Skip to content

opencode.md

Latest commit

 

History

History
172 lines (139 loc) · 5.48 KB

opencode.md

File metadata and controls

172 lines (139 loc) · 5.48 KB
Raw

OpenCode

Connect OpenCode to tuxevil-rotator to use Google Antigravity models as your OpenCode provider.

Configuration

Create or edit ~/.config/opencode/opencode.json (global) or opencode.json in your project root:

{
  "$schema": "https://opencode.ai/config.json",
  "provider": {
    "antigravity": {
      "npm": "@ai-sdk/openai-compatible",
      "name": "Antigravity (local)",
      "options": {
        "baseURL": "http://localhost:51200/v1"
      },
      "models": {
        "gemini-3.8-flash-high": {
          "name": "Gemini 3.8 Flash High",
          "limit": { "context": 1000000, "output": 65536 }
        },
        "gemini-3.8-flash-medium": {
          "name": "Gemini 3.8 Flash Medium",
          "limit": { "context": 1000000, "output": 65536 }
        },
        "gemini-3.8-flash-low": {
          "name": "Gemini 3.8 Flash Low",
          "limit": { "context": 1000000, "output": 65536 }
        },
        "gemini-3.7-flash-tiered": {
          "name": "Gemini 3.7 Flash Tiered (Thinking)",
          "limit": { "context": 1000000, "output": 65536 },
          "options": {
            "reasoningEffort": "high"
          }
        },
        "gemini-3.6-flash-high": {
          "name": "Gemini 3.6 Flash High",
          "limit": { "context": 1000000, "output": 65536 }
        },
        "gemini-3.6-flash-medium": {
          "name": "Gemini 3.6 Flash Medium",
          "limit": { "context": 1000000, "output": 65536 }
        },
        "gemini-3.1-pro-low": {
          "name": "Gemini 3.1 Pro",
          "limit": { "context": 1000000, "output": 65536 }
        },
        "claude-sonnet-4-6": {
          "name": "Claude Sonnet 4.6",
          "limit": { "context": 1000000, "output": 64000 }
        },
        "claude-opus-4-6-thinking": {
          "name": "Claude Opus 4.6 Thinking",
          "limit": { "context": 1000000, "output": 64000 }
        }
      }
    }
  },
  "model": "antigravity/gemini-3.8-flash-high"
}

Setting the API Key

Option A: Via /connect command inside OpenCode (recommended)

/connect

Select "Other", enter provider ID antigravity, then enter your key (no-key if no Virtual Keys are configured, or your rk-... Virtual Key).

Option B: Environment variable in config

{
  "provider": {
    "antigravity": {
      "options": {
        "baseURL": "http://localhost:51200/v1",
        "apiKey": "{env:TUXEVIL_ROTATOR_KEY}"
      }
    }
  }
}

Then set TUXEVIL_ROTATOR_KEY=no-key (or your rk-... key) in your shell.

Selecting a Model

Set as default in config:

{ "model": "antigravity/gemini-3.8-flash-high" }

Or via CLI flag:

opencode --model antigravity/gemini-3.8-flash-high

Or use /models inside OpenCode to pick interactively.

Thinking & Reasoning Levels

Antigravity models handle reasoning and internal thinking in three ways:

1. Named Reasoning Levels (Gemini 3.8)

Gemini 3.8 exposes separate native model IDs for low, medium, and high reasoning. Select the desired level through the model ID; the rotator leaves thinking adaptive within that upstream level.

Model ID Context Window Reasoning Level
gemini-3.8-flash-high 1,000,000 High
gemini-3.8-flash-medium 1,000,000 Medium
gemini-3.8-flash-low 1,000,000 Low

2. Adaptive Thinking (gemini-3.7-flash-tiered)

gemini-3.7-flash-tiered uses Google's dynamic thinking level. By default, it operates adaptively (the model decides how many tokens to think). You can control the thinking depth using standard reasoning_effort:

Level Upstream thinkingLevel Use Case
low LOW Quick responses, straightforward edits, simple questions.
medium MEDIUM Balanced reasoning for standard coding and multi-file refactors.
high HIGH Deep architectural design, complex bug diagnosis, deep multi-step planning.
(default) Adaptive Model dynamically adjusts thinking tokens based on prompt complexity.

3. Fixed Token Budgets (Gemini 3.6 / 3.1 & Claude)

For earlier Gemini models and Claude, the thinking budget is predetermined by model ID:

Model ID Context Window Thinking Budget
gemini-3.7-flash-tiered 1,000,000 Dynamic / Tiered (low / medium / high)
gemini-3.6-flash-high 1,000,000 10,000 tokens (fixed)
gemini-3.6-flash-medium 1,000,000 4,000 tokens (fixed)
gemini-3.6-flash-low 1,000,000 1,000 tokens (fixed)
gemini-3.1-pro-low 1,000,000 1,001 tokens (fixed)
claude-sonnet-4-6 1,000,000 32,768 tokens (fixed)
claude-opus-4-6-thinking 1,000,000 32,768 tokens (fixed)

Advanced Settings & Prompt Compression

You can pass custom headers in opencode.json to enable features like token compression:

{
  "provider": {
    "antigravity": {
      "options": {
        "baseURL": "http://localhost:51200/v1",
        "headers": {
          "X-Rotator-Compression": "rtk+lite"
        }
      }
    }
  }
}

Notes

  • The npm field must be "@ai-sdk/openai-compatible" for the OpenAI Chat Completions endpoint
  • Model IDs in the config must match what GET /v1/models returns from the rotator
  • Add as many models as you want to the models map — only listed ones appear in the picker
  • tuxevil-rotator automatically sanitizes complex JSON schemas sent by OpenCode tools (read, edit, shell, task, grep, glob), stripping unsupported keywords so Antigravity executes them seamlessly.