Skip to content

Use Claude Code, Codex, VSCode, Pi, and OpenCode (and 6 other harnesses) for free (1.3B+ free tokens) from your terminal, app, IDE, or phone, and now from the browser with native browser sessions (multi-harness + multi-model) like OpenClaw (voice supported + ToS friendly)

Python

Forward vision attachments for DeepSeek vision models (#1498)

wocessade4c7cced57ce6
diff --git a/src/free_claude_code/providers/deepseek/compat.py b/src/free_claude_code/providers/deepseek/compat.py
index d941c06d..f62503ff 100644
--- a/src/free_claude_code/providers/deepseek/compat.py
+++ b/src/free_claude_code/providers/deepseek/compat.py
@@ -39,12 +39,25 @@ _UNSUPPORTED_MESSAGE_BLOCK_TYPES = frozenset(
}
)
_STRIPPABLE_MESSAGE_BLOCK_TYPES = frozenset({"image", "document"})
+_FORWARDABLE_ATTACHMENT_BLOCK_TYPES = frozenset({"image"})
_OMITTED_ATTACHMENT_TEXT = (
"[attachment omitted: DeepSeek does not support image or document inputs]"
)
_OMITTED_ATTACHMENT_BLOCK = {"type": "text", "text": _OMITTED_ATTACHMENT_TEXT}
+def _is_vision_capable_model(model: str | None) -> bool:
+ """True when the DeepSeek model accepts OpenAI image parts.
+
+ DeepSeek's vision models (e.g. ``deepseek-v4-flash-vision-exp``) support
+ image inputs natively. Document blocks are still stripped because the
+ shared OpenAI conversion path has no document parts.
+ """
+ if not model:
+ return False
+ return "vision" in str(model).rsplit("/", 1)[-1].lower()
+
+
def build_deepseek_request_body(
request_data: MessagesRequest, *, reasoning: ReasoningPolicy
) -> dict:
@@ -55,10 +68,13 @@ def build_deepseek_request_body(
len(request_data.messages),
)
+ vision_capable = _is_vision_capable_model(request_data.model)
data = dump_messages_request(request_data)
if "messages" in data:
- data["messages"] = _strip_unsupported_attachment_blocks(data["messages"])
- _validate_deepseek_request_dict(data)
+ data["messages"] = _strip_unsupported_attachment_blocks(
+ data["messages"], allow_attachments=vision_capable
+ )
+ _validate_deepseek_request_dict(data, allow_attachments=vision_capable)
_downgrade_forced_tool_choice(data)
has_tool_history = _has_tool_history(data)
@@ -155,7 +171,9 @@ def sanitize_deepseek_messages_for_openai(messages: Any) -> Any:
return sanitized
-def _strip_unsupported_attachment_blocks(messages: Any) -> Any:
+def _strip_unsupported_attachment_blocks(
+ messages: Any, *, allow_attachments: bool = False
+) -> Any:
if not isinstance(messages, list):
return messages
@@ -179,6 +197,12 @@ def _strip_unsupported_attachment_blocks(messages: Any) -> Any:
if isinstance(block, dict):
btype = block.get("type")
if btype in _STRIPPABLE_MESSAGE_BLOCK_TYPES:
+ if (
+ allow_attachments
+ and btype in _FORWARDABLE_ATTACHMENT_BLOCK_TYPES
+ ):
+ new_content.append(block)
+ continue
top_level_dropped[btype] = top_level_dropped.get(btype, 0) + 1
message_dropped_attachment = True
continue
@@ -192,6 +216,12 @@ def _strip_unsupported_attachment_blocks(messages: Any) -> Any:
and sub.get("type") in _STRIPPABLE_MESSAGE_BLOCK_TYPES
):
sub_type = sub["type"]
+ if (
+ allow_attachments
+ and sub_type in _FORWARDABLE_ATTACHMENT_BLOCK_TYPES
+ ):
+ filtered_inner.append(sub)
+ continue
nested_dropped[sub_type] = (
nested_dropped.get(sub_type, 0) + 1
)
@@ -234,7 +264,9 @@ def _is_server_listed_tool(tool: Mapping[str, Any]) -> bool:
return False
-def _walk_block_list_for_unsupported(blocks: Any, *, where: str) -> None:
+def _walk_block_list_for_unsupported(
+ blocks: Any, *, where: str, allow_attachments: bool = False
+) -> None:
if not isinstance(blocks, list):
return
for block in blocks:
@@ -242,16 +274,22 @@ def _walk_block_list_for_unsupported(blocks: Any, *, where: str) -> None:
continue
btype = block.get("type")
if btype in _UNSUPPORTED_MESSAGE_BLOCK_TYPES:
+ if allow_attachments and btype in _FORWARDABLE_ATTACHMENT_BLOCK_TYPES:
+ continue
raise InvalidRequestError(
f"DeepSeek native does not support {btype!r} blocks ({where})."
)
if btype == "tool_result" and "content" in block:
_walk_block_list_for_unsupported(
- block["content"], where=f"{where} (tool_result content)"
+ block["content"],
+ where=f"{where} (tool_result content)",
+ allow_attachments=allow_attachments,
)
