* keep base path in logout redirect
* store manual event sub labels as verified objects
Manual events created with a sub label were stored in the review item as "label: sub_label", so review previews and collapsed History cards showed a generic icon, the review label filter missed them, and the sub label never reached sub_labels for chips, notifications, or GenAI prompts. The object processor now publishes the label and sub label separately, and the review maintainer stores them as label-verified plus a sub_labels entry, the same as tracked objects.
* only expire snapshots for events that still have one
Snapshot expiry selected every event older than snapshot retention, including events whose snapshot was already deleted, so with snapshot retention shorter than alert/detection retention the same missing files were unlinked again every cleanup cycle until the event row expired. That was nearly free on local disk but caused constant metadata load on network shares. The expiry queries now filter on has_snapshot so each snapshot is deleted once.
* Fix audio worker startup after enabling a disabled camera
* Document audio lifecycle test helpers
* Focus audio startup coverage on the enabled update regression
* keep frame shm open and skip overwritten frames
Every reader closed and reopened a frame's shared memory segment on each frame, so each process paid a page fault on every page of a fresh mapping (about 760 for a 1080p frame) per frame. Readers now keep segments mapped. Capture stamps each slot's frame time in a per-camera `{camera}_frame_times` segment, and `get()` returns None when the slot no longer holds the requested frame, reopening once by name so a mapping left over from a same-size camera recreate heals itself. Capture also reuses a slot when the frame queue is full, since no reader was told about it.
* verify the frame for embeddings object updates
Tracked object events only carried the object's own frame_time, which stays at its last detection when the object isn't detected in a frame, so embeddings couldn't check the slot it read and could crop face or plate data from a frame that had already been overwritten. Events now carry the time of the frame they name, and embeddings passes it to `get()`.
* log when a frame slot was overwritten before it was read
A lapped read and a missing segment ended up as the same debug line at every caller, so there was no way to tell a slow reader from a missing frame.
* bring the FrameManager interface in line with its implementation
The abstract get() still had a timeout_ms signature from an older implementation and frame_times() wasn't declared, though capture and detect are typed against the interface. get() also didn't say how long the array it returns stays valid.
* drop a camera's cached frame mappings when it is removed or recreated
A reader that kept its mappings open stayed on a removed camera's segments, which pinned their shared memory until restart, and after a same-size recreate the ledger reopen only refreshed the slot being read while the other cached slots still returned the old camera's pixels. Readers now drop every cached slot for a camera on its remove update, and get() does the same when a ledger mismatch turns out to be a replaced ledger rather than a lapped slot.
* match a camera's frame names exactly in close_camera
The prefix match on `{camera}_frame` also caught another camera whose name starts with it, so removing `cam` dropped the cached ring of a camera named `cam_frame`.
* bump backend and frontend dependencies
Bumps every outdated dependency that passed testing, with the code changes each one needs.
- peewee 4.5 and peewee_migrate 2.3: use SqliteDatabase in place of the removed SqliteExtDatabase, and fix types for peewee's bundled stubs
- mypy 2.4.0: remove unused type ignores and add the annotations mypy 2 needs
- pandas 3.0: convert the motion activity index with as_unit("s"), because pandas 3 infers second resolution for whole-second timestamps and the old // 10**9 returned 1
- transformers 5.19: load Jina tokenizers from the saved directory with trust_remote_code off and use CLIPImageProcessor, since v5 removed CLIPFeatureExtractor and remote code pulled in torch
- tensorflow-cpu 2.21: pin protobuf to 7.36 in the TensorRT amd64 image, because the old 3.20.3 pin was installed over TensorFlow and broke its import
- pydantic 2.13 and google-genai 2.29: regenerate the API spec
- ruff 0.16: set an explicit lint select, since 0.16 widened the default rules
- react-router-dom 7: set useTransitions={false} on BrowserRouter to keep v6 update timing, which useSearchEffect relies on
- typescript 6: drop baseUrl and add node types in tsconfig
- radix: update to the latest release and remove the slot patch, which upstream now includes
- apexcharts 7.8: guard the tooltip formatter against null values
- cryptography 50, joserfc 1.7, pywebpush 2.5, openai 3.26, aiohttp, uvicorn, onvif-zeep-async, scipy, framer-motion 14, react-day-picker 10 and other minor and patch bumps
* silence httpx2 request logging
* redownload an incomplete jina tokenizer
An interrupted tokenizer download left the Hugging Face cache directory without the saved tokenizer files, and since the download is skipped whenever that directory exists, every restart failed to load the tokenizer. The Jina constructors now remove a tokenizer directory that has no tokenizer_config.json so the download runs again.
