AWSTemplateFormatVersion: "2010-09-09" Description: >- ALOS for AWS (AWS Marketplace): the ALOS app in your AWS account (portable mode, docs/aws.md) on ECS Fargate, its SQLite database and files on encrypted EFS, HTTPS at a load balancer, generated secrets in Secrets Manager, and a yearly licence checked with AWS License Manager. Optional: ALOS Mediate (its own Marketplace listing) as an advisory pre-check on agent actions, from your model package subscription, reachable only with the app's role. Identity, governance metadata and the agent registry stay on the ALOS platform (ALOS_API_URL); records, files, the audit trail and Mediate's decisions stay in this account. One copy only: no high availability yet. # ALOS for AWS 1.0.0, as linked from the AWS Marketplace listing (product prod-vgjwh4i6thkyw). Generated from # deploy/aws/alos-for-aws.yaml with its two placeholders filled in: ImageUri is the 1.0.0 image by digest in AWS # Marketplace's registry, and MarketplaceProductSku is the listing's product id (to be confirmed by the test # subscription: AWS Marketplace documents License Manager's ProductSKU as the product's "Product ID"; the buyer # account's ListReceivedLicenses shows the value the licence carries). Change the source, then regenerate this file. # # Source: TEKIMAX/alos-v4 deploy/aws/alos-for-aws.yaml. The app's network, storage, load balancer, service and alarms # are the same blocks as deploy/aws/alos-app.yaml (scripts/aws-templates.test.mjs keeps them identical); Mediate's # are deploy/aws/mediate/gateway/mediate-gateway.yaml's, in IAM mode only. One file, no nested stacks, so the # Marketplace listing ships a single template. Parameters: # ---------------------------------------------------------------- where ALOS answers PublicUrl: Type: String Description: The address people reach ALOS at, e.g. https://alos.example.com. Must match the certificate's name. For a first look without a domain, leave CertificateArn empty and set this to http:// after the stack is up (update the stack); sign-in stays off over http. AllowedPattern: "^https?://[^/]+$" CertificateArn: Type: String Default: "" Description: ACM certificate for PublicUrl's host, in this region. Empty = HTTP only (evaluation; sign-in cannot work). TlsMode: Type: String Default: standard AllowedValues: [standard, strict, fips] Description: >- HTTPS policy (only with a certificate). All three use post-quantum key exchange (hybrid ML-KEM) with every client that offers it. standard = TLS 1.3, or TLS 1.2 with restricted ciphers for older clients (AWS's recommended policy). strict = TLS 1.3 only; old clients and some inspecting proxies cannot connect. fips = standard with FIPS 140-3 validated cryptography. # ---------------------------------------------------------------- the app ImageUri: Type: String Default: 709825985650.dkr.ecr.us-east-1.amazonaws.com/tekimax/tekimax-alos-for-aws@sha256:a585c3084b9bef6ee8651d8856353c57b2165fbb940e1832e386263fdc64a9ac Description: Leave as is. The ALOS app image in AWS Marketplace's registry, set by TEKIMAX for this listing version. Cpu: Type: Number Default: 1024 AllowedValues: [512, 1024, 2048, 4096] MemoryMiB: Type: Number Default: 2048 AllowedValues: [1024, 2048, 4096, 8192] SettingsSecretArn: Type: String Default: "" Description: >- Leave empty on a first install: the stack generates the cookie