Quiet expected preview failures; rename pyvips to vips

Worker logs expected PreviewError failures as warnings; tracebacks are
reserved for unanticipated internal errors. pdf read errors report
backend "pdf" (vips not yet reached); the stage field is dropped as the
backend name now encodes the failing pipeline step. Backend label
"pyvips" renamed to "vips" everywhere.

mediapreview/__init__.py

@@ -17,12 +17,14 @@ from mediapreview.backends import (
     process_video,
 )
 from mediapreview.cache import CachedPreview, PreviewCache
+from mediapreview.exceptions import PreviewError
 from mediapreview.formats import is_previewable_path
 from mediapreview.protocol import PreviewRequest, PreviewResponse
 
 __all__ = [
     "CachedPreview",
     "PreviewCache",
+    "PreviewError",
     "PreviewRequest",
     "PreviewResponse",
     "dispatch",

mediapreview/__main__.py

@@ -31,10 +31,19 @@ from pathlib import Path
 from docopt import docopt
 
 from mediapreview.backends import dispatch
+from mediapreview.exceptions import PreviewError
+from mediapreview.util.logformat import EmojiFormatter
+
+
+def _configure_logging() -> None:
+    handler = logging.StreamHandler(sys.stderr)
+    handler.setFormatter(EmojiFormatter())
+    logging.basicConfig(level=logging.INFO, handlers=[handler])
+    # pyvips is chatty at INFO ("threadpool completed ..." per operation).
+    logging.getLogger("pyvips").setLevel(logging.WARNING)
 
 
 def _oosetup(name: str, port: int) -> None:
-    logging.basicConfig(level=logging.INFO, format="%(message)s")
     try:
         # Lazy import: keeps the base CLI free of office-extra concerns.
         from mediapreview.office import setup_docker  # noqa: PLC0415
@@ -58,12 +67,16 @@ def _preview(args: dict) -> None:
         sys.stderr.write(f"error: no such file: {path}\n")
         sys.exit(2)
 
-    result, resp = dispatch(
+    try:
-        path,
+        result, resp = dispatch(
-        quality=int(args["-q"]),
+            path,
-        maxsize=int(args["--maxsize"]),
+            quality=int(args["-q"]),
-        maxzoom=float(args["--maxzoom"]),
+            maxsize=int(args["--maxsize"]),
-    )
+            maxzoom=float(args["--maxzoom"]),
+        )
+    except PreviewError as e:
+        sys.stderr.write(f"error: {e}\n")
+        sys.exit(1)
     if not resp.ok or result is None:
         sys.stderr.write(f"error: {resp.error or 'preview failed'}\n")
         if resp.stderr:
@@ -84,6 +97,7 @@ def _preview(args: dict) -> None:
 
 
 def main() -> None:
+    _configure_logging()
     # docopt matches usage patterns in order, so `oosetup` would be swallowed
     # by the <path> pattern if it came second. Dispatch it before parsing;
     # the main help above still documents both modes.

mediapreview/backends/__init__.py

@@ -15,8 +15,8 @@ from mediapreview.backends.image import (
 )
 from mediapreview.backends.pdf import process_pdf
 from mediapreview.backends.video import process_video
+from mediapreview.exceptions import PreviewError, backend_error
 from mediapreview.formats import DOC_PREVIEW_SUFFIXES
-from mediapreview.protocol import PreviewResponse
 
 __all__ = [
     "dispatch",
@@ -34,7 +34,7 @@ def dispatch(path, quality, maxsize, maxzoom, data=None):
     backend = "unknown"
     try:
         if data:
-            backend = "pyvips"
+            backend = "vips"
             return process_image_buffer(
                 data, quality=quality, maxsize=maxsize, maxzoom=maxzoom
             )
@@ -47,15 +47,19 @@ def dispatch(path, quality, maxsize, maxzoom, data=None):
             backend = "video"
             return process_video(path, quality=quality, maxsize=maxsize)
         if mime_type and mime_type.startswith("image/"):
-            backend = "pyvips"
+            backend = "vips"
             return process_image(path, quality=quality, maxsize=maxsize)
+    except PreviewError:
+        # Already structured (e.g. a failing stage of a combined pipeline
+        # like pdf+vips) — keep the original backend identity.
+        raise
     except ValueError as e:
-        return None, PreviewResponse(ok=False, backend=backend, error=str(e))
+        raise backend_error(backend, str(e)) from e
     except ImportError as e:
         # Missing optional extra — expected, so a plain message, no traceback.
         logger.error("Preview dispatch failed for %s: %s", path, e)  # noqa: TRY400
-        return None, PreviewResponse(ok=False, backend=backend, error=str(e))
+        raise backend_error(backend, str(e)) from e
     except Exception as e:
         logger.exception("Preview dispatch failed for %s", path)
-        return None, PreviewResponse(ok=False, backend=backend, error=str(e))
+        raise backend_error(backend, str(e)) from e
-    return None, PreviewResponse(ok=False, backend=backend, error="preview unsupported")
+    raise backend_error(backend, "preview unsupported")

mediapreview/backends/image.py

@@ -148,7 +148,7 @@ def process_image_pyvips(path, *, maxsize, quality):
         )
     except pyvips.error.Error as e:
         raise ValueError(f"cannot decode image: {e}") from e
-    backend = "pyvips"
+    backend = "vips"
     t_end = perf_counter()
 
     return ret, PreviewResponse(
@@ -181,7 +181,7 @@ def process_image_buffer(data: bytes, *, quality, maxsize, maxzoom):
     return ret, PreviewResponse(
         ok=True,
         mime="image/avif",
-        backend="pyvips",
+        backend="vips",
         timings=[round((t_end - t_start) * 1000, 1)],
         width=orig_w,
         height=orig_h,

mediapreview/backends/pdf.py

@@ -1,10 +1,11 @@
-"""PDF/XPS/EPUB preview conversion via PyMuPDF + pyvips."""
+"""PDF/XPS/EPUB preview conversion via PyMuPDF + vips."""
 
 from time import perf_counter
 
 import pyvips
 
 from mediapreview.backends.image import AVIF_FAST_EFFORT
+from mediapreview.exceptions import backend_error
 from mediapreview.protocol import PreviewResponse
 
 try:
@@ -12,6 +13,8 @@ try:
 except ImportError:  # pragma: no cover - optional pdf extra
     pymupdf = None
 
