn, owner app:com.h3.probe). The set of platform functions a hook may name is empty, before and after.",
"H4": "Refused at bind time, in bindHooksToEngine's unresolved branch. No coded refusal existed for this condition. INVALID_REFERENCE / 400 is a standard-catalog member, so no ledger code was added and nothing widens. PERMISSION_DENIED / 403 from the sibling stored-metadata refusal was not reused: it is a different condition.",
"H5": "Re-measured at the base. Zero writers, as the census rows show."
},
"tests": "At 1eb671b (the later merges b43c6fe and 308ae94 bring only service-analytics and cli files, no overlap). Build order: turbo build --filter=@objectstack/runtime^..., then --filter=@objectstack/dogfood^... --filter=@objectstack/rest --filter=@objectstack/service-automation, all under os-verify-lock with VERDICT command-exit 0. Results: objectql local 370 files / 7441 passed, repo 1 / 5 passed, typecheck green. runtime local 319 / 4534 passed, 19 skipped; repo 3 / 751 passed; typecheck green. rest local 260 / 4897 passed, 326 skipped; repo 5 / 177 passed, 1 skipped. service-automation 168 / 2078 passed. dogfood hook files (hook-error-format, hook-refusal-user-facing-marking, hook-runas-fls, webhook-materialization) 4 / 13 passed. spec check:generated: all 15 up to date. New pins: objectql hook-binder-package-scope.test.ts (5 refusals, each asserting code, status and not bound, plus 2 controls) and runtime hook-handler-package-scope.pin.test.ts (composed kernel: multi-app refusal, metadata-door refusal via PUT /api/v1/meta/hook, own-functions control, --artifact runtime-module control). Ablation, committed first at 1eb671b: ablation-replace WRAP removed the owner check (anchor 1 to 0, replacement 0 to 1, blob 9301e0130c to 49bf4c96cc); objectql rebuilt with exit 0; ablation-dist-preflight found the marker in all 4 JS dist files. Readings: objectql 4 red (cross-package, strict, metadata-door, unowned) and 30 green (typo refusal, controls, existing suite); runtime pin 2 red (Y's insert came back '|x-fn'; authored row '|x-fn|authored-body|x-fn') and 2 controls green. Restore: blob back to the HEAD blob, git diff HEAD empty, whole-tree porcelain empty, rebuild, preflight --absent shows the marker gone from all 14 dist files, pins green again (34/34 and 4/4). A first ablation run read the same split but its DTS step failed on an unused parameter; it was rerun with void packageId, and the clean run is the one quoted.",
"gates": "At 308ae94: dispatch-gates --commands --repo objectstack-ai/objectstack, with no paths, derived 91 commands (the dispatch list plus check-empty-changeset x2, release-rehearsal-clone and release-pending-publish self-tests, check:engine-double-contract, check:objectql-double-limit, check:objectui-changeset, check:pm-changeset-deadline-census, check:query-options-erasure, check:stack-collection-maps, check:swallow-census-controls, check:type-check-coverage, check:type-check-debt, check:where-matcher). All 91 ran with exit codes captured before any pipe, and all exited 0. --ran reconciliation: 91 derived, 91 run, 0 NOT-MEASURED, 0 UNRUN (a derived zero). The same 91 also ran green at b43c6fe; there, check:dual-build-cjs-loads first answered PREREQUISITE NOT MET because 8 packages unrelated to this diff had no dist/, and after building them it measured green. Partial sub-readings inside green gates: check:entry-nameability reports no callable export on spec/api-assembled and spec/qa (inherent to those entries); the check:objectui-pin-citations self-test skipped its live round trip (no objectui checkout). Lint, narrowed: eslint --no-inline-config --format json over the 6 touched TS files gives 6 files, 0 errors, 0 warnings at 308ae94. eslint.config.mjs enables no type-aware linting, so no untouched file's verdict can move; repo-wide pnpm lint belongs to CI. NOT MEASURED (CI-only): Test Core shards, Dogfood Regression Gate, Dogfood Verify CLI, Build Core, Temporal Conformance, workspace type-check lanes. CLI integration tier declared to CI (no CLI file touched). CI convergence: in_progress at report time.",
"line_budget": "n/a",
"deviations": "None against the order's rulings. Notes: (1) The commit trailers use AGENTS.md's model-free pair (Claude-Session plus Co-authored-by Claude), and the PR footer uses AGENTS.md's session-URL form; both instead of the harness reminder's spellings, because AGENTS.md takes precedence. (2) The refusal logs at error, where an unresolved name logged warn before; this is boundary flag 1 in the PR. (3) Gates were run twice, at b43c6fe and again at 308ae94 after a second merge of origin/main. The second batch was split because the runner's own timeout cut gate #78 mid-run, and #78 was rerun. (4) Writes: git push x6 (empty probe, 3 commits, 2 merges), pr_create x1, label-write assign x1, this comment x1, all within the four-kind budget.",
