[Decision digest] maintainer inbox — generated view (please pin)

[os-support-ai]

Generated view — not a second tracker. Authority lives on each card (one-board rule); this body is refreshed by the triage-seat Routine each round from a label query (needs-user-decision, all repos in scope) + literal-text extraction of each card's os-decision-facets block. Edit history = archive. Batch adjudication: reply in ONE comment below with one pair per line, each pair spelled like #9807 B (card number, space, option letter); the triage seat transcribes each ruling back onto its card citing your comment as provenance. (Standing per the 2026-08-18 ruling: standard block, pinned digest, batch channel — 「同意」.)

Last refresh: 2026-08-19 ~11:2xZ, session session_015BAKXkmpGhLRSys9AjNZMX. Current inbox: objectstack 7 (+#9968, +#9952 new this round; +#9617 re-entered by the skills seat — see its provenance note), objectui 1, objectos 5. Total 13. Cards with open downstream dependents: objectos#68 (← objectos#135, and via the chain objectos#75).

---

objectstack #9968 — impersonate_user / set_user_role still 403 every platform admin; neither vendor route is safely raw-mountable (auth/session boundary + ADR-0068 Option-3 veto tension)

Recommendation: B (retire set_user_role; for impersonate_user, measure the plugin-endpoint override question first, then re-implement as a better-auth plugin endpoint or retire) · A = status quo (dead buttons, fails closed) · C = re-implement set-role with a vendor-excluding vocabulary + plugin-endpoint impersonate.

<!-- os-decision-facets -->

• platform long-term coherence: Retiring set_user_role (B) shrinks the retired dual-identity surface — ADR-0068 moved platform-admin membership to sys_user_permission_set, and a supported UI writer for the legacy scalar re-grows exactly that contract. For impersonation, a better-auth plugin endpoint (not a raw Hono mount) is the only shape that keeps the signed-cookie contract with /admin/stop-impersonating and the better-auth bearer plugin lets a bearer session silently shadow an impersonation the server just created — /admin/impersonate-user returns 200 and is a no-op for any bearer client #8243 rotation hook attached.
• measured business pull: Both buttons are measured broken today (byte-identical 403 for admin and plain member). But nothing measured pulls for set-role's only effect — writing the vetoed scalar; impersonation is a real admin-support workflow, though no usage telemetry exists.
• AI-agent error-resistance: A raw re-implementation must hand-roll better-auth's signed-cookie format — a subtly wrong signature is a broken exit path or a forgeable cookie, and the detached path-keyed hook re-introduces better-auth bearer plugin lets a bearer session silently shadow an impersonation the server just created — /admin/impersonate-user returns 200 and is a no-op for any bearer client #8243 with no failing test. Removing the set-role text box removes a one-user-at-a-time resurrection channel for the dual identity.
• startup scope discipline: Remove beats declare-and-maintain for set-role. For impersonate, the cheap first step is the unmeasured probe (checkEndpointConflicts override behaviour), not a commitment to re-implement.
  Confidence gap: this analysis cannot see whether any real deployment relies on impersonation (zero usage telemetry), and the plugin-endpoint conflict behaviour on better-auth 1.7.1 is explicitly unmeasured — if the probe fails, the impersonate half of B collapses back into A vs. a much more expensive path.

---

objectstack #9952 — [field report] rc→GA envs: package-declared permission sets frozen at first-boot snapshot; provenance skip is silent, no adoption/migration path (stored-data migration shape)

Recommendation: B (explicit operator command, e.g. os meta adopt-permission-sets, + surface the skip in Setup) · A = boot-time auto-adoption + surfacing · C = document SQL remediation only.

<!-- os-decision-facets -->

• platform long-term coherence: The env-authored protection is sound; the defect is the platform's own historical rows landing in that bucket with no exit. A one-time adoption path restores the [cli/plugin-security] dev 元数据物化不一致:默认权限集 insert-if-not-exists 不重物化(声明集走 upsert) #2705 reconcile intent without weakening provenance; leaving it means "declared = enforced" is permanently false for every rc→GA env.
• measured business pull: Production field report, measured on a live env — grants shipped after first boot return 403 to real users, remediation today is raw SQL. Every env upgraded along the same rc line is affected.
• AI-agent error-resistance: The current failure signal is a log counter — the exact silent-tolerance shape that lets drift run for weeks. Surfacing the skip loudly (Setup badge/warning on the set) is the contract-tightening half and looks safe under every option.
• startup scope discipline: Keep it a narrow, one-time adoption keyed on name-match + provenance-predates-scheme; do not build a general reconciliation framework.
  Confidence gap: cannot see how many envs sit on the rc→GA line, nor whether any genuinely admin-authored set shares a package set's name in the wild — the collision likelihood is what would flip B to A.

