Always Questions

An OpenCode plugin that ensures the AI always asks for your next steps after completing any task by appending a fixed directive suffix to the system prompt and every user message.

What it does

After the AI finishes any task, it is required to call the built-in questions tool to present context-aware options for what to do next, rather than simply ending the response with a passive "let me know if you need anything."

Mechanism

The plugin hooks into two experimental transform events:

Hook	Effect
experimental.chat.system.transform	Appends a mandatory behavior directive to the system prompt
experimental.chat.messages.transform	Appends a per-message reminder to the last user message before it reaches the LLM

Both hooks are invisible in the TUI — the user does not see the injected text, but the model always receives it.

Installation

Option 1: npm (global or project)

Add to your opencode.json:

{
  "$schema": "https://opencode.ai/config.json",
  "plugin": ["opencode-plugin-always-questions"]
}

OpenCode will install the package automatically via Bun at startup.

Option 2: Local plugin file

Copy index.js (or index.ts) to your plugin directory:

• Project-level: .opencode/plugins/always-questions.js
• Global: ~/.config/opencode/plugins/always-questions.js

Files in these directories are loaded automatically — no config entry needed.

Configuration

The plugin currently has no configuration options. The injected suffix text is defined as constants in index.ts:

• SYSTEM_SUFFIX — directive added to the system prompt once per session
• MESSAGE_SUFFIX — reminder added to each user message

To customize, clone the repo and edit these constants before building.

Building from source

bun install
bun build index.ts --outfile index.js --target bun --format esm

Requirements

• OpenCode with plugin support
• Bun runtime (used internally by OpenCode)

Notes

• This plugin uses experimental OpenCode APIs (experimental.chat.system.transform and experimental.chat.messages.transform) which may change in future versions.
• The injected text counts toward the model's context window. The total overhead is approximately 200 tokens per session (system suffix, added once) plus ~30 tokens per user message.

License

MIT