Bug: opencode run fails with "Unexpected server error" while API and TUI work fine

Bug Report: opencode run fails with "Unexpected server error" while API and TUI work fine

Description

The opencode run command consistently fails with UnknownError: Unexpected server error when using the OpenCode Zen provider. However:

1. The API works perfectly when called directly via curl with the same API key and model
2. The TUI works when run interactively in a terminal (human input)
3. The same configuration that fails in run mode worked in previous sessions

This appears to be a bug in how the run (batch/non-interactive) command constructs or handles the API request, not a server-side or API key issue.

Environment

• OpenCode version: 1.18.5 (auto-upgraded from 1.18.4 during session)
• OS: Linux 7.1.3-ogc5.1.fc44.x86_64 (Fedora/Bazzite, x86_64)
• Node.js: v20.20.2
• Provider: OpenCode Zen (opencode_zen)
• Model: big-pickle

Configuration

// ~/.config/opencode/opencode.jsonc
{
  "$schema": "https://opencode.ai/config.json",
  "provider": {
    "opencode_zen": {
      "name": "OpenCode Zen",
      "npm": "@ai-sdk/openai",
      "options": {
        "baseURL": "https://opencode.ai/zen/v1",
        "apiKey": "sk-..."
      }
    }
  },
  "model": "opencode_zen/big-pickle"
}

Steps to Reproduce

1. Configure OpenCode with OpenCode Zen provider and big-pickle model
2. Run: opencode run 'Say OK'
3. Observe the error

Actual Result

{
  "name": "UnknownError",
  "data": {
    "message": "Unexpected server error. Check server logs for details.",
    "ref": "err_<unique-id>"
  }
}

This error occurs consistently with every opencode run attempt. Each attempt generates a unique ref ID.

Expected Result

The command should execute the prompt and return the model's response, as it does via:

• Direct API call (curl)
• Interactive TUI mode

Evidence: API Works Via curl

curl -s "https://opencode.ai/zen/v1/chat/completions" \
  -H "Authorization: Bearer sk-..." \
  -H "Content-Type: application/json" \
  -d '{"model":"big-pickle","messages":[{"role":"user","content":"Say OK"}],"max_tokens":10}'

Response (HTTP 200):

{
  "id": "gen-...",
  "model": "big-pickle",
  "choices": [{
    "index": 0,
    "finish_reason": "length",
    "message": {
      "role": "assistant",
      "content": null,
      "reasoning": "The user wants me to say \"OK\"."
    }
  }],
  "usage": {
    "prompt_tokens": 249,
    "completion_tokens": 10,
    "total_tokens": 259
  }
}

Evidence: TUI Works Interactively

Running opencode (no arguments) opens the TUI, which successfully connects to the same provider/model and processes messages when typed interactively by a human.

Evidence: run Command Fails

$ opencode run 'Say OK'
Error: {
  "name": "UnknownError",
  "data": {
    "message": "Unexpected server error. Check server logs for details.",
    "ref": "err_fa563afb"
  }
}

Also tried with explicit model flag:

$ opencode run 'Say OK' --model opencode_zen/claude-sonnet-4
Error: {
  "name": "UnknownError",
  "data": {
    "message": "Unexpected server error. Check server logs for details.",
    "ref": "err_e4073cfe"
  }
}

Log Analysis

The OpenCode log (~/.local/share/opencode/log/opencode.log) shows:

• TUI sessions initialize and stream successfully with providerID=opencode modelID=big-pickle
• The run command attempts are not logged with stream details — they fail before reaching the streaming stage
• No specific error details are logged for the run failures

# Successful TUI session (from earlier):
timestamp=2026-07-24T19:54:04.438Z level=INFO run=b3c41545 message=stream providerID=opencode modelID=big-pickle session.id=... small=false agent=build mode=primary
timestamp=2026-07-24T19:54:04.440Z level=INFO run=b3c41545 message="llm runtime selected" llm.runtime=ai-sdk llm.provider=opencode llm.model=big-pickle

# Failed `run` attempt — only shows init, no stream:
timestamp=2026-07-25T03:18:13.940Z level=INFO run=f77bccff message="creating instance" directory=...
timestamp=2026-07-25T03:18:14.155Z level=INFO run=f77bccff message=init
timestamp=2026-07-25T03:19:14.481Z level=INFO run=f77bccff message=cleanup prune=7.days

Additional Context

• The API key is valid and was tested immediately before/after the run failures
• The model big-pickle is confirmed available in the /v1/models endpoint
• Multiple models were tested (big-pickle, claude-sonnet-4) — all fail with the same error in run mode
• The issue started on 2026-07-25; the same configuration worked on 2026-07-24

Hypothesis

The run command may be:

1. Adding different headers or request parameters compared to the TUI/curl
2. Using a different authentication flow that the Zen API rejects
3. Having an issue with how it handles the streaming response in non-interactive mode
4. Experiencing a regression in how the AI SDK constructs the request in batch mode