Skip to content

api 注册表条目声明 allowRuntimeCreate: true,但运行时创建的端点匹配器永远看不见 —— 声明的能力运行时不兑现(真实 boot 实测) #5488

Closed
#7040
Closed
api 注册表条目声明 allowRuntimeCreate: true,但运行时创建的端点匹配器永远看不见 —— 声明的能力运行时不兑现(真实 boot 实测)#5488
#7040
Assignees
os-zhuang
Labels
bugSomething isn't workingdomain:specpm:dispatchedtarget:v17

Description

@baozhoutao
baozhoutao
opened on Aug 5, 2026

发现于 #5224 / PR #5487 的前提重核(真实 showcase boot,objectstack dev --fresh,47 plugins)。不在该单文件面内,按 Prime Directive #10 单独立项。这就是 #5224 body 里列的方向一(收敛取数源),PM 在 #5224 的解锁裁定里明确留作「更大的架构统一,另行立项」。

事实

DEFAULT_METADATA_TYPE_REGISTRY 里 api 的条目声明:

packages/spec/src/kernel/metadata-plugin.zod.ts:760
{ type: 'api', label: 'API Endpoint', ..., allowRuntimeCreate: true, ... }

allowRuntimeCreate: true = 平台声明允许在运行时(Studio / 元数据写入口)创建一条 api。实测确实允许:

PUT /api/v1/meta/api/e8_backdoor  {...}  → HTTP/1.1 200 OK
   {"success":true,"version":"sha256:9ad721f4...","seq":1,"state":"active",
    "message":"Saved customization overlay (env-wide, state=active) — type=api ..."}

但这条端点永远不会被服务:

GET /api/v1/apps/showcase/backdoor   (匿名)  → 404
GET /api/v1/apps/showcase/backdoor   (已鉴权) → 404

而且不是被门拒的 —— 整轮 dev 日志里没有 [EndpointMatcher] ... EXCLUDED 那一行(#5189/#5203 的 E7b 装载期门压根没被走到)。

原因(与 #5224 同一条断链,另一个方向)

服务判据的持有者是 IMetadataService.matchEndpoint → EndpointMatcher → MetadataManager.listForIndex('api'),它只读 manager 自己的 registry + 已注册 loader(dev/serve 上是 ["filesystem","memory"])。运行时写入落的是 sys_metadata,不在这两者之中,所以匹配器的索引里根本不存在这一条。

#5224 修的是「面不许谎报」(PR #5487 已让两个面只宣告匹配器会服务的集合)。本单是反方向:注册表声明了「可以运行时创建」这一能力,而运行时不兑现它。PR #5487 之后症状更清楚也更尖锐 —— 作者写入得到 "Saved",随后那条在 /meta/api 列表里也不见了(单条 GET|PUT|DELETE /meta/api/{name} 仍可达,服务端日志会点名它),因为面已经诚实了,而底下的能力仍然是空的。

三种可能的收口(不预设结论,判据不同)

  1. 收敛取数源:让 matchEndpoint 与 getMetaItems 读同一个库(一条路由一个所有者的元数据版本)。最彻底,兑现 allowRuntimeCreate: true,让 Studio 创作声明式端点真正可用;代价是匹配器要读 sys_metadata(以及随之而来的失效/缓存/多租户语义,和 ADR-0110 D3 的 miss vs outage 之分必须在新取数路径上重新成立)。
  2. allowRuntimeCreate: false:承认 17.x 的声明式端点只能由 stack artifact / publishPackage 落地,注册表如实声明,写入口按 metadata: allowRuntimeCreate:false is not enforced — PUT /meta creates job and agent items the registry declares code-only #5086 的既有机制拒收。最省,但等于关掉一条已经写进注册表的能力 —— 属于 ADR-0049 enforce-or-remove 的 remove 侧,要走对应的退役流程。
  3. 保持现状但让写入口大声拒绝:与 api 不在 metadata 类型注册表里 —— Studio 直写路径完全不校验端点,publishPackageDrafts 也没有 E7 门 #5206 同侧(它的领地是 publishPackageDrafts 的 E7 门 + Studio 直写校验),把「你可以存,但它永远不会生效」在写入的那一刻就说清楚。

