power-loop 6.3.0__tar.gz → 6.5.0__tar.gz

{power_loop-6.3.0 → power_loop-6.5.0}/PKG-INFO

@@ -1,6 +1,6 @@
 Metadata-Version: 2.4
 Name: power-loop
-Version: 6.3.0
+Version: 6.5.0
 Summary: Embeddable agent execution kernel — LLM loop, hooks, events, tools, dynamic sub-agents.
 Author-email: zhangran <zhangran24@126.com>
 License: MIT

{power_loop-6.3.0 → power_loop-6.5.0}/power_loop/__init__.py

@@ -15,7 +15,7 @@ Stability tiers
 无版本承诺，可随时变更或删除。
 """
 
-__version__ = "6.3.0"
+__version__ = "6.5.0"
 
 # Public LLM contract (SDK-free) re-exported so callers (e.g. writing llm.* hooks or
 # a custom LLMService) don't reach into the internal vendored transport package (H3.4).

{power_loop-6.3.0 → power_loop-6.5.0}/power_loop/_vendor/llm_client/multimodal.py

@@ -39,8 +39,25 @@ class PreparedAttachment:
     strategy: str = "text"
 
 
+#: 扩展名 → MIME 的**内置**兜底表。不能只靠 `mimetypes`：它读的是系统的 mime.types，
+#: 不同镜像装的不一样。真实事故（DeepTalk conv-201）：宿主 Python 认得 `.webp`，
+#: 生产容器里 `guess_type(".webp")` 返回 None → 落到 application/octet-stream →
+#: 渲染层判定「这不是图片」抛 ModelCapabilityError → 重试耗尽 → 整个 run 降级终止。
+#: agent 刚把配图处理成 webp、正要看一眼，会话就死在这里。本地全绿、生产炸掉的典型。
+_EXT_MIME_FALLBACK = {
+    ".webp": "image/webp", ".png": "image/png", ".jpg": "image/jpeg", ".jpeg": "image/jpeg",
+    ".gif": "image/gif", ".bmp": "image/bmp", ".tif": "image/tiff", ".tiff": "image/tiff",
+    ".avif": "image/avif", ".heic": "image/heic", ".heif": "image/heif", ".svg": "image/svg+xml",
+    ".pdf": "application/pdf",
+}
+
+
 def _guess_mime_type(path: Path) -> str:
+    """文件的 MIME。系统表优先，认不出来再查内置表 —— 顺序不能反：
+    系统表更全，内置表只补它在某些镜像里缺的那几个。"""
     mime_type, _ = mimetypes.guess_type(str(path))
+    if not mime_type:
+        mime_type = _EXT_MIME_FALLBACK.get(path.suffix.lower())
     return mime_type or "application/octet-stream"
 
 
@@ -152,6 +169,10 @@ def _downscale_to_data_url(path: Path, mime_type: str, max_edge: int | None) ->
             resized.save(buf, format="JPEG", quality=85, optimize=True)
         payload = base64.b64encode(buf.getvalue()).decode("ascii")
         return f"data:image/jpeg;base64,{payload}"
+    except OSError:
+        # 文件读不到不是「缩放失败」——报 "sending it at original size" 会把调用方引到
+        # 图像解码上，而真正的毛病在路径。交给上层按「读不到」降级。
+        raise
     except Exception:  # noqa: BLE001 — a decode failure must not lose the image entirely
         logger.warning("could not downscale %s; sending it at original size", path.name,
                        exc_info=True)
@@ -171,12 +192,27 @@ def _render_image_attachment(ref: AttachmentRef, path: Path, capabilities: Model
     调用方想要「发图给看不了图的模型就报错」，用 ``capabilities.require_image_input()`` 自查。
     """
     if capabilities.supports_image_input is True:
-        part = {
-            "type": "image_url",
-            "image_url": {
-                "url": _downscale_to_data_url(path, ref.mime_type, capabilities.max_image_edge)
-            },
-        }
+        try:
+            url = _downscale_to_data_url(path, ref.mime_type, capabilities.max_image_edge)
+        except OSError as exc:
+            # 文件读不到（路径解析错、被清理、权限）——**降级，不抛**。抛出去的代价与
+            # 能力不匹配那一路完全一样：一次 render 同时渲染历史与本轮，一个读不到的
+            # 附件会让每一次 send 都失败，重试耗尽后整个 run 终止。DeepTalk conv-198
+            # 真实发生过：两张刚生成的图因相对路径被按进程 cwd 解析而找不到，agent 就此
+            # 停在半路，图既没被看到也没发出去。占位文本同样必须说清「你没有看到」。
+            logger.warning("image %s could not be read (%s); surfaced as a text placeholder",
+                           ref.filename, exc)
+            text = (
+                f"[图片 {describe_attachment_ref(ref)}——**这个文件读不到，你没有看到它的内容**。"
+                "不要凭空描述这张图；需要看就按上面的坐标用视觉工具重新取它。]"
+            )
+            return PreparedAttachment(
+                ref=ref,
+                text_fallback=text,
+                rendered_parts=({"type": "text", "text": text},),
+                strategy="image-unreadable",
+            )
+        part = {"type": "image_url", "image_url": {"url": url}}
         return PreparedAttachment(ref=ref, rendered_parts=(part,), strategy="native-image")
 
