The old test put the retained segments after 59 older unretained ones, so cleanup freed its hourly target and stopped before it ever reached them. It passed even with the retention check removed. Retained segments now come first, and cleanup needs every unretained segment after them, so it has to skip the retained ones to finish.
The locale load path was relative, so any URL with a trailing slash or more than one segment (like /admin/) fetched locales from under that path, got index.html back from the nginx fallback, and broke every UI string. Locales now load from the same base url as the API, which also covers HA ingress and FRIGATE_BASE_PATH.
Export thumbnails under /clips/export/ weren't classified by media auth, so roles with a camera allow-list got a 403 on every thumbnail. The thumbnail now resolves to its export's camera through a thumb_path lookup, and the camera access check applies as usual.
* don't crash review maintainer when a manual event ends with no active segment
The review maintainer read the label of every manual event message before checking its state, so an end message, which carries no label, raised a KeyError and killed the review process when the camera had no active segment. This happened when a recordings or enabled toggle, a completed manual event, or audio closed the segment of a still-open indefinite event, and on every plate expiry on dedicated LPR cameras with detections disabled. With the process dead, no review items were created and alert/detection-only retention deleted all new recordings until restart. End messages with no active segment now drop the event from the indefinite events and are otherwise ignored.
* Update the default timeout for detection subscriber to avoid blocking indefinitely
* run embeddings requests off the api event loop
set_description, regenerate_review_description, and the chat semantic search tools called the embeddings requestor directly from async handlers, so the API event loop waited on the embeddings process with no timeout. In normal operation the wait is short (about 80 to 140 ms of text embedding with Jina V1 on CPU, and only queue time for review regeneration), so this is mostly defensive. The cases it guards against are a request sent while the embeddings process is still downloading or loading models, which froze the whole UI until it finished, and the embeddings process dying with a request in flight, which left the API unresponsive until Frigate restarted. The requestor calls now run in asyncio.to_thread, matching the other blocking calls in the API.
* Handle other fields for context size from api providers
* fix stream checks missing non-aac audio
The ffprobe API never asked ffprobe for codec_type, so the stream checks could only spot audio and video by codec name. G.711, Opus, and other non-AAC audio showed as no audio, and H.265 detect streams showed an unknown resolution. ffprobe_stream now requests codec_type, and the stream check matches tracks on it alone.
* fix timezone-dependent http tests
The devcontainer mounts the host's /etc/localtime over zoneinfo/Etc/UTC, so the API's ZoneInfo("UTC") used the host's local zone and the recordings summary and event time_range tests failed at any hour. The review summary tests also built their expected day key from the local date, so they failed whenever the local day differed from the UTC day. BaseTestHttp now writes a minimal UTC zone file to a temp dir and puts it first on the zoneinfo search path, which works without the tzdata package, and the review tests convert their expected day keys to UTC.
---------
Co-authored-by: Nicolas Mowen <nickmowen213@gmail.com>
* Add Gemini embeddings for semantic search
gemini-embedding-2 maps text and images into one space. Request 768 dimensions, send one input per call, and transcode WebP thumbnails to JPEG. Check model supported_actions for embedContent before advertising embeddings support, and raise ValueError on failure so the embeddings process stays up.
* Fix unreachable Gemini provider checks
* update nginx to 1.30.7 and backport nginx module changes from dev
nginx 1.27.4 was vulnerable to CVE-2026-123007 (HTTP/2 flow control heap overflow) and CVE-2026-123633 (HTTP/2 use-after-free), both reachable before auth on the internal and external ports because http2 is on for the whole server block. Kaltura also deleted the nginx-vod-module and nginx-secure-token-module repos, so the build could no longer download either module.
