Solon 4.0.3 Released: AI Loop, Code Talent, JDK 25 Prep

On July 2, 2026, the Solon team released v4.0.3 - just 22 days after v4.0.0 and 7 days after v4.0.2. Three releases in under a month. The pace tells a clear story: Solon's iteration speed is accelerating, and the focus is firmly on AI Agent infrastructure. Here's what's new and why it matters.

What's in v4.0.3

The Three Changes That Matter Most

1. solon-ai-loop: The Agent's "Iteration Engine"

This is the most significant new module in this release. What problem does it solve? In ReActAgent and TeamAgent, reasoning steps are linear - think, act, observe, done. But many real-world scenarios require agents to loop:

• Code generation: write code → compile fails → read error → fix → recompile → until green
• Automated testing: write tests → run → analyze failure → fix → rerun → until pass
• Data cleaning: clean → validation fails → adjust params → clean again → verify → until合格

solon-ai-loop provides a controlled loop execution framework, allowing agents to iterate until a termination condition is met. It works with the existing StopLoopInterceptor to control max iterations, timeouts, and even manual interruption.

2. solon-ai-talent-code: Code Abilities as a Standalone Module

Before v4.0.3, code-related talents (file reading/writing, shell execution, LSP analysis, etc.) were bundled inside Harness. If you only needed ChatModel + code abilities, you had to pull in the entire Harness. Now solon-ai-talent-code is independent, giving you a lightweight, focused dependency:

<!-- Code abilities only, no Harness needed -->
<dependency>
    <groupId>org.noear</groupId>
    <artifactId>solon-ai-talent-code</artifactId>
    <version>4.0.3</version>
</dependency>

This makes Harness leaner, and makes code talents reusable in other contexts - like SolonCode CLI, which is itself Harness-driven.

3. JDK 25 ScopeLocal Auto-Loading

JDK 25 is the upcoming LTS release (September 2026), and Solon has a tradition of pre-adapting before the official JDK release. The ScopeLocalJdk25 auto-load mechanism means: when you upgrade to JDK 25, Solon automatically switches to JDK 25's scoped value implementation - zero code changes, zero config. This is "compatible with Java 8 ~ Java 25" in action - not just running on new JDKs, but actually leveraging their capabilities.

From M1 to GA: 11 Days

v4.0.3-M1 (milestone) was released in late June 2026, and it took just 11 days to reach GA status. The v4.0.x release cadence:

Three GA releases in under four months tells us two things:

• The v4.x core stack is stable - each release is incremental, not firefighting
• AI ecosystem is the primary iteration focus - all three releases center on AI modules

v4.0.x LTS Support

Per the official LTS plan (solon.noear.org/article/lts):

v4.0.x is the current main line, with full LTS guarantees. Expect a minor release roughly every 20 days, with each minor version line having independent LTS maintenance. If you're on v3.x in production, v3.10.x is still actively maintained - no rush to migrate.

What This Means for the AI Ecosystem

Modularization Progressing

• v4.0.0 → skill renamed to talent (unified naming)
• v4.0.2 → multiple new AI plugins added
• v4.0.3 → talent-code extracted, AI Loop added

Solon AI's modularization follows a "unify then split" pattern - first standardize the concepts, then extract reusable modules.

Loop Capability Filled

Previously, Solon AI lacked a loop abstraction. Developers had to write custom loop logic inside onObservation, leading to code duplication and poor control over termination, timeout, and resource cleanup. solon-ai-loop fills this gap. Combined with StopLoopInterceptor, you now have clean, controlled iteration:

// Conceptual example
Harness.harness(agent)
    .addInterceptor(new StopLoopInterceptor(maxIterations))
    .loop(context -> {
        return agent.step(context);
    })
    .start();

Paving the Way for SolonCode

The extraction of solon-ai-talent-code directly benefits SolonCode CLI - Solon's AI coding assistant. SolonCode is Harness-driven under the hood. With code talents as a standalone module, SolonCode can depend on talent-code without pulling in the full Harness weight, improving CLI startup time and reducing package size.

Upgrade Path

New Projects

Start directly with v4.0.3:

<dependency>
    <groupId>org.noear</groupId>
    <artifactId>solon</artifactId>
    <version>4.0.3</version>
</dependency>

From v4.0.2

Compatible upgrade - just bump the version number. No breaking changes.

From v4.0.0

If you're still using the old solon-ai-skill-* package names (pre-v4.0.0), migrate to solon-ai-talent-*. The migration guide was published alongside v4.0.0.

From v3.x

v3.x → v4.x changes are concentrated in AI modules and package restructuring. The core framework (solon, solon-web, etc.) is largely compatible. Validate on a test project first.

What's Next

Based on the current iteration direction and community discussions:

• solon-ai-talent-code will continue evolving - expect deeper LSP integration and Git operations
• MCP protocol tracking - the MCP protocol itself is evolving fast, Solon will keep pace
• v4.0.4 is already on the roadmap - roughly late July at the current cadence

Final Thoughts

v4.0.3 isn't a "big" release. But it's a revealing one: the Solon team is using high-frequency iteration to rapidly fill capability gaps in the AI ecosystem. In the 22 days from v4.0.0 to v4.0.3, we've seen:

• Unified naming (skill → talent)
• Core capabilities extracted (talent-code)
• Loop engine built (AI Loop)
• Next-gen JDK support (JDK 25)

Each release is small. But each one moves the needle.

Links

• GitHub Release v4.0.3
• Solon Official Site
• LTS Plan
• Solon AI Ecosystem Overview