     logger.warning(
@@ -246,11 +282,24 @@ def prepare_attachment(ref_payload: dict[str, Any], capabilities: ModelCapabilit
     if ref.kind == "pdf":
         return _render_pdf_attachment(ref, path, capabilities)
 
-    raise ModelCapabilityError(
-        f"Cannot send attachment {ref.filename!r}: unsupported type {ref.mime_type!r}. "
-        "This library renders images (declared models only) and text-extractable PDFs; "
-        "anything else must be converted by the caller. It is not summarised into a "
-        "placeholder line — a placeholder reads to the model as if the file had been read."
+    # 认不出的类型：降级成**明说没读到**的占位，不抛。
+    #
+    # 原先这里抛 ModelCapabilityError，理由是「占位符读起来像文件已经被读过」——那个理由
+    # 对含糊的占位成立，对下面这句不成立：它明确写着模型没有拿到内容，并给出回取坐标。
+    # 而抛出去的代价是整个会话不可用：一次 render 同时渲染历史与本轮，历史里躺着一个认不出
+    # 类型的附件，之后**每一次** send 都失败。DeepTalk conv-201 就是这么断的——容器里
+    # `.webp` 猜不出 MIME（见 _EXT_MIME_FALLBACK），agent 处理完配图正要看一眼，会话就死了。
+    logger.warning("attachment %s has unsupported type %r; surfaced as a text placeholder",
+                   ref.filename, ref.mime_type)
+    text = (
+        f"[附件 {describe_attachment_ref(ref)}（类型 {ref.mime_type}）——**这个类型发不进上下文，"
+        "你没有读到它的内容**。不要凭空描述它；需要的话先把它转成图片或可提取文本的 PDF。]"
+    )
+    return PreparedAttachment(
+        ref=ref,
+        text_fallback=text,
+        rendered_parts=({"type": "text", "text": text},),
+        strategy="unsupported-type",
     )
 
 

{power_loop-6.3.0 → power_loop-6.5.0}/power_loop/runtime/image_recall.py

@@ -28,6 +28,7 @@ Pass ``durable=False`` for the look-once case (request-only, never stored).
 from __future__ import annotations
 
 import threading
+from collections.abc import Sequence
 from typing import Any
 
 LoopMessage = dict[str, Any]
@@ -77,6 +78,43 @@ def queue_image_for_next_round(
     return True
 
 
+def queue_images_for_next_round(
+    session_id: str | None,
+    images: Sequence[tuple[str, str]],
+    *,
+    note: str = "",
+    durable: bool = True,
+) -> int:
+    """Queue a BATCH as ONE user message: a single note + one attachment block per image.
+
+    Queuing them one by one produces N separate user turns, each repeating the note — three
+    screenshots asked about with one question became three copies of that question in the
+    transcript. A batch is one turn, which is also what the provider APIs expect.
+
+    ``images`` is ``[(path, ref)]``. Returns how many were accepted (0 = nothing queued, so the
+    caller must say so rather than claim the pictures were delivered).
+    """
+    if not session_id:
+        return 0
+    blocks: list[dict[str, Any]] = []
+    if note:
+        blocks.append({"type": "text", "text": note})
+    accepted = 0
+    for path, ref in images:
+        if not path:
+            continue
+        blocks.append({"type": "attachment", "attachment": _attachment_ref(path, ref=ref)})
+        accepted += 1
+    if not accepted:
+        return 0
+    with _lock:
+        queue = _pending.setdefault(session_id, [])
+        if len(queue) >= MAX_PENDING_PER_SESSION:
+            return 0
+        queue.append({"role": "user", "content": blocks, "__durable__": durable})
+    return accepted
+
+
 def drain_queued_images(session_id: str | None) -> tuple[list[LoopMessage], list[LoopMessage]]:
     """Take everything queued for this session (and clear it).
 

{power_loop-6.3.0 → power_loop-6.5.0}/power_loop.egg-info/PKG-INFO

@@ -1,6 +1,6 @@
 Metadata-Version: 2.4
 Name: power-loop
-Version: 6.3.0
+Version: 6.5.0
 Summary: Embeddable agent execution kernel — LLM loop, hooks, events, tools, dynamic sub-agents.
 Author-email: zhangran <zhangran24@126.com>
 License: MIT