Original Summary

I am desperate for feedback on my homebrewed harness, &quot;Plurnk.&quot;<p>I&#x27;ve been dogfooding it on my RTX5070Ti with only 16GB VRAM as my daily driver for weeks now and it&#x27;s genuinely better than everything else available (I&#x27;m biased).<p>1. Curation, not Compaction<p>The model&#x27;s fully responsible (bitter lesson) for deterministically curating its own context.<p>2. ANTLR Grammar<p>The dozen verbs for the harness are optimized for model pretraining, looking like markdown while being fully integrated with an EBNF grammar that plugs into an AST for superior workflow and recovery.<p>3. Universal Resources<p>Everything&#x27;s addressable by the model through a pseudo-URI interface, including the log entries.<p>4. Omnipatterns<p>Everything in the repo goes through treesitters which build a complete graph that the model can search with glob, regex, jsonpath, xpath, graph, and sqlite fulltext.<p>5. Standards<p>MCP2 tools, AG-UI client interface, A2A &quot;agent to agent&quot;, plugs into everything through OAI spec + models.dev<p>6. Interfaces<p>CLI, TUI, and a Neovim plugin, easy to extend, very pluggable<p>7. Lean Agent<p>Service&#x2F;Client architecture with lean TUI allows dozens of agents to run on weak hardware. 200MB footprint.<p>8. Lean Context<p>Achieves all of the &quot;batteries includes&quot; features of a fat harness while the sysprompt is only slightly larger than pi agent&#x27;s.<p>It started as a proof of concept for my bespoke opinions on what &quot;bitter lesson&quot; actually means, and it&#x27;s gone so well that I believe others trying to build agentic workflows, especially with local and humble constraints, could benefit from the proven concepts.

中文概览

中文标题： Show HN：Plurnk（又一个 AI Harness）

作者展示自制AI harness Plurnk，称其在16GB显存的RTX5070Ti上作为日常工具使用数周，并列出精选上下文、ANTLR语法、伪URI资源、Omnipatterns、标准、CLI/TUI/Neovim接口及轻量代理等特性。

  • 情报分类：开源项目与落地
  • 分类依据：Show HN 展示自制AI harness及其特性，属于项目展示。
  • 信息来源：Hacker News 新项目
  • 发布时间：2026/9/16 07:57:14