opencode-triage

OpenCode already lazy-loads full skill bodies — but it still injects every skill’s name and description (from SKILL.md frontmatter) into the prompt on every message. With a growing skill library and verbose descriptions, that can mean hundreds of tokens burned before you type a word. opencode-triage eliminates that XML list entirely, replacing it with a compact tool definition and routing on-demand via keyword matching.

Quick Start

npm install -g opencode-triage

Restart OpenCode, then type /triage on. That’s it.

What It Does

OpenCode lists every skill’s name and description in the prompt on every message. With many skills, that burns hundreds of tokens before you type a word.

Triage hides all skills from the prompt. When you need one, the LLM calls triage() — the plugin finds the best match by keyword scoring and loads only that skill’s instructions.

No triage:   [name+desc of A] [name+desc of B] [name+desc of C] ...  ← every message
With triage: triage({ query }) → keyword match → loads one skill body when needed

The more skills you install, the more triage saves. Each new skill adds tokens to the native prompt, but zero tokens when triage is on.

LLM Quick-Install

Copy-paste this into any LLM:

Install opencode-triage (https://github.com/cascharly/opencode-triage) — a deterministic skill router for OpenCode.

1. Run: npm install -g opencode-triage
2. If any errors occur, visit the repo link above and resolve them.
3. Tell me to restart OpenCode.

Commands

Note: Skills are hidden via hooks at the LLM prompt level — SKILL.md files stay intact on disk, so other AI tools (Cursor, Claude Code, Windsurf) still see them. File renaming (SKILL.md ↔ SKILL.md.disabled) is only used as a fallback for older OpenCode versions without hook support. When files are renamed to .disabled, they will also be hidden from other AI tools that scan the same directories.

Flags

All commands also work in your terminal: npx opencode-triage . No OpenCode session needed.

Configuration Combinations