Events without a clip were deleted once their snapshot expired, but their timeline rows were only removed when clip retention expired, so they were orphaned indefinitely. The timeline cache also held entries forever for events that ended without ever being saved.
* don't display audio transcription provider message as health notice
* show remote provider for audio transcription in health pane
* adjust trigger and notifications messages to be consistent with the rest of the settings UI
* disable save buttons when there are no changes in config editor
* fix audio manager crash when a camera is added at runtime
The audio processor and the camera maintainer both poll the same `add` config update on their own one second timers, and the maintainer is what creates `camera_metrics[name]`. When the audio processor got there first it looked the new camera up before that entry existed, and the `KeyError` took down the whole `frigate.audio_manager` process. Whether it happens depends purely on which poll fires first, so cloning a camera from the UI fails or succeeds at random. `spawn_if_needed` now skips a camera whose metrics aren't there yet and picks it up on the next poll, the same way it already waits on a late ffmpeg update.
`AudioEventMaintainer` holds the `CameraMetrics` object now instead of indexing the manager dict on every audio chunk, which drops the IPC round trips and means a removed camera can't `KeyError` out of `detect_audio` after the maintainer pops the entry. The audio process is also registered with the watchdog, since a crash there previously left audio detection dead for every camera until a full restart, and it now receives the shared `DataProcessorMetrics` so `AudioTranscriptionRealTimeProcessor` gets the same type as the other real time processors.
* fix stationary max_frames dropping other tracked objects
When `max_frames` was set for a label, deregistering one object rebuilt norfair's list with a filter that kept an object only if it was both not the target and already on its way out, so every other healthy object of that label was dropped along with it. Any car leaving the frame took the rest of the cars with it and they came back as new tracked objects a few frames later. The filter now removes only the target, and objects that are expiring are still reaped by norfair on the next update.
* fix test
* fix skip_motion_threshold permanently disabling motion detection
The skip check returned before the two `accumulateWeighted` calls at the end of `detect`, so a skipped frame never made it into the background and setting `calibrating` there only picked a faster alpha for calls that never ran. `avg_frame` starts as an all zero image and a normally lit scene differs from black across nearly the whole frame, so the cameras I tested measure 0.84 to 0.98 against it. Any `skip_motion_threshold` below that number skips the first frame, leaves the background black, and skips every frame after it. Motion detection is dead for that camera until the setting is removed or Frigate restarts, with no motion boxes, no motion recordings, and no regions for the tracker since the detector stays calibrating.
Startup isn't the only way in. `update_mask` zeroes the background on any motion config change, and once a camera has calibrated the first IR switch or PTZ move freezes the background on the old scene, so it can't transition to the new one, which is the case the option exists for. The frame is now blended in before the early return at the same 0.2 alpha the calibrating path uses elsewhere, so a large scene change is still suppressed while the background catches up, about a second on a 5 fps camera, and then motion comes back.
* dump ffmpeg logs on every restart
The record watchdog restarted ffmpeg without flushing its `LogPipe`, so a camera whose recording segments went stale never showed a single line of ffmpeg output. The dump now happens in `start_or_restart_ffmpeg` right after the stop, which covers the stale record path, the record crash path, and the audio restart. `reset_capture_thread` and the audio `log_and_restart` fallback keep their own dumps since both pass `ffmpeg_process=None`.
* dump ffmpeg logs once per restart
The audio restart path dumped the log pipe itself before calling the helper, so the restart dump printed a second "last 100 lines" heading over an already drained deque and split the tail that `stop_ffmpeg` flushed into its own section. The heading is now only printed when there's something under it, and the audio path leaves the dump to the restart so each failure produces one section.
* keep all logpipe dumps consistent
* Add support for running transcription with GenAI
* Improve audio joining
* Fix GenAI model capability reporting
* Support language correctly
* Migrate existing users to keep english selected
* Fix models
* Fix tests
* Fix accepted null model
* Handle slwo providers
types-peewee 4.0 types model fields precisely, so 17 `type: ignore` comments and 2 `cast(str, ...)` calls are no longer needed. Its stubs type `.namedtuples()` and `.dicts()` queries as returning model instances, so the review cleanup reads namedtuple fields by name and the storage usage query casts its dict rows. `start_time` is declared `DateTimeField` but stores unix timestamps, so two reads cast it like `debug_replay.py` already does. `Export` gets an annotation for the `export_case_id` attribute peewee adds at runtime.