"files_changed": [
".changeset/21604-hook-handler-package-scope.md",
"packages/objectql/src/engine.ts",
"packages/objectql/src/hook-binder.ts",
"packages/objectql/src/hook-binder.test.ts",
"packages/objectql/src/hook-binder-package-scope.test.ts",
"packages/runtime/src/hook-handler-package-scope.pin.test.ts",
"packages/spec/src/data/hook.zod.ts"
],
"mcp_calls": "1: mcp__claude-code-remote__add_repo (objectstack-ai/cloud, access read), refused with no access. 0 GitHub MCP calls, no MCP write tool. GitHub reads went through gh api (issue 21604 and its comments; comments 5970542332, 5970562606 and 5965059068; PR 21615 and its files; PR 21653 read-back).",
"api_writes": "3 REST writes, each through the fleet-write relay (POST /repos/objectstack-ai/objectstack/dispatches executed as objectstack-fleet[bot]): pr_create, which is POST /repos/objectstack-ai/objectstack/pulls (PR 21653, draft, body read back byte-identical, 16322 bytes); assign, which is POST /repos//issues/21653/assignees via scripts/pm/label-write.mjs (read-back matches); comment, which is POST /repos//issues/21604/comments via scripts/pm/post-stamped.mjs (this report). git push is not a REST write.",
"open_questions": [
{
"question": "Reading of 'the hook's own package' on the two doors the ruling's words do not spell out. The PR implements A and flags it (boundary flags 5 and 6); no measured composition is affected either way.",
"options": [
"A (implemented): the hook's package is the owner it is bound under. Metadata-door hooks are bound under metadata-service, which holds no functions, so a handler-only authored hook is refused even when its sys_metadata row is stamped with a package_id. A bind with no packageId resolves only the functions handed to it.",
"B: on the metadata door, resolve a row stamped with package_id P against app:P's functions. This needs objectql to learn the runtime's app: owner prefix, and it lets any metadata author claim a package's code by stamping a package id.",
"C: treat 'no owner' as a scope of its own, so an unowned bind resolves unowned engine entries."
],
"recommendation": "A. Real need: zero string handlers in the census, and no first-party door binds without an owner. Long-term soundness: the owner is already the engine's unit of hooks and functions, so no new mechanism is needed. Guarding against AI mistakes: A refuses loudly at bind, while B makes a stamped package id a key to that package's code. No sprawl: B adds a cross-package mapping and C keeps an unowned channel open."
}
],
"out_of_scope_findings": [
"class: c · reach: public door. PUT /api/v1/meta/hook/scope_authored_cross with handler 'x_stamp' and no body answered 200 (the composed pin's printed reading, at 573e9a2 with this PR's change). The door's re-sync then refuses the hook at bind, and it never runs: a runtime-authored hook holds no functions, so the deprecated handler form can never bind on this door. The save door accepts metadata the runtime refuses. Before this PR the same shape answered 200 and was skipped with a warn for an unknown name, or bound across packages. Route: refuse a body-less handler hook at the metadata save door, as the stored-metadata body boundary got its save-side half through HookSchema; here HookSchema cannot do it, because build artifacts legitimately carry handler strings, so the check belongs to the door. The metadata-protocol save door has an in-flight card in this lane (#21639), so this is not fixed in place; for the seat to file. Dedupe words: metadata door saves handler-only hook 200 never binds; PUT meta hook handler without body accepted refused at bind; runtime-authored hook handler string no functions",
"carrier: triage (承接者:无 if none) · noted, not filed. Action.target and a flow script node's config.function still resolve through the engine function registry by bare name (service-automation plugin.ts bridges objectql.resolveFunction). The ruling covers a hook's handler only. Same family on other surfaces; code-read only, no measured reach.",
"carrier: 承接者:无 · noted, not filed. The engine function registry stays keyed by bare name: registerFunction from two packages with one name keeps the later one's entry (code-read, pre-existing). Hooks bound in the same call resolve their own bundle first, so boot binding is unaffected."
]
}

