opencode-multireview-plugin

Local-first, npm-ready OpenCode plugin that bundles the multireview agent set and two Plannotator-backed review skills:

• multireview-explainer
• multireview-diff

The package injects the four agents directly through the OpenCode config hook. Bundled skills are installed by the package postinstall script, which copies them into ${XDG_CONFIG_HOME:-$HOME/.config}/opencode/skills.

What multireview does

@multireview is an adversarial code review coordinator. Invoking it spawns three specialist reviewers in parallel, each scoped to a single concern and blind to the others' findings:

• multireview_correctness — logic soundness, edge cases, error handling, concurrency, performance, and OWASP-style security issues.
• multireview_codestyle — naming, function design, comments, DRY violations, and file/code organisation.
• multireview_testing — unit and integration test coverage for the changed code, plus test-quality anti-patterns (asserting on mocks, hardcoded sleeps, over-testing).

Each specialist reviews only the given change scope (uncommitted changes, a branch diff, or a PR) and returns structured findings with a severity, exact code location, and justification — no fixes, no fluff, no praise.

The coordinator then acts as final arbiter: it discards hallucinated, out-of-scope, or overly pedantic findings, merges duplicates across reviewers, and writes everything to REVIEW_FINDINGS.md in the repo root (git-excluded), split into ## Valid Findings and ## Ignored Findings. Every ignored finding carries a one-line Wontfix: justification so later runs don't re-flag it.

From there, two optional skills turn findings into an interactive review pass using Plannotator:

• multireview-explainer renders the findings and a narrative PR summary as a static HTML page. Reviewers select text and leave native comments (including wontfix) directly on findings.
• multireview-diff instead preloads findings as line comments in Plannotator's native git-diff UI, so reviewers work against the real file tree and diff rather than a rendered page.

Both skills triage the human feedback back into REVIEW_FINDINGS.md — confirming, dismissing, or re-scoping findings — and hand the confirmed valid findings off to a fixer agent to implement.

Install

cd ~/.config/opencode
npm install opencode-multireview-plugin

Then configure OpenCode with Plannotator first:

{
  "plugin": [
    ["@plannotator/opencode@latest", { "workflow": "all-agents" }],
    "opencode-multireview-plugin"
  ]
}

Restart OpenCode after changing plugin configuration.

Install from a local checkout

For development or before a version is published, install directly from a local path:

cd ~/.config/opencode
npm install /path/to/opencode-multireview-plugin

Or reference the built plugin file directly:

{
  "plugin": [
    ["@plannotator/opencode@latest", { "workflow": "all-agents" }],
    "file:///path/to/opencode-multireview-plugin/dist/index.js"
  ]
}

Bundled skills

The bundled skills are copied during npm install, respecting $XDG_CONFIG_HOME when set:

• ~/.config/opencode/skills/multireview-explainer/SKILL.md
• ~/.config/opencode/skills/multireview-diff/SKILL.md

Plannotator requirement

This plugin depends on Plannotator but does not vendor it. By default it fails during the config hook if @plannotator/opencode is not present in cfg.plugin. The bundled skills also expect the plannotator CLI to be available on PATH.

For development only, the runtime check can be disabled:

{
  "plannotator": {
    "requirePlugin": false
  }
}

Config overrides

Defaults:

{
  "models": {
    "coordinator": "github-copilot/claude-opus-4.8",
    "codestyle": "github-copilot/claude-sonnet-5",
    "correctness": "github-copilot/gpt-5.4",
    "testing": "github-copilot/gemini-3.5-flash"
  },
  "plannotator": {
    "requirePlugin": true
  }
}

Create ~/.config/opencode/multireview-plugin.json to override them locally:

{
  "models": {
    "correctness": "github-copilot/gpt-5.2"
  }
}

Plugin tuple options override the local file:

{
  "plugin": [
    ["@plannotator/opencode@latest", { "workflow": "all-agents" }],
    [
      "opencode-multireview-plugin",
      {
        "configPath": "~/.config/opencode/multireview-plugin.json",
        "models": { "testing": "github-copilot/gemini-3.1-pro-preview" }
      }
    ]
  ]
}

CLI helpers

The skills call package bins rather than absolute local paths:

• opencode-multireview-parse-findings
• opencode-multireview-build-explainer
• opencode-multireview-build-code-annotations

Development

npm install
npm run typecheck
npm test
npm pack --dry-run

CI

GitHub Actions runs .github/workflows/ci.yml on pushes to main and on pull requests:

npm install
npm test

Publishing

Releases are published to npm via a GitHub Actions workflow (.github/workflows/publish.yml) using npm's trusted publishing (OIDC), triggered when a vX.Y.Z tag is pushed. The same workflow also creates the GitHub Release automatically. No long-lived npm token is stored in the repository.

Cut releases with release-it, which bumps package.json, updates CHANGELOG.md, and creates and pushes the vX.Y.Z tag:

npm run release

To preview the release without changing anything:

npm run release:dry-run

Once the tag is pushed, .github/workflows/publish.yml creates the GitHub Release, builds, tests, and publishes the package to npm.