environment.yml

name: multi_agent_cad
channels:
  - conda-forge
  - defaults

dependencies:
  # ── Runtime: Python ───────────────────────────────────────────────────
  - python=3.11.*

  # ── Core scientific / math (C extensions — stability from conda-forge) ──
  - numpy>=1.24,<2.3
  - scipy>=1.10
  - scikit-learn>=1.3

  # ── Mesh analysis (C extensions — pyembree, rtree) ────────────────────
  - trimesh>=4.0
  - rtree>=1.1

  # ── Toolchain ─────────────────────────────────────────────────────────
  - pip>=23.0
  - setuptools>=68.0
  - wheel>=0.41

  # ── pip-installed packages (fast-moving, PyPI-first) ──────────────────
  - pip:
      # --- CAD kernel ---
      # build123d pulls in cadquery-ocp-novtk (provides `OCP`) transitively
      - build123d>=0.8

      # --- LangGraph orchestration (pinned for Python 3.11 compat) ---
      - langgraph>=0.2,<0.3
      - langgraph-checkpoint>=2.0,<3.0

      # --- Data contracts ---
      - pydantic>=2.5

      # --- LLM clients (DashScope / Qwen via OpenAI-compatible endpoint) ---
      - openai>=1.20.0

      # NOTE: aider-chat is intentionally NOT installed here. Every version
      # on PyPI (0.50.0 through 0.86.2+) hard-pins numpy==1.26.4 (1.x),
      # which is incompatible with build123d's numpy>=2 requirement, and
      # conda's pip subprocess can't bypass that pin. Install it manually
      # AFTER `conda env create` succeeds:
      #   pip install --no-deps "aider-chat==0.82.3"
      # --no-deps skips the numpy==1.26.4 pin; aider 0.82.3 imports
      # cleanly on numpy 2.x (the pin is over-cautious upstream).

      # --- Anthropic SDK (备用 LLM 接口) ---
      - anthropic>=0.30

      # --- Dev / debug (optional but useful) ---
      - ipython>=8.15
      - pytest>=7.4