fix: align Anthropic provider with Messages API and packaging

Author: Achieve3318

paperbanana/providers/vlm/anthropic.py

@@ -72,7 +72,7 @@ async def generate(
                 b64 = image_to_base64(img)
                 content.append(
                     {
-                        "type": "input_image",
+                        "type": "image",
                         "source": {
                             "type": "base64",
                             "media_type": "image/png",
@@ -101,8 +101,17 @@ async def generate(
             params["system"] = system_prompt
 
         if response_format == "json":
-            # Constrain the model to emit a JSON object if supported by the SDK.
-            params["response_format"] = {"type": "json_object"}
+            # Use structured outputs with a permissive JSON schema so callers
+            # receive machine-parseable JSON while retaining flexibility.
+            params["output_config"] = {
+                "format": {
+                    "type": "json_schema",
+                    "schema": {
+                        "type": "object",
+                        "additionalProperties": True,
+                    },
+                }
+            }
 
         response = await client.messages.create(**params)
 

pyproject.toml

@@ -46,7 +46,13 @@ dependencies = [
 google = ["google-genai>=1.65"]
 openai = ["openai>=1.0"]
 bedrock = ["boto3>=1.34"]
-all-providers = ["google-genai>=1.65", "openai>=1.0", "boto3>=1.34"]
+anthropic = ["anthropic>=0.83"]
+all-providers = [
+    "google-genai>=1.65",
+    "openai>=1.0",
+    "boto3>=1.34",
+    "anthropic>=0.83",
+]
 mcp = ["fastmcp>=2.0"]
 dev = [
     "pytest>=8.0",

tests/test_providers/test_anthropic_vlm.py

@@ -86,8 +86,12 @@ def _fake_image_to_base64(_img: Image.Image) -> str:
     msg = captured["messages"][0]
     assert msg["role"] == "user"
     content = msg["content"]
-    assert content[0]["type"] == "input_image"
+    assert content[0]["type"] == "image"
     assert content[0]["source"]["data"] == "base64-image-data"
     assert content[-1]["type"] == "text"
     assert content[-1]["text"] == "Hi with image"
-    assert captured["response_format"] == {"type": "json_object"}
+
+    output_config = captured["output_config"]
+    fmt = output_config["format"]
+    assert fmt["type"] == "json_schema"
+    assert isinstance(fmt["schema"], dict)