杨澜访谈录

New Zealand mayor's ‘China‑inspired’ solar project highlights potential for cooperation_我的网站

nba

一 |

Solar rooftop photovoltaic power generation facilities are seen on a building in Qingdao, East China's Shandong Province, on June 23, 2026. Photo: VCG
Some Western politicians have been smearing Chinese green products and even advocating for a "decoupling," but reality tells a different story. Recently, a mayor in New Zealand is reportedly pushing for a solar upgrade at energy-hungry public facilities after witnessing advanced green infrastructure during his China tour.
Christchurch Mayor Phil Mauger wants to turn some of the city's biggest power-hungry facilities into their own electricity generators, with a giant solar project at the wastewater treatment plant the first step in an ambitious plan inspired by a trip to China, the New Zealand Herald reported.
The mayor has ordered a report into whether solar panels could be suspended above the treatment plant in Bromley, creating a vast power-generating canopy capable of helping run one of the city's largest energy users. If successful, Mauger sees potential for similar installations at major facilities, where vast roofs and car parks could be harnessed to generate solar power.
Mauger's push for solar infrastructure was inspired by his recent visit to China, during which an innovative solar installation above a wastewater treatment plant caught his attention and sparked ideas for Christchurch, according to the New Zealand Herald.
Back in June, Mauger toured the wastewater treatment plant in Wuhan, Central China's Hubei Province, according to Hubei Daily. His itinerary included the central control station, solar-powered equipment room and sewage-treatment pools.
"What impressed us most were the solar panels mounted on the top, which supply electricity for the whole system. We have learned a great deal about its simple operation procedures and will hold relevant discussions afterwards," Mauger was quoted as saying in a local media report.
Chen Hong, a researcher at the Australian Studies Center at East China Normal University, told the Global Times on Wednesday Christchurch's planned project demonstrates that China-New Zealand cooperation can extend beyond conventional trade into city-level governance.
"Collaborative initiatives in clean-energy deployment, sewage treatment and waste disposal can create an innovative cooperation framework integrating technical solutions, engineering capacity and on-the-ground operational expertise," Chen said.
He added that this real-world case lays bare the fundamental divide between geopolitical narratives advocating so-called decoupling from China and the practical realities of local governance.
"If a Chinese solution actually works, ideological labels shouldn't get in the way. This case shows that the whole 'decoupling' talk is mostly political rhetoric. At the end of the day, economies run on cost, efficiency, and results," Chen said.
Global Times
。智猩猩AI整理编辑：没方最近在折腾几款 AI 编程工具，很多人第一反应都是问背后用的是GPT-5.6 Sol还是Claude Fable 5。但相同模型，放进不同的编程工具里，表现可能完全不一样。有的 Agent 能连续工作几个小时，写完代码还知道运行测试、检查结果、继续优化。有的写了几十行代码，就告诉你任务已经完成。等真正运行项目，才发现测试没跑、依赖没装，甚至文件都改错了。模型明明没有变，差距从哪来？答案可能藏在模型外面的那层 Agent Harness 里，可以理解为包裹在大模型外部的“智能体运行系统”。它负责把模型、上下文、记忆、工具、权限、终端界面、任务调度和多智能体协作串联起来。模型决定 Agent 的能力下限，而 Harness 则在很大程度上决定模型能够稳定工作多久、一次能处理多复杂的任务，以及最终能把多少智能转化成真实生产力。今天刷github日趋榜，看到了个名叫 jcode 的开源项目，正试图把 Agent 这层外壳推到更远。它是一套使用 Rust 构建的终端 Coding Agent Harness，主打多会话工作流、持久记忆、多智能体协作、高度自定义和低资源占用。在github上已经收获10k stars。

