OpenCode Design Lab

OpenCode Design Lab is an OpenCode plugin for running the same planning or review task across multiple configured models in one session. A single primary agent coordinates model-specific subagents, writes artifacts to disk, and summarizes results back into chat.

Overview

• Single coordinator agent: design_lab coordinates every workflow and can run directly or through delegation.
• Model-specific subagents: design_lab_model_* agents are generated from config.
• Current model inheritance: design_lab does not set a fixed model, so OpenCode can use the active UI/default model for coordination.
• Default-agent activation: the bundled design-lab skill can be loaded from the normal build agent when your prompt mentions multi-model design or review.
• Arbitrary variants: model variants are passed through as configured, including values like xhigh.
• Review-only subagents: model subagents write review files and never modify source code during review.
• File-first outputs: full model responses, plans, reviews, manifests, and summaries are saved under .design-lab/.

Commands

| Command | Description | | -------------------------- | ----------------------------------------------- | | /design-lab:init | Initialize .opencode/design-lab.json | | /design-lab:ask <prompt> | Run the unified multi-model Design Lab workflow | | /design-lab:journal | Document recent decisions in .journal/ | | /design-lab:repowiki | Generate comprehensive repository documentation |

You can also switch to the design_lab agent and ask directly without using /design-lab:ask. Direct agent usage follows the same workflow rules.

You do not have to switch agents for common prompts. The plugin registers a bundled design-lab skill and nudges the current agent to load it when prompts mention multi-model design, multi-model review, blind review, or current-code review. The nudge is applied to the latest matching user turn, so it also works in an existing chat. After loading the skill, the current agent should call the native Task tool directly for each selected design_lab_model_* agent; this keeps every model worker visible as a top-level task card in the OpenCode UI. If the current agent is already design_lab, it runs the coordinator workflow directly instead of delegating back to itself. If native Task fanout is unavailable, the current agent can fall back to the design_lab_run tool, which starts a child design_lab session and returns the summary. Restart OpenCode after rebuilding or installing the plugin so the bundled skill and tool are available.

Configuration

Create .opencode/design-lab.json:

{
  "$schema": "https://raw.githubusercontent.com/HuakunShen/opencode-design-lab/main/schemas/design-lab-config.schema.json",
  "models": [
    { "model": "openai/gpt-5.2-codex", "variant": "xhigh" },
    { "model": "opencode/kimi-k2.5-free", "variant": "max" },
    { "model": "local/model-without-variant", "variant": null }
  ],
  "default_variant": "max",
  "base_output_dir": ".design-lab"
}

Options

| Option | Type | Default | Description | | ----------------- | ---------------------- | ------------- | ----------------------------------------------- | | models | (string \| object)[] | Required | Models used for all Design Lab workflows, min 2 | | default_variant | string \| null | "max" | Variant applied to plain string model entries | | base_output_dir | string | .design-lab | Output directory for run artifacts |

Model entries can be plain strings or objects:

{
  "models": [
    "openai/gpt-5.2-codex",
    { "model": "openai/gpt-5.2-codex", "variant": "xhigh" },
    { "model": "local/model-without-variant", "variant": null }
  ]
}

• Plain strings use default_variant.
• variant can be any non-empty string supported by your provider/OpenCode setup.
• variant: null means invoke that model without a variant.

Workflows

Generate Plans

/design-lab:ask generate plans for adding OAuth login

Each selected model writes its own plan under plans/, and manifest.json records the model-to-file mapping.

Revise Plans

/design-lab:ask revise the latest plans to support multi-tenant auth

The primary agent reads manifest.json and asks each model to revise only its own plan file.

Blind Review Plans

/design-lab:ask anonymously review the latest plans

The primary agent copies plan files to anonymous names under blinds/plans-blind/, writes blinds/mapping.json, and sends only anonymous files to reviewers.

Review Current Code Changes

/design-lab:ask review current changes with gpt and kimi only

From the normal build agent, this also works as a plain prompt:

review current changes with multiple models

The primary agent collects git status, git diff, changed files, and your review focus into a context packet. Selected model subagents read the packet and write review files under reviews/. They do not edit source code.

Reviewer selection can use full model names, short names, file stems, agent names, or ordinals such as 1, 3, 5. Ambiguous selectors require clarification.

Output Structure

Plan and blind review runs use:

.design-lab/YYYY-MM-DD-topic/
├── prompt.md
├── manifest.json
├── plans/
│   ├── gpt-5-2-codex.md
│   └── kimi-k2-5-free.md
├── blinds/
│   ├── mapping.json
│   └── plans-blind/
│       ├── plan-a.md
│       └── plan-b.md
├── reviews/
│   ├── review-gpt-5-2-codex.md
│   └── review-kimi-k2-5-free.md
└── summary.md

Current-code review runs use:

.design-lab/YYYY-MM-DD-code-review/
├── context/
│   ├── review-request.md
│   ├── git-status.txt
│   ├── diff.patch
│   └── changed-files.txt
├── reviews/
│   ├── code-review-gpt-5-2-codex.md
│   └── code-review-kimi-k2-5-free.md
├── manifest.json
└── summary.md

Development

bun run build
bun run dev
bun run test
bun run typecheck
bun run export-schemas