Merge pull request #321 from kerberos-io/fix/discard-failed-recordings

fix/discard-failed-recordings

.devcontainer/Dockerfile

@@ -1,4 +1,4 @@
-FROM mcr.microsoft.com/devcontainers/go:1.24-bookworm
+FROM mcr.microsoft.com/devcontainers/go:1.25-trixie
 
 # Install node environment
 RUN apt-get update && \

.devcontainer/devcontainer.json

@@ -1,7 +1,7 @@
 // For format details, see https://aka.ms/devcontainer.json. For config options, see the
 // README at: https://github.com/devcontainers/templates/tree/main/src/python
 {
-	"name": "go:1.24-bookworm",
+	"name": "go:1.24-trixie",
 	"runArgs": [
 		"--name=agent",
 		"--network=host"
@@ -20,5 +20,5 @@
 		3000,
 		8080
 	],
-	"postCreateCommand": "cd ui && yarn install && yarn build && cd ../machinery && go mod download"
+	"postCreateCommand": "cd ui && yarn install && yarn build && cd ../machinery && go mod download && bash ./verify-moq-devcontainer.sh"
 }

.github/workflows/pr-build.yml

@@ -16,10 +16,8 @@ jobs:
         include:
           - architecture: amd64
             runner: ubuntu-24.04
-            dockerfile: Dockerfile
           - architecture: arm64
             runner: ubuntu-24.04-arm
-            dockerfile: Dockerfile.arm64
     steps:
       - name: Checkout
         uses: actions/checkout@v3
@@ -34,7 +32,7 @@ jobs:
           length: 7
       - name: Run Build
         run: |
-          docker build -t ${{ matrix.architecture }} -f ${{ matrix.dockerfile }} .
+          docker build -t ${{ matrix.architecture }} .
           CID=$(docker create ${{matrix.architecture}})
           docker cp ${CID}:/home/agent ./output-${{matrix.architecture}}
           docker rm ${CID}

.github/workflows/release-bump.yml

@@ -108,7 +108,7 @@ jobs:
           length: 7
       - name: Run Build
         run: |
-          docker build --provenance=false --build-arg VERSION=${{ needs.bump-release.outputs.tag }} -t ${{matrix.architecture}} -f Dockerfile.arm64 .
+          docker build --provenance=false --build-arg VERSION=${{ needs.bump-release.outputs.tag }} -t ${{matrix.architecture}} .
           CID=$(docker create ${{matrix.architecture}})
           docker cp ${CID}:/home/agent ./output-${{matrix.architecture}}
           docker rm ${CID}

.github/workflows/release-create.yml

@@ -71,7 +71,7 @@ jobs:
           length: 7
       - name: Run Build
         run: |
-          docker build --provenance=false --build-arg VERSION=${{github.event.inputs.tag || github.ref_name}} -t ${{matrix.architecture}} -f Dockerfile.arm64 .
+          docker build --provenance=false --build-arg VERSION=${{github.event.inputs.tag || github.ref_name}} -t ${{matrix.architecture}} .
           CID=$(docker create ${{matrix.architecture}})
           docker cp ${CID}:/home/agent ./output-${{matrix.architecture}}
           docker rm ${CID}

.vscode/launch.json

@@ -17,7 +17,7 @@
                 "8080"
             ],
             "envFile": "${workspaceFolder}/machinery/.env.local",
-            "buildFlags": "--tags dynamic",
+            "buildFlags": "--tags dynamic,moq",
             "env": {
                 "GOWORK": "off"
             },

.vscode/tasks.json

@@ -0,0 +1,14 @@
+{
+  "version": "2.0.0",
+  "tasks": [
+    {
+      "label": "agent: moq verify",
+      "type": "shell",
+      "command": "bash ./verify-moq-devcontainer.sh",
+      "options": {
+        "cwd": "${workspaceFolder}/machinery"
+      },
+      "problemMatcher": []
+    }
+  ]
+}

Dockerfile

@@ -1,13 +1,15 @@
 
-ARG BASE_IMAGE_VERSION=amd64-ddbe40e
+ARG GO_IMAGE=golang:1.25-trixie
+ARG RUNTIME_IMAGE=debian:trixie-slim
 ARG VERSION=0.0.0
-FROM kerberos/base:${BASE_IMAGE_VERSION} AS build-machinery
+FROM ${GO_IMAGE} AS build-machinery
 LABEL AUTHOR=uug.ai
 
 # Re-declare VERSION inside this stage so the value passed via
 # `--build-arg VERSION=...` (e.g. the release tag) is available below.
 # ARGs declared before the first FROM are not visible inside build stages.
 ARG VERSION
+ARG TARGETARCH
 
 ENV GOROOT=/usr/local/go
 ENV GOPATH=/go
@@ -17,9 +19,10 @@ ENV GOSUMDB=off
 ##########################################
 # Installing some additional dependencies.
 
-RUN apt-get upgrade -y && apt-get update && apt-get install -y --fix-missing --no-install-recommends \
+RUN apt-get update && apt-get install -y --fix-missing --no-install-recommends \
 	git build-essential cmake pkg-config unzip libgtk2.0-dev \
-	curl ca-certificates libcurl4-openssl-dev libssl-dev libjpeg62-turbo-dev && \
+	curl ca-certificates libavcodec-dev libavutil-dev libcurl4-openssl-dev \
+	libssl-dev libjpeg62-turbo-dev libswscale-dev && \
 	rm -rf /var/lib/apt/lists/*
 
 ##############################################################################
@@ -43,7 +46,9 @@ RUN cd /go/src/github.com/kerberos-io/agent/machinery && \
 	if [ -z "${VERSION}" ] || [ "${VERSION}" = "0.0.0" ]; then \
 		VERSION=$(cd /go/src/github.com/kerberos-io/agent && git describe --tags --always 2>/dev/null || echo "0.0.0"); \
 	fi && \
-	go build -tags timetzdata,netgo,osusergo --ldflags "-s -w -X github.com/kerberos-io/agent/machinery/src/utils.VERSION=${VERSION} -extldflags '-static -latomic'" main.go && \
+	BUILD_TAGS=timetzdata,netgo,osusergo && \
+	case "${TARGETARCH:-$(go env GOARCH)}" in amd64|arm64) BUILD_TAGS="moq,${BUILD_TAGS}" ;; esac && \
+	go build -tags "${BUILD_TAGS}" --ldflags "-s -w -X github.com/kerberos-io/agent/machinery/src/utils.VERSION=${VERSION}" main.go && \
 	mkdir -p /agent && \
 	mv main /agent && \
 	mv version /agent && \
@@ -89,12 +94,17 @@ RUN mkdir -p ./agent && cp -r /go/src/github.com/kerberos-io/agent/machinery/www
 ############################################
 # Publish main binary to GitHub release
 
-FROM alpine:latest
+FROM ${RUNTIME_IMAGE}
 
 ############################
 # Protect by non-root user.
 
-RUN addgroup -S kerberosio && adduser -S agent -G kerberosio && addgroup agent video
+RUN apt-get update && apt-get install -y --no-install-recommends \
+	ca-certificates curl ffmpeg libatomic1 libcap2-bin libstdc++6 && \
+	rm -rf /var/lib/apt/lists/* && \
+	groupadd --system kerberosio && \
+	useradd --system --gid kerberosio --groups video --create-home agent && \
+	chmod 0755 /home/agent
 
 #################################
 # Copy files from previous images
@@ -102,8 +112,6 @@ RUN addgroup -S kerberosio && adduser -S agent -G kerberosio && addgroup agent v
 COPY --chown=0:0 --from=build-machinery /dist /
 COPY --chown=0:0 --from=build-ui /dist /
 
-RUN apk update && apk add ca-certificates curl ffmpeg libstdc++ libc6-compat --no-cache && rm -rf /var/cache/apk/*
-
 ##################
 # Try running agent
 
@@ -123,7 +131,7 @@ RUN chown -R agent:kerberosio /home/agent/www
 
 ###########################
 # Grant the necessary root capabilities to the process trying to bind to the privileged port
-RUN apk add libcap && setcap 'cap_net_bind_service=+ep' /home/agent/main
+RUN setcap 'cap_net_bind_service=+ep' /home/agent/main
 
 ###################
 # Run non-root user

Dockerfile.arm64

@@ -1,147 +0,0 @@
-
-ARG BASE_IMAGE_VERSION=arm64-ddbe40e
-ARG VERSION=0.0.0
-FROM kerberos/base:${BASE_IMAGE_VERSION} AS build-machinery
-LABEL AUTHOR=uug.ai
-
-# Re-declare VERSION inside this stage so the value passed via
-# `--build-arg VERSION=...` (e.g. the release tag) is available below.
-# ARGs declared before the first FROM are not visible inside build stages.
-ARG VERSION
-
-ENV GOROOT=/usr/local/go
-ENV GOPATH=/go
-ENV PATH=$GOPATH/bin:$GOROOT/bin:/usr/local/lib:$PATH
-ENV GOSUMDB=off
-
-##########################################
-# Installing some additional dependencies.
-
-RUN apt-get upgrade -y && apt-get update && apt-get install -y --fix-missing --no-install-recommends \
-	git build-essential cmake pkg-config unzip libgtk2.0-dev \
-	curl ca-certificates libcurl4-openssl-dev libssl-dev libjpeg62-turbo-dev && \
-	rm -rf /var/lib/apt/lists/*
-
-##############################################################################
-# Copy all the relevant source code in the Docker image, so we can build this.
-
-RUN mkdir -p /go/src/github.com/kerberos-io/agent
-COPY machinery /go/src/github.com/kerberos-io/agent/machinery
-RUN rm -rf /go/src/github.com/kerberos-io/agent/machinery/.env
-
-##################################################################
-# Get the latest commit hash, so we know which version we're running
-COPY .git /go/src/github.com/kerberos-io/agent/.git
-RUN cd /go/src/github.com/kerberos-io/agent/.git && git log --format="%H" -n 1 | head -c7 > /go/src/github.com/kerberos-io/agent/machinery/version
-RUN cat /go/src/github.com/kerberos-io/agent/machinery/version
-
-##################
-# Build Machinery
-
-RUN cd /go/src/github.com/kerberos-io/agent/machinery && \
-	go mod download && \
-	if [ -z "${VERSION}" ] || [ "${VERSION}" = "0.0.0" ]; then \
-		VERSION=$(cd /go/src/github.com/kerberos-io/agent && git describe --tags --always 2>/dev/null || echo "0.0.0"); \
-	fi && \
-	go build -tags timetzdata,netgo,osusergo --ldflags "-s -w -X github.com/kerberos-io/agent/machinery/src/utils.VERSION=${VERSION} -extldflags '-static -latomic'" main.go && \
-	mkdir -p /agent && \
-	mv main /agent && \
-	mv version /agent && \
-	mv data /agent && \
-	mkdir -p /agent/data/cloud && \
-	mkdir -p /agent/data/snapshots && \
-	mkdir -p /agent/data/log && \
-	mkdir -p /agent/data/recordings && \
-	mkdir -p /agent/data/capture-test && \
-	mkdir -p /agent/data/config
-
-####################################
-# Let's create a /dist folder containing just the files necessary for runtime.
-# Later, it will be copied as the / (root) of the output image.
-
-WORKDIR /dist
-RUN cp -r /agent ./
-
-####################################################################################
-# This will collect dependent libraries so they're later copied to the final image.
-
-RUN /dist/agent/main version
-
-FROM node:22-alpine AS build-ui
-
-RUN apk update && apk upgrade --available && sync
-
-########################
-# Build Web (React app)
-
-RUN mkdir -p /go/src/github.com/kerberos-io/agent/machinery/www
-COPY ui /go/src/github.com/kerberos-io/agent/ui
-RUN cd /go/src/github.com/kerberos-io/agent/ui && rm -rf yarn.lock && yarn config set network-timeout 300000 && \
-	yarn && yarn build
-
-####################################
-# Let's create a /dist folder containing just the files necessary for runtime.
-# Later, it will be copied as the / (root) of the output image.
-
-WORKDIR /dist
-RUN mkdir -p ./agent && cp -r /go/src/github.com/kerberos-io/agent/machinery/www ./agent/
-
-############################################
-# Publish main binary to GitHub release
-
-FROM alpine:latest
-
-############################
-# Protect by non-root user.
-
-RUN addgroup -S kerberosio && adduser -S agent -G kerberosio && addgroup agent video
-
-#################################
-# Copy files from previous images
-
-COPY --chown=0:0 --from=build-machinery /dist /
-COPY --chown=0:0 --from=build-ui /dist /
-
-RUN apk update && apk add ca-certificates curl ffmpeg libstdc++ libc6-compat --no-cache && rm -rf /var/cache/apk/*
-
-##################
-# Try running agent
-
-RUN mv /agent/* /home/agent/
-RUN /home/agent/main version
-
-#######################
-# Make template config
-
-RUN cp /home/agent/data/config/config.json /home/agent/data/config.template.json
-
-###########################
-# Set permissions correctly
-
-RUN chown -R agent:kerberosio /home/agent/data
-RUN chown -R agent:kerberosio /home/agent/www
-
-###########################
-# Grant the necessary root capabilities to the process trying to bind to the privileged port
-RUN apk add libcap && setcap 'cap_net_bind_service=+ep' /home/agent/main
-
-###################
-# Run non-root user
-
-USER agent
-
-######################################
-# By default the app runs on port 80
-
-EXPOSE 80
-
-######################################
-# Check if agent is still running
-
-HEALTHCHECK CMD curl --fail http://localhost:80 || exit 1   
-
-###################################################
-# Leeeeettttt'ssss goooooo!!!
-# Run the shizzle from the right working directory.
-WORKDIR /home/agent
-CMD ["./main", "-action", "run", "-port", "80"]

README-RTSPS-TLS.md

@@ -0,0 +1,645 @@
+# RTSPS and TLS certificates
+
+This guide explains how Kerberos Agent connects to an IP camera over RTSPS,
+how to issue a camera certificate with a private CA, and how to validate the
+complete trust path. It also explains why some apparently corrupted trust
+bundles can still allow a connection.
+
+The camera-specific steps were verified with a Bosch FLEXIDOME micro 3100i.
+Other Bosch firmware versions may use different labels or ports.
+
+The commands were tested with Smallstep CLI `0.30.6` and OpenSSL `3.5.6` on
+Debian. Check `step certificate sign --help` when using an older Smallstep CLI.
+The OpenSSL isolation flags `-no-CApath` and `-no-CAstore` require a version that
+lists them in `openssl s_client -help`.
+
+## Tested configuration
+
+| Setting | Value |
+| --- | --- |
+| Camera | Bosch FLEXIDOME micro 3100i |
+| Example camera address | `10.0.30.11` |
+| RTSPS port | `9554` |
+| Main stream | `rtsps://<user>:<password>@10.0.30.11:9554/?inst=1` |
+| Sub stream | `rtsps://<user>:<password>@10.0.30.11:9554/?inst=2` |
+| Certificate SAN | `IP Address:10.0.30.11` |
+| Bosch certificate usage | `HTTPS` |
+| Agent trust input | Issuing intermediate plus root CA |
+
+Replace the example address and certificate names throughout this guide. Keep
+camera credentials out of source control and percent-encode reserved URL
+characters in usernames and passwords.
+
+## Mental model
+
+### RTSPS, SRTSP, TLS, and SRTP
+
+- The standard URL scheme is `rtsps://`. Do not use `srtsp://`.
+- Bosch interfaces and documentation may use SRTSP or Secure RTSP as product
+  terminology.
+- RTSPS carries the RTSP control connection over TLS. With gortsplib, media is
+  normally interleaved over the same TCP/TLS connection for this camera.
+- SRTP is a separate media protection mechanism and is negotiated only when the
+  camera advertises an appropriate secure RTP profile.
+
+Encryption alone does not prove which camera the Agent reached. Verified TLS
+also checks that:
+
+1. The camera certificate is signed by a trusted authority.
+2. The certificate is valid at the current time.
+3. The URL host matches a certificate Subject Alternative Name (SAN).
+
+Modern Go verification uses SANs for identity. A Common Name alone is not
+sufficient. Connecting to `10.0.30.11` requires an IP SAN with that exact value,
+not `DNS:10.0.30.11` and not only a device-name DNS SAN.
+
+### Agent behavior
+
+Kerberos Agent uses gortsplib for RTSP and RTSPS. With the normal configuration,
+gortsplib receives a nil custom TLS configuration and Go performs standard
+certificate and hostname verification with the process trust pool.
+
+`AGENT_CAPTURE_IPCAMERA_RTSPS_INSECURE=true` is an explicit escape hatch that
+sets `InsecureSkipVerify` for camera clients. It should be false in a verified
+deployment.
+
+## Communication and certificate flow
+
+The certificate is used during the TLS handshake, before the first RTSP command
+is exchanged. It is not attached to `DESCRIBE`, `SETUP`, or `PLAY`, and the CA
+trust bundle is never sent to the camera.
+
+```mermaid
+sequenceDiagram
+  participant Agent as Kerberos Agent
+  participant Trust as Go trust pool
+  participant Camera as Camera RTSPS :9554
+
+  Agent->>Trust: Load system roots and append AGENT_CAPTURE_IPCAMERA_RTSPS_CA_FILE
+  Agent->>Camera: Open TCP connection
+  Agent->>Camera: Send TLS ClientHello
+  Camera-->>Agent: Send TLS ServerHello and camera certificate
+  Agent->>Trust: Verify chain, validity, serverAuth, and URL host against SAN
+  Trust-->>Agent: Accept or reject the camera identity
+  Agent->>Camera: Complete TLS handshake
+  Note over Agent,Camera: All following traffic is encrypted by TLS
+  Agent->>Camera: DESCRIBE with RTSP authentication
+  Camera-->>Agent: Return SDP and available media tracks
+  Agent->>Camera: SETUP selected video and audio tracks over TCP
+  Agent->>Camera: PLAY
+  Camera-->>Agent: Send interleaved RTP and RTCP media over TLS
+```
+
+The files and keys have distinct roles:
+
+| Material | Location | Purpose | Sent over the connection |
+| --- | --- | --- | --- |
+| Camera leaf certificate | Camera | Identifies the camera and binds its public key to its SAN | Yes, by the camera during the TLS handshake |
+| Camera private key | Camera | Proves that the camera owns the presented certificate | No |
+| Intermediate and root CA PEM bundle | Agent | Lets Go build and trust the camera certificate chain | No |
+| RTSP username and password | Agent configuration or URL | Authenticates the Agent to the RTSP service after TLS succeeds | An authentication response is sent inside TLS; its form depends on the RTSP authentication method |
+
+For an `rtsps://` URL, the Agent parses the URL and gives gortsplib the host and
+TLS settings. gortsplib opens the TCP connection and starts TLS. Go compares the
+certificate presented by the camera with the local trust pool, checks its
+validity period and server usage, and matches the URL hostname or IP address to
+the certificate SAN. Only a successful handshake creates the encrypted channel
+needed for the RTSP exchange.
+
+The Agent then sends `DESCRIBE`, selects the advertised video and audio tracks,
+sends `SETUP`, and starts delivery with `PLAY`. For the tested camera, gortsplib
+uses interleaved TCP, so the RTSP control messages and RTP/RTCP media remain
+inside the same encrypted TLS connection. Main stream, sub stream, and enabled
+audio backchannel clients each establish and verify their own connection.
+
+If certificate verification fails, the TLS handshake does not complete and no
+usable RTSP session is established. Setting
+`AGENT_CAPTURE_IPCAMERA_RTSPS_INSECURE=true` keeps traffic encrypted but skips
+certificate-chain and hostname verification, so an attacker could impersonate
+the camera. It is not equivalent to trusting the camera certificate.
+
+## Decide the certificate identity first
+
+Choose the stable name used in every Agent URL before creating the certificate:
+
+- For an IP URL, add that address as an IP SAN.
+- For a DNS URL, add the exact hostname as a DNS SAN.
+- Add both when clients legitimately use both forms.
+
+A certificate stops matching if the camera IP changes. Use a static address,
+DHCP reservation, or stable DNS name.
+
+## Configure RTSPS in the Bosch UI
+
+1. Sign in to the camera as an administrator.
+2. Open **Configuration**.
+3. Open **Network > Network Services**.
+4. Enable **RTSPS**.
+5. Confirm port `9554`, or record the configured alternative.
+6. Click **Set**.
+
+RTSP on port `554` and RTSPS on port `9554` are separate services. Enabling
+RTSPS does not make an `rtsp://` URL secure.
+
+## Generate the private key and CSR on the camera
+
+Keeping the TLS private key on the camera avoids exporting it to an operator
+workstation or deployment system.
+
+1. Open **Service > Certificates**.
+2. Click **Add**.
+3. Select **Generate signing request**.
+4. Select `RSA 2048bit` or the stronger option supported by all clients.
+5. Enter a unique file name, such as `agent-rtsps`.
+6. Enter a descriptive Common Name and any required organization fields.
+7. Click **Generate**.
+8. Download the resulting CSR from the certificate table.
+
+On the tested firmware, this form contains no SAN field. The downloaded CSR
+therefore has no IP SAN. The CA must add the SAN while signing.
+
+Inspect the CSR before signing:
+
+```bash
+openssl req -in camera.csr.pem -noout -verify -subject
+openssl req -in camera.csr.pem -noout -text
+```
+
+The first command must report `Certificate request self-signature verify OK`.
+An absent `Subject Alternative Name` section is expected for this firmware.
+
+## Prepare Smallstep
+
+Use an existing organizational CA when one is available. Creating a new CA
+creates a new long-lived trust domain that must be distributed, protected,
+backed up, and eventually rotated.
+
+### Install the CLI on Debian amd64
+
+```bash
+curl -fsSL \
+  https://dl.smallstep.com/cli/docs-ca-install/latest/step-cli_amd64.deb \
+  -o /tmp/step-cli_amd64.deb
+sudo dpkg -i /tmp/step-cli_amd64.deb
+rm /tmp/step-cli_amd64.deb
+step version
+```
+
+Use the official package matching the host architecture on other systems.
+
+### Create a dedicated offline CA
+
+Skip this section when using an existing CA.
+
+```bash
+umask 077
+mkdir -p "$HOME/.step/secrets" "$HOME/.step/camera"
+
+openssl rand -base64 48 > "$HOME/.step/secrets/camera_ca_password"
+chmod 600 "$HOME/.step/secrets/camera_ca_password"
+
+step ca init \
+  --pki \
+  --name "UUG Camera CA" \
+  --password-file "$HOME/.step/secrets/camera_ca_password"
+```
+
+This produces:
+
+```text
+$HOME/.step/certs/root_ca.crt
+$HOME/.step/certs/intermediate_ca.crt
+$HOME/.step/secrets/root_ca_key
+$HOME/.step/secrets/intermediate_ca_key
+$HOME/.step/secrets/camera_ca_password
+```
+
+The files under `secrets/` are sensitive. Keep them mode `600`, never commit
+them, and back them up to encrypted persistent storage. A devcontainer can be
+rebuilt or deleted; it is not sufficient as the only CA backup.
+
+## Add the SAN while signing
+
+Copy the camera CSR into a protected working directory:
+
+```bash
+cp camera.csr.pem "$HOME/.step/camera/camera.csr.pem"
+```
+
+Create `$HOME/.step/camera/bosch-rtsps.tpl`:
+
+```json
+{
+  "subject": {
+    "commonName": {{ toJson .Insecure.CR.Subject.CommonName }}
+  },
+  "ipAddresses": ["10.0.30.11"],
+  "keyUsage": ["keyEncipherment", "digitalSignature"],
+  "extKeyUsage": ["serverAuth", "clientAuth"]
+}
+```
+
+The template preserves the camera CSR public key, sets the IP identity, and
+creates a TLS leaf rather than a CA certificate.
+
+Sign it with a validity period that ends before the intermediate CA expires.
+A one-year leaf is preferable to a ten-year leaf when automated renewal is
+available:
+
+```bash
+step certificate sign \
+  --template "$HOME/.step/camera/bosch-rtsps.tpl" \
+  --bundle \
+  --not-after 8760h \
+  --password-file "$HOME/.step/secrets/camera_ca_password" \
+  "$HOME/.step/camera/camera.csr.pem" \
+  "$HOME/.step/certs/intermediate_ca.crt" \
+  "$HOME/.step/secrets/intermediate_ca_key" \
+  > "$HOME/.step/camera/bosch-rtsps-chain.pem"
+```
+
+For an online `step-ca`, do not assume `step ca sign` accepts a `--san` flag. It
+does not. Authorize SANs in the one-time token or configure a provisioner
+template that produces the required SANs.
+
+## Validate before upload
+
+Inspect the leaf certificate, which is the first PEM block in the chain file:
+
+```bash
+openssl x509 \
+  -in "$HOME/.step/camera/bosch-rtsps-chain.pem" \
+  -noout -subject -issuer -dates -ext subjectAltName -ext extendedKeyUsage
+```
+
+Confirm the SAN separately because some OpenSSL versions display only the last
+requested extension:
+
+```bash
+openssl x509 \
+  -in "$HOME/.step/camera/bosch-rtsps-chain.pem" \
+  -noout -ext subjectAltName
+```
+
+Verify the path and IP identity:
+
+```bash
+openssl verify \
+  -CAfile "$HOME/.step/certs/root_ca.crt" \
+  -untrusted "$HOME/.step/certs/intermediate_ca.crt" \
+  -verify_ip 10.0.30.11 \
+  "$HOME/.step/camera/bosch-rtsps-chain.pem"
+```
+
+Confirm that the signed leaf uses the exact public key from the camera CSR:
+
+```bash
+csr_key=$(
+  openssl req -in "$HOME/.step/camera/camera.csr.pem" -pubkey -noout |
+    openssl pkey -pubin -outform DER 2>/dev/null |
+    sha256sum | cut -d' ' -f1
+)
+
+cert_key=$(
+  openssl x509 -in "$HOME/.step/camera/bosch-rtsps-chain.pem" -pubkey -noout |
+    openssl pkey -pubin -outform DER 2>/dev/null |
+    sha256sum | cut -d' ' -f1
+)
+
+test "$csr_key" = "$cert_key"
+```
+
+Do not upload a certificate when any of these checks fail.
+
+## Upload and assign the certificate
+
+1. Return to **Service > Certificates**.
+2. Click **Add > Upload certificate**.
+3. Select the leaf-plus-intermediate PEM chain.
+4. Click **Upload** and wait for `100%`.
+5. Confirm that the former CSR row is now a `Certificate`.
+6. Confirm that the key icon is present. It proves that the camera associated
+   the certificate with its retained private key.
+7. Open the new certificate's **Usage** selector.
+8. Select only **HTTPS**.
+9. Leave **CBS client** assigned to the original Bosch `DeviceCertificate`.
+10. Click **Set** and wait for the table to reload.
+
+On the tested firmware, there is no separate SRTSP usage. RTSPS presents the
+certificate assigned to HTTPS. Reassigning HTTPS therefore changes both the
+web interface and RTSPS certificate.
+
+After saving, the expected split is:
+
+| Certificate | Usage |
+| --- | --- |
+| Private-CA camera certificate | `HTTPS` |
+| Bosch `DeviceCertificate` | `CBS client` |
+
+The browser may warn about the new HTTPS certificate until the private root CA
+is trusted by the workstation.
+
+## Account for the Bosch chain behavior
+
+The tested firmware served only the leaf certificate on ports `443` and `9554`,
+even when the uploaded file contained the leaf and intermediate. Uploading the
+intermediate separately as a trusted camera certificate did not change the
+served chain.
+
+Confirm the behavior:
+
+```bash
+openssl s_client \
+  -connect 10.0.30.11:9554 \
+  -showcerts </dev/null 2>/dev/null |
+  grep -c '^-----BEGIN CERTIFICATE-----$'
+```
+
+A result of `1` means the client must already have the issuing intermediate.
+Create a portable trust bundle containing the intermediate and root:
+
+```bash
+step certificate bundle \
+  "$HOME/.step/certs/intermediate_ca.crt" \
+  "$HOME/.step/certs/root_ca.crt" \
+  "$HOME/.step/camera/uug-camera-trust-bundle.pem"
+
+chmod 644 "$HOME/.step/camera/uug-camera-trust-bundle.pem"
+```
+
+The trust bundle is public material. The CA private keys and password are not.
+
+## Configure Kerberos Agent
+
+For a process running directly in the same environment:
+
+```dotenv
+AGENT_CAPTURE_IPCAMERA_RTSP="rtsps://<user>:<password>@10.0.30.11:9554/?inst=1"
+AGENT_CAPTURE_IPCAMERA_SUB_RTSP="rtsps://<user>:<password>@10.0.30.11:9554/?inst=2"
+AGENT_CAPTURE_IPCAMERA_RTSPS_INSECURE=false
+AGENT_CAPTURE_IPCAMERA_RTSPS_CA_FILE=/home/agent/data/config/uug-camera-trust-bundle.pem
+```
+
+For a container, mount the public bundle read-only at the exact path visible
+inside the container. The Agent image creates `/home/agent/data/config` and
+includes Debian's `ca-certificates` package:
+
+```bash
+docker run \
+  -v /secure/config/uug-camera-trust-bundle.pem:/home/agent/data/config/uug-camera-trust-bundle.pem:ro \
+  -e AGENT_CAPTURE_IPCAMERA_RTSPS_CA_FILE=/home/agent/data/config/uug-camera-trust-bundle.pem \
+  -e AGENT_CAPTURE_IPCAMERA_RTSPS_INSECURE=false \
+  -e 'AGENT_CAPTURE_IPCAMERA_RTSP=rtsps://<user>:<password>@10.0.30.11:9554/?inst=1' \
+  -e 'AGENT_CAPTURE_IPCAMERA_SUB_RTSP=rtsps://<user>:<password>@10.0.30.11:9554/?inst=2' \
+  kerberos/agent:latest
+```
+
+`AGENT_CAPTURE_IPCAMERA_RTSPS_CA_FILE` starts with the operating system's roots
+and appends the camera bundle only to the gortsplib TLS configuration. Other
+clients, including MoQ, Hub, and Vault, retain the normal public CA chain.
+Restart the Agent after changing trust files.
+
+Do not set `SSL_CERT_FILE` or `SSL_CERT_DIR` in production solely for camera
+trust. They are process-wide and can prevent other clients from validating
+public services. For a deliberate process-wide isolation test, mount an empty
+directory and set `SSL_CERT_DIR` to its path:
+
+```bash
+-v /secure/config/empty-ca-dir:/home/agent/data/config/empty-ca-dir:ro \
+-e SSL_CERT_DIR=/home/agent/data/config/empty-ca-dir
+```
+
+Do not use `SSL_CERT_DIR=`. Go treats an empty value as unset and scans its
+default certificate directories.
+
+Only use that mode when the Agent does not need public roots for other TLS
+connections.
+
+## Validate the live endpoints
+
+### Strict TLS and identity check
+
+Use only the specified bundle, without OpenSSL's default CA locations:
+
+```bash
+openssl s_client \
+  -brief \
+  -connect 10.0.30.11:9554 \
+  -verify_ip 10.0.30.11 \
+  -verify_return_error \
+  -CAfile "$HOME/.step/camera/uug-camera-trust-bundle.pem" \
+  -no-CApath \
+  -no-CAstore \
+  </dev/null
+```
+
+Repeat with port `443`. Both must report `Verification: OK`.
+
+Confirm that identity checking is active by repeating the command with a wrong
+address, such as `-verify_ip 10.0.30.12`. It must fail with an IP address
+mismatch.
+
+### Confirm the live leaf is the generated leaf
+
+```bash
+live_fingerprint=$(
+  openssl s_client -connect 10.0.30.11:9554 -showcerts </dev/null 2>/dev/null |
+    openssl x509 -noout -fingerprint -sha256 |
+    cut -d= -f2
+)
+
+local_fingerprint=$(
+  openssl x509 \
+    -in "$HOME/.step/camera/bosch-rtsps-chain.pem" \
+    -noout -fingerprint -sha256 |
+    cut -d= -f2
+)
+
+test "$live_fingerprint" = "$local_fingerprint"
+```
+
+### Validate the media path
+
+A successful TLS handshake does not prove that RTSP authentication, DESCRIBE,
+SETUP, PLAY, and RTP delivery work. Start a fresh Agent with verified TLS and
+confirm that it connects without an x509 error and receives frames. During the
+verified setup described here, a gortsplib probe completed all RTSP operations
+and received an RTP packet over TCP.
+
+## Why a tampered bundle may still connect
+
+Editing PEM text is not always a useful negative TLS test.
+
+### A certificate can still parse after a byte change
+
+Base64 can remain syntactically valid when one character changes. OpenSSL may
+still list the certificate subject and issuer even though a signature is now
+invalid. Parsing and signature verification are different operations.
+
+### Trust anchors are not validated through a parent
+
+Every certificate loaded into Go's root pool is a trust anchor, including a
+non-self-signed intermediate CA. Verification can terminate at that certificate.
+
+If tampering changes only the intermediate's signature from its parent root,
+but does not change its public key, that intermediate can still validate the
+camera leaf when it is trusted directly. Its now-invalid parent signature is
+not consulted at the trust boundary.
+
+This is equivalent to OpenSSL's partial-chain behavior:
+
+```bash
+openssl s_client \
+  -connect 10.0.30.11:9554 \
+  -verify_ip 10.0.30.11 \
+  -verify_return_error \
+  -partial_chain \
+  -CAfile tampered-bundle.pem \
+  -no-CApath \
+  -no-CAstore \
+  </dev/null
+```
+
+### `SSL_CERT_FILE` does not isolate Go from CA directories
+
+On Unix, Go uses `SSL_CERT_FILE` instead of its default aggregate CA file, but it
+still scans default certificate directories such as `/etc/ssl/certs`. Setting
+`SSL_CERT_FILE` alone therefore does not remove CA certificates installed with
+`update-ca-certificates`. `AGENT_CAPTURE_IPCAMERA_RTSPS_CA_FILE` is appended
+after this system pool is loaded; it does not replace the system roots.
+
+Blank values do not select empty trust sources. Both `SSL_CERT_FILE=` and
+`SSL_CERT_DIR=` are treated as unset, so Go falls back to its default aggregate
+CA file and certificate directories. To test with no trusted certificates on
+Linux, use a non-empty file path that contains no certificates and a non-empty
+directory path that contains no certificates:
+
+```bash
+mkdir -p /tmp/empty-ca-dir
+SSL_CERT_FILE=/dev/null \
+SSL_CERT_DIR=/tmp/empty-ca-dir \
+AGENT_CAPTURE_IPCAMERA_RTSPS_CA_FILE= \
+AGENT_CAPTURE_IPCAMERA_RTSPS_INSECURE=false \
+GOWORK=off \
+go run -tags moq . -action run -port 8080
+```
+
+That fresh process must fail with `x509: certificate signed by unknown
+authority`.
+
+Use exactly one camera trust-distribution approach when possible:
+
+1. Mount a private trust bundle and set
+  `AGENT_CAPTURE_IPCAMERA_RTSPS_CA_FILE`; or
+2. Install the CA certificates into the operating-system trust store.
+
+Using both is valid, but makes isolation tests less obvious.
+
+### Running processes can retain old roots
+
+A long-running Go process may already have loaded and cached the trust pool.
+Always start a new process after changing trust configuration during a negative
+test.
+
+## Perform a meaningful negative test
+
+Do not corrupt only the root or intermediate signature. Instead, give a fresh
+Agent process a completely unrelated CA and hide the default CA directories.
+
+```bash
+mkdir -p /tmp/empty-ca-dir
+
+openssl req \
+  -x509 -newkey rsa:2048 -nodes -days 1 \
+  -subj '/CN=Unrelated Test Root' \
+  -keyout /tmp/unrelated-test-root.key \
+  -out /tmp/unrelated-test-root.crt
+
+SSL_CERT_FILE=/dev/null \
+SSL_CERT_DIR=/tmp/empty-ca-dir \
+AGENT_CAPTURE_IPCAMERA_RTSPS_CA_FILE=/tmp/unrelated-test-root.crt \
+AGENT_CAPTURE_IPCAMERA_RTSPS_INSECURE=false \
+GOWORK=off \
+go run -tags moq . -action run -port 8080
+```
+
+The connection must fail with an unknown-authority or chain-building error.
+Delete the temporary test key and certificate afterward.
+
+To test bundle integrity rather than client distrust, validate the intermediate
+against the root explicitly:
+
+```bash
+openssl verify \
+  -CAfile "$HOME/.step/certs/root_ca.crt" \
+  -no-CApath \
+  -no-CAstore \
+  "$HOME/.step/certs/intermediate_ca.crt"
+```
+
+Store and compare approved SHA-256 fingerprints when detecting unauthorized
+certificate-file changes is a requirement.
+
+Restore an accidentally edited bundle from the protected CA certificates, then
+restart the Agent:
+
+```bash
+step certificate bundle -f \
+  "$HOME/.step/certs/intermediate_ca.crt" \
+  "$HOME/.step/certs/root_ca.crt" \
+  "$HOME/.step/camera/uug-camera-trust-bundle.pem"
+
+openssl verify \
+  -CAfile "$HOME/.step/certs/root_ca.crt" \
+  -no-CApath \
+  -no-CAstore \
+  "$HOME/.step/certs/intermediate_ca.crt"
+```
+
+## Optional system trust installation
+
+On Debian, install both public CA certificates when every process in the system
+should trust this camera PKI:
+
+```bash
+sudo install -m 0644 \
+  "$HOME/.step/certs/root_ca.crt" \
+  /usr/local/share/ca-certificates/uug-camera-ca.crt
+
+sudo install -m 0644 \
+  "$HOME/.step/certs/intermediate_ca.crt" \
+  /usr/local/share/ca-certificates/uug-camera-intermediate-ca.crt
+
+sudo update-ca-certificates
+```
+
+This creates links below `/etc/ssl/certs`. Remove those files and rerun
+`update-ca-certificates` before attempting an isolated trust-bundle test.
+
+## Renewal and recovery
+
+- Renew before the leaf or intermediate expires.
+- Generate a new camera CSR if the firmware cannot renew the existing key.
+- Sign the new CSR with all required SANs.
+- Upload and validate the new certificate before deleting the old one.
+- Preserve an alternate administrative access path while changing HTTPS usage.
+- Back up the CA certificates, encrypted CA keys, and password separately.
+- If the CA private keys are lost, create a new CA and redistribute its trust
+  before replacing camera certificates.
+
+## Production checklist
+
+- [ ] The Agent URL uses `rtsps://`, not `rtsp://` or `srtsp://`.
+- [ ] RTSPS is enabled on the camera and the configured port is reachable.
+- [ ] The certificate SAN exactly matches the Agent URL host.
+- [ ] The leaf public key matches the camera-generated CSR.
+- [ ] The certificate has `serverAuth` extended key usage.
+- [ ] The certificate expires before its issuer.
+- [ ] HTTPS is assigned to the private-CA certificate.
+- [ ] CBS client remains assigned to the Bosch device certificate.
+- [ ] The Agent has the intermediate and root CA certificates it needs.
+- [ ] `AGENT_CAPTURE_IPCAMERA_RTSPS_INSECURE=false`.
+- [ ] The Agent was restarted after trust changes.
+- [ ] A strict TLS check reports `Verification: OK`.
+- [ ] A real Agent connection receives RTP packets.
+- [ ] CA private keys and passwords are backed up outside the devcontainer.

README.md

@@ -189,6 +189,21 @@ Next to attaching the configuration file, it is also possible to override the co
     -e AGENT_CAPTURE_CONTINUOUS=true \
     -d --restart=always kerberos/agent:latest
 
+### Secure camera streams (RTSPS)
+
+The Agent accepts `rtsps://` camera URLs. Do not use `srtsp://`; RTSPS is RTSP over TLS. For a Bosch FLEXIDOME micro 3100i, enable **Secure RTSP** under **Network > Network Services** and use port `9554`:
+
+```bash
+AGENT_CAPTURE_IPCAMERA_RTSP='rtsps://username:password@camera.example:9554/?inst=1'
+AGENT_CAPTURE_IPCAMERA_SUB_RTSP='rtsps://username:password@camera.example:9554/?inst=2'
+```
+
+Certificate verification is enabled by default. The URL hostname or IP address must match the camera certificate SAN. On this Bosch firmware, RTSPS presents the certificate assigned to **HTTPS**; there is no separate SRTSP certificate usage. Leave **CBS client** assigned to the Bosch device certificate.
+
+For a private CA, mount a PEM trust bundle containing every CA certificate needed to build the camera certificate chain and set `AGENT_CAPTURE_IPCAMERA_RTSPS_CA_FILE` to its path inside the Agent. The bundle is appended to the system roots for camera RTSPS connections only. This Bosch firmware presents only its leaf certificate, so include both the issuing intermediate and root certificates in the bundle. As a temporary fallback, `AGENT_CAPTURE_IPCAMERA_RTSPS_INSECURE=true` disables certificate verification for camera streams only.
+
+See [RTSPS and TLS certificates](README-RTSPS-TLS.md) for the complete Bosch UI, private-CA, deployment, validation, and troubleshooting procedure.
+
 | Name                                    | Description                                                                                     | Default Value                  |
 | --------------------------------------- | ----------------------------------------------------------------------------------------------- | ------------------------------ |
 | `LOG_LEVEL`                                 | Level for logging, could be "info", "warning", "debug", "error" or "fatal".                     | "info"                         |
@@ -208,8 +223,10 @@ Next to attaching the configuration file, it is also possible to override the co
 | `AGENT_TIME`                                | Enable the timetable for Kerberos Agent                                                         | "false"                        |
 | `AGENT_TIMETABLE`                           | A (weekly) time table to specify when to make recordings "start1,end1,start2,end2;start1..      | ""                             |
 | `AGENT_REGION_POLYGON`                      | A single polygon set for motion detection: "x1,y1;x2,y2;x3,y3;...                               | ""                             |
-| `AGENT_CAPTURE_IPCAMERA_RTSP`               | Full-HD RTSP endpoint to the camera you're targetting.                                          | ""                             |
-| `AGENT_CAPTURE_IPCAMERA_SUB_RTSP`           | Sub-stream RTSP endpoint used for livestreaming (WebRTC).                                       | ""                             |
+| `AGENT_CAPTURE_IPCAMERA_RTSP`               | Full-HD RTSP or RTSPS endpoint for the target camera.                                           | ""                             |
+| `AGENT_CAPTURE_IPCAMERA_SUB_RTSP`           | RTSP or RTSPS sub-stream endpoint used for livestreaming (WebRTC).                              | ""                             |
+| `AGENT_CAPTURE_IPCAMERA_RTSPS_CA_FILE`      | PEM CA bundle appended to the system roots for RTSPS camera certificate verification.           | ""                             |
+| `AGENT_CAPTURE_IPCAMERA_RTSPS_INSECURE`     | Disable RTSPS camera certificate verification; use only when a trusted CA cannot be installed.  | "false"                        |
 | `AGENT_CAPTURE_IPCAMERA_BASE_WIDTH`         | Force a specific width resolution for live view processing.                                     | ""                             |
 | `AGENT_CAPTURE_IPCAMERA_BASE_HEIGHT`        | Force a specific height resolution for live view processing.                                    | ""                             |
 | `AGENT_CAPTURE_IPCAMERA_ONVIF`              | Mark as a compliant ONVIF device.                                                               | ""                             |
@@ -412,13 +429,53 @@ Remember the build step of the `web` part, during build time we move the build d
 
 ## Building for Docker
 
-Inside the root of this `agent` repository, you will find a `Dockerfile`. This file contains the instructions for building and shipping a **Kerberos Agent**. Important to note is that you start from a prebuilt base image, `kerberos/base:xxx`.
-This base image already contains a couple of tools, such as Golang, FFmpeg and OpenCV. We do this for faster compilation times.
+Inside the root of this `agent` repository, you will find a `Dockerfile`. This file contains the instructions for building and shipping a **Kerberos Agent**. It uses Debian Trixie to support the native dependencies used by the Agent, including Media over QUIC.
 
 By running the `docker build` command, you will create the Kerberos Agent Docker image. After building you can simply run the image as a Docker container.
 
     docker build -t kerberos/agent .
 
+### Media over QUIC
+
+The standard AMD64 and ARM64 images include the optional MoQ publisher. Its Rust
+FFI archive requires CGO and glibc 2.38 or newer, which is why the standard image
+uses Debian Trixie. The publisher is disabled unless explicitly enabled at runtime:
+
+    docker run --rm -p 80:80 \
+      -e AGENT_LIVE_MOQ_ENABLED=true \
+      -e AGENT_LIVE_MOQ_URL=https://relay.uug.ai/anon \
+      kerberos/agent
+
+`AGENT_LIVE_MOQ_BROADCAST_PREFIX` defaults to `devices`. MoQ viewers subscribe to
+a relay and never negotiate with the Agent, so every quality tier is published as
+its own broadcast and switching quality is simply a resubscribe:
+
+| Tier   | Broadcast                             | Source                                     |
+| ------ | ------------------------------------- | ------------------------------------------ |
+| `high` | `devices/<agent-key>/live.hang`       | highest-resolution camera stream           |
+| `low`  | `devices/<agent-key>/live-low.hang`   | sub stream (main stream when none is set)  |
+
+Each tier only uploads while it has at least one subscriber, so the tier nobody
+watches costs virtually no bandwidth. `AGENT_LIVE_MOQ_QUALITY` accepts `high` or
+`low` to pin the Agent to a single tier; viewers requesting the other tier then
+find no broadcast. Any other value (including the default) publishes both. The
+initial implementation publishes H.264 video only.
+
+The `/anon` relay route is intended for interoperability testing. Production
+deployments must set `AGENT_LIVE_MOQ_URL` to a short-lived, device-scoped
+publisher URL issued by Hub API.
+
+To verify the native SDK in a development container, rebuild the Agent or shared
+monorepo devcontainer so it uses the Trixie base, then run the VS Code task
+`agent: moq verify`. The same check is available from a terminal:
+
+    cd machinery
+    bash ./verify-moq-devcontainer.sh
+
+The check requires glibc 2.38 or newer, runs the tagged package tests, links the
+complete Agent with `-tags moq`, and executes the resulting binary's version
+command. Both devcontainers also run this check during their post-create setup.
+
 ## What is new?
 
 This repository contains the next generation of Kerberos.io, **Kerberos Agent (v3)**, and is the successor of the machinery and web repositories. A switch in technologies and architecture has been made. This version is still under active development and can be followed on the [develop branch](https://github.com/kerberos-io/agent/tree/develop) and [project overview](https://github.com/kerberos-io/agent/projects/1).

machinery/data/tus/1786715908_3-737_encrypted_0-0-0-0_-1_25498.mp4.primary.json

@@ -0,0 +1 @@
+{"upload_url":"https://vault.kerberos.io/api/storage/tus/19e42fbc666a38064904caf8c46d182a","vault_uri":"https://vault.kerberos.io/api/storage/tus/","size":1591581}

machinery/go.mod

@@ -1,6 +1,6 @@
 module github.com/kerberos-io/agent/machinery
 
-go 1.24.2
+go 1.25.0
 
 replace google.golang.org/genproto => google.golang.org/genproto v0.0.0-20250519155744-55703ea1f237
 
@@ -8,7 +8,7 @@ require (
 	github.com/Eyevinn/mp4ff v0.48.0
 	github.com/InVisionApp/conjungo v1.1.0
 	github.com/appleboy/gin-jwt/v2 v2.10.3
-	github.com/bluenviron/gortsplib/v4 v4.14.1
+	github.com/bluenviron/gortsplib/v5 v5.6.3
 	github.com/bluenviron/mediacommon v1.14.0
 	github.com/cedricve/go-onvif v0.0.0-20200222191200-567e8ce298f6
 	github.com/dromara/carbon/v2 v2.6.8
@@ -24,14 +24,16 @@ require (
 	github.com/gorilla/websocket v1.5.3
 	github.com/kellydunn/golang-geo v0.7.0
 	github.com/kerberos-io/joy4 v1.0.64
-	github.com/kerberos-io/onvif v1.2.1
+	github.com/kerberos-io/onvif v1.2.2
 	github.com/minio/minio-go/v6 v6.0.57
+	github.com/moq-dev/moq-go v0.5.7
 	github.com/nfnt/resize v0.0.0-20180221191011-83c6a9932646
 	github.com/op/go-logging v0.0.0-20160315200505-970db520ece7
-	github.com/pion/interceptor v0.1.40
-	github.com/pion/rtp v1.8.19
-	github.com/pion/webrtc/v4 v4.1.2
+	github.com/pion/interceptor v0.1.47
+	github.com/pion/rtp v1.10.5
+	github.com/pion/webrtc/v4 v4.2.18
 	github.com/sirupsen/logrus v1.9.3
+	github.com/stretchr/testify v1.11.1
 	github.com/swaggo/files v1.0.1
 	github.com/swaggo/gin-swagger v1.6.0
 	github.com/swaggo/swag v1.16.4
@@ -51,13 +53,14 @@ require (
 	github.com/PuerkitoBio/purell v1.1.1 // indirect
 	github.com/PuerkitoBio/urlesc v0.0.0-20170810143723-de5bf2ad4578 // indirect
 	github.com/beevik/etree v1.2.0 // indirect
-	github.com/bluenviron/mediacommon/v2 v2.2.0 // indirect
+	github.com/bluenviron/mediacommon/v2 v2.9.2 // indirect
 	github.com/bytedance/sonic v1.13.2 // indirect
 	github.com/bytedance/sonic/loader v0.2.4 // indirect
 	github.com/cenkalti/backoff/v5 v5.0.2 // indirect
 	github.com/clbanning/mxj v1.8.4 // indirect
 	github.com/clbanning/mxj/v2 v2.7.0 // indirect
 	github.com/cloudwego/base64x v0.1.5 // indirect
+	github.com/davecgh/go-spew v1.1.1 // indirect
 	github.com/elastic/go-windows v1.0.2 // indirect
 	github.com/elgs/gostrgen v0.0.0-20161222160715-9d61ae07eeae // indirect
 	github.com/erikstmartin/go-testdb v0.0.0-20160219214506-8d10e4a1bae5 // indirect
@@ -93,21 +96,23 @@ require (
 	github.com/modern-go/concurrent v0.0.0-20180306012644-bacd9c7ef1dd // indirect
 	github.com/modern-go/reflect2 v1.0.2 // indirect
 	github.com/montanaflynn/stats v0.7.1 // indirect
+	github.com/moq-dev/moq-go-ffi v0.3.7 // indirect
 	github.com/nxadm/tail v1.4.11 // indirect
 	github.com/pelletier/go-toml/v2 v2.2.3 // indirect
-	github.com/pion/datachannel v1.5.10 // indirect
-	github.com/pion/dtls/v3 v3.0.6 // indirect
-	github.com/pion/ice/v4 v4.0.10 // indirect
-	github.com/pion/logging v0.2.3 // indirect
-	github.com/pion/mdns/v2 v2.0.7 // indirect
+	github.com/pion/datachannel v1.6.2 // indirect
+	github.com/pion/dtls/v3 v3.1.5 // indirect
+	github.com/pion/ice/v4 v4.4.0 // indirect
+	github.com/pion/logging v0.2.4 // indirect
+	github.com/pion/mdns/v2 v2.1.0 // indirect
 	github.com/pion/randutil v0.1.0 // indirect
-	github.com/pion/rtcp v1.2.15 // indirect
-	github.com/pion/sctp v1.8.39 // indirect
-	github.com/pion/sdp/v3 v3.0.13 // indirect
-	github.com/pion/srtp/v3 v3.0.5 // indirect
-	github.com/pion/stun/v3 v3.0.0 // indirect
-	github.com/pion/transport/v3 v3.0.7 // indirect
-	github.com/pion/turn/v4 v4.0.0 // indirect
+	github.com/pion/rtcp v1.2.17 // indirect
+	github.com/pion/sctp v1.11.1 // indirect
+	github.com/pion/sdp/v3 v3.0.19 // indirect
+	github.com/pion/srtp/v3 v3.0.12 // indirect
+	github.com/pion/stun/v3 v3.1.6 // indirect
+	github.com/pion/transport/v4 v4.0.2 // indirect
+	github.com/pion/turn/v5 v5.0.12 // indirect
+	github.com/pmezard/go-difflib v1.0.0 // indirect
 	github.com/prometheus/procfs v0.15.1 // indirect
 	github.com/twitchyliquid64/golang-asm v0.15.1 // indirect
 	github.com/ugorji/go/codec v1.2.12 // indirect
@@ -121,13 +126,14 @@ require (
 	go.opentelemetry.io/otel/metric v1.36.0 // indirect
 	go.opentelemetry.io/proto/otlp v1.6.0 // indirect
 	golang.org/x/arch v0.16.0 // indirect
-	golang.org/x/crypto v0.38.0 // indirect
-	golang.org/x/net v0.40.0 // indirect
+	golang.org/x/crypto v0.54.0 // indirect
+	golang.org/x/net v0.57.0 // indirect
 	golang.org/x/oauth2 v0.30.0 // indirect
-	golang.org/x/sync v0.14.0 // indirect
-	golang.org/x/sys v0.33.0 // indirect
-	golang.org/x/text v0.25.0 // indirect
-	golang.org/x/tools v0.30.0 // indirect
+	golang.org/x/sync v0.22.0 // indirect
+	golang.org/x/sys v0.47.0 // indirect
+	golang.org/x/text v0.40.0 // indirect
+	golang.org/x/time v0.14.0 // indirect
+	golang.org/x/tools v0.47.0 // indirect
 	google.golang.org/genproto/googleapis/api v0.0.0-20250519155744-55703ea1f237 // indirect
 	google.golang.org/genproto/googleapis/rpc v0.0.0-20250519155744-55703ea1f237 // indirect
 	google.golang.org/grpc v1.72.1 // indirect

machinery/go.sum

@@ -393,12 +393,12 @@ github.com/appleboy/gofight/v2 v2.1.2/go.mod h1:frW+U1QZEdDgixycTj4CygQ48yLTUhpl
 github.com/bazelbuild/rules_go v0.49.0/go.mod h1:Dhcz716Kqg1RHNWos+N6MlXNkjNP2EwZQ0LukRKJfMs=
 github.com/beevik/etree v1.2.0 h1:l7WETslUG/T+xOPs47dtd6jov2Ii/8/OjCldk5fYfQw=
 github.com/beevik/etree v1.2.0/go.mod h1:aiPf89g/1k3AShMVAzriilpcE4R/Vuor90y83zVZWFc=
-github.com/bluenviron/gortsplib/v4 v4.14.1 h1:v99NmXeeJFfbrO+ipPzPxYGibQaR5ZOUESOA9UQZhsI=
-github.com/bluenviron/gortsplib/v4 v4.14.1/go.mod h1:3LaEcg0d47+kfXju5KSlsSxCiZ3IKBI/sqIrBPcsS64=
+github.com/bluenviron/gortsplib/v5 v5.6.3 h1:OXvHthQZ9fZbLh6r3Go2wuF4XQ4/QW4WTIM2f4bv/W4=
+github.com/bluenviron/gortsplib/v5 v5.6.3/go.mod h1:kzHgUtvl8NWNsQ5Vsez6Vuugk6ItFT4ByCPm8J/kbSQ=
 github.com/bluenviron/mediacommon v1.14.0 h1:lWCwOBKNKgqmspRpwpvvg3CidYm+XOc2+z/Jw7LM5dQ=
 github.com/bluenviron/mediacommon v1.14.0/go.mod h1:z5LP9Tm1ZNfQV5Co54PyOzaIhGMusDfRKmh42nQSnyo=
-github.com/bluenviron/mediacommon/v2 v2.2.0 h1:fGXEX0OEvv5VhGHOv3Q2ABzOtSkIpl9UbwOHrnKWNTk=
-github.com/bluenviron/mediacommon/v2 v2.2.0/go.mod h1:a6MbPmXtYda9mKibKVMZlW20GYLLrX2R7ZkUE+1pwV0=
+github.com/bluenviron/mediacommon/v2 v2.9.2 h1:jvYeBjvhHKFOBRMTMm4hvrSjyOlCelkOkx6708DidQM=
+github.com/bluenviron/mediacommon/v2 v2.9.2/go.mod h1:jMf/OJDaJl02xRgkLM2zbidUHnDYqLnO1dMMveCmyyU=
 github.com/boombuler/barcode v1.0.0/go.mod h1:paBWMcWSl3LHKBqUq+rly7CNSldXjb2rDl3JlRe0mD8=
 github.com/boombuler/barcode v1.0.1/go.mod h1:paBWMcWSl3LHKBqUq+rly7CNSldXjb2rDl3JlRe0mD8=
 github.com/bytedance/sonic v1.13.2 h1:8/H1FempDZqC4VqjptGo14QQlJx8VdZJegxs6wwfqpQ=
@@ -774,8 +774,8 @@ github.com/kellydunn/golang-geo v0.7.0 h1:A5j0/BvNgGwY6Yb6inXQxzYwlPHc6WVZR+Mrar
 github.com/kellydunn/golang-geo v0.7.0/go.mod h1:YYlQPJ+DPEzrHx8kT3oPHC/NjyvCCXE+IuKGKdrjrcU=
 github.com/kerberos-io/joy4 v1.0.64 h1:gTUSotHSOhp9mNqEecgq88tQHvpj7TjmrvPUsPm0idg=
 github.com/kerberos-io/joy4 v1.0.64/go.mod h1:nZp4AjvKvTOXRrmDyAIOw+Da+JA5OcSo/JundGfOlFU=
-github.com/kerberos-io/onvif v1.2.1 h1:+vxyHPylt0ufK8gv7FL+KzhJUeComMGrTmP5KxT2YEc=
-github.com/kerberos-io/onvif v1.2.1/go.mod h1:XSgEQXmEDjUQTbdXvsaRJt6Az8YPGj7L+j5iXKEGijU=
+github.com/kerberos-io/onvif v1.2.2 h1:QnxITps7xvAVD2abWRsa3+p9QexjKESQQldUFMx/mYA=
+github.com/kerberos-io/onvif v1.2.2/go.mod h1:XSgEQXmEDjUQTbdXvsaRJt6Az8YPGj7L+j5iXKEGijU=
 github.com/kisielk/errcheck v1.5.0/go.mod h1:pFxgyoBC7bSaBwPgfKdkLd5X25qrDl4LWUI2bnpBCr8=
 github.com/kisielk/gotool v1.0.0/go.mod h1:XhKaO+MFFWcvkIS/tQcRk01m1F5IRFswLeQ+oQHNcck=
 github.com/klauspost/asmfmt v1.3.2/go.mod h1:AG8TuvYojzulgDAMCnYn50l/5QV3Bs/tp6j0HLHbNSE=
@@ -845,6 +845,10 @@ github.com/modern-go/reflect2 v1.0.2 h1:xBagoLtFs94CBntxluKeaWgTMpvLxC4ur3nMaC9G
 github.com/modern-go/reflect2 v1.0.2/go.mod h1:yWuevngMOJpCy52FWWMvUC8ws7m/LJsjYzDa0/r8luk=
 github.com/montanaflynn/stats v0.7.1 h1:etflOAAHORrCC44V+aR6Ftzort912ZU+YLiSTuV8eaE=
 github.com/montanaflynn/stats v0.7.1/go.mod h1:etXPPgVO6n31NxCd9KQUMvCM+ve0ruNzt6R8Bnaayow=
+github.com/moq-dev/moq-go v0.5.7 h1:LfFpgAU8FRMcnU85L5Lb03HzxhYbQ2+BhT/pJQdjy5U=
+github.com/moq-dev/moq-go v0.5.7/go.mod h1:5K8zjKKjWe5lzfCtlfxoAfdCr3KX6EBuTOZsz0WBnzw=
+github.com/moq-dev/moq-go-ffi v0.3.7 h1:+xwPOzTJHvB0tuTnW6znbvvOgT4yLnN7sMqXXZYh02M=
+github.com/moq-dev/moq-go-ffi v0.3.7/go.mod h1:zxpOlUetvaoxWBnbXTdILtLLUytaoVuLvO36lftwWO0=
 github.com/nfnt/resize v0.0.0-20180221191011-83c6a9932646 h1:zYyBkD/k9seD2A7fsi6Oo2LfFZAehjjQMERAvZLEDnQ=
 github.com/nfnt/resize v0.0.0-20180221191011-83c6a9932646/go.mod h1:jpp1/29i3P1S/RLdc7JQKbRpFeM1dOBd8T9ki5s+AY8=
 github.com/niemeyer/pretty v0.0.0-20200227124842-a10e7caefd8e/go.mod h1:zD1mROLANZcx1PVRCS0qkT7pwLkGfwJo4zjcN/Tysno=
@@ -862,38 +866,40 @@ github.com/phpdave11/gofpdf v1.4.2/go.mod h1:zpO6xFn9yxo3YLyMvW8HcKWVdbNqgIfOOp2
 github.com/phpdave11/gofpdi v1.0.12/go.mod h1:vBmVV0Do6hSBHC8uKUQ71JGW+ZGQq74llk/7bXwjDoI=
 github.com/phpdave11/gofpdi v1.0.13/go.mod h1:vBmVV0Do6hSBHC8uKUQ71JGW+ZGQq74llk/7bXwjDoI=
 github.com/pierrec/lz4/v4 v4.1.18/go.mod h1:gZWDp/Ze/IJXGXf23ltt2EXimqmTUXEy0GFuRQyBid4=
-github.com/pion/datachannel v1.5.10 h1:ly0Q26K1i6ZkGf42W7D4hQYR90pZwzFOjTq5AuCKk4o=
-github.com/pion/datachannel v1.5.10/go.mod h1:p/jJfC9arb29W7WrxyKbepTU20CFgyx5oLo8Rs4Py/M=
-github.com/pion/dtls/v3 v3.0.6 h1:7Hkd8WhAJNbRgq9RgdNh1aaWlZlGpYTzdqjy9x9sK2E=
-github.com/pion/dtls/v3 v3.0.6/go.mod h1:iJxNQ3Uhn1NZWOMWlLxEEHAN5yX7GyPvvKw04v9bzYU=
-github.com/pion/ice/v4 v4.0.10 h1:P59w1iauC/wPk9PdY8Vjl4fOFL5B+USq1+xbDcN6gT4=
-github.com/pion/ice/v4 v4.0.10/go.mod h1:y3M18aPhIxLlcO/4dn9X8LzLLSma84cx6emMSu14FGw=
-github.com/pion/interceptor v0.1.40 h1:e0BjnPcGpr2CFQgKhrQisBU7V3GXK6wrfYrGYaU6Jq4=
-github.com/pion/interceptor v0.1.40/go.mod h1:Z6kqH7M/FYirg3frjGJ21VLSRJGBXB/KqaTIrdqnOic=
-github.com/pion/logging v0.2.3 h1:gHuf0zpoh1GW67Nr6Gj4cv5Z9ZscU7g/EaoC/Ke/igI=
-github.com/pion/logging v0.2.3/go.mod h1:z8YfknkquMe1csOrxK5kc+5/ZPAzMxbKLX5aXpbpC90=
-github.com/pion/mdns/v2 v2.0.7 h1:c9kM8ewCgjslaAmicYMFQIde2H9/lrZpjBkN8VwoVtM=
-github.com/pion/mdns/v2 v2.0.7/go.mod h1:vAdSYNAT0Jy3Ru0zl2YiW3Rm/fJCwIeM0nToenfOJKA=
+github.com/pion/datachannel v1.6.2 h1:7EXQ8TH3vTouBUdRWYbcX2edSx9Yj6k5zl5P+qyxEPc=
+github.com/pion/datachannel v1.6.2/go.mod h1:pzbdAZvyGtXbcHM1hBbsFaOTf40lZizU/dNlvVOak6E=
+github.com/pion/dtls/v3 v3.1.5 h1:9xJtVsHwMYeSjPp5Hh1FTis4DchnQWtnOa5o+6ygqfc=
+github.com/pion/dtls/v3 v3.1.5/go.mod h1:gz1K4jg6c+fq86oQMH4pilpCEOEPwmEr2jY+VcF/mkU=
+github.com/pion/ice/v4 v4.4.0 h1:wvHDDqimaC38Y7MVpD46Y63p246ChvXd87VKoLYS5b4=
+github.com/pion/ice/v4 v4.4.0/go.mod h1:obAyD+J+Hzs7QA7Y8YXHp5uIn6gb7z87pKedXZkrcFU=
+github.com/pion/interceptor v0.1.47 h1:yw8t5pJ2f8t78NgU+8EmxhaqYLXS7uFCC/tAGOaSDBo=
+github.com/pion/interceptor v0.1.47/go.mod h1:7yoRBzaIDETPC6cIN8Zj9EyGqHv1ImOpcTFPha6MuOM=
+github.com/pion/logging v0.2.4 h1:tTew+7cmQ+Mc1pTBLKH2puKsOvhm32dROumOZ655zB8=
+github.com/pion/logging v0.2.4/go.mod h1:DffhXTKYdNZU+KtJ5pyQDjvOAh/GsNSyv1lbkFbe3so=
+github.com/pion/mdns/v2 v2.1.0 h1:3IJ9+Xio6tWYjhN6WwuY142P/1jA0D5ERaIqawg/fOY=
+github.com/pion/mdns/v2 v2.1.0/go.mod h1:pcez23GdynwcfRU1977qKU0mDxSeucttSHbCSfFOd9A=
 github.com/pion/randutil v0.1.0 h1:CFG1UdESneORglEsnimhUjf33Rwjubwj6xfiOXBa3mA=
 github.com/pion/randutil v0.1.0/go.mod h1:XcJrSMMbbMRhASFVOlj/5hQial/Y8oH/HVo7TBZq+j8=
-github.com/pion/rtcp v1.2.15 h1:LZQi2JbdipLOj4eBjK4wlVoQWfrZbh3Q6eHtWtJBZBo=
-github.com/pion/rtcp v1.2.15/go.mod h1:jlGuAjHMEXwMUHK78RgX0UmEJFV4zUKOFHR7OP+D3D0=
-github.com/pion/rtp v1.8.19 h1:jhdO/3XhL/aKm/wARFVmvTfq0lC/CvN1xwYKmduly3c=
-github.com/pion/rtp v1.8.19/go.mod h1:bAu2UFKScgzyFqvUKmbvzSdPr+NGbZtv6UB2hesqXBk=
-github.com/pion/sctp v1.8.39 h1:PJma40vRHa3UTO3C4MyeJDQ+KIobVYRZQZ0Nt7SjQnE=
-github.com/pion/sctp v1.8.39/go.mod h1:cNiLdchXra8fHQwmIoqw0MbLLMs+f7uQ+dGMG2gWebE=
-github.com/pion/sdp/v3 v3.0.13 h1:uN3SS2b+QDZnWXgdr69SM8KB4EbcnPnPf2Laxhty/l4=
-github.com/pion/sdp/v3 v3.0.13/go.mod h1:88GMahN5xnScv1hIMTqLdu/cOcUkj6a9ytbncwMCq2E=
-github.com/pion/srtp/v3 v3.0.5 h1:8XLB6Dt3QXkMkRFpoqC3314BemkpMQK2mZeJc4pUKqo=
-github.com/pion/srtp/v3 v3.0.5/go.mod h1:r1G7y5r1scZRLe2QJI/is+/O83W2d+JoEsuIexpw+uM=
-github.com/pion/stun/v3 v3.0.0 h1:4h1gwhWLWuZWOJIJR9s2ferRO+W3zA/b6ijOI6mKzUw=
-github.com/pion/stun/v3 v3.0.0/go.mod h1:HvCN8txt8mwi4FBvS3EmDghW6aQJ24T+y+1TKjB5jyU=
-github.com/pion/transport/v3 v3.0.7 h1:iRbMH05BzSNwhILHoBoAPxoB9xQgOaJk+591KC9P1o0=
-github.com/pion/transport/v3 v3.0.7/go.mod h1:YleKiTZ4vqNxVwh77Z0zytYi7rXHl7j6uPLGhhz9rwo=
-github.com/pion/turn/v4 v4.0.0 h1:qxplo3Rxa9Yg1xXDxxH8xaqcyGUtbHYw4QSCvmFWvhM=
-github.com/pion/turn/v4 v4.0.0/go.mod h1:MuPDkm15nYSklKpN8vWJ9W2M0PlyQZqYt1McGuxG7mA=
-github.com/pion/webrtc/v4 v4.1.2 h1:mpuUo/EJ1zMNKGE79fAdYNFZBX790KE7kQQpLMjjR54=
-github.com/pion/webrtc/v4 v4.1.2/go.mod h1:xsCXiNAmMEjIdFxAYU0MbB3RwRieJsegSB2JZsGN+8U=
+github.com/pion/rtcp v1.2.17 h1:PxiT6L79yPZKtXIsXdG1eakBl6dtBj4x+4oVEL0DlSw=
+github.com/pion/rtcp v1.2.17/go.mod h1:7kBpuBJaWwax4hzc/pgexY8vkOpvh8atgYDbaKZq0iU=
+github.com/pion/rtp v1.10.5 h1:ip0HhO/wYZqQ4bKS+R99KnZh/GRCmIT0jDXikub7vlE=
+github.com/pion/rtp v1.10.5/go.mod h1:Au8fc6cEByy8RLTwKTQTEeQqDB/SJDxwL4mZuxYA5Pk=
+github.com/pion/sctp v1.11.1 h1:O4dIFyURw1KTST7w+gtD4gLeYXkhPa0xXLHMMoe/OSA=
+github.com/pion/sctp v1.11.1/go.mod h1:7KFmTwLcoYgJs/Z+99nJvsWL0qDpuyloSI0RbAqlrz0=
+github.com/pion/sdp/v3 v3.0.19 h1:1VMKs3gIkTQV5M3hNKfTAPrDXSNrYtOlmOD8+mSZUGQ=
+github.com/pion/sdp/v3 v3.0.19/go.mod h1:dE5WOSlzXrtiE/iuZqe9n+AcEbOjtAd3k5m5NtlV/qU=
+github.com/pion/srtp/v3 v3.0.12 h1:U7V17bckl7sI4mb3sepiojByDuBY0wNCqQE+6IlQBbc=
+github.com/pion/srtp/v3 v3.0.12/go.mod h1:EeZOi/sd6glM1EXapg051gdNWO9yWT1YSsgQ4SlJkns=
+github.com/pion/stun/v3 v3.1.6 h1:WnhsD0eHCiwCfKNkVx0VJJwr2Y3eV4Ueih3KJ+dfZy8=
+github.com/pion/stun/v3 v3.1.6/go.mod h1:zRUghXSQU32Lx5orJsz3uYMkIihweXb3mu5gIns02fs=
+github.com/pion/transport/v3 v3.1.1 h1:Tr684+fnnKlhPceU+ICdrw6KKkTms+5qHMgw6bIkYOM=
+github.com/pion/transport/v3 v3.1.1/go.mod h1:+c2eewC5WJQHiAA46fkMMzoYZSuGzA/7E2FPrOYHctQ=
+github.com/pion/transport/v4 v4.0.2 h1:ifYlPqNwsy6aKQ9y8yzxXlHae5431ZrH2avkD/Rn6Tk=
+github.com/pion/transport/v4 v4.0.2/go.mod h1:06hFI+jCFcok2X2MekVufNZ/uzNZXivGBPfviSVcjgM=
+github.com/pion/turn/v5 v5.0.12 h1:6+b69ivQQXSlyfkp2AKripqD2k3W32qXK8QzCzpJWPI=
+github.com/pion/turn/v5 v5.0.12/go.mod h1:CQACsRDJtjQ+6RSrGHrS2PCIerLwbW3uqXRqOvtjAFg=
+github.com/pion/webrtc/v4 v4.2.18 h1:smA/3g6Gy4RohM0VIZ5KKY/12TQbxv3XFgpUMyb2EUI=
+github.com/pion/webrtc/v4 v4.2.18/go.mod h1:vmzi6s+rvhoIuT94DPqivB+0xJXs9rG4QRD+4MgBtlY=
 github.com/pkg/diff v0.0.0-20210226163009-20ebb0f2a09e/go.mod h1:pJLUxLENpZxwdsKMEsNbx1VGcRFpLqf3715MtcvvzbA=
 github.com/pkg/errors v0.8.1/go.mod h1:bwawxfHBFNV+L2hUp1rHADufV3IMtnDRdf1r5NINEl0=
 github.com/pkg/errors v0.9.1/go.mod h1:bwawxfHBFNV+L2hUp1rHADufV3IMtnDRdf1r5NINEl0=
@@ -953,8 +959,9 @@ github.com/stretchr/testify v1.8.2/go.mod h1:w2LPCIKwWwSfY2zedu0+kehJoqGctiVI29o
 github.com/stretchr/testify v1.8.3/go.mod h1:sz/lmYIOXD/1dqDmKjjqLyZ2RngseejIcXlSw2iwfAo=
 github.com/stretchr/testify v1.8.4/go.mod h1:sz/lmYIOXD/1dqDmKjjqLyZ2RngseejIcXlSw2iwfAo=
 github.com/stretchr/testify v1.9.0/go.mod h1:r2ic/lqez/lEtzL7wO/rwa5dbSLXVDPFyf8C91i36aY=
-github.com/stretchr/testify v1.10.0 h1:Xv5erBjTwe/5IxqUQTdXv5kgmIvbHo3QQyRwhJsOfJA=
 github.com/stretchr/testify v1.10.0/go.mod h1:r2ic/lqez/lEtzL7wO/rwa5dbSLXVDPFyf8C91i36aY=
+github.com/stretchr/testify v1.11.1 h1:7s2iGBzp5EwR7/aIZr8ao5+dra3wiQyKjjFuvgVKu7U=
+github.com/stretchr/testify v1.11.1/go.mod h1:wZwfW3scLgRK+23gO65QZefKpKQRnfz6sD981Nm4B6U=
 github.com/substrait-io/substrait-go v0.4.2/go.mod h1:qhpnLmrcvAnlZsUyPXZRqldiHapPTXC3t7xFgDi3aQg=
 github.com/swaggo/files v1.0.1 h1:J1bVJ4XHZNq0I46UU90611i9/YzdrF7x92oX1ig5IdE=
 github.com/swaggo/files v1.0.1/go.mod h1:0qXmMNH6sXNf+73t65aKeB+ApmgxdnkQzVTAj2uaMUg=
@@ -1162,8 +1169,8 @@ golang.org/x/crypto v0.33.0/go.mod h1:bVdXmD7IV/4GdElGPozy6U7lWdRXA4qyRVGJV57uQ5
 golang.org/x/crypto v0.35.0/go.mod h1:dy7dXNW32cAb/6/PRuTNsix8T+vJAqvuIy5Bli/x0YQ=
 golang.org/x/crypto v0.36.0/go.mod h1:Y4J0ReaxCR1IMaabaSMugxJES1EpwhBHhv2bDHklZvc=
 golang.org/x/crypto v0.37.0/go.mod h1:vg+k43peMZ0pUMhYmVAWysMK35e6ioLh3wB8ZCAfbVc=
-golang.org/x/crypto v0.38.0 h1:jt+WWG8IZlBnVbomuhg2Mdq0+BBQaHbtqHEFEigjUV8=
-golang.org/x/crypto v0.38.0/go.mod h1:MvrbAqul58NNYPKnOra203SB9vpuZW0e+RRZV+Ggqjw=
+golang.org/x/crypto v0.54.0 h1:YLIA59K4fiNzHzjnZt2tUJQjQtUWfWbeHBqKtk3eScw=
+golang.org/x/crypto v0.54.0/go.mod h1:KWL8ny2AZdGR2cWmzeHrp2azQPGogOv+HeQaVEXC2dk=
 golang.org/x/exp v0.0.0-20180321215751-8460e604b9de/go.mod h1:CJ0aWSM057203Lf6IL+f9T1iT9GByDxfZKAQTCR3kQA=
 golang.org/x/exp v0.0.0-20180807140117-3d87b88a115f/go.mod h1:CJ0aWSM057203Lf6IL+f9T1iT9GByDxfZKAQTCR3kQA=
 golang.org/x/exp v0.0.0-20190121172915-509febef88a4/go.mod h1:CJ0aWSM057203Lf6IL+f9T1iT9GByDxfZKAQTCR3kQA=
@@ -1232,8 +1239,9 @@ golang.org/x/mod v0.15.0/go.mod h1:hTbmBsO62+eylJbnUtE2MGJUyE7QWk4xUqPFrRgJ+7c=
 golang.org/x/mod v0.17.0/go.mod h1:hTbmBsO62+eylJbnUtE2MGJUyE7QWk4xUqPFrRgJ+7c=
 golang.org/x/mod v0.18.0/go.mod h1:hTbmBsO62+eylJbnUtE2MGJUyE7QWk4xUqPFrRgJ+7c=
 golang.org/x/mod v0.19.0/go.mod h1:hTbmBsO62+eylJbnUtE2MGJUyE7QWk4xUqPFrRgJ+7c=
-golang.org/x/mod v0.23.0 h1:Zb7khfcRGKk+kqfxFaP5tZqCnDZMjC5VtUBs87Hr6QM=
 golang.org/x/mod v0.23.0/go.mod h1:6SkKJ3Xj0I0BrPOZoBy3bdMptDDU9oJrpohJ3eWZ1fY=
+golang.org/x/mod v0.37.0 h1:vF1DjpVEshcIqoEaauuHebaLk1O1forxjxBaVn884JQ=
+golang.org/x/mod v0.37.0/go.mod h1:m8S8VeM9r4dzDwjrKO0a1sZP3YjeMamRRlD+fmR2Q/0=
 golang.org/x/net v0.0.0-20180724234803-3673e40ba225/go.mod h1:mL1N/T3taQHkDXs73rZJwtUhF3w3ftmwwsq0BUmARs4=
 golang.org/x/net v0.0.0-20180826012351-8a410e7b638d/go.mod h1:mL1N/T3taQHkDXs73rZJwtUhF3w3ftmwwsq0BUmARs4=
 golang.org/x/net v0.0.0-20190108225652-1e06a53dbb7e/go.mod h1:mL1N/T3taQHkDXs73rZJwtUhF3w3ftmwwsq0BUmARs4=
@@ -1321,8 +1329,8 @@ golang.org/x/net v0.34.0/go.mod h1:di0qlW3YNM5oh6GqDGQr92MyTozJPmybPK4Ev/Gm31k=
 golang.org/x/net v0.35.0/go.mod h1:EglIi67kWsHKlRzzVMUD93VMSWGFOMSZgxFjparz1Qk=
 golang.org/x/net v0.37.0/go.mod h1:ivrbrMbzFq5J41QOQh0siUuly180yBYtLp+CKbEaFx8=
 golang.org/x/net v0.39.0/go.mod h1:X7NRbYVEA+ewNkCNyJ513WmMdQ3BineSwVtN2zD/d+E=
-golang.org/x/net v0.40.0 h1:79Xs7wF06Gbdcg4kdCCIQArK11Z1hr5POQ6+fIYHNuY=
-golang.org/x/net v0.40.0/go.mod h1:y0hY0exeL2Pku80/zKK7tpntoX23cqL3Oa6njdgRtds=
+golang.org/x/net v0.57.0 h1:K5+3DljvIuDG9/Jv9rvyMywYNFCQ9RSUY6OOTTkT+tE=
+golang.org/x/net v0.57.0/go.mod h1:KpXc8iv+r3XplLAG/f7Jsf9RPszJzdR0f58q9vGOuEU=
 golang.org/x/oauth2 v0.0.0-20180821212333-d2e6202438be/go.mod h1:N/0e6XlmueqKjAGxoOufVs8QHGRruUQn6yWY3a++T0U=
 golang.org/x/oauth2 v0.0.0-20190226205417-e64efc72b421/go.mod h1:gOpvHmFTYa4IltrdGE7lF6nIHvwfUNPOp7c8zoXwtLw=
 golang.org/x/oauth2 v0.0.0-20190604053449-0f29369cfe45/go.mod h1:gOpvHmFTYa4IltrdGE7lF6nIHvwfUNPOp7c8zoXwtLw=
@@ -1401,8 +1409,9 @@ golang.org/x/sync v0.10.0/go.mod h1:Czt+wKu1gCyEFDUtn0jG5QVvpJ6rzVqr5aXyt9drQfk=
 golang.org/x/sync v0.11.0/go.mod h1:Czt+wKu1gCyEFDUtn0jG5QVvpJ6rzVqr5aXyt9drQfk=
 golang.org/x/sync v0.12.0/go.mod h1:1dzgHSNfp02xaA81J2MS99Qcpr2w7fw1gpm99rleRqA=
 golang.org/x/sync v0.13.0/go.mod h1:1dzgHSNfp02xaA81J2MS99Qcpr2w7fw1gpm99rleRqA=
-golang.org/x/sync v0.14.0 h1:woo0S4Yywslg6hp4eUFjTVOyKt0RookbpAHG4c1HmhQ=
 golang.org/x/sync v0.14.0/go.mod h1:1dzgHSNfp02xaA81J2MS99Qcpr2w7fw1gpm99rleRqA=
+golang.org/x/sync v0.22.0 h1:SZjpbeLmrCk4xhRSZFNZW5gFUeCeFgjekvI/+gfScek=
+golang.org/x/sync v0.22.0/go.mod h1:9xrNwdLfx4jkKbNva9FpL6vEN7evnE43NNNJQ2LF3+0=
 golang.org/x/sys v0.0.0-20180830151530-49385e6e1522/go.mod h1:STP8DvDyc/dI5b8T5hshtkjS+E42TnysNCUPdjciGhY=
 golang.org/x/sys v0.0.0-20190215142949-d0b11bdaac8a/go.mod h1:STP8DvDyc/dI5b8T5hshtkjS+E42TnysNCUPdjciGhY=
 golang.org/x/sys v0.0.0-20190312061237-fead79001313/go.mod h1:h1NjWce9XRLGQEsW7wpKNCjG9DtNlClVuFLEZdDNbEs=
@@ -1511,8 +1520,8 @@ golang.org/x/sys v0.29.0/go.mod h1:/VUhepiaJMQUp4+oa/7Zr1D23ma6VTLIYjOOTFZPUcA=
 golang.org/x/sys v0.30.0/go.mod h1:/VUhepiaJMQUp4+oa/7Zr1D23ma6VTLIYjOOTFZPUcA=
 golang.org/x/sys v0.31.0/go.mod h1:BJP2sWEmIv4KK5OTEluFJCKSidICx8ciO85XgH3Ak8k=
 golang.org/x/sys v0.32.0/go.mod h1:BJP2sWEmIv4KK5OTEluFJCKSidICx8ciO85XgH3Ak8k=
-golang.org/x/sys v0.33.0 h1:q3i8TbbEz+JRD9ywIRlyRAQbM0qF7hu24q3teo2hbuw=
-golang.org/x/sys v0.33.0/go.mod h1:BJP2sWEmIv4KK5OTEluFJCKSidICx8ciO85XgH3Ak8k=
+golang.org/x/sys v0.47.0 h1:o7XGOvZQCADBQQ4Y7VNq2dRWQR7JmOUW8Kxx4ZsNgWs=
+golang.org/x/sys v0.47.0/go.mod h1:4GL1E5IUh+htKOUEOaiffhrAeqysfVGipDYzABqnCmw=
 golang.org/x/telemetry v0.0.0-20240228155512-f48c80bd79b2/go.mod h1:TeRTkGYfJXctD9OcfyVLyj2J3IxLnKwHJR8f4D8a3YE=
 golang.org/x/telemetry v0.0.0-20240521205824-bda55230c457/go.mod h1:pRgIJT+bRLFKnoM1ldnzKoxTIn14Yxz928LQRYYgIN0=
 golang.org/x/term v0.0.0-20201126162022-7de9c90e9dd1/go.mod h1:bj7SfCRtBDWHUb9snDiAeCFNEtKQo2Wmx5Cou7ajbmo=
@@ -1579,8 +1588,8 @@ golang.org/x/text v0.21.0/go.mod h1:4IBbMaMmOPCJ8SecivzSH54+73PCFmPWxNTLm+vZkEQ=
 golang.org/x/text v0.22.0/go.mod h1:YRoo4H8PVmsu+E3Ou7cqLVH8oXWIHVoX0jqUWALQhfY=
 golang.org/x/text v0.23.0/go.mod h1:/BLNzu4aZCJ1+kcD0DNRotWKage4q2rGVAg4o22unh4=
 golang.org/x/text v0.24.0/go.mod h1:L8rBsPeo2pSS+xqN0d5u2ikmjtmoJbDBT1b7nHvFCdU=
-golang.org/x/text v0.25.0 h1:qVyWApTSYLk/drJRO5mDlNYskwQznZmkpV2c8q9zls4=
-golang.org/x/text v0.25.0/go.mod h1:WEdwpYrmk1qmdHvhkSTNPm3app7v4rsT8F2UD6+VHIA=
+golang.org/x/text v0.40.0 h1:Ub2Z6/xjgF1WrYQz2nuITOEegKFtiIy+rieRJ5lHZKs=
+golang.org/x/text v0.40.0/go.mod h1:hpnzDAfGV753zIKo+wk3u1bVKCGPbrnF7+7LBF/UHVY=
 golang.org/x/time v0.0.0-20181108054448-85acf8d2951c/go.mod h1:tRJNPiyCQ0inRvYxbN9jk5I+vvW/OXSQhTDSoE431IQ=
 golang.org/x/time v0.0.0-20190308202827-9d24e82272b4/go.mod h1:tRJNPiyCQ0inRvYxbN9jk5I+vvW/OXSQhTDSoE431IQ=
 golang.org/x/time v0.0.0-20191024005414-555d28b269f0/go.mod h1:tRJNPiyCQ0inRvYxbN9jk5I+vvW/OXSQhTDSoE431IQ=
@@ -1592,6 +1601,8 @@ golang.org/x/time v0.8.0/go.mod h1:3BpzKBy/shNhVucY/MWOyx10tF3SFh9QdLuxbVysPQM=
 golang.org/x/time v0.9.0/go.mod h1:3BpzKBy/shNhVucY/MWOyx10tF3SFh9QdLuxbVysPQM=
 golang.org/x/time v0.10.0/go.mod h1:3BpzKBy/shNhVucY/MWOyx10tF3SFh9QdLuxbVysPQM=
 golang.org/x/time v0.11.0/go.mod h1:CDIdPxbZBQxdj6cxyCIdrNogrJKMJ7pr37NYpMcMDSg=
+golang.org/x/time v0.14.0 h1:MRx4UaLrDotUKUdCIqzPC48t1Y9hANFKIRpNx+Te8PI=
+golang.org/x/time v0.14.0/go.mod h1:eL/Oa2bBBK0TkX57Fyni+NgnyQQN4LitPmob2Hjnqw4=
 golang.org/x/tools v0.0.0-20180525024113-a5b4c53f6e8b/go.mod h1:n7NCudcB/nEzxVGmLbDWY5pfWTLqBcC2KZ6jyYvM4mQ=
 golang.org/x/tools v0.0.0-20180917221912-90fa682c2a6e/go.mod h1:n7NCudcB/nEzxVGmLbDWY5pfWTLqBcC2KZ6jyYvM4mQ=
 golang.org/x/tools v0.0.0-20190114222345-bf090417da8b/go.mod h1:n7NCudcB/nEzxVGmLbDWY5pfWTLqBcC2KZ6jyYvM4mQ=
@@ -1665,8 +1676,9 @@ golang.org/x/tools v0.15.0/go.mod h1:hpksKq4dtpQWS1uQ61JkdqWM3LscIS6Slf+VVkm+wQk
 golang.org/x/tools v0.21.1-0.20240508182429-e35e4ccd0d2d/go.mod h1:aiJjzUbINMkxbQROHiO6hDPo2LHcIPhhQsa9DLh0yGk=
 golang.org/x/tools v0.22.0/go.mod h1:aCwcsjqvq7Yqt6TNyX7QMU2enbQ/Gt0bo6krSeEri+c=
 golang.org/x/tools v0.23.0/go.mod h1:pnu6ufv6vQkll6szChhK3C3L/ruaIv5eBeztNG8wtsI=
-golang.org/x/tools v0.30.0 h1:BgcpHewrV5AUp2G9MebG4XPFI1E2W41zU1SaqVA9vJY=
 golang.org/x/tools v0.30.0/go.mod h1:c347cR/OJfw5TI+GfX7RUPNMdDRRbjvYTS0jPyvsVtY=
+golang.org/x/tools v0.47.0 h1:7Kn5x/d1svx/PzryTsqeoZN4TZwqeH5pGWjefhLi/1Q=
+golang.org/x/tools v0.47.0/go.mod h1:dFHnyTvFWY212G+h7ZY4Vsp/K3U4/7W9TyVaAul8uCA=
 golang.org/x/xerrors v0.0.0-20190717185122-a985d3407aa7/go.mod h1:I/5z698sn9Ka8TeJc9MKroUUfqBBauWjQqLJ2OPfmY0=
 golang.org/x/xerrors v0.0.0-20191011141410-1b5146add898/go.mod h1:I/5z698sn9Ka8TeJc9MKroUUfqBBauWjQqLJ2OPfmY0=
 golang.org/x/xerrors v0.0.0-20191204190536-9bdfabe68543/go.mod h1:I/5z698sn9Ka8TeJc9MKroUUfqBBauWjQqLJ2OPfmY0=

machinery/src/capture/cleanup_test.go

@@ -28,7 +28,8 @@ func writeRecording(t *testing.T, recordingsDir, name string, ageMinutes int) {
 // marking it as still queued for upload.
 func markPending(t *testing.T, cloudDir, name string) {
 	t.Helper()
-	if err := os.WriteFile(filepath.Join(cloudDir, name), nil, 0o644); err != nil {
+	markerName := models.RecordingUploadMetadataFileName(name)
+	if err := os.WriteFile(filepath.Join(cloudDir, markerName), nil, 0o644); err != nil {
 		t.Fatalf("write marker %s: %v", name, err)
 	}
 }
@@ -71,6 +72,24 @@ func TestPickRecordingToCleanup_PrefersUploaded(t *testing.T) {
 	}
 }
 
+func TestPickRecordingToCleanup_RecognizesLegacyMarkerName(t *testing.T) {
+	recordingsDir, cloudDir := newCleanupDirs(t)
+
+	writeRecording(t, recordingsDir, "legacy_pending.mp4", 30)
+	if err := os.WriteFile(filepath.Join(cloudDir, "legacy_pending.mp4"), nil, 0o644); err != nil {
+		t.Fatalf("write legacy marker: %v", err)
+	}
+	writeRecording(t, recordingsDir, "uploaded.mp4", 10)
+
+	name, pending, err := pickRecordingToCleanup(recordingsDir, cloudDir)
+	if err != nil {
+		t.Fatalf("unexpected error: %v", err)
+	}
+	if pending || name != "uploaded.mp4" {
+		t.Fatalf("cleanup picked name=%q pending=%v, want uploaded.mp4 pending=false", name, pending)
+	}
+}
+
 // Among several already-uploaded recordings, the oldest uploaded one is chosen.
 func TestPickRecordingToCleanup_OldestUploadedFirst(t *testing.T) {
 	recordingsDir, cloudDir := newCleanupDirs(t)

machinery/src/capture/gortsplib.go

@@ -8,24 +8,27 @@ import "C"
 
 import (
 	"context"
+	"crypto/tls"
+	"crypto/x509"
 	"errors"
 	"fmt"
 	"image"
+	"os"
 	"reflect"
 	"strconv"
 	"sync"
 	"time"
 	"unsafe"
 
-	"github.com/bluenviron/gortsplib/v4"
-	"github.com/bluenviron/gortsplib/v4/pkg/base"
-	"github.com/bluenviron/gortsplib/v4/pkg/description"
-	"github.com/bluenviron/gortsplib/v4/pkg/format"
-	"github.com/bluenviron/gortsplib/v4/pkg/format/rtph264"
-	"github.com/bluenviron/gortsplib/v4/pkg/format/rtph265"
-	"github.com/bluenviron/gortsplib/v4/pkg/format/rtplpcm"
-	"github.com/bluenviron/gortsplib/v4/pkg/format/rtpmpeg4audio"
-	"github.com/bluenviron/gortsplib/v4/pkg/format/rtpsimpleaudio"
+	"github.com/bluenviron/gortsplib/v5"
+	"github.com/bluenviron/gortsplib/v5/pkg/base"
+	"github.com/bluenviron/gortsplib/v5/pkg/description"
+	"github.com/bluenviron/gortsplib/v5/pkg/format"
+	"github.com/bluenviron/gortsplib/v5/pkg/format/rtph264"
+	"github.com/bluenviron/gortsplib/v5/pkg/format/rtph265"
+	"github.com/bluenviron/gortsplib/v5/pkg/format/rtplpcm"
+	"github.com/bluenviron/gortsplib/v5/pkg/format/rtpmpeg4audio"
+	"github.com/bluenviron/gortsplib/v5/pkg/format/rtpsimpleaudio"
 	"github.com/bluenviron/mediacommon/pkg/codecs/h264"
 	"github.com/bluenviron/mediacommon/pkg/codecs/h265"
 	"github.com/bluenviron/mediacommon/pkg/codecs/mpeg4audio"
@@ -38,6 +41,36 @@ import (
 
 var tracer = otel.Tracer("github.com/kerberos-io/agent/machinery/src/capture")
 
+const (
+	rtspsCAFileEnv   = "AGENT_CAPTURE_IPCAMERA_RTSPS_CA_FILE"
+	rtspsInsecureEnv = "AGENT_CAPTURE_IPCAMERA_RTSPS_INSECURE"
+)
+
+func rtspsTLSConfig() (*tls.Config, error) {
+	if os.Getenv(rtspsInsecureEnv) == "true" {
+		return &tls.Config{InsecureSkipVerify: true}, nil // #nosec G402 -- explicit opt-in for cameras with self-signed certificates
+	}
+
+	caFile := os.Getenv(rtspsCAFileEnv)
+	if caFile == "" {
+		return nil, nil
+	}
+
+	rootCAs, err := x509.SystemCertPool()
+	if err != nil {
+		return nil, fmt.Errorf("load system CA pool: %w", err)
+	}
+	caPEM, err := os.ReadFile(caFile)
+	if err != nil {
+		return nil, fmt.Errorf("read RTSPS CA file %q: %w", caFile, err)
+	}
+	if !rootCAs.AppendCertsFromPEM(caPEM) {
+		return nil, fmt.Errorf("RTSPS CA file %q contains no certificates", caFile)
+	}
+
+	return &tls.Config{RootCAs: rootCAs}, nil
+}
+
 // Implements the RTSPClient interface.
 type Golibrtsp struct {
 	RTSPClient
@@ -317,11 +350,17 @@ func (g *Golibrtsp) Connect(ctx context.Context, ctxOtel context.Context) (err e
 	_, span := tracer.Start(ctxOtel, "Connect")
 	defer span.End()
 
-	transport := gortsplib.TransportTCP
+	tlsConfig, err := rtspsTLSConfig()
+	if err != nil {
+		return fmt.Errorf("configure RTSPS TLS: %w", err)
+	}
+
+	protocol := gortsplib.ProtocolTCP
 	g.health = newStreamHealth()
 	g.Client = gortsplib.Client{
 		RequestBackChannels: false,
-		Transport:           &transport,
+		Protocol:            &protocol,
+		TLSConfig:           tlsConfig,
 		// Route gortsplib's packet-loss / decode-error reporting through our
 		// structured logger with stream context (replaces its plain stdout
 		// logging). These hooks are what let us tell whether the camera is
@@ -342,7 +381,9 @@ func (g *Golibrtsp) Connect(ctx context.Context, ctxOtel context.Context) (err e
 	}
 
 	// connect to the server
-	err = g.Client.Start(u.Scheme, u.Host)
+	g.Client.Scheme = u.Scheme
+	g.Client.Host = u.Host
+	err = g.Client.Start()
 	if err != nil {
 		log.Log.Debug("capture.golibrtsp.Connect(Start): " + err.Error())
 	}
@@ -592,11 +633,17 @@ func (g *Golibrtsp) ConnectBackChannel(ctx context.Context, ctxRunAgent context.
 	_, span := tracer.Start(ctxRunAgent, "ConnectBackChannel")
 	defer span.End()
 
+	tlsConfig, err := rtspsTLSConfig()
+	if err != nil {
+		return fmt.Errorf("configure RTSPS TLS: %w", err)
+	}
+
 	// Transport TCP
-	transport := gortsplib.TransportTCP
+	protocol := gortsplib.ProtocolTCP
 	g.Client = gortsplib.Client{
 		RequestBackChannels: true,
-		Transport:           &transport,
+		Protocol:            &protocol,
+		TLSConfig:           tlsConfig,
 	}
 	// parse URL
 	u, err := base.ParseURL(g.Url)
@@ -606,7 +653,9 @@ func (g *Golibrtsp) ConnectBackChannel(ctx context.Context, ctxRunAgent context.
 	}
 
 	// connect to the server
-	err = g.Client.Start(u.Scheme, u.Host)
+	g.Client.Scheme = u.Scheme
+	g.Client.Host = u.Host
+	err = g.Client.Start()
 	if err != nil {
 		log.Log.Error("capture.golibrtsp.ConnectBackChannel(): " + err.Error())
 	}
@@ -680,6 +729,12 @@ func compositionOffsetMs(ext dtsExtractor, au [][]byte, pts int64, clockRate int
 	return offset * 1000 / int64(clockRate)
 }
 
+func ptsToDuration(pts int64, clockRate int) time.Duration {
+	rate := int64(clockRate)
+	return time.Duration(pts/rate)*time.Second +
+		time.Duration(pts%rate)*time.Second/time.Duration(rate)
+}
+
 // Start the RTSP client, and start reading packets.
 func (g *Golibrtsp) Start(ctx context.Context, streamType string, queue *packets.Queue, configuration *models.Configuration, communication *models.Communication) (err error) {
 	log.Log.Debug("capture.golibrtsp.Start(): started")
@@ -693,13 +748,13 @@ func (g *Golibrtsp) Start(ctx context.Context, streamType string, queue *packets
 	// called when a MULAW audio RTP packet arrives
 	if g.AudioG711Media != nil && g.AudioG711Forma != nil {
 		g.Client.OnPacketRTP(g.AudioG711Media, g.AudioG711Forma, func(rtppkt *rtp.Packet) {
-			pts, ok := g.Client.PacketPTS(g.AudioG711Media, rtppkt)
 			// decode timestamp
-			pts2, ok := g.Client.PacketPTS2(g.AudioG711Media, rtppkt)
+			pts2, ok := g.Client.PacketPTS(g.AudioG711Media, rtppkt)
 			if !ok {
 				log.Log.Debug("capture.golibrtsp.Start(): " + "unable to get PTS")
 				return
 			}
+			pts := ptsToDuration(pts2, g.AudioG711Forma.ClockRate())
 
 			// extract LPCM samples from RTP packets
 			op, err := g.AudioG711Decoder.Decode(rtppkt)
@@ -729,12 +784,12 @@ func (g *Golibrtsp) Start(ctx context.Context, streamType string, queue *packets
 	if g.AudioMPEG4Media != nil && g.AudioMPEG4Forma != nil {
 		g.Client.OnPacketRTP(g.AudioMPEG4Media, g.AudioMPEG4Forma, func(rtppkt *rtp.Packet) {
 			// decode timestamp
-			pts, ok := g.Client.PacketPTS(g.AudioMPEG4Media, rtppkt)
-			pts2, ok := g.Client.PacketPTS2(g.AudioMPEG4Media, rtppkt)
+			pts2, ok := g.Client.PacketPTS(g.AudioMPEG4Media, rtppkt)
 			if !ok {
 				log.Log.Error("capture.golibrtsp.Start(): " + "unable to get PTS")
 				return
 			}
+			pts := ptsToDuration(pts2, g.AudioMPEG4Forma.ClockRate())
 
 			// Encode the AAC samples from RTP packets
 			// extract access units from RTP packets
@@ -788,12 +843,12 @@ func (g *Golibrtsp) Start(ctx context.Context, streamType string, queue *packets
 			if len(rtppkt.Payload) > 0 {
 
 				// decode timestamps — validate each call separately
-				pts, okPTS := g.Client.PacketPTS(g.VideoH264Media, rtppkt)
-				pts2, okPTS2 := g.Client.PacketPTS2(g.VideoH264Media, rtppkt)
+				pts2, okPTS2 := g.Client.PacketPTS(g.VideoH264Media, rtppkt)
 				if !okPTS2 {
-					log.Log.Debug("capture.golibrtsp.Start(): unable to get PTS2 from PacketPTS2")
+					log.Log.Debug("capture.golibrtsp.Start(): unable to get PTS")
 					return
 				}
+				pts := ptsToDuration(pts2, g.VideoH264Forma.ClockRate())
 
 				// Extract access units from RTP packets.
 				// We need a complete access unit to determine whether
@@ -807,15 +862,13 @@ func (g *Golibrtsp) Start(ctx context.Context, streamType string, queue *packets
 				}
 
 				// Frame is complete — update per-stream FPS from PTS.
-				if okPTS {
-					ft := g.fpsTrackers[g.VideoH264Index]
-					if ft == nil {
-						ft = newFPSTracker(30)
-						g.fpsTrackers[g.VideoH264Index] = ft
-					}
-					if ptsFPS := ft.update(pts); ptsFPS > 0 && ptsFPS <= 120 {
-						g.Streams[g.VideoH264Index].FPS = ptsFPS
-					}
+				ft := g.fpsTrackers[g.VideoH264Index]
+				if ft == nil {
+					ft = newFPSTracker(30)
+					g.fpsTrackers[g.VideoH264Index] = ft
+				}
+				if ptsFPS := ft.update(pts); ptsFPS > 0 && ptsFPS <= 120 {
+					g.Streams[g.VideoH264Index].FPS = ptsFPS
 				}
 
 				// We'll need to read out a few things.
@@ -1035,12 +1088,12 @@ func (g *Golibrtsp) Start(ctx context.Context, streamType string, queue *packets
 			if len(rtppkt.Payload) > 0 {
 
 				// decode timestamps — validate each call separately
-				pts, okPTS := g.Client.PacketPTS(g.VideoH265Media, rtppkt)
-				pts2, okPTS2 := g.Client.PacketPTS2(g.VideoH265Media, rtppkt)
+				pts2, okPTS2 := g.Client.PacketPTS(g.VideoH265Media, rtppkt)
 				if !okPTS2 {
 					log.Log.Debug("capture.golibrtsp.Start(): unable to get PTS")
 					return
 				}
+				pts := ptsToDuration(pts2, g.VideoH265Forma.ClockRate())
 
 				// Extract access units from RTP packets.
 				// We need a complete access unit to determine whether
@@ -1054,16 +1107,14 @@ func (g *Golibrtsp) Start(ctx context.Context, streamType string, queue *packets
 				}
 
 				// Frame is complete — update per-stream FPS from PTS.
-				if okPTS {
-					ft := g.fpsTrackers[g.VideoH265Index]
-					if ft == nil {
-						ft = newFPSTracker(30)
-						g.fpsTrackers[g.VideoH265Index] = ft
-					}
-					if ptsFPS := ft.update(pts); ptsFPS > 0 && ptsFPS <= 120 {
-						g.Streams[g.VideoH265Index].FPS = ptsFPS
-						g.persistStreamFPS(configuration, streamType, ptsFPS)
-					}
+				ft := g.fpsTrackers[g.VideoH265Index]
+				if ft == nil {
+					ft = newFPSTracker(30)
+					g.fpsTrackers[g.VideoH265Index] = ft
+				}
+				if ptsFPS := ft.update(pts); ptsFPS > 0 && ptsFPS <= 120 {
+					g.Streams[g.VideoH265Index].FPS = ptsFPS
+					g.persistStreamFPS(configuration, streamType, ptsFPS)
 				}
 
 				// Preserve the decoded access unit (in decode order) for DTS

machinery/src/capture/gortsplib_test.go

@@ -0,0 +1,79 @@
+package capture
+
+import (
+	"bytes"
+	"encoding/pem"
+	"net/http/httptest"
+	"os"
+	"path/filepath"
+	"testing"
+)
+
+func TestRTSPSTLSConfig(t *testing.T) {
+	t.Run("verifies certificates by default", func(t *testing.T) {
+		t.Setenv(rtspsCAFileEnv, "")
+		t.Setenv(rtspsInsecureEnv, "")
+
+		got, err := rtspsTLSConfig()
+		if err != nil {
+			t.Fatalf("rtspsTLSConfig() error = %v", err)
+		}
+		if got != nil {
+			t.Fatalf("rtspsTLSConfig() = %#v, want nil", got)
+		}
+	})
+
+	t.Run("allows explicit insecure mode", func(t *testing.T) {
+		t.Setenv(rtspsCAFileEnv, "/missing/ignored-in-insecure-mode.pem")
+		t.Setenv(rtspsInsecureEnv, "true")
+
+		got, err := rtspsTLSConfig()
+		if err != nil {
+			t.Fatalf("rtspsTLSConfig() error = %v", err)
+		}
+		if got == nil || !got.InsecureSkipVerify {
+			t.Fatalf("rtspsTLSConfig() = %#v, want InsecureSkipVerify enabled", got)
+		}
+	})
+
+	t.Run("adds a camera CA to system roots", func(t *testing.T) {
+		t.Setenv(rtspsInsecureEnv, "")
+		server := httptest.NewTLSServer(nil)
+		defer server.Close()
+
+		certificate := server.Certificate()
+		caFile := filepath.Join(t.TempDir(), "camera-ca.pem")
+		caPEM := pem.EncodeToMemory(&pem.Block{Type: "CERTIFICATE", Bytes: certificate.Raw})
+		if err := os.WriteFile(caFile, caPEM, 0o600); err != nil {
+			t.Fatal(err)
+		}
+		t.Setenv(rtspsCAFileEnv, caFile)
+
+		got, err := rtspsTLSConfig()
+		if err != nil {
+			t.Fatalf("rtspsTLSConfig() error = %v", err)
+		}
+		if got == nil || got.RootCAs == nil {
+			t.Fatalf("rtspsTLSConfig() = %#v, want custom RootCAs", got)
+		}
+		for _, subject := range got.RootCAs.Subjects() {
+			if bytes.Equal(subject, certificate.RawSubject) {
+				return
+			}
+		}
+		t.Fatal("camera CA was not added to RootCAs")
+	})
+
+	t.Run("rejects an invalid camera CA file", func(t *testing.T) {
+		t.Setenv(rtspsInsecureEnv, "")
+		caFile := filepath.Join(t.TempDir(), "camera-ca.pem")
+		if err := os.WriteFile(caFile, []byte("not a certificate"), 0o600); err != nil {
+			t.Fatal(err)
+		}
+		t.Setenv(rtspsCAFileEnv, caFile)
+
+		if _, err := rtspsTLSConfig(); err == nil {
+			t.Fatal("rtspsTLSConfig() error = nil, want invalid CA error")
+		}
+	})
+}

machinery/src/capture/main.go

@@ -4,8 +4,11 @@ package capture
 import (
 	"context"
 	"encoding/base64"
+	"encoding/json"
 	"image"
+	"math"
 	"os"
+	"path/filepath"
 	"strconv"
 	"time"
 
@@ -50,6 +53,52 @@ func publishRecordingState(mqttClient mqtt.Client, hubKey string, configuration
 	}
 }
 
+func recordingUploadMetadata(name, deviceKey string, timestamp int64, mp4Video *video.MP4) models.RecordingUploadMetadata {
+	metadata := models.RecordingUploadMetadata{
+		FileName:  filepath.Base(name),
+		DeviceKey: deviceKey,
+		Timestamp: timestamp,
+		Duration:  mp4Video.VideoTotalDuration,
+	}
+	value := mp4Video.AverageFPS()
+	if value > 0 && value <= 240 && !math.IsInf(value, 0) && !math.IsNaN(value) {
+		metadata.FPS = value
+	}
+	return metadata
+}
+
+// queueRecordingForUpload creates the marker consumed by the upload worker and
+// stores metadata captured from the finalized recording.
+func queueRecordingForUpload(configDirectory string, metadata models.RecordingUploadMetadata) {
+	payload, err := json.Marshal(metadata)
+	if err != nil {
+		log.Log.Error("capture.main.queueRecordingForUpload(): " + err.Error())
+		return
+	}
+
+	// Publish the marker with a same-filesystem rename. Writing directly to the
+	// watched directory would briefly expose an empty file to the upload poller.
+	marker, err := os.CreateTemp(filepath.Join(configDirectory, "data"), ".upload-marker-*")
+	if err == nil {
+		_, err = marker.Write(payload)
+	}
+	if err == nil {
+		err = marker.Chmod(0644)
+	}
+	if marker != nil {
+		if closeErr := marker.Close(); err == nil {
+			err = closeErr
+		}
+		defer os.Remove(marker.Name())
+	}
+	if err == nil {
+		err = os.Rename(marker.Name(), filepath.Join(configDirectory, "data", "cloud", models.RecordingUploadMetadataFileName(metadata.FileName)))
+	}
+	if err != nil {
+		log.Log.Error("capture.main.queueRecordingForUpload(): " + err.Error())
+	}
+}
+
 const (
 	// manualRecordingHeartbeatTimeout is how long the agent keeps a manual
 	// (live-view / remote) recording alive after the LAST viewer heartbeat. The
@@ -152,8 +201,10 @@ func CleanupRecordingDirectory(configDirectory string, configuration *models.Con
 		// now-dangling upload marker so the upload loop doesn't keep trying to
 		// upload a file that no longer exists.
 		log.Log.Warning("HandleRecordStream: removed oldest recording as part of cleanup, but it was STILL PENDING UPLOAD (disk full of un-uploaded recordings) - " + recordingsDirectory + "/" + name)
-		if err := os.Remove(cloudDirectory + "/" + name); err != nil && !os.IsNotExist(err) {
-			log.Log.Info("HandleRecordStream: could not remove dangling upload marker " + name + ", " + err.Error())
+		for _, markerName := range uploadMarkerNames(name) {
+			if err := os.Remove(filepath.Join(cloudDirectory, markerName)); err != nil && !os.IsNotExist(err) {
+				log.Log.Info("HandleRecordStream: could not remove dangling upload marker " + markerName + ", " + err.Error())
+			}
 		}
 	} else {
 		log.Log.Info("HandleRecordStream: removed oldest file as part of cleanup - " + recordingsDirectory + "/" + name)
@@ -246,9 +297,9 @@ func pickRecordingToCleanup(recordingsDirectory, cloudDirectory string) (string,
 			oldestAnyTime = modTime
 		}
 
-		// A recording is still pending upload if a marker with the same name
-		// exists in the cloud directory. Skip those when picking a safe candidate.
-		if _, statErr := os.Stat(cloudDirectory + "/" + entry.Name()); statErr == nil {
+		// A recording is still pending upload if either its current .metadata
+		// marker or a marker created by an older agent exists.
+		if recordingPendingUpload(cloudDirectory, entry.Name()) {
 			continue
 		}
 
@@ -267,6 +318,19 @@ func pickRecordingToCleanup(recordingsDirectory, cloudDirectory string) (string,
 	return "", false, os.ErrNotExist
 }
 
+func uploadMarkerNames(recordingName string) []string {
+	return []string{models.RecordingUploadMetadataFileName(recordingName), filepath.Base(recordingName)}
+}
+
+func recordingPendingUpload(cloudDirectory, recordingName string) bool {
+	for _, markerName := range uploadMarkerNames(recordingName) {
+		if _, err := os.Stat(filepath.Join(cloudDirectory, markerName)); err == nil {
+			return true
+		}
+	}
+	return false
+}
+
 func HandleRecordStream(queue *packets.Queue, configDirectory string, configuration *models.Configuration, communication *models.Communication, rtspClient RTSPClient, mqttClient mqtt.Client) {
 
 	config := configuration.Config
@@ -438,9 +502,7 @@ func HandleRecordStream(queue *packets.Queue, configDirectory string, configurat
 						}
 					}
 
-					// Create a symbol link.
-					fc, _ := os.Create(configDirectory + "/data/cloud/" + name)
-					fc.Close()
+					queueRecordingForUpload(configDirectory, recordingUploadMetadata(name, config.Key, startRecording, mp4Video))
 
 					recordingStatus = "idle"
 
@@ -597,9 +659,7 @@ func HandleRecordStream(queue *packets.Queue, configDirectory string, configurat
 						}
 					}
 
-					// Create a symbol link.
-					fc, _ := os.Create(configDirectory + "/data/cloud/" + name)
-					fc.Close()
+					queueRecordingForUpload(configDirectory, recordingUploadMetadata(name, config.Key, startRecording, mp4Video))
 
 					recordingStatus = "idle"
 
@@ -869,9 +929,7 @@ func HandleRecordStream(queue *packets.Queue, configDirectory string, configurat
 					}
 				}
 
-				// Create a symbol linc.
-				fc, _ := os.Create(configDirectory + "/data/cloud/" + name)
-				fc.Close()
+				queueRecordingForUpload(configDirectory, recordingUploadMetadata(name, config.Key, displayTime, mp4Video))
 
 				// Clean up the recording directory if necessary.
 				CleanupRecordingDirectory(configDirectory, configuration)

machinery/src/capture/main_test.go

@@ -0,0 +1,88 @@
+package capture
+
+import (
+	"encoding/json"
+	"math"
+	"os"
+	"path/filepath"
+	"testing"
+	"time"
+
+	"github.com/kerberos-io/agent/machinery/src/models"
+	"github.com/kerberos-io/agent/machinery/src/video"
+)
+
+func TestPTSToDuration(t *testing.T) {
+	tests := []struct {
+		name      string
+		pts       int64
+		clockRate int
+		want      time.Duration
+	}{
+		{name: "one video second", pts: 90_000, clockRate: 90_000, want: time.Second},
+		{name: "one audio frame", pts: 1_024, clockRate: 8_000, want: 128 * time.Millisecond},
+		{name: "fractional millisecond", pts: 45_045, clockRate: 90_000, want: 500*time.Millisecond + 500*time.Microsecond},
+		{name: "negative timestamp", pts: -45_045, clockRate: 90_000, want: -500*time.Millisecond - 500*time.Microsecond},
+		{name: "large timestamp", pts: 90_000 * 60 * 60 * 24, clockRate: 90_000, want: 24 * time.Hour},
+	}
+
+	for _, test := range tests {
+		t.Run(test.name, func(t *testing.T) {
+			if got := ptsToDuration(test.pts, test.clockRate); got != test.want {
+				t.Fatalf("ptsToDuration(%d, %d) = %s, want %s", test.pts, test.clockRate, got, test.want)
+			}
+		})
+	}
+}
+
+func TestQueueRecordingForUploadStoresFinalizedMetadata(t *testing.T) {
+	configDirectory := t.TempDir()
+	if err := os.MkdirAll(filepath.Join(configDirectory, "data", "cloud"), 0o755); err != nil {
+		t.Fatalf("mkdir cloud queue: %v", err)
+	}
+
+	mp4Video := &video.MP4{VideoTotalDuration: 20452, SampleCount: 613}
+	metadata := recordingUploadMetadata("recording.mp4", "device-key", 1785934709414, mp4Video)
+	queueRecordingForUpload(configDirectory, metadata)
+
+	got, err := os.ReadFile(filepath.Join(configDirectory, "data", "cloud", "recording.metadata"))
+	if err != nil {
+		t.Fatalf("read upload marker: %v", err)
+	}
+	var stored models.RecordingUploadMetadata
+	if err := json.Unmarshal(got, &stored); err != nil {
+		t.Fatalf("decode upload marker: %v", err)
+	}
+	expectedFPS := mp4Video.AverageFPS()
+	if stored.FileName != "recording.mp4" || stored.DeviceKey != "device-key" || stored.Timestamp != 1785934709414 || stored.Duration != 20452 || math.Abs(stored.FPS-expectedFPS) > 1e-9 {
+		t.Fatalf("upload marker = %+v", stored)
+	}
+	if stored.FPS == math.Floor(stored.FPS) {
+		t.Fatalf("upload marker FPS = %v, want fractional precision", stored.FPS)
+	}
+}
+
+func TestQueueRecordingForUploadKeepsUnknownFPSCompatible(t *testing.T) {
+	for _, fps := range []float64{0, 0.99, -1, math.NaN(), math.Inf(1), 241} {
+		t.Run("invalid FPS", func(t *testing.T) {
+			configDirectory := t.TempDir()
+			if err := os.MkdirAll(filepath.Join(configDirectory, "data", "cloud"), 0o755); err != nil {
+				t.Fatalf("mkdir cloud queue: %v", err)
+			}
+
+			metadata := models.RecordingUploadMetadata{FileName: "recording.mp4"}
+			if fps >= 1 && fps <= 240 && !math.IsNaN(fps) && !math.IsInf(fps, 0) {
+				metadata.FPS = fps
+			}
+			queueRecordingForUpload(configDirectory, metadata)
+
+			got, err := os.ReadFile(filepath.Join(configDirectory, "data", "cloud", "recording.metadata"))
+			if err != nil {
+				t.Fatalf("read upload marker: %v", err)
+			}
+			if string(got) != `{"filename":"recording.mp4","device_key":"","timestamp":0,"duration":0}` {
+				t.Fatalf("upload marker = %q, want metadata without FPS", got)
+			}
+		})
+	}
+}

machinery/src/cloud/cloud.go

@@ -78,7 +78,8 @@ func HandleUpload(configDirectory string, configuration *models.Configuration, c
 					default:
 					}
 
-					fileName := f.Name()
+					markerFileName := f.Name()
+					fileName := models.RecordingFileNameFromUploadMarker(markerFileName)
 					uploaded := false
 					configured := false
 					err = nil
@@ -113,7 +114,7 @@ func HandleUpload(configDirectory string, configuration *models.Configuration, c
 					// Check if the file is uploaded, if so, remove it.
 					if uploaded {
 						delay = 500 * time.Millisecond // reset
-						err := os.Remove(watchDirectory + fileName)
+						err := os.Remove(watchDirectory + markerFileName)
 						if err != nil {
 							log.Log.Error("HandleUpload: " + err.Error())
 						}
@@ -127,7 +128,7 @@ func HandleUpload(configDirectory string, configuration *models.Configuration, c
 							}
 						}
 					} else if !configured {
-						err := os.Remove(watchDirectory + fileName)
+						err := os.Remove(watchDirectory + markerFileName)
 						if err != nil {
 							log.Log.Error("HandleUpload: " + err.Error())
 						}
@@ -228,17 +229,71 @@ func rawJSONOrEmptyArray(b []byte) json.RawMessage {
 	return json.RawMessage(b)
 }
 
