修复多轮/引用识图串图：历史不再持久化原图，失败图片 URL 丢弃，请求图本地优先

astrbot/core/agent/message.py

@@ -302,6 +302,72 @@ def get_checkpoint_id(message: Message | dict) -> str | None:
     return None
 
 
+
+_IMAGE_HISTORY_PLACEHOLDER = "[Image]"
+
+
+def _is_image_content_part(part: Any) -> bool:
+    if isinstance(part, ImageURLPart):
+        return True
+    if isinstance(part, dict) and part.get("type") == "image_url":
+        return True
+    return getattr(part, "type", None) == "image_url"
+
+
+def _image_placeholder_part_dict() -> dict[str, Any]:
+    return {"type": "text", "text": _IMAGE_HISTORY_PLACEHOLDER}
+
+
+def strip_images_from_content_parts(content: Any) -> Any:
+    """Replace image parts with a short text placeholder.
+
+    Used when persisting or reloading conversation history so previous-turn
+    image pixels / remote URLs do not keep leaking into later requests.
+    """
+    if not isinstance(content, list):
+        return content
+
+    stripped: list[Any] = []
+    previous_was_image_placeholder = False
+    for part in content:
+        if _is_image_content_part(part):
+            if previous_was_image_placeholder:
+                continue
+            if isinstance(part, ContentPart):
+                stripped.append(TextPart(text=_IMAGE_HISTORY_PLACEHOLDER))
+            else:
+                stripped.append(_image_placeholder_part_dict())
+            previous_was_image_placeholder = True
+            continue
+
+        previous_was_image_placeholder = False
+        stripped.append(part)
+    return stripped
+
+
+def strip_images_from_history_message_dict(message: dict[str, Any]) -> dict[str, Any]:
+    """Return a shallow-copied history message dict without image payloads."""
+    if not isinstance(message, dict):
+        return message
+    if is_checkpoint_message(message):
+        return message
+
+    content = message.get("content")
+    if not isinstance(content, list):
+        return message
+
+    new_message = dict(message)
+    new_message["content"] = strip_images_from_content_parts(content)
+    return new_message
+
+
+def strip_images_from_history_messages(history: list[dict] | None) -> list[dict]:
+    """Strip image payloads from a full persisted history list."""
+    if not history:
+        return []
+    return [strip_images_from_history_message_dict(item) for item in history]
+
+
 def strip_checkpoint_messages(history: list[dict]) -> list[dict]:
     """Remove internal checkpoint messages from provider-facing history."""
     return [message for message in history if not is_checkpoint_message(message)]
@@ -327,7 +393,8 @@ def _get_checkpoint_data(message: Message | dict) -> CheckpointData | None:
 def bind_checkpoint_messages(history: list[dict]) -> list[Message]:
     """Load persisted history and bind checkpoint segments to prior messages."""
     messages: list[Message] = []
-    for item in history:
+    for raw_item in history:
+        item = strip_images_from_history_message_dict(raw_item)
         if is_checkpoint_message(item):
             checkpoint = _get_checkpoint_data(item)
             if checkpoint is not None and messages:
@@ -348,10 +415,22 @@ def dump_messages_with_checkpoints(messages: list[Message]) -> list[dict]:
     for message in messages:
         message_data = message.model_dump()
         if isinstance(message.content, list):
+            persisted_parts: list[Any] = []
+            previous_was_image_placeholder = False
+            for part in message.content:
+                if getattr(part, "_no_save", False):
+                    continue
+                if _is_image_content_part(part):
+                    if previous_was_image_placeholder:
+                        continue
+                    persisted_parts.append(TextPart(text=_IMAGE_HISTORY_PLACEHOLDER))
+                    previous_was_image_placeholder = True
+                    continue
+                previous_was_image_placeholder = False
+                persisted_parts.append(part)
             message_data["content"] = [
-                part.model_dump()
-                for part in message.content
-                if not getattr(part, "_no_save", False)
+                part.model_dump() if isinstance(part, ContentPart) else part
+                for part in persisted_parts
             ]
         dumped.append(message_data)
         if message._checkpoint_after is not None:

astrbot/core/astr_main_agent.py

@@ -115,6 +115,7 @@
     extract_quoted_message_text,
 )
 from astrbot.core.utils.string_utils import normalize_and_dedupe_strings
+from astrbot.core.agent.message import strip_images_from_history_messages
 from astrbot.core.workspace import (
     normalize_umo_for_workspace,
     resolve_workspace_root_for_umo,
@@ -1351,6 +1352,43 @@ def _select_image_chat_provider(
     return provider
 