+BACKEND = "pdf+vips"
+
 
 def process_pdf(path, *, maxsize, maxzoom, quality, page_number=0):
     if pymupdf is None:
@@ -19,26 +22,31 @@ def process_pdf(path, *, maxsize, maxzoom, quality, page_number=0):
             "PDF previews require the 'pdf' extra: pip install mediapreview[pdf]"
         )
     t_load_start = perf_counter()
-    with pymupdf.open(path) as pdf:
+    try:
-        page = pdf.load_page(page_number)
+        with pymupdf.open(path) as pdf:
-        w, h = page.rect[2:4]
+            page = pdf.load_page(page_number)
-        zoom = min(maxsize / w, maxsize / h, maxzoom)
+            w, h = page.rect[2:4]
-        mat = pymupdf.Matrix(zoom, zoom)
+            zoom = min(maxsize / w, maxsize / h, maxzoom)
-        pix = page.get_pixmap(matrix=mat)
+            mat = pymupdf.Matrix(zoom, zoom)
-        t_load_end = perf_counter()
+            pix = page.get_pixmap(matrix=mat)
+            samples, width, height, n = pix.samples_mv, pix.width, pix.height, pix.n
+    except Exception as e:
+        # vips was never reached — this is a plain pdf error.
+        raise backend_error("pdf", str(e)) from e
+    t_load_end = perf_counter()
 
-        t_save_start = perf_counter()
+    t_save_start = perf_counter()
-        img = pyvips.Image.new_from_memory(
+    try:
-            pix.samples_mv, pix.width, pix.height, pix.n, "uchar"
+        img = pyvips.Image.new_from_memory(samples, width, height, n, "uchar")
-        )
+        ret = img.write_to_buffer(".avif", Q=quality, effort=AVIF_FAST_EFFORT, keep="none")
-    ret = img.write_to_buffer(".avif", Q=quality, effort=AVIF_FAST_EFFORT, keep="none")
+    except Exception as e:
-    backend = "pdf+pyvips"
+        raise backend_error(BACKEND, str(e)) from e
     t_save_end = perf_counter()
 
     return ret, PreviewResponse(
         ok=True,
         mime="image/avif",
-        backend=backend,
+        backend=BACKEND,
         timings=[
             round((t_load_end - t_load_start) * 1000, 1),
             round((t_save_end - t_save_start) * 1000, 1),

mediapreview/backends/video.py

@@ -15,8 +15,8 @@ try:
     import av
     import numpy as np
 except ImportError:  # pragma: no cover - optional video extra
-    av = None
+    av = None  # type: ignore[assignment]
-    np = None
+    np = None  # type: ignore[assignment]
 
 
 def _rotate_frame_yuv(frame, k):
@@ -78,21 +78,22 @@ def process_video(path, *, maxsize, quality):
         istream.codec_context.skip_frame = "NONKEY"
         icontainer.seek((icontainer.duration or 0) // 8)
         for frame in icontainer.decode(istream):
-            if frame.dts is not None:
+            if frame.dts is not None or frame.pts is not None:
                 break
         else:
             raise RuntimeError("No frames found in video")
 
-        # Resize frame to thumbnail size
+        # Resize frame to thumbnail size. Keep dimensions even for planar
+        # 4:2:0 chroma subsampling, which _rotate_frame_yuv expects.
         # Capture display dimensions before resize (accounting for rotation)
         disp_w = frame.width
         disp_h = frame.height
-        if frame.rotation in (90, 270):
+        if abs(frame.rotation) in (90, 270):
             disp_w, disp_h = disp_h, disp_w
         if frame.width > maxsize or frame.height > maxsize:
             scale_factor = min(maxsize / frame.width, maxsize / frame.height)
-            new_width = int(frame.width * scale_factor)
+            new_width = int(frame.width * scale_factor) // 2 * 2
-            new_height = int(frame.height * scale_factor)
+            new_height = int(frame.height * scale_factor) // 4 * 4
             frame = frame.reformat(width=new_width, height=new_height)
 