+const heartbeatResponseBodyLogLimit = 4 * 1024
+
+func readHeartbeatResponseBody(response *http.Response) (string, bool, error) {
+	if response == nil || response.Body == nil {
+		return "", false, nil
+	}
+	defer response.Body.Close()
+
+	body, err := io.ReadAll(io.LimitReader(response.Body, heartbeatResponseBodyLogLimit+1))
+	if err != nil {
+		return "", false, err
+	}
+
+	truncated := len(body) > heartbeatResponseBodyLogLimit
+	if truncated {
+		body = body[:heartbeatResponseBodyLogLimit]
+	}
+
+	return strings.TrimSpace(string(body)), truncated, nil
+}
+
+func formatHeartbeatFailureLog(response *http.Response, requestErr error, responseBody string, responseBodyTruncated bool, responseBodyErr error, elapsed time.Duration) string {
+	details := make([]string, 0, 6)
+	if response != nil {
+		details = append(details, "status_code="+strconv.Itoa(response.StatusCode))
+		if response.Status != "" {
+			details = append(details, "status="+strconv.Quote(response.Status))
+		}
+	} else {
+		details = append(details, "status_code=none")
+	}
+	details = append(details, "duration="+elapsed.Round(time.Millisecond).String())
+	if requestErr != nil {
+		details = append(details, "request_error="+strconv.Quote(requestErr.Error()))
+	}
+	if responseBody != "" {
+		details = append(details, "response_body="+strconv.Quote(responseBody))
+	}
+	if responseBodyTruncated {
+		details = append(details, "response_body_truncated=true")
+	}
+	if responseBodyErr != nil {
+		details = append(details, "response_body_error="+strconv.Quote(responseBodyErr.Error()))
+	}
+
+	return "cloud.HandleHeartBeat(): heartbeat request to Kerberos Hub failed: " + strings.Join(details, ", ")
+}
+
 func HandleHeartBeat(configuration *models.Configuration, communication *models.Communication, uptimeStart time.Time) {
 	log.Log.Debug("cloud.HandleHeartBeat(): started")
 
+	// Bound every heartbeat POST so a stalled connection (e.g. a saturated uplink
+	// or an unresponsive Hub/Vault) fails fast on this cycle instead of blocking
+	// the whole heartbeat loop indefinitely. A hung POST would otherwise stop all
+	// further heartbeats, and Hub marks a camera offline once its last heartbeat is
+	// older than 180s even while capture is healthy.
+	const heartbeatHTTPTimeout = 30 * time.Second
 	var client *http.Client
 	if os.Getenv("AGENT_TLS_INSECURE") == "true" {
 		tr := &http.Transport{
 			TLSClientConfig: &tls.Config{InsecureSkipVerify: true},
 		}
-		client = &http.Client{Transport: tr}
+		client = &http.Client{Transport: tr, Timeout: heartbeatHTTPTimeout}
 	} else {
-		client = &http.Client{}
+		client = &http.Client{Timeout: heartbeatHTTPTimeout}
 	}
 
 	kerberosAgentVersion := utils.VERSION
@@ -637,20 +692,26 @@ loop:
 
 				var jsonStr = []byte(object)
 				buffy := bytes.NewBuffer(jsonStr)
-				req, _ := http.NewRequest("POST", hubURI, buffy)
-				req.Header.Set("Content-Type", "application/json")
-				resp, err := client.Do(req)
-				if resp != nil {
-					resp.Body.Close()
+				requestStarted := time.Now()
+				req, requestErr := http.NewRequest("POST", hubURI, buffy)
+				var resp *http.Response
+				if requestErr == nil {
+					req.Header.Set("Content-Type", "application/json")
+					resp, requestErr = client.Do(req)
 				}
-				if err == nil && resp.StatusCode == 200 {
+				if requestErr == nil && resp != nil && resp.StatusCode == http.StatusOK {
+					if resp.Body != nil {
+						_, _ = io.Copy(io.Discard, resp.Body)
+						resp.Body.Close()
+					}
 					communication.CloudTimestamp.Store(time.Now().Unix())
 					log.Log.Info("cloud.HandleHeartBeat(): (200) Heartbeat received by Kerberos Hub.")
 				} else {
+					responseBody, responseBodyTruncated, responseBodyErr := readHeartbeatResponseBody(resp)
 					if communication.CloudTimestamp != nil && communication.CloudTimestamp.Load() != nil {
 						communication.CloudTimestamp.Store(int64(0))
 					}
-					log.Log.Error("cloud.HandleHeartBeat(): (400) Something went wrong while sending to Kerberos Hub.")
+					log.Log.Error(formatHeartbeatFailureLog(resp, requestErr, responseBody, responseBodyTruncated, responseBodyErr, time.Since(requestStarted)))
 				}
 			} else {
 				log.Log.Error("cloud.HandleHeartBeat(): Disabled as we do not have a public key defined.")

machinery/src/cloud/cloud_test.go

@@ -0,0 +1,51 @@
+package cloud
+
+import (
+	"io"
+	"net/http"
+	"strings"
+	"testing"
+	"time"
+)
+
+func TestHeartbeatFailureLogIncludesHubResponse(t *testing.T) {
+	response := &http.Response{
+		StatusCode: http.StatusBadRequest,
+		Status:     "400 Bad Request",
+		Body:       io.NopCloser(strings.NewReader(`{"error":"invalid heartbeat"}`)),
+	}
+
+	responseBody, truncated, err := readHeartbeatResponseBody(response)
+	if err != nil {
+		t.Fatalf("readHeartbeatResponseBody() error = %v", err)
+	}
+	message := formatHeartbeatFailureLog(response, nil, responseBody, truncated, nil, 125*time.Millisecond)
+
+	for _, expected := range []string{
+		"status_code=400",
+		`status="400 Bad Request"`,
+		"duration=125ms",
+		`response_body="{\"error\":\"invalid heartbeat\"}"`,
+	} {
+		if !strings.Contains(message, expected) {
+			t.Errorf("formatHeartbeatFailureLog() = %q, want it to contain %q", message, expected)
+		}
+	}
+}
+
+func TestReadHeartbeatResponseBodyTruncatesLargeBody(t *testing.T) {
+	response := &http.Response{
+		Body: io.NopCloser(strings.NewReader(strings.Repeat("x", heartbeatResponseBodyLogLimit+1))),
+	}
+
+	body, truncated, err := readHeartbeatResponseBody(response)
+	if err != nil {
+		t.Fatalf("readHeartbeatResponseBody() error = %v", err)
+	}
+	if !truncated {
+		t.Fatal("readHeartbeatResponseBody() truncated = false, want true")
+	}
+	if len(body) != heartbeatResponseBodyLogLimit {
+		t.Fatalf("len(body) = %d, want %d", len(body), heartbeatResponseBodyLogLimit)
+	}
+}

machinery/src/cloud/kerberos_hub.go

@@ -85,6 +85,7 @@ func UploadKerberosHub(configuration *models.Configuration, fileName string) (bo
 	req.Header.Set("X-Kerberos-Hub-PublicKey", config.HubKey)
 	req.Header.Set("X-Kerberos-Hub-PrivateKey", config.HubPrivateKey)
 	req.Header.Set("X-Kerberos-Hub-Region", config.S3.Region)
+	setQueuedRecordingMetadataHeaders(req.Header, fileName)
 
 	var client *http.Client
 	if os.Getenv("AGENT_TLS_INSECURE") == "true" {
@@ -128,6 +129,7 @@ func UploadKerberosHub(configuration *models.Configuration, fileName string) (bo
 	req.Header.Set("X-Kerberos-Hub-PublicKey", config.HubKey)
 	req.Header.Set("X-Kerberos-Hub-PrivateKey", config.HubPrivateKey)
 	req.Header.Set("X-Kerberos-Hub-Region", config.S3.Region)
+	setQueuedRecordingMetadataHeaders(req.Header, fileName)
 	resp, err = client.Do(req)
 	if resp != nil {
 		defer resp.Body.Close()

machinery/src/cloud/kerberos_vault.go

@@ -35,10 +35,15 @@ func UploadKerberosVault(configuration *models.Configuration, fileName string) (
 	// This can happen when the file was already removed (e.g. cleanup, or an
 	// earlier successful upload). Skip it so the watcher drops the marker
 	// instead of retrying indefinitely.
-	if _, err := os.Stat("data/recordings/" + fileName); err != nil {
+	info, err := os.Stat("data/recordings/" + fileName)
+	if err != nil {
 		log.Log.Info("UploadKerberosVault: skipping " + fileName + ", file doesn't exist anymore")
 		return false, false, nil
 	}
+	if info.Size() == 0 {
+		log.Log.Warning("UploadKerberosVault: skipping " + fileName + ", recording is empty")
+		return false, false, nil
+	}
 
 	// timestamp_microseconds_instanceName_regionCoordinates_numberOfChanges_token
 	// 1564859471_6-474162_oprit_577-283-727-375_1153_27.mp4
@@ -165,6 +170,7 @@ func uploadVaultLegacy(vault models.KStorage, publicKey, deviceKey, fileName, la
 	}
 	req.Header.Set("Content-Type", "video/mp4")
 	setVaultHeaders(req.Header, vault, publicKey, deviceKey, fileName)
+	setQueuedRecordingMetadataHeaders(req.Header, fileName)
 
 	client := newVaultHTTPClient(0)
 	resp, err := client.Do(req)

machinery/src/cloud/livemoq/annexb.go

@@ -0,0 +1,103 @@
+package livemoq
+
+import (
+	"bytes"
+	"strings"
+
+	"github.com/bluenviron/mediacommon/pkg/codecs/h264"
+	"github.com/kerberos-io/agent/machinery/src/models"
+)
+
+var annexBStartCode = []byte{0x00, 0x00, 0x00, 0x01}
+
+// H264NormalizationStats reports malformed duplication removed from an access unit.
+type H264NormalizationStats struct {
+	DuplicateIDRNALUs int
+}
+
+// EnsureAnnexB restores the start code stripped by the Agent capture queue.
+func EnsureAnnexB(payload []byte) []byte {
+	if hasAnnexBStartCode(payload) {
+		return payload
+	}
+
+	framed := make([]byte, 0, len(annexBStartCode)+len(payload))
+	framed = append(framed, annexBStartCode...)
+	return append(framed, payload...)
+}
+
+// NormalizeH264AccessUnit removes delimiters and exact duplicate parameter-set
+// or IDR NALUs that can confuse older MoQ splitters and decoders.
+func NormalizeH264AccessUnit(payload []byte) ([]byte, error) {
+	normalized, _, err := NormalizeH264AccessUnitWithStats(payload)
+	return normalized, err
+}
+
+// NormalizeH264AccessUnitWithStats also reports exact duplicate IDR NALUs.
+func NormalizeH264AccessUnitWithStats(payload []byte) ([]byte, H264NormalizationStats, error) {
+	nalus, err := h264.AnnexBUnmarshal(EnsureAnnexB(payload))
+	if err != nil {
+		return nil, H264NormalizationStats{}, err
+	}
+
+	stats := H264NormalizationStats{}
+	normalized := make([][]byte, 0, len(nalus))
+	for _, nalu := range nalus {
+		if len(nalu) == 0 || nalu[0]&0x1f == 9 {
+			continue
+		}
+		naluType := nalu[0] & 0x1f
+		if naluType == 7 || naluType == 8 || naluType == 5 {
+			duplicate := false
+			for _, existing := range normalized {
+				if bytes.Equal(existing, nalu) {
+					duplicate = true
+					break
+				}
+			}
+			if duplicate {
+				if naluType == 5 {
+					stats.DuplicateIDRNALUs++
+				}
+				continue
+			}
+		}
+		normalized = append(normalized, nalu)
+	}
+
+	result, err := h264.AnnexBMarshal(normalized)
+	return result, stats, err
+}
+
+// BroadcastPath returns the relay path a quality tier is published on. Every
+// tier gets its own broadcast so a viewer switches between the camera's main and
+// sub stream by resubscribing to another path, without any control channel back
+// to the Agent. The high tier keeps the historical ".../live.hang" path so
+// existing viewers keep working; the low tier lives next to it on
+// ".../live-low.hang".
+func BroadcastPath(prefix string, deviceKey string, quality string) string {
+	prefix = strings.Trim(prefix, "/")
+	if prefix == "" {
+		prefix = "devices"
+	}
+	name := "live.hang"
+	if quality == models.StreamQualityLow {
+		name = "live-low.hang"
+	}
+	return prefix + "/" + strings.Trim(deviceKey, "/") + "/" + name
+}
+
+// TimestampUs converts the capture presentation timestamp from milliseconds.
+// CompositionTime must not be added: it is already represented in the PTS and
+// is only used by muxers to derive DTS for streams containing B-frames.
+func TimestampUs(presentationTimeMs int64) uint64 {
+	if presentationTimeMs < 0 {
+		return 0
+	}
+	return uint64(presentationTimeMs) * 1000
+}
+
+func hasAnnexBStartCode(payload []byte) bool {
+	return len(payload) >= 4 && payload[0] == 0 && payload[1] == 0 &&
+		((payload[2] == 0 && payload[3] == 1) || payload[2] == 1)
+}

machinery/src/cloud/livemoq/annexb_test.go

@@ -0,0 +1,116 @@
+package livemoq
+
+import (
+	"bytes"
+	"testing"
+
+	"github.com/kerberos-io/agent/machinery/src/models"
+)
+
+func TestEnsureAnnexB(t *testing.T) {
+	tests := []struct {
+		name    string
+		payload []byte
+		want    []byte
+	}{
+		{
+			name:    "missing start code",
+			payload: []byte{0x41, 0x01},
+			want:    []byte{0x00, 0x00, 0x00, 0x01, 0x41, 0x01},
+		},
+		{
+			name:    "four byte start code",
+			payload: []byte{0x00, 0x00, 0x00, 0x01, 0x65},
+			want:    []byte{0x00, 0x00, 0x00, 0x01, 0x65},
+		},
+		{
+			name:    "three byte start code",
+			payload: []byte{0x00, 0x00, 0x01, 0x41},
+			want:    []byte{0x00, 0x00, 0x01, 0x41},
+		},
+	}
+
+	for _, test := range tests {
+		t.Run(test.name, func(t *testing.T) {
+			if got := EnsureAnnexB(test.payload); !bytes.Equal(got, test.want) {
+				t.Fatalf("EnsureAnnexB() = %x, want %x", got, test.want)
+			}
+		})
+	}
+}
+
+func TestNormalizeH264AccessUnit(t *testing.T) {
+	startCode := []byte{0x00, 0x00, 0x00, 0x01}
+	sps := []byte{0x67, 0x42, 0x00, 0x1f}
+	pps := []byte{0x68, 0xce, 0x06, 0xe2}
+	aud := []byte{0x09, 0xf0}
+	idr := []byte{0x65, 0x88, 0x84}
+
+	payload := make([]byte, 0)
+	for _, nalu := range [][]byte{sps, pps, aud, sps, pps, idr} {
+		payload = append(payload, startCode...)
+		payload = append(payload, nalu...)
+	}
+
+	got, err := NormalizeH264AccessUnit(payload)
+	if err != nil {
+		t.Fatal(err)
+	}
+	want := make([]byte, 0)
+	for _, nalu := range [][]byte{sps, pps, idr} {
+		want = append(want, startCode...)
+		want = append(want, nalu...)
+	}
+	if !bytes.Equal(got, want) {
+		t.Fatalf("NormalizeH264AccessUnit() = %x, want %x", got, want)
+	}
+}
+
+func TestNormalizeH264AccessUnitRemovesOnlyExactDuplicateIDRSlices(t *testing.T) {
+	startCode := []byte{0x00, 0x00, 0x00, 0x01}
+	idrSlice1 := []byte{0x65, 0x88, 0x84}
+	idrSlice2 := []byte{0x65, 0x44, 0x22}
+
+	payload := make([]byte, 0)
+	for _, nalu := range [][]byte{idrSlice1, idrSlice1, idrSlice2} {
+		payload = append(payload, startCode...)
+		payload = append(payload, nalu...)
+	}
+
+	got, stats, err := NormalizeH264AccessUnitWithStats(payload)
+	if err != nil {
+		t.Fatal(err)
+	}
+	want := make([]byte, 0)
+	for _, nalu := range [][]byte{idrSlice1, idrSlice2} {
+		want = append(want, startCode...)
+		want = append(want, nalu...)
+	}
+	if !bytes.Equal(got, want) {
+		t.Fatalf("NormalizeH264AccessUnitWithStats() = %x, want %x", got, want)
+	}
+	if stats.DuplicateIDRNALUs != 1 {
+		t.Fatalf("DuplicateIDRNALUs = %d, want 1", stats.DuplicateIDRNALUs)
+	}
+}
+
+func TestBroadcastPath(t *testing.T) {
+	if got := BroadcastPath("/devices/", "/camera-1/", models.StreamQualityHigh); got != "devices/camera-1/live.hang" {
+		t.Fatalf("BroadcastPath() high = %q", got)
+	}
+	if got := BroadcastPath("", "camera-1", models.StreamQualityHigh); got != "devices/camera-1/live.hang" {
+		t.Fatalf("BroadcastPath() default = %q", got)
+	}
+	if got := BroadcastPath("", "camera-1", models.StreamQualityLow); got != "devices/camera-1/live-low.hang" {
+		t.Fatalf("BroadcastPath() low = %q", got)
+	}
+}
+
+func TestTimestampUs(t *testing.T) {
+	if got := TimestampUs(1234); got != 1_234_000 {
+		t.Fatalf("TimestampUs() = %d, want 1234000", got)
+	}
+	if got := TimestampUs(-1); got != 0 {
+		t.Fatalf("TimestampUs() negative = %d, want 0", got)
+	}
+}

machinery/src/cloud/livemoq/dedup.go

@@ -0,0 +1,41 @@
+package livemoq
+
+import (
+	"crypto/sha256"
+	"time"
+)
+
+type KeyframeDeduplicator struct {
+	hasPrevious  bool
+	timestampMs  int64
+	capturedAtMs int64
+	observedAt   time.Time
+	digest       [sha256.Size]byte
+}
+
+func (d *KeyframeDeduplicator) Reset() {
+	*d = KeyframeDeduplicator{}
+}
+
+// IsDuplicate reports exact repeated keyframe access units observed close
+// together. Distinct IDR slices within one access unit remain untouched.
+func (d *KeyframeDeduplicator) IsDuplicate(timestampMs int64, capturedAtMs int64, payload []byte, observedAt time.Time, window time.Duration) bool {
+	digest := sha256.Sum256(payload)
+	duplicate := d.hasPrevious && d.timestampMs == timestampMs && d.digest == digest
+	if duplicate {
+		if capturedAtMs > 0 && d.capturedAtMs > 0 {
+			gap := time.Duration(capturedAtMs-d.capturedAtMs) * time.Millisecond
+			duplicate = gap >= 0 && gap <= window
+		} else {
+			gap := observedAt.Sub(d.observedAt)
+			duplicate = gap >= 0 && gap <= window
+		}
+	}
+
+	d.hasPrevious = true
+	d.timestampMs = timestampMs
+	d.capturedAtMs = capturedAtMs
+	d.observedAt = observedAt
+	d.digest = digest
+	return duplicate
+}

machinery/src/cloud/livemoq/dedup_test.go

@@ -0,0 +1,60 @@
+package livemoq
+
+import (
+	"testing"
+	"time"
+)
+
+func TestKeyframeDeduplicator(t *testing.T) {
+	now := time.UnixMilli(10_000)
+	window := 500 * time.Millisecond
+	payload := []byte{0x00, 0x00, 0x00, 0x01, 0x65, 0x88}
+	deduplicator := KeyframeDeduplicator{}
+
+	if deduplicator.IsDuplicate(1_000, 10_000, payload, now, window) {
+		t.Fatal("first keyframe reported as duplicate")
+	}
+	if !deduplicator.IsDuplicate(1_000, 10_020, payload, now.Add(20*time.Millisecond), window) {
+		t.Fatal("exact repeated keyframe was not reported as duplicate")
+	}
+	if deduplicator.IsDuplicate(2_000, 11_000, payload, now.Add(time.Second), window) {
+		t.Fatal("same payload with a new timestamp reported as duplicate")
+	}
+	if deduplicator.IsDuplicate(2_000, 11_020, append(payload, 0x01), now.Add(1020*time.Millisecond), window) {
+		t.Fatal("different payload with the same timestamp reported as duplicate")
+	}
+}
+
+func TestKeyframeDeduplicatorAllowsTimestampReuseOutsideWindow(t *testing.T) {
+	now := time.UnixMilli(10_000)
+	payload := []byte{0x00, 0x00, 0x00, 0x01, 0x65, 0x88}
+	deduplicator := KeyframeDeduplicator{}
+
+	deduplicator.IsDuplicate(1_000, 10_000, payload, now, 500*time.Millisecond)
+	if deduplicator.IsDuplicate(1_000, 20_000, payload, now.Add(10*time.Second), 500*time.Millisecond) {
+		t.Fatal("later keyframe after timestamp reset reported as duplicate")
+	}
+}
+
+func TestKeyframeDeduplicatorFallsBackToObservationTime(t *testing.T) {
+	now := time.UnixMilli(10_000)
+	payload := []byte{0x65, 0x88}
+	deduplicator := KeyframeDeduplicator{}
+
+	deduplicator.IsDuplicate(1_000, 0, payload, now, 500*time.Millisecond)
+	if !deduplicator.IsDuplicate(1_000, 0, payload, now.Add(20*time.Millisecond), 500*time.Millisecond) {
+		t.Fatal("duplicate without capture time was not reported")
+	}
+}
+
+func TestKeyframeDeduplicatorReset(t *testing.T) {
+	now := time.UnixMilli(10_000)
+	payload := []byte{0x65, 0x88}
+	deduplicator := KeyframeDeduplicator{}
+
+	deduplicator.IsDuplicate(1_000, 10_000, payload, now, 500*time.Millisecond)
+	deduplicator.Reset()
+	if deduplicator.IsDuplicate(1_000, 10_020, payload, now.Add(20*time.Millisecond), 500*time.Millisecond) {
+		t.Fatal("first keyframe after reset reported as duplicate")
+	}
+}

machinery/src/cloud/livemoq/recovery.go

@@ -0,0 +1,55 @@
+package livemoq
+
+import "time"
+
+type FrameGateEvent uint8
+
+const (
+	FrameGateEventNone FrameGateEvent = iota
+	FrameGateEventStarted
+	FrameGateEventLagging
+	FrameGateEventRecovered
+)
+
+// FrameGate keeps publication on a decodable, recent GOP.
+type FrameGate struct {
+	started    bool
+	recovering bool
+}
+
+// Reset closes the gate so publication resumes on the next keyframe. It is used
+// when the publisher stopped writing for a reason unrelated to the stream health
+// (no subscribers), so the next viewer never receives a partial GOP.
+func (g *FrameGate) Reset() {
+	g.started = false
+	g.recovering = false
+}
+
+// Allow rejects stale frames and waits for a fresh keyframe before reopening.
+func (g *FrameGate) Allow(isKeyFrame bool, capturedAtMs int64, now time.Time, maxAge time.Duration) (bool, FrameGateEvent) {
+	if capturedAtMs > 0 && now.Sub(time.UnixMilli(capturedAtMs)) > maxAge {
+		event := FrameGateEventNone
+		if g.started {
+			if !g.recovering {
+				event = FrameGateEventLagging
+			}
+			g.started = false
+			g.recovering = true
+		}
+		return false, event
+	}
+
+	if !g.started {
+		if !isKeyFrame {
+			return false, FrameGateEventNone
+		}
+		g.started = true
+		if g.recovering {
+			g.recovering = false
+			return true, FrameGateEventRecovered
+		}
+		return true, FrameGateEventStarted
+	}
+
+	return true, FrameGateEventNone
+}

machinery/src/cloud/livemoq/recovery_test.go

@@ -0,0 +1,49 @@
+package livemoq
+
+import (
+	"testing"
+	"time"
+)
+
+func TestFrameGateRecoversAtFreshKeyframe(t *testing.T) {
+	now := time.UnixMilli(10_000)
+	maxAge := 1500 * time.Millisecond
+	gate := FrameGate{}
+
+	tests := []struct {
+		name         string
+		isKeyFrame   bool
+		capturedAtMs int64
+		wantAllowed  bool
+		wantEvent    FrameGateEvent
+	}{
+		{name: "waits for initial keyframe", capturedAtMs: 10_000},
+		{name: "starts at initial keyframe", isKeyFrame: true, capturedAtMs: 10_000, wantAllowed: true, wantEvent: FrameGateEventStarted},
+		{name: "publishes fresh delta", capturedAtMs: 10_020, wantAllowed: true},
+		{name: "detects stale packet", capturedAtMs: 8_000, wantEvent: FrameGateEventLagging},
+		{name: "rejects fresh delta while recovering", capturedAtMs: 10_040},
+		{name: "rejects stale keyframe without duplicate event", isKeyFrame: true, capturedAtMs: 8_000},
+		{name: "recovers at fresh keyframe", isKeyFrame: true, capturedAtMs: 10_060, wantAllowed: true, wantEvent: FrameGateEventRecovered},
+		{name: "publishes delta after recovery", capturedAtMs: 10_080, wantAllowed: true},
+	}
+
+	for _, test := range tests {
+		t.Run(test.name, func(t *testing.T) {
+			allowed, event := gate.Allow(test.isKeyFrame, test.capturedAtMs, now, maxAge)
+			if allowed != test.wantAllowed {
+				t.Fatalf("Allow() allowed = %t, want %t", allowed, test.wantAllowed)
+			}
+			if event != test.wantEvent {
+				t.Fatalf("Allow() event = %d, want %d", event, test.wantEvent)
+			}
+		})
+	}
+}
+
+func TestFrameGateAllowsMissingCaptureTime(t *testing.T) {
+	gate := FrameGate{}
+	allowed, event := gate.Allow(true, 0, time.Now(), time.Second)
+	if !allowed || event != FrameGateEventStarted {
+		t.Fatalf("Allow() = (%t, %d), want (true, %d)", allowed, event, FrameGateEventStarted)
+	}
+}

machinery/src/cloud/livemoq_disabled.go

@@ -0,0 +1,8 @@
+//go:build !moq
+
+package cloud
+
+import "github.com/kerberos-io/agent/machinery/src/models"
+
+// StartLiveStreamMoQ is disabled in the standard Agent build.
+func StartLiveStreamMoQ(_ *models.Configuration, _ *models.Communication, _ bool) {}