- Update nginx to 1.30.7
- Backport the switch to the maintained dio-az nginx-vod-module fork (#24123)
- Backport the nginx-secure-token-module removal (#24893)
* bump playwright to 1.65.0 for ubuntu 26.04 runners
* tweak agents file
* add serbian latin
* fix serbian latin components
weblate put them in the wrong directory (sr_Latn). the component style was updated in weblate itself so this shouldn't happen for any future translations
* raise on empty genai embedding response so the maintainer thread survives
* add kebab menu to export cases
* disable download while case exports are still encoding
* stop case menu clicks from opening the case
* check case download readiness against jobs on every camera
Currently translated at 100.0% (64 of 64 strings)
Translated using Weblate (Serbian (Latin script))
Currently translated at 100.0% (46 of 46 strings)
Translated using Weblate (Serbian (Latin script))
Currently translated at 100.0% (60 of 60 strings)
Translated using Weblate (Serbian (Latin script))
Currently translated at 100.0% (66 of 66 strings)
Translated using Weblate (Serbian (Latin script))
Currently translated at 100.0% (129 of 129 strings)
Translated using Weblate (Serbian (Latin script))
Currently translated at 100.0% (500 of 500 strings)
Translated using Weblate (Serbian (Latin script))
Currently translated at 100.0% (775 of 775 strings)
Translated using Weblate (Serbian (Latin script))
Currently translated at 100.0% (522 of 522 strings)
Translated using Weblate (Serbian (Latin script))
Currently translated at 100.0% (74 of 74 strings)
Translated using Weblate (Serbian (Latin script))
Currently translated at 100.0% (1409 of 1409 strings)
Translated using Weblate (Serbian (Latin script))
Currently translated at 100.0% (49 of 49 strings)
Translated using Weblate (Serbian (Latin script))
Currently translated at 100.0% (130 of 130 strings)
Translated using Weblate (Serbian (Latin script))
Currently translated at 100.0% (145 of 145 strings)
Translated using Weblate (Serbian (Latin script))
Currently translated at 100.0% (775 of 775 strings)
Translated using Weblate (Serbian (Latin script))
Currently translated at 100.0% (522 of 522 strings)
Translated using Weblate (Serbian (Latin script))
Currently translated at 100.0% (37 of 37 strings)
Translated using Weblate (Serbian (Latin script))
Currently translated at 100.0% (74 of 74 strings)
Translated using Weblate (Serbian (Latin script))
Currently translated at 100.0% (282 of 282 strings)
Translated using Weblate (Serbian (Latin script))
Currently translated at 100.0% (1409 of 1409 strings)
Translated using Weblate (Serbian (Latin script))
Currently translated at 100.0% (46 of 46 strings)
Translated using Weblate (Serbian (Latin script))
Currently translated at 100.0% (64 of 64 strings)
Translated using Weblate (Serbian (Latin script))
Currently translated at 100.0% (130 of 130 strings)
Translated using Weblate (Serbian (Latin script))
Currently translated at 100.0% (68 of 68 strings)
Translated using Weblate (Serbian (Latin script))
Currently translated at 100.0% (141 of 141 strings)
Translated using Weblate (Serbian (Latin script))
Currently translated at 100.0% (775 of 775 strings)
Translated using Weblate (Serbian (Latin script))
Currently translated at 100.0% (522 of 522 strings)
Translated using Weblate (Serbian (Latin script))
Currently translated at 100.0% (37 of 37 strings)
Translated using Weblate (Serbian (Latin script))
Currently translated at 100.0% (74 of 74 strings)
Translated using Weblate (Serbian (Latin script))
Currently translated at 100.0% (120 of 120 strings)
Translated using Weblate (Serbian (Latin script))
Currently translated at 100.0% (50 of 50 strings)
Translated using Weblate (Serbian (Latin script))
