@toninho09/opencode-continue

An opencode plugin that adds a /continue command: it resumes an interrupted run (aborted with ESC) without adding a new message to the session and without the model ever noticing the interruption.

Useful when you interrupt the agent mid-task and want it to redo the aborted response as if the interruption had never happened — no orphaned partial text, no "[Tool execution was interrupted]" in the context, no "continuing from where I left off" meta-commentary, and no polluting the history with messages like "continue" or "go on".

Requirements

• A recent opencode version (plugin API >= 0.6.0).

Installation

Add the plugin to your opencode.json (project or global):

{
  "$schema": "https://opencode.ai/config.json",
  "plugin": ["@toninho09/opencode-continue"]
}

Or install from this repository (local checkout):

{
  "$schema": "https://opencode.ai/config.json",
  "plugin": ["/path/to/opencode-continue/src/index.ts"]
}

Restart opencode after saving the config. That's it — the plugin self-registers (the /continue command appears automatically, no need to create a command file).

Usage

/continue

• A run was started and interrupted with ESC → the plugin removes the interrupted tail from the session (the aborted assistant message, with its partial text and interrupted tools) and the agent regenerates the full response from the last completed step (or from the original prompt) — for the model and for the transcript, it's as if the interruption never existed. File edits already made are not undone.
• Session busy → warns and does nothing.
• Last run already completed → warns and does nothing (idempotent).
• Session with no user prompt → warns and does nothing.

How it works

1. A config hook registers the /continue command.
2. When /continue runs, a command.execute.before hook validates the session (idle, has a user prompt, last run unfinished), identifies the interrupted tail after the last real user prompt — assistant messages with an error (the aborted response) and the plugin's own old neutral messages — and deletes them from the session via DELETE /session/:id/message/:messageID (an endpoint that does NOT revert file changes). Completed steps of the turn (tool calls with results) are preserved: the model naturally resumes after them. If deletion fails (server without the endpoint), the removed IDs are hidden from the context via the experimental.chat.messages.transform hook.
3. Finally, the hook replaces the command's own parts, in place, with a single empty, ignored text part. The command creates that message (effectively invisible) and runs the agent loop. Since opencode filters ignored/empty parts out of the provider conversion (MessageV2.toModelMessagesEffect), the model receives the clean history — ending at the last completed step or at the original prompt — and regenerates the response from scratch, with no trace of the interruption.

The original user message is never rewritten nor re-sent — neither in the UI (its time.created is preserved, nothing jumps around) nor to the model (the text does not go into the context again).

The mechanism mirrors opencode's own loop exit condition (SessionPrompt.runLoop): a run without a valid finish = work pending; a clean finish = nothing to continue.

Known limitations

• No-op cases: busy session, already-completed run, or a session with no prompt cancel the command with a throw — the TUI shows the generic "failed to send command" toast in those cases (alongside the plugin's explanatory toast). On the normal path (an interrupted run) there is no error at all.
• Continuation model: the resume uses the session's current model/agent (not necessarily the ones from the aborted run).
• Tool re-execution: the response is regenerated from the last completed step — tools from the aborted step may run again (file edits remain; the model sees the current state of the files).
• Name collision: on opencode versions where /continue or /resume resolve to the session picker before custom commands, register one of the alternate names the plugin also intercepts: retry, go or again (e.g. .opencode/command/retry.md with any template). The hook intercepts those names automatically.
• The "real" solution would be a native /session/:id/continue endpoint exposing SessionPrompt.loop() — this plugin is the 100% plugin-API way until then.

Development

bun install
bun run typecheck

License

MIT © toninho09