Followups

An OpenCode v2 TUI plugin. After every assistant turn it asks the model for 3-5 prompts you might send next. It lists them in a toast and puts the top one in your prompt, ready to edit or send:

Follow-ups (1.1s)
1. Add a subtract function to calc.py.
2. Write unit tests for the add function.
3. Add error handling for non-numeric inputs.
...
/followup [n] to use one

/followup opens a picker, and /followup 2 puts suggestion 2 straight into the prompt. Nothing is ever sent for you. The suggestion only goes in when the prompt is focused and empty, so anything you've started typing stays as it is. If it can't go in, /followup opens an editable dialog instead: Enter sends, Esc cancels.

Requirements

• OpenCode v2 (tested with 2.0.27), using the terminal TUI.
• The plugin does nothing in the desktop or web app. Those don't load TUI plugins.

Install

Local copy

mkdir -p ~/.config/opencode/plugins/opencode-followups
cp -R tui.ts src package.json ~/.config/opencode/plugins/opencode-followups/

The TUI picks up every directory under ~/.config/opencode/plugins/, or under a project's .opencode/plugins/, that has a tui.ts. Note that it's plugins/, not plugin/: v2's TUI only scans plugins/.

Configuration

The defaults work without any config. To set options, register the plugin in the TUI config, ~/.config/opencode/cli.json, rather than relying on auto-discovery, because a directory found under plugins/ gets no options. Keep the copy outside plugins/ so it doesn't load twice:

{
  "plugins": [
    { "package": "./followups", "options": { "model": "local/qwen3#fast", "messages": 4, "timeoutMs": 10000 } }
  ]
}

The relative path resolves against the directory that holds cli.json. Environment variables override options for a single run.

Option	Env	Default	Meaning
enabled	OPENCODE_FOLLOWUPS=0/1	true	Master switch
model	OPENCODE_FOLLOWUPS_MODEL	prompt's current model	provider/model or provider/model#variant
variant		"fast"	Variant to use when model names none. Applied only if the model has a variant with this id. Point it at a variant that turns thinking off. "" disables it.
messages		6	How many recent user/assistant messages are sent
maxCharsPerMessage		1500	Each message is clipped to this many characters
timeoutMs	OPENCODE_FOLLOWUPS_TIMEOUT_MS	20000	The call is aborted after this
count		5	Max suggestions (1-5)
maxSuggestionChars		160	Each suggestion is clipped to this many characters
toastMs		12000	How long the toast stays up
insert		true	Put the top suggestion in the prompt when it is focused and empty

OPENCODE_FOLLOWUPS_DEBUG=1 turns on a warning toast when a call fails. Without it, failures are silent.

Turning thinking off

The extra call goes through OpenCode's own model layer (client.generate.text), so it reuses your provider, URL and key. OpenCode doesn't expose a per-call "no thinking" switch, so add a model variant that does it. With llama.cpp / Qwen3:

"qwen3": {
  "body": { "chat_template_kwargs": { "enable_thinking": true } },
  "variants": [{ "id": "fast", "body": { "chat_template_kwargs": { "enable_thinking": false } } }]
}

The default variant: "fast" then picks it up automatically.

How it works

1. It listens for session.execution.succeeded, which is v2's "turn finished" event. Failed and interrupted turns are skipped.
2. It skips subagent sessions (any session with a parentID), and it never runs two calls at once for the same session.
3. It reads the session context and keeps only user text and assistant text; tool calls and reasoning are dropped. It takes the last messages turns and sends one stateless generate.text request. That request creates no session, so it can't trigger itself.
4. It parses the reply the way Open WebUI does: it takes the text from the first { to the last }, then reads .follow_ups. Any parse problem produces no toast.
5. If you start a new turn while suggestions are still generating, the result is dropped.

Limits

• Prompt insertion uses an internal detail. OpenCode v2 has no prompt-append API: the tui.prompt.append event exists, but the server never publishes it. So the plugin calls insertText on the focused OpenTUI textarea (renderer.currentFocusedRenderable), and only when the keymap is in base mode with no dialog open. A future OpenCode release could change this. If it does, insertion fails quietly and the toast plus the /followup dialog still work.
• No hard output cap. generate.text takes only a prompt and a model, with no max_tokens. Output length is limited by the instructions in the prompt, by clipping each suggestion, and by the timeout.
• A timeout doesn't stop the server. The plugin stops waiting, but OpenCode's server may still finish the request on the GPU.
• It shares your model. On a single-slot local server (llama.cpp -np 1), the extra request runs between your turns. If your server has no host prompt cache (--cache-ram), the next turn may have to re-read the whole context.

Development

bun install
bun test          # parsing, loop guard, config, prompt building
bunx tsc --noEmit # typecheck against @opencode/plugin 2.0.27

Similar plugins

• opencode-ghost and opencode-next-prompt show one predicted next prompt as inline ghost text, accepted with Tab or the right arrow.
• opencode-quick-reply (v1 only) suggests a reply when the agent is waiting on you.

This plugin instead offers a numbered set of 3-5 options after every turn, uses your own configured model with thinking off, and has a picker.

Credits

The feature, and the {"follow_ups": [...]} reply format with its first-{-to-last-} parsing, come from Open WebUI's follow-up generation task. The prompt text here is original; no Open WebUI code or template text is included.

License

MIT