opencode.jsonc

{
  "$schema": "https://opencode.ai/config.json",
  "instructions": ["AGENTS.md", "UPGRADE_PLAN.md"],

  // ── Multi-tier loop ──────────────────────────────────────────────
  // YOU → Commander (Ollama OSS-20B, local) → Planner (Qwen3-8B on NPU)
  //        → Worker (cloud) → Validator (cargo test/check) → Reviewer (cloud)
  //
  // Start the NPU server first:
  //   C:\Users\claimoar\Documents\npu-test\run_npu_server.bat
  // Health: curl http://localhost:11435/health  or  /v1/models
  //
  // Commander decides WHAT, Planner decomposes HOW (file-level tasks,
  // deps, risks — ~5 tok/s is fine, one-shot). Worker codes fast.
  // Validator proves (deterministic), Reviewer judges (cloud).

  "agent": {
    "commander": {
      "model": "ollama/gpt-oss:20b",
      "prompt": "You are the Commander. Understand the mission from UPGRADE_PLAN.md and AGENTS.md, control the loop, and delegate decomposition to the Planner. Never write code yourself — delegate to Worker."
    },
    "planner": {
      "model": "npu/Qwen3-8B",
      "prompt": "You are the Planner (Qwen3-8B on Intel AI Boost NPU). Decompose the Commander's objective into file-level tasks with dependencies, risks, and an ordered checklist. Do not write code. Be precise and grounded in the repo state (read AGENTS.md, UPGRADE_PLAN.md, Cargo.toml, crate src)."
    },
    "plan": {
      "model": "npu/Qwen3-8B"
    }
  },

  // NPU provider is defined globally in ~/.config/opencode/opencode.jsonc
  // (npu/Qwen3-8B → http://localhost:11435/v1). This project file only
  // routes agents; no provider duplication needed.
}