0.52.9

.github/workflows/containers.yml

@@ -15,7 +15,6 @@ on:
   push:
     branches:
       - master
-      - dev
 
 jobs:
   metadata:
@@ -93,17 +92,28 @@ jobs:
           version: latest
           driver-opts: image=moby/buildkit:master
 
-      # dev branch -> :dev container tag
+      # master branch -> :dev container tag
+      - name: Docker meta :dev
+        if: ${{ github.ref == 'refs/heads/master' && github.event_name != 'release' }}
+        uses: docker/metadata-action@v5
+        id: meta_dev
+        with:
+          images: |
+            ${{ secrets.DOCKER_HUB_USERNAME }}/changedetection.io
+            ghcr.io/${{ github.repository }}
+          tags: |
+            type=raw,value=dev
+
       - name: Build and push :dev
         id: docker_build
-        if: ${{ github.ref == 'refs/heads/dev' }}
+        if: ${{ github.ref == 'refs/heads/master' && github.event_name != 'release' }}
         uses: docker/build-push-action@v6
         with:
           context: ./
           file: ./Dockerfile
           push: true
-          tags: |
-            ${{ secrets.DOCKER_HUB_USERNAME }}/changedetection.io:dev,ghcr.io/${{ github.repository }}:dev
+          tags: ${{ steps.meta_dev.outputs.tags }}
+          labels: ${{ steps.meta_dev.outputs.labels }}
           platforms: linux/amd64,linux/arm64,linux/arm/v7,linux/arm/v8
           cache-from: type=gha
           cache-to: type=gha,mode=max
@@ -142,6 +152,7 @@ jobs:
           file: ./Dockerfile
           push: true
           tags: ${{ steps.meta.outputs.tags }}
+          labels: ${{ steps.meta.outputs.labels }}
           platforms: linux/amd64,linux/arm64,linux/arm/v7,linux/arm/v8
           cache-from: type=gha
           cache-to: type=gha,mode=max

Dockerfile

@@ -78,6 +78,12 @@ RUN --mount=type=cache,id=pip,sharing=locked,target=/tmp/pip-cache \
 # Final image stage
 FROM python:${PYTHON_VERSION}-slim-bookworm
 LABEL org.opencontainers.image.source="https://github.com/dgtlmoon/changedetection.io"
+LABEL org.opencontainers.image.url="https://changedetection.io"
+LABEL org.opencontainers.image.documentation="https://changedetection.io/tutorials"
+LABEL org.opencontainers.image.title="changedetection.io"
+LABEL org.opencontainers.image.description="Self-hosted web page change monitoring and notification service"
+LABEL org.opencontainers.image.licenses="Apache-2.0"
+LABEL org.opencontainers.image.vendor="changedetection.io"
 
 RUN apt-get update && apt-get install -y --no-install-recommends \
     libxslt1.1 \

changedetectionio/__init__.py

@@ -2,7 +2,7 @@
 
 # Read more https://github.com/dgtlmoon/changedetection.io/wiki
 # Semver means never use .01, or 00. Should be .1.
-__version__ = '0.52.7'
+__version__ = '0.52.9'
 
 from changedetectionio.strtobool import strtobool
 from json.decoder import JSONDecodeError

changedetectionio/async_update_worker.py

@@ -163,8 +163,10 @@ async def async_update_worker(worker_id, q, notification_q, app, datastore, exec
                 except ProcessorException as e:
                     if e.screenshot:
                         watch.save_screenshot(screenshot=e.screenshot)
+                        e.screenshot = None  # Free memory immediately
                     if e.xpath_data:
                         watch.save_xpath_data(data=e.xpath_data)
+                        e.xpath_data = None  # Free memory immediately
                     datastore.update_watch(uuid=uuid, update_obj={'last_error': e.message})
                     process_changedetection_results = False
 
@@ -184,9 +186,11 @@ async def async_update_worker(worker_id, q, notification_q, app, datastore, exec
 
                     if e.screenshot:
                         watch.save_screenshot(screenshot=e.screenshot, as_error=True)
+                        e.screenshot = None  # Free memory immediately
 
                     if e.xpath_data:
                         watch.save_xpath_data(data=e.xpath_data)
+                        e.xpath_data = None  # Free memory immediately
                         
                     process_changedetection_results = False
 
@@ -205,8 +209,10 @@ async def async_update_worker(worker_id, q, notification_q, app, datastore, exec
 
                     if e.screenshot:
                         watch.save_screenshot(screenshot=e.screenshot, as_error=True)
+                        e.screenshot = None  # Free memory immediately
                     if e.xpath_data:
                         watch.save_xpath_data(data=e.xpath_data, as_error=True)
+                        e.xpath_data = None  # Free memory immediately
                     if e.page_text:
                         watch.save_error_text(contents=e.page_text)
 
@@ -223,9 +229,11 @@ async def async_update_worker(worker_id, q, notification_q, app, datastore, exec
                     # Filter wasnt found, but we should still update the visual selector so that they can have a chance to set it up again
                     if e.screenshot:
                         watch.save_screenshot(screenshot=e.screenshot)
+                        e.screenshot = None  # Free memory immediately
 
                     if e.xpath_data:
                         watch.save_xpath_data(data=e.xpath_data)
+                        e.xpath_data = None  # Free memory immediately
 
                     # Only when enabled, send the notification
                     if watch.get('filter_failure_notification_send', False):
@@ -317,6 +325,7 @@ async def async_update_worker(worker_id, q, notification_q, app, datastore, exec
                     err_text = "Error running JS Actions - Page request - "+e.message
                     if e.screenshot:
                         watch.save_screenshot(screenshot=e.screenshot, as_error=True)
+                        e.screenshot = None  # Free memory immediately
                     datastore.update_watch(uuid=uuid, update_obj={'last_error': err_text,
                                                                 'last_check_status': e.status_code})
                     process_changedetection_results = False
@@ -328,6 +337,7 @@ async def async_update_worker(worker_id, q, notification_q, app, datastore, exec
 
                     if e.screenshot:
                         watch.save_screenshot(screenshot=e.screenshot, as_error=True)
+                        e.screenshot = None  # Free memory immediately
 
                     datastore.update_watch(uuid=uuid, update_obj={'last_error': err_text,
                                                                 'last_check_status': e.status_code,
@@ -369,9 +379,17 @@ async def async_update_worker(worker_id, q, notification_q, app, datastore, exec
                         if changed_detected or not watch.history_n:
                             if update_handler.screenshot:
                                 watch.save_screenshot(screenshot=update_handler.screenshot)
+                                # Free screenshot memory immediately after saving
+                                update_handler.screenshot = None
+                                if hasattr(update_handler, 'fetcher') and hasattr(update_handler.fetcher, 'screenshot'):
+                                    update_handler.fetcher.screenshot = None
 
                             if update_handler.xpath_data:
                                 watch.save_xpath_data(data=update_handler.xpath_data)
+                                # Free xpath data memory
+                                update_handler.xpath_data = None
+                                if hasattr(update_handler, 'fetcher') and hasattr(update_handler.fetcher, 'xpath_data'):
+                                    update_handler.fetcher.xpath_data = None
 