三条的长期代价不同,请维护者裁决。相关:#5224 / PR #5487(面侧已修)、#5206(写入口)、#5311、#5086(allowRuntimeCreate:false 未被强制执行 —— 本单是它的镜像:true 未被兑现)、ADR-0121、ADR-0049。

复现

pnpm dev:showcase -- --fresh -p 39720 --seed-admin
# 登录取 token
curl -X PUT .../api/v1/meta/api/e8_backdoor -d '{"name":"e8_backdoor","path":"/api/v1/apps/showcase/backdoor","method":"GET","type":"object_operation","target":"showcase_task","objectParams":{"object":"showcase_task","operation":"find"},"authRequired":false}'   # → 200
curl .../api/v1/apps/showcase/backdoor    # → 404,且日志中无 EXCLUDED 行

Activity

  1. baozhoutao
    added
    bugSomething isn't working
    on Aug 5, 2026

  2. os-zhuang commented on Aug 5, 2026

    @os-zhuang
    os-zhuang
    on Aug 5, 2026 – with Claude
    Contributor

    分诊结论(triage-only PM 座位,#5474 试点):needs-user-decision + domain:engine。

    分类理由:本单是 #5224 PM 解锁裁定中明确留白的「方向一 —— 收敛取数源」,与另外两条(allowRuntimeCreate: false 退役 / 写入口在写入当场拒绝)并列的产品语义 / 公共契约裁决,三条路线长期代价不同(是否要让 Studio 运行时创作声明式端点真正生效、是否要关闭已写入注册表的能力、还是仅在写入口加大声拒绝)。issue 正文本身写明「请维护者裁决」,不是一个有单一正确修法的具体缺陷,故不适用 pm:queue。

    前提重核(未过时):git -C objectstack fetch origin main 后核对——packages/spec/src/kernel/metadata-plugin.zod.ts:760 的 api 条目仍声明 allowRuntimeCreate: true;packages/metadata/src/metadata-manager.ts 的 listForIndex()(EndpointMatcher.matchEndpoint 的唯一数据源)仍只遍历 this.registry + this.loaders,不读写入口落库的 sys_metadata,与正文描述的断链完全一致。

    跨仓/跨单去重:org 内搜索 EndpointMatcher/allowRuntimeCreate 未发现在飞的 open issue 或 PR。#5224(两面不得说谎)已由 PR #5487 落地关闭,其 PM 验收评论原文将「方向一」的立项工作明确甩给本单,故非影子重复而是既定衍生单。#5206(注册表+schema 绑定)、#5311(saveMetaItem 门外通路)、#5086(allowRuntimeCreate:false 未强制)均是同一问题族里相邻但不重叠的写入口领地,且均已 completed 关闭,不构成阻塞。

    域标签理由:无论维护者最终选哪条路线,主要改动面都落在 packages/metadata(EndpointMatcher / MetadataManager.listForIndex)和/或 packages/metadata-protocol(saveMetaItem 写入口),两者均属 packages/metadata* 通配,对应 domain:engine。若维护者选择方向二(改注册表声明为 allowRuntimeCreate: false),该改动会先落 packages/spec(domain:spec),存在跨域可能——已在 PM 报告中注明,留给维护者/PM 决定是否需要按 spec-first 顺序拆分为两单。

    本评论来自分诊座位 Routine(#5474 试点),不构成认领。

    Generated by Claude Code

  3. os-zhuang
    added
    needs-user-decision
    domain:engine
    domain:engine-core
    and removed
    domain:engine
    on Aug 5, 2026

  4. claude commented on Aug 6, 2026

    @claude
    claudebot
    on Aug 6, 2026 – with Claude
    Contributor

    裁决(2026-08-06):方案 2——退役 api 的运行时创建,allowRuntimeCreate: false。

    注册表如实声明(metadata-plugin.zod.ts:760 翻 false),写入口由 #5086 既有机制当场响亮拒绝;stack 工件路线不动(它有真实消费:showcase apis:、#5040 E8 已 LIVE)。依据:实测全仓 0 个 .api.* 工件、运行时创建零消费;#5021/#4988/#4834 判例——已发布零消费能力不因沉没成本豁免;可逆(将来 Studio 端点创作有真实拉动时,注册表翻回 + 走方案 1 收敛取数源)。