---

objectstack #9930 — retire the legacy sys_webhook.definition_json.headers cleartext read path (stored-data migration + accepted-input removal)

Recommendation: A (three-step retirement: measure population → boot-sweep migration into headers_secret → flip readLegacyHeaders to loud rejection; migration isolated in its own flagged PR) · B = migrate but keep warn-and-accept · C = leave as-is.

<!-- os-decision-facets -->

• platform long-term coherence: retiring the cleartext read path completes [security] webhook custom headers are still cleartext in two JSON blobs — the sibling of #7799 that PR #7901 did not close #7986 — one storage location for webhook headers instead of a permanent legacy fork; shrinks a special case.
• measured business pull: chore(plugin-webhooks): declare sys_webhook's data-API exposure explicitly — and record that it narrows nothing (#9756) #9927's census measured this as the only actionable residue of sys_webhook still declares no enable block — the read-surface narrowing #7799 and #7986 both pointed at was never delivered, and #7986 closed completed #9756's reachable-cleartext concern (the enable.apiMethods remedy measured as a no-op; url masking ruled won't-fix in [security] decide explicitly: a webhook URL can BE the credential (Slack/Discord-style endpoints), and it is stored plain on two objects #8025). No user asked for legacy rows to keep working silently.
• AI-agent error-resistance: warn-and-accept is the tolerance pattern that hides authored mistakes; after content/docs/automation/webhooks.mdx still documents custom headers as living in sys_webhook.definition_json, and omits headers_secret entirely — the doc points authors at the cleartext column #7986 moved them out of #9928 fixes the docs, a loud dated rejection (post-migration) makes writing headers into definition_json impossible to get silently wrong.
• startup scope discipline: the work is a bounded three-step retirement with an in-package precedent (migrate-webhook-secrets.ts); the alternative — keeping the legacy path forever — is an open-ended maintenance obligation.
  Confidence gap: deployed-instance population of pre-[security] webhook custom headers are still cleartext in two JSON blobs — the sibling of #7799 that PR #7901 did not close #7986 legacy rows unmeasured; if large, the migration's blast radius and idempotency proof deserve extra scrutiny before A executes.

---

objectstack #9901 — federation datasource routes admit any authenticated caller while the admin twin now requires a capability (security boundary; two WRITE routes included)

Recommendation: A (manage_platform_settings on the read twins per #9593; manage_metadata on the write routes import/refresh-catalog; fold the equivalence-test divergence row back) · B = mirror #9593 exactly, no read/write split · C = rule the federation family deliberately lower-privilege and document it.

<!-- os-decision-facets -->

• platform long-term coherence: the twin-equivalence family is a maintained invariant (The same three datasource-admin routes answer 400 with the wrong service's error code (#4225 follow-up, one field over) #4249 failure contract, finding: the two listRemoteTables route twins diverge on ?schema= — the admin spelling drops it #7955 request shape, finding: the external-datasource federation HTTP family mounts outside the enforceAuth seam with no guard of its own — read and write #9686 anonymous axis); the capability axis is the first pinned disagreement — converging shrinks a special case, and the write routes (import, refresh-catalog) creating live metadata behind auth-only is the sharpest edge.
• measured business pull: the asymmetry is measured and pinned in remote-tables-twin.equivalence.test.ts; no user has asked for a lower-privilege introspection door — but datasources.external.* IS published SDK surface, so a new capability requirement is a behaviour change for any real caller presenting a plain credential.
• AI-agent error-resistance: one admission policy for one operation family is the shape agents can't get wrong; two doors with different gates invite the "use the other spelling" bypass an agent will find.
• startup scope discipline: the fix reuses an existing seam (resolveExecutionContext + systemPermissions, already read by package-routes.ts) — no new mechanism; the decision cost is the SDK-caller break, not the code.
  Confidence gap: real-world datasources.external.* SDK callers invisible from the repo — A and B both break a non-admin production caller if one exists, and no staged-migration story has been designed.

---

objectstack #9885 — sys_position.permissions is a declared "JSON-serialized array of permission strings" column with zero producers and zero readers (enforce-or-remove)

Recommendation: A (REMOVE via the ADR-0049/ADR-0087 retirement flow, incl. the clone_position copy entry and docs riders; objectui designer-preview check folded into the retirement checklist) · B = enforce (design position-level direct grants — a real security feature, only on real pull) · C = leave declared-and-inert.