Currently translated at 100.0% (266 of 266 strings)
Translated using Weblate (Serbian (Latin script))
Currently translated at 100.0% (282 of 282 strings)
Translated using Weblate (Serbian (Latin script))
Currently translated at 67.1% (946 of 1409 strings)
Translated using Weblate (Serbian (Latin script))
Currently translated at 100.0% (130 of 130 strings)
Translated using Weblate (Serbian (Latin script))
Currently translated at 100.0% (68 of 68 strings)
Translated using Weblate (Serbian (Latin script))
Currently translated at 100.0% (129 of 129 strings)
Translated using Weblate (Serbian (Latin script))
Currently translated at 100.0% (500 of 500 strings)
Co-authored-by: Branislav Maricic <b.n.maricic@users.noreply.hosted.weblate.org>
Co-authored-by: Hosted Weblate <hosted@weblate.org>
Translate-URL: https://hosted.weblate.org/projects/frigate-nvr/audio/sr_Latn/
Translate-URL: https://hosted.weblate.org/projects/frigate-nvr/common/sr_Latn/
Translate-URL: https://hosted.weblate.org/projects/frigate-nvr/components-camera/sr_Latn/
Translate-URL: https://hosted.weblate.org/projects/frigate-nvr/components-dialog/sr_Latn/
Translate-URL: https://hosted.weblate.org/projects/frigate-nvr/components-filter/sr_Latn/
Translate-URL: https://hosted.weblate.org/projects/frigate-nvr/components-player/sr_Latn/
Translate-URL: https://hosted.weblate.org/projects/frigate-nvr/config-cameras/sr_Latn/
Translate-URL: https://hosted.weblate.org/projects/frigate-nvr/config-global/sr_Latn/
Translate-URL: https://hosted.weblate.org/projects/frigate-nvr/objects/sr_Latn/
Translate-URL: https://hosted.weblate.org/projects/frigate-nvr/views-chat/sr_Latn/
Translate-URL: https://hosted.weblate.org/projects/frigate-nvr/views-classificationmodel/sr_Latn/
Translate-URL: https://hosted.weblate.org/projects/frigate-nvr/views-events/sr_Latn/
Translate-URL: https://hosted.weblate.org/projects/frigate-nvr/views-explore/sr_Latn/
Translate-URL: https://hosted.weblate.org/projects/frigate-nvr/views-facelibrary/sr_Latn/
Translate-URL: https://hosted.weblate.org/projects/frigate-nvr/views-live/sr_Latn/
Translate-URL: https://hosted.weblate.org/projects/frigate-nvr/views-motionsearch/sr_Latn/
Translate-URL: https://hosted.weblate.org/projects/frigate-nvr/views-replay/sr_Latn/
Translate-URL: https://hosted.weblate.org/projects/frigate-nvr/views-search/sr_Latn/
Translate-URL: https://hosted.weblate.org/projects/frigate-nvr/views-settings/sr_Latn/
Translate-URL: https://hosted.weblate.org/projects/frigate-nvr/views-system/sr_Latn/
Translation: Frigate NVR/Config - Cameras
Translation: Frigate NVR/Config - Global
Translation: Frigate NVR/audio
Translation: Frigate NVR/common
Translation: Frigate NVR/components-camera
Translation: Frigate NVR/components-dialog
Translation: Frigate NVR/components-filter
Translation: Frigate NVR/components-player
Translation: Frigate NVR/objects
Translation: Frigate NVR/views-chat
Translation: Frigate NVR/views-classificationmodel
Translation: Frigate NVR/views-events
Translation: Frigate NVR/views-explore
Translation: Frigate NVR/views-facelibrary
Translation: Frigate NVR/views-live
Translation: Frigate NVR/views-motionSearch
Translation: Frigate NVR/views-replay
Translation: Frigate NVR/views-search
Translation: Frigate NVR/views-settings
Translation: Frigate NVR/views-system
* Resize classification training images the same way as inference
Training let Keras resize dataset images to 224x224, which uses nearest
neighbor by default. Inference resizes the crop with cv2.resize
(bilinear). For crops much smaller than 224 px the two results look very
different, so a model could fit its training images perfectly and still
report the same state for every live frame.
