chore(deps): make vLLM an optional extra so base install works on macOS

README.md

@@ -98,6 +98,12 @@ make lint      # ruff check .
 
 For local vLLM inference (NVIDIA GPU, CUDA 12.8, ≥ 24 GB VRAM):
 
+```bash
+uv sync --extra vllm
+```
+
+vLLM is an optional extra so the base install resolves on macOS / non-CUDA hosts. If you don't use `uv`:
+
 ```bash
 uv pip install vllm --extra-index-url https://download.pytorch.org/whl/cu128
 ```

pyproject.toml

@@ -20,12 +20,14 @@ dependencies = [
     "transformers>=4.40.0",
     "sentence-transformers>=5.2.3",
     "torch>=2.10.0",
-    "vllm==0.17.0",
     "openai>=2.24.0",
     "anthropic>=0.75.0",
     "google-genai>=1.0.0",
     "pyarrow>=16.0.0",
 ]
 
+[project.optional-dependencies]
+vllm = ["vllm==0.17.0"]
+
 [tool.pytest.ini_options]
 pythonpath = ["."]

sob/providers/anthropic_native.py

@@ -36,15 +36,20 @@ def _infer_one(
     candidate = None
     input_tokens = output_tokens = 0
 
+    # Opus 4.7 deprecated temperature/top_p/top_k — passing any of them
+    # returns a 400. See https://platform.claude.com/docs/en/about-claude/models/whats-new-claude-4-7
+    kwargs: dict = dict(
+        model=config.model_id,
+        max_tokens=config.max_tokens,
+        system=SYSTEM_PROMPT,
+        messages=[{"role": "user", "content": user_msg}],
+    )
+    if "opus-4-7" not in config.model_id:
+        kwargs["temperature"] = config.temperature
+
     for attempt in range(config.max_retries):
         try:
-            response = client.messages.create(
-                model=config.model_id,
-                max_tokens=config.max_tokens,
-                system=SYSTEM_PROMPT,
-                messages=[{"role": "user", "content": user_msg}],
-                temperature=config.temperature,
-            )
+            response = client.messages.create(**kwargs)
             # Anthropic returns a list of content blocks; text lives on the
             # first one (TextBlock) for plain JSON tasks.
             raw = ""