ByteDance Seed models hang due to non-standard reasoning token streaming

Description

ByteDance Seed models (bytedance-seed/seed-2.0-lite, seed-2.0-mini, seed-1.6, seed-1.6-flash) via OpenRouter hang indefinitely when used with opencode run. The issue is that these models return reasoning and reasoning_details fields in their streaming response delta chunks, which opencode's parser does not handle.

Reproduction

opencode run "Reply OK" -m openrouter/bytedance-seed/seed-2.0-lite

The command hangs at > build · bytedance-seed/seed-2.0-lite and never completes.

Root Cause

The OpenRouter streaming response for ByteDance Seed models includes non-standard fields in the delta:

{
  "choices": [{
    "delta": {
      "content": "",
      "role": "assistant",
      "reasoning": "Got it, the user asked...",
      "reasoning_details": [{"type": "reasoning.text", "text": "...", "format": "unknown", "index": 0}]
    }
  }]
}

During the reasoning phase, content is "" (empty string) while reasoning carries the thinking tokens. This causes opencode's parser to hang — it sees empty content and waits indefinitely for more.

Workaround

Using include_reasoning: false in the API request body disables reasoning output and the model works via curl. However, there's no way to pass this parameter through opencode's config.

Expected Behavior

Either:

1. Parse the reasoning field correctly (like reasoning_content for DeepSeek models)
2. Ignore unknown fields in the streaming delta and process content only
3. Expose providerOptions in the config schema so users can pass reasoning: {exclude: true}

Environment

• OpenCode version: 1.18.5
• Platform: macOS (darwin, arm64)
• Provider: OpenRouter
• Models: openrouter/bytedance-seed/seed-2.0-lite, seed-2.0-mini, seed-1.6, seed-1.6-flash