Add resize_classification_image() and use it everywhere an image is
resized for a classification model. Before training, the dataset is
copied to a temporary directory through that same function, so Keras is
given images that are already the model input size and does no resizing
of its own.
The new test runs a frame through the state processor and the crop it
saves through the training process (stopped at model.fit), and checks
the model receives the same pixels from both, for random frame and crop
sizes. Adds hypothesis as a dev dependency.
Co-Authored-By: Claude Fable 5.2 <noreply@anthropic.com>
Co-Authored-By: Claude Opus 5.5 <noreply@anthropic.com>
* Fail training when a prepared image cannot be written
cv2.imwrite returns False instead of raising, so a full temporary
directory would have dropped images from training without any error.
Raise with the file path instead; the training process already logs
the exception and fails.
Co-Authored-By: Claude Opus 5.5 <noreply@anthropic.com>
* Trim classification parity tests and drop hypothesis
Per review: the parity test is now a plain loop over fixed crop sizes
(64, 79, 224 and 600 px) on a fixed 1280x720 frame, which covers both
upscaled and downscaled crops. The 224x224 check and the failed-write
test are removed; the parity test already covers the prepared image
size. hypothesis is no longer a dev dependency.
The 64, 79 and 600 px cases still fail on dev and pass with the fix.
Co-Authored-By: Claude Opus 5.5 <noreply@anthropic.com>
---------
Co-authored-by: Victor Bajanov <victorb@spunmonkey.org>
Co-authored-by: Claude Fable 5.2 <noreply@anthropic.com>
* only remove the deleted camera's media files
Camera deletion globbed snapshot, clean copy, and review thumbnail files by "{camera}-*", and camera names can contain "-", so deleting "front" also removed every one of those files for "front-door". Files are now matched on the full name, the camera followed by an event or review id.
* only use the camera's own preview frames for review descriptions
get_cache_frames matched cached preview frames by "preview_{camera}-", so when "front" had no frame after the review ended, the first "front-door" frame became the trailing frame and went into front's GenAI prompt. It now uses is_camera_preview_frame like the preview recorder and exports.
* keep cleaning up when a media directory can't be listed
* Update third_party_extensions.md with new integrations
Added new integrations for Auto Arm and Camera AutoNotify to the documentation.
* Enhance integration descriptions for clarity
Updated descriptions for Auto Arm and Camera AutoNotify integrations to include additional features and synchronization options.
go2rtc merges its config files, and the HomeKit file and anything saved through go2rtc's unauthenticated config API load before Frigate's generated config. A planted exec: or echo: source, or an ngrok.command, ran whatever program it named, and as root under FRIGATE_ROOT_SERVICES=go2rtc. Unless GO2RTC_ALLOW_ARBITRARY_EXEC is set, the generated config now sets exec.allow_paths and echo.allow_paths to Frigate's ffmpeg and sets an empty ngrok.command, which override earlier files, and the config-save API rejects new exec and echo keys. The non-root docs note that a root go2rtc still isn't separated from the frigate user, since ffmpeg with arbitrary arguments can overwrite files as root.
* disable onnxruntime telemetry
* hold the last frame on screen when the live mode changes
Turning on two-way talk while watching MSE switched the player to WebRTC and showed a buffering spinner until WebRTC played. The frame capture now runs on a mode change as well as a stream change, so the outgoing frame stays up until the new player plays.
Mobile has no equivalent of the Ctrl+A shortcut, so selecting many
review items means long-pressing one and tapping every other card.
Show an "All" control next to "Unselect" on mobile while some items
are still unselected, matching Explore, Face Library, and the
classification training view.
The guide said a 320 model gets smaller regions than a 640 model. Since 0.17 the minimum region is 320 px for any model, so both models get the same crops and a 640 model usually receives an upscaled one.