         # Apply display-matrix rotation if present

mediapreview/exceptions.py

@@ -0,0 +1,175 @@
+"""Structured preview exceptions.
+
+Preview failures are represented by ``PreviewError`` and a small number of
+subclasses, carrying their fields directly: ``short`` (concise, for UI with
+limited space — the backend name is usually printed in front of it, so it is
+left out), the exception message itself (for logs), ``backend``, and
+optional subclass-specific metadata such as ``code`` or ``timeout_seconds``
+for callers that wish to do their own processing.
+
+The worker pool ships exceptions between processes with pickle (same
+trust domain — the pool unpickles only data from its own workers), so any
+exception arrives intact on the caller side, no per-class serialization
+machinery needed. Because the initializers are keyword-heavy, pickling is
+routed through ``__dict__`` via ``PreviewError.__reduce__``.
+
+The hierarchy is intentionally small:
+
+- ``OnlyOfficeError`` covers all OnlyOffice failures; optional fields
+  (``code``, ``status``, ``url``, ``snippet``) describe the specific failure.
+- ``PreviewBackendError`` covers backend conversion failures (ffmpeg, vips,
+  pdf, etc.). Combined pipelines report the failing step in ``backend``
+  (e.g. "pdf" if pdf reading failed before vips was reached, "pdf+vips"
+  for a vips write failure).
+- ``PreviewTimeoutError`` covers timeouts for any backend.
+- ``PreviewCancelledError`` covers cancellations (e.g. pool shutdown).
+
+BaseExceptions such as ``KeyboardInterrupt``, ``SystemExit`` and
+``asyncio.CancelledError`` are never wrapped in these types.
+"""
+
+from __future__ import annotations
+
+
+class PreviewError(Exception):
+    """Base preview exception.
+
+    ``short`` is concise text for UIs with limited space (the backend name
+    is usually printed in front of it); ``str(err)`` is the full
+    human-readable message; ``backend`` identifies the backend
+    (e.g. "onlyoffice", "ffmpeg").
+    """
+
+    def __init__(
+        self,
+        message: str = "preview failed",
+        short: str = "error",
+        *,
+        backend: str | None = None,
+    ):
+        super().__init__(message)
+        self.short = short
+        self.backend = backend
+
+    def __reduce__(self):
+        # Keyword-heavy initializers do not unpickle via args; pass the message
+        # positionally and restore the rest from __dict__.
+        return (type(self), (str(self),), self.__dict__)
+
+
+class OnlyOfficeError(PreviewError):
+    """OnlyOffice conversion failed. Specifics are in the extra fields."""
+
+    def __init__(  # noqa: PLR0913 - metadata fields are independent
+        self,
+        message: str = "OnlyOffice conversion failed",
+        short: str = "error",
+        *,
+        code: str | None = None,
+        status: int | None = None,
+        url: str | None = None,
+        snippet: str | None = None,
+        backend: str | None = "onlyoffice",
+    ):
+        super().__init__(message, short, backend=backend)
+        self.code = code
+        self.status = status
+        self.url = url
+        self.snippet = snippet
+
+
+class PreviewBackendError(PreviewError):
+    """Backend conversion failure (image/video/pdf/etc)."""
+
+
+class PreviewTimeoutError(PreviewError):
+    """Preview conversion exceeded its timeout for a given backend."""
+
+    def __init__(
+        self,
+        message: str = "preview timed out",
+        short: str = "timeout",
+        *,
+        timeout_seconds: float = 0.0,
+        backend: str | None = None,
+    ):
+        super().__init__(message, short, backend=backend)
+        self.timeout_seconds = timeout_seconds
+
+
+class PreviewCancelledError(PreviewError):
+    """Preview was cancelled (e.g. pool shut down)."""
+
+    def __init__(
+        self,
+        message: str = "Preview cancelled (pool closed)",
+        short: str = "cancelled",
+        *,
+        reason: str = "pool closed",
+        backend: str | None = None,
+    ):
+        super().__init__(message, short, backend=backend)
+        self.reason = reason
+
+
+# ---------------------------------------------------------------------------
+# Factory helpers
+# ---------------------------------------------------------------------------
+
+_OO_CODE_ERRORS = {
+    "-8": ("jwt error", "OnlyOffice JWT authentication failed"),
+    "-4": ("input error", "OnlyOffice input error"),
+    "-2": ("timeout error", "OnlyOffice conversion timed out"),
+    "-1": ("unknown error", "OnlyOffice conversion failed with unknown error"),
+}
+
+
+def onlyoffice_error_from_code(code: str | None = None) -> OnlyOfficeError:
+    """Build an OnlyOfficeError from a conversion status code (e.g. "-8")."""
+    if code in _OO_CODE_ERRORS:
+        short, log = _OO_CODE_ERRORS[code]
+    elif code:
+        short, log = f"{code} error", f"OnlyOffice conversion failed: {code}"
+    else:
+        short, log = "unknown error", "OnlyOffice conversion failed with unknown error"
+    return OnlyOfficeError(log, short, code=code)
+
+
+def onlyoffice_unavailable_error(url: str | None = None) -> OnlyOfficeError:
+    log = "OnlyOffice document server not reachable"
+    if url:
+        log = f"{log} at {url}"
+    return OnlyOfficeError(log, "unavailable", url=url)
+
+
+def onlyoffice_http_error(status: int) -> OnlyOfficeError:
+    return OnlyOfficeError(f"OnlyOffice HTTP error: {status}", "http error", status=status)
+
+
+def onlyoffice_no_fileurl_error(snippet: str | None = None) -> OnlyOfficeError:
+    log = "OnlyOffice response did not contain FileUrl"
+    if snippet:
+        log = f"{log}: {snippet}"
+    return OnlyOfficeError(log, "no-fileurl error", snippet=snippet)
+
+
+def backend_error(backend: str, message: str) -> PreviewBackendError:
+    short = message.splitlines()[0][:60]
+    return PreviewBackendError(
+        f"[{backend}] preview failed: {message}",
+        short,
+        backend=backend,
+    )
+
+
+def preview_timeout_error(backend: str, timeout_seconds: float) -> PreviewTimeoutError:
+    return PreviewTimeoutError(
+        f"{backend.capitalize()} preview timed out after {timeout_seconds}s",
+        "timeout",
+        backend=backend,
+        timeout_seconds=timeout_seconds,
+    )
+
+
+def preview_cancelled_error(reason: str = "pool closed") -> PreviewCancelledError:
+    return PreviewCancelledError(f"Preview cancelled ({reason})", "cancelled", reason=reason)

mediapreview/formats.py

@@ -90,5 +90,5 @@ def expected_backend(path: Path) -> str:
     if mime_type and mime_type.startswith("video/"):
         return "video"
     if mime_type and mime_type.startswith("image/"):
-        return "pyvips"
+        return "vips"
     return "preview"

mediapreview/office.py

@@ -30,6 +30,14 @@ from pathlib import Path
 from time import perf_counter
 from urllib.parse import quote
 
+from mediapreview.exceptions import (
+    onlyoffice_error_from_code,
+    onlyoffice_http_error,
+    onlyoffice_no_fileurl_error,
+    onlyoffice_unavailable_error,
+    preview_timeout_error,
+)
+
 try:
     import httpx
     import jwt
@@ -45,6 +53,7 @@ logger = logging.getLogger(__name__)
 
 
 _httpx_client: httpx.AsyncClient | None = None
+_httpx_client_loop: asyncio.AbstractEventLoop | None = None
 
 
 def _get_onlyoffice_url() -> str:
@@ -82,35 +91,27 @@ def _get_callback_host() -> str:
     return "127.0.0.1"
 
 
-def onlyoffice_error_short_text(detail: str) -> str:
+# ---------------------------------------------------------------------------
-    """Short human-readable label for an OnlyOffice failure, for log annotation."""
-    if detail.startswith("OnlyOffice conversion error:"):
-        code = detail.rsplit(":", 1)[-1].strip()
-        return {
-            "-8": "onlyoffice jwt error",
-            "-4": "onlyoffice input error",
-            "-2": "onlyoffice timeout error",
-            "-1": "onlyoffice unknown error",
-        }.get(code, f"onlyoffice {code} error")
-    if "OnlyOffice response did not contain FileUrl" in detail:
-        return "onlyoffice no-fileurl error"
-    return "onlyoffice error"
-
-
 # ---------------------------------------------------------------------------
 # Async HTTP client
 # ---------------------------------------------------------------------------
 
 
 def get_httpx_client() -> httpx.AsyncClient:
-    """Return the shared async HTTP client for OnlyOffice requests."""
+    """Return the shared async HTTP client for OnlyOffice requests.
+
+    The client is recreated if the running event loop changes, because an
+    ``httpx.AsyncClient`` is bound to the loop that created it.
+    """
     if httpx is None:
         raise ImportError(
             "OnlyOffice integration requires the 'office' extra: pip install mediapreview[office]"
         )
-    global _httpx_client
+    global _httpx_client, _httpx_client_loop
-    if _httpx_client is None:
+    current_loop = asyncio.get_running_loop()
+    if _httpx_client is None or _httpx_client_loop is not current_loop:
         _httpx_client = httpx.AsyncClient()
+        _httpx_client_loop = current_loop
     return _httpx_client
 
