Why AI Coding Is Slower in Java and Five Steps to Build a Harness Environment

1. Experience Gap

In lightweight projects (frontend, CLI tools, local Python scripts) the AI coding loop—edit, run, test, AI reads results, auto‑fixes, re‑verify—runs entirely locally, allowing the AI to iterate dozens of rounds autonomously. In a Java micro‑service project the loop breaks because the code depends on OSS, remote sandboxes, and HSF services that are unavailable locally, forcing a manual push‑to‑pre‑release, wait, and feedback cycle that can take 5‑10 minutes per iteration.

2. Problem Essence

The root cause is the heavy reliance of Java micro‑services on cloud infrastructure (HSF, TDDL, Diamond/Switch, MetaQ). An @Autowired injection often pulls in a concrete cloud implementation, making the whole service impossible to run locally without those services.

3. Refactoring Principles

3.1 Dependency Inversion, Interface First

Abstract the upper‑level logic to depend on interfaces rather than concrete implementations. Replace direct OSS and sandbox calls with StorageAdapter and CommandExecutor interfaces, providing LocalStorageAdapter (java.nio.file) and LocalCommandExecutor (ProcessBuilder) for local runs. The factory selects the implementation based on parameters, leaving upper layers unchanged.

3.2 Zero‑Intrusion, Profile Isolation

Use Spring @Profile("local") beans for local implementations and @Profile("!local") for cloud ones, ensuring no extra code paths are added to the production flow. Component scanning excludes cloud‑only packages, and spring.autoconfigure.exclude disables auto‑configuration of middleware such as EagleEye, HSF, and OpenTelemetry.

3.3 Tool AI‑ification: CLI First

Expose internal tools via CLI so the AI can invoke them. Examples include mw diamond get … and mw hsf address … scripts that fetch Switch configuration and HSF service addresses, producing structured JSON that the AI can parse. Register these scripts as Skills in the AI Agent’s toolbox.

4. Practice Case: From "Push‑Pre‑Release" to "Local Closed‑Loop"

4.1 Before Refactor

The Agent platform relied on OSS for file storage, a remote sandbox for command execution, and cloud middleware for configuration and monitoring, making local development impossible.

4.2 Refactor Plan

Introduce a dual‑implementation architecture with interfaces, add @Profile beans, replace TDDL/MySQL with an H2 file database, substitute GroupSequence with an AtomicLong based LocalSequence, and fetch Switch/Diamond configs via the fetch-switch-config.sh script. Exclude cloud‑only auto‑configurations and use a simple start-local.sh script to launch the service.

4.3 Refactor Effect

After refactoring, the service starts locally with mvn spring-boot:run -Dspring.profiles.active=local, files are written to /tmp/agentfs/…, Bash commands run via ProcessBuilder, and the AI can now verify its own changes in seconds instead of minutes. The full bug‑fix cycle drops from 30‑minute, multi‑round manual work to a few‑second autonomous iteration.

5. Harness Engineering Checklist

Can the project start with a single command (e.g., mvn spring-boot:run)?

Are external middleware dependencies replaced with local substitutes (H2, in‑memory queues)?

Are external dependencies abstracted behind interfaces with profile‑based implementations?

6. Methodology Summary

Identify the minimal runnable subset (core request → LLM → tool → response).

Replace cloud services with real local equivalents (H2, local file system, real Bash).

Script every manual step (config fetch, start, verification).

Layered verification: compile → start → core API → end‑to‑end test.

Let the AI participate in the refactor, using the scripts as its hands.

7. Future Directions

Further automate JVM diagnostics (e.g., jstack, arthas watch) and feed their structured output to the AI, enabling it to understand failures without human intervention.

8. Appendix: Localisation Checklist

Key changes include:

Replace TDDL/MySQL with H2 ( application-local.properties).

Replace GroupSequence with LocalSequence using AtomicLong.

Fetch Switch config via fetch-switch-config.sh into switch-config-local.properties.

Exclude cloud auto‑configurations via spring.autoconfigure.exclude and ComponentScan filters.

Provide start-local.sh for one‑click startup and verify-local.sh for automated health checks.