落地私聊图片询价：识图当口播复用三线文字流程，图上航线三字码保持原样，避免被改写成中文地名。

Co-authored-by: Cursor <cursoragent@cursor.com>

inquiry-agent/agent/llm/mode_vision.py

@@ -1,30 +1,166 @@
 """
-千问视觉：读图 mode 壳。
+千问视觉：只读图上可见原文。
+
+本文件职责：把图片交给 qwen3-vl-plus，抄出看得见的文字。
+属于 llm 包。依赖：LlmHttpClient（生产）；单测注入 chat_fn。
+禁止：填询价字段 JSON、算最终缺项、发明港口/重量/日期；禁止在回调线程长等（由 Worker 槽调用）。
 """
 
 from __future__ import annotations
 
-from typing import Any
+import logging
+from typing import Any, Optional
 
 from agent.llm.modes import LlmMode
 
+logger = logging.getLogger(__name__)
+
 MODE = LlmMode.VISION
 
+VISION_SYSTEM = (
+    "只读出图片上能看见的文字和数字，原样抄写，一行不漏。"
+    "英文、机场三字码、航线（如 SZX-CGK、SZX/CGK）必须保持原样，禁止翻译成中文城市名。"
+    "看不清的字不要猜。不要补询价缺项，不要发明港口、重量或日期。"
+    "只有整张图完全没有文字时才输出空；只要看得见字就必须抄出来。"
+)
+
+
+def sniff_image_mime(raw: bytes) -> str:
+    """
+    按文件头认图片类型。企微截图经常是 PNG，不能一律当 JPEG 送给千问。
+    """
+    blob = raw or b""
+    if len(blob) < 12:
+        return ""
+    if blob.startswith(b"\x89PNG\r\n\x1a\n"):
+        return "image/png"
+    if blob.startswith(b"\xff\xd8\xff"):
+        return "image/jpeg"
+    if blob[:4] == b"RIFF" and blob[8:12] == b"WEBP":
+        return "image/webp"
+    if blob.startswith(b"GIF87a") or blob.startswith(b"GIF89a"):
+        return "image/gif"
+    if blob[4:8] == b"ftyp":
+        return "image/heic"
+    if blob.startswith(b"BM"):
+        return "image/bmp"
+    return ""
+
 
 def mode_meta() -> dict[str, Any]:
     return {
         "mode": MODE.value,
         "provider": "qwen",
         "purpose": "vision",
-        "implemented": False,
+        "implemented": True,
+    }
+
+
+def _pack_text(text: str, *, stub: bool = False) -> dict[str, Any]:
+    body = (text or "").strip()
+    return {
+        "ok": True,
+        "stub": stub,
+        "mode": MODE.value,
+        "text": body,
+        "empty": not bool(body),
+        "error": "",
     }
 
 
 def invoke_shell(*, payload: dict[str, Any]) -> dict[str, Any]:
+    """
+    无网/单测入口：assistant_text 当作图上可见原文。
+
+    没有该键时保持失败码，避免假装已经读图。
+    """
+    if "assistant_text" in (payload or {}):
+        packed = _pack_text(str(payload.get("assistant_text") or ""), stub=True)
+        return packed
     return {
         "ok": False,
         "stub": True,
         "mode": MODE.value,
+        "text": "",
+        "empty": True,
         "error": "vision_shell_not_implemented",
-        "echo_keys": sorted(payload.keys()),
+        "echo_keys": sorted((payload or {}).keys()),
     }
+
+
+def _vision_messages(*, image_b64: str, mime: str, image_url: str) -> list[dict[str, Any]]:
+    """拼兼容 OpenAI 的图文消息。优先 data URL，其次 http 图链。"""
+    url = (image_url or "").strip()
+    if not url and image_b64:
+        kind = (mime or "image/jpeg").strip() or "image/jpeg"
+        url = f"data:{kind};base64,{image_b64}"
+    user_content: list[dict[str, Any]] = []
+    if url:
+        user_content.append({"type": "image_url", "image_url": {"url": url}})
+    user_content.append({"type": "text", "text": "请只抄写图上可见文字。"})
+    return [
+        {"role": "system", "content": VISION_SYSTEM},
+        {"role": "user", "content": user_content},
+    ]
+
+
+def invoke_vision(
+    *,
+    payload: dict[str, Any],
+    chat_fn: Optional[Any] = None,
+    allow_network: bool = False,
+) -> dict[str, Any]:
+    """
+    读一张图，返回可见原文。
+
+    chat_fn：单测注入，签名同 LlmHttpClient.chat。
+    allow_network=False 且无 chat_fn：走 invoke_shell，不打真网。
+    """
+    data = dict(payload or {})
+    image_b64 = str(data.get("image_b64") or "")
+    image_url = str(data.get("image_url") or "")
+    mime = str(data.get("mime") or "").strip() or "image/jpeg"
+    if chat_fn is None and not allow_network:
+        return invoke_shell(payload=data)
+    if not image_b64 and not image_url:
+        logger.info("vision 拒绝：没有图")
+        return {
+            "ok": False,
+            "stub": False,
+            "mode": MODE.value,
+            "text": "",
+            "empty": True,
+            "error": "no_image",
+        }
+    if chat_fn is None:
+        from agent.llm.http_client import LlmHttpClient
+
+        client = LlmHttpClient.from_settings()
+        chat_fn = client.chat
+    messages = _vision_messages(
+        image_b64=image_b64,
+        mime=mime,
+        image_url=image_url,
+    )
+    result = chat_fn(mode=MODE, messages=messages)
+    if not getattr(result, "ok", False):
+        err = str(getattr(result, "error", "") or "vision_failed")
+        logger.info("vision 失败 err=%s", err)
+        return {
+            "ok": False,
+            "stub": False,
+            "mode": MODE.value,
+            "text": "",
+            "empty": True,
+            "error": err,
+        }
+    packed = _pack_text(str(getattr(result, "content", "") or ""))
+    packed["stub"] = False
+    logger.info(
+        "vision 读完 ok=%s empty=%s text_len=%s mime=%s",
+        True,
+        packed.get("empty"),
+        len(str(packed.get("text") or "")),
+        mime,
+    )
+    return packed