-def _validate_deepseek_request_dict(data: dict[str, Any]) -> None:
+def _validate_deepseek_request_dict(
+ data: dict[str, Any], *, allow_attachments: bool = False
+) -> None:
mcp = data.get("mcp_servers")
if mcp:
raise InvalidRequestError("DeepSeek does not support mcp_servers on requests.")
@@ -270,11 +308,17 @@ def _validate_deepseek_request_dict(data: dict[str, Any]) -> None:
continue
content = message.get("content")
if isinstance(content, list):
- _walk_block_list_for_unsupported(content, where=f"messages[{i}].content")
+ _walk_block_list_for_unsupported(
+ content,
+ where=f"messages[{i}].content",
+ allow_attachments=allow_attachments,
+ )
system = data.get("system")
if isinstance(system, list):
- _walk_block_list_for_unsupported(system, where="system")
+ _walk_block_list_for_unsupported(
+ system, where="system", allow_attachments=allow_attachments
+ )
def _has_tool_history_blocks(message: Mapping[str, Any]) -> bool:
diff --git a/tests/providers/test_deepseek.py b/tests/providers/test_deepseek.py
index 7748a926..89e7ac20 100644
--- a/tests/providers/test_deepseek.py
+++ b/tests/providers/test_deepseek.py
@@ -10,6 +10,7 @@ from free_claude_code.application.errors import InvalidRequestError
from free_claude_code.config.constants import ANTHROPIC_DEFAULT_MAX_OUTPUT_TOKENS
from free_claude_code.config.provider_catalog import DEEPSEEK_DEFAULT_BASE
from free_claude_code.core.anthropic.models import (
+ ContentBlockDocument,
ContentBlockImage,
Message,
MessagesRequest,
@@ -716,6 +717,106 @@ def test_preflight_strips_user_image():
assert "image or document inputs" in content.lower()
+def test_vision_model_forwards_user_image():
+ """Vision-capable DeepSeek models receive image blocks as OpenAI image_url parts."""
+ request = MessagesRequest(
+ model="deepseek-v4-flash-vision-exp",
+ messages=[
+ Message(
+ role="user",
+ content=[
+ ContentBlockImage(
+ type="image",
+ source={
+ "type": "base64",
+ "media_type": "image/png",
+ "data": "YQ==",
+ },
+ ),
+ {"type": "text", "text": "Describe this image"},
+ ],
+ )
+ ],
+ )
+ provider = DeepSeekProvider(
+ make_provider_config(
+ api_key="k",
+ base_url=DEEPSEEK_DEFAULT_BASE,
+ rate_limit=1,
+ rate_window=1,
+ ),
+ admission=immediate_admission(),
+ )
+ # Must not raise on preflight (no InvalidRequestError for image blocks).
+ provider.preflight_stream(request, reasoning=REASONING_ON)
+ body = provider._build_request_body(request, reasoning=reasoning_for(request))
+ content = body["messages"][0]["content"]
+ assert isinstance(content, list)
+ image_parts = [
+ part
+ for part in content
+ if isinstance(part, dict) and part.get("type") == "image_url"
+ ]
+ assert len(image_parts) == 1
+ assert image_parts[0]["image_url"]["url"] == "data:image/png;base64,YQ=="
+ text_parts = [
+ part
+ for part in content
+ if isinstance(part, dict) and part.get("type") == "text"
+ ]
+ assert text_parts[0]["text"] == "Describe this image"
+
+
+def test_vision_model_strips_user_document():
+ """Vision models still omit documents; conversion has no document parts."""
+ request = MessagesRequest(
+ model="deepseek-v4-flash-vision-exp",
+ messages=[
+ Message(
+ role="user",
+ content=[
+ ContentBlockDocument(
+ type="document",
+ source={
+ "type": "base64",
+ "media_type": "application/pdf",
+ "data": "YQ==",
+ },
+ )
+ ],
+ )
+ ],
+ )
+ provider = DeepSeekProvider(
+ make_provider_config(
+ api_key="k",
+ base_url=DEEPSEEK_DEFAULT_BASE,
+ rate_limit=1,
+ rate_window=1,
+ ),
+ admission=immediate_admission(),
+ )
+ provider.preflight_stream(request, reasoning=REASONING_ON)
+ body = provider._build_request_body(request, reasoning=reasoning_for(request))
+ content = body["messages"][0]["content"]
+ assert content
+ if isinstance(content, str):
+ lowered = content.lower()
+ else:
+ texts = [
+ part.get("text", "")
+ for part in content
+ if isinstance(part, dict) and part.get("type") == "text"
+ ]
+ lowered = "\n".join(texts).lower()
+ assert not any(
+ isinstance(part, dict) and part.get("type") == "image_url"
+ for part in content
+ )
+ assert "attachment omitted" in lowered
+ assert "image or document inputs" in lowered
+
+
def test_preflight_rejects_mcp_servers():
request = MessagesRequest(
model="m",