feat: refine screenshot catalogs through SubLLM vision

.github/workflows/ci.yml

@@ -0,0 +1,24 @@
+name: CI
+
+on:
+  pull_request:
+    branches: [main]
+  push:
+    branches: [main]
+
+jobs:
+  test:
+    runs-on: ubuntu-latest
+    steps:
+      - uses: actions/checkout@v4
+      - uses: actions/setup-python@v5
+        with:
+          python-version: "3.13"
+      - name: Install test subjects
+        run: |
+          python -m pip install --upgrade pip
+          python -m pip install -e .
+          python -m pip install 'pytest>=8.0'
+          python -m pip install -e packages/mcp2imgl --no-deps
+      - name: Run tests
+        run: python -m pytest tests -q

imgl/cli.py

@@ -331,7 +331,7 @@ def build_parser() -> argparse.ArgumentParser:
     interact_parser.add_argument(
         "--llm",
         action="store_true",
-        help="Use vision LLM catalog (requires OPENROUTER_API_KEY, pip install litellm)",
+        help="Use vision LLM catalog (requires OPENROUTER_API_KEY, pip install -e '.[llm]')",
     )
     interact_parser.add_argument(
         "--no-filter",

imgl/llm_catalog.py

@@ -6,6 +6,7 @@
 import json
 import os
 import re
+import sys
 from io import BytesIO
 from pathlib import Path
 from typing import Any
@@ -86,13 +87,21 @@ def llm_available() -> bool:
 
 
 def llm_dependencies_ok() -> tuple[bool, str | None]:
+    if sys.version_info < (3, 11):
+        return False, "SubLLM vision requires Python 3.11+"
     try:
-        import litellm  # type: ignore  # noqa: F401
+        _subllm_complete()
     except ImportError:
-        return False, "litellm not installed (pip install -e '.[llm]')"
+        return False, "subactor-subllm not installed (pip install -e '.[llm]')"
     return True, None
 
 
+def _subllm_complete():
+    from subllm import complete as subllm_complete
+
+    return subllm_complete
+
+
 def refine_catalog_with_llm(
     scene: Scene,
     *,
@@ -219,12 +228,7 @@ def _call_vision_llm(
     crop_bbox: BBox | None = None,
     window_title: str | None = None,
 ) -> dict[str, Any]:
-    os.environ.setdefault("LITELLM_LOG", "ERROR")
-    import litellm  # type: ignore
-
-    litellm.set_verbose = False
-    if hasattr(litellm, "suppress_debug_info"):
-        litellm.suppress_debug_info = True
+    subllm_complete = _subllm_complete()
     image_b64 = _image_to_base64(image_path, crop_bbox=crop_bbox)
     scope_hint = (
         f"This image shows one application window{f' ({window_title})' if window_title else ''}. "
@@ -251,13 +255,15 @@ def _call_vision_llm(
             ],
         },
     ]
-    response = litellm.completion(
-        model=model,
-        messages=messages,
-        temperature=0.1,
+    api_key = os.getenv("OPENROUTER_API_KEY", "").strip()
+    response = subllm_complete(
+        "autogrammar-imgl",
+        "vision",
+        messages,
         response_format={"type": "json_object"},
+        credentials={"openrouter": api_key} if api_key else None,
     )
-    content = (response.choices[0].message.content or "").strip()
+    content = (response.content or "").strip()
     return _parse_json_payload(content)
 
 

pyproject.toml

@@ -45,8 +45,8 @@ diagnose = [
     "numpy>=1.24",
 ]
 llm = [
-    "litellm>=1.30",
     "python-dotenv>=1.0",
+    "subactor-subllm @ git+https://github.com/subactor/subllm.git@067732fcd2fab3b36453160bb4497012b13bc298 ; python_version >= '3.11'",
 ]
 web = [
     "fastapi>=0.110",

tests/test_llm_catalog.py

@@ -124,3 +124,26 @@ def test_merge_heuristic_inputs_appends_missing_fields():
         for opt in inputs
         for term in ("Chat", "Terminal", "Pole", "Editor", "Input")
     )
+
+
+def test_call_vision_llm_uses_central_subllm(tmp_path: Path, monkeypatch):
+    from imgl import llm_catalog
+
+    captured: dict = {}
+    from PIL import Image
+
+    image = tmp_path / "shot.png"
+    Image.new("RGB", (8, 8), color=(255, 255, 255)).save(image)
+
+    def fake_complete(application, function, messages, **kwargs):
+        captured.update(application=application, function=function, messages=messages, kwargs=kwargs)
+        return type("Response", (), {"content": '{"elements":[]}'})()
+
+    monkeypatch.setenv("OPENROUTER_API_KEY", "sk-test")
+    monkeypatch.setattr(llm_catalog, "_subllm_complete", lambda: fake_complete)
+    payload = llm_catalog._call_vision_llm(str(image), model="ignored", max_elements=5)
+    assert payload == {"elements": []}
+    assert captured["application"] == "autogrammar-imgl"
+    assert captured["function"] == "vision"
+    assert captured["messages"][1]["content"][0]["type"] == "image_url"
+    assert captured["kwargs"]["credentials"] == {"openrouter": "sk-test"}