machinery/src/cloud/livemoq_enabled.go

@@ -0,0 +1,298 @@
+//go:build moq
+
+package cloud
+
+import (
+	"context"
+	"fmt"
+	"os"
+	"strings"
+	"sync/atomic"
+	"time"
+
+	"github.com/kerberos-io/agent/machinery/src/cloud/livemoq"
+	"github.com/kerberos-io/agent/machinery/src/log"
+	"github.com/kerberos-io/agent/machinery/src/models"
+	"github.com/kerberos-io/agent/machinery/src/packets"
+	"github.com/moq-dev/moq-go/moq"
+)
+
+const (
+	defaultMoQRelayURL      = "https://relay.uug.ai/anon"
+	minMoQRetryDelay        = time.Second
+	maxMoQRetryDelay        = 30 * time.Second
+	maxMoQLivePacketAge     = 1500 * time.Millisecond
+	slowMoQWriteThreshold   = 100 * time.Millisecond
+	moQWriteWarningInterval = 10 * time.Second
+	duplicateKeyframeWindow = 500 * time.Millisecond
+)
+
+type liveMoQConfig struct {
+	relayURL    string
+	broadcast   string
+	quality     string
+	sourceLabel string
+	queue       *packets.Queue
+}
+
+// label identifies the tier in log lines, since one Agent runs a publisher per
+// quality tier.
+func (c liveMoQConfig) label() string {
+	return c.quality + " (" + c.sourceLabel + " stream)"
+}
+
+// StartLiveStreamMoQ starts the publisher only in the dedicated MoQ build and
+// only when explicitly enabled by the deployment.
+//
+// Unlike WebRTC and HLS — where a viewer negotiates a session with the Agent and
+// can therefore ask for another quality on the fly — MoQ viewers subscribe to a
+// relay and never talk to the Agent. The quality selector is honoured by
+// publishing each tier as its OWN broadcast (see livemoq.BroadcastPath): the
+// high tier from the camera's highest-resolution stream and the low tier from
+// its sub stream, so switching quality in the frontend is a resubscribe to the
+// other path. Each tier only uploads while it actually has subscribers, so the
+// second broadcast is close to free when nobody watches it.
+func StartLiveStreamMoQ(configuration *models.Configuration, communication *models.Communication, subStreamEnabled bool) {
+	if os.Getenv("AGENT_LIVE_MOQ_ENABLED") != "true" {
+		return
+	}
+
+	config := configuration.Config
+	if config.Offline == "true" || config.Capture.Liveview == "false" {
+		log.Log.Info("cloud.StartLiveStreamMoQ(): disabled by Agent live-view configuration")
+		return
+	}
+	if config.Key == "" {
+		log.Log.Warning("cloud.StartLiveStreamMoQ(): AGENT_KEY is required")
+		return
+	}
+
+	// Both tiers are published by default. AGENT_LIVE_MOQ_QUALITY pins the Agent
+	// to a single tier for deployments that must never publish the other one
+	// (viewers asking for the pinned-away tier then find no broadcast).
+	qualities := []string{models.StreamQualityHigh, models.StreamQualityLow}
+	switch strings.ToLower(strings.TrimSpace(os.Getenv("AGENT_LIVE_MOQ_QUALITY"))) {
+	case models.StreamQualityHigh:
+		qualities = []string{models.StreamQualityHigh}
+	case models.StreamQualityLow:
+		qualities = []string{models.StreamQualityLow}
+	}
+
+	relayURL := os.Getenv("AGENT_LIVE_MOQ_URL")
+	if relayURL == "" {
+		relayURL = defaultMoQRelayURL
+	}
+	broadcastPrefix := os.Getenv("AGENT_LIVE_MOQ_BROADCAST_PREFIX")
+
+	ctx := context.Background()
+	if communication.Context != nil {
+		ctx = *communication.Context
+	}
+
+	for _, quality := range qualities {
+		queue := communication.Queue
+		sourceLabel := "main"
+		if models.SelectSubStreamForQuality(config, quality, subStreamEnabled) && communication.SubQueue != nil {
+			queue = communication.SubQueue
+			sourceLabel = "sub"
+		}
+		if queue == nil {
+			log.Log.Warning("cloud.StartLiveStreamMoQ(): packet queue for the " + quality + " tier is unavailable")
+			continue
+		}
+		go runLiveStreamMoQ(ctx, liveMoQConfig{
+			relayURL:    relayURL,
+			broadcast:   livemoq.BroadcastPath(broadcastPrefix, config.Key, quality),
+			quality:     quality,
+			sourceLabel: sourceLabel,
+			queue:       queue,
+		})
+	}
+}
+
+func runLiveStreamMoQ(ctx context.Context, config liveMoQConfig) {
+	log.Log.Info(fmt.Sprintf(
+		"cloud.runLiveStreamMoQ(): publishing %s stream (quality=%s) to %s/%s",
+		config.sourceLabel, config.quality, strings.TrimRight(config.relayURL, "/"), config.broadcast,
+	))
+
+	retryDelay := minMoQRetryDelay
+	for ctx.Err() == nil {
+		connectedAt := time.Now()
+		err := publishLiveStreamMoQ(ctx, config)
+		if ctx.Err() != nil {
+			return
+		}
+		log.Log.Warning("cloud.runLiveStreamMoQ(): publisher stopped: " + err.Error())
+		if time.Since(connectedAt) >= time.Minute {
+			retryDelay = minMoQRetryDelay
+		}
+
+		timer := time.NewTimer(retryDelay)
+		select {
+		case <-ctx.Done():
+			timer.Stop()
+			return
+		case <-timer.C:
+		}
+		if retryDelay < maxMoQRetryDelay {
+			retryDelay *= 2
+			if retryDelay > maxMoQRetryDelay {
+				retryDelay = maxMoQRetryDelay
+			}
+		}
+	}
+}
+
+func publishLiveStreamMoQ(ctx context.Context, config liveMoQConfig) error {
+	client, err := moq.Dial(ctx, config.relayURL)
+	if err != nil {
+		return fmt.Errorf("connect to relay: %w", err)
+	}
+	defer client.Close()
+
+	sessionCtx, cancelSessionWatch := context.WithCancel(ctx)
+	defer cancelSessionWatch()
+	sessionClosed := make(chan error, 1)
+	go func() {
+		sessionClosed <- client.Session().Closed(sessionCtx)
+	}()
+
+	broadcast, err := client.CreateBroadcast(config.broadcast)
+	if err != nil {
+		return fmt.Errorf("create broadcast: %w", err)
+	}
+	defer broadcast.Finish()
+
+	stream, err := broadcast.PublishMedia("avc3", nil)
+	if err != nil {
+		return fmt.Errorf("create H.264 media stream: %w", err)
+	}
+	defer stream.Finish()
+
+	// Only upload while this tier is actually being watched. `publishing` starts
+	// true so the track becomes discoverable on the relay even before the first
+	// subscriber ever arrives; from the moment a viewer has attached once, the
+	// subscriber watcher takes over and idles the tier again when everybody left.
+	watchCtx, cancelWatch := context.WithCancel(ctx)
+	defer cancelWatch()
+	publishing := &atomic.Bool{}
+	publishing.Store(true)
+	go watchLiveStreamMoQSubscribers(watchCtx, stream, publishing, config)
+
+	cursor := config.queue.Latest()
+	gate := livemoq.FrameGate{}
+	deduplicator := livemoq.KeyframeDeduplicator{}
+	var lastSlowWriteWarning time.Time
+	var lastDuplicateKeyframeWarning time.Time
+	idle := false
+	for {
+		select {
+		case err := <-sessionClosed:
+			return fmt.Errorf("relay session closed: %w", err)
+		default:
+		}
+
+		packet, err := cursor.ReadPacket()
+		if err != nil {
+			return fmt.Errorf("read packet: %w", err)
+		}
+		if !publishing.Load() {
+			// Keep draining the cursor so we stay at the live edge, but publish
+			// nothing. The gate is closed so the next viewer resumes on a keyframe.
+			if !idle {
+				gate.Reset()
+				deduplicator.Reset()
+				idle = true
+			}
+			continue
+		}
+		idle = false
+		if !packet.IsVideo || len(packet.Data) == 0 || !strings.EqualFold(packet.Codec, "H264") {
+			continue
+		}
+		allowed, event := gate.Allow(packet.IsKeyFrame, packet.CurrentTime, time.Now(), maxMoQLivePacketAge)
+		switch event {
+		case livemoq.FrameGateEventStarted:
+			log.Log.Info("cloud.publishLiveStreamMoQ(): first H.264 keyframe received; " + config.label() + " broadcast is live")
+		case livemoq.FrameGateEventLagging:
+			log.Log.Warning("cloud.publishLiveStreamMoQ(): " + config.label() + " stream is lagging; dropping packets until a recent keyframe")
+		case livemoq.FrameGateEventRecovered:
+			log.Log.Info("cloud.publishLiveStreamMoQ(): caught up with the " + config.label() + " live stream at a recent keyframe")
+		}
+		if !allowed {
+			continue
+		}
+		payload, normalizationStats, err := livemoq.NormalizeH264AccessUnitWithStats(packet.Data)
+		if err != nil {
+			return fmt.Errorf("normalize H.264 access unit: %w", err)
+		}
+		if normalizationStats.DuplicateIDRNALUs > 0 && time.Since(lastDuplicateKeyframeWarning) >= moQWriteWarningInterval {
+			log.Log.Warning(fmt.Sprintf(
+				"cloud.publishLiveStreamMoQ(): %s removed %d duplicate IDR NALU(s) from H.264 keyframe (timestamp_ms=%d)",
+				config.label(), normalizationStats.DuplicateIDRNALUs, packet.Time,
+			))
+			lastDuplicateKeyframeWarning = time.Now()
+		}
+		if packet.IsKeyFrame && deduplicator.IsDuplicate(packet.Time, packet.CurrentTime, payload, time.Now(), duplicateKeyframeWindow) {
+			if time.Since(lastDuplicateKeyframeWarning) >= moQWriteWarningInterval {
+				log.Log.Warning(fmt.Sprintf(
+					"cloud.publishLiveStreamMoQ(): %s dropping duplicate H.264 keyframe (timestamp_ms=%d)",
+					config.label(), packet.Time,
+				))
+				lastDuplicateKeyframeWarning = time.Now()
+			}
+			continue
+		}
+		frame := moq.Frame{
+			Payload:     payload,
+			TimestampUs: livemoq.TimestampUs(packet.Time),
+		}
+		writeStartedAt := time.Now()
+		if err := stream.WriteFrame(frame); err != nil {
+			return fmt.Errorf("write H.264 access unit: %w", err)
+		}
+		writeDuration := time.Since(writeStartedAt)
+		if writeDuration >= slowMoQWriteThreshold && time.Since(lastSlowWriteWarning) >= moQWriteWarningInterval {
+			packetAge := time.Duration(0)
+			if packet.CurrentTime > 0 {
+				packetAge = time.Since(time.UnixMilli(packet.CurrentTime))
+				if packetAge < 0 {
+					packetAge = 0
+				}
+			}
+			log.Log.Warning(fmt.Sprintf(
+				"cloud.publishLiveStreamMoQ(): %s WriteFrame blocked for %s (packet_age=%s keyframe=%t)",
+				config.label(), writeDuration.Round(time.Millisecond), packetAge.Round(time.Millisecond), packet.IsKeyFrame,
+			))
+			lastSlowWriteWarning = time.Now()
+		}
+	}
+}
+
+// watchLiveStreamMoQSubscribers flips the publisher between uploading and idling
+// as viewers subscribe to and leave this tier's broadcast. Used and Unused both
+// block, so they are followed from their own goroutine.
+//
+// It deliberately never turns publishing off before the first subscriber has
+// been observed: the relay catalog is only complete once media has flowed, so
+// going idle up front could keep the tier undiscoverable. On any error it fails
+// open (keeps publishing) — a stalled watcher must never take the live view down.
+func watchLiveStreamMoQSubscribers(ctx context.Context, stream *moq.MediaProducer, publishing *atomic.Bool, config liveMoQConfig) {
+	for ctx.Err() == nil {
+		if err := stream.Used(ctx); err != nil {
+			publishing.Store(true)
+			return
+		}
+		if publishing.CompareAndSwap(false, true) {
+			log.Log.Info("cloud.watchLiveStreamMoQSubscribers(): viewer subscribed, resuming the " + config.label() + " broadcast")
+		}
+
+		if err := stream.Unused(ctx); err != nil {
+			publishing.Store(true)
+			return
+		}
+		publishing.Store(false)
+		log.Log.Info("cloud.watchLiveStreamMoQSubscribers(): no viewers left, idling the " + config.label() + " broadcast")
+	}
+}

machinery/src/cloud/recording_metadata.go

@@ -0,0 +1,88 @@
+package cloud
+
+import (
+	"encoding/json"
+	"math"
+	"net/http"
+	"os"
+	"path/filepath"
+	"strconv"
+	"strings"
+
+	"github.com/kerberos-io/agent/machinery/src/models"
+)
+
+const recordingFPSHeader = "X-Kerberos-Storage-Fps"
+const recordingDurationHeader = "X-Kerberos-Storage-Duration"
+const recordingTimestampHeader = "X-Kerberos-Storage-Timestamp"
+
+// queuedRecordingFPS reads the FPS snapshot written into the upload marker
+// when the recording was finalized. Historical empty markers intentionally
+// return no value so receivers can retain their existing MP4-derived fallback.
+func queuedRecordingFPS(fileName string) string {
+	value, ok := readRecordingUploadMetadata(fileName)
+	if !ok {
+		return ""
+	}
+
+	marker := strings.TrimSpace(string(value))
+	if strings.HasPrefix(marker, "{") {
+		metadata, ok := decodeRecordingUploadMetadata(value)
+		if !ok || metadata.FPS <= 0 || metadata.FPS > 240 || math.IsInf(metadata.FPS, 0) || math.IsNaN(metadata.FPS) {
+			return ""
+		}
+		return strconv.FormatFloat(metadata.FPS, 'f', -1, 64)
+	}
+
+	// Compatibility with markers created before upload metadata used JSON.
+	fps := marker
+	parsed, err := strconv.ParseFloat(fps, 64)
+	if err != nil || parsed <= 0 || parsed > 240 || math.IsInf(parsed, 0) || math.IsNaN(parsed) {
+		return ""
+	}
+	return fps
+}
+
+func queuedRecordingMetadata(fileName string) (models.RecordingUploadMetadata, bool) {
+	value, ok := readRecordingUploadMetadata(fileName)
+	if !ok || !strings.HasPrefix(strings.TrimSpace(string(value)), "{") {
+		return models.RecordingUploadMetadata{}, false
+	}
+	return decodeRecordingUploadMetadata(value)
+}
+
+func decodeRecordingUploadMetadata(value []byte) (models.RecordingUploadMetadata, bool) {
+	var metadata models.RecordingUploadMetadata
+	if err := json.Unmarshal(value, &metadata); err != nil {
+		return models.RecordingUploadMetadata{}, false
+	}
+	return metadata, true
+}
+
+func readRecordingUploadMetadata(fileName string) ([]byte, bool) {
+	markerNames := []string{
+		models.RecordingUploadMetadataFileName(fileName),
+		filepath.Base(fileName),
+	}
+	for _, markerName := range markerNames {
+		value, err := os.ReadFile(filepath.Join("data", "cloud", markerName))
+		if err == nil {
+			return value, true
+		}
+	}
+	return nil, false
+}
+
+func setQueuedRecordingMetadataHeaders(header http.Header, fileName string) {
+	if fps := queuedRecordingFPS(fileName); fps != "" {
+		header.Set(recordingFPSHeader, fps)
+	}
+	if metadata, ok := queuedRecordingMetadata(fileName); ok {
+		if metadata.Duration > 0 {
+			header.Set(recordingDurationHeader, strconv.FormatUint(metadata.Duration, 10))
+		}
+		if metadata.Timestamp > 0 {
+			header.Set(recordingTimestampHeader, strconv.FormatInt(metadata.Timestamp, 10))
+		}
+	}
+}

machinery/src/cloud/tus_client.go

@@ -163,6 +163,15 @@ func runTusUpload(baseURL, metadata, fileName, label, slot string, setHeaders tu
 	// never consume the retry budget (matching the legacy single-POST behaviour).
 	lastStatus := 0
 
+	// highWaterOffset is the furthest server-acknowledged offset observed across
+	// all attempts (via HEAD or PATCH). It lets the retry budget be refreshed only
+	// on GENUINE net forward progress. Without it, a server that keeps resetting the
+	// offset — e.g. a persistent 409 ERR_MISMATCHED_OFFSET where HEAD reports 0 again
+	// while the first chunk still "succeeds" — would refresh the budget every attempt
+	// and loop forever, wedging the upload worker on one recording and saturating the
+	// uplink.
+	highWaterOffset := int64(0)
+
 	for attempt := 0; attempt < maxAttempts; attempt++ {
 		// (1) Ensure we have an active upload URL, creating one if needed.
 		if uploadURL == "" {
@@ -196,6 +205,15 @@ func runTusUpload(baseURL, metadata, fileName, label, slot string, setHeaders tu
 			continue
 		}
 
+		// The furthest offset any previous attempt reached. If this attempt pushes
+		// past it (via HEAD showing server-side progress or a successful PATCH) we made
+		// genuine net progress and may refresh the retry budget; if not, a repeated
+		// failure at the same spot must count against maxAttempts.
+		startHighWater := highWaterOffset
+		if offset > highWaterOffset {
+			highWaterOffset = offset
+		}
+
 		// (3) All bytes are present but the upload was not finalized (e.g. the
 		// completion hook failed). A completed tus upload cannot be re-finalized
 		// with another PATCH, so delete it and re-upload to force a clean finalize.
@@ -216,7 +234,6 @@ func runTusUpload(baseURL, metadata, fileName, label, slot string, setHeaders tu
 		// checkpointing the offset after each one so an interruption resumes from the
 		// last completed chunk instead of re-uploading everything.
 		chunkSize := tusChunkSize()
-		progressed := false
 		patchFailed := false
 		var lastBody string
 		loggedProgressBucket := tusProgressBucket(offset, size)
@@ -244,10 +261,10 @@ func runTusUpload(baseURL, metadata, fileName, label, slot string, setHeaders tu
 				patchFailed = true
 				break
 			}
-			if newOffset > offset {
-				progressed = true
-			}
 			offset = newOffset
+			if offset > highWaterOffset {
+				highWaterOffset = offset
+			}
 			lastBody = respBody
 			logTusUploadProgress(label, offset, size, &loggedProgressBucket)
 			if offset < size {
@@ -256,10 +273,13 @@ func runTusUpload(baseURL, metadata, fileName, label, slot string, setHeaders tu
 			}
 		}
 		if patchFailed {
-			if progressed {
-				// Forward progress refreshes the retry budget: maxAttempts bounds the
-				// number of consecutive failures, not the number of chunks needed for
-				// a large recording.
+			if highWaterOffset > startHighWater {
+				// Genuine net progress (we advanced past the furthest point any previous
+				// attempt reached) refreshes the retry budget: maxAttempts bounds the
+				// number of consecutive *non-progressing* failures, not the number of
+				// chunks needed for a large recording. A server that keeps rejecting the
+				// same offset (no net progress, e.g. a persistent ERR_MISMATCHED_OFFSET)
+				// therefore gives up after maxAttempts instead of retrying forever.
 				attempt = -1
 			}
 			continue
@@ -285,14 +305,17 @@ func runTusUpload(baseURL, metadata, fileName, label, slot string, setHeaders tu
 // is additionally carried in the tus Upload-Metadata.
 func uploadVaultResumable(vault models.KStorage, publicKey, deviceKey, fileName, label, slot string) (bool, bool, bool, string, error) {
 	baseURL := strings.TrimRight(vault.URI, "/") + tusUploadPath
-	metadata := encodeTusMetadata(map[string]string{
+	metadataValues := map[string]string{
 		"filename":  fileName,
 		"device":    deviceKey,
 		"directory": vault.Directory,
 		"provider":  vault.Provider,
 		"capture":   "IPCamera",
 		"cloudkey":  publicKey,
-	})
+		"fps":       queuedRecordingFPS(fileName),
+	}
+	addRecordingTusMetadata(metadataValues, fileName)
+	metadata := encodeTusMetadata(metadataValues)
 	setHeaders := func(h http.Header, fn string) {
 		setVaultTusHeaders(h, vault, publicKey, deviceKey, fn)
 	}
@@ -306,17 +329,33 @@ func uploadVaultResumable(vault models.KStorage, publicKey, deviceKey, fileName,
 // intentionally omitted from the metadata here.
 func uploadHubResumable(config *models.Config, fileName, label, slot string) (bool, bool, bool, string, error) {
 	baseURL := strings.TrimRight(config.HubURI, "/") + tusUploadPath
-	metadata := encodeTusMetadata(map[string]string{
+	metadataValues := map[string]string{
 		"filename": fileName,
 		"device":   config.Key,
 		"capture":  "IPCamera",
-	})
+		"fps":      queuedRecordingFPS(fileName),
+	}
+	addRecordingTusMetadata(metadataValues, fileName)
+	metadata := encodeTusMetadata(metadataValues)
 	setHeaders := func(h http.Header, fn string) {
 		setHubTusHeaders(h, config, fn)
 	}
 	return runTusUpload(baseURL, metadata, fileName, label, slot, setHeaders)
 }
 
+func addRecordingTusMetadata(values map[string]string, fileName string) {
+	metadata, ok := queuedRecordingMetadata(fileName)
+	if !ok {
+		return
+	}
+	if metadata.Duration > 0 {
+		values["duration"] = strconv.FormatUint(metadata.Duration, 10)
+	}
+	if metadata.Timestamp > 0 {
+		values["timestamp"] = strconv.FormatInt(metadata.Timestamp, 10)
+	}
+}
+
 // tusCreate performs the tus "creation" request (POST). On success it returns
 // the resolved upload URL the agent should use for subsequent HEAD/PATCH calls.
 func tusCreate(client *http.Client, baseURL string, size int64, metadata string, setHeaders tusHeaderFunc, fileName string) (string, int, error) {

machinery/src/cloud/tus_client_test.go

@@ -49,6 +49,14 @@ type fakeTus struct {
 	// after storing the bytes, simulating a failed completion hook.
 	failFinalize int
 
+	// loseProgress simulates a vault that never durably retains the in-progress
+	// upload: every PATCH is acknowledged (the response advertises the advanced
+	// offset) but the stored offset is immediately reset to 0. HEAD therefore
+	// keeps reporting 0 and the next chunk — sent at the advanced offset — is
+	// rejected with 409, reproducing the cross-replica ERR_MISMATCHED_OFFSET
+	// loop that previously wedged the agent's upload worker forever.
+	loseProgress bool
+
 	// requests records the headers of every received request (in order) so
 	// tests can assert which auth/routing headers the client sent per method.
 	requests []recordedRequest
@@ -155,6 +163,31 @@ func (s *fakeTus) ServeHTTP(w http.ResponseWriter, r *http.Request) {
 			w.WriteHeader(http.StatusNotFound)
 			return
 		}
+		if s.loseProgress {
+			reqOffset, _ := strconv.ParseInt(r.Header.Get("Upload-Offset"), 10, 64)
+			s.mu.Lock()
+			cur := u.offset
+			if reqOffset != cur {
+				// The offset the client resumes from no longer matches what this
+				// "replica" retained, so reject like a vault returning
+				// ERR_MISMATCHED_OFFSET.
+				s.mu.Unlock()
+				w.Header().Set("Upload-Offset", strconv.FormatInt(cur, 10))
+				w.WriteHeader(http.StatusConflict)
+				return
+			}
+			n, _ := io.Copy(io.Discard, r.Body)
+			s.lastPatchBytes = n
+			s.patchSizes = append(s.patchSizes, n)
+			// Advertise progress to the client, then immediately forget it so the
+			// next chunk (sent at the advanced offset) mismatches again.
+			reported := cur + n
+			u.offset = 0
+			s.mu.Unlock()
+			w.Header().Set("Upload-Offset", strconv.FormatInt(reported, 10))
+			w.WriteHeader(http.StatusNoContent)
+			return
+		}
 		n, _ := io.Copy(io.Discard, r.Body)
 		s.mu.Lock()
 		u.offset += n
@@ -210,6 +243,17 @@ func withRecording(t *testing.T, fileName string, payload []byte) {
 	}
 }
 
+func withQueuedRecordingFPS(t *testing.T, fileName, fps string) {
+	t.Helper()
+	if err := os.MkdirAll("data/cloud", 0o755); err != nil {
+		t.Fatalf("mkdir cloud queue: %v", err)
+	}
+	markerName := models.RecordingUploadMetadataFileName(fileName)
+	if err := os.WriteFile(filepath.Join("data/cloud", markerName), []byte(fps), 0o644); err != nil {
+		t.Fatalf("write cloud queue marker: %v", err)
+	}
+}
+
 func testVault(uri string) models.KStorage {
 	return models.KStorage{
 		URI:             uri,
@@ -220,6 +264,36 @@ func testVault(uri string) models.KStorage {
 	}
 }
 
+func TestUploadKerberosVaultSkipsEmptyRecording(t *testing.T) {
+	fileName := "1787015373_3-654_office-camera17_0-0-0-0_-1_1960.mp4"
+	withRecording(t, fileName, nil)
+
+	requestCount := 0
+	server := httptest.NewServer(http.HandlerFunc(func(w http.ResponseWriter, r *http.Request) {
+		requestCount++
+		w.WriteHeader(http.StatusInternalServerError)
+	}))
+	defer server.Close()
+
+	vault := testVault(server.URL)
+	configuration := &models.Configuration{Config: models.Config{
+		Key:               "device-key",
+		KStorage:          &vault,
+		KStorageSecondary: &models.KStorage{},
+	}}
+
+	uploaded, configured, err := UploadKerberosVault(configuration, fileName)
+	if err != nil {
+		t.Fatalf("UploadKerberosVault() error = %v", err)
+	}
+	if uploaded || configured {
+		t.Fatalf("UploadKerberosVault() uploaded/configured = %v/%v, want false/false", uploaded, configured)
+	}
+	if requestCount != 0 {
+		t.Fatalf("Vault received %d requests, want 0", requestCount)
+	}
+}
+
 func TestUploadVaultResumable_HappyPath(t *testing.T) {
 	srv := newFakeTus()
 	ts := httptest.NewServer(srv)
@@ -228,6 +302,7 @@ func TestUploadVaultResumable_HappyPath(t *testing.T) {
 	fileName := "1564859471_6-474162_oprit_577-283-727-375_1153_27.mp4"
 	payload := bytes.Repeat([]byte("x"), 4096)
 	withRecording(t, fileName, payload)
+	withQueuedRecordingFPS(t, fileName, `{"filename":"recording.mp4","device_key":"device-key","timestamp":1785934709414,"duration":20452,"fps":29.97}`)
 
 	uploaded, responded, supported, _, err := uploadVaultResumable(testVault(ts.URL), "pk", "dev", fileName, "test", "primary")
 	if err != nil {
@@ -242,6 +317,103 @@ func TestUploadVaultResumable_HappyPath(t *testing.T) {
 	if _, err := os.Stat(tusSidecarPath(fileName, "primary")); !os.IsNotExist(err) {
 		t.Fatalf("expected sidecar to be removed after success, stat err = %v", err)
 	}
+	posts := srv.requestsForMethod(http.MethodPost)
+	metadata := decodeTusMetadata(posts[0].header.Get("Upload-Metadata"))
+	if got := metadata["fps"]; got != "29.97" {
+		t.Fatalf("POST metadata fps = %q, want %q", got, "29.97")
+	}
+	if got := metadata["duration"]; got != "20452" {
+		t.Fatalf("POST metadata duration = %q, want %q", got, "20452")
+	}
+	if got := metadata["timestamp"]; got != "1785934709414" {
+		t.Fatalf("POST metadata timestamp = %q, want %q", got, "1785934709414")
+	}
+}
+
+func TestQueuedRecordingFPSValidation(t *testing.T) {
+	for _, test := range []struct {
+		name string
+		fps  string
+		want string
+	}{
+		{name: "json", fps: `{"fps":29}`, want: "29"},
+		{name: "json fractional", fps: `{"fps":17.35}`, want: "17.35"},
+		{name: "json with future field", fps: `{"fps":29,"codec":"h264"}`, want: "29"},
+		{name: "json without fps", fps: `{}`},
+		{name: "json invalid fps", fps: `{"fps":241}`},
+		{name: "legacy fractional", fps: "29.97", want: "29.97"},
+		{name: "legacy trimmed", fps: " 25 \n", want: "25"},
+		{name: "empty"},
+		{name: "invalid", fps: "invalid"},
+		{name: "zero", fps: "0"},
+		{name: "negative", fps: "-1"},
+		{name: "nan", fps: "NaN"},
+		{name: "infinite", fps: "+Inf"},
+		{name: "unreasonable", fps: "241"},
+	} {
+		t.Run(test.name, func(t *testing.T) {
+			fileName := "recording.mp4"
+			withRecording(t, fileName, []byte("recording"))
+			withQueuedRecordingFPS(t, fileName, test.fps)
+
+			if got := queuedRecordingFPS(fileName); got != test.want {
+				t.Fatalf("queuedRecordingFPS() = %q, want %q", got, test.want)
+			}
+
+			header := make(http.Header)
+			setQueuedRecordingMetadataHeaders(header, fileName)
+			if got := header.Get(recordingFPSHeader); got != test.want {
+				t.Fatalf("legacy FPS header = %q, want %q", got, test.want)
+			}
+		})
+	}
+}
+
+func TestQueuedRecordingFPSAllowsMissingHistoricalMarker(t *testing.T) {
+	fileName := "recording.mp4"
+	withRecording(t, fileName, []byte("recording"))
+
+	if got := queuedRecordingFPS(fileName); got != "" {
+		t.Fatalf("queuedRecordingFPS() = %q, want empty for missing marker", got)
+	}
+	header := make(http.Header)
+	setQueuedRecordingMetadataHeaders(header, fileName)
+	if got := header.Get(recordingFPSHeader); got != "" {
+		t.Fatalf("legacy FPS header = %q, want empty for missing marker", got)
+	}
+}
+
+func TestQueuedRecordingMetadataHeaders(t *testing.T) {
+	fileName := "recording.mp4"
+	withRecording(t, fileName, []byte("recording"))
+	withQueuedRecordingFPS(t, fileName, `{"filename":"recording.mp4","device_key":"device-key","timestamp":1785934709414,"duration":20452,"fps":25}`)
+
+	header := make(http.Header)
+	setQueuedRecordingMetadataHeaders(header, fileName)
+	if got := header.Get(recordingFPSHeader); got != "25" {
+		t.Fatalf("FPS header = %q", got)
+	}
+	if got := header.Get(recordingDurationHeader); got != "20452" {
+		t.Fatalf("duration header = %q", got)
+	}
+	if got := header.Get(recordingTimestampHeader); got != "1785934709414" {
+		t.Fatalf("timestamp header = %q", got)
+	}
+}
+
+func TestQueuedRecordingFPSAllowsLegacyMarkerFileName(t *testing.T) {
+	fileName := "recording.mp4"
+	withRecording(t, fileName, []byte("recording"))
+	if err := os.MkdirAll("data/cloud", 0o755); err != nil {
+		t.Fatalf("mkdir cloud queue: %v", err)
+	}
+	if err := os.WriteFile(filepath.Join("data/cloud", fileName), []byte("25"), 0o644); err != nil {
+		t.Fatalf("write legacy cloud queue marker: %v", err)
+	}
+
+	if got := queuedRecordingFPS(fileName); got != "25" {
+		t.Fatalf("queuedRecordingFPS() = %q, want legacy marker FPS", got)
+	}
 }
 
 func TestUploadVaultResumable_Chunked(t *testing.T) {
@@ -393,6 +565,61 @@ func TestUploadVaultResumable_NetworkErrorKeepsRetryBudget(t *testing.T) {
 	}
 }
 
