Skip to content

Your lightweight open-source personal agent.

fix(providers): preserve image inputs for deepseek-flash

Kail Tian8cc5748464b9
diff --git a/nanobot/providers/openai_compat_provider.py b/nanobot/providers/openai_compat_provider.py
index 2d3de903a..35e65dfd3 100644
--- a/nanobot/providers/openai_compat_provider.py
+++ b/nanobot/providers/openai_compat_provider.py
@@ -117,6 +117,7 @@ _KIMI_SERVER_MANAGED_TEMPERATURE_MODELS: frozenset[str] = frozenset({
"kimi-k2.6",
})
_DEEPSEEK_MULTIMODAL_MODELS: frozenset[str] = frozenset({
+ "deepseek-flash",
"deepseek-v4-flash-vision-exp",
})
_TEXT_TOOL_CALL_RE = re.compile(r"<tool_call>\s*(.*?)\s*</tool_call>", re.DOTALL)
diff --git a/tests/providers/test_litellm_kwargs.py b/tests/providers/test_litellm_kwargs.py
index 917463f82..7253e347d 100644
--- a/tests/providers/test_litellm_kwargs.py
+++ b/tests/providers/test_litellm_kwargs.py
@@ -1624,31 +1624,49 @@ def test_deepseek_coerces_list_content_to_string() -> None:
assert "world" in kw["messages"][0]["content"]
-def test_deepseek_vision_preserves_multimodal_content() -> None:
- """DeepSeek's vision model requires OpenAI-compatible content blocks."""
+ "deepseek-flash",
+ "deepseek-v4-flash-vision-exp",
+])
[email protected]("prefixed", [False, True])
[email protected]("with_text", [False, True])
[email protected]("responses", [False, True])
+def test_deepseek_vision_preserves_multimodal_content(
+ model: str, prefixed: bool, with_text: bool, responses: bool,
+) -> None:
+ """DeepSeek vision models must preserve images on both API surfaces."""
+ if prefixed:
+ model = f"deepseek/{model}"
spec = find_by_name("deepseek")
with patch("nanobot.providers.openai_compat_provider.AsyncOpenAI"):
p = OpenAICompatProvider(
api_key="k",
- default_model="deepseek-v4-flash-vision-exp",
+ default_model=model,
spec=spec,
)
- content = [
- {"type": "text", "text": "describe this image"},
- {"type": "image_url", "image_url": {"url": "data:image/png;base64,AA=="}},
- ]
+ urls = ["data:image/png;base64,AA==", "https://example.com/image.png"]
+ content = [{"type": "text", "text": "describe these images"}] if with_text else []
+ content.extend({"type": "image_url", "image_url": {"url": url}} for url in urls)
- kw = p._build_kwargs(
+ build_request = p._build_responses_body if responses else p._build_kwargs
+ kw = build_request(
messages=[{"role": "user", "content": content}],
tools=None,
- model="deepseek-v4-flash-vision-exp",
+ model=model,
max_tokens=1024,
temperature=0.7,
reasoning_effort=None,
tool_choice=None,
)
- assert kw["messages"][0]["content"] == content
+ if responses:
+ expected = [{"type": "input_text", "text": "describe these images"}] if with_text else []
+ expected.extend(
+ {"type": "input_image", "image_url": url, "detail": "auto"} for url in urls
+ )
+ assert kw["input"] == [{"role": "user", "content": expected}]
+ else:
+ assert kw["messages"][0]["content"] == content
def test_non_deepseek_keeps_list_content() -> None: