[Suggestion] Use one tool to fetch different skill details

[gkzhb]

This plugin creates a tool for each skill which may lead to many tools injected into llm prompts.

I found this from google:
Reverse Engineering Claude Code: How Skills different from Agents, Commands and Styles | by Vuong Ngo | Oct, 2025 | Level Up Coding and https://github.com/AgiFlow/claude-code-prompt-analysis/blob/a4ae566e796a193f3f593273e329c522dd416d35/data/skills.log for the llm prompt messages.

We can see Claude Code uses one tool "Skill" to fetch the details of a skill with the tool name passed as arguments.

So I modified the code to try to follow Claude Code's prompts and tool actions on https://github.com/gkzhb/opencode-skills/tree/feat/merge-tool

[malhashemi]

Hi @gkzhb Thanks for trying out the plugin and for the thoughtful feedback.

I actually gave this a lot of thought during the design phase, whether to have one unified skill tool or one tool per skill. Here's my thinking:

On the token/context argument:

If we merge into one skill tool, we'd still need to pass every skill name and description in that single tool's description (exactly as you proposed). So we're not actually saving tokens or reducing the LLM's decision-making load, the model still sees all the same information, just packaged differently.

The key benefit of separate tools:

OpenCode's strength is fine-grained agent control. With one tool per skill, you get precise control over which agents can use which skills.

Example use case in opencode.json:

{
  "tools": {
    "skills*": false
  }
}

Skills disabled for all agents by default.

Then in your content-creator agent:

mode: subagent
description: content creator
tools:
  skills_document_skills_pptx: true

This way, each agent only sees the skills it needs, which actually does reduce context overload vs the one skill tool.

---

That said, I'm keeping this discussion open because I'm genuinely open to counter-arguments. Let me know if I'm missing something or if there's an angle I haven't considered!

Really appreciate you taking the time to review and open this discussion.

[airtonix]

I think this is a key difference between how opencode permissions work vs how claudecode permissions work.