opencode-ghost

Next-prompt suggestions for the opencode TUI, in the style of Claude Code: after each turn a small model reads the recent conversation and writes the message you would most likely send next, shown as dimmed ghost text you accept with Tab (or →).

Why "ghost": it writes the line you were about to type, then waits for you to take it.

What it does

• Suggests your next message. On session.idle it debounces, builds a short transcript of the recent turns, and asks a model (your small_model by default) for one line that sounds like you.
• Ghost text, Tab to accept. The suggestion renders under the prompt; Tab or → drops it into the input so you can edit and send.
• Slash-command argument ghosts while typing. After /name , the accepted argument options (from argHints) are listed as [a | b | c] inside the prompt input box, right after the caret. Once you start typing an argument, the list gives way to a dimmed inline completion of the matching option (e.g. /keepwarm alw ghosts ays). Running out of options or naming a command with no configured hints shows no ghost — command and skill descriptions are never suggested as if they were arguments. Tab completes the next argument. While the command name is still being typed, opencode's own slash menu stays in charge. Purely local — no model calls.
• History ghost while typing. If the line you are typing is a prefix of a message you already sent in the same session, the rest of it is ghosted in the box as well; Tab pulls the full line back. Also purely local.
• Grey, dimmed ghost styling. Every ghost is grey and dimmed so it cannot be mistaken for typed text. A next-message suggestion that is wider than the prompt wraps onto extra rows and grows the prompt box (up to five rows) so it stays fully readable; other ghosts are clipped to the box and never spill past it.
• /suggest toggles it. State is stored in the plugin KV.

Requirements

• opencode (TUI plugins; tested on 1.18.x).
• A cheap/fast model for small_model — or set model explicitly.

Install

From this repo (recommended)

git clone https://github.com/ozandogrultan/opencode-ghost.git
cd opencode-ghost
./install.sh

Then restart opencode — plugins are loaded at startup only. install.sh copies the plugin into <config>/plugins/opencode-ghost/ and registers it in <config>/tui.json (backing the file up first).

Manually

Add the checkout to your tui.json (global at ~/.config/opencode/tui.json, or a project's .opencode/). Plugins are loaded as source:

{
  "$schema": "https://opencode.ai/tui.json",
  "plugin": ["/absolute/path/to/opencode-ghost/src/tui.tsx"],
}

From npm

opencode plugin opencode-ghost       # current project
opencode plugin -g opencode-ghost    # global

Restart opencode after changing tui.json.

Options

Pass an options object as the second element of the plugin tuple:

{
  "plugin": [
    ["opencode-ghost", {
      "enabled": true,
      "model": "anthropic/claude-haiku-4-5",
      "acceptKeys": ["tab", "right"],
      "maxChars": 120,
      "idleDelayMs": 500,
      "recentMessages": 10
    }]
  ]
}

| Option | Type | Default | Notes | | ---------------- | ---------- | --------------------- | --------------------------------------------------------------------- | | enabled | boolean | true | Initial state; /suggest toggles it at runtime (persisted). | | model | string | session small_model | provider/model. Falls back to small_model, then the session model. | | endpoint | string | unset | Custom OpenAI-compatible or gateway endpoint (baseURL or /chat/completions). Enables stateless direct generation without creating OpenCode sessions. | | apiKey | string | unset | API key for endpoint, supporting {file:...}, {env:...}, or raw token. | | acceptKeys | string[] | ["tab", "right"] | Key names as reported by the terminal (tab, right, ...). | | maxChars | number | 120 | Maximum suggestion length. | | idleDelayMs | number | 500 | Debounce after a turn finishes before generating. | | recentMessages | number | 10 | How many recent messages feed the suggestion prompt. | | system | string | built-in | Override the system prompt sent to the suggestion model. | | backOnEmptyLeft | boolean | true | Return to the home screen with Left when the prompt is empty. | | internalSessionMarkerDir | string | unset | Optional extra crash-recovery tracking. A configuration-free sweep already reaps leftover hidden sessions, so this is only an additional safety net. | | argHints | Record<string, string[]> | {} | Argument option lists per slash command, shown as [a \| b \| c] after a complete /name and completed with Tab, e.g. "/keepwarm": ["6h", "always", "off", "status"]. Commands without an entry suggest no arguments. |

Commands

• /suggest — toggle suggestions on/off (state is stored in the plugin KV).

When enabling, a suggestion is generated immediately for the current session.

How it works

• Listens for session.idle, debounces, and builds a short transcript from the session's recent messages.
• Asks a model (your small_model by default) for one short next user message. When provider credentials or an endpoint are available, it generates suggestions statelessly via direct HTTP fetch (zero server sessions created, zero cmux/session hooks emitted). Otherwise, falls back to a throwaway hidden session with tools disabled that is deleted immediately afterwards. Every hidden-session call is scoped to the current session's project directory, hidden sessions carry a metadata tag that survives the server's automatic title generation, and a periodic sweep (plus shutdown cleanup) reaps any leftovers — including sessions orphaned by a crash — with no configuration required.
• Renders the result through the session_prompt host slot, ghosted inside the prompt box at the caret, and accepts it into the input via the prompt ref.

Caveats

• Owns the session_prompt slot. opencode only exposes the prompt ref through that slot, so this plugin replaces the default prompt component. It forwards on_submit, ref, right, visible, and disabled, and cannot be combined with another plugin that also renders session_prompt.
• Ghost placement. All ghosts (next-message suggestion, typing completions) render inside the prompt box right after the caret; opencode exposes no inline-completion API, so the plugin overlays the prompt's editor renderable and never touches the prompt's hint row.
• Cost. One small-model call per turn while enabled. Point model at a cheap or free model.
• acceptKeys. tab is also opencode's agent-cycle key; this plugin only captures it while a suggestion is visible and the input is empty.
• Suggestions fail silently (no ghost text) if the model is unavailable, the provider is out of balance, or the reply is NONE.

Development

bun install
bun run typecheck
bun test

No build step: like other opencode TUI plugins, the package ships TSX source and opencode's runtime transpiles it.

Contributing

See CONTRIBUTING.md for the check suite and commit conventions, AGENTS.md for the design rules, and CHANGELOG.md for what changed.

License

MIT