Per the architecture-review redline (EIP3 决策会纪要: AI claims limited to 'AI-assisted, human-reviewed'; never promise 'AI full-auto development'), soften endpoint @Operation summaries, the default task objective, and AI docs that read as committed full-automation. Endpoint paths/method names unchanged (API contract). Added a capability-boundary disclaimer at the top of AI_AGENT_AUTO_DEVELOPMENT.md. 33 platform-ai tests green.
Co-Authored-By: Claude Opus 4.8 (1M context) <noreply@anthropic.com>
AiAgentController's endpoints relied only on the SecurityFilterChain (prod: /api/** authenticated; dev: permitAll), so any authenticated user could trigger AI writes/actions (auto-develop, local-apply, change-sets, deployment packages) - a lower bar than the comparable overlay/standard-page write endpoints which require admin.
Gate the 14 mutating/action POST endpoints with requireAdmin() (UserSession CURRENTUSER == ADMIN), aligned with PageSchemaController.requireAdmin / AdminWriteController. Read GETs and the pure validate-schema/access-check endpoints stay open. Actual three-truth writes are now admin + localhost double-gated.
Tests: AiAgentControllerTest + AiAgentApiPipelineTest run under an ADMIN session (+ AfterEach clear); added a 403 test for non-admin. 61 platform-ai tests green.
Co-Authored-By: Claude Opus 4.8 (1M context) <noreply@anthropic.com>