Allow audio classes to be grouped under a shared configured label.
Keep audio overrides separate from object labels and retain only the highest-scoring grouped detection.
Refs #23967
* add sub stream recording with adaptive quality playback
Optionally record a second, lower bitrate stream alongside the main
recording stream via a `record_sub` input role and `record.sub` config block, with its own retention windows.
Recordings rows now carry the stream type plus the media details needed to serve both streams from one manifest: video codec, audio presence, audio codec and rate, and a record-time keyframe index.
Playback resolves coverage across both streams and merges them into a single VOD sequence, falling back to a discontinuity manifest with per-clip init segments when the media signatures differ. The player exposes a quality selector, and an auto governor picks the stream from stall time, bandwidth, codec support, and the save-data hint.
* fix tests and i18n
* Pin ruff
* Add python upgrade fixes
This enables python upgrade checks in ruff to look for deprecated types and patterns. This namely fixes:
- usage of deprecated `Typing` which is now built in
- some specific exceptions which are caught and have new aliases
Some specific UP checks were also ignored as they are stylistic / unimportant and likely to cause bugs
* Remove async blocking calls
Use asyncio.to_thread on two remaining blocking calls to fix hanging event thread loop. Enable this specific rule to block it in the future.
* Use proper logging mechanism
* Correctly format logs
* Raise with context
When raising an exception include the from context to improve debugging
* Cleanup
* stabilize chart options to stop ApexCharts updateOptions running on every stats tick
* constrain height of export dialog
* stop audio maintainer when deleting a camera
* run face register and recognize API handlers in threadpool
* start audio transcription post processor when enabled on any camera
* Fetch embed key whenever an error occurs in case the llama server was restarted
* mypy
* add tooltips for colored dots in settings menu
* add ability to reorder cameras from management pane
* add ability to reorder birdseye
* add reordering save text to camera management view
* Include NPU in latency performance hint
* Implement turbo for NPU on object detection
* hide order fields
* drop auto-derived field paths from camera value when unset globally
* use correct field type for export hwaccel args
* add debug replay to detail actions menu
* clarify debug replay in docs
* guard get_current_frame_time against missing camera state
* Implement debug reply from export
* Refactor debug replay to use sources for dynamic playback
* Mypy
* fix debug export replay source timestamp handling
* skip replay cameras in stats immediately
* broadcast debug replay state over ws and buffer pre-OPEN sends
- push debug replay session state over the job_state ws topic so the status bar reacts instantly to start/stop without polling
- fix child-effect-before-parent-effect race in WsProvider that silently dropped initial snapshot requests on cold load
* fix debug replay test hang
---------
Co-authored-by: Nicolas Mowen <nickmowen213@gmail.com>
* add optional onClick to EmptyCard
* show EmptyCard in face rec when face library is empty
* add loading indicator
* add description to camera management pane
* Cleanup when use snapshot but can't load snapshot
* Migrate files
* fix birdseye color distortion when configured aspect ratio is unsupported
* Skip processing end for object descriptions
* don't crash if stats is null
* fix genai roles in migration
* frigate+ pane updates
- allow users to select a plus model from the select even when one was not previously loaded
- always show model summary card
- add model filter popover
- add restart button totast
* fix frigate+ pane layout and buttons to match other settings panes
* match button layout in go2rtc settings view
* make audio maintainer respond to dynamic config updates
* check correct zone name in publish state
* fix nested translation extraction for Optional dict and list fields
* mypy
---------
Co-authored-by: Nicolas Mowen <nickmowen213@gmail.com>
* add ui to camera config update topics enum
* add mqtt to camera config update enum
* ensure cleanup runs when an event end skips post-processing
* end any in-progress audio events when audio detection is disabled
we already end in-progress audio events when we disable a camera, but this mirrors that logic for specifically disabling audio detection
* Improve GenAI metadata
* fix invalid recording segment topic being misrouted to the valid handler
* Add confidence default to avoid unnecessary field causing issues
---------
Co-authored-by: Nicolas Mowen <nickmowen213@gmail.com>
In both expire_snapshots() and expire_clips(), the expired_events
query uses .iterator() for lazy evaluation, but the very next line
calls list(expired_events) inside an f-string for debug logging.