password, redaction key and audit seal key, each in its own secret, kept if the stack is deleted. To reinstall over a kept data volume, give a secret (JSON) with WORKOS_COOKIE_PASSWORD, REDACTION_KEY and AUDIT_SEAL_HMAC_KEY holding the old values, or old redactions and checkpoint signatures cannot be read. AuditSignAlg: Type: String Default: hmac-sha256 AllowedValues: [none, hmac-sha256] Description: Audit checkpoint signatures. hmac-sha256 signs each checkpoint with this deployment's own generated key, so a checkpoint holds against the database alone; each is still timestamped by the external authority. none signs nothing (the seal still runs). # ---------------------------------------------------------------- ALOS Mediate (optional) MediateModelPackageArn: Type: String Default: "" Description: >- Optional. Your ALOS Mediate model package ARN for this Region, from its own AWS Marketplace subscription. Empty = no Mediate. Set, the stack runs Mediate here (a SageMaker endpoint and a gateway that only this app's role may call) and ALOS asks it before an agent makes a change on its own: anything but "proceed" waits for a person. MediateInstanceType: Type: String Default: ml.g6.xlarge AllowedValues: [ml.g6.xlarge, ml.g6.2xlarge, ml.g6.4xlarge, ml.g6e.xlarge, ml.g6e.2xlarge, ml.m5.xlarge, ml.m5.2xlarge, ml.m5.4xlarge, ml.c5.2xlarge, ml.c5.4xlarge] Description: GPU (ml.g6*) to check agent actions live (under 100 ms each through the API); CPU (ml.m5/ml.c5) for testing (seconds each). MediateMinInstances: Type: Number Default: 1 MinValue: 1 MaxValue: 20 Description: Instances always running. Use 2 or more in production. Your SageMaker quota for the instance type (" for endpoint usage") must allow MediateMaxInstances. MediateMaxInstances: Type: Number Default: 4 MinValue: 1 MaxValue: 40 Description: Upper limit for automatic scale-out under load. One ml.g6.xlarge handled about 50 decisions per second in TEKIMAX's load test. MediateThrottleRatePerSecond: Type: Number Default: 100 MinValue: 1 Description: Steady requests per second the Mediate API accepts; above it callers get HTTP 429. MediateLatencyAlarmMs: Type: Number Default: 2000 MinValue: 100 Description: Alert when the slowest 10% of Mediate decisions take longer than this (milliseconds, inside the model) for 5 minutes. MediateCodeBucket: Type: String Default: tekimax-mediate-downloads Description: Leave as is. TEKIMAX's bucket with the Mediate gateway code. It is in us-east-1, and Lambda reads code from its own Region, so Mediate deploys in us-east-1 unless TEKIMAX gives you a bucket for your Region. MediateCodeKey: Type: String Default: mediate-gateway/1.2.0/mediate-gateway.zip Description: Leave as is. # ---------------------------------------------------------------- licence and platform (leave as is) MarketplaceProductSku: Type: String Default: prod-vgjwh4i6thkyw Description: Leave as is. The listing's product id, which the app checks the licence against with AWS License Manager. Empty = no licence check. MarketplaceEntitlement: Type: String Default: organization Description: Leave as is. The listing's pricing dimension (its API name). MarketplaceEntitlementUnit: Type: String Default: Count AllowedValues: [Count, None] Description: Leave as is. Count for a quantity dimension (one organization), None for a tier. AlosApiUrl: Type: String Default: