opencode-advisor

An opencode plugin that implements the advisor strategy: a smaller model working on a task gets access to an ask_advisor tool that routes questions to a larger, more capable model.

The plugin automatically gathers the full session transcript (including tool inputs and outputs) so the advisor model sees complete context without the working model having to describe its own situation.

How it works

1. The working model hits a hard problem — architectural decision, tricky bug, non-obvious tradeoff
2. It calls ask_advisor with a concise question
3. The plugin fetches the full session transcript and sends it along with the question to the advisor model
4. The advisor responds with targeted, actionable guidance
5. The working model synthesizes the advice and continues

The advisor runs as a separate, ephemeral session — it can't take actions, only give advice.

Installation

Option A: local plugin (recommended for now)

Copy src/index.ts into your project as .opencode/plugins/advisor.ts, then install its dependency:

cd .opencode
npm install @opencode-ai/plugin

No changes to opencode.json needed — opencode auto-loads files in .opencode/plugins/.

Option B: npm package

npm install opencode-advisor

Then add to your opencode.json:

{
  "$schema": "https://opencode.ai/config.json",
  "plugins": ["opencode-advisor"]
}

Configuration

Create .opencode/advisor.json in your project:

{
  "model": "anthropic/claude-opus-4-7",
  "maxToolOutput": 10000
}

| Field | Default | Description | |---|---|---| | model | anthropic/claude-opus-4-7 | Advisor model in provider/model format. Must be a model available in your opencode setup. | | maxToolOutput | (none) | Truncate tool call outputs in the transcript at this many characters. Useful to prevent token overload when sessions contain large file reads. Omit to include full outputs. |

You can also set ADVISOR_MODEL=provider/model as an environment variable — this takes precedence over the config file.

Choosing a model

The advisor model must be a provider/model pair configured in opencode (e.g. via your opencode.json or environment variables). Examples:

• anthropic/claude-opus-4-7
• anthropic/claude-opus-4-5
• openai/o3
• google/gemini-2-5-pro

The whole point is to use a more capable model than your working model — pick accordingly.

When should the working model use it?

The tool description instructs the model to use it sparingly:

Consult a more powerful model for guidance when you are genuinely stuck, need architectural advice, or face a critical decision with non-obvious tradeoffs. Use sparingly — only for hard problems where expert input materially changes the outcome.

Good uses: stuck on a hard bug, uncertain between two architectures, need a second opinion on a risky refactor.

Bad uses: simple questions the model can answer itself, routine tasks, fishing for validation.

Viewing logs

The plugin logs to opencode's structured log. On Windows:

%APPDATA%\opencode\log\

On macOS/Linux:

~/.local/share/opencode/log/

Or run opencode with --print-logs to stream logs to stdout.

Filter for plugin activity:

grep "opencode-advisor" opencode.log

Development

npm install       # installs @opencode-ai/plugin in root node_modules/
npm run build     # compile to dist/
npm run dev       # watch mode

src/index.ts is the single source of truth. .opencode/plugins/advisor.ts is a thin re-export that lets opencode load the plugin when working in this repo itself — it relies on npm install having been run at the root so that @opencode-ai/plugin is resolvable from src/.

License

MIT