                             # Ensure unique timestamp for history
                             if watch.newest_history_key and int(fetch_start_time) == int(watch.newest_history_key):
@@ -438,6 +456,20 @@ async def async_update_worker(worker_id, q, notification_q, app, datastore, exec
                     update_handler.fetcher.clear_content()
                     logger.debug(f"Cleared fetcher content for UUID {uuid}")
 
+                # Explicitly delete update_handler to free all references
+                if update_handler:
+                    del update_handler
+                    update_handler = None
+
+                # Force aggressive memory cleanup after clearing
+                import gc
+                gc.collect()
+                try:
+                    import ctypes
+                    ctypes.CDLL('libc.so.6').malloc_trim(0)
+                except Exception:
+                    pass
+
         except Exception as e:
             logger.error(f"Worker {worker_id} unexpected error processing {uuid}: {e}")
             logger.error(f"Worker {worker_id} traceback:", exc_info=True)

changedetectionio/blueprint/ui/__init__.py

@@ -1,6 +1,6 @@
 import time
 import threading
-from flask import Blueprint, request, redirect, url_for, flash, render_template, session
+from flask import Blueprint, request, redirect, url_for, flash, render_template, session, current_app
 from flask_babel import gettext
 from loguru import logger
 
@@ -404,4 +404,25 @@ def construct_blueprint(datastore: ChangeDetectionStore, update_q, worker_handle
 
         return redirect(url_for('watchlist.index'))
 
+    @ui_blueprint.route("/language/auto-detect", methods=['GET'])
+    def delete_locale_language_session_var_if_it_exists():
+        """Clear the session locale preference to auto-detect from browser Accept-Language header"""
+        if 'locale' in session:
+            session.pop('locale', None)
+            # Refresh Flask-Babel to clear cached locale
+            from flask_babel import refresh
+            refresh()
+            flash(gettext("Language set to auto-detect from browser"))
+
+        # Check if there's a redirect parameter to return to the same page
+        redirect_url = request.args.get('redirect')
+
+        # If redirect is provided and safe, use it
+        from changedetectionio.is_safe_url import is_safe_url
+        if redirect_url and is_safe_url(redirect_url, current_app):
+            return redirect(redirect_url)
+
+        # Otherwise redirect to watchlist
+        return redirect(url_for('watchlist.index'))
+
     return ui_blueprint

changedetectionio/content_fetchers/base.py

@@ -71,10 +71,19 @@ class Fetcher():
     supports_screenshots = False        # Can capture page screenshots
     supports_xpath_element_data = False # Can extract xpath element positions/data for visual selector
 
+    # Screenshot element locking - prevents layout shifts during screenshot capture
+    # Only needed for visual comparison (image_ssim_diff processor)
+    # Locks element dimensions in the first viewport to prevent headers/ads from resizing
+    lock_viewport_elements = False      # Default: disabled for performance
+
     def __init__(self, **kwargs):
         if kwargs and 'screenshot_format' in kwargs:
             self.screenshot_format = kwargs.get('screenshot_format')
 
+        # Allow lock_viewport_elements to be set via kwargs
+        if kwargs and 'lock_viewport_elements' in kwargs:
+            self.lock_viewport_elements = kwargs.get('lock_viewport_elements')
+
 
     @classmethod
     def get_status_icon_data(cls):

changedetectionio/content_fetchers/playwright.py

@@ -10,18 +10,20 @@ from changedetectionio.content_fetchers import SCREENSHOT_MAX_HEIGHT_DEFAULT, vi
 from changedetectionio.content_fetchers.base import Fetcher, manage_user_agent
 from changedetectionio.content_fetchers.exceptions import PageUnloadable, Non200ErrorCodeReceived, EmptyReply, ScreenshotUnavailable
 
-async def capture_full_page_async(page, screenshot_format='JPEG'):
+async def capture_full_page_async(page, screenshot_format='JPEG', watch_uuid=None, lock_viewport_elements=False):
     import os
     import time
-    import multiprocessing
 
     start = time.time()
+    watch_info = f"[{watch_uuid}] " if watch_uuid else ""
 
+    setup_start = time.time()
     page_height = await page.evaluate("document.documentElement.scrollHeight")
     page_width = await page.evaluate("document.documentElement.scrollWidth")
     original_viewport = page.viewport_size
+    dimensions_time = time.time() - setup_start
 
-    logger.debug(f"Playwright viewport size {page.viewport_size} page height {page_height} page width {page_width}")
+    logger.debug(f"{watch_info}Playwright viewport size {page.viewport_size} page height {page_height} page width {page_width} (got dimensions in {dimensions_time:.2f}s)")
 
     # Use an approach similar to puppeteer: set a larger viewport and take screenshots in chunks
     step_size = SCREENSHOT_SIZE_STITCH_THRESHOLD # Size that won't cause GPU to overflow
@@ -29,25 +31,31 @@ async def capture_full_page_async(page, screenshot_format='JPEG'):
     y = 0
     elements_locked = False
 
-    if page_height > page.viewport_size['height']:
-
-        # Lock all element dimensions BEFORE screenshot to prevent CSS media queries from resizing
-        # capture_full_page_async() changes viewport height which triggers @media (min-height) rules
+    # Only lock viewport elements if explicitly enabled (for image_ssim_diff processor)
+    # This prevents headers/ads from resizing when viewport changes
+    if lock_viewport_elements and page_height > page.viewport_size['height']:
+        lock_start = time.time()
         lock_elements_js_path = os.path.join(os.path.dirname(__file__), 'res', 'lock-elements-sizing.js')
         with open(lock_elements_js_path, 'r') as f:
             lock_elements_js = f.read()
         await page.evaluate(lock_elements_js)
         elements_locked = True
+        lock_time = time.time() - lock_start
+        logger.debug(f"{watch_info}Viewport element locking enabled (took {lock_time:.2f}s)")
 
-        logger.debug("Element dimensions locked before screenshot capture")
-
+    if page_height > page.viewport_size['height']:
         if page_height < step_size:
             step_size = page_height # Incase page is bigger than default viewport but smaller than proposed step size
-        logger.debug(f"Setting bigger viewport to step through large page width W{page.viewport_size['width']}xH{step_size} because page_height > viewport_size")
+        viewport_start = time.time()
+        logger.debug(f"{watch_info}Setting bigger viewport to step through large page width W{page.viewport_size['width']}xH{step_size} because page_height > viewport_size")
         # Set viewport to a larger size to capture more content at once
         await page.set_viewport_size({'width': page.viewport_size['width'], 'height': step_size})
+        viewport_time = time.time() - viewport_start
+        logger.debug(f"{watch_info}Viewport changed to {page.viewport_size['width']}x{step_size} (took {viewport_time:.2f}s)")
 
     # Capture screenshots in chunks up to the max total height
+    capture_start = time.time()
+    chunk_times = []
     # Use PNG for better quality (no compression artifacts), JPEG for smaller size
     screenshot_type = screenshot_format.lower() if screenshot_format else 'jpeg'
     # PNG should use quality 100, JPEG uses configurable quality
@@ -69,7 +77,11 @@ async def capture_full_page_async(page, screenshot_format='JPEG'):
         if screenshot_type == 'jpeg':
             screenshot_kwargs['quality'] = screenshot_quality
 
