识图与文件上传/预览:可配置视觉模型 + 修复上传与回显

后端:
- 新增可配置识图模型(config: vision.model_name/prompt/max_tokens)。配置后 view_image 委托该模型识别并返回文字结果,文本主模型也能“看图”;未配置时回退到主模型自带视觉能力。新增 deerflow/models/vision.py 与 VisionConfig。
- 修复 Windows 上传全部报 unsafe:open_upload_file_no_symlink 不再强制要求 POSIX 专有的 O_NOFOLLOW,缺失时降级(仍用 lstat/fstat 防符号链接),Windows 追加 O_BINARY。
- 修复 normalize_input 丢弃 human 消息 additional_kwargs:保留 files/client_ts 等,使上传文件能回传到 UI 并被 UploadsMiddleware 识别。
- 约束 agent 行为:上传图片时引导只用 view_image,禁止用 read_file/grep/bash/python 解析图片字节;view_image 工具说明同步加固。
- 启用文档转换(uploads.auto_convert_documents),PDF/Word/Excel/PPT 自动转 Markdown 供 read_file/grep 读取。
- 订正若干模型错误的 supports_vision 标记。

前端:
- 修复上传文件不回显:parseUploadedFiles 不再因 "(empty)" 提前返回。
- 修复 artifact 401 破图:改用带 Authorization 头的 apiFetch 取 blob 显示(新增 useAuthedObjectUrl)。
- 新增上传文件预览弹窗:点击卡片可查看图片/文本/Markdown、Office(读转换后的 .md)、PDF;文件卡片改为可点击。

Co-Authored-By: Claude Opus 4.8 (1M context) <noreply@anthropic.com>
This commit is contained in:
吉鑫升andClaude Opus 4.8 committed 2026-06-02 16:43:53 +08:00
1 parent 43cbcd86be
commit 5aa1fdca04
17 files changed
+706 -74

No files matched your search

@@ -51,20 +51,33 @@ def view_image_tool(
runtime: ToolRuntime[ContextT, ThreadState],
image_path: str,
tool_call_id: Annotated[str, InjectedToolCallId],
question: str | None = None,
) -> Command:
"""Read an image file.
"""Read and understand an image file.
Use this tool to read an image file and make it available for display.
This is the ONLY correct way to understand an image. The image is analyzed by
the configured vision model and its description/answer is returned to you as
the tool result. Treat that result as authoritative.
When to use the view_image tool:
- When you need to view an image file.
- Whenever you need to view, describe, or answer a question about an image file.
When NOT to use the view_image tool:
- For non-image files (use present_files instead)
- For multiple files at once (use present_files instead)
IMPORTANT — do NOT attempt to understand images any other way:
- Never read, open, decode, or parse an image with read_file, grep, bash,
python, PIL/Pillow, OpenCV, EXIF parsing, or byte/hex inspection. Those
cannot interpret visual content and will only waste turns.
- If `view_image` returns a weak or incomplete description, do not fall back
to code. Either call `view_image` again with a more specific `question`,
or tell the user the configured vision model could not extract more detail
(the vision model may need to be changed in config).
Args:
image_path: Absolute /mnt/user-data virtual path to the image file. Common formats supported: jpg, jpeg, png, webp.
question: Optional specific question to ask about the image (e.g. "What is the total on this invoice?"). When omitted, a general detailed description is produced.
"""
from deerflow.sandbox.exceptions import SandboxRuntimeError
from deerflow.sandbox.tools import (
@@ -153,8 +166,27 @@ def view_image_tool(
mime_type = detected_mime_type
image_base64 = base64.b64encode(image_data).decode("utf-8")
# Update viewed_images in state
# The merge_viewed_images reducer will handle merging with existing images
# Delegated recognition: when a dedicated vision model is configured, send
# the image to that model and return its description as the tool result, so
# a text-only main chat model can still "see" the image.
from deerflow.models.vision import is_delegated_vision, recognize_image
if is_delegated_vision():
try:
description = recognize_image(image_base64, mime_type, question=question)
except Exception as e: # noqa: BLE001 - surface as a recoverable tool error
return Command(
update={"messages": [ToolMessage(f"Error recognizing image: {_sanitize_image_error(e, thread_data)}", tool_call_id=tool_call_id)]},
)
if not description:
description = "The vision model returned an empty description."
return Command(
update={"messages": [ToolMessage(f"Image recognition result for {image_path}:\n\n{description}", tool_call_id=tool_call_id)]},
)
# Native vision: store the image so ViewImageMiddleware injects it into the
# (vision-capable) main model before the next LLM call.
# The merge_viewed_images reducer will handle merging with existing images.
new_viewed_images = {image_path: {"base64": image_base64, "mime_type": mime_type}}
return Command(
@@ -138,11 +138,14 @@ def get_available_tools(
if model_name is None and config.models:
model_name = config.models[0].name
# Add view_image_tool only if the model supports vision
model_config = config.get_model_config(model_name) if model_name else None
if model_config is not None and model_config.supports_vision:
# Add view_image_tool when vision is available — either the main model
# supports vision natively, or a dedicated vision model is configured
# (vision.model_name) so recognition can be delegated to it.
from deerflow.models.vision import vision_enabled
if vision_enabled(model_name, app_config=config):
builtin_tools.append(view_image_tool)
logger.info(f"Including view_image_tool for model '{model_name}' (supports_vision=True)")
logger.info(f"Including view_image_tool for model '{model_name}' (vision available)")
# Get cached MCP tools if enabled
# NOTE: We use ExtensionsConfig.from_file() instead of config.extensions