prompt-router

Classify user prompts into routing tiers and rewrite messages with this OpenCode plugin. It uses chat.message to transform user text.

Install

Run these commands to install:

cd /home/dzack/opencode-plugins/prompt-router
just install

Register the plugin via file: in your OpenCode config:

{
  "plugin": [
    "file:///home/dzack/opencode-plugins/prompt-router/src/index.ts"
  ]
}

View a sample configuration here: prompt-router/.config/opencode.json

MCP: None. This package provides a chat-transform hook rather than a tool server.

Agent Surface

This plugin intercepts chat messages without exposing tool names. It performs these actions:

• Reads the latest user text.
• Classifies input into tiers: model-self, knowledge, C, B, A, or S.
• Injects instructions from the canonical response template.

Dependencies:

• Runtime: Bun, @opencode-ai/plugin, yaml
• External local assets: ~/ai/prompts/...
• External local runtime: ~/ai/opencode/.venv, which provides llm-run and llm-template-render from the GitHub-backed llm-runner and llm-templating-engine dependencies declared in ~/ai/opencode/pyproject.toml

LLM Integration

prompt-router does not call the legacy ~/ai/scripts/llm bridge. It shells into the standalone JSON CLIs instead:

• llm-run for prompt execution and structured classifier output
• llm-template-render for response-template rendering

That means the local OpenCode environment must be synced first:

cd /home/dzack/ai/opencode
uv sync --dev

For ad hoc smoke tests outside the project environment, the canonical upstream CLIs are also available directly from GitHub via uvx --from:

uvx --from git+https://github.com/dzackgarza/llm-runner.git llm-run --help
uvx --from git+https://github.com/dzackgarza/llm-templating-engine.git llm-template-render --help

Checks

Run checks with just:

just typecheck
just test