Personal blog / Field notes

JavaYu.

Notes on technology, ideas, and continuous learning.

Recent writing

Oct 8, 2026OpenCode / Plugin development / TypeScript / Open source
Inside OpenCode V1/V2 Plugin Loading: One Package for Two Versions and Local Debugging
OpenCode V1 and V2 use fundamentally different plugin mechanisms, and their plugin definitions are not interchangeable. However, the official migration guide provides a dual-version plugin pattern: by exposing separate V1 and V2 entry points, one npm package can support both versions. This article explains the differences between local directories, npm packages, and server entry points through concrete examples, then presents a packaging strategy for dual-version plugins.
Read article
Aug 14, 2026OpenCode / LLM / Local models / Open source
opencode-models-discovery 1.4.0: Better Model Metadata Injection, Caching, and Performance
From 1.0.0 to 1.4.0, opencode-models-discovery added multiple provider metadata enrichment modes. Alongside caching, timeout controls, and explicit configuration merging, these changes make dynamic model discovery in OpenCode more complete and reliable.
Read article
Aug 12, 2026Philosophy of Technology / Cybernetics / AI / Agent / Industrial Revolution / Productivity
AI: A New Kind of ‘Entropy Reduction’
From steam engines to large language models, explore how four industrial revolutions countered entropy increase—and how AI could become the core engine of next-generation civilisation by autonomously generating rules.
Read article
Aug 6, 2026LLM / Agents / Skills / Tool Calling
How LLM Initiative Works: From Language Intent to Controlled Execution
Initiative comes from the model's mathematical computation: it generates content directly or requests supplemental content; a client recognises the structure, controls execution, and returns results.
Read article
Aug 6, 2026LLM / Context Engineering / Tool Calling / Agents
How to Use LLMs to Their Strengths: Give Them a Real Environment
Use task context, history, persona, behavioural rules, and controlled tools to connect an LLM's language ability to real work.
Read article

View all writing

JavaYu.

Recent writing

Inside OpenCode V1/V2 Plugin Loading: One Package for Two Versions and Local Debugging

opencode-models-discovery 1.4.0: Better Model Metadata Injection, Caching, and Performance

AI: A New Kind of ‘Entropy Reduction’

How LLM Initiative Works: From Language Intent to Controlled Execution

How to Use LLMs to Their Strengths: Give Them a Real Environment