+func TestUploadVaultResumable_MismatchedOffsetGivesUp(t *testing.T) {
+	// A vault that never durably retains the in-progress upload (offset resets to
+	// 0 between chunks) makes every resume "progress" by one chunk and then fail
+	// the next chunk with 409. Before the high-water gating fix this refreshed the
+	// retry budget every attempt and looped forever, wedging the upload worker and
+	// saturating the uplink (which starved heartbeats and reported the camera
+	// offline). The loop must now be bounded: give up after a fixed number of
+	// non-progressing attempts and report responded=true so the caller re-queues.
+	srv := newFakeTus()
+	srv.loseProgress = true
+	ts := httptest.NewServer(srv)
+	defer ts.Close()
+
+	// Keep the between-attempt back-off tiny so the test stays fast.
+	oldDelay := tusBackoffBaseDelay
+	tusBackoffBaseDelay = time.Millisecond
+	defer func() { tusBackoffBaseDelay = oldDelay }()
+
+	// Force multiple chunks so there is always a second chunk to be rejected.
+	t.Setenv("AGENT_TUS_CHUNK_SIZE_BYTES", "4096")
+
+	fileName := "1564859471_6-474162_oprit_577-283-727-375_1153_27.mp4"
+	withRecording(t, fileName, bytes.Repeat([]byte("m"), 12288))
+
+	done := make(chan struct{})
+	var uploaded, responded bool
+	var upErr error
+	go func() {
+		uploaded, responded, _, _, upErr = uploadVaultResumable(testVault(ts.URL), "pk", "dev", fileName, "test", "primary")
+		close(done)
+	}()
+
+	select {
+	case <-done:
+	case <-time.After(30 * time.Second):
+		t.Fatal("resumable upload did not terminate: the retry loop is unbounded on a persistent mismatched offset")
+	}
+
+	if uploaded {
+		t.Fatal("expected uploaded=false when the vault never retains the offset")
+	}
+	if !responded {
+		t.Fatal("expected responded=true (the vault answered) so the caller re-queues the recording")
+	}
+	if upErr == nil {
+		t.Fatal("expected an error when the upload cannot complete")
+	}
+
+	// The bounded retry budget must cap the number of PATCH requests. Two PATCHes
+	// per attempt across a handful of attempts stays comfortably below this.
+	if count, _ := srv.patchCounts(); count > 50 {
+		t.Fatalf("expected a bounded number of PATCH requests, got %d (retry loop not bounded)", count)
+	}
+}
+
 func TestUploadVaultResumable_FinalizeRetry(t *testing.T) {
 	srv := newFakeTus()
 	srv.failFinalize = 1
@@ -491,6 +718,7 @@ func TestUploadHubResumable_HappyPath(t *testing.T) {
 	fileName := "1564859471_6-474162_oprit_577-283-727-375_1153_27.mp4"
 	payload := bytes.Repeat([]byte("h"), 4096)
 	withRecording(t, fileName, payload)
+	withQueuedRecordingFPS(t, fileName, "29.97")
 
 	uploaded, _, supported, _, err := uploadHubResumable(testHubConfig(ts.URL), fileName, "test", "hub")
 	if err != nil {
@@ -561,6 +789,9 @@ func TestUploadHubResumable_HappyPath(t *testing.T) {
 	if meta["capture"] != "IPCamera" {
 		t.Errorf("hub metadata capture = %q, want %q", meta["capture"], "IPCamera")
 	}
+	if meta["fps"] != "29.97" {
+		t.Errorf("hub metadata fps = %q, want %q", meta["fps"], "29.97")
+	}
 }
 
 func TestUploadHubResumable_Unsupported(t *testing.T) {

machinery/src/components/backchannel.go

@@ -2,7 +2,9 @@ package components
 
 import (
 	"bufio"
+	"context"
 	"fmt"
+	"math/rand"
 	"os"
 	"time"
 
@@ -15,6 +17,62 @@ import (
 	"github.com/zaf/g711"
 )
 
+const (
+	backchannelSampleRate       = 8000
+	backchannelTalkspurtGap     = 500 * time.Millisecond
+	backchannelReconnectInitial = time.Second
+	backchannelReconnectMax     = 30 * time.Second
+)
+
+type backchannelClient interface {
+	ConnectBackChannel(ctx context.Context, otelContext context.Context) error
+	StartBackChannel(ctx context.Context, otelContext context.Context) error
+	WritePacket(pkt packets.Packet) error
+	Close(otelContext context.Context) error
+}
+
+type backchannelPacketizer struct {
+	sequenceNumber uint16
+	timestamp      uint32
+	ssrc           uint32
+	lastPacketAt   time.Time
+}
+
+func newBackchannelPacketizer() backchannelPacketizer {
+	return backchannelPacketizer{
+		sequenceNumber: uint16(rand.Uint32()),
+		timestamp:      rand.Uint32(),
+		ssrc:           rand.Uint32(),
+	}
+}
+
+func (p *backchannelPacketizer) packet(audio models.AudioDataPartial, now time.Time) packets.Packet {
+	bufferUlaw := make([]byte, len(audio.Data))
+	for index, sample := range audio.Data {
+		bufferUlaw[index] = g711.EncodeUlawFrame(sample)
+	}
+
+	pkt := packets.Packet{
+		Packet: &rtp.Packet{
+			Header: rtp.Header{
+				Version:        2,
+				Marker:         p.lastPacketAt.IsZero() || now.Sub(p.lastPacketAt) >= backchannelTalkspurtGap,
+				PayloadType:    0,
+				SequenceNumber: p.sequenceNumber,
+				Timestamp:      p.timestamp,
+				SSRC:           p.ssrc,
+			},
+			Payload: bufferUlaw,
+		},
+	}
+
+	p.timestamp += uint32(len(bufferUlaw))
+	p.sequenceNumber++
+	p.lastPacketAt = now
+
+	return pkt
+}
+
 func GetBackChannelAudioCodec(streams []av.CodecData, communication *models.Communication) av.AudioCodecData {
 	for _, stream := range streams {
 		if stream.Type().IsAudio() {
@@ -31,41 +89,115 @@ func GetBackChannelAudioCodec(streams []av.CodecData, communication *models.Comm
 }
 
 func WriteAudioToBackchannel(communication *models.Communication, rtspClient capture.RTSPClient) {
-	log.Log.Info("Audio.WriteAudioToBackchannel(): writing to backchannel audio codec")
-	length := uint32(0)
-	sequenceNumber := uint16(0)
-	for audio := range communication.HandleAudio {
-		// Encode PCM to MULAW
-		var bufferUlaw []byte
-		for _, v := range audio.Data {
-			b := g711.EncodeUlawFrame(v)
-			bufferUlaw = append(bufferUlaw, b)
-		}
-
-		pkt := packets.Packet{
-			Packet: &rtp.Packet{
-				Header: rtp.Header{
-					Version:        2,
-					Marker:         true, // should be true
-					PayloadType:    0,    //packet.PayloadType, // will be owerwriten
-					SequenceNumber: sequenceNumber,
-					Timestamp:      uint32(length),
-					SSRC:           1293847657,
-				},
-				Payload: bufferUlaw,
-			},
-		}
-		err := rtspClient.WritePacket(pkt)
-		if err != nil {
-			log.Log.Error("Audio.WriteAudioToBackchannel(): error writing packet to backchannel")
-		}
-
-		length = (length + uint32(len(bufferUlaw))) % 65536
-		sequenceNumber = (sequenceNumber + 1) % 65535
-		time.Sleep(128 * time.Millisecond)
+	ctx := context.Background()
+	if communication.Context != nil {
+		ctx = *communication.Context
 	}
-	log.Log.Info("Audio.WriteAudioToBackchannel(): finished")
 
+	writeAudioToBackchannel(ctx, ctx, communication.HandleAudio, rtspClient)
+}
+
+func writeAudioToBackchannel(ctx context.Context, otelContext context.Context, audioChannel <-chan models.AudioDataPartial, rtspClient backchannelClient) {
+	log.Log.Info("Audio.WriteAudioToBackchannel(): writing to backchannel audio codec")
+
+	if err := rtspClient.StartBackChannel(ctx, otelContext); err != nil {
+		log.Log.Error("Audio.WriteAudioToBackchannel(): error starting backchannel: " + err.Error())
+		if !reconnectBackchannel(ctx, otelContext, rtspClient) {
+			log.Log.Info("Audio.WriteAudioToBackchannel(): stopped while reconnecting")
+			return
+		}
+	}
+
+	packetizer := newBackchannelPacketizer()
+	for {
+		select {
+		case <-ctx.Done():
+			log.Log.Info("Audio.WriteAudioToBackchannel(): stopped")
+			return
+		case audio, ok := <-audioChannel:
+			if !ok {
+				log.Log.Info("Audio.WriteAudioToBackchannel(): finished")
+				return
+			}
+
+			audio = latestBackchannelAudio(audio, audioChannel)
+			if len(audio.Data) == 0 {
+				continue
+			}
+
+			pkt := packetizer.packet(audio, time.Now())
+			if err := rtspClient.WritePacket(pkt); err != nil {
+				log.Log.Error("Audio.WriteAudioToBackchannel(): error writing packet to backchannel: " + err.Error())
+				if !reconnectBackchannel(ctx, otelContext, rtspClient) {
+					log.Log.Info("Audio.WriteAudioToBackchannel(): stopped while reconnecting")
+					return
+				}
+				packetizer = newBackchannelPacketizer()
+				continue
+			}
+
+			if !waitForBackchannel(ctx, time.Duration(len(audio.Data))*time.Second/backchannelSampleRate) {
+				log.Log.Info("Audio.WriteAudioToBackchannel(): stopped")
+				return
+			}
+		}
+	}
+}
+
+func latestBackchannelAudio(audio models.AudioDataPartial, audioChannel <-chan models.AudioDataPartial) models.AudioDataPartial {
+	for {
+		select {
+		case next, ok := <-audioChannel:
+			if !ok {
+				return audio
+			}
+			audio = next
+		default:
+			return audio
+		}
+	}
+}
+
+func reconnectBackchannel(ctx context.Context, otelContext context.Context, rtspClient backchannelClient) bool {
+	backoff := backchannelReconnectInitial
+	for {
+		if err := rtspClient.Close(otelContext); err != nil {
+			log.Log.Error("Audio.WriteAudioToBackchannel(): error closing failed backchannel: " + err.Error())
+		}
+		if ctx.Err() != nil {
+			return false
+		}
+
+		err := rtspClient.ConnectBackChannel(ctx, otelContext)
+		if err == nil {
+			err = rtspClient.StartBackChannel(ctx, otelContext)
+		}
+		if err == nil {
+			log.Log.Info("Audio.WriteAudioToBackchannel(): reconnected backchannel")
+			return true
+		}
+
+		log.Log.Error("Audio.WriteAudioToBackchannel(): error reconnecting backchannel: " + err.Error())
+		if !waitForBackchannel(ctx, backoff) {
+			return false
+		}
+		backoff *= 2
+		if backoff > backchannelReconnectMax {
+			backoff = backchannelReconnectMax
+		}
+	}
+}
+
+func waitForBackchannel(ctx context.Context, duration time.Duration) bool {
+	timer := time.NewTimer(duration)
+	defer timer.Stop()
+
+	select {
+	case <-ctx.Done():
+		return false
+	case <-timer.C:
+		return true
+	}
 }
 
 func WriteFileToBackChannel(infile av.DemuxCloser) {

machinery/src/components/backchannel_test.go

@@ -0,0 +1,209 @@
+package components
+
+import (
+	"context"
+	"errors"
+	"sync"
+	"testing"
+	"time"
+
+	"github.com/kerberos-io/agent/machinery/src/models"
+	"github.com/kerberos-io/agent/machinery/src/packets"
+)
+
+type fakeBackchannelClient struct {
+	mutex           sync.Mutex
+	startErrors     []error
+	connectError    error
+	writeErrors     []error
+	startCalls      int
+	connectCalls    int
+	closeCalls      int
+	writeCalls      int
+	connectAttempt  chan struct{}
+	successfulWrite chan packets.Packet
+}
+
+func (f *fakeBackchannelClient) ConnectBackChannel(context.Context, context.Context) error {
+	f.mutex.Lock()
+	f.connectCalls++
+	err := f.connectError
+	f.mutex.Unlock()
+
+	select {
+	case f.connectAttempt <- struct{}{}:
+	default:
+	}
+	return err
+}
+
+func (f *fakeBackchannelClient) StartBackChannel(context.Context, context.Context) error {
+	f.mutex.Lock()
+	defer f.mutex.Unlock()
+	f.startCalls++
+	if len(f.startErrors) == 0 {
+		return nil
+	}
+	err := f.startErrors[0]
+	f.startErrors = f.startErrors[1:]
+	return err
+}
+
+func (f *fakeBackchannelClient) WritePacket(pkt packets.Packet) error {
+	f.mutex.Lock()
+	f.writeCalls++
+	var err error
+	if len(f.writeErrors) != 0 {
+		err = f.writeErrors[0]
+		f.writeErrors = f.writeErrors[1:]
+	}
+	f.mutex.Unlock()
+
+	if err == nil {
+		select {
+		case f.successfulWrite <- pkt:
+		default:
+		}
+	}
+	return err
+}
+
+func (f *fakeBackchannelClient) Close(context.Context) error {
+	f.mutex.Lock()
+	f.closeCalls++
+	f.mutex.Unlock()
+	return nil
+}
+
+func (f *fakeBackchannelClient) callCounts() (start, connect, close, write int) {
+	f.mutex.Lock()
+	defer f.mutex.Unlock()
+	return f.startCalls, f.connectCalls, f.closeCalls, f.writeCalls
+}
+
+func TestBackchannelPacketizerUsesFullRTPClock(t *testing.T) {
+	packetizer := backchannelPacketizer{ssrc: 1}
+	audio := models.AudioDataPartial{Data: make([]int16, 1024)}
+	startedAt := time.Unix(1, 0)
+
+	var timestamp uint32
+	for index := 0; index <= 64; index++ {
+		pkt := packetizer.packet(audio, startedAt.Add(time.Duration(index)*128*time.Millisecond))
+		timestamp = pkt.Packet.Timestamp
+	}
+
+	if timestamp != 65536 {
+		t.Fatalf("timestamp after 64 frames = %d, want 65536", timestamp)
+	}
+}
+
+func TestBackchannelPacketizerUsesNaturalSequenceRollover(t *testing.T) {
+	packetizer := backchannelPacketizer{sequenceNumber: ^uint16(0), ssrc: 1}
+	audio := models.AudioDataPartial{Data: []int16{0}}
+	startedAt := time.Unix(1, 0)
+
+	last := packetizer.packet(audio, startedAt)
+	firstAfterRollover := packetizer.packet(audio, startedAt.Add(time.Millisecond))
+
+	if last.Packet.SequenceNumber != ^uint16(0) {
+		t.Fatalf("last sequence number = %d, want %d", last.Packet.SequenceNumber, ^uint16(0))
+	}
+	if firstAfterRollover.Packet.SequenceNumber != 0 {
+		t.Fatalf("first sequence number after rollover = %d, want 0", firstAfterRollover.Packet.SequenceNumber)
+	}
+}
+
+func TestBackchannelPacketizerMarksTalkspurtStart(t *testing.T) {
+	packetizer := backchannelPacketizer{ssrc: 1}
+	audio := models.AudioDataPartial{Data: []int16{0}}
+	startedAt := time.Unix(1, 0)
+
+	first := packetizer.packet(audio, startedAt)
+	continuous := packetizer.packet(audio, startedAt.Add(128*time.Millisecond))
+	afterGap := packetizer.packet(audio, startedAt.Add(backchannelTalkspurtGap+128*time.Millisecond))
+
+	if !first.Packet.Marker {
+		t.Fatal("first packet must mark the start of a talkspurt")
+	}
+	if continuous.Packet.Marker {
+		t.Fatal("continuous packet must not carry the marker bit")
+	}
+	if !afterGap.Packet.Marker {
+		t.Fatal("packet after an audio gap must mark a new talkspurt")
+	}
+}
+
+func TestWriteAudioToBackchannelReconnectsAfterWriteFailure(t *testing.T) {
+	writeFailure := errors.New("EOF")
+	client := &fakeBackchannelClient{
+		writeErrors:     []error{writeFailure, nil},
+		connectAttempt:  make(chan struct{}, 1),
+		successfulWrite: make(chan packets.Packet, 1),
+	}
+	audioChannel := make(chan models.AudioDataPartial, 2)
+	ctx, cancel := context.WithCancel(context.Background())
+	done := make(chan struct{})
+	go func() {
+		writeAudioToBackchannel(ctx, ctx, audioChannel, client)
+		close(done)
+	}()
+
+	audioChannel <- models.AudioDataPartial{Data: make([]int16, 1024)}
+	select {
+	case <-client.connectAttempt:
+	case <-time.After(time.Second):
+		t.Fatal("backchannel was not reconnected after the write failure")
+	}
+
+	audioChannel <- models.AudioDataPartial{Data: make([]int16, 1024)}
+	select {
+	case pkt := <-client.successfulWrite:
+		if !pkt.Packet.Marker {
+			t.Fatal("first packet after reconnect must mark a new talkspurt")
+		}
+	case <-time.After(time.Second):
+		t.Fatal("fresh audio was not written after reconnect")
+	}
+
+	cancel()
+	select {
+	case <-done:
+	case <-time.After(time.Second):
+		t.Fatal("backchannel writer did not stop after cancellation")
+	}
+
+	startCalls, connectCalls, closeCalls, writeCalls := client.callCounts()
+	if startCalls != 2 || connectCalls != 1 || closeCalls != 1 || writeCalls != 2 {
+		t.Fatalf("calls (start, connect, close, write) = (%d, %d, %d, %d), want (2, 1, 1, 2)", startCalls, connectCalls, closeCalls, writeCalls)
+	}
+}
+
+func TestWriteAudioToBackchannelCancellationStopsReconnect(t *testing.T) {
+	client := &fakeBackchannelClient{
+		connectError:    errors.New("camera unavailable"),
+		writeErrors:     []error{errors.New("EOF")},
+		connectAttempt:  make(chan struct{}, 1),
+		successfulWrite: make(chan packets.Packet, 1),
+	}
+	audioChannel := make(chan models.AudioDataPartial, 1)
+	ctx, cancel := context.WithCancel(context.Background())
+	done := make(chan struct{})
+	go func() {
+		writeAudioToBackchannel(ctx, ctx, audioChannel, client)
+		close(done)
+	}()
+
+	audioChannel <- models.AudioDataPartial{Data: make([]int16, 1024)}
+	select {
+	case <-client.connectAttempt:
+	case <-time.After(time.Second):
+		t.Fatal("expected a reconnect attempt")
+	}
+	cancel()
+
+	select {
+	case <-done:
+	case <-time.After(250 * time.Millisecond):
+		t.Fatal("cancellation did not interrupt reconnect backoff")
+	}
+}

machinery/src/components/kerberos.go

@@ -74,6 +74,7 @@ func Bootstrap(ctx context.Context, configDirectory string, configuration *model
 	communication.HandleLiveHDKeepalive = make(chan string, 1)
 	communication.HandleLiveHDPeers = make(chan string, 1)
 	communication.HandleLiveHLS = make(chan string, 1)
+	communication.HandleAudio = make(chan models.AudioDataPartial, 10)
 	communication.IsConfiguring = abool.New()
 	communication.IsRecordingManual = abool.New()
 	communication.RecordingManualHeartbeat = &atomic.Int64{}
@@ -269,11 +270,11 @@ func RunAgent(configDirectory string, configuration *models.Configuration, commu
 	communication.MainStreamConnected = true
 
 	// Try to create backchannel
+	communication.HasBackChannel = false
 	rtspBackChannelClient := captureDevice.SetBackChannelClient(rtspUrl)
 	err = rtspBackChannelClient.ConnectBackChannel(ctx, ctxRunAgent)
 	if err == nil {
 		log.Log.Info("components.Kerberos.RunAgent(): opened RTSP backchannel stream: " + rtspUrl)
-		go rtspBackChannelClient.StartBackChannel(ctx, ctxRunAgent)
 	}
 
 	rtspSubClient := captureDevice.RTSPSubClient
@@ -305,6 +306,10 @@ func RunAgent(configDirectory string, configuration *models.Configuration, commu
 	// watching.
 	go cloud.HandleLiveStreamHLS(configuration, communication, mqttClient, subStreamEnabled)
 
+	// MoQ is available only in the dedicated CGO/glibc build. The standard
+	// static Alpine build resolves this hook to a no-op.
+	cloud.StartLiveStreamMoQ(configuration, communication, subStreamEnabled)
+
 	// Handle livestream HD (high resolution over WEBRTC). Both the main and sub
 	// stream are exposed as separate broadcasters so a viewer can request the
 	// high (main) or low (sub) resolution per peer connection; "auto" prefers the
@@ -341,7 +346,11 @@ func RunAgent(configDirectory string, configuration *models.Configuration, commu
 	communication.HandleONVIF = make(chan models.OnvifAction, 10)
 	go onvif.HandleONVIFActions(configuration, communication)
 
-	communication.HandleAudio = make(chan models.AudioDataPartial, 10)
+	// Handle ONVIF event stream — opt-in via Capture.ONVIFMotion="true".
+	// Stops when the agent's shared context is cancelled. The function
+	// is a no-op if ONVIFMotion is not enabled.
+	go onvif.HandleONVIFEventStream(*communication.Context, configuration, communication)
+
 	if rtspBackChannelClient.HasBackChannel {
 		communication.HasBackChannel = true
 		go WriteAudioToBackchannel(communication, rtspBackChannelClient)
@@ -432,9 +441,6 @@ func RunAgent(configDirectory string, configuration *models.Configuration, commu
 	close(communication.HandleMotion)
 	communication.HandleMotion = nil
 
-	close(communication.HandleAudio)
-	communication.HandleAudio = nil
-
 	close(communication.HandleONVIF)
 	communication.HandleONVIF = nil
 

machinery/src/computervision/main.go

@@ -24,17 +24,12 @@ func ProcessMotion(motionCursor *packets.QueueCursor, configuration *models.Conf
 	var motionRectangle models.MotionRectangle
 	var motionRectangles []models.MotionRectangle
 
-	// Resolve the motion sensitivity (pixel-change threshold):
-	//   nil (unset) -> default 150
-	//   0           -> motion detection DISABLED (temporary off switch from the UI)
-	//   > 0         -> trigger when the number of changed pixels exceeds it
+	// Resolve the motion sensitivity (pixel-change threshold). Nil, zero, and
+	// negative values use the historical default so older configurations keep
+	// recording after an upgrade.
 	pixelThreshold := 150
-	motionDisabled := false
-	if config.Capture.PixelChangeThreshold != nil {
+	if config.Capture.PixelChangeThreshold != nil && *config.Capture.PixelChangeThreshold > 0 {
 		pixelThreshold = *config.Capture.PixelChangeThreshold
-		if pixelThreshold <= 0 {
-			motionDisabled = true
-		}
 	}
 	// In motion mode we always run detection. In CONTINUOUS mode recording is
 	// 24/7 so motion detection is normally skipped, BUT if a motion region is
@@ -45,11 +40,7 @@ func ProcessMotion(motionCursor *packets.QueueCursor, configuration *models.Conf
 	continuousMode := config.Capture.Continuous == "true"
 	hasMotionRegion := config.Region != nil && len(config.Region.Polygon) > 0
 
-	if motionDisabled {
-
-		log.Log.Info("computervision.main.ProcessMotion(): motion detection disabled (pixelChangeThreshold set to 0), skipping.")
-
-	} else if continuousMode && !hasMotionRegion {
+	if continuousMode && !hasMotionRegion {
 
 		log.Log.Info("computervision.main.ProcessMotion(): continuous recording enabled and no motion region configured, so no motion detection required.")
 
@@ -231,12 +222,13 @@ func ProcessMotion(motionCursor *packets.QueueCursor, configuration *models.Conf
 													"mainWidth":  configuration.Config.Capture.IPCamera.Width,
 													"mainHeight": configuration.Config.Capture.IPCamera.Height,
 													"regions":    motionRectangles,
-													"polygon":    regionPolygons,												// Motion sensitivity = the pixel-change threshold that must
-												// be exceeded before motion triggers. The live view renders
-												// a reference square of sqrt(threshold) px (in this MOTION
-												// frame's pixel space) so the user can visually gauge how
-												// large a moving object must be before it is detected.
-												"pixelChangeThreshold": pixelThreshold,												},
+													"polygon":    regionPolygons, // Motion sensitivity = the pixel-change threshold that must
+													// be exceeded before motion triggers. The live view renders
+													// a reference square of sqrt(threshold) px (in this MOTION
+													// frame's pixel space) so the user can visually gauge how
+													// large a moving object must be before it is detected.
+													"pixelChangeThreshold": pixelThreshold,
+												},
 											},
 										}
 										payload, err := models.PackageMQTTMessage(configuration, message)

machinery/src/config/main.go

@@ -377,6 +377,9 @@ func applyAgentEnvVars(configuration *models.Configuration, prefix string, apply
 			case "AGENT_CAPTURE_MOTION":
 				configuration.Config.Capture.Motion = value
 				break
+			case "AGENT_CAPTURE_ONVIF_MOTION":
+				configuration.Config.Capture.ONVIFMotion = value
+				break
 			case "AGENT_CAPTURE_SNAPSHOTS":
 				configuration.Config.Capture.Snapshots = value
 				break
@@ -648,6 +651,16 @@ func applyAgentEnvVars(configuration *models.Configuration, prefix string, apply
 		}
 	}
 
+	// Motion sensitivity historically used 0 to mean "use the default". Preserve
+	// that behaviour for configurations created before this field became a
+	// pointer, and also recover invalid negative values. Only apply this to the
+	// effective configuration so missing values can still be inherited between
+	// the separate global and custom layers.
+	if applyDefaults && (configuration.Config.Capture.PixelChangeThreshold == nil || *configuration.Config.Capture.PixelChangeThreshold <= 0) {
+		defaultPixelChangeThreshold := 150
+		configuration.Config.Capture.PixelChangeThreshold = &defaultPixelChangeThreshold
+	}
+
 	// Signing is a new feature, so if empty we set default values. Only applied
 	// for the effective configuration (applyDefaults), not for the separate
 	// global/custom views.

machinery/src/config/main_test.go

@@ -0,0 +1,40 @@
+package config
+
+import (
+	"testing"
+
+	"github.com/kerberos-io/agent/machinery/src/models"
+)
+
+func TestApplyAgentEnvVarsPixelChangeThresholdDefault(t *testing.T) {
+	tests := []struct {
+		name      string
+		threshold *int
+		want      int
+	}{
+		{name: "missing", want: 150},
+		{name: "legacy zero", threshold: intPointer(0), want: 150},
+		{name: "negative", threshold: intPointer(-1), want: 150},
+		{name: "positive", threshold: intPointer(275), want: 275},
+	}
+
+	for _, test := range tests {
+		t.Run(test.name, func(t *testing.T) {
+			configuration := &models.Configuration{}
+			configuration.Config.Capture.PixelChangeThreshold = test.threshold
+
+			applyAgentEnvVars(configuration, "TEST_", true)
+
+			if configuration.Config.Capture.PixelChangeThreshold == nil {
+				t.Fatal("PixelChangeThreshold is nil after applying defaults")
+			}
+			if got := *configuration.Config.Capture.PixelChangeThreshold; got != test.want {
+				t.Fatalf("PixelChangeThreshold = %d, want %d", got, test.want)
+			}
+		})
+	}
+}
+
+func intPointer(value int) *int {
+	return &value
+}

machinery/src/models/config.go

@@ -75,6 +75,14 @@ type Capture struct {
 	Fragmented            string      `json:"fragmented,omitempty" bson:"fragmented,omitempty"`
 	FragmentedDuration    int64       `json:"fragmentedduration,omitempty" bson:"fragmentedduration,omitempty"`
 	PixelChangeThreshold  *int        `json:"pixelChangeThreshold,omitempty"`
+	// ONVIFMotion routes the camera's ONVIF motion events into the
+	// agent's motion-triggered recording pipeline. When "true" the
+	// agent opens an event/stream against the configured ONVIF
+	// endpoint and forwards Motion+Active events to HandleMotion.
+	// Requires Capture.IPCamera.ONVIFXAddr / ONVIFUsername /
+	// ONVIFPassword to be set. Default empty (disabled) keeps the
+	// existing pixel-diff motion detection as the only source.
+	ONVIFMotion string `json:"onvif_motion,omitempty" bson:"onvif_motion,omitempty"`
 }
 
 // IPCamera configuration, such as the RTSP url of the IPCamera and the FPS.