<!-- os-decision-facets -->

• platform long-term coherence: grants flow through permission-set bindings only (sys_position_permission_set); a parallel free-text permissions catalogue on Position is a second, never-implemented grant vocabulary — removing it shrinks the security model to the one real path. Same class as the ruled [finding] sys_user_permission_set.delegated_from is declared and data-door-writable, but the runtime delegation gate structurally never reads it on this object #9730 (delegated_from, REMOVE 2026-08-18) and sys_permission_set.active and sys_position.active are unenforced too — both Deactivate dialogs promise access stops, and it does not #8613.
• measured business pull: zero — 97 sys_position-naming files read; no producer, no reader; the only in-file reference is clone_position copying a value nothing writes. Nobody is using it.
• AI-agent error-resistance: the declared textarea tells an AI author "grant permission strings directly on a position" — a capability the runtime does not have; an agent following the declaration authors security config that silently does nothing. Removal makes the wrong model unwritable.
• startup scope discipline: implementing position-level direct grants (the enforce arm) is a whole security feature with no pull; remove wins by default under the standing enforce-or-remove posture.
  Confidence gap: objectui not searched (evidenceScope: framework repo only) — a console position-designer consumer would need a rider; hand-filled deployment data invisible.

---

objectstack #9864 — duplicate plugin registration: LiteKernel throws, ObjectKernel silently overwrites (fourth two-kernel semantic split; #9863 rides the overwrite)

Recommendation: A (converge on refuse-loudly in both kernels + an explicit supersede spelling for #9863's config-over-CLI case, pinned dual-kernel) · B = converge on overwrite-with-warning · C = declare the kernels intentionally different and document it.

<!-- os-decision-facets -->

• platform long-term coherence: fourth measured instance of the two-kernel semantic split (core: kernel:ready 钩子抛错在 ObjectKernel 上失败 boot、在 LiteKernel 上被吞成一条 error 日志 —— 同一钩子两套失败语义 #5170, core(finding): ObjectKernel 不继承 ObjectKernelBase —— 钩子分发语义在两处各写一遍,#5170/#5257/#5274 都是同一条结构缝 #5282, finding: serve.ts's cloud-connected marketplace arm can silently replace a host config's own marketplace plugins #8357 adjacent); whichever direction is picked, one declared registration contract pinned against BOTH kernels shrinks the seam.
• measured business pull: the overwrite is load-bearing today only by accident (finding: record-view auditing has no configuration path on the os serve boot path — the CLI registers AuditPlugin with no options #9863's config-supersedes-CLI path); nobody has asked for silent replacement semantics, and the failure direction (a differently-configured instance silently replacing a plugin, logging "registered" twice) is the expensive one.
• AI-agent error-resistance: loud refusal on duplicate names + an explicit override/supersede API is the shape an agent cannot misuse; silent last-writer-wins is exactly the tolerance that hides authored mistakes.
• startup scope discipline: option A is small (a has() check + an explicit registration option); option B (document divergence) creates a permanent two-contract obligation.
  Confidence gap: no census of real boot paths relying on the overwrite (only finding: record-view auditing has no configuration path on the os serve boot path — the CLI registers AuditPlugin with no options #9863's collision is measured); A needs that census first or quiet boots become hard failures.

---

objectstack #9617 — checklist-test's documented QA-run title convention matches 0 of 23 real records (governed-surface contract; ⚠️ has a recorded ruling AND a revert — see provenance)

Recommendation: 2, new-records-only (one canonical title shape, strict parser, no migration of the 23 records) · 1 = document bends to measured reality. Provenance: comment 5334893052 (08-18 22:31) recorded Option 2 as ruled via the live-chat batch acceptance 「其他接受你的建议」; the skills seat (comment 5340733624, 08-19 10:19) re-entered it for an explicit per-card verdict. A one-word confirm resolves it.

<!-- os-decision-facets -->

• platform long-term coherence: a documented convention matching 0 of 23 real records is declared≠enforced in the agent-instruction layer; only direction 2 (one canonical shape, strict parser) prevents recurrence.
• measured business pull: real and current — PR tooling(qa): roll-up view over qa-run records — selector, latest verdict, staleness #9616's roll-up parses these titles today and had to be written against the measured shape, not the documented one.
• AI-agent error-resistance: one emitted shape + a strict parser that prints what it cannot read makes deviation loud on first occurrence; canonizing five phrasings is the tolerant-consumer shape that hides authoring errors.
• startup scope discipline: direction 2 with new-records-only costs one SKILL.md edit and zero migration; the "Not parsed" roll-up section already exists as the enforcement surface.
  Confidence gap: none material on substance — the open question is provenance, above.

---

objectui #5301 — WidgetConfigPanel's resolveLabel reads only the RETIRED label form; opening the panel on a locale-map title shows EMPTY and a save writes '' over the author's map

Recommendation: A (display via pickLocalized; un-edited fields round-trip byte-for-byte; an edit writes the active locale's map entry, preserving other locales; map-editing UX stays a future card) · B = read fix + lossy collapse-to-string on edit · C = full multi-locale editing UI now.