https://alos.tekimax.ai/api/v1 Description: The ALOS platform API this deployment is governed by. WorkosClientId: Type: String Default: client_01KTD8WBQ93V72NFBGW4B7BZBR Description: The sign-in client (public). The deployment's https callback must be registered with it (send TEKIMAX the address). WorkosIssuerHosts: Type: String Default: api.tekimax.ai WorkosAuthkitHost: Type: String Default: signin.tekimax.ai # ---------------------------------------------------------------- operations AlarmEmail: Type: String Default: "" Description: Optional. Who is told when ALOS (or Mediate) is unhealthy, answering errors, or failing closed. LogRetentionDays: Type: Number Default: 365 AllowedValues: [30, 90, 180, 365, 731, 1827, 3653] VpcCidr: Type: String Default: 10.42.0.0/16 Metadata: AWS::CloudFormation::Interface: ParameterGroups: - Label: {default: Where ALOS answers} Parameters: [PublicUrl, CertificateArn, TlsMode] - Label: {default: ALOS app} Parameters: [Cpu, MemoryMiB, AuditSignAlg, SettingsSecretArn] - Label: {default: "ALOS Mediate (optional, needs its own subscription)"} Parameters: [MediateModelPackageArn, MediateInstanceType, MediateMinInstances, MediateMaxInstances, MediateThrottleRatePerSecond, MediateLatencyAlarmMs] - Label: {default: Operations} Parameters: [AlarmEmail, LogRetentionDays, VpcCidr] - Label: {default: Listing and platform (leave as is)} Parameters: [ImageUri, MarketplaceProductSku, MarketplaceEntitlement, MarketplaceEntitlementUnit, AlosApiUrl, WorkosClientId, WorkosIssuerHosts, WorkosAuthkitHost, MediateCodeBucket, MediateCodeKey] ParameterLabels: PublicUrl: {default: ALOS address} CertificateArn: {default: Certificate (ACM ARN)} TlsMode: {default: HTTPS policy} Cpu: {default: CPU units} MemoryMiB: {default: Memory (MiB)} AuditSignAlg: {default: Audit checkpoint signature} SettingsSecretArn: {default: Existing settings secret (reinstall only)} MediateModelPackageArn: {default: Mediate model package ARN} MediateInstanceType: {default: Mediate instance type} MediateMinInstances: {default: Mediate minimum instances} MediateMaxInstances: {default: Mediate maximum instances} MediateThrottleRatePerSecond: {default: Mediate requests per second} MediateLatencyAlarmMs: {default: Mediate slow decision alarm (ms)} AlarmEmail: {default: Alarm email} LogRetentionDays: {default: Log retention (days)} VpcCidr: {default: Network range (CIDR)} ImageUri: {default: App image} MarketplaceProductSku: {default: Listing product id} MarketplaceEntitlement: {default: Listing pricing dimension} MarketplaceEntitlementUnit: {default: Dimension unit} AlosApiUrl: {default: ALOS platform API} WorkosClientId: {default: Sign-in client} WorkosIssuerHosts: {default: Sign-in issuer hosts} WorkosAuthkitHost: {default: Sign-in page host} MediateCodeBucket: {default: Mediate code bucket} MediateCodeKey: {default: Mediate code key} Rules: ListingFilledIn: Assertions: - Assert: !Not [!Equals [!Ref ImageUri, MARKETPLACE_IMAGE_URI]] AssertDescription: The app image is still the placeholder. Launch ALOS from its AWS Marketplace listing, which fills it in. - Assert: !Not [!Equals [!Ref MarketplaceProductSku, MARKETPLACE_PRODUCT_SKU]] AssertDescription: The listing's product id is still the placeholder. Launch ALOS from its AWS Marketplace listing, which fills it in. Mappings: # Load balancer security policies (docs.aws.amazon.com/elasticloadbalancing/latest/application/describe-ssl-policies.html). TlsPolicy: standard: {Name: ELBSecurityPolicy-TLS13-1-2-Res-PQ-2025-09} strict: {Name: ELBSecurityPolicy-TLS13-1-3-PQ-2025-09} fips: {Name: ELBSecurityPolicy-TLS13-1-2-Res-FIPS-PQ-2025-09} # Mediate scales out when an instance averages this many decisions per minute: about 60% of what one instance # sustained in TEKIMAX's load test (deploy/aws/mediate/gateway/mediate-gateway.yaml). MediateScaleTarget: ml.g6.xlarge: {PerMinute: 1800} ml.g6.2xlarge: {PerMinute: 1800} ml.g6.4xlarge: {PerMinute: 1800} ml.g6e.xlarge: {PerMinute: 2400} ml.g6e.2xlarge: {PerMinute: 2400} ml.m5.xlarge: {PerMinute: 25} ml.m5.2xlarge: {PerMinute: 50} ml.m5.4xlarge: {PerMinute: 90} ml.c5.2xlarge: {PerMinute: 50} ml.c5.4xlarge: {PerMinute: 90} Conditions: HasCertificate: !Not [!Equals [!Ref CertificateArn, ""]] NoCertificate: !Equals [!Ref CertificateArn, ""] HasAlarmEmail: !Not [!Equals [!Ref AlarmEmail, ""]] GenerateSettings: !Equals [!Ref SettingsSecretArn, ""] HasMediate: !Not [!Equals [!Ref MediateModelPackageArn, ""]] ChecksLicence: !Not [!Equals [!Ref MarketplaceProductSku, ""]] Resources: # ---------------------------------------------------------------- network: two public subnets, no NAT # The task needs outbound HTTPS (the ALOS API, sign-in, the image registry). A public IP on the task is the # cheapest way; its security group admits only the load balancer, so nothing reaches it directly. Vpc: Type: AWS::EC2::VPC Properties: CidrBlock: !Ref VpcCidr EnableDnsSupport: true EnableDnsHostnames: true Tags: [{Key: Name, Value: !Sub "${AWS::StackName}-alos"}] InternetGateway: Type: AWS::EC2::InternetGateway GatewayAttachment: Type: AWS::EC2::VPCGatewayAttachment Properties: {VpcId: !Ref Vpc, InternetGatewayId: !Ref InternetGateway} SubnetA: Type: AWS::EC2::Subnet Properties: VpcId: !Ref Vpc AvailabilityZone: !Select [0, !GetAZs ""] CidrBlock: !Select [0, !Cidr [!Ref VpcCidr, 4, 12]] MapPublicIpOnLaunch: false SubnetB: Type: AWS::EC2::Subnet Properties: VpcId: !Ref Vpc AvailabilityZone: !Select [1, !GetAZs ""] CidrBlock: !Select [1, !Cidr [!Ref VpcCidr, 4, 12]] MapPublicIpOnLaunch: false RouteTable: Type: AWS::EC2::RouteTable Properties: {VpcId: !Ref Vpc} DefaultRoute: Type: AWS::EC2::Route DependsOn: GatewayAttachment Properties: {RouteTableId: !Ref RouteTable, DestinationCidrBlock: 0.0.0.0/0, GatewayId: !Ref InternetGateway} SubnetARoutes: Type: AWS::EC2::SubnetRouteTableAssociation Properties: {SubnetId: !Ref SubnetA, RouteTableId: !Ref RouteTable} SubnetBRoutes: Type: AWS::EC2::SubnetRouteTableAssociation Properties: {SubnetId: !Ref SubnetB, RouteTableId: !Ref RouteTable} LoadBalancerSecurityGroup: Type: AWS::EC2::SecurityGroup Properties: GroupDescription: ALOS load balancer VpcId: !Ref Vpc SecurityGroupIngress: - {IpProtocol: tcp, FromPort: 443, ToPort: 443, CidrIp: 0.0.0.0/0} - {IpProtocol: tcp, FromPort: 80, ToPort: 80, CidrIp: 0.0.0.0/0} AppSecurityGroup: Type: AWS::EC2::SecurityGroup Properties: GroupDescription: ALOS app - reachable only from its load balancer VpcId: !Ref Vpc SecurityGroupIngress: - {IpProtocol: tcp, FromPort: 8080, ToPort: 8080, SourceSecurityGroupId: !Ref LoadBalancerSecurityGroup} StorageSecurityGroup: Type: AWS::EC2::SecurityGroup Properties: GroupDescription: ALOS data volume - reachable only from the app VpcId: !Ref