Engineering notes on the Starchild agent platform — architecture, systems, and design decisionsStarchild Agent 平台工程档案——架构、子系统与设计决策
A self-maintaining layered knowledge system: Rules / Wiki / Raw three-tier storage, write gating, weekly lint, and daily distillation forming a fully automated knowledge loop.自维护的分层知识系统：Rules / Wiki / Raw 三层存储，写入门控、每周 lint、每日蒸馏，构成全自动知识闭环。
Read →阅读 → SystemIn-session write → background consolidation → promotion / retirement → audit: distill, weekly promotion, and skill curation all piggyback on bg_review, closing the loop from write to archive.对话内写入 → 后台巩固 → 晋升/淘汰 → 审计：蒸馏、每周晋升与技能整理全部搭载 bg_review，闭环贯穿写入到归档。
Read →阅读 → SystemThree-tier capability discovery: local platform tools & installed skills → community skills & x402 marketplace → web search. Pre-install interception, upgrade chains, and soft recommendation.三层能力发现：本地平台工具与已装 skills → 社区 skills 与 x402 marketplace → Web 搜索。装前拦截、空结果升级链与软性推荐。
Read →阅读 → SystemConnect external tool servers over stdio, streamable HTTP or SSE — including remote servers behind an OAuth 2.1 login, with tokens held out of the agent's reach and third-party tool descriptions treated as untrusted data.通过 stdio、streamable HTTP 或 SSE 接入外部工具服务器——包括需要 OAuth 2.1 登录的远程服务器，token 存放在 agent 够不到的地方，第三方工具描述一律按不可信数据处理。
Read →阅读 → SystemThe spoken front door to the same threads and tasks: GPT-Live client delegation, a bridge with per-scenario bindings (thread · onboarding · orchestrator), four data flows that keep the thread the single source of truth, and §12: the deployed path as of 2026-09-18 — proxied sessions, Sideband gate, one message per exchange, DB-confirmed refresh.通往同一批线程与任务的语音正门：GPT-Live 客户端委托、带场景 binding（thread · onboarding · orchestrator）的 bridge、让线程保持唯一事实源的四条数据流，以及 §12：截至 2026-09-18 的已部署链路——代理会话、Sideband 门控、一轮一条消息、数据库确认后刷新。
Read →阅读 → SystemThe thread is the stable identity: many sessions — interactive, push, scheduled, sub-agent — flow into one conversation that spans web and messaging channels, plus cross-session handoff with ownership and loop guards.线程才是稳定身份：交互、推送、定时、子代理等多种会话汇入同一条横跨 Web 与消息渠道的对话，并支持带所有权与防循环守卫的跨会话交接。
Read →阅读 → SystemThree escalating triggers (0.50 hygiene / 0.78 summarize / 0.92 hard), newest-first token-budget retention that truncates instead of dropping, and the validation guards that stop a bad summary from silently erasing history.三级递进触发（0.50 清理 / 0.78 摘要 / 0.92 强制）、最近优先的 token 预算保留（截断而非丢弃），以及防止坏摘要静默抹掉历史的校验守卫。
Read →阅读 → SystemA lightweight classifier picks the model per turn from an editable Markdown rules file — structure-over-surface judgement, tier-snap guards, and lane inheritance for terse follow-ups.轻量级分类器依据可编辑的 Markdown 规则文件为每轮选模型——看结构不看表面的判据、层级吸附守卫，以及简短追问的车道继承。
Read →阅读 → APIAn OpenAI-compatible endpoint that picks the model for you: send model:"auto", get routed across four difficulty lanes plus vision and video, and pay the real cost of whichever model answered.一个 OpenAI 兼容接口，替你选模型：发 model:"auto"，请求会被路由到四条难度腿及视觉、视频腿，按实际回答的模型真实成本计费。
One persistent agent per user — a real machine with a surviving workspace, credential injection via sc-proxy, deferred tool loading for cache discipline, and one identity across web and messaging channels.每个用户一个持久 agent——真实的机器：可持久化的 workspace、sc-proxy 凭证注入、为缓存纪律设计的延迟工具加载，以及横跨 Web 与消息渠道的同一身份。
Read →阅读 → PlatformProcedural memory as versioned markdown: three-tier discovery, patch-in-place lifecycle, community registry with PR-based governance — few comprehensive domain skills over many narrow ones.以版本化 Markdown 承载的程序性记忆：三层发现、就地修补的生命周期、基于 PR 治理的社区仓库——少而全的领域 skill 优于多而窄。
Read →阅读 → SystemOne dial (off / on / anchor / guard) for who does the grunt work: contract-first delegation to cheap executors, budget-stop salvage with evidence-gated partials, advisor consult/verify/takeover, and cache-safe anchor nudges.一个档位（off / on / anchor / guard）决定谁干粗活：契约先行的廉价委派、证据门控的预算截停打捞、advisor 咨询、裁决与接管，以及不破坏缓存的 anchor 软引导。
Read →阅读 → SystemScheduled tasks in two execution modes with mandatory cost estimates, the promise rule for background work, scripted delivery chains, and three-surface observability for credit drain.两种执行模式的定时任务与强制成本预估、后台工作的“承诺规则”、脚本化交付链，以及排查 credit 消耗的三面可观测性。
Read →阅读 → Platform2,100+ services agents sell and buy per call in USDC over the open x402 protocol — one-command monetization, fail-closed autonomous payments, and indexing of the Coinbase CDP / Bazaar ecosystem.2,100+ 服务，agent 之间通过开放的 x402 协议按次以 USDC 买卖——一条命令变现、失败关闭的自主支付，并索引 Coinbase CDP / Bazaar 生态。
Read →阅读 →