+        chunk_start = time.time()
         screenshot_chunks.append(await page.screenshot(**screenshot_kwargs))
+        chunk_time = time.time() - chunk_start
+        chunk_times.append(chunk_time)
+        logger.debug(f"{watch_info}Chunk {len(screenshot_chunks)} captured in {chunk_time:.2f}s")
         y += step_size
 
     # Restore original viewport size
@@ -81,40 +93,54 @@ async def capture_full_page_async(page, screenshot_format='JPEG'):
         with open(unlock_elements_js_path, 'r') as f:
             unlock_elements_js = f.read()
         await page.evaluate(unlock_elements_js)
-        logger.debug("Element dimensions unlocked after screenshot capture")
+        logger.debug(f"{watch_info}Element dimensions unlocked after screenshot capture")
+
+    capture_time = time.time() - capture_start
+    total_capture_time = sum(chunk_times)
+    logger.debug(f"{watch_info}All {len(screenshot_chunks)} chunks captured in {capture_time:.2f}s (total chunk time: {total_capture_time:.2f}s)")
 
     # If we have multiple chunks, stitch them together
     if len(screenshot_chunks) > 1:
-        logger.debug(f"Screenshot stitching {len(screenshot_chunks)} chunks together")
+        stitch_start = time.time()
+        logger.debug(f"{watch_info}Starting stitching of {len(screenshot_chunks)} chunks")
 
-        # For small number of chunks (2-3), stitch inline to avoid multiprocessing overhead
-        # Only use separate process for many chunks (4+) to avoid blocking the event loop
-        if len(screenshot_chunks) <= 3:
-            from changedetectionio.content_fetchers.screenshot_handler import stitch_images_inline
-            screenshot = stitch_images_inline(screenshot_chunks, page_height, SCREENSHOT_MAX_TOTAL_HEIGHT)
-        else:
-            # Use separate process for many chunks to avoid blocking
-            # Always use spawn for thread safety - consistent behavior in tests and production
-            from changedetectionio.content_fetchers.screenshot_handler import stitch_images_worker
-            ctx = multiprocessing.get_context('spawn')
-            parent_conn, child_conn = ctx.Pipe()
-            p = ctx.Process(target=stitch_images_worker, args=(child_conn, screenshot_chunks, page_height, SCREENSHOT_MAX_TOTAL_HEIGHT))
-            p.start()
-            screenshot = parent_conn.recv_bytes()
-            p.join()
-            # Explicit cleanup
-            del p
-            del parent_conn, child_conn
+        # Always use spawn subprocess for ANY stitching (2+ chunks)
+        # PIL allocates at C level and Python GC never releases it - subprocess exit forces OS to reclaim
+        # Trade-off: 35MB resource_tracker vs 500MB+ PIL leak in main process
+        from changedetectionio.content_fetchers.screenshot_handler import stitch_images_worker_raw_bytes
+        import multiprocessing
+        import struct
 
+        ctx = multiprocessing.get_context('spawn')
+        parent_conn, child_conn = ctx.Pipe()
+        p = ctx.Process(target=stitch_images_worker_raw_bytes, args=(child_conn, page_height, SCREENSHOT_MAX_TOTAL_HEIGHT))
+        p.start()
+
+        # Send via raw bytes (no pickle)
+        parent_conn.send_bytes(struct.pack('I', len(screenshot_chunks)))
+        for chunk in screenshot_chunks:
+            parent_conn.send_bytes(chunk)
+
+        screenshot = parent_conn.recv_bytes()
+        p.join()
+
+        parent_conn.close()
+        child_conn.close()
+        del p, parent_conn, child_conn
+
+        stitch_time = time.time() - stitch_start
+        total_time = time.time() - start
+        setup_time = total_time - capture_time - stitch_time
         logger.debug(
-            f"Screenshot (chunked/stitched) - Page height: {page_height} Capture height: {SCREENSHOT_MAX_TOTAL_HEIGHT} - Stitched together in {time.time() - start:.2f}s")
-        # Explicit cleanup
-        del screenshot_chunks
-        screenshot_chunks = None
+            f"{watch_info}Screenshot complete - Page height: {page_height}px, Capture height: {SCREENSHOT_MAX_TOTAL_HEIGHT}px | "
+            f"Setup: {setup_time:.2f}s, Capture: {capture_time:.2f}s, Stitching: {stitch_time:.2f}s, Total: {total_time:.2f}s")
         return screenshot
 
+    total_time = time.time() - start
+    setup_time = total_time - capture_time
     logger.debug(
-        f"Screenshot Page height: {page_height} Capture height: {SCREENSHOT_MAX_TOTAL_HEIGHT} - Stitched together in {time.time() - start:.2f}s")
+        f"{watch_info}Screenshot complete - Page height: {page_height}px, Capture height: {SCREENSHOT_MAX_TOTAL_HEIGHT}px | "
+        f"Setup: {setup_time:.2f}s, Single chunk: {capture_time:.2f}s, Total: {total_time:.2f}s")
 
     return screenshot_chunks[0]
 
@@ -184,7 +210,8 @@ class fetcher(Fetcher):
 
     async def screenshot_step(self, step_n=''):
         super().screenshot_step(step_n=step_n)
-        screenshot = await capture_full_page_async(page=self.page, screenshot_format=self.screenshot_format)
+        watch_uuid = getattr(self, 'watch_uuid', None)
+        screenshot = await capture_full_page_async(page=self.page, screenshot_format=self.screenshot_format, watch_uuid=watch_uuid, lock_viewport_elements=self.lock_viewport_elements)
 
         # Request GC immediately after screenshot to free memory
         # Screenshots can be large and browser steps take many of them
@@ -233,6 +260,7 @@ class fetcher(Fetcher):
         import playwright._impl._errors
         import time
         self.delete_browser_steps_screenshots()
+        self.watch_uuid = watch_uuid  # Store for use in screenshot_step
         response = None
 
         async with async_playwright() as p:
@@ -318,7 +346,7 @@ class fetcher(Fetcher):
                     logger.error(f"Error fetching FavIcon info {str(e)}, continuing.")
 
             if self.status_code != 200 and not ignore_status_codes:
-                screenshot = await capture_full_page_async(self.page, screenshot_format=self.screenshot_format)
+                screenshot = await capture_full_page_async(self.page, screenshot_format=self.screenshot_format, watch_uuid=watch_uuid, lock_viewport_elements=self.lock_viewport_elements)
                 # Cleanup before raising to prevent memory leak
                 await self.page.close()
                 await context.close()
@@ -374,7 +402,17 @@ class fetcher(Fetcher):
                 # which will significantly increase the IO size between the server and client, it's recommended to use the lowest
                 # acceptable screenshot quality here
                 # The actual screenshot - this always base64 and needs decoding! horrible! huge CPU usage
-                self.screenshot = await capture_full_page_async(page=self.page, screenshot_format=self.screenshot_format)
+                self.screenshot = await capture_full_page_async(page=self.page, screenshot_format=self.screenshot_format, watch_uuid=watch_uuid, lock_viewport_elements=self.lock_viewport_elements)
+
+                # Force aggressive memory cleanup - screenshots are large and base64 decode creates temporary buffers
+                await self.page.request_gc()
+                gc.collect()
+                # Release C-level memory from base64 decode back to OS
+                try:
+                    import ctypes
+                    ctypes.CDLL('libc.so.6').malloc_trim(0)
+                except Exception:
+                    pass
 
             except ScreenshotUnavailable:
                 # Re-raise screenshot unavailable exceptions

changedetectionio/content_fetchers/puppeteer.py

@@ -20,18 +20,20 @@ from changedetectionio.content_fetchers.exceptions import PageUnloadable, Non200
 # Screenshots also travel via the ws:// (websocket) meaning that the binary data is base64 encoded
 # which will significantly increase the IO size between the server and client, it's recommended to use the lowest
 # acceptable screenshot quality here
-async def capture_full_page(page, screenshot_format='JPEG'):
+async def capture_full_page(page, screenshot_format='JPEG', watch_uuid=None, lock_viewport_elements=False):
     import os
     import time
-    import multiprocessing
 
     start = time.time()
+    watch_info = f"[{watch_uuid}] " if watch_uuid else ""
 
