multi: cross-model code review without a reviewer vote

[szarkans]

I built multi around a simple problem: a model that just wrote a change is also the model most tempted to say the change is done.

The plugin can call Claude subagents, Codex, OpenCode, Gemini, or headless Claude Code, depending on the command and which backends are connected. Code review uses Claude, Codex, and OpenCode, then reconciles their findings into one report. The reviewers do not vote. When they conflict, the workflow goes back to the code and checks the claim.

That is the main difference from role-based review boards that run several personas through one model. multi uses different model families when they are available and keeps the disagreement visible in the report. /multi:ask is even simpler: it shows every answer separately with no synthesis.

The stricter workflow is /multi:check-if-done. Reviewers that did not write the change inspect it, and the workflow will not report success without an executed verification command. /multi:code-review does the independent review pass. /multi:ask shows several answers side by side, and /multi:adhd applies the same approach to planning.

Several models can still be wrong together. I do not think reviewer count is a correctness proof. The useful signal is where independent reviewers split and which claim survives an actual command.

Repo and install instructions: https://github.com/szarkans/multi

I would like feedback from people who have tried cross-model review: do separate verdicts expose useful blind spots, or does the disagreement mostly add noise and cost?