 
@@ -330,26 +331,34 @@ async def convert_to_png_async(file_path: Path, request_timeout: float = 5.0) ->
             headers["Authorization"] = token
 
         t_start = perf_counter()
-        response = await client.post(
+        try:
-            convert_url,
+            response = await client.post(
-            content=json.dumps(payload).encode(),
+                convert_url,
-            headers=headers,
+                content=json.dumps(payload).encode(),
-            timeout=request_timeout,
+                headers=headers,
-        )
+                timeout=request_timeout,
-        response.raise_for_status()
+            )
+            response.raise_for_status()
+        except httpx.TimeoutException as e:
+            raise preview_timeout_error("onlyoffice", request_timeout) from e
+        except httpx.HTTPStatusError as e:
+            raise onlyoffice_http_error(e.response.status_code) from e
+        except httpx.RequestError as e:
+            raise onlyoffice_unavailable_error(_get_onlyoffice_url()) from e
         body = response.content
         t_end = perf_counter()
 
         # Parse XML response
         text = body.decode("utf-8", errors="replace")
         if "<Error>" in text:
-            code = "unknown"
+            code = None
-            if "<Error>" in text and "</Error>" in text:
+            if "</Error>" in text:
                 code = text.split("<Error>")[1].split("</Error>")[0]
-            raise RuntimeError(f"OnlyOffice conversion error: {code}")
+            raise onlyoffice_error_from_code(code)
 
         if "<FileUrl>" not in text:
-            raise RuntimeError("OnlyOffice response did not contain FileUrl")
+            snippet = text if len(text) <= 200 else text[:200] + "..."
+            raise onlyoffice_no_fileurl_error(snippet)
 
         file_url = text.split("<FileUrl>")[1].split("</FileUrl>")[0]
         file_url = file_url.replace("&amp;", "&")
@@ -357,8 +366,17 @@ async def convert_to_png_async(file_path: Path, request_timeout: float = 5.0) ->
         logger.debug("OnlyOffice converted in %.2fs: %s", t_end - t_start, file_url)
 
         # Download converted PNG
-        png_response = await client.get(file_url, timeout=request_timeout)
+        try:
-        png_response.raise_for_status()
+            png_response = await client.get(file_url, timeout=request_timeout)
+            png_response.raise_for_status()
+        except httpx.TimeoutException as e:
+            raise preview_timeout_error("onlyoffice", request_timeout) from e
+        except httpx.HTTPStatusError as e:
+            raise onlyoffice_http_error(e.response.status_code) from e
+        except httpx.RequestError as e:
+            # The converted file lives on the OO server, so a request failure here
+            # usually means OO itself could not be reached after conversion.
+            raise onlyoffice_unavailable_error(_get_onlyoffice_url()) from e
         return png_response.content
     finally:
         await asyncio.to_thread(httpd.shutdown)
@@ -385,7 +403,7 @@ class OOConversionManager:
     async def convert(self, filepath: Path) -> bytes:
         """Return PNG bytes for *filepath*, deduplicating concurrent requests."""
         if not await is_available_cached():
-            raise RuntimeError("OnlyOffice server not reachable")
+            raise onlyoffice_unavailable_error(_get_onlyoffice_url())
         stat = await asyncio.to_thread(filepath.stat)
         key = f"{filepath}:{stat.st_mtime_ns}"
 

mediapreview/pool.py

@@ -4,6 +4,7 @@ import asyncio
 import contextlib
 import logging
 import os
+import pickle
 import signal
 import struct
 import sys
@@ -20,6 +21,13 @@ except ImportError as e:  # pragma: no cover - optional worker extra
         "The worker pool requires the 'worker' extra: pip install mediapreview[worker]"
     ) from e
 
+from mediapreview.exceptions import (
+    PreviewError,
+    PreviewTimeoutError,
+    backend_error,
+    preview_cancelled_error,
+    preview_timeout_error,
+)
 from mediapreview.formats import (
     expected_backend as _expected_preview_backend,
 )
@@ -35,7 +43,6 @@ from mediapreview.protocol import PreviewRequest, PreviewResponse
 __all__ = [
     "PREVIEW_TIMEOUT",
     "PreviewError",
-    "PreviewPoolClosedError",
     "PreviewTimeoutError",
     "generate_office_preview",
     "is_previewable_path",
@@ -47,33 +54,6 @@ __all__ = [
 logger = logging.getLogger(__name__)
 
 
-class PreviewTimeoutError(Exception):
-    """Raised when the preview subprocess exceeds PREVIEW_TIMEOUT."""
-
-    def __init__(self, message: str, *, backend: str | None = None):
-        super().__init__(message)
-        self.backend = backend
-
-
-class PreviewError(Exception):
-    """Raised when the preview subprocess exits with a non-zero status."""
-
-    def __init__(
-        self,
-        message: str,
-        *,
-        stderr: str | None = None,
-        backend: str | None = None,
-    ):
-        super().__init__(message)
-        self.stderr = stderr
-        self.backend = backend
-
-
-class PreviewPoolClosedError(PreviewError):
-    """The preview worker pool has been shut down."""
-
-
 class WorkerChecksumError(Exception):
     """Raised when worker response checksum does not match the packet."""
 
@@ -82,6 +62,20 @@ class WorkerProtocolError(Exception):
     """Raised when worker response packet is malformed."""
 
