opencode-max-mode

An OpenCode V2 plugin that adds a best-of-N tool. It runs several candidates for a hard question and a judge picks the winner.

OpenCode

This plugin runs on OpenCode. New accounts through my referral link get $5 in usage credits, and I get $5 too:

https://opencode.ai/go?ref=N9H3ZEP22A

Install

mkdir -p ~/.config/opencode/plugins
curl -fsSL \
  https://raw.githubusercontent.com/jadmadi/opencode-max-mode/main/max-mode.ts \
  -o ~/.config/opencode/plugins/max-mode.ts

For one project, put it in .opencode/plugins/. Tested against OpenCode v2.0.3.

To pin a release, replace main in the URL with a tag such as v0.1.0.

Use

Command	Effect
/max <n>	Set the default candidate count for the session
/max	Show the current default
/max off	Clear the default

The plugin registers a best_of_n tool. Ask the model to use it, or call it with { prompt, candidates? }. The count defaults to the session setting, or 3. The cap is 8.

Why a tool

No plugin hook can replace the main turn's response. The context hook only edits inputs, and only the compaction and title hooks may supply a result. So best-of-N is a tool the model can call, not an interceptor for the next turn. The tool runs n parallel ctx.generate.text calls, then one judge call, and returns the winning text.

Cost multiplies with the candidate count: n calls plus one judge.

Model

Candidates and the judge use the session's model. Some providers do not support transient generation: OpenCode Go returned Request is missing x-opencode-session in testing. Set MAX_MODE_MODEL to a working model, for example:

MAX_MODE_MODEL=deepseek/deepseek-flash

Tests

bun test

Attribution

Inspired by MiMoCode's max mode. See NOTICE.

License

MIT