Prompt Router

An OpenCode plugin that automatically matches user prompts to relevant skills (SKILL.md files) using TF-IDF text scoring. When a user sends a message, the router scores all discovered skills and injects a preamble instructing the AI to load the best matches.

How it works

1. Discovery -- Recursively finds SKILL.md files in ~/.agents/skills/, ~/.claude/skills/, and <workdir>/.opencode/skills/
2. Parsing -- Extracts YAML frontmatter (name, description, tags) from each skill file
3. Enrichment -- Auto-derives tags from skill body content for skills without explicit tags
4. Scoring -- Two-stage TF-IDF scoring:
   • Stage 1: Weighted field matching (name ×3, tags ×2, description ×1) with suppressor filtering and IDF floor
   • Stage 2: Body scan bonus for borderline matches
5. Routing -- Returns top N matches above the minimum score, formatted as a preamble prepended to the user's message

Setup

bun install

Add to your opencode.json:

{
  "plugin": [
    ["opencode-prompt-router", { "minScore": 15, "debug": true }]
  ]
}

Or using a GitHub reference directly:

{
  "plugin": [
    ["github:anderssv/opencode-prompt-router", { "minScore": 15, "debug": true }]
  ]
}

Note: Using a GitHub reference in the plugin array means OpenCode runs bun install at startup for each workspace. On the Desktop version this can add noticeable load time. If that's a problem, install via a package.json in ~/.config/opencode/ and use a one-line shim in ~/.config/opencode/plugins/ instead:

// ~/.config/opencode/package.json
{ "dependencies": { "opencode-prompt-router": "github:anderssv/opencode-prompt-router" } }

// ~/.config/opencode/plugins/prompt-router.ts
import type { Plugin } from "@opencode-ai/plugin";
import { PromptRouter as _PromptRouter } from "opencode-prompt-router";
export const PromptRouter: Plugin = (ctx) => _PromptRouter(ctx, { debug: true });

Configuration

Option	Default	Description
minScore	15	Minimum TF-IDF score to surface a skill
maxPromptLength	500	Prompts longer than this are skipped
debug	false	Enable visible preamble output and detailed logging

Set PROMPT_ROUTER_DEBUG=1 as an alternative to enable debug mode via environment variable.

All matches are logged to ~/prompt-router.log.

Testing

bun test

Tests include unit tests for each module, precision regression tests against real skill corpora, and approval-based false-positive tests.

Inspiration

This project was inspired by The Prompt Router — a 47ms keyword classifier for context selection.

Project structure

index.ts              Plugin entry point (hooks into chat.message)
core/
  tokenizer.ts        Text tokenization + basic stemming
  parser.ts           YAML frontmatter parser for SKILL.md
  discovery.ts        Recursive SKILL.md file finder
  cache.ts            Mtime-based skill cache
  corpus.ts           IDF index builder
  enrich.ts           Auto-derives tags from body content
  scorer.ts           Two-stage TF-IDF scoring
  router.ts           Orchestrator: discover → score → format
  config.ts           Default weights, suppressors, thresholds
  types.ts            Type definitions
tests/
  *.test.ts           Unit and regression tests
  fixtures/skills/    Fixture SKILL.md files
  approvals/          Approval test snapshots