"gentleman-local": {
  "mode": "primary",
  "model": "ollama/qwen2.5:7b",
  "prompt": "Your working directory is /Users/me/project. Always use absolute paths when calling file tools."
}
2. Start a session with this agent
3. Ask it to read a file: read pom.xml
4. The model calls read with a placeholder path like /path/to/pom.xml instead of /Users/me/project/pom.xml
Expected Behavior
The prompt field should be sent as the system message in the chat completion request, exactly as it works for cloud models (e.g. Claude, GPT).
Actual Behavior
The model receives no system prompt. Verified by:
- Checking opencode logs: the prompt content never appears
- Testing the same model directly via curl with the prompt as system message → model behaves correctly
- Testing with inline prompt string in opencode.json (not {file:...}) → same result, prompt is ignored


### Screenshot and/or share link

Evidence
Direct curl with system prompt works correctly:
curl http://localhost:11434/v1/chat/completions \
  -H "Content-Type: application/json" \
  -d '{
    "model": "qwen2.5:7b",
    "messages": [
      {"role": "system", "content": "Working directory: /Users/me/project. Always prepend it to file paths."},
      {"role": "user", "content": "read pom.xml"}
    ],
    "tools": [...]
  }'
# → calls read with /Users/me/project/pom.xml ✅
Same model via opencode agent with prompt set:
# → calls read with /path/to/pom.xml ❌ (prompt was not sent)
Environment
- opencode version: 1.14.24
- OS: macOS (Apple M4, darwin)
- Provider: Ollama 0.21.1 via @ai-sdk/openai-compatible
- Model tested: qwen2.5:7b, qwen2.5-coder:7b
- Cloud agents (Claude Sonnet) work correctly with the same prompt field

### Operating System

26.3.1

### Terminal

iTerm2