 
+def _reraise_worker_error(resp: PreviewResponse, payload: bytes) -> None:
+    """Re-raise the exception the worker sent back in the payload, if any.
+
+    The payload is pickled by our own worker processes (same trust domain), so
+    the original exception type arrives intact on the caller side. Falls back
+    to a plain PreviewBackendError built from the error message.
+    """
+    if payload:
+        exc = pickle.loads(payload)  # noqa: S301 - trusted: our own workers
+        if isinstance(exc, Exception):
+            raise exc
+    raise backend_error(resp.backend or "unknown", resp.error or "preview worker error")
+
+
 PREVIEW_TIMEOUT = 10.0  # seconds until preview subprocess is killed
 PREVIEW_WORKERS = max(2, min(8, cpu_count()))
 WORKER_KILL_GRACE = 5.0  # max seconds to wait for a killed worker to be reaped
@@ -139,11 +133,7 @@ class _PreviewWorker:
 
         resp = msgspec.json.decode(meta_raw, type=PreviewResponse)
         if not resp.ok:
-            raise PreviewError(
+            _reraise_worker_error(resp, payload)
-                resp.error or "preview worker error",
-                stderr=resp.stderr,
-                backend=resp.backend,
-            )
         return payload or None, resp
 
     async def kill(self) -> None:
@@ -279,9 +269,9 @@ class _PreviewWorkerPool:
             )
             if not future.done():
                 future.set_exception(
-                    PreviewTimeoutError(
+                    preview_timeout_error(
-                        args[0].name,
+                        _expected_preview_backend(args[0]),
-                        backend=_expected_preview_backend(args[0]),
+                        PREVIEW_TIMEOUT,
                     )
                 )
             return
@@ -305,9 +295,9 @@ class _PreviewWorkerPool:
             )
             if not future.done():
                 future.set_exception(
-                    PreviewTimeoutError(
+                    preview_timeout_error(
-                        filepath.name,
+                        _expected_preview_backend(filepath),
-                        backend=_expected_preview_backend(filepath),
+                        PREVIEW_TIMEOUT,
                     )
                 )
         except WorkerChecksumError:
@@ -319,7 +309,10 @@ class _PreviewWorkerPool:
             )
             if not future.done():
                 future.set_exception(
-                    PreviewError(f"worker checksum mismatch for {filepath.name}")
+                    backend_error(
+                        _expected_preview_backend(filepath),
+                        f"worker checksum mismatch for {filepath.name}",
+                    )
                 )
         except PreviewError as e:
             if not future.done():
@@ -342,14 +335,20 @@ class _PreviewWorkerPool:
             )
             if not future.done():
                 future.set_exception(
-                    PreviewError(f"worker protocol failure for {filepath.name}: {e}")
+                    backend_error(
+                        _expected_preview_backend(filepath),
+                        f"worker protocol failure for {filepath.name}: {e}",
+                    )
                 )
         except Exception:
             replace = True
             logger.exception("Unexpected preview worker error for %s", filepath.name)
             if not future.done():
                 future.set_exception(
-                    PreviewError(f"unexpected worker error for {filepath.name}")
+                    backend_error(
+                        _expected_preview_backend(filepath),
+                        f"unexpected worker error for {filepath.name}",
+                    )
                 )
         finally:
             if replace:
@@ -378,7 +377,7 @@ class _PreviewWorkerPool:
         data: bytes | None = None,
     ):
         if self._closed:
-            raise PreviewPoolClosedError("preview worker pool closed")
+            raise preview_cancelled_error("preview worker pool closed")
         loop = asyncio.get_running_loop()
         future = loop.create_future()
         self._in_flight.add(future)
@@ -410,18 +409,14 @@ class _PreviewWorkerPool:
         # out their timeouts during server shutdown.
         for future in list(self._in_flight):
             if not future.done():
-                future.set_exception(
+                future.set_exception(preview_cancelled_error("pool closed"))
-                    PreviewPoolClosedError("preview worker pool closed")
-                )
         while not self._pending.empty():
             try:
                 _priority, _seq, future, _args = self._pending.get_nowait()
             except asyncio.QueueEmpty:
                 break
             if not future.done():
-                future.set_exception(
+                future.set_exception(preview_cancelled_error("pool closed"))
-                    PreviewPoolClosedError("preview worker pool closed")
-                )
         while not self._idle.empty():
             try:
                 self._idle.get_nowait()
@@ -469,7 +464,11 @@ async def shutdown_preview_workers() -> None:
 async def generate_office_preview(
     filepath: Path, quality: int, maxsize: int, maxzoom: float
 ) -> tuple[bytes | None, PreviewResponse | None]:
-    """Generate a preview for an office file using OnlyOffice + worker AVIF conversion."""
+    """Generate a preview for an office file using OnlyOffice + worker AVIF conversion.
+
+    Raises:
+        OnlyOfficeError: If the OnlyOffice Document Server cannot convert the file.
+    """
     manager = get_oo_manager()
     t_oo_start = perf_counter()
     png_bytes = await manager.convert(filepath)
@@ -478,7 +477,7 @@ async def generate_office_preview(
     img, resp = await run_preview(filepath, quality, maxsize, maxzoom, data=png_bytes)
 
     if resp is not None:
-        resp.backend = "onlyoffice+" + (resp.backend or "pyvips")
+        resp.backend = "onlyoffice+" + (resp.backend or "vips")
         if resp.timings:
             resp.timings = [round((t_oo_end - t_oo_start) * 1000, 1), *resp.timings]
     return img, resp
@@ -490,5 +489,5 @@ async def run_preview(
     """Run preview request in a persistent worker process."""
     await start_preview_workers()
     if _preview_pool is None:
-        raise PreviewPoolClosedError("preview worker pool closed")
+        raise preview_cancelled_error("preview worker pool closed")
     return await _preview_pool.run(filepath, quality, maxsize, maxzoom, data)

mediapreview/protocol.py

@@ -11,7 +11,7 @@ class PreviewRequest(msgspec.Struct, omit_defaults=True):
 
 
 class PreviewResponse(msgspec.Struct, omit_defaults=True):
-    ok: bool
+    ok: bool  # Indicates whether binary payload is the file or a pickled exception
     mime: str | None = None
     backend: str | None = None
     timings: list[float] | None = None

mediapreview/worker.py

@@ -19,6 +19,7 @@ import contextlib
 import io
 import logging
 import os
+import pickle
 import signal
 import struct
 import sys
@@ -37,12 +38,25 @@ except ImportError:  # pragma: no cover - optional worker extra
     sys.exit(1)
 
 from mediapreview.backends import dispatch
+from mediapreview.exceptions import PreviewError
 from mediapreview.protocol import PreviewRequest, PreviewResponse
 from mediapreview.util.logformat import format_level_prefix
 
 logger = logging.getLogger(__name__)
 
 
+def _serialize_exception(e: BaseException) -> bytes:
+    """Pickle an exception for re-raising across the pool wire.
+
+    Falls back to an empty payload (caller uses the plain error message) in
+    the unlikely case the exception cannot be pickled.
+    """
+    try:
+        return pickle.dumps(e)
+    except Exception:
+        return b""
+
+
 class _WorkerLogFormatter(logging.Formatter):
     """Emoji level prefix like the main process, tagged with the worker pid."""
 