Vpc SecurityGroupIngress: - {IpProtocol: tcp, FromPort: 2049, ToPort: 2049, SourceSecurityGroupId: !Ref AppSecurityGroup} # ---------------------------------------------------------------- data: encrypted EFS, backed up daily # Holds alos.sqlite (and its WAL files) and the blob folders. Retained if the stack is deleted: a customer's # records are never removed by an infrastructure change. DataVolume: Type: AWS::EFS::FileSystem DeletionPolicy: Retain UpdateReplacePolicy: Retain Properties: Encrypted: true PerformanceMode: generalPurpose ThroughputMode: elastic BackupPolicy: {Status: ENABLED} FileSystemTags: [{Key: Name, Value: !Sub "${AWS::StackName}-alos-data"}] DataMountA: Type: AWS::EFS::MountTarget Properties: {FileSystemId: !Ref DataVolume, SubnetId: !Ref SubnetA, SecurityGroups: [!Ref StorageSecurityGroup]} DataMountB: Type: AWS::EFS::MountTarget Properties: {FileSystemId: !Ref DataVolume, SubnetId: !Ref SubnetB, SecurityGroups: [!Ref StorageSecurityGroup]} # The image runs as `node` (uid 1000); the access point gives it /data as its own and nothing else. DataAccessPoint: Type: AWS::EFS::AccessPoint Properties: FileSystemId: !Ref DataVolume PosixUser: {Uid: "1000", Gid: "1000"} RootDirectory: Path: /alos CreationInfo: {OwnerUid: "1000", OwnerGid: "1000", Permissions: "700"} # ---------------------------------------------------------------- load balancer LoadBalancer: Type: AWS::ElasticLoadBalancingV2::LoadBalancer Properties: Scheme: internet-facing Type: application Subnets: [!Ref SubnetA, !Ref SubnetB] SecurityGroups: [!Ref LoadBalancerSecurityGroup] LoadBalancerAttributes: - {Key: idle_timeout.timeout_seconds, Value: "300"} - {Key: routing.http.drop_invalid_header_fields.enabled, Value: "true"} TargetGroup: Type: AWS::ElasticLoadBalancingV2::TargetGroup Properties: VpcId: !Ref Vpc TargetType: ip Protocol: HTTP Port: 8080 HealthCheckPath: /api/health HealthCheckIntervalSeconds: 15 HealthyThresholdCount: 2 UnhealthyThresholdCount: 4 TargetGroupAttributes: - {Key: deregistration_delay.timeout_seconds, Value: "20"} HttpsListener: Type: AWS::ElasticLoadBalancingV2::Listener Condition: HasCertificate Properties: LoadBalancerArn: !Ref LoadBalancer Protocol: HTTPS Port: 443 SslPolicy: !FindInMap [TlsPolicy, !Ref TlsMode, Name] Certificates: [{CertificateArn: !Ref CertificateArn}] DefaultActions: [{Type: forward, TargetGroupArn: !Ref TargetGroup}] HttpRedirectListener: Type: AWS::ElasticLoadBalancingV2::Listener Condition: HasCertificate Properties: LoadBalancerArn: !Ref LoadBalancer Protocol: HTTP Port: 80 DefaultActions: - Type: redirect RedirectConfig: {Protocol: HTTPS, Port: "443", StatusCode: HTTP_301} # Evaluation only: with no certificate the app is served over plain HTTP, and sign-in stays off. HttpListener: Type: AWS::ElasticLoadBalancingV2::Listener Condition: NoCertificate Properties: LoadBalancerArn: !Ref LoadBalancer Protocol: HTTP Port: 80 DefaultActions: [{Type: forward, TargetGroupArn: !Ref TargetGroup}] # ---------------------------------------------------------------- the app LogGroup: Type: AWS::Logs::LogGroup Properties: {RetentionInDays: !Ref LogRetentionDays} Cluster: Type: AWS::ECS::Cluster Properties: ClusterSettings: [{Name: containerInsights, Value: enabled}] # ---------------------------------------------------------------- settings: generated here, never typed in # Three