+    setup_start = time.time()
     page_height = await page.evaluate("document.documentElement.scrollHeight")
     page_width = await page.evaluate("document.documentElement.scrollWidth")
     original_viewport = page.viewport
+    dimensions_time = time.time() - setup_start
 
-    logger.debug(f"Puppeteer viewport size {page.viewport} page height {page_height} page width {page_width}")
+    logger.debug(f"{watch_info}Puppeteer viewport size {page.viewport} page height {page_height} page width {page_width} (got dimensions in {dimensions_time:.2f}s)")
 
     # Bug 3 in Playwright screenshot handling
     # Some bug where it gives the wrong screenshot size, but making a request with the clip set first seems to solve it
@@ -50,20 +52,35 @@ async def capture_full_page(page, screenshot_format='JPEG'):
     screenshot_chunks = []
     y = 0
     elements_locked = False
-    if page_height > page.viewport['height']:
-        # Lock all element dimensions BEFORE screenshot to prevent CSS media queries from resizing
-        # capture_full_page() changes viewport height which triggers @media (min-height) rules
+
+    # Only lock viewport elements if explicitly enabled (for image_ssim_diff processor)
+    # This prevents headers/ads from resizing when viewport changes
+    if lock_viewport_elements and page_height > page.viewport['height']:
+        lock_start = time.time()
         lock_elements_js_path = os.path.join(os.path.dirname(__file__), 'res', 'lock-elements-sizing.js')
+        file_read_start = time.time()
         with open(lock_elements_js_path, 'r') as f:
             lock_elements_js = f.read()
-        await page.evaluate(lock_elements_js)
-        elements_locked = True
-        logger.debug("Element dimensions locked before screenshot capture")
+        file_read_time = time.time() - file_read_start
 
+        evaluate_start = time.time()
+        await page.evaluate(lock_elements_js)
+        evaluate_time = time.time() - evaluate_start
+
+        elements_locked = True
+        lock_time = time.time() - lock_start
+        logger.debug(f"{watch_info}Viewport element locking enabled - File read: {file_read_time:.3f}s, Browser evaluate: {evaluate_time:.2f}s, Total: {lock_time:.2f}s")
+
+    if page_height > page.viewport['height']:
         if page_height < step_size:
             step_size = page_height # Incase page is bigger than default viewport but smaller than proposed step size
+        viewport_start = time.time()
         await page.setViewport({'width': page.viewport['width'], 'height': step_size})
+        viewport_time = time.time() - viewport_start
+        logger.debug(f"{watch_info}Viewport changed to {page.viewport['width']}x{step_size} (took {viewport_time:.2f}s)")
 
+    capture_start = time.time()
+    chunk_times = []
     while y < min(page_height, SCREENSHOT_MAX_TOTAL_HEIGHT):
         # better than scrollTo incase they override it in the page
         await page.evaluate(
@@ -82,7 +99,11 @@ async def capture_full_page(page, screenshot_format='JPEG'):
         if screenshot_type == 'jpeg':
             screenshot_kwargs['quality'] = screenshot_quality
 
+        chunk_start = time.time()
         screenshot_chunks.append(await page.screenshot(**screenshot_kwargs))
+        chunk_time = time.time() - chunk_start
+        chunk_times.append(chunk_time)
+        logger.debug(f"{watch_info}Chunk {len(screenshot_chunks)} captured in {chunk_time:.2f}s")
         y += step_size
 
     await page.setViewport({'width': original_viewport['width'], 'height': original_viewport['height']})
@@ -93,26 +114,53 @@ async def capture_full_page(page, screenshot_format='JPEG'):
         with open(unlock_elements_js_path, 'r') as f:
             unlock_elements_js = f.read()
         await page.evaluate(unlock_elements_js)
-        logger.debug("Element dimensions unlocked after screenshot capture")
+        logger.debug(f"{watch_info}Element dimensions unlocked after screenshot capture")
+
+    capture_time = time.time() - capture_start
+    total_capture_time = sum(chunk_times)
+    logger.debug(f"{watch_info}All {len(screenshot_chunks)} chunks captured in {capture_time:.2f}s (total chunk time: {total_capture_time:.2f}s)")
 
     if len(screenshot_chunks) > 1:
-        # Always use spawn for thread safety - consistent behavior in tests and production
-        from changedetectionio.content_fetchers.screenshot_handler import stitch_images_worker
-        logger.debug(f"Screenshot stitching {len(screenshot_chunks)} chunks together")
+        stitch_start = time.time()
+        logger.debug(f"{watch_info}Starting stitching of {len(screenshot_chunks)} chunks")
+
+        # Always use spawn subprocess for ANY stitching (2+ chunks)
+        # PIL allocates at C level and Python GC never releases it - subprocess exit forces OS to reclaim
+        # Trade-off: 35MB resource_tracker vs 500MB+ PIL leak in main process
+        from changedetectionio.content_fetchers.screenshot_handler import stitch_images_worker_raw_bytes
+        import multiprocessing
+        import struct
+
         ctx = multiprocessing.get_context('spawn')
         parent_conn, child_conn = ctx.Pipe()
-        p = ctx.Process(target=stitch_images_worker, args=(child_conn, screenshot_chunks, page_height, SCREENSHOT_MAX_TOTAL_HEIGHT))
+        p = ctx.Process(target=stitch_images_worker_raw_bytes, args=(child_conn, page_height, SCREENSHOT_MAX_TOTAL_HEIGHT))
         p.start()
+
+        # Send via raw bytes (no pickle)
+        parent_conn.send_bytes(struct.pack('I', len(screenshot_chunks)))
+        for chunk in screenshot_chunks:
+            parent_conn.send_bytes(chunk)
+
         screenshot = parent_conn.recv_bytes()
         p.join()
-        logger.debug(
-            f"Screenshot (chunked/stitched) - Page height: {page_height} Capture height: {SCREENSHOT_MAX_TOTAL_HEIGHT} - Stitched together in {time.time() - start:.2f}s")
 