@@ -125,21 +139,23 @@ def _run_loop() -> None:
                 result, resp = dispatch(
                     Path(req.path), req.quality, req.maxsize, req.maxzoom, data
                 )
-            if not resp.ok:
-                captured = stderr_capture.getvalue().strip()
-                if captured:
-                    resp = PreviewResponse(
-                        ok=False,
-                        backend=resp.backend,
-                        error=resp.error,
-                        stderr=captured,
-                    )
             _write_response(resp, result or b"")
         except Exception as e:
-            logger.exception("Preview worker error for %s", req.path)
+            # PreviewError is an expected failure (broken input, missing
+            # extra, backend error) — a warning suffices. Tracebacks are
+            # reserved for internal errors we did not anticipate.
+            if isinstance(e, PreviewError):
+                logger.warning("Preview failed for %s: %s", req.path, e)
+            else:
+                logger.exception("Preview worker error for %s", req.path)
             captured = stderr_capture.getvalue().strip()
             _write_response(
-                PreviewResponse(ok=False, error=str(e), stderr=captured or None), b""
+                PreviewResponse(
+                    ok=False,
+                    error=str(e),
+                    stderr=captured or None,
+                ),
+                _serialize_exception(e),
             )
         finally:
             root_logger.removeHandler(handler)

tests/files/generate_fixtures.py

@@ -0,0 +1,180 @@
+#!/usr/bin/env python3
+"""Generate missing test fixtures for the low-level preview tests.
+
+This script only creates files that are not already present in tests/files/.
+Run it whenever you need to rebuild the rotated/HDR video fixtures or the
+sample PDF.
+"""
+
+from __future__ import annotations
+
+import logging
+import subprocess
+import sys
+from pathlib import Path
+
+try:
+    import pymupdf
+except ImportError as e:  # pragma: no cover - optional pdf extra
+    raise ImportError(
+        "Fixture generation requires pymupdf: pip install mediapreview[pdf]"
+    ) from e
+
+ROOT = Path(__file__).resolve().parent
+FILES = ROOT
+logger = logging.getLogger(__name__)
+
+
+def run(cmd: list[str]) -> None:
+    subprocess.run(cmd, check=True, capture_output=True, text=True)
+
+
+def _find_box(data: bytearray, box_type: bytes, start: int = 0, end: int | None = None) -> int:
+    end = end or len(data)
+    i = start
+    while i + 8 <= end:
+        size = int.from_bytes(data[i : i + 4], "big")
+        btype = data[i + 4 : i + 8]
+        if btype == box_type:
+            return i
+        if size == 0:
+            break
+        if size == 1:
+            size = int.from_bytes(data[i + 8 : i + 16], "big")
+        if size < 8:
+            raise ValueError(f"Invalid box size {size} for {btype.decode('ascii', errors='replace')}")
+        i += size
+    return -1
+
+
+def _matrix_90_cw() -> list[int]:
+    return [0, 1 << 16, 0, -(1 << 16), 0, 0, 0, 0, 1 << 16]
+
+
+def _matrix_270_cw() -> list[int]:
+    return [0, -(1 << 16), 0, 1 << 16, 0, 0, 0, 0, 1 << 16]
+
+
+def _patch_tkhd_rotation(in_path: Path, out_path: Path, degrees: int) -> None:
+    """Write a copy of *in_path* with the video track display matrix rotated."""
+    data = bytearray(in_path.read_bytes())
+
+    moov_idx = _find_box(data, b"moov")
+    if moov_idx < 0:
+        raise ValueError("moov box not found")
+    moov_end = moov_idx + int.from_bytes(data[moov_idx : moov_idx + 4], "big")
+
+    trak_idx = _find_box(data, b"trak", moov_idx + 8, moov_end)
+    while trak_idx >= 0:
+        trak_size = int.from_bytes(data[trak_idx : trak_idx + 4], "big")
+        trak_end = trak_idx + trak_size
+        mdia_idx = _find_box(data, b"mdia", trak_idx + 8, trak_end)
+        if mdia_idx < 0:
+            trak_idx = _find_box(data, b"trak", trak_end, moov_end)
+            continue
+        mdia_end = mdia_idx + int.from_bytes(data[mdia_idx : mdia_idx + 4], "big")
+        hdlr_idx = _find_box(data, b"hdlr", mdia_idx + 8, mdia_end)
+        if hdlr_idx < 0:
+            trak_idx = _find_box(data, b"trak", trak_end, moov_end)
+            continue
+        handler_type = data[hdlr_idx + 16 : hdlr_idx + 20]
+        if handler_type == b"vide":
+            tkhd_idx = _find_box(data, b"tkhd", trak_idx + 8, trak_end)
+            if tkhd_idx < 0:
+                raise ValueError("video track has no tkhd")
+            version = data[tkhd_idx + 8]
+            if version != 0:
+                raise ValueError(f"unsupported tkhd version {version}")
+            matrix_offset = tkhd_idx + 48
+            matrix = _matrix_90_cw() if degrees == 90 else _matrix_270_cw()
+            for i, val in enumerate(matrix):
+                data[matrix_offset + i * 4 : matrix_offset + (i + 1) * 4] = val.to_bytes(
+                    4, "big", signed=True
+                )
+            out_path.write_bytes(data)
+            return
+        trak_idx = _find_box(data, b"trak", trak_end, moov_end)
+    raise ValueError("video track not found")
+
+
+def generate_pdf() -> None:
+    pdf_path = FILES / "sample.pdf"
+    if pdf_path.exists():
+        return
+    doc = pymupdf.open()
+    page = doc.new_page(width=595, height=842)
+    page.insert_text((100, 100), "Hello, PDF preview test!")
+    page.draw_rect((100, 150, 400, 250), color=(1, 0, 0), width=2)
+    doc.save(str(pdf_path))
+    doc.close()
+    logger.info("generated %s", pdf_path)
+
+
+def generate_sdr_rotated() -> None:
+    source = FILES / "sample-1mb.mp4"
+    if not source.exists():
+        raise FileNotFoundError(f"Missing source fixture: {source}")
+    for degrees in (90, 270):
+        out = FILES / f"rotated_{degrees}.mp4"
+        if out.exists():
+            continue
+        _patch_tkhd_rotation(source, out, degrees)
+        logger.info("generated %s", out)
+
+
+def generate_hdr_video() -> None:
+    out = FILES / "hdr_video.mp4"
+    if out.exists():
+        return
+    run(
+        [
+            "ffmpeg",
+            "-y",
+            "-hide_banner",
+            "-loglevel",
+            "error",
+            "-nostdin",
+            "-f",
+            "lavfi",
+            "-i",
+            "testsrc=duration=1:size=320x240:rate=30",
+            "-c:v",
+            "libx265",
+            "-preset",
+            "fast",
+            "-crf",
+            "30",
+            "-pix_fmt",
+            "yuv420p10le",
+            "-x265-params",
+            "colorprim=bt2020:transfer=smpte2084:colormatrix=bt2020nc:repeat-headers=1",
+            "-an",
+            str(out),
+        ]
+    )
+    logger.info("generated %s", out)
+
+
+def generate_hdr_rotated() -> None:
+    source = FILES / "hdr_video.mp4"
+    if not source.exists():
+        raise FileNotFoundError(f"Generate hdr_video.mp4 first: {source}")
+    for degrees in (90, 270):
+        out = FILES / f"hdr_rotated_{degrees}.mp4"
+        if out.exists():
+            continue
+        _patch_tkhd_rotation(source, out, degrees)
+        logger.info("generated %s", out)
+
+
+def main() -> int:
+    logging.basicConfig(level=logging.INFO, format="%(message)s")
+    generate_pdf()
+    generate_sdr_rotated()
+    generate_hdr_video()
+    generate_hdr_rotated()
+    return 0
+
+
+if __name__ == "__main__":
+    sys.exit(main())