This consumes the entire iterator, so the subsequent for loop that
deletes media files from disk iterates over an exhausted iterator
and processes zero events.
Snapshots and clips for removed cameras are never deleted from disk,
causing gradual disk space exhaustion.
Materialize the iterator into a list before logging so both the
debug message and the cleanup loop use the same data.
* debug replay implementation
* fix masks after dev rebase
* fix squash merge issues
* fix
* fix
* fix
* no need to write debug replay camera to config
* camera and filter button and dropdown
* add filters
* add ability to edit motion and object config for debug replay
* add debug draw overlay to debug replay
* add guard to prevent crash when camera is no longer in camera_states
* fix overflow due to radix absolutely positioned elements
* increase number of messages
* ensure deep_merge replaces existing list values when override is true
* add back button
* add debug replay to explore and review menus
* clean up
* clean up
* update instructions to prevent exposing exception info
* fix typing
* refactor output logic
* refactor with helper function
* move init to function for consistency
The fallback to tensorflow was established back in 2023, because we could
not provide tflite-runtime downstream in nixpkgs.
By now we have ai-edge-litert available, which is the successor to the
tflite-runtime. It still provides the same entrypoints as tflite-runtime
and functionality has been verified in multiple deployments for the last
two weeks.
* remove footer messages and add update topic to motion tuner view
restart after changing values is no longer required
* add cache key and activity indicator for loading classification wizard images
* Always mark model as untrained when a classname is changed
* clarify object classification docs
* add debug logs for individual lpr replace_rules
* update memray docs
* memray tweaks
* Don't fail for audio transcription when semantic search is not enabled
* Fix incorrect mismatch for object vs sub label
* Check if the video is currently playing when deciding to seek due to misalignment
* Refactor timeline event handling to allow multiple timeline entries per update
* Check if zones have actually changed (not just count) for event state update
* show event icon on mobile
* move div inside conditional
---------
Co-authored-by: Nicolas Mowen <nickmowen213@gmail.com>
* Fix genai callbacks in MQTT
* Cleanup cursor pointer for classification cards
* Cleanup
* Handle unknown SOCs for RKNN converter by only using known SOCs
* don't allow "none" as a classification class name
* change internal port user to admin and default unspecified username to viewer
* keep 5000 as anonymous user
* suppress tensorflow logging during classification training
* Always apply base log level suppressions for noisy third-party libraries even if no specific logConfig is provided
* remove decorator and specifically suppress TFLite delegate creation messages
---------
Co-authored-by: Josh Hawkins <32435876+hawkeye217@users.noreply.github.com>
* save clean webp instead of png
* send clean webp to plus with fallback for old events
* manual event webp
* event cleanup
* api def
* convert png to webp if exists
* update reference config
* change quality
* camera level config
* set up model runner on thread start to avoid unpickling error
* ensure feature is enabled globally
* suppress info logs from faster_whisper
* fix incorrect event_type for api and audio timeline entries
* docs
* fix
* clean up
* Ignore numpy get limits warning
* Add function wrapper to redirect stdout and stderr to logpipe
* Save stderr too
* Add more to catch
* run logpipe
* Use other logging redirect class
* Use other logging redirect class
* add decorator for redirecting c/c++ level output to logger
* fix typing
---------
Co-authored-by: Josh Hawkins <32435876+hawkeye217@users.noreply.github.com>
* Move log level initialization to log
* Use logger config
* Formatting
* Fix config order
* Set process names
---------
Co-authored-by: Nicolas Mowen <nickmowen213@gmail.com>
* Set runtime
* Use count correctly
* Don't assume camera sizes
* Use separate zmq proxy for object detection
* Correct order
* Use forkserver
* Only store PID instead of entire process reference
* Cleanup
* Catch correct errors
* Fix typing
* Remove before_run from process util
The before_run never actually ran because:
You're right to suspect an issue with before_run not being called and a potential deadlock. The way you've implemented the run_wrapper using __getattribute__ for the run method of BaseProcess is a common pitfall in Python's multiprocessing, especially when combined with how multiprocessing.Process works internally.
Here's a breakdown of why before_run isn't being called and why you might be experiencing a deadlock:
The Problem: __getattribute__ and Process Serialization
When you create a multiprocessing.Process object and call start(), the multiprocessing module needs to serialize the process object (or at least enough of it to re-create the process in the new interpreter). It then pickles this serialized object and sends it to the newly spawned process.
