[objectstack-fleet]

Requested by Warren · project thread

Refs objectstack-ai/cloud#3974

维护者速读

• 改了什么：reference/security.mdx 三处说多了的地方改成和企业版能力清单、INSTALL.md 一致。
• 认证：删掉「已经有很多部署通过了」，保留「ObjectOS 本身未认证，部署可以认证」。
• AI 审计事件：原来列的 5 个 ai.* 事件名在框架和 cloud 里都不存在，换成 sys_ai_audit_log 真正写入的 tool_call / quota_refusal / guardrail_refusal。
• 新增一句：审计只覆盖产品内置 AI；外部 agent 走 /mcp 不写审计行，RLS / FLS 和脱敏仍然生效。
• 「可以导出」改成「审计日志导出目前不是已交付功能」；token 与成本记在 ai_traces，不在审计行上。
• 译文：按翻译流程第 3 步删掉 ja / zh-Hans，重生成后 zh-Hant 一并删除。需要你做的：无。

What changed

content/docs/reference/security.mdx (English only):

1. Certification (L152–153). "Your deployment can be certified — many already are." → "Your deployment can be certified." No source backs "many already are".
2. ### Audit events (L253–271). The five listed names are written by nothing:
   • objectstack at cloud's .objectstack-sha 6ed533f6: git grep for ai\.chat\.message, ai\.tool\.call, ai\.pending_action, ai\.metadata\.applied (code, excluding .md/.mdx) → 0 hits each. The framework has no sys_ai_audit_log writer at all.
   • The trail lives in cloud: packages/security-enterprise/src/ai-governance/audit-log.object.ts:44-50 declares tool_call, turn, quota_refusal, guardrail_refusal. Writers: tool-interceptor.ts:354 (tool_call), ai-governance-plugin.ts:196,237 (quota_refusal, turn quota and env token guardrail), ai-governance-plugin.ts:288 (guardrail_refusal). turn has no writer (git grep "event: 'turn'" → 0), so it is not listed.
   • The list now names those three and says the rows are append-only (the object's own wording, audit-log.object.ts:12).
   • Added: the trail covers the product's built-in AI; an external agent calling data tools over /mcp writes no sys_ai_audit_log row, while RLS / FLS and masking still apply (cloud docs/design/ee-feature-catalog.md L55; objectstack packages/runtime/src/domains/mcp.ts:604 callData).
3. Export. "These rows are immutable and can be exported for security review or chargeback" → "Exporting the audit log is not an included feature yet" (cloud apps/objectos-ee/deploy/INSTALL.md L164–169).
   • Cost sentence kept in corrected form: token counts and cost per model exist on ai_traces (cloud packages/service-ai/src/objects/ai-trace.object.ts:77 model, :91 prompt_tokens, :135 total_cost), not on audit rows (audit-log.object.ts:20-21). The old "(provider, model, user)" key is dropped: ai_traces has no user column.

Same-claim grep (English content, locale files excluded)

• "many already are" → only reference/security.mdx:154 (fixed).
• AI audit "exported" → only reference/security.mdx:264 (fixed). operate/observability.mdx:10 ("the deployment decides where to export them") is about request-id/metrics/audit primitives in general, not AI audit export; not an exact repeat, left as is.
• "every AI call is audited" or equivalent → no exact repeat. Near hits, not changed (outside this card's scope; flag for a follow-up card if wanted):
  • deploy/licensing.mdx:73 "An audit trail of every AI tool call" (lacks the "built-in AI" scope the catalog uses).
  • why.mdx:58 "every AI-driven change is itself an audit entry" (record changes go through sys_audit_log; reads as broader than the catalog).

Translations

Step 3 of the Translation workflow (capability assertions removed): deleted security.ja.mdx and security.zh-Hans.mdx; pnpm --filter @objectos/docs gen:zh-hant removed the stray security.zh-Hant.mdx. No locale file modified.

Gates

See gates-3974.txt in the seat folder; exit codes are listed in the os-dev-report on cloud#3974.

🤖 Generated with Claude Code

https://claude.ai/code/session_01JQwxbZ78dWmZbQWL1kT2JM

[objectstack-fleet]

repo:cloud#3 · os-warren · session_01JQwxbZqfEd7pcckjz3pGqn

席位核对

ACCEPT · head 10e297b · repo:cloud#3 · TIER_DEFAULT seat review (2026-11-09T03:02Z; content-only: content/docs/reference/security.mdx +16/−11, its ja / zh-Hans / zh-Hant siblings deleted; no workflow, no apps/docs code)

1. Shape: draft → main, body opens Refs objectstack-ai/cloud#3974 (no closing keyword; the seat closes the card at landing), six-line 维护者速读 present. Surface matches claim 6731146357.
2. Premise measured: on docs main the page said "many already are" (certified), listed five ai.* audit events, and said the rows "can be exported". The seat re-read the sources:
   • cloud packages/security-enterprise/src/ai-governance/audit-log.object.ts L10–21: one append-only row per governed AI event, registered only on licensed Enterprise runtimes, with token/cost on ai_traces, not here.
   • L42–52: event options are tool_call / turn / quota_refusal / guardrail_refusal. turn has no writer in cloud packages (seat grep: 0).
   • ee-feature-catalog.md L55: external /mcp writes no row.
   • INSTALL.md L164–169: export is not included.
     The new text says exactly this, with the Enterprise licence scope, the three written events, the external-AI exclusion (with RLS/FLS/masking still holding), export not included, and cost on ai_traces. The thread's 0-hit greps for the old ai.* names at pin 6ed533f6 are in the body.
3. Locale deletions: Translation step 3 (capability claims removed). zh-Hant was pruned by gen:zh-hant; no locale file was modified.
4. Gates (thread, gates-3974.txt): gen-zh-hant --check, check-translations, check-translation-ownership, check-translation-output, docs build, check:positioning, check:locale-surface all exit 0 (report cloud#3974 6731184557). The ready run's CI is the gate.
5. Out of scope, noted: deploy/licensing.mdx:73 ("every AI tool call", which lacks "built-in") and why.mdx:58 are near-misses, not exact repeats. The seat takes them as one follow-up wording card. Skills-docs: n/a.
6. Landing: CI-gated deploy with live smoke and automatic rollback (ci.yml → deploy-docs.yml) is the verification layer. pr_ready now, auto-merge after green; landed = deployed + smoke green. Then the seat closes cloud#3974.