 
+
+def _is_remote_image_ref(image_ref: str) -> bool:
+    return image_ref.startswith("http://") or image_ref.startswith("https://")
+
+
+def _finalize_request_image_urls(
+    image_urls: list[str] | None,
+    *,
+    max_total: int = 4,
+) -> list[str]:
+    """Dedupe and prioritize local image refs over remote URLs.
+
+    Remote CDN links (especially QQ multimedia URLs) expire easily. Keeping them
+    beside already-downloaded local paths causes models to "see" missing/wrong
+    images in later turns.
+    """
+    urls = normalize_and_dedupe_strings(image_urls or [])
+    if not urls:
+        return []
+
+    local_refs: list[str] = []
+    embedded_refs: list[str] = []
+    remote_refs: list[str] = []
+    for ref in urls:
+        if _is_remote_image_ref(ref):
+            remote_refs.append(ref)
+        elif ref.startswith("data:") or ref.startswith("base64://"):
+            embedded_refs.append(ref)
+        else:
+            local_refs.append(ref)
+
+    ordered = local_refs + embedded_refs + remote_refs
+    if max_total > 0:
+        ordered = ordered[:max_total]
+    return ordered
+
+
 async def build_main_agent(
     *,
     event: AstrMessageEvent,
@@ -1381,7 +1419,7 @@ async def build_main_agent(
                 "provider_request 必须是 ProviderRequest 类型。"
             )
             if req.conversation:
-                req.contexts = json.loads(req.conversation.history)
+                req.contexts = strip_images_from_history_messages(json.loads(req.conversation.history))
         else:
             req = ProviderRequest()
             req.prompt = ""
@@ -1514,7 +1552,7 @@ async def build_main_agent(
 
             conversation = await _get_session_conv(event, plugin_context)
             req.conversation = conversation
-            req.contexts = json.loads(conversation.history)
+            req.contexts = strip_images_from_history_messages(json.loads(conversation.history))
             event.set_extra("provider_request", req)
 
     if isinstance(req.contexts, str):
@@ -1530,7 +1568,10 @@ async def build_main_agent(
                 )
             )
         )
-    req.image_urls = normalize_and_dedupe_strings(req.image_urls)
+    req.image_urls = _finalize_request_image_urls(
+        normalize_and_dedupe_strings(req.image_urls),
+        max_total=max(int(getattr(config, "max_quoted_fallback_images", 4) or 4), 4),
+    )
     req.audio_urls = normalize_and_dedupe_strings(req.audio_urls)
 
     if config.file_extract_enabled:

astrbot/core/provider/sources/openai_source.py

@@ -273,21 +273,37 @@ async def _transform_content_part(self, part: dict) -> dict:
         if part.get("type") == "image_url":
             url, image_detail = self._extract_image_part_info(part)
             if not url:
-                return part
+                return {
+                    "type": "text",
+                    "text": "[Image unavailable]",
+                }
 
             try:
                 resolved_part = await self._resolve_image_part(
                     url, image_detail=image_detail
                 )
             except Exception as exc:
                 logger.warning(
-                    "图片 %s 预处理失败，将保留原始内容。错误: %s",
+                    "Image preprocess failed; dropping image content. url=%s err=%s",
                     url,
                     exc,
                 )
-                return part
+                return {
+                    "type": "text",
+                    "text": "[Image unavailable]",
+                }
 
-            return resolved_part or part
+            if resolved_part:
+                return resolved_part
+
+            logger.warning(
+                "Image preprocess returned empty; dropping image content. url=%s",
+                url,
+            )
+            return {
+                "type": "text",
+                "text": "[Image unavailable]",
+            }
 
         if part.get("type") == "audio_url":
             audio_ref = self._extract_audio_part_info(part)

tests/test_image_history_strip.py

@@ -0,0 +1,57 @@
+from astrbot.core.agent.message import (
+    ImageURLPart,
+    Message,
+    TextPart,
+    bind_checkpoint_messages,
+    dump_messages_with_checkpoints,
+    strip_images_from_history_messages,
+)
+from astrbot.core.astr_main_agent import _finalize_request_image_urls
+
+
+def test_dump_messages_strips_image_parts():
+    msg = Message(
+        role="user",
+        content=[
+            TextPart(text="see this"),
+            ImageURLPart(image_url=ImageURLPart.ImageURL(url="data:image/png;base64,AAAA")),
+            ImageURLPart(image_url=ImageURLPart.ImageURL(url="https://example.com/x.jpg")),
+        ],
+    )
+    dumped = dump_messages_with_checkpoints([msg])
+    content = dumped[0]["content"]
+    assert all(part.get("type") != "image_url" for part in content)
+    assert sum(1 for part in content if part.get("text") == "[Image]") == 1
+
+
+def test_bind_checkpoint_messages_strips_polluted_history():
+    hist = [
+        {
+            "role": "user",
+            "content": [
+                {"type": "text", "text": "old"},
+                {"type": "image_url", "image_url": {"url": "data:image/jpeg;base64,BBB"}},
+                {"type": "image_url", "image_url": {"url": "https://example.com/a.jpg"}},
+            ],
+        }
+    ]
+    bound = bind_checkpoint_messages(hist)
+    types = [getattr(part, "type", None) for part in bound[0].content]
+    assert "image_url" not in types
+    stripped = strip_images_from_history_messages(hist)
+    assert all(part.get("type") != "image_url" for part in stripped[0]["content"])
+
+
+def test_finalize_request_image_urls_prefers_local_and_limits():
+    out = _finalize_request_image_urls(
+        [
+            "https://example.com/a",
+            r"C:\\tmp\\a.jpg",
+            "https://example.com/a",
+            r"C:\\tmp\\b.jpg",
+            "https://example.com/c.jpg",
+        ],
+        max_total=2,
+    )
+    assert len(out) == 2
+    assert all(not u.startswith("http") for u in out)