-        screenshot_chunks = None
+        parent_conn.close()
+        child_conn.close()
+        del p, parent_conn, child_conn
+
+        stitch_time = time.time() - stitch_start
+        total_time = time.time() - start
+        setup_time = total_time - capture_time - stitch_time
+        logger.debug(
+            f"{watch_info}Screenshot complete - Page height: {page_height}px, Capture height: {SCREENSHOT_MAX_TOTAL_HEIGHT}px | "
+            f"Setup: {setup_time:.2f}s, Capture: {capture_time:.2f}s, Stitching: {stitch_time:.2f}s, Total: {total_time:.2f}s")
         return screenshot
 
+    total_time = time.time() - start
+    setup_time = total_time - capture_time
     logger.debug(
-        f"Screenshot Page height: {page_height} Capture height: {SCREENSHOT_MAX_TOTAL_HEIGHT} - Stitched together in {time.time() - start:.2f}s")
+        f"{watch_info}Screenshot complete - Page height: {page_height}px, Capture height: {SCREENSHOT_MAX_TOTAL_HEIGHT}px | "
+        f"Setup: {setup_time:.2f}s, Single chunk: {capture_time:.2f}s, Total: {total_time:.2f}s")
     return screenshot_chunks[0]
 
 
@@ -357,7 +405,7 @@ class fetcher(Fetcher):
                 logger.error(f"Error fetching FavIcon info {str(e)}, continuing.")
 
         if self.status_code != 200 and not ignore_status_codes:
-            screenshot = await capture_full_page(page=self.page, screenshot_format=self.screenshot_format)
+            screenshot = await capture_full_page(page=self.page, screenshot_format=self.screenshot_format, watch_uuid=watch_uuid, lock_viewport_elements=self.lock_viewport_elements)
 
             raise Non200ErrorCodeReceived(url=url, status_code=self.status_code, screenshot=screenshot)
 
@@ -387,7 +435,17 @@ class fetcher(Fetcher):
 
         # Now take screenshot (scrolling may trigger layout changes, but measurements are already captured)
         logger.debug(f"Screenshot format {self.screenshot_format}")
-        self.screenshot = await capture_full_page(page=self.page, screenshot_format=self.screenshot_format)
+        self.screenshot = await capture_full_page(page=self.page, screenshot_format=self.screenshot_format, watch_uuid=watch_uuid, lock_viewport_elements=self.lock_viewport_elements)
+
+        # Force aggressive memory cleanup - pyppeteer base64 decode creates temporary buffers
+        import gc
+        gc.collect()
+        # Release C-level memory from base64 decode back to OS
+        try:
+            import ctypes
+            ctypes.CDLL('libc.so.6').malloc_trim(0)
+        except Exception:
+            pass
         self.xpath_data = await self.page.evaluate(XPATH_ELEMENT_JS, {
             "visualselector_xpath_selectors": visualselector_xpath_selectors,
             "max_height": MAX_TOTAL_HEIGHT

changedetectionio/content_fetchers/res/lock-elements-sizing.js

@@ -1,5 +1,5 @@
 /**
- * Lock Element Dimensions for Screenshot Capture
+ * Lock Element Dimensions for Screenshot Capture (First Viewport Only)
  *
  * THE PROBLEM:
  * When taking full-page screenshots of tall pages, Chrome/Puppeteer/Playwright need to:
@@ -10,40 +10,31 @@
  * However, changing the viewport height triggers CSS media queries like:
  *   @media (min-height: 860px) { .ad { height: 250px; } }
  *
- * This causes elements (especially ads) to resize during screenshot capture, creating a mismatch:
- * - Screenshot shows element at NEW size (after media query triggered)
- * - xpath element coordinates measured at OLD size (before viewport change)
- * - Visual selector overlays don't align with screenshot
- *
- * EXAMPLE BUG:
- * - Initial viewport: 1280x800, ad height: 138px, article position: 279px ✓
- * - Viewport changes to 1280x3809 for screenshot
- * - Media query triggers: ad expands to 250px
- * - All content below shifts down by 112px (250-138)
- * - Article now at position: 391px (279+112)
- * - But xpath data says 279px → 112px mismatch! ✗
+ * This causes elements (especially ads/headers) to resize during screenshot capture.
  *
  * THE SOLUTION:
- * Before changing viewport, lock ALL element dimensions with !important inline styles.
- * Inline styles with !important override media query CSS, preventing layout changes.
+ * Lock element dimensions in the FIRST VIEWPORT ONLY with !important inline styles.
+ * This prevents headers, navigation, and top ads from resizing when viewport changes.
+ * We only lock the visible portion because:
+ * - Most layout shifts happen in headers/navbars/top ads
+ * - Locking only visible elements is 100x+ faster (100-200 elements vs 10,000+)
+ * - Below-fold content shifts don't affect visual comparison accuracy
  *
  * WHAT THIS SCRIPT DOES:
- * 1. Iterates through every element on the page
- * 2. Captures current computed dimensions (width, height)
- * 3. Sets inline styles with !important to freeze those dimensions
+ * 1. Gets current viewport height
+ * 2. Finds elements within first viewport (top of page to bottom of screen)
+ * 3. Locks their dimensions with !important inline styles
  * 4. Disables ResizeObserver API (for JS-based resizing)
- * 5. When viewport changes for screenshot, media queries can't resize anything
- * 6. Layout remains consistent → xpath coordinates match screenshot ✓
  *
  * USAGE:
  * Execute this script BEFORE calling capture_full_page() / screenshot functions.
- * The page must be fully loaded and settled at its initial viewport size.
- * No need to restore state afterward - page is closed after screenshot.
+ * Only enabled for image_ssim_diff processor (visual comparison).
+ * Default: OFF for performance.
  *
  * PERFORMANCE:
- * - Iterates all DOM elements (can be 1000s on complex pages)
- * - Typically completes in 50-200ms
- * - One-time cost before screenshot, well worth it for coordinate accuracy
+ * - Only processes 100-300 elements (first viewport) vs 10,000+ (entire page)
+ * - Typically completes in 10-50ms
+ * - 100x+ faster than locking entire page
  *
  * @see https://github.com/dgtlmoon/changedetection.io/issues/XXXX
  */
@@ -52,11 +43,34 @@
     // Store original styles in a global WeakMap for later restoration
     window.__elementSizingRestore = new WeakMap();
 
-    // Lock ALL element dimensions to prevent media query layout changes
-    document.querySelectorAll('*').forEach(el => {
-        const computed = window.getComputedStyle(el);
-        const rect = el.getBoundingClientRect();
+    const start = performance.now();
 
+    // Get current viewport height (visible portion of page)
+    const viewportHeight = window.innerHeight;
+
+    // Get all elements and filter to FIRST VIEWPORT ONLY
+    // This dramatically reduces elements to process (100-300 vs 10,000+)
+    const allElements = Array.from(document.querySelectorAll('*'));
+
+    // BATCH READ PHASE: Get bounding rects and filter to viewport
+    const measurements = allElements.map(el => {
+        const rect = el.getBoundingClientRect();
+        const computed = window.getComputedStyle(el);
+
+        // Only lock elements in the first viewport (visible on initial page load)
+        // rect.top < viewportHeight means element starts within visible area
+        const inViewport = rect.top < viewportHeight && rect.top >= 0;
+        const hasSize = rect.height > 0 && rect.width > 0;
+
+        return inViewport && hasSize ? { el, computed, rect } : null;
+    }).filter(Boolean);  // Remove null entries
+
+    const elapsed = performance.now() - start;
+    console.log(`Locked first viewport elements: ${measurements.length} of ${allElements.length} total elements (viewport height: ${viewportHeight}px, took ${elapsed.toFixed(0)}ms)`);
+
+    // BATCH WRITE PHASE: Apply all inline styles without triggering layout
+    // No interleaved reads means browser can optimize style application
+    measurements.forEach(({el, computed, rect}) => {
         // Save original inline style values BEFORE locking
         const properties = ['height', 'min-height', 'max-height', 'width', 'min-width', 'max-width'];
         const originalStyles = {};
@@ -89,5 +103,5 @@
         disconnect() {}
     };
 