* analytics notifications flag and device grant warnings
- report notifications as on when any camera enables them, not only the global setting
- don't warn about device nodes the runtime users can already use, such as world-rw nodes in an unprivileged LXC
* clarify which stream the All Cameras dashboard uses
* add exception catchall for autotracking calibration
* only default review api after to 24 hours when no bounds are given
The review endpoint replaced a missing or zero `after` with 24 hours ago, so a `before`-only query older than 24 hours returned nothing and `after=0` couldn't mean no lower bound. The default now applies only when neither `before` nor `after` is sent, a `before`-only query has no lower bound, and both bounds use `is not None` checks. The Live dashboard alerts bar sends no bounds and keeps the 24 hour window.
* hold the last frame on screen while switching live streams
Switching streams in the single camera view blanked the player until the new stream played, because the old player unmounts in the same render and the hidden new player gave the card no height. The outgoing video's last frame is now drawn to a canvas before the remount and covers the gap until the new stream plays, and the card takes the camera's aspect ratio so it keeps its size while the video is hidden.
* show live streams as a one-tap grid in the mobile drawer
Switching streams on mobile meant scrolling past the camera toggles, opening a dropdown, and picking a stream. The drawer now opens with the stream section, and streams are a wrapping grid of toggles sized to their names, so a switch is two taps. A single stream shows as plain text. The audio and two-way talk lines show a spinner while the stream's go2rtc metadata loads instead of reporting audio as unavailable, which showed on the first switch to a transcoded stream.
* don't capture the switch frame on tiles that show the camera still
* allow switching the detect scene of a debug replay camera without a restart
- show the scene field in the replay detect settings
- drop the replay camera's cached model on a detect change so the refresh recycle picks up the new scene's model
- recreate the tracked object state on refresh when the camera's model changed
- reject a scene no model can serve with a 400 instead of crashing the tracker and maintainer threads
- size each camera's detection input shm for the largest model and recreate leftover smaller segments, so a camera re-added under a larger model no longer overflows it
- ignore event ends for objects already dropped by a state reset
* end tracked events before replacing replay camera state on a scene switch
- test the replay model rebind against two configured models
* move genai review frame extraction off the embeddings loop
- Run frame gathering for finished review items on a thread so per-frame ffmpeg calls don't stall the embeddings maintainer
- Share one path between live and regenerated review descriptions
- Stop padding review start and end times in place, which widened the stored review item
- Lock timestamp expiry in EventsPerSecond since update() and eps() now run on different threads
* serialize recording frame extraction for review descriptions
* Validate MQTT TLS verification options before client startup
* Test MQTT configuration recovery with authentication available
* Cover both MQTT TLS verification values in safe-mode recovery
* Use a documentation address in MQTT regression tests
* Skip MQTT TLS validation when MQTT is disabled
* Simplify MQTT TLS validation regression coverage
---------
Co-authored-by: Peter Kay <dougle.peter.kay@googlemail.com>
* pin debug replay to one recording stream
Debug replay stream-copied the merged VOD playlist into one mp4. When main had gaps that sub filled, the clip switched streams mid-file, and a camera with an H.265 main and an H.264 sub produced an mp4 labeled H.264 that also held HEVC packets. The replay camera's decoder then failed and the watchdog killed ffmpeg. Replay now follows the clip download rule: it pins the playlist to main and only falls back to sub when the range has no main recordings.
* choose the replay stream with the vod coverage resolver
Stream selection counted a main row that only touched the range, while the pinned playlist served nothing for it, so replay failed instead of using the sub footage. Selection now uses resolve_coverage, the same resolver the pinned vod and export use.
* fix sub label filter for non-ascii names
The sub_labels filter lowercased the requested name in Python but the stored value with SQLite's LOWER(), which only folds ASCII. Cyrillic and other non-ASCII names never matched, so filtering by a face name returned no events. Both sides are now lowered in SQLite, which keeps ASCII matching case-insensitive and non-ASCII matching exact.
* cover events/search in sub label filter test