Model Configuration

The agent connects to the OpenCode Go gateway, which routes requests to available models. The gateway client lives in internal/llm/gateway/client.go and the model catalog is embedded from internal/llm/gateway/models.json.

Architecture

All communication uses SSE (Server-Sent Events) streaming for real-time output.

Configuration

Environment variables control model selection:

# Required
AI_GATEWAY_API_KEY=...

# Optional
AI_GATEWAY_BASE_URL=https://opencode.ai/zen/go/v1
AI_GATEWAY_MODEL=opencode-go/deepseek-v4-pro

Model resolution order: explicit /model slash-command choice → AI_GATEWAY_MODEL env → models.json default → compile-time defaultModel constant.

The gateway client strips the opencode-go/ prefix before sending the bare model id to the API. For example, opencode-go/deepseek-v4-pro becomes deepseek-v4-pro on the wire.

Model Catalog

The model catalog (models.json) is embedded into the binary at compile time via //go:embed. Available models:

Default: opencode-go/deepseek-v4-pro

Runtime Model Switching

Type /model in the TUI composer to open the interactive model picker:

• /model with no argument opens the full picker
• /model <provider>/<model> switches directly
• /model <provider> lists that provider's models

Switching is live: no restart required. The status bar updates to show the new model name.

Gateway Retry Logic

The gateway HTTP client retries on transient failures with exponential backoff and jitter:

Only transient errors trigger retries: HTTP 429/500/502/503/504, connection resets, refused connections, EOF, broken pipe, DNS failures, and timeouts.

Gateway URL Security

The config loader (internal/config/gateway.go) enforces HTTPS for non-loopback hosts. If AI_GATEWAY_BASE_URL uses http:// for anything other than localhost / 127.x.x.x / ::1, the URL is rewritten to https:// and a warning is emitted. Invalid URLs fall back to the default OpenCode Go endpoint.