machinery/src/models/recording_upload_metadata.go

@@ -0,0 +1,37 @@
+package models
+
+import (
+	"path/filepath"
+	"strings"
+)
+
+const RecordingUploadMetadataExtension = ".metadata"
+
+// RecordingUploadMetadata is persisted in the upload queue marker associated
+// with a recording. New optional fields can be added without changing the queue
+// mechanism or breaking older agents.
+type RecordingUploadMetadata struct {
+	FileName  string  `json:"filename"`
+	DeviceKey string  `json:"device_key"`
+	Timestamp int64   `json:"timestamp"` // Unix milliseconds.
+	Duration  uint64  `json:"duration"`  // Milliseconds.
+	FPS       float64 `json:"fps,omitempty"`
+}
+
+// RecordingUploadMetadataFileName returns the queue marker name associated
+// with a recording, replacing the recording extension with .metadata.
+func RecordingUploadMetadataFileName(recordingFileName string) string {
+	name := filepath.Base(recordingFileName)
+	extension := filepath.Ext(name)
+	return strings.TrimSuffix(name, extension) + RecordingUploadMetadataExtension
+}
+
+// RecordingFileNameFromUploadMarker resolves a queue entry to its recording.
+// Markers created by older agents used the recording filename directly.
+func RecordingFileNameFromUploadMarker(markerFileName string) string {
+	name := filepath.Base(markerFileName)
+	if strings.HasSuffix(name, RecordingUploadMetadataExtension) {
+		return strings.TrimSuffix(name, RecordingUploadMetadataExtension) + ".mp4"
+	}
+	return name
+}

machinery/src/models/recording_upload_metadata_test.go

@@ -0,0 +1,15 @@
+package models
+
+import "testing"
+
+func TestRecordingUploadMetadataFileNames(t *testing.T) {
+	if got := RecordingUploadMetadataFileName("141245_x_x_.mp4"); got != "141245_x_x_.metadata" {
+		t.Fatalf("RecordingUploadMetadataFileName() = %q", got)
+	}
+	if got := RecordingFileNameFromUploadMarker("141245_x_x_.metadata"); got != "141245_x_x_.mp4" {
+		t.Fatalf("RecordingFileNameFromUploadMarker() = %q", got)
+	}
+	if got := RecordingFileNameFromUploadMarker("legacy.mp4"); got != "legacy.mp4" {
+		t.Fatalf("legacy RecordingFileNameFromUploadMarker() = %q", got)
+	}
+}

machinery/src/onvif/events.go

@@ -0,0 +1,225 @@
+package onvif
+
+import (
+	"context"
+	"errors"
+	"strconv"
+	"strings"
+	"time"
+
+	"github.com/kerberos-io/agent/machinery/src/log"
+	"github.com/kerberos-io/agent/machinery/src/models"
+	"github.com/kerberos-io/onvif/event/stream"
+)
+
+// The library handles in-stream reconnect; these guards cover the
+// initial-connect path the library cannot see.
+const (
+	initialBackoff = time.Second
+	maxBackoff     = 5 * time.Minute
+)
+
+// HandleONVIFEventStream opens an event/stream against the configured
+// ONVIF camera and routes Motion events into communication.HandleMotion.
+//
+// Behind the Capture.ONVIFMotion flag; the goroutine returns
+// immediately when not enabled. The flag is read once at start, so
+// toggling at runtime requires an agent restart. On transient
+// construction failure (camera not yet ready at boot, brief network
+// blip, credential reload) the goroutine retries with exponential
+// backoff. Exits when ctx is cancelled.
+func HandleONVIFEventStream(ctx context.Context, configuration *models.Configuration, communication *models.Communication) {
+	log.Log.Debug("onvif.HandleONVIFEventStream(): started")
+	defer log.Log.Debug("onvif.HandleONVIFEventStream(): finished")
+
+	if !isONVIFMotionEnabled(configuration.Config.Capture.ONVIFMotion) {
+		return
+	}
+	if configuration.Config.Capture.IPCamera.ONVIFXAddr == "" {
+		log.Log.Warning("onvif.HandleONVIFEventStream(): ONVIFMotion enabled but ONVIFXAddr is empty; nothing to do")
+		return
+	}
+
+	backoff := initialBackoff
+	for {
+		if ctx.Err() != nil {
+			return
+		}
+		recoverable := runStreamOnce(ctx, configuration, communication)
+		if !recoverable {
+			return
+		}
+		if !sleepCtx(ctx, backoff) {
+			return
+		}
+		backoff *= 2
+		if backoff > maxBackoff {
+			backoff = maxBackoff
+		}
+	}
+}
+
+// runStreamOnce returns true when the caller should retry construction
+// (transient failure), false on clean ctx-driven shutdown.
+func runStreamOnce(ctx context.Context, configuration *models.Configuration, communication *models.Communication) (retry bool) {
+	camera := configuration.Config.Capture.IPCamera
+
+	device, _, err := ConnectToOnvifDevice(&camera)
+	if err != nil {
+		log.Log.Error("onvif.HandleONVIFEventStream(): connect: " + err.Error())
+		return true
+	}
+
+	deviceID := resolveDeviceID(configuration.Name, camera.ONVIFXAddr)
+	s, err := stream.NewStream(ctx, device, stream.Options{DeviceID: deviceID})
+	if err != nil {
+		log.Log.Error("onvif.HandleONVIFEventStream(): open stream: " + err.Error())
+		return true
+	}
+	defer func() {
+		if err := s.Close(); err != nil {
+			log.Log.Debug("onvif.HandleONVIFEventStream(): close: " + err.Error())
+		}
+	}()
+
+	log.Log.Info("onvif.HandleONVIFEventStream(): consuming events for " + deviceID)
+
+	// recovering = the first successful event after an error streak
+	// logs a recovery line so on-call operators see the clear-of-
+	// condition for the ERROR they were paged on.
+	var recovering bool
+	for {
+		select {
+		case <-ctx.Done():
+			return false
+		case ev, ok := <-s.Events():
+			if !ok {
+				return false
+			}
+			if recovering {
+				log.Log.Info("onvif.HandleONVIFEventStream(): event stream recovered for " + deviceID)
+				recovering = false
+			}
+			dispatchEvent(ctx, ev, configuration, communication)
+		case e, ok := <-s.Errors():
+			if !ok {
+				return false
+			}
+			recovering = true
+			logStreamError(e)
+		}
+	}
+}
+
+// dispatchEvent routes motion-active events to HandleMotion.
+//
+// The ctx pre-check + ctx-in-select guards a shutdown race: the agent
+// closes HandleMotion shortly after cancelling ctx, and a stale event
+// reaching the send would otherwise panic on a closed channel.
+func dispatchEvent(ctx context.Context, ev stream.Event, configuration *models.Configuration, communication *models.Communication) {
+	topic := sanitiseTopic(ev.Topic)
+	if ev.Kind != stream.KindMotion {
+		log.Log.Debug("onvif.dispatchEvent(): non-motion event " + ev.Kind.String() + " topic=" + topic)
+		return
+	}
+	if ev.State != stream.StateActive {
+		return
+	}
+	if !isTransition(ev.Operation) {
+		log.Log.Debug("onvif.dispatchEvent(): " + ev.Operation.String() + " is not a transition, not a trigger: topic=" + topic)
+		return
+	}
+	if configuration.Config.Capture.Recording == "false" {
+		return
+	}
+	if ctx.Err() != nil {
+		return
+	}
+
+	dataToPass := models.MotionDataPartial{
+		Timestamp:       time.Now().Unix(),
+		NumberOfChanges: 0, // ONVIF does not quantify motion area.
+	}
+	select {
+	case <-ctx.Done():
+	case communication.HandleMotion <- dataToPass:
+		// Logged on the send, not before it: this line records that a
+		// recording started, so a dropped event must not leave one.
+		log.Log.Debug("onvif.dispatchEvent(): recording trigger " + ev.Kind.String() + " topic=" + topic)
+	default:
+		log.Log.Debug("onvif.dispatchEvent(): HandleMotion full, dropping ONVIF motion event")
+	}
+}
+
+// isTransition reports whether an operation represents a state change.
+// A camera replays every property's current state as Initialized on
+// each new subscription and announces removals as Deleted; neither is
+// motion starting. Absent (Unknown) counts — PropertyOperation is
+// optional per WS-Notification and many non-property events omit it.
+func isTransition(op stream.PropertyOperation) bool {
+	return op == stream.PropertyChanged || op == stream.PropertyUnknown
+}
+
+// maxLoggedTopic bounds a topic in the log; the wire imposes no limit,
+// and the reject path logs every event received.
+const maxLoggedTopic = 256
+
+// sanitiseTopic makes a camera-controlled topic safe to concatenate
+// into a log line. logrus's coloured text formatter writes the message
+// unquoted, so a raw newline would let a camera forge entries in the
+// log being used to diagnose it.
+func sanitiseTopic(topic string) string {
+	if len(topic) > maxLoggedTopic {
+		topic = topic[:maxLoggedTopic] + "…(truncated)"
+	}
+	quoted := strconv.Quote(topic)
+	return quoted[1 : len(quoted)-1]
+}
+
+// logStreamError logs at a level matching severity: recreate is loud
+// because it usually means the camera is offline; pull and renew are
+// debug because the library recovers from them automatically.
+func logStreamError(e error) {
+	var recreate stream.ErrRecreateFailed
+	var pull stream.ErrPullFailed
+	var renew stream.ErrRenewFailed
+	switch {
+	case errors.As(e, &recreate):
+		log.Log.Error("onvif.HandleONVIFEventStream(): subscription recreate failed (camera may be offline): " + recreate.Err.Error())
+	case errors.As(e, &renew):
+		log.Log.Debug("onvif.HandleONVIFEventStream(): renew failed (will recover via pull/recreate): " + renew.Err.Error())
+	case errors.As(e, &pull):
+		log.Log.Debug("onvif.HandleONVIFEventStream(): pull failed (will retry): " + pull.Err.Error())
+	default:
+		log.Log.Info("onvif.HandleONVIFEventStream(): stream error: " + e.Error())
+	}
+}
+
+func isONVIFMotionEnabled(v string) bool {
+	return strings.EqualFold(strings.TrimSpace(v), "true")
+}
+
+// resolveDeviceID falls back from operator-supplied name to ONVIF
+// endpoint to a constant placeholder so log lines always have
+// something to grep.
+func resolveDeviceID(configName, xaddr string) string {
+	if n := strings.TrimSpace(configName); n != "" {
+		return n
+	}
+	if x := strings.TrimSpace(xaddr); x != "" {
+		return x
+	}
+	return "unknown"
+}
+
+// sleepCtx returns false if ctx was cancelled, true if d elapsed.
+func sleepCtx(ctx context.Context, d time.Duration) bool {
+	t := time.NewTimer(d)
+	defer t.Stop()
+	select {
+	case <-ctx.Done():
+		return false
+	case <-t.C:
+		return true
+	}
+}

machinery/src/onvif/events_test.go

@@ -0,0 +1,375 @@
+package onvif
+
+import (
+	"bytes"
+	"context"
+	"strings"
+	"testing"
+	"time"
+
+	"github.com/kerberos-io/agent/machinery/src/models"
+	"github.com/kerberos-io/onvif/event/stream"
+	"github.com/sirupsen/logrus"
+	"github.com/stretchr/testify/assert"
+	"github.com/stretchr/testify/require"
+)
+
+func makeConfig(recording, onvifMotion, name string) *models.Configuration {
+	return &models.Configuration{
+		Name: name,
+		Config: models.Config{
+			Capture: models.Capture{
+				Recording:   recording,
+				ONVIFMotion: onvifMotion,
+			},
+		},
+	}
+}
+
+func makeCommunication(buffer int) *models.Communication {
+	return &models.Communication{
+		HandleMotion: make(chan models.MotionDataPartial, buffer),
+	}
+}
+
+// --- dispatchEvent ---------------------------------------------------
+
+func TestDispatchEvent_MotionActive_SendsToHandleMotion(t *testing.T) {
+	cfg := makeConfig("true", "true", "cam-1")
+	comm := makeCommunication(1)
+	ev := stream.Event{Kind: stream.KindMotion, State: stream.StateActive}
+
+	ctx, cancel := context.WithCancel(context.Background())
+	defer cancel()
+	dispatchEvent(ctx, ev, cfg, comm)
+
+	select {
+	case m := <-comm.HandleMotion:
+		assert.NotZero(t, m.Timestamp)
+	case <-time.After(time.Second):
+		t.Fatal("expected motion data on HandleMotion")
+	}
+}
+
+func TestDispatchEvent_MotionInactive_DoesNotSend(t *testing.T) {
+	cfg := makeConfig("true", "true", "cam-1")
+	comm := makeCommunication(1)
+	ev := stream.Event{Kind: stream.KindMotion, State: stream.StateInactive}
+
+	ctx, cancel := context.WithCancel(context.Background())
+	defer cancel()
+	dispatchEvent(ctx, ev, cfg, comm)
+
+	select {
+	case <-comm.HandleMotion:
+		t.Fatal("inactive motion must not reach HandleMotion (motion-stop is a follow-up)")
+	case <-time.After(100 * time.Millisecond):
+	}
+}
+
+func TestDispatchEvent_NonMotionKindIgnored(t *testing.T) {
+	cfg := makeConfig("true", "true", "cam-1")
+	comm := makeCommunication(1)
+	ev := stream.Event{Kind: stream.KindDigitalInput, State: stream.StateActive}
+
+	ctx, cancel := context.WithCancel(context.Background())
+	defer cancel()
+	dispatchEvent(ctx, ev, cfg, comm)
+
+	select {
+	case <-comm.HandleMotion:
+		t.Fatal("non-motion kinds must not reach HandleMotion")
+	case <-time.After(100 * time.Millisecond):
+	}
+}
+
+// captureDebugLog redirects logrus to a buffer at debug level for the
+// duration of a test and returns what was written. It mutates package
+// globals, so callers must not run in parallel.
+func captureDebugLog(t *testing.T) *bytes.Buffer {
+	t.Helper()
+	var buf bytes.Buffer
+	prevOut, prevLevel := logrus.StandardLogger().Out, logrus.GetLevel()
+	logrus.SetOutput(&buf)
+	logrus.SetLevel(logrus.DebugLevel)
+	t.Cleanup(func() {
+		logrus.SetOutput(prevOut)
+		logrus.SetLevel(prevLevel)
+	})
+	return &buf
+}
+
+// TestDispatchEvent_LogsTheTriggeringTopic — a dispatched event is what
+// actually starts a recording, so its topic is the one an operator needs
+// when a camera records for the wrong reason (or the right reason and
+// nobody can prove which). Rejected events were already logged; without
+// this the triggering topic is only knowable by elimination.
+func TestDispatchEvent_LogsTheTriggeringTopic(t *testing.T) {
+	buf := captureDebugLog(t)
+
+	cfg := makeConfig("true", "true", "cam-1")
+	comm := makeCommunication(1)
+	ev := stream.Event{
+		Kind:  stream.KindMotion,
+		State: stream.StateActive,
+		Topic: "tns1:RuleEngine/tnsaxis:VMD3/vmd3_video_1",
+	}
+
+	ctx, cancel := context.WithCancel(context.Background())
+	defer cancel()
+	dispatchEvent(ctx, ev, cfg, comm)
+
+	assert.Contains(t, buf.String(), "tns1:RuleEngine/tnsaxis:VMD3/vmd3_video_1",
+		"the dispatched event's topic must appear in the log")
+	assert.Contains(t, buf.String(), "Motion",
+		"the dispatched event's Kind must appear in the log")
+}
+
+// TestDispatchEvent_PropertyOperation — a camera replays the current
+// state of every property topic as Initialized whenever a pull-point
+// subscription is created. If that counts as a trigger, every
+// reconnect restarts a recording for any motion property that happens
+// to be active, and a flapping subscription manufactures motion out of
+// nothing. Only reject Initialized specifically: PropertyOperation is
+// optional per WS-Notification and absent on many non-property events,
+// which decode reports as PropertyUnknown.
+func TestDispatchEvent_PropertyOperation(t *testing.T) {
+	tests := []struct {
+		name     string
+		op       stream.PropertyOperation
+		wantSend bool
+	}{
+		{"changed is a real transition", stream.PropertyChanged, true},
+		{"absent attribute still counts", stream.PropertyUnknown, true},
+		{"initialized is a subscription state replay", stream.PropertyInitialized, false},
+	}
+
+	for _, tt := range tests {
+		t.Run(tt.name, func(t *testing.T) {
+			cfg := makeConfig("true", "true", "cam-1")
+			comm := makeCommunication(1)
+			ev := stream.Event{
+				Kind:      stream.KindMotion,
+				State:     stream.StateActive,
+				Operation: tt.op,
+				Topic:     "tns1:RuleEngine/tnsaxis:VMD3/vmd3_video_1",
+			}
+
+			ctx, cancel := context.WithCancel(context.Background())
+			defer cancel()
+			dispatchEvent(ctx, ev, cfg, comm)
+
+			select {
+			case <-comm.HandleMotion:
+				if !tt.wantSend {
+					t.Fatalf("%v must not trigger a recording", tt.op)
+				}
+			case <-time.After(100 * time.Millisecond):
+				if tt.wantSend {
+					t.Fatalf("%v must trigger a recording", tt.op)
+				}
+			}
+		})
+	}
+}
+
+func TestDispatchEvent_RecordingDisabled_DoesNotSend(t *testing.T) {
+	cfg := makeConfig("false", "true", "cam-1")
+	comm := makeCommunication(1)
+	ev := stream.Event{Kind: stream.KindMotion, State: stream.StateActive}
+
+	ctx, cancel := context.WithCancel(context.Background())
+	defer cancel()
+	dispatchEvent(ctx, ev, cfg, comm)
+
+	select {
+	case <-comm.HandleMotion:
+		t.Fatal("Recording=false must gate the send (matches computervision behaviour)")
+	case <-time.After(100 * time.Millisecond):
+	}
+}
+
+func TestDispatchEvent_HandleMotionFull_DropsRatherThanBlocks(t *testing.T) {
+	cfg := makeConfig("true", "true", "cam-1")
+	// Pre-fill the buffer so the next send would block.
+	comm := &models.Communication{HandleMotion: make(chan models.MotionDataPartial, 1)}
+	comm.HandleMotion <- models.MotionDataPartial{}
+	ev := stream.Event{Kind: stream.KindMotion, State: stream.StateActive}
+
+	ctx, cancel := context.WithCancel(context.Background())
+	defer cancel()
+
+	done := make(chan struct{})
+	go func() {
+		dispatchEvent(ctx, ev, cfg, comm)
+		close(done)
+	}()
+
+	select {
+	case <-done:
+	case <-time.After(time.Second):
+		t.Fatal("dispatchEvent must drop when HandleMotion is full, not block")
+	}
+}
+
+func TestDispatchEvent_CtxCancelledAndHandleMotionClosed_DoesNotPanic(t *testing.T) {
+	// Regression for the shutdown race: between cancel() and
+	// close(HandleMotion) the agent leaves a 3s window. If dispatchEvent
+	// runs in that window AFTER the channel is closed, a non-protected
+	// send would panic. The ctx pre-check must short-circuit before the
+	// send is attempted.
+	cfg := makeConfig("true", "true", "cam-1")
+	comm := &models.Communication{HandleMotion: make(chan models.MotionDataPartial, 1)}
+	close(comm.HandleMotion)
+	ev := stream.Event{Kind: stream.KindMotion, State: stream.StateActive}
+
+	ctx, cancel := context.WithCancel(context.Background())
+	cancel() // already cancelled, matching the shutdown sequence
+
+	assert.NotPanics(t, func() {
+		dispatchEvent(ctx, ev, cfg, comm)
+	})
+}
+
+// --- isONVIFMotionEnabled --------------------------------------------
+
+func TestIsONVIFMotionEnabled_CaseAndWhitespace(t *testing.T) {
+	tests := []struct {
+		in   string
+		want bool
+	}{
+		{"true", true},
+		{"True", true},
+		{"TRUE", true},
+		{" true", true},
+		{"true ", true},
+		{"  true  ", true},
+		{"false", false},
+		{"False", false},
+		{"", false},
+		{"yes", false},
+		{"1", false},
+	}
+	for _, tc := range tests {
+		t.Run(tc.in, func(t *testing.T) {
+			assert.Equal(t, tc.want, isONVIFMotionEnabled(tc.in))
+		})
+	}
+}
+
+// --- resolveDeviceID -------------------------------------------------
+
+func TestResolveDeviceID_FallbackChain(t *testing.T) {
+	tests := []struct {
+		name    string
+		cfgName string
+		xaddr   string
+		want    string
+	}{
+		{"name_set", "front-door", "192.168.1.10", "front-door"},
+		{"name_empty_xaddr_set", "", "192.168.1.10", "192.168.1.10"},
+		{"name_whitespace_only_xaddr_set", "  ", "192.168.1.10", "192.168.1.10"},
+		{"both_empty", "", "", "unknown"},
+		{"name_with_trailing_whitespace", "cam-2  ", "192.168.1.10", "cam-2"},
+	}
+	for _, tc := range tests {
+		t.Run(tc.name, func(t *testing.T) {
+			assert.Equal(t, tc.want, resolveDeviceID(tc.cfgName, tc.xaddr))
+		})
+	}
+}
+
+// TestDispatchEvent_OnlyRealTransitionsTrigger — a camera replays every
+// property topic's state on each new subscription (Initialized) and
+// announces removals (Deleted). Neither is a motion transition, and a
+// flapping pull-point would otherwise manufacture recordings out of
+// replayed state. PropertyOperation is optional per WS-Notification, so
+// absent (Unknown) still counts — many non-property events omit it.
+func TestDispatchEvent_OnlyRealTransitionsTrigger(t *testing.T) {
+	tests := []struct {
+		op       stream.PropertyOperation
+		wantSend bool
+	}{
+		{stream.PropertyChanged, true},
+		{stream.PropertyUnknown, true},
+		{stream.PropertyInitialized, false},
+		{stream.PropertyDeleted, false},
+	}
+
+	for _, tt := range tests {
+		t.Run(tt.op.String(), func(t *testing.T) {
+			cfg := makeConfig("true", "true", "cam-1")
+			comm := makeCommunication(1)
+			ev := stream.Event{Kind: stream.KindMotion, State: stream.StateActive, Operation: tt.op}
+
+			ctx, cancel := context.WithCancel(context.Background())
+			defer cancel()
+			dispatchEvent(ctx, ev, cfg, comm)
+
+			if tt.wantSend {
+				require.Len(t, comm.HandleMotion, 1, "%v must trigger a recording", tt.op)
+				return
+			}
+			require.Empty(t, comm.HandleMotion, "%v must not trigger a recording", tt.op)
+		})
+	}
+}
+
+// TestSanitiseTopic — ev.Topic is camera-controlled and reaches the log
+// unmodified. logrus's coloured text formatter (the default) writes the
+// message without quoting, so an embedded newline forges whole log
+// lines: a compromised camera can fabricate ERROR entries or spoof
+// another device's id, in the logs an operator is reading to diagnose
+// that very camera. Length is also unbounded on the wire, and the
+// reject path logs every event, so an oversized topic is a cheap way to
+// evict a container's whole retained history.
+func TestSanitiseTopic(t *testing.T) {
+	tests := []struct {
+		name string
+		in   string
+		want string
+	}{
+		{"ordinary topic passes through", "tns1:RuleEngine/tnsaxis:VMD3/vmd3_video_1", "tns1:RuleEngine/tnsaxis:VMD3/vmd3_video_1"},
+		{"newline cannot forge a line", "a\nERRO[fake] boom", `a\nERRO[fake] boom`},
+		{"carriage return", "a\rb", `a\rb`},
+		{"tab", "a\tb", `a\tb`},
+		{"NUL", "a\x00b", `a\x00b`},
+		{"empty", "", ""},
+	}
+
+	for _, tt := range tests {
+		t.Run(tt.name, func(t *testing.T) {
+			got := sanitiseTopic(tt.in)
+			assert.Equal(t, tt.want, got)
+			assert.NotContains(t, got, "\n", "no raw newline may survive")
+			assert.NotContains(t, got, "\r", "no raw carriage return may survive")
+		})
+	}
+}
+
+func TestSanitiseTopic_Truncates(t *testing.T) {
+	got := sanitiseTopic(strings.Repeat("x", maxLoggedTopic*2))
+	assert.LessOrEqual(t, len(got), maxLoggedTopic+len("…(truncated)"))
+	assert.Contains(t, got, "truncated")
+}
+
+// TestDispatchEvent_LogsTriggerOnlyWhenSent — the trigger line is the
+// record that a recording started. Logging it before the send means a
+// dropped event (full channel, or shutdown) leaves a line claiming a
+// recording that never began.
+func TestDispatchEvent_LogsTriggerOnlyWhenSent(t *testing.T) {
+	buf := captureDebugLog(t)
+
+	cfg := makeConfig("true", "true", "cam-1")
+	comm := &models.Communication{HandleMotion: make(chan models.MotionDataPartial, 1)}
+	comm.HandleMotion <- models.MotionDataPartial{} // full
+	ev := stream.Event{Kind: stream.KindMotion, State: stream.StateActive, Topic: "tns1:VideoSource/MotionAlarm"}
+
+	ctx, cancel := context.WithCancel(context.Background())
+	defer cancel()
+	dispatchEvent(ctx, ev, cfg, comm)
+
+	assert.NotContains(t, buf.String(), "recording trigger",
+		"a dropped event must not be logged as a trigger")
+	assert.Contains(t, buf.String(), "dropping", "the drop itself must still be logged")
+}

machinery/src/routers/mqtt/main.go

@@ -332,7 +332,7 @@ func MQTTListenerHandler(mqttClient mqtt.Client, hubKey string, configDirectory
 				case "record":
 					go HandleRecording(mqttClient, hubKey, payload, configuration, communication)
 				case "get-audio-backchannel":
-					go HandleAudio(mqttClient, hubKey, payload, configuration, communication)
+					HandleAudio(mqttClient, hubKey, payload, configuration, communication)
 				case "get-ptz-position":
 					go HandleGetPTZPosition(mqttClient, hubKey, payload, configuration, communication)
 				case "update-ptz-position":
@@ -442,10 +442,31 @@ func HandleAudio(mqttClient mqtt.Client, hubKey string, payload models.Payload,
 			Timestamp: audioPayload.Timestamp,
 			Data:      audioPayload.Data,
 		}
-		communication.HandleAudio <- audioDataPartial
+		if enqueueLatestAudio(communication.HandleAudio, audioDataPartial) {
+			log.Log.Debug("routers.mqtt.main.HandleAudio(): dropped stale audio because the backchannel queue was full")
+		}
 	}
 }
 
+func enqueueLatestAudio(audioChannel chan models.AudioDataPartial, audio models.AudioDataPartial) bool {
+	select {
+	case audioChannel <- audio:
+		return false
+	default:
+	}
+
+	select {
+	case <-audioChannel:
+	default:
+	}
+
+	select {
+	case audioChannel <- audio:
+	default:
+	}
+	return true
+}
+
 func HandleGetPTZPosition(mqttClient mqtt.Client, hubKey string, payload models.Payload, configuration *models.Configuration, communication *models.Communication) {
 	value := payload.Value
 

machinery/src/routers/mqtt/main_test.go

@@ -0,0 +1,39 @@
+package mqtt
+
+import (
+	"testing"
+	"time"
+
+	"github.com/kerberos-io/agent/machinery/src/models"
+)
+
+func TestEnqueueLatestAudioReplacesOldestFrameWhenFull(t *testing.T) {
+	audioChannel := make(chan models.AudioDataPartial, 2)
+	audioChannel <- models.AudioDataPartial{Timestamp: 1}
+	audioChannel <- models.AudioDataPartial{Timestamp: 2}
+
+	dropped := enqueueLatestAudio(audioChannel, models.AudioDataPartial{Timestamp: 3})
+	if !dropped {
+		t.Fatal("enqueueLatestAudio() dropped = false, want true")
+	}
+
+	first := <-audioChannel
+	second := <-audioChannel
+	if first.Timestamp != 2 || second.Timestamp != 3 {
+		t.Fatalf("queued timestamps = (%d, %d), want (2, 3)", first.Timestamp, second.Timestamp)
+	}
+}
+
+func TestEnqueueLatestAudioDoesNotBlockNilChannel(t *testing.T) {
+	done := make(chan struct{})
+	go func() {
+		enqueueLatestAudio(nil, models.AudioDataPartial{Timestamp: 1})
+		close(done)
+	}()
+
+	select {
+	case <-done:
+	case <-time.After(100 * time.Millisecond):
+		t.Fatal("enqueueLatestAudio() blocked on a nil channel")
+	}
+}

machinery/src/video/mp4.go

@@ -283,6 +283,8 @@ func (mp4 *MP4) flushPendingVideoSample(nextPTS uint64) bool {
 	err := mp4.MultiTrackFragment.AddFullSampleToTrack(*mp4.VideoFullSample, uint32(mp4.VideoTrack))
 	if err != nil {
 		log.Log.Error("mp4.flushPendingVideoSample(): error adding sample: " + err.Error())
+	} else {
+		mp4.SampleCount++
 	}
 	if isKF {
 		mp4.TotalKeyframesWritten++
@@ -296,6 +298,15 @@ func (mp4 *MP4) flushPendingVideoSample(nextPTS uint64) bool {
 	return true
 }
 
+// AverageFPS returns the average frame rate of the video samples actually
+// committed to this recording.
+func (mp4 *MP4) AverageFPS() float64 {
+	if mp4.SampleCount == 0 || mp4.VideoTotalDuration == 0 {
+		return 0
+	}
+	return float64(mp4.SampleCount) * 1000 / float64(mp4.VideoTotalDuration)
+}
+
 // AddSampleToTrack appends a sample to the given track.
 //
 // For video, pts is the decode timestamp (DTS, in milliseconds) and

machinery/src/video/mp4_duration_test.go

@@ -2,6 +2,7 @@ package video
 
 import (
 	"fmt"
+	"math"
 	"os"
 	"testing"
 
@@ -173,4 +174,10 @@ func TestMP4Duration(t *testing.T) {
 		t.Errorf("MISMATCH: mdhd.Duration should be 0 for fragmented MP4, got %d",
 			parsedFile.Moov.Traks[0].Mdia.Mdhd.Duration)
 	}
+	if mp4Video.SampleCount != sampleCount {
+		t.Errorf("SampleCount = %d, finalized MP4 contains %d video samples", mp4Video.SampleCount, sampleCount)
+	}
+	if fps := mp4Video.AverageFPS(); math.Abs(fps-25) > 0.001 {
+		t.Errorf("AverageFPS() = %.3f, want 25", fps)
+	}
 }

machinery/verify-moq-devcontainer.sh

@@ -0,0 +1,23 @@
+#!/usr/bin/env bash
+set -euo pipefail
+
+required_glibc="2.38"
+current_glibc="$(getconf GNU_LIBC_VERSION | awk '{print $2}')"
+
+if ! dpkg --compare-versions "$current_glibc" ge "$required_glibc"; then
+    echo "MoQ requires glibc ${required_glibc}+; this container has ${current_glibc}." >&2
+    echo "Rebuild the devcontainer with the Trixie base, then run this check again." >&2
+    exit 1
+fi
+
+echo "==> Testing MoQ packages (glibc ${current_glibc})"
+GOWORK=off go test -tags moq ./src/cloud/livemoq ./src/cloud
+
+binary="${TMPDIR:-/tmp}/agent-moq"
+trap 'rm -f "$binary"' EXIT
+
+echo "==> Linking the MoQ Agent"
+GOWORK=off go build -tags moq -o "$binary" ./main.go
+
+echo "==> Running the linked binary"
+"$binary" -action version