ZCode Sparks the Next Wave of AI Programming: From Code Generation to Full Delivery

AI programming moves beyond “write a method”

Two years ago AI coding tools were used mainly for code completion: generating method bodies from signatures, creating Controllers/DTOs/Services from brief requirements, or explaining error logs. That kept the human in charge of the engineering process while AI handled isolated snippets.

Now tools such as Codex, Claude Code, Cursor and ZCode aim for developers to hand over an entire task rather than a single piece of code.

Example instruction (Chinese) to fix duplicate refund records in order-service with numbered requirements (analysis, root‑cause identification, minimal changes, add concurrency tests, run Maven tests, output modified files, test results, remaining risks).

The agent must search the codebase, understand call chains, modify files, run commands, read failure logs, iterate until the task is completed, delivering a merge‑ready change set.

Why Java projects need “complete delivery”

Spring Boot services consist of many interconnected components (Controllers, DTOs, Services, domain objects, repositories, XML mappers, databases, Redis, MQ, tests). Changing an order status can affect front‑end APIs, scheduled jobs, and message consumers.

@Transactional
public RefundResult createRefund(RefundCommand command) {
  Order order = orderRepository.findByOrderNo(command.orderNo());
  order.changeStatus(OrderStatus.REFUNDING);
  refundRepository.save(RefundRecord.create(order, command.amount()));
  refundProducer.send(order.getOrderNo());
  return RefundResult.success(order.getOrderNo());
}

Beyond the method, a real project must verify user ownership, amount limits, duplicate‑request handling, MQ failure handling, idempotent status updates, and compatibility with historical orders—concerns that span multiple files and services.

ZCode emphasizes long‑context and long‑task capabilities to keep track of project structure, constraints, terminal output, and incremental modifications across files.

ZCode’s core capabilities

Self‑built agents

Long‑task execution

Skills and plugins

Multi‑agent coordination (MCP, sub‑agents)

Remote development via Feishu, WeChat, mobile

These capabilities correspond to features in Codex (multi‑agent worktrees), Claude Code (terminal and sub‑agents), and Cursor (cloud agents and mobile), but ZCode integrates model, tooling, remote control, and team workflow into a single desktop environment.

1 M token context for Java projects

GLM‑5.2 is marketed with a million‑token context window. For large multi‑module Spring Boot projects containing thousands of files and extensive configuration, a longer context reduces the risk of forgetting earlier constraints.

Effective agents first search the repository, then retrieve only relevant files, balancing capacity with accuracy. Loading the entire repository can introduce irrelevant noise.

Preparing AGENTS.md and Skills

Agents need explicit documentation of team engineering rules. Teams should provide an AGENTS.md that lists project conventions, module boundaries, development standards, prohibited actions, and verification commands.

# 项目规则
# 模块边界
- order-service负责订单业务。
- payment-service负责支付与退款流水。
- common-core禁止在普通业务任务中修改。

## 开发规范
- Controller只负责参数校验和协议转换。
- 业务逻辑必须进入Service或Domain层。
- 金额统一使用BigDecimal，数据库以分存储。
- MyBatis XML不得拼接未经校验的用户输入。

## 禁止事项
- 禁止修改application-prod.yml。
- 禁止删除已有测试。
- 禁止为了通过测试降低业务校验。
- 禁止直接执行生产数据库写操作。

## 验证命令
mvn -pl <module> -am test

Repeated tasks can be encapsulated as Skills (e.g., a Spring Boot interface creation skill that enforces parameter validation, permission checks, exception handling, testing, and compatibility). ZCode also supports a SKILL.md to describe trigger scenarios, steps, and output requirements.

Example ZCode task template for Java

任务背景：order-service新增了订单来源筛选条件，但部分请求返回空列表。
任务目标：定位参数传递或MyBatis动态SQL中的问题，完成最小修复。
允许修改：
- order-service中的查询DTO
- OrderQueryService
- 对应Mapper接口和XML
- 相关单元测试
禁止修改：
- common-core
- payment-service
- application-prod.yml
- 数据库现有字段
执行步骤：
1. 先分析Controller到Mapper的调用链。
2. 找到根因后输出修改计划。
3. 确认修改范围后再编辑文件。
4. 修改完成后运行： mvn -pl order-service -am test
5. 不得删除或降低已有测试断言。
6. 最后输出根因、修改文件、测试结果和风险。

This bounded task lets teams evaluate whether an agent respects module boundaries, runs the correct tests, and produces an audit‑ready Git diff.

Remote control via Feishu and WeChat

Bot integration allows developers to monitor and supplement tasks away from the IDE, useful for long‑running operations such as Maven tests, CI failure analysis, dependency upgrades, and log investigation.

Security considerations include enforcing permissions: allow read‑only code access, test execution, branch creation, pull‑request generation; forbid reading production secrets, connecting to production databases, direct deployments, or high‑risk deletions.

Enterprise considerations for domestic AI programming tools

Key concerns for adoption: handling of private repository code, log and context privacy, enterprise account and permission management, private Maven repository access, intranet environment support, auditability of agent commands, and cost control.

Prudent adoption strategy

Start with a non‑critical project, define 5‑10 realistic tasks (bug fixes, test additions, fault analysis, small feature development), record success rate, modification scope, test outcomes, manual rework time, and cost. Expand to core modules only after stable results.

The next competition in AI programming

The trend is shifting from code assistants to full engineering platforms. Future tools will be judged on their ability to understand project rules, orchestrate multiple agents, connect terminals, browsers, and MCPs, and operate across desktop, mobile, and collaboration software.

For Java developers, the skill shift is from “how to get AI to generate code” to “how to build a reliable, agent‑driven engineering workflow”.