---

Generated by Claude Code

[objectstack-fleet]

os-dev-report — patch round 1
{
"issue": 21604,
"status": "done",
"patch_round": 1,
"branch": "claude/issue-21604-hook-handler-package-scope",
"pr": "#21653",
"head": "b8c4e5edc4",
"previous_head": "308ae946b9",
"session": "session_017ErfyP2Rx7XWHJA27QjyUi (subagent: the dispatching PM's session; identity is the branch)",
"premise_still_valid": true,
"summary": "Pure regeneration for CI step 119 (check:error-status-conformance). The worktree was recreated from the remote branch at exactly 308ae94, and the failure reproduced there: gate exit 1, 'INVALID_REFERENCE: baselined as unpinned, but a producer now declares its status'. The only write was node scripts/check-error-status-conformance.mjs --update (exit 0, 'Baseline rewritten: 24 unpinned code(s)'), which removed the one line INVALID_REFERENCE from scripts/error-status-unpinned-baseline.json. That one file was committed as b8c4e5e, on top of 308ae94, with no merge of origin/main, and pushed as a fast-forward (308ae94..b8c4e5e, no force). Nothing else changed: no source file, comment, changeset or PR body. The ratchet-down remedy is the author's own; the gate's MAINTAINER-ONLY marker covers only the weakening remedy (gate battery 11b).",
"purity": {
"diff_name_only": "git diff --name-only 308ae94 HEAD lists exactly 1 file, scripts/error-status-unpinned-baseline.json (1 file changed, 1 deletion: the line INVALID_REFERENCE).",
"rerun_update": "Re-running node scripts/check-error-status-conformance.mjs --update on the new HEAD b8c4e5e exits 0 ('Baseline rewritten: 24 unpinned code(s)') and leaves git status --porcelain empty and git diff HEAD empty."
},
"gate": "pnpm check:error-status-conformance (self-test, then gate) at b8c4e5e: exit 0, captured before any pipe. Self-test: 70 cases pass. Gate: 'every derivable runtime status is documented, and every documented status is reachable'. At 308ae94, before the regeneration, the gate alone exited 1 with exactly the one ratchet finding.",
"tests": "No code changed, so no suite is owed this round. The reading above is the gate's.",
"gates": "Why the first round missed this: at 308ae94, dispatch-gates placed check:error-status-conformance in its 'Artifact rosters' block (53 families). That block sits outside the runnable 91, and the tool says to run those gates rather than read their silence. Round 1 named the block as outside the total but did not run it. This round, read-only, at b8c4e5e: --commands now derives 100 commands, the 91 plus 9 matched by the baseline path (check:error-status-conformance, check:engine-double-contract, check:cli-command-ids, check:entry-guard, check:parse-guard, check:pnpm-filter-targets, check:agent-test-spelling, check:bash32-floor, check-scripts-symbol-anchors with its self-test). I ran those 9 plus the 52 remaining roster commands, 61 in all. 59 measured with exit 0. The first pass had non-zero exits that were not measurements (no node_modules in the recreated worktree, and annotation text I had failed to strip from some commands); after pnpm install, clean command strings, check:partof-closing-keyword fed PR 21653's body through PR_BODY, and a cached turbo build of the packages check:published-readme-exports reads (92/92 tasks from cache), each of those reran with exit 0. 2 are NOT MEASURED, reason: they need PR context and a token, and run in their own workflows: node scripts/check-closing-target-claim.mjs and node scripts/check-single-claim-paths.mjs. Rows the derivation marks checker-health (self-tests) grade the checker, not this diff. The other 91 were green at 308ae94 (round 1), and the only new path, the baseline JSON, is read by check:error-status-conformance alone. CI convergence on b8c4e5e: in_progress at report time.",
"line_budget": "n/a",
"deviations": "(1) To measure the roster gates (not asked, read-only), I ran pnpm install and cached turbo builds in the recreated worktree. They wrote only gitignored paths (node_modules, dist); git status --porcelain was empty after each, and the committed trees are unaffected. The one tracked write remains the --update. (2) The comment's first line is 'os-dev-report — patch round 1', as the coordinator asked.",
"files_changed": [
"scripts/error-status-unpinned-baseline.json"
],
"mcp_calls": "0 this round.",
"api_writes": "1 REST write this round: this comment, POST /repos//issues/21604/comments through scripts/pm/post-stamped.mjs (the fleet-write relay). Plus 1 git push (fast-forward, not a REST write).",
"open_questions": [],
"out_of_scope_findings": []
}

---

Generated by Claude Code