secrets, one generated value each (CloudFormation generates one key per secret). Kept if the stack is # deleted, like the data volume: the redaction key reads old redactions and the seal key checks old checkpoints. CookieSecret: Type: AWS::SecretsManager::Secret Condition: GenerateSettings DeletionPolicy: Retain UpdateReplacePolicy: Retain Properties: Description: ALOS session cookie password (generated) GenerateSecretString: SecretStringTemplate: "{}" GenerateStringKey: WORKOS_COOKIE_PASSWORD PasswordLength: 48 ExcludePunctuation: true RedactionSecret: Type: AWS::SecretsManager::Secret Condition: GenerateSettings DeletionPolicy: Retain UpdateReplacePolicy: Retain Properties: Description: ALOS redaction key (generated; rotating it makes stored redactions unreadable) GenerateSecretString: SecretStringTemplate: "{}" GenerateStringKey: REDACTION_KEY PasswordLength: 48 ExcludePunctuation: true SealSecret: Type: AWS::SecretsManager::Secret Condition: GenerateSettings DeletionPolicy: Retain UpdateReplacePolicy: Retain Properties: Description: ALOS audit seal key (generated; rotating it makes earlier checkpoint signatures uncheckable) GenerateSecretString: SecretStringTemplate: "{}" GenerateStringKey: AUDIT_SEAL_HMAC_KEY PasswordLength: 48 ExcludePunctuation: true # Pulls the image, reads the settings secrets, writes logs. Nothing else. ExecutionRole: Type: AWS::IAM::Role Properties: AssumeRolePolicyDocument: Version: "2012-10-17" Statement: [{Effect: Allow, Principal: {Service: ecs-tasks.amazonaws.com}, Action: sts:AssumeRole}] ManagedPolicyArns: [!Sub "arn:${AWS::Partition}:iam::aws:policy/service-role/AmazonECSTaskExecutionRolePolicy"] Policies: - PolicyName: read-settings PolicyDocument: Version: "2012-10-17" Statement: - Effect: Allow Action: secretsmanager:GetSecretValue Resource: !If [GenerateSettings, [!Ref CookieSecret, !Ref RedactionSecret, !Ref SealSecret], [!Ref SettingsSecretArn]] # What the running app may do in AWS: mount its own data volume through its own access point; check out and # return its own licence (License Manager has no resource-level permissions for these calls); and, with # Mediate, call that one gateway and nothing else. TaskRole: Type: AWS::IAM::Role Properties: AssumeRolePolicyDocument: Version: "2012-10-17" Statement: [{Effect: Allow, Principal: {Service: ecs-tasks.amazonaws.com}, Action: sts:AssumeRole}] Policies: - PolicyName: own-data-volume PolicyDocument: Version: "2012-10-17" Statement: - Effect: Allow Action: [elasticfilesystem:ClientMount, elasticfilesystem:ClientWrite] Resource: !GetAtt DataVolume.Arn Condition: {StringEquals: {"elasticfilesystem:AccessPointArn": !GetAtt DataAccessPoint.Arn}} - !If - ChecksLicence - PolicyName: marketplace-licence PolicyDocument: Version: "2012-10-17" Statement: - Effect: Allow Action: [license-manager:CheckoutLicense, license-manager:CheckInLicense, license-manager:GetLicense, license-manager:ExtendLicenseConsumption, license-manager:ListReceivedLicenses] Resource: "*" - !Ref AWS::NoValue - !If - HasMediate - PolicyName: ask-mediate PolicyDocument: Version: "2012-10-17" Statement: - Effect: Allow Action: execute-api:Invoke Resource: !Sub "arn:${AWS::Partition}:execute-api:${AWS::Region}:${AWS::AccountId}:${MediateApi}/*/POST/mcp" - !Ref AWS::NoValue TaskDefinition: Type: AWS::ECS::TaskDefinition Properties: Family: !Sub "${AWS::StackName}-alos" RequiresCompatibilities: [FARGATE] NetworkMode: awsvpc Cpu: !Ref Cpu Memory: !Ref MemoryMiB RuntimePlatform: {CpuArchitecture: X86_64, OperatingSystemFamily: LINUX} ExecutionRoleArn: !GetAtt ExecutionRole.Arn TaskRoleArn: !GetAtt TaskRole.Arn Volumes: - Name: data EFSVolumeConfiguration: FilesystemId: !Ref DataVolume TransitEncryption: ENABLED AuthorizationConfig: {AccessPointId: !Ref DataAccessPoint, IAM: ENABLED} - Name: tmp ContainerDefinitions: - Name: alos Image: !Ref ImageUri Essential: true ReadonlyRootFilesystem: true User: "1000:1000" PortMappings: [{ContainerPort: 8080, Protocol: tcp}] MountPoints: - {SourceVolume: data, ContainerPath: /data, ReadOnly: false} - {SourceVolume: tmp, ContainerPath: /tmp, ReadOnly: false} LinuxParameters: Capabilities: {Drop: [ALL]} InitProcessEnabled: true StopTimeout: 30 Environment: - {Name: ALOS_PUBLIC_URL, Value: !Ref PublicUrl} - {Name: ALOS_TRUST_PROXY, Value: "1"} - {Name: ALOS_EMAIL, Value: "off"} - {Name: ALOS_CRON, Value: "1"} - {Name: ALOS_API_URL, Value: !Ref AlosApiUrl} - {Name: WORKOS_CLIENT_ID, Value: !Ref WorkosClientId} - {Name: WORKOS_ISSUER_HOSTS, Value: !Ref WorkosIssuerHosts} - {Name: WORKOS_AUTHKIT_HOST, Value: !Ref WorkosAuthkitHost} - {Name: AUDIT_SIGN_ALG, Value: !Ref AuditSignAlg} - {Name: DEV_OPEN, Value: "0"} # Sign-in without the identity provider's secret key (docs/portability.md), on HTTPS only: over plain # http (evaluation) sign-in stays off and every signed-in route answers 401. - {Name: WORKOS_KEYLESS, Value: !If [HasCertificate, "1", "0"]} # The licence check (worker/marketplace/licence.ts): empty SKU = no check. - {Name: AWS_REGION, Value: !Ref AWS::Region} - {Name: MARKETPLACE_PRODUCT_SKU, Value: !Ref MarketplaceProductSku} - {Name: MARKETPLACE_ENTITLEMENT, Value: !Ref MarketplaceEntitlement} - {Name: MARKETPLACE_ENTITLEMENT_UNIT, Value: !Ref MarketplaceEntitlementUnit} # Mediate as an advisory pre-check on agent changes (worker/mediate/advisor.ts): empty = not asked. - {Name: MEDIATE_API_URL, Value: !If [HasMediate, !Sub "https://${MediateApi}.execute-api.${AWS::Region}.amazonaws.com", ""]} Secrets: - {Name: WORKOS_COOKIE_PASSWORD, ValueFrom: !If [GenerateSettings, !Sub "${CookieSecret}:WORKOS_COOKIE_PASSWORD::", !Sub "${SettingsSecretArn}:WORKOS_COOKIE_PASSWORD::"]} - {Name: REDACTION_KEY, ValueFrom: !If [GenerateSettings, !Sub "${RedactionSecret}:REDACTION_KEY::", !Sub "${SettingsSecretArn}:REDACTION_KEY::"]} - {Name: AUDIT_SEAL_HMAC_KEY, ValueFrom: !If [GenerateSettings, !Sub "${SealSecret}:AUDIT_SEAL_HMAC_KEY::", !Sub "${SettingsSecretArn}:AUDIT_SEAL_HMAC_KEY::"]} HealthCheck: Command: ["CMD", "node", "-e", "fetch('http://127.0.0.1:8080/api/health').then(r=>process.exit(r.ok?0:1),()=>process.exit(1))"] Interval: 30 Timeout: 5 Retries: 3 StartPeriod: 60 LogConfiguration: LogDriver: awslogs Options: awslogs-group: !Ref LogGroup awslogs-region: !Ref AWS::Region awslogs-stream-prefix: alos # ECS refuses a target group until a listener joins it to the load balancer. Both listeners are conditional, so # the service waits on this handle, which names whichever one exists. ListenerReady: Type: AWS::CloudFormation::WaitConditionHandle Metadata: Listener: !If [HasCertificate, !Ref HttpsListener, !Ref HttpListener] # Exactly one copy, and never two at once: SQLite has one writer, and limits and jobs live in memory. A