tests/test_office.py

@@ -0,0 +1,105 @@
+"""Tests for OnlyOffice error handling and structured responses."""
+
+from __future__ import annotations
+
+import pickle
+from pathlib import Path
+
+import pytest
+
+from mediapreview import office
+from mediapreview.exceptions import (
+    OnlyOfficeError,
+    PreviewBackendError,
+    backend_error,
+    onlyoffice_error_from_code,
+    onlyoffice_http_error,
+    onlyoffice_no_fileurl_error,
+    onlyoffice_unavailable_error,
+)
+from mediapreview.pool import generate_office_preview
+
+FILES = Path(__file__).resolve().parent / "files"
+
+
+@pytest.mark.parametrize(
+    ("code", "short", "message"),
+    [
+        ("-8", "jwt error", "OnlyOffice JWT authentication failed"),
+        ("-4", "input error", "OnlyOffice input error"),
+        ("-2", "timeout error", "OnlyOffice conversion timed out"),
+        ("-1", "unknown error", "OnlyOffice conversion failed with unknown error"),
+        ("-99", "-99 error", "OnlyOffice conversion failed: -99"),
+        (None, "unknown error", "OnlyOffice conversion failed with unknown error"),
+    ],
+)
+def test_onlyoffice_error_from_code(code, short, message):
+    """Conversion status codes map to clean labels and messages."""
+    err = onlyoffice_error_from_code(code)
+    assert str(err) == message
+    assert err.short == short
+    assert err.code == code
+    assert err.backend == "onlyoffice"
+    assert isinstance(err, OnlyOfficeError)
+
+
+def test_onlyoffice_http_error():
+    err = onlyoffice_http_error(502)
+    assert str(err) == "OnlyOffice HTTP error: 502"
+    assert err.status == 502
+    assert err.short == "http error"
+
+
+def test_onlyoffice_unavailable_error():
+    err = onlyoffice_unavailable_error("http://localhost:8988")
+    assert "http://localhost:8988" in str(err)
+    assert err.url == "http://localhost:8988"
+    assert err.short == "unavailable"
+
+
+def test_onlyoffice_no_fileurl_error():
+    err = onlyoffice_no_fileurl_error("<empty />")
+    assert "<empty />" in str(err)
+    assert err.snippet == "<empty />"
+    assert err.short == "no-fileurl error"
+
+
+def test_backend_error_pipeline_backend():
+    """Combined pipelines report the failing step in the backend name."""
+    err = backend_error("pdf", "cannot read document")
+    assert err.backend == "pdf"
+    assert err.short == "cannot read document"
+    assert isinstance(err, PreviewBackendError)
+
+
+def test_error_pickle_round_trip():
+    """Exceptions survive pickling (the worker pool wire) intact."""
+    err = onlyoffice_error_from_code("-8")
+    restored = pickle.loads(pickle.dumps(err))
+    assert type(restored) is type(err)
+    assert str(restored) == str(err)
+    assert restored.code == err.code
+    assert restored.short == err.short
+    assert restored.backend == err.backend
+
+
+@pytest.mark.asyncio
+async def test_generate_office_preview_raises_structured_error(monkeypatch):
+    """On OnlyOffice failure, generate_office_preview raises OnlyOfficeError."""
+    async def fake_convert(_filepath: Path, request_timeout: float = 5.0) -> bytes:
+        raise onlyoffice_error_from_code("-8")
+
+    monkeypatch.setattr(office, "convert_to_png_async", fake_convert)
+
+    with pytest.raises(OnlyOfficeError) as exc_info:
+        await generate_office_preview(
+            FILES / "file-sample_100kB.docx",
+            quality=60,
+            maxsize=512,
+            maxzoom=2.0,
+        )
+
+    err = exc_info.value
+    assert err.code == "-8"
+    assert err.short == "jwt error"
+    assert str(err) == "OnlyOffice JWT authentication failed"