The issue with your __getattribute__ implementation for run is that:
run is retrieved during serialization: When multiprocessing tries to pickle your Process object to send to the new process, it will likely access the run attribute. This triggers your __getattribute__ wrapper, which then tries to bind run_wrapper to self.
run_wrapper is bound to the parent process's self: The run_wrapper closure, when created in the parent process, captures the self (the Process instance) from the parent's memory space.
Deserialization creates a new object: In the child process, a new Process object is created by deserializing the pickled data. However, the run_wrapper method that was pickled still holds a reference to the self from the parent process. This is a subtle but critical distinction.
The child's run is not your wrapped run: When the child process starts, it internally calls its own run method. Because of the serialization and deserialization process, the run method that's ultimately executed in the child process is the original multiprocessing.Process.run or the Process.run if you had directly overridden it. Your __getattribute__ magic, which wraps run, isn't correctly applied to the Process object within the child's context.
* Cleanup
* Logging bugfix (#18465)
* use mp Manager to handle logging queues
A Python bug (https://github.com/python/cpython/issues/91555) was preventing logs from the embeddings maintainer process from printing. The bug is fixed in Python 3.14, but a viable workaround is to use the multiprocessing Manager, which better manages mp queues and causes the logging to work correctly.
* consolidate
* fix typing
* Fix typing
* Use global log queue
* Move to using process for logging
* Convert camera tracking to process
* Add more processes
* Finalize process
* Cleanup
* Cleanup typing
* Formatting
* Remove daemon
---------
Co-authored-by: Josh Hawkins <32435876+hawkeye217@users.noreply.github.com>
* Add base class for global config updates
* Add or remove camera states
* Move camera process management to separate thread
* Move camera management fully to separate class
* Cleanup
* Stop camera processes when stop command is sent
* Start processes dynamically when needed
* Adjust
* Leave extra room in tracked object queue for two cameras
* Dynamically set extra config pieces
* Add some TODOs
* Fix type check
* Simplify config updates
* Improve typing
* Correctly handle indexed entries
* Cleanup
* Create out SHM
* Use ZMQ for signaling object detectoin is completed
* Get camera correctly created
* Cleanup for updating the cameras config
* Cleanup
* Don't enable audio if no cameras have audio transcription
* Use exact string so similar camera names don't interfere
* Add ability to update config via json body to config/set endpoint
Additionally, update the config in a single rather than multiple calls for each updated key
* fix autotracking calibration to support new config updater function
---------
Co-authored-by: Josh Hawkins <32435876+hawkeye217@users.noreply.github.com>
* install new packages for transcription support
* add config options
* audio maintainer modifications to support transcription
* pass main config to audio process
* embeddings support
* api and transcription post processor
* embeddings maintainer support for post processor
* live audio transcription with sherpa and faster-whisper
* update dispatcher with live transcription topic
* frontend websocket
* frontend live transcription
* frontend changes for speech events
* i18n changes
* docs
* mqtt docs
* fix linter
* use float16 and small model on gpu for real-time
* fix return value and use requestor to embed description instead of passing embeddings
* run real-time transcription in its own thread
* tweaks
* publish live transcriptions on their own topic instead of tracked_object_update
* config validator and docs
* clarify docs
* Fix the `Any` typing hint treewide
There has been confusion between the Any type[1] and the any function[2]
in typing hints.
[1] https://docs.python.org/3/library/typing.html#typing.Any
[2] https://docs.python.org/3/library/functions.html#any
* Fix typing for various frame_shape members
Frame shapes are most likely defined by height and width, so a single int
cannot express that.
* Wrap gpu stats functions in Optional[]
These can return `None`, so they need to be `Type | None`, which is what
`Optional` expresses very nicely.
* Fix return type in get_latest_segment_datetime
Returns a datetime object, not an integer.
* Make the return type of FrameManager.write optional
This is necessary since the SharedMemoryFrameManager.write function can
return None.
* Fix total_seconds() return type in get_tz_modifiers
The function returns a float, not an int.
https://docs.python.org/3/library/datetime.html#datetime.timedelta.total_seconds
* Account for floating point results in to_relative_box
Because the function uses division the return types may either be int or
float.
* Resolve ruff deprecation warning
The config has been split into formatter and linter, and the global
options are deprecated.