-    console.log('✓ Element dimensions locked to prevent media query changes during screenshot');
+    console.log(`✓ Element dimensions locked (${measurements.length} elements) to prevent media query changes during screenshot`);
 })();

changedetectionio/content_fetchers/screenshot_handler.py

@@ -8,92 +8,42 @@ from loguru import logger
 
 from changedetectionio.content_fetchers import SCREENSHOT_MAX_HEIGHT_DEFAULT, SCREENSHOT_DEFAULT_QUALITY
 
-# Cache font to avoid loading on every stitch
-_cached_font = None
-
-def _get_caption_font():
-    """Get or create cached font for caption text."""
-    global _cached_font
-    if _cached_font is None:
-        from PIL import ImageFont
-        try:
-            _cached_font = ImageFont.truetype("arial.ttf", 35)
-        except IOError:
-            _cached_font = ImageFont.load_default()
-    return _cached_font
-
-
-def stitch_images_inline(chunks_bytes, original_page_height, capture_height):
-    """
-    Stitch image chunks together inline (no multiprocessing).
-    Optimized for small number of chunks (2-3) to avoid process creation overhead.
-
-    Args:
-        chunks_bytes: List of JPEG image bytes
-        original_page_height: Original page height in pixels
-        capture_height: Maximum capture height
-
-    Returns:
-        bytes: Stitched JPEG image
-    """
-    import os
-    import io
-    from PIL import Image, ImageDraw
-
-    # Load images from byte chunks
-    images = [Image.open(io.BytesIO(b)) for b in chunks_bytes]
-    total_height = sum(im.height for im in images)
-    max_width = max(im.width for im in images)
-
-    # Create stitched image
-    stitched = Image.new('RGB', (max_width, total_height))
-    y_offset = 0
-    for im in images:
-        stitched.paste(im, (0, y_offset))
-        y_offset += im.height
-        im.close()  # Close immediately after pasting
-
-    # Draw caption only if page was trimmed
-    if original_page_height > capture_height:
-        draw = ImageDraw.Draw(stitched)
-        caption_text = f"WARNING: Screenshot was {original_page_height}px but trimmed to {capture_height}px because it was too long"
-        padding = 10
-        font = _get_caption_font()
-
-        bbox = draw.textbbox((0, 0), caption_text, font=font)
-        text_width = bbox[2] - bbox[0]
-        text_height = bbox[3] - bbox[1]
-
-        # Draw white background rectangle
-        draw.rectangle([(0, 0), (max_width, text_height + 2 * padding)], fill=(255, 255, 255))
-
-        # Draw text centered
-        text_x = (max_width - text_width) // 2
-        draw.text((text_x, padding), caption_text, font=font, fill=(255, 0, 0))
-
-    # Encode to JPEG
-    output = io.BytesIO()
-    stitched.save(output, format="JPEG", quality=int(os.getenv("SCREENSHOT_QUALITY", SCREENSHOT_DEFAULT_QUALITY)), optimize=True)
-    result = output.getvalue()
-
-    # Cleanup
-    stitched.close()
-
-    return result
-
-
-def stitch_images_worker(pipe_conn, chunks_bytes, original_page_height, capture_height):
+def stitch_images_worker_raw_bytes(pipe_conn, original_page_height, capture_height):
     """
     Stitch image chunks together in a separate process.
-    Used for large number of chunks (4+) to avoid blocking the main event loop.
+
+    Uses spawn multiprocessing to isolate PIL's C-level memory allocation.
+    When the subprocess exits, the OS reclaims ALL memory including C-level allocations
+    that Python's GC cannot release. This prevents the ~50MB per stitch from accumulating
+    in the main process.
+
+    Trade-off: Adds 35MB resource_tracker subprocess, but prevents 500MB+ memory leak
+    in main process (much better at scale: 35GB vs 500GB for 1000 instances).
+
+    Args:
+        pipe_conn: Pipe connection to receive data and send result
+        original_page_height: Original page height in pixels
+        capture_height: Maximum capture height
     """
     import os
     import io
+    import struct
     from PIL import Image, ImageDraw, ImageFont
 
     try:
+        # Receive chunk count as 4-byte integer (no pickle!)
+        count_bytes = pipe_conn.recv_bytes()
+        chunk_count = struct.unpack('I', count_bytes)[0]
+
+        # Receive each chunk as raw bytes (no pickle!)
+        chunks_bytes = []
+        for _ in range(chunk_count):
+            chunks_bytes.append(pipe_conn.recv_bytes())
+
         # Load images from byte chunks
         images = [Image.open(io.BytesIO(b)) for b in chunks_bytes]
+        del chunks_bytes
+
         total_height = sum(im.height for im in images)
         max_width = max(im.width for im in images)
 
@@ -103,15 +53,14 @@ def stitch_images_worker(pipe_conn, chunks_bytes, original_page_height, capture_
         for im in images:
             stitched.paste(im, (0, y_offset))
             y_offset += im.height
-            im.close()  # Close immediately after pasting
+            im.close()
+        del images
 
         # Draw caption only if page was trimmed
         if original_page_height > capture_height:
             draw = ImageDraw.Draw(stitched)
             caption_text = f"WARNING: Screenshot was {original_page_height}px but trimmed to {capture_height}px because it was too long"
             padding = 10
-
-            # Try to load font
             try:
                 font = ImageFont.truetype("arial.ttf", 35)
             except IOError:
@@ -120,23 +69,26 @@ def stitch_images_worker(pipe_conn, chunks_bytes, original_page_height, capture_
             bbox = draw.textbbox((0, 0), caption_text, font=font)
             text_width = bbox[2] - bbox[0]
             text_height = bbox[3] - bbox[1]
-
-            # Draw white background rectangle
             draw.rectangle([(0, 0), (max_width, text_height + 2 * padding)], fill=(255, 255, 255))
-
-            # Draw text centered
             text_x = (max_width - text_width) // 2
             draw.text((text_x, padding), caption_text, font=font, fill=(255, 0, 0))
 