tests/test_previews.py

@@ -0,0 +1,191 @@
+"""Success tests for low-level preview conversion functions.
+
+These tests exercise the concrete backend converters (image, video, PDF,
+office) against the fixture files in tests/files/.  The only thing they
+assert is that the conversion succeeds and returns non-empty AVIF bytes, plus
+basic metadata sanity checks.
+
+Office tests are skipped unless an OnlyOffice Document Server is reachable.
+Configure them with environment variables before running pytest:
+
+    ONLYOFFICE_URL=http://localhost:8988
+    ONLYOFFICE_JWT_SECRET=<same-secret-you-gave-the-oo-container>
+    ONLYOFFICE_CALLBACK_HOST=<host the OO container can reach>
+"""
+
+from __future__ import annotations
+
+import os
+from pathlib import Path
+
+import pytest
+
+from mediapreview import dispatch
+from mediapreview.backends.image import (
+    process_image,
+    process_image_buffer,
+    process_image_pyvips,
+)
+from mediapreview.backends.pdf import process_pdf
+from mediapreview.backends.video import process_video
+from mediapreview.office import is_available_async
+from mediapreview.pool import generate_office_preview
+
+FILES = Path(__file__).resolve().parent / "files"
+
+
+# ---------------------------------------------------------------------------
+# Helpers
+# ---------------------------------------------------------------------------
+
+
+def _assert_ok(data, resp, backend: str | None = None) -> None:
+    assert resp.ok, f"conversion failed: {resp.error}"
+    assert data
+    assert resp.mime == "image/avif"
+    if backend:
+        assert resp.backend == backend
+
+
+# ---------------------------------------------------------------------------
+# Images
+# ---------------------------------------------------------------------------
+
+
+@pytest.mark.parametrize("path", sorted(FILES.glob("Landscape_*.jpg")))
+def test_process_image_exif_orientations(path: Path) -> None:
+    """Every EXIF orientation fixture must produce a valid preview."""
+    data, resp = process_image(path, maxsize=512, quality=60)
+    _assert_ok(data, resp, backend="vips")
+    assert resp.width in (1200, 1800)
+    assert resp.height in (1200, 1800)
+
+
+def test_process_image_hdr_avif() -> None:
+    """HDR AVIF images are routed through ffmpeg to preserve colour metadata."""
+    path = FILES / "hdr_cosmos01650_cicp9-16-9_yuv444_full_qp10.avif"
+    data, resp = process_image(path, maxsize=512, quality=60)
+    _assert_ok(data, resp, backend="ffmpeg")
+    assert resp.width is not None
+    assert resp.height is not None
+
+
+def test_process_image_pyvips() -> None:
+    """The pyvips-only image backend works on a plain JPEG."""
+    path = FILES / "Landscape_1.jpg"
+    data, resp = process_image_pyvips(path, maxsize=512, quality=60)
+    _assert_ok(data, resp, backend="vips")
+
+
+def test_process_image_buffer() -> None:
+    """Processing a JPEG from a buffer produces the same valid preview."""
+    path = FILES / "Landscape_1.jpg"
+    data, resp = process_image_buffer(
+        path.read_bytes(), maxsize=512, quality=60, maxzoom=2.0
+    )
+    _assert_ok(data, resp, backend="vips")
+
+
+# ---------------------------------------------------------------------------
+# Videos
+# ---------------------------------------------------------------------------
+
+
+VIDEO_FIXTURES = [
+    ("sample-1mb.mp4", 854, 480),
+    ("rotated_90.mp4", 480, 854),
+    ("rotated_270.mp4", 480, 854),
+    ("hdr_video.mp4", 320, 240),
+    ("hdr_rotated_90.mp4", 240, 320),
+    ("hdr_rotated_270.mp4", 240, 320),
+]
+
+
+@pytest.mark.parametrize(
+    ("filename", "expected_width", "expected_height"),
+    VIDEO_FIXTURES,
+    ids=[f[0] for f in VIDEO_FIXTURES],
+)
+def test_process_video(filename: str, expected_width: int, expected_height: int) -> None:
+    """SDR and HDR video clips, with and without rotation, convert successfully."""
+    data, resp = process_video(FILES / filename, maxsize=512, quality=60)
+    _assert_ok(data, resp, backend="video")
+    assert resp.width == expected_width
+    assert resp.height == expected_height
+
+
+# ---------------------------------------------------------------------------
+# PDF
+# ---------------------------------------------------------------------------
+
+
+def test_process_pdf() -> None:
+    """A simple single-page PDF converts to a valid preview."""
+    data, resp = process_pdf(
+        FILES / "sample.pdf",
+        maxsize=512,
+        maxzoom=2.0,
+        quality=60,
+    )
+    _assert_ok(data, resp, backend="pdf+vips")
+    assert resp.width == 595
+    assert resp.height == 842
+
+
+# ---------------------------------------------------------------------------
+# Public dispatch entry point
+# ---------------------------------------------------------------------------
+
+
+DISPATCH_FIXTURES = [
+    ("Landscape_1.jpg", "vips", 1800, 1200),
+    ("sample-1mb.mp4", "video", 854, 480),
+    ("sample.pdf", "pdf+vips", 595, 842),
+]
+
+
+@pytest.mark.parametrize(
+    ("filename", "backend", "expected_width", "expected_height"),
+    DISPATCH_FIXTURES,
+    ids=[f[0] for f in DISPATCH_FIXTURES],
+)
+def test_dispatch(
+    filename: str, backend: str, expected_width: int, expected_height: int
+) -> None:
+    """The public dispatch() wrapper routes to the correct backend and succeeds."""
+    data, resp = dispatch(FILES / filename, quality=60, maxsize=512, maxzoom=2.0)
+    _assert_ok(data, resp, backend=backend)
+    assert resp.width == expected_width
+    assert resp.height == expected_height
+
+
+# ---------------------------------------------------------------------------
+# Office previews via OnlyOffice
+# ---------------------------------------------------------------------------
+
+
+@pytest.mark.asyncio
+async def test_generate_office_preview() -> None:
+    """DOCX preview through OnlyOffice.
+
+    Skipped unless an OnlyOffice server is reachable and ONLYOFFICE_JWT_SECRET
+    is set to the same secret the OO container is using. The shared secret must
+    be at least 32 bytes long so PyJWT doesn't warn about weak HMAC keys.
+    """
+    secret = os.environ.get("ONLYOFFICE_JWT_SECRET", "")
+    if len(secret.encode()) < 32:
+        pytest.skip("ONLYOFFICE_JWT_SECRET not set or shorter than 32 bytes")
+    if not await is_available_async():
+        pytest.skip("OnlyOffice Document Server not reachable")
+
+    data, resp = await generate_office_preview(
+        FILES / "file-sample_100kB.docx",
+        quality=60,
+        maxsize=512,
+        maxzoom=2.0,
+    )
+    assert data is not None
+    assert resp is not None
+    _assert_ok(data, resp)
+    assert resp.width is not None
+    assert resp.height is not None