better-LVGL-for-dsh

English | 中文

An offline LVGL 8.4 knowledge skill pack for DeepSeek Harness (dsh) — it teaches the agent to write correct LVGL v8 API code and, above all, to stop mixing in v9 APIs (the single most common failure mode of LLM-generated LVGL code).

• Plugin form: dsh bundle (dsh.bundle declaration in package.json + cordis.patch.yml patch row)
• Capability path: Cordis plugin → runtime skills registered on ctx.skills → the stock dsh-tool-skill consumer → the model-facing skill catalog and skill loader
• Skill format: Agent Skills standard (SKILL.md + YAML frontmatter); pure ESM, zero dependencies, no build step

Why

Large language models have seen far more LVGL v9 content than v8.4, so they routinely produce code that mixes the two — e.g. lv_display_create() (v9) next to lv_disp_drv_t (v8), or lv_async_call() which does not exist in v8.4 at all. This pack pins the agent to v8.4 facts with nine focused, on-demand skills.

Skills

Installation

# From GitHub
dsh plugin --profile web add github:sa998aaron/better-LVGL-for-dsh

# Or from a local clone
dsh plugin --profile web add <absolute path to this repo>

Reload the profile afterwards. The lvgl84-* skills then appear in the model-facing skill catalog and load on demand; you can also invoke one explicitly with a /lvgl84-style-style gesture.

Verify it works

Ask:

Write an LVGL 8.4 UI with a title bar and three buttons.

If the generated code uses lv_obj_create / lv_obj_set_flex_flow / lv_obj_add_event_cb — and not v9's lv_display_create — the plugin is doing its job.

Layout

better-LVGL-for-dsh/
├── package.json        # dsh.bundle declaration
├── cordis.patch.yml    # Cordis composition patch row
├── index.js            # Entry: scans skills/ and registers each on ctx.skills
└── skills/<name>/SKILL.md

Adding your own skills

Create skills/<kebab-case-name>/SKILL.md:

---
name: my-skill
description: One line on what it is and when to use it.
whenToUse: Optional trigger description
---

Body (Markdown)…

Reload the plugin; the directory is rescanned at apply time.

License

MIT