-        # Encode and send image with optimization
+        # Encode and send
         output = io.BytesIO()
         stitched.save(output, format="JPEG", quality=int(os.getenv("SCREENSHOT_QUALITY", SCREENSHOT_DEFAULT_QUALITY)), optimize=True)
-        pipe_conn.send_bytes(output.getvalue())
+        result_bytes = output.getvalue()
 
         stitched.close()
+        del stitched
+        output.close()
+        del output
+
+        pipe_conn.send_bytes(result_bytes)
+        del result_bytes
+
     except Exception as e:
-        pipe_conn.send(f"error:{e}")
+        logger.error(f"Error in stitch_images_worker_raw_bytes: {e}")
+        error_msg = f"error:{e}".encode('utf-8')
+        pipe_conn.send_bytes(error_msg)
     finally:
         pipe_conn.close()
-
-

changedetectionio/flask_app.py

@@ -94,6 +94,14 @@ if os.getenv('FLASK_SERVER_NAME'):
 app.config['BABEL_TRANSLATION_DIRECTORIES'] = str(Path(__file__).parent / 'translations')
 app.config['BABEL_DEFAULT_LOCALE'] = 'en_GB'
 
+# Session configuration
+# NOTE: Flask session (for locale, etc.) is separate from Flask-Login's remember-me cookie
+# - Flask session stores data like session['locale'] in a signed cookie
+# - Flask-Login's remember=True creates a separate authentication cookie
+# - Setting PERMANENT_SESSION_LIFETIME controls how long the Flask session cookie lasts
+from datetime import timedelta
+app.config['PERMANENT_SESSION_LIFETIME'] = timedelta(days=3650)  # ~10 years (effectively unlimited)
+
 #app.config["EXPLAIN_TEMPLATE_LOADING"] = True
 
 
@@ -550,6 +558,9 @@ def changedetection_app(config=None, datastore_o=None):
 
         # Validate the locale against available languages
         if locale in language_codes:
+            # Make session permanent so language preference persists across browser sessions
+            # NOTE: This is the Flask session cookie (separate from Flask-Login's remember-me auth cookie)
+            session.permanent = True
             session['locale'] = locale
 
             # CRITICAL: Flask-Babel caches the locale in the request context (ctx.babel_locale)

changedetectionio/model/Watch.py

@@ -20,8 +20,9 @@ mtable = {'seconds': 1, 'minutes': 60, 'hours': 3600, 'days': 86400, 'weeks': 86
 
 def _brotli_save(contents, filepath, mode=None, fallback_uncompressed=False):
     """
-    Save compressed data using native brotli.
-    Testing shows no memory leak when using gc.collect() after compression.
+    Save compressed data using native brotli with streaming compression.
+    Uses chunked compression to minimize peak memory usage and malloc_trim()
+    to force release of C-level memory back to the OS.
 
     Args:
         contents: data to compress (str or bytes)
@@ -37,27 +38,52 @@ def _brotli_save(contents, filepath, mode=None, fallback_uncompressed=False):
     """
     import brotli
     import gc
+    import ctypes
 
     # Ensure contents are bytes
     if isinstance(contents, str):
         contents = contents.encode('utf-8')
 
     try:
-        logger.debug(f"Starting brotli compression of {len(contents)} bytes.")
+        original_size = len(contents)
+        logger.debug(f"Starting brotli streaming compression of {original_size} bytes.")
 
-        if mode is not None:
-            compressed_data = brotli.compress(contents, mode=mode)
-        else:
-            compressed_data = brotli.compress(contents)
+        # Create streaming compressor
+        compressor = brotli.Compressor(quality=6, mode=mode if mode is not None else brotli.MODE_GENERIC)
+
+        # Stream compress in chunks to minimize memory usage
+        chunk_size = 65536  # 64KB chunks
+        total_compressed_size = 0
 
         with open(filepath, 'wb') as f:
-            f.write(compressed_data)
+            # Process data in chunks
+            offset = 0
+            while offset < len(contents):
+                chunk = contents[offset:offset + chunk_size]
+                compressed_chunk = compressor.process(chunk)
+                if compressed_chunk:
+                    f.write(compressed_chunk)
+                    total_compressed_size += len(compressed_chunk)
+                offset += chunk_size
 
-        logger.debug(f"Finished brotli compression - From {len(contents)} to {len(compressed_data)} bytes.")
+            # Finalize compression - critical for proper cleanup
+            final_chunk = compressor.finish()
+            if final_chunk:
+                f.write(final_chunk)
+                total_compressed_size += len(final_chunk)
 
-        # Force garbage collection to prevent memory buildup
+        logger.debug(f"Finished brotli compression - From {original_size} to {total_compressed_size} bytes.")
+
+        # Cleanup: Delete compressor, force Python GC, then force C-level memory release
+        del compressor
         gc.collect()
 
+        # Force release of C-level memory back to OS (since brotli is a C library)
+        try:
+            ctypes.CDLL('libc.so.6').malloc_trim(0)
+        except Exception:
+            pass  # malloc_trim not available on all systems (e.g., macOS)
+
         return filepath
 
     except Exception as e:

changedetectionio/static/favicons/site.webmanifest

@@ -1,19 +1,25 @@
 {
-    "name": "",
-    "short_name": "",
+    "name": "ChangeDetection.io",
+    "short_name": "ChangeDetect",
+    "description": "Self-hosted website change detection and monitoring",
     "icons": [
         {
             "src": "android-chrome-192x192.png",
             "sizes": "192x192",
-            "type": "image/png"
+            "type": "image/png",
+            "purpose": "any maskable"
         },
         {
             "src": "android-chrome-256x256.png",
             "sizes": "256x256",
-            "type": "image/png"
+            "type": "image/png",
+            "purpose": "any maskable"
         }
     ],
-    "theme_color": "#ffffff",
+    "start_url": "/",
+    "theme_color": "#5bbad5",
     "background_color": "#ffffff",
-    "display": "standalone"
+    "display": "standalone",
+    "categories": ["utilities", "productivity"],
+    "orientation": "any"
 }

changedetectionio/templates/base.html

@@ -27,7 +27,7 @@
     <link rel="apple-touch-icon" sizes="180x180" href="{{url_for('static_content', group='favicons', filename='apple-touch-icon.png')}}">
     <link rel="icon" type="image/png" sizes="32x32" href="{{url_for('static_content', group='favicons', filename='favicon-32x32.png')}}">
     <link rel="icon" type="image/png" sizes="16x16" href="{{url_for('static_content', group='favicons', filename='favicon-16x16.png')}}">
-    <link rel="manifest" href="{{url_for('static_content', group='favicons', filename='site.webmanifest')}}">
+    <link rel="manifest" href="{{url_for('static_content', group='favicons', filename='site.webmanifest')}}" crossorigin="use-credentials">
     <link rel="mask-icon" href="{{url_for('static_content', group='favicons', filename='safari-pinned-tab.svg')}}" color="#5bbad5">
     <link rel="shortcut icon" href="{{url_for('static_content', group='favicons', filename='favicon.ico')}}">
     <meta name="msapplication-TileColor" content="#da532c">
@@ -265,6 +265,9 @@
           </a>
           {% endfor %}
         </div>
+        <div>
+            <a href="{{ url_for('ui.delete_locale_language_session_var_if_it_exists', redirect=request.path) }}" >{{ _('Auto-detect from browser') }}</a>
+        </div>
         <div>
             {{ _('Language support is in beta, please help us improve by opening a PR on GitHub with any updates.') }}
         </div>

changedetectionio/templates/menu.html

@@ -50,7 +50,9 @@
             <span class="visually-hidden">{{ _('Change language') }}</span>
             <span class="{{ get_flag_for_locale(get_locale()) }}" id="language-selector-flag"></span>
         </button>
-        <a class="github-link" href="https://github.com/dgtlmoon/changedetection.io">
+        <a class="github-link" href="https://github.com/dgtlmoon/changedetection.io"
+           target="_blank" 
+          rel="noopener" >
             {% include "svgs/github.svg" %}
         </a>
     </li>

changedetectionio/tests/test_i18n.py

@@ -160,7 +160,7 @@ def test_invalid_locale(client, live_server, measure_memory_usage, datastore_pat
 def test_language_persistence_in_session(client, live_server, measure_memory_usage, datastore_path):
     """
     Test that the language preference persists across multiple requests
-    within the same session.
+    within the same session, and that auto-detect properly clears the preference.
     """
 