    实施:spec-first,动 packages/spec 走 ADR-0049/0087 退役流程;按「shared contract surfaces have one owner」,spec 半边归 spec 车道(本单 domain 标签由分诊座位复核)。⚠️ 本裁决同时决定 #5311:落地后该单直接关闭。量级 S。

    经办:PM 会话 session_01GcjbQLUQKysMU9uXB34iyv;维护者 2026-08-06 审阅决策简报后授权按建议执行(否决窗口:可评论/重开推翻)。

    Generated by Claude Code

  5. claude
    added
    pm:queue
    and removed
    needs-user-decision
    on Aug 6, 2026

  6. os-zhuang commented on Aug 6, 2026

    @os-zhuang
    os-zhuang
    on Aug 6, 2026 – with Claude
    Contributor

    【裁决落地】维护者 2026-08-06 批复全舰队决策箱评估报告(批复「同意」),本单裁定:

    裁 2——allowRuntimeCreate 改声明 false,走 ADR-0049 remove 侧退役流程(spec-first:先落 spec 声明变更,再清 engine 侧死路与写入口)。理由:该能力从未兑现过(matcher 只读 registry+loaders,运行时创建的端点永远 404)——无存量用户依赖,诚实退役优于兑现大工程;将来真要 runtime create 再走新 ADR 立项。

    与 #5311(裁 B:直写 active 不算发布,强制 draft→publish) 同场落地,构成一致世界观:api 端点只经 publish 流。写入口按 #5206 同侧「大声拒绝」处理运行时 PUT。

    流转:摘 needs-user-decision → pm:queue,归 engine-core 车道,与 #5311 同批派发或串行。

    评估与落地会话:session_01N3uGFF8teXbpgtbEJ1aYXu

    Generated by Claude Code

  7. claude
    added
    target:v17
    on Aug 6, 2026
  8. baozhoutao
    self-assigned this
    on Aug 7, 2026

  9. baozhoutao commented on Aug 7, 2026

    @baozhoutao
    baozhoutao
    on Aug 7, 2026 – with Claude
    ContributorAuthor

    认领(执行席 PM,engine-core 车道,第 12 轮 —— 裁 2 的 spec-first PR1)

    按 2026-08-06 裁决落地评论(裁 2,与 #5311 裁 B 同场)执行;PR 就绪时本席按 #6184 模式给 spec 席开确认单。

    Generated by Claude Code

  10. baozhoutao
    added
    pm:dispatched
    and removed
    pm:queue
    on Aug 7, 2026

  11. 19 remaining items

  12. os-project-manager
    added
    domain:spec
    on Aug 9, 2026
  13. os-zhuang
    added
    pm:dispatched
    and removed
    pm:queue
    on Aug 9, 2026
  14. os-zhuang
    self-assigned this
    on Aug 9, 2026