<!-- os-decision-facets -->

• platform long-term coherence: fourth private copy of the retired-form resolver (fix(runtime,rest): advertise mcp only when it is actually serveable #4032 swept three); whichever write-back rule is chosen should be stated once and shared, not re-derived per panel.
• measured business pull: the destructive path is the ordinary one — open a widget whose title is an inline locale map, change anything, save ⇒ the author's map is replaced by ''. Studio authors hit this on every map-titled widget today.
• AI-agent error-resistance: an editor that silently collapses a multi-locale map to one string (or to empty) destroys data with no failure signal; write-back-only-the-edited-locale keeps saves idempotent for untouched fields.
• startup scope discipline: a full map-editing UI is the expensive road; the minimal honest fix needs no new UI.
  Confidence gap: not checked whether the dashboard save path round-trips unknown keys untouched or rebuilds config from the draft — decides whether A needs a diff mechanism.

---

objectos inbox (5 open — options and recommendations on the cards; standard facet blocks pending backfill, a low-frequency sub-round item)

• objectos Fix test data to match evolved schemas #137 — root engines.node >=22 claims a window (22.0.0–22.11.x) yargs does not support. Card recommendation: 2+1 (strengthen the gate to "declared floor satisfies every dependency range" + tighten the declaration to >=22.12.0).
• objectos 🔗 Broken links detected in documentation #138 — fourth Node floor declaration: tools/ci-scripts says >=20.0.0 while the repo says 22. Fork: one repo-wide floor vs. accurate per-package floors; either way is a one-line follow-up.
• objectos Release v0.3.3 - ObjectAI documentation + critical build fixes #102 — Studio is a third surface name the 🔗 Broken links detected in documentation #79 naming decision does not cover.
• objectos Organize JSON schema output by protocol category #90 — how why.mdx sells the multi-tenant internal-portal scenario now that per-project kernels are gone.
• objectos Complete protocol review and Q1-Q4 2026 action plan (Chinese + English) - Verified 2026-01-23 #68 — ownership check inert until TRANSLATION_BOT_LOGIN is set (also blocks objectos#135 and, via the unlock chain, objectos#75).

[os-warren]

退役裁定(维护者 2026-08-19,原话:「之前定期生成的决策汇总 issue没什么用」「就是这个 #9822」)—— 本视图停止刷新。

替代机制当日已实测落地(skills 座,session ae0d8bd1-5667-5f58-8ff8-fa22ec154325):

• needs-user-decision 卡 ⇒ assign os-zhuang(手机推送 + Assigned 队列即决策收件箱,关卡自动消项)。skills(checklist-test): the documented QA-run title convention matches 0 of 23 real records #9617/finding: SKILL.md's governed-PR fork says "never flip ready", while the devx seat post's standing practice flips governed ADR PRs ready-for-review so the maintainer's merge button works #9900/finding: CLAUDE.md carries no line-ratchet ceiling, while AGENTS.md does — same read path, same governance #9965 已指派。
• 治理面 PR 等人合 ⇒ review request os-zhuang(Review-requested 队列;draft 照样可点名,PR docs(agents): name all six blocking contexts in the merge-queue rule #9940 实测推送到达)。
• 各卡 os-decision-facets 块原地保留(本视图本就是从卡上抽取的,无信息丢失);批量裁决走聊天或逐卡回复,分诊誊录职责不变。

@分诊席:下轮起 ⛔ 勿再刷新本贴 —— 2026-08-18 的 standing 裁定已被上述 2026-08-19 裁定取代;SKILL.md 的对应文本修订已立卡走治理面(人合),落地后本贴关闭,edit history 即归档。收件箱里尚未指派 os-zhuang 的存量决策卡(objectstack #9968/#9952/#9930/#9901/#9885/#9864、objectui #5301、objectos 5 张),请在下轮 sweep 时按新机制逐张补指派(objectui/objectos 由各自可达席位执行)。

---

Generated by Claude Code

[os-warren]

退役完成,本贴关闭(skills 座,session ae0d8bd1-5667-5f58-8ff8-fa22ec154325)。

裁定链:2026-08-19 维护者「之前定期生成的决策汇总 issue没什么用」→ 成文于 PR #9982(12:19Z 由维护者合并):digest 逐轮刷新职责与批量决裁通道自 SKILL.md 删除;各卡 os-decision-facets 块与分诊誊录职责保留不变。决策收件箱后续形态(同日第二次裁定收窄):不推送、不指派,维护者定期与 AI 在聊天中讨论后逐卡裁决;需维护者动手的只有治理面 PR 的 review request。

权威恒在各卡;本贴编辑历史即存档。

---

Generated by Claude Code