⚡ OpenCode (v1/v2) Async Prompt Injector

Slide a sticky note under your agent's nose while it's busy thinking.

No interruption. No restart. The agent picks it up at the earliest opportunity.

You know that feeling when someone's mid-story, and you suddenly remember something important… but you don't want to cut them off?

Working with an AI agent is the same — except the agent is the one on a roll. You forgot a detail. A new idea just hit. You want to nudge the direction. Waiting feels wasteful. Interrupting just burns credits.

This plugin lets you drop a message that the agent reads mid-task, without breaking its flow.

✨ How it works

1. Agent is thinking / using tools.
2. You type /inject_prompt <your message> and send.
3. The agent reads it and acts on it within the current flow.

That’s it.

📦 Installation

Add it to your opencode.json — v1 and v2 both load this package:

// OpenCode v1
{ "plugin": ["opencode-async-prompt-injector"] }

// OpenCode v2
{ "plugins": ["opencode-async-prompt-injector"] }

Restart OpenCode.

🚀 Usage

• /inject_prompt Hey, I just updated the tests — check them too
• /inject_prompt Change of plans: postpone item #7 until the next round

Your message stays visible exactly as you typed it. The agent picks it up the moment it’s free to look, without you having to wait for the current task to finish.

🛠️ Development

git clone https://github.com/theOnlyBoy/opencode-async-prompt-injector.git
cd opencode-async-prompt-injector

npm install        # or yarn / pnpm install
npm run build      # or yarn build / pnpm build

opencode           # run OpenCode from the project dir to test the plugin

The plugin is registered in ./opencode.json as "./dist/index.js", so after rebuilding you can test changes right away — just restart OpenCode.

⚠️ Model compatibility

This plugin injects your message into the conversation as cleanly as possible. Whether the agent actually acts on it depends on the model.

Most models handle it well. Some (especially smaller or more rigid ones) may treat the injected message as the end of the turn and stop, or respond more cautiously than usual.

There’s no plugin-side fix for that — it’s just how those models interpret turn boundaries. If something feels off, try a different model.

🤝 Contributing

Found a bug? Have an idea? PRs and issues are welcome.

Open an issue or submit a pull request — happy to take a look.

License

MIT