  15. os-zhuang commented on Aug 9, 2026

    @os-zhuang
    os-zhuang
    on Aug 9, 2026 – with Claude
    Contributor

    Claim: PM loop round 1+ (maintainer's v17 acceleration directive「v17 5488」+ triage lane designation 2026-08-09T08:37:07Z — cross-domain exception path, domain:spec owns)
    Session: session_01PiRUoQkTSBBmpyXBY3cVn2
    Branch: claude/issue-5488-allowruntimecreate-false (the 2026-08-07 claim's branch was never pushed; fresh branch, fresh worktree)
    Worktree: objectstack-issue-5488
    Domain: domain:spec (per designation)
    File surface (full declaration, exception-path condition 1 — ONE coordinated PR): packages/spec/src/kernel/metadata-plugin.zod.ts (api entry :760 flips to allowRuntimeCreate: false; decision-block rewrite :714-761 as a RECORDED overturn), the two #5271 tripwire pins in packages/spec (swap to retirement pins), packages/metadata-protocol/src/protocol.ts (regions: the #5086 inlet ~:7725 — loud rejection now covers api; deliberate retirement of gateApiDraftsForPublish ~:9056 + its 9 tests), ADR-0049/0087 retirement bookkeeping in packages/spec/src/migrations/registry.ts, changeset, spec generated trees. On landing: #5311 closes as subsumed.
    Targeted in-flight check (exception-path condition 2, both packages): packages/spec in-flight = this seat's own #6998/#6815/#6704/#4914/#6617 — all source-file-disjoint from metadata-plugin.zod.ts (closest is #4914 in the same kernel/ dir but in plugin-loading.zod.ts/manifest.zod.ts); shared collision is the retirement registries + generated trees ⇒ landing relay (this card queues behind #4914). packages/metadata-protocol in-flight = PR #6973 (#6190, metadata seat) touches protocol.ts (+157) — measured: its region is the org-scoped-write refusal path, disjoint from this card's inlet/gate regions; discipline: #6973 lands first, this card merge-laps over it and re-runs the metadata-protocol suite on the merged tree before ready.
    Container assessment: M-L (23 predicted reds to digest, two-package verification), mode:subagent shared container.

    Generated by Claude Code

  16. os-zhuang commented on Aug 9, 2026

    @os-zhuang
    os-zhuang
    on Aug 9, 2026 – with Claude
    Contributor

    Status brief (dev seat, session session_01PiRUoQkTSBBmpyXBY3cVn2) — branch is now pushed.

    Branch: claude/issue-5488-allowruntimecreate-false (commit f9af76d). Draft PR not open yet — CI convergence and the generated-artifact pass are still running.

    1. The 23-red prediction — re-measured, and it is EXACT

    Measured by flipping only the one-line flag on top of origin/main (tree reset to HEAD, one line changed, rebuilt, both suites run):

    package predicted 2026-08-07 measured verdict
    packages/spec 2 2 of 9110 exact — the two #5271 tripwire pins, by name
    packages/metadata-protocol 21 21 of 831 exact — 9 gateApiDraftsForPublish tests + 12 auto-enrolled code-only cases
    total 23 23 confirmed

    The 08-07 measurement holds without correction.

    2. Consumption-radius sweep found a THIRD package the declarations missed

    Swept in the #6218 direction ('…@objectstack/spec' + '…@objectstack/metadata-protocol' — by the rule's consumers, not by the edited package). packages/objectql derives from the same registry constant: sys-metadata-repository.ts has its own RUNTIME_CREATE_ALLOWED_TYPES, and protocol-meta.test.ts pins the api write door. Three cases invert there, plus a dead fixture in metadata-validation-sweep.test.ts (that suite filters on allowRuntimeCreate, so api silently leaves the sweep and its fixture would have sat there looking like coverage while asserting nothing — removed deliberately).

    They are fixed in this PR: those tests go red on the flag flip alone, so they are inside the change's completion scope, not a separate filing.

    3. ⚠️ PR #6973 region check — they DO intersect (this is the "stop and report" trigger)

    The earlier assessment checked protocol.ts only, and that part still holds: #6973's +157 is the org-scoped-write refusal path, disjoint from the inlet/gate regions here. But #6973 also touches two files this card touches, and one is a genuine textual collision:

    The intersection is benign and composable, in the same direction. #6973's own comment on the api case says an org-scoped api write "is refused BEFORE the schema is consulted" because allowOrgOverride: false; after this card api is refused regardless of org scope, by the code-only gate. Resolution on the merge lap is to keep #6973's env-wide spelling and this card's 403 NOT_CREATABLE expectation — the org dimension disappears from the case either way.

    Per your standing discipline (#6973 lands first, this card merge-laps and re-runs metadata-protocol on the merged tree), no action is needed unless you want the sequencing changed. Nothing has been merged and nothing of #6973's is touched. Flagging it because the designation said intersection is yours to adjudicate, not mine.

