Subagent Router

Dynamic sub-agent model router for OpenCode — cost-optimized model selection with escalation and quality guardrails.

Provider Support

Provider	Cost tracking	Default rules	Notes
DeepSeek	Built-in	Built-in	Full support
OpenAI	Config-based	Manual	Add via providers config
Anthropic	Config-based	Manual	Add via providers config
Google	Config-based	Manual	Add via providers config
Others	Config-based	Manual	Add via providers config

Roadmap: #1 Add built-in cost data + default rules for OpenAI, Anthropic, Google. PRs welcome — see CONTRIBUTING.md.

Features

• Dynamic routing: chat.message hook intercepts sub-agent invocations and overrides the model based on task type
• Cost-optimized: Routes cheap tasks (explore, title, summary, compaction) to cheap models and complex tasks (general, build, plan) to powerful models
• Provider-agnostic: Works with any provider supporting multiple models via single API key. Built-in rules for DeepSeek; config-based rules for everything else.
• Smart escalation: If a cheap model receives a complex task, auto-promotes to the powerful model with cooldown
• Quality guardrails: Tracks when cheap models get re-tasked (signal of inadequate output) and auto-reverts to flagship
• Session summaries: Reports cost saved at session end via toast
• TUI sidebar: Live cost savings visible in the right-side panel
• /route-stats command: Show routing stats on demand

Install

npm install opencode-subagent-router

Configure

In .opencode/opencode.json:

{
  "plugin": ["opencode-subagent-router"]
}

In .opencode/tui.json:

{
  "plugin": ["opencode-subagent-router"]
}

Configuration

Add a model-router section to opencode.json:

{
  "model-router": {
    "rules": [
      {
        "agents": ["explore", "title", "summary", "compaction"],
        "model": "deepseek-v4-flash",
        "provider": "deepseek",
        "label": "Cheap with reasoning"
      },
      {
        "agents": ["general", "build", "plan"],
        "model": "deepseek-v4-pro",
        "provider": "deepseek",
        "label": "Flagship for code"
      }
    ],
    "defaultModel": "deepseek-v4-flash",
    "escalation": {
      "enabled": true,
      "threshold": 2,
      "cooldownMs": 300000
    },
    "guardrails": {
      "enabled": true,
      "maxRetries": 2,
      "retryWindowMs": 600000
    }
  }
}

Adding a new provider

Add model costs via the providers field. The router uses this data to calculate per-session savings:

{
  "model-router": {
    "rules": [
      {
        "agents": ["explore", "title", "summary", "compaction"],
        "model": "gpt-4o-mini",
        "provider": "openai",
        "label": "Cheap for simple tasks"
      },
      {
        "agents": ["general", "build", "plan"],
        "model": "gpt-4o",
        "provider": "openai",
        "label": "Powerful for code"
      }
    ],
    "defaultModel": "gpt-4o-mini",
    "providers": {
      "openai": {
        "gpt-4o-mini": {
          "cost": { "input": 0.15, "output": 0.60 },
          "reasoning": false,
          "context": 128000
        },
        "gpt-4o": {
          "cost": { "input": 2.50, "output": 10.00 },
          "reasoning": true,
          "context": 128000
        }
      }
    }
  }
}

Architecture

chat.message hook
  └─ scan output.parts for SubtaskPart
       ├─ No subtask → pass through
       ├─ Subtask found → lookup agent in rules
       │    ├─ Match found → override part.model
       │    └─ No match → use defaultModel
       └─ Track decision + cost saved

License

MIT