tui: add custom provider via /connect command palette #48935

Summary

Add a "Custom Provider" flow in the TUI /connect command palette that lets users register any provider by entering an endpoint and optional API key, auto-discovers available models, and lets users select which models to enable.

Motivation

OpenCode has many built-in providers, but users frequently need to add:

• Self-hosted models (Ollama, vLLM, LocalAI, etc.)
• Third-party platforms (DeepSeek, Qwen, Kimi, etc.)
• Corporate/internal API gateways
• Any OpenAI-compatible endpoint

Currently this requires manually editing opencode.jsonc with the correct schema, npm package name, and model capabilities — which is error-prone and not discoverable.

Proposed Behavior

In the TUI, invoke /connect (or a new command like /add-provider) and present an interactive flow:

Step 1: Endpoint

Provider endpoint URL: https://api.deepseek.com/v1

• Accept any URL (local or remote)
• Validate the URL is reachable

Step 2: API Key (optional)

API key (leave empty for no auth): ****

• Key is optional because:
  • Local providers (Ollama, LocalAI) typically don't need a key
  • Some platforms return different model lists based on whether a key is provided
• Key should be stored securely (not in plain text in config if possible)

Step 3: Auto-fetch Models

After entering endpoint + key, automatically call the provider's model list endpoint (e.g., /v1/models) to discover available models.

Fetching models from https://api.deepseek.com/v1/models ...

Available models:
  [x] deepseek-chat          (context: 64000, output: 8000)
  [x] deepseek-reasoner      (context: 64000, output: 8000)
  [ ] deepseek-v3            (context: 128000, output: 16000)
  [ ] deepseek-v2.5          (context: 32000, output: 4000)

Select models to enable (space to toggle, enter to confirm):

• Pre-select all models, let user deselect unwanted ones
• Show model name and basic info (context window, output limit) when available
• If the endpoint doesn't support /v1/models, fall back to manual model name input

Step 4: Confirm & Save

Provider: deepseek
Endpoint: https://api.deepseek.com/v1
Models: deepseek-chat, deepseek-reasoner

Save to: [Global ~/.config/opencode/opencode.jsonc] / [Project .opencode/opencode.jsonc]
Confirm? (y/n)

Technical Notes

• The /v1/models endpoint is part of the OpenAI API spec and supported by most compatible providers
• For providers that don't support model listing, fall back to manual model name entry
• The config should use @ai-sdk/openai-compatible npm package as the adapter
• Consider supporting other adapter packages in the future (e.g., @ai-sdk/anthropic, @ai-sdk/google) if the endpoint is detected

Expected Config Output

After completing the flow, the following should be written to opencode.jsonc:

{
  "providers": {
    "deepseek": {
      "npm": "@ai-sdk/openai-compatible",
      "options": {
        "apiKey": "sk-xxx",
        "baseURL": "https://api.deepseek.com/v1"
      },
      "models": {
        "deepseek-chat": {
          "name": "deepseek-chat",
          "tool_call": true,
          "limit": { "context": 64000, "output": 8000 }
        },
        "deepseek-reasoner": {
          "name": "deepseek-reasoner",
          "reasoning": true,
          "tool_call": true,
          "limit": { "context": 64000, "output": 8000 }
        }
      }
    }
  }
}

Environment

• opencode version: 2.0.3
• OS: Windows_NT 10.0.26200 (win32 x64)
• Terminal: TUI mode

Additional Context

This feature would make OpenCode much more accessible for users who don't want to manually edit JSON configs. The /connect palette is the natural place for this since it already handles provider connections.