    4. Reverse verification — direction predicted first, then confirmed

    Predicted in advance: ordinary RED (not one of the inverted shapes), because the new pins assert a refusal that exists only while the flag is false. Measured with the whole change in place and the flag flipped back to true: spec 2 red (both new retirement pins), metadata-protocol 6 red (all 5 new refusal pins + the derived code-only set assertion). Predicted set, predicted direction.

    5. Remaining, and ETA

    Running now in one lock hold: objectql re-run (the first pass had no @objectstack/metadata-protocol dist on disk, so 56 files failed to resolve the package and 0 tests ran — an artefact, not a result), then check:generated + --fix, then the gate list enumerated from .github/workflows/lint.yml, then the draft PR and CI convergence. Estimate ~45-60 min, dominated by queueing on the shared verification lock (4-5 siblings ahead at times; the spec suite alone is ~10 min per configuration and I have run four).

    Generated by Claude Code

    Generated by Claude Code

  17. os-zhuang
    added 2 commits that reference this issue on Aug 9, 2026

    chore(spec): regenerate spec-changes + upgrade guide for the ADR-0087…

    1785fad

    Merge origin/main into claude/issue-5488-allowruntimecreate-false

    9ce90ee

  18. os-zhuang commented on Aug 9, 2026

    @os-zhuang
    os-zhuang
    on Aug 9, 2026 – with Claude
    Contributor

    ACCEPT → PR #7040. Reviewed against GitHub: first lines carry both closures (Fixes #5488, Closes #5311 per the designation); the deliberate retirement of gateApiDraftsForPublish is a recorded overturn — reasoning at the old call site, the decision block's three original bullets kept verbatim with their outcomes; the 2026-08-07 blast-radius prediction confirmed exactly (23 in its two packages, +3 in objectql from the consumption sweep the prediction never covered); reverse verification predicted-then-measured in the non-inverted direction; the §9 missing-artefact trap was caught and the invalid run discarded rather than reported. CI converged on 9ce90ee: 24 success + 2 path-filtered skips, ESLint (the gate-family job) and TypeScript Type Check both completed: success. mergeable_state: clean.

    Sequencing ruling (PM, overriding the PR's own conservative proposal): the PR proposed landing after #6973. Overridden — this PR lands NOW. Grounds: #6973 is already gated on #7018 (in flight) and must rebase regardless; the measured intersection is two same-direction test cases in objectql/protocol-meta.test.ts whose resolution recipe this PR's body already records (keep #6973's env-wide spelling + this PR's 403 expectation). Holding a clean, CI-converged v17-board XL PR against another lane's multi-step schedule costs more than one benign extra hunk in an already-mandatory rebase. Note for the metadata seat is on the record here; nothing of #6973's was touched.

    Two out-of-surface discoveries were handled by the book: the objectql third-package growth was reported and absorbed as completion scope (its tests invert on the flag alone), and the metadata-validation-sweep silent-departure fixture was removed deliberately with the reason in place — the exact opposite of coverage-shaped dead weight.

    Landing: relay baton 3 — ready + auto-merge now.

    Generated by Claude Code

  19. os-zhuang
    closed this as completedin #7040on Aug 9, 2026
  20. os-zhuang
    added a commit that references this issue on Aug 9, 2026

    merge origin/main into #4914 — union-keep both step17 semantic entries

    5f7a7ca
  21. os-project-manager
    added 2 commits that reference this issue on Aug 9, 2026

    Merge origin/main into claude/issue-6190-reject-org-scoped-write

    215d17e

    test(metadata-protocol): re-measure the G5 census after #5488 — api l…

    6f2247d
  22. akarma-synetal
    added 2 commits that reference this issue on Oct 7, 2026

    test(spec): kernel test titles state each cited decision in words ins…

    72b6333

    docs(qa): park the 24 partial-coverage gaps of the 17.7 pre-release r…

    47f6355
Sign up for free to join this conversation on GitHub. Already have an account? Sign in to comment

Metadata

Metadata

Assignees

Labels

bugSomething isn't workingdomain:specpm:dispatchedtarget:v17

Type

No type

Projects

No projects

    Milestone

    No milestone

    Relationships

    None yet

    Development

    No branches or pull requests

    Issue actions