I build agent systems in Go & Rust — production AIGC pipelines, open-source execution infrastructure, and the Sui / Walrus layer that gives agents wallets, memory, and tamper-evident audit trails.
Most agent projects stop at the demo. I care about what comes after: how an autonomous workload gets bounded, audited, and explained once it runs unattended.
🔭 Now — shipping the aigc-server agent pipeline (V1.0) · AgentOS audit / replay layer · Rust async internals
Go & Rust backends, agent systems you can operate
| Project | Stack | What it does |
|---|---|---|
| Production · agent systems | ||
| aigc-server production · private | Go Eino |
AIGC video platform for a social app — controlled-autonomy generation pipeline: the LLM proposes, the backend decides (full walkthrough below) |
| Open source · agent infrastructure | ||
| agentos | Go Rust |
Self-hosted agent execution platform — task orchestration & lifecycle, local / NATS dual scheduling, audit & replay APIs, SSE telemetry, agent loop with tool calling |
| sentinel-protocol | Go Move |
Verifiable pre-execution security for OpenClaw agents — policy gate + multi-signal risk engine → ALLOW / REQUIRE_APPROVAL / BLOCK, decisions anchored to Sui as tamper-evident evidence |
| Open source · on-chain agents | ||
| sui-nexus | Go Move |
Settlement infrastructure for the AI agent economy on Sui — HMAC + zkLogin auth, Move-enforced wallet policy (not middleware), Walrus-backed memory, PTB atomic execution. Sui Overflow 2026, live on testnet |
| memwal · PyPI | Python Move |
One import gives any AI agent persistent, cross-session, verifiable memory — Walrus storage + Sui on-chain references, framework-agnostic (pip install memwal) |
| Go / Rust tooling | ||
| git-ai-cli · npm | Rust |
Branch-aware AI Git assistant — commit generation, Copilot guard, weekly digests. DeepSeek / Qwen / GLM / Moonshot / Ollama, fully local-first |
| ai-tracker private | Rust |
Observability for AI coding agents — sessions, streaks and tool / model usage across Claude Code, Codex, Cursor |
| alpha-research-agent private | Go |
Crypto research agent — aggregates market, on-chain and DEX data into briefing / deep-dive / long-form outputs |
| release-pilot private | Go TypeScript |
GitHub Release notification bot — subscribe to repos, filter by rules, render templates, push to channels |
Also: EcoPilot Python — sustainable CI/CD agent for GitLab Duo · MyNote — agent-development knowledge base (private) · the rest in the full repository list.
aigc-server — agent-driven AIGC video platform (code private)
A social app's AIGC video platform, built on one principle: the LLM proposes, the backend decides. The model only ever emits a structured decision contract — every side effect (spending credits, submitting a generation, cancelling, redoing) passes backend validation, a confirmation gate, and a deterministic execution chain.
What makes it hold up in production:
- Idempotent end to end — SHA-256 request-hash dedup at the command queue, per-tool audit replay inside the runtime loop, and
agent-{session}-rev{r}-step{n}request IDs so a director-mode re-run resumes per shot instead of paying twice. - Multi-shot by default — any video plan with more than one shot is promoted to director mode at compile time, server-side; every video plan always has a shot layer, credits accumulate per shot, and a partial failure returns to the confirmation gate for resumable continuation.
- Bounded autonomy — ≤4 LLM calls and ≤8 tool calls per turn under a hard deadline; read-only tools run automatically while side-effect tools require approval; the terminal-decision validator gets exactly one repair round.
- Pluggable providers — ComfyUI / Bailian / Vidu / RunningHub behind one abstraction: submission consumes credits, failure refunds, and swapping a provider never touches the core.
The execution path I keep rebuilding
Every action an agent takes passes a policy gate before it runs — and every decision leaves evidence:
Stats, PR graph & the snake
|
|
|
|
- GoPolymarket/polymarket-go-sdk: fix(clob): restore json.Number salt for POLY_1271 order payloads
- hoh-zone/Overflow2026-CNNo1: add project description
- GoPolymarket/polymarket-go-sdk: fix(clob): support flexible order time fields & make govulncheck advisory
- MystenLabs/sui: [types][cleanup] fix typo: barreier -> barrier in governance.rs
- MystenLabs/sui: fix(kvstore): decode balance_changes and unchanged_loaded_runtime_objects from KV store
🀄 中文简介
我是 dongowu，用 Go / Rust 做 Agent 系统：生产环境的 AIGC 业务 pipeline、开源的 Agent 执行基础设施，以及用 Sui / Walrus 给 agent 加钱包、记忆与不可篡改审计锚点——关心调度、治理、可审计、可观测这些"演示之后"的工程问题，让一个自主运行的工作负载可以被约束、审计和解释。
- aigc-server · Agent 子系统（生产项目，代码私有，架构图见上方 Production system）：社交 App 的 AIGC 视频平台，一条受控自主创作 pipeline — 模型只出结构化 Decision，所有副作用都过后端校验、确认门与确定性执行链；PlanCompiler 服务端编译多镜方案并自动提升 director 模式，幂等键到镜级、断点续跑不重复扣费，可插拔视频 Provider（ComfyUI / 阿里百炼 / Vidu / RunningHub）
- AgentOS：自托管 Agent 执行平台 — 任务编排与执行生命周期、本地 / NATS 双调度路径、审计与回放 API、SSE 遥测、agent loop 与工具调用
- Sentinel Protocol：面向 OpenClaw + Sui 的可验证执行前安全层 — 策略门 + 多信号风险引擎给出 ALLOW / 需审批 / 拦截，决策经哈希链与 Merkle 批次锚定上链，形成不可篡改的审计证据
- sui-nexus：Sui 上 AI agent 经济的结算基础设施 — HMAC + zkLogin 认证、Move 层强制的钱包策略、Walrus 记忆、PTB 原子执行，已部署 testnet（Sui Overflow 2026）
- memwal（PyPI）：一行导入即给任意 AI agent 持久、跨会话、可验证的记忆 — Walrus 存储 + Sui 链上引用，框架无关
- git-ai-cli（npm）：分支感知的 AI Git 助手，支持 DeepSeek / 通义千问 / 智谱 GLM / Moonshot / Ollama，可完全本地部署，代码不出内网
其余项目见上方 Featured work。
🐍 snake by Platane/snk · 📊 cards by readme-tools · 🏷️ badges by shields.io