二 | 项目链接： https://github.com/1jehuang/jcode 01 从单次对话，走向长期运行的Agent工作系统记忆系统是jcode的一大杀手锏，下面这段演示视频可以直观看到它的记忆召回效果：播放下一个打开循环播放 00:00 / 00:00 倍速 3.0X 2.0X 1.5X 1.25X 1.0X 0.75X 0.5X 多音轨 AirPlay 0 静音播放中，点击恢复音量画中画网页全屏全屏你可以刷新试试视频信息 1.33.6 播放信息上传日志调试信息 [X] 视频ID VID - 播放流水 Flowid - 播放内核 Kernel - 显示器信息 Res - 帧数 - 缓冲健康度 - 网络活动 net - 视频分辨率 - 编码 Codec - mystery mystery - 按住画面移动小窗 jcode 记忆功能演示它会将每轮对话转换为语义向量，并利用记忆图谱管理这些信息。当新任务到来时，系统会通过余弦相似度检索与当前对话有关的记忆，再将命中的信息注入上下文。为了降低“相似但不相关”的记忆干扰，jcode还可以调用独立的Memory Sideagent，对召回内容进行二次判断。这个Sideagent既可以验证记忆是否真正相关，也可以继续执行信息检索，再把整理后的结果交给主Agent。记忆的写入也不完全依赖用户操作。当对话发生明显语义漂移、间隔一定轮次或者会话结束时，记忆Sideagent可以从对话中抽取值得长期保存的信息，写入记忆图谱。

三 | 在后台模式（Ambient Mode）下，系统还会定期整理和整合记忆，检查过期信息与潜在冲突。 jcode将多会话工作流作为核心设计目标。用户可以让jcode作为后台服务器持续运行，再从不同终端连接到同一服务。多个会话可以复用同一个后台服务，并由服务端统一管理会话连接、Agent通信和协作状态，而不需要为每个终端都重新启动一套完整的运行环境。根据项目公布的自测结果，按照PSS内存指标计算，在关闭本地嵌入功能时，jcode运行一个活跃会话的内存占用约为27.8MB。而Claude Code约为386.6MB，约是 jcode 的13.9倍。当活跃会话增加到10个时，两者的资源差距更加明显：jcode的内存占用约为117MB，Claude Code则达到约2300.6MB，约为jcode的19.7倍。从会话扩展成本来看，jcode每增加一个会话，额外内存占用约为9.9MB；Claude Code每增加一个会话则需要约212.7MB，约为jcode的21.5倍。 jcode的另一项重要能力，是原生Swarm多智能体协作。

四 | 当Agent A修改了Agent B此前读取过的文件，服务器会主动通知Agent B，提醒它当前代码已经发生变化。Agent B可以根据任务相关性忽略提醒，也可以查看Diff，确认自己的修改是否需要调整。 Swarm协作文件变更通知不同Agent之间还可以发送私信、广播消息和同步任务状态。主Agent也能够自主创建新的Agent团队，把工作拆分给不同成员并行处理，自己则转变为负责规划、调度和汇总结果的协调者。

五 | 为了降低大仓库中的上下文消耗，jcode还开发了面向Agent的代码搜索工具Agent Grep。该工具会在搜索结果中加入文件结构、函数列表和函数所在位置等信息，让模型在不读取全文的情况下，大致判断相关代码承担什么作用。还会根据Agent已经看过的内容，自适应地截断重复搜索结果。 Skills也采用了类似的按需加载思路。jcode不会在启动时把所有Skill完整写入上下文，而是对当前对话进行语义匹配，在检索到相关Skill时再自动注入。用户和Agent仍然可以通过工具或斜杠命令手动激活Skill，但无关的说明不会长期占据上下文窗口。浏览器能力同样纳入了Harness设计。目前jcode内置的浏览器工具主要通过Firefox Agent Bridge工作，可以执行打开页面、获取内容、点击、输入、填写表单、截图、滚动和上传文件等操作。

六 | jcode最有趣的功能是Self-Dev模式。进入这一模式后，Agent可以直接修改jcode自身的源代码，随后完成构建和测试，重新加载生成的二进制程序，并继续处理原有的一个或多个会话。如果用户不满意界面布局、工具行为或者某项工作流，可以让Agent直接修改Harness本身，而不必完全等待项目维护者提供新功能。这让人想到翁荔最近提出的一个判断，在模型能够稳定修改自身权重之前，AI的自我改进可能会率先发生在Harness层。

七 | 《DeepSeek崔添翼力挺翁荔最新长文：AI自进化先从Harness开始》 jcode的Self-Dev正是提供了象征意义的工程样本。 02 安装教程（1）macOS 和 Linux `curl -fsSL https://jcode.sh/install | bash` （2）Windows 11 `irm https://jcode.sh/install.ps1 | iex` 安装完成后，在终端输入 `jcode` 即可启动交互式界面。 03 总结总的来说，jcode能够为Coding Agent搭建一套更完整的工作系统，让智能体记得更久、协作得更好，也能在用户授权下按需改造自己的运行环境。