     # Establish session cookie
@@ -184,6 +184,34 @@ def test_language_persistence_in_session(client, live_server, measure_memory_usa
         assert res.status_code == 200
         assert b"Annulla" in res.data, "Italian text should persist across requests"
 
+    # Verify locale is in session
+    with client.session_transaction() as sess:
+        assert sess.get('locale') == 'it', "Locale should be set in session"
+
+    # Call auto-detect to clear the locale
+    res = client.get(
+        url_for("ui.delete_locale_language_session_var_if_it_exists"),
+        follow_redirects=True
+    )
+
+    assert res.status_code == 200
+    # Verify the flash message appears (in English since we cleared the locale)
+    assert b"Language set to auto-detect from browser" in res.data, "Should show flash message"
+
+    # Verify locale was removed from session
+    with client.session_transaction() as sess:
+        assert 'locale' not in sess, "Locale should be removed from session after auto-detect"
+
+    # Now requests should use browser default (English in test environment)
+    res = client.get(
+        url_for("watchlist.index"),
+        follow_redirects=True
+    )
+
+    assert res.status_code == 200
+    assert b"Cancel" in res.data, "Should show English after auto-detect clears Italian"
+    assert b"Annulla" not in res.data, "Should not show Italian after auto-detect"
+
 
 def test_set_language_with_redirect(client, live_server, measure_memory_usage, datastore_path):
     """

changedetectionio/translations/README.md

@@ -0,0 +1,101 @@
+# Translation Guide
+
+## Updating Translations
+
+To maintain consistency and minimize unnecessary changes in translation files, run these commands:
+
+```bash
+python setup.py extract_messages   # Extract translatable strings
+python setup.py update_catalog     # Update all language files
+python setup.py compile_catalog    # Compile to binary .mo files
+```
+
+## Configuration
+
+All translation settings are configured in **`../../setup.cfg`** (single source of truth).
+
+The configuration below is shown for reference - **edit `setup.cfg` to change settings**:
+
+```ini
+[extract_messages]
+# Extract translatable strings from source code
+mapping_file = babel.cfg
+output_file = changedetectionio/translations/messages.pot
+input_paths = changedetectionio
+keywords = _ _l gettext
+# Options to reduce unnecessary changes in .pot files
+sort_by_file = true       # Keeps entries ordered by file path
+width = 120               # Consistent line width (prevents rewrapping)
+add_location = file       # Show file path only (not line numbers)
+
+[update_catalog]
+# Update existing .po files with new strings from .pot
+# Note: 'locale' is omitted - Babel auto-discovers all catalogs in output_dir
+input_file = changedetectionio/translations/messages.pot
+output_dir = changedetectionio/translations
+domain = messages
+# Options for consistent formatting
+width = 120               # Consistent line width
+no_fuzzy_matching = true  # Avoids incorrect automatic matches
+
+[compile_catalog]
+# Compile .po files to .mo binary format
+directory = changedetectionio/translations
+domain = messages
+```
+
+**Key formatting options:**
+- `sort_by_file = true` - Orders entries by file path (consistent ordering)
+- `width = 120` - Fixed line width prevents text rewrapping
+- `add_location = file` - Shows file path only, not line numbers (reduces git churn)
+- `no_fuzzy_matching = true` - Prevents incorrect automatic fuzzy matches
+
+## Why Use These Commands?
+
+Running pybabel commands directly without consistent options causes:
+- ❌ Entries get reordered differently each time
+- ❌ Text gets rewrapped at different widths
+- ❌ Line numbers change every edit (if not configured)
+- ❌ Large diffs that make code review difficult
+
+Using `python setup.py` commands ensures:
+- ✅ Consistent ordering (by file path, not alphabetically)
+- ✅ Consistent line width (120 characters, no rewrapping)
+- ✅ File-only locations (no line number churn)
+- ✅ No fuzzy matching (prevents incorrect auto-translations)
+- ✅ Minimal diffs (only actual changes show up)
+- ✅ Easier code review and git history
+
+These commands read settings from `../../setup.cfg` automatically.
+
+## Supported Languages
+
+- `cs` - Czech (Čeština)
+- `de` - German (Deutsch)
+- `en_GB` - English (UK)
+- `en_US` - English (US)
+- `fr` - French (Français)
+- `it` - Italian (Italiano)
+- `ko` - Korean (한국어)
+- `zh` - Chinese Simplified (中文简体)
+- `zh_Hant_TW` - Chinese Traditional (繁體中文)
+
+## Adding a New Language
+
+1. Initialize the new language catalog:
+   ```bash
+   pybabel init -i changedetectionio/translations/messages.pot -d changedetectionio/translations -l NEW_LANG_CODE
+   ```
+2. Compile it:
+   ```bash
+   python setup.py compile_catalog
+   ```
+
+Babel will auto-discover the new language on subsequent translation updates.
+
+## Translation Notes
+
+From CLAUDE.md:
+- Always use "monitor" or "watcher" terminology (not "clock")
+- Use the most brief wording suitable
+- When finding issues in one language, check ALL languages for the same issue

requirements.txt

@@ -91,7 +91,7 @@ jq~=1.3; python_version >= "3.8" and sys_platform == "linux"
 
 # playwright is installed at Dockerfile build time because it's not available on all platforms
 
-pyppeteer-ng==2.0.0rc11
+pyppeteer-ng==2.0.0rc12
 pyppeteerstealth>=0.0.4
 
 # Include pytest, so if theres a support issue we can ask them to run these tests on their setup

setup.cfg

@@ -0,0 +1,28 @@
+# Translation configuration for changedetection.io
+# See changedetectionio/translations/README.md for full documentation on updating translations
+
+[extract_messages]
+# Extract translatable strings from source code
+mapping_file = babel.cfg
+output_file = changedetectionio/translations/messages.pot
+input_paths = changedetectionio
+keywords = _ _l gettext
+# Options to reduce unnecessary changes in .pot files
+sort_by_file = true
+width = 120
+add_location = file
+
+[update_catalog]
+# Update existing .po files with new strings from .pot
+# Note: Omitting 'locale' makes Babel auto-discover all catalogs in output_dir
+input_file = changedetectionio/translations/messages.pot
+output_dir = changedetectionio/translations
+domain = messages
+# Options for consistent formatting
+width = 120
+no_fuzzy_matching = true
+
+[compile_catalog]
+# Compile .po files to .mo binary format
+directory = changedetectionio/translations
+domain = messages