new # version stops the old task before starting the new one (a short gap during updates, by design). Service: Type: AWS::ECS::Service DependsOn: [DataMountA, DataMountB, ListenerReady] Properties: Cluster: !Ref Cluster LaunchType: FARGATE PlatformVersion: LATEST DesiredCount: 1 DeploymentConfiguration: MinimumHealthyPercent: 0 MaximumPercent: 100 DeploymentCircuitBreaker: {Enable: true, Rollback: true} HealthCheckGracePeriodSeconds: 120 EnableExecuteCommand: false NetworkConfiguration: AwsvpcConfiguration: AssignPublicIp: ENABLED Subnets: [!Ref SubnetA, !Ref SubnetB] SecurityGroups: [!Ref AppSecurityGroup] LoadBalancers: [{ContainerName: alos, ContainerPort: 8080, TargetGroupArn: !Ref TargetGroup}] TaskDefinition: !Ref TaskDefinition # ---------------------------------------------------------------- alerts AlarmTopic: Type: AWS::SNS::Topic Properties: {DisplayName: ALOS alerts} AlarmSubscription: Type: AWS::SNS::Subscription Condition: HasAlarmEmail Properties: {TopicArn: !Ref AlarmTopic, Protocol: email, Endpoint: !Ref AlarmEmail} UnhealthyAlarm: Type: AWS::CloudWatch::Alarm Properties: AlarmDescription: ALOS is not answering its health check. Namespace: AWS/ApplicationELB MetricName: HealthyHostCount Dimensions: - {Name: TargetGroup, Value: !GetAtt TargetGroup.TargetGroupFullName} - {Name: LoadBalancer, Value: !GetAtt LoadBalancer.LoadBalancerFullName} Statistic: Minimum Period: 60 EvaluationPeriods: 5 DatapointsToAlarm: 3 Threshold: 1 ComparisonOperator: LessThanThreshold TreatMissingData: breaching AlarmActions: [!Ref AlarmTopic] OKActions: [!Ref AlarmTopic] ServerErrorAlarm: Type: AWS::CloudWatch::Alarm Properties: AlarmDescription: ALOS is answering server errors (5xx). Namespace: AWS/ApplicationELB MetricName: HTTPCode_Target_5XX_Count Dimensions: [{Name: LoadBalancer, Value: !GetAtt LoadBalancer.LoadBalancerFullName}] Statistic: Sum Period: 60 EvaluationPeriods: 5 DatapointsToAlarm: 2 Threshold: 5 ComparisonOperator: GreaterThanOrEqualToThreshold TreatMissingData: notBreaching AlarmActions: [!Ref AlarmTopic] OKActions: [!Ref AlarmTopic] # ---------------------------------------------------------------- ALOS Mediate (only with MediateModelPackageArn) # deploy/aws/mediate/gateway/mediate-gateway.yaml in IAM mode: a network-isolated SageMaker endpoint from the # buyer's own model package subscription, the gateway Lambda, and an HTTP API whose every route needs SigV4. Only # the app's task role is granted execute-api:Invoke (TaskRole above). Advisory and fail-closed: an error or a # timeout answers escalate, and the app treats anything but proceed as "wait for a person". MediateModelRole: Type: AWS::IAM::Role Condition: HasMediate Properties: AssumeRolePolicyDocument: Version: "2012-10-17" Statement: [{Effect: Allow, Principal: {Service: sagemaker.amazonaws.com}, Action: sts:AssumeRole}] Policies: - PolicyName: run PolicyDocument: Version: "2012-10-17" Statement: - Effect: Allow Action: [logs:CreateLogGroup, logs:CreateLogStream, logs:PutLogEvents, logs:DescribeLogStreams] Resource: !Sub arn:${AWS::Partition}:logs:${AWS::Region}:${AWS::AccountId}:log-group:/aws/sagemaker/* - Effect: Allow Action: cloudwatch:PutMetricData Resource: "*" Condition: {StringEquals: {"cloudwatch:namespace": ["/aws/sagemaker/Endpoints", "AWS/SageMaker"]}} MediateModel: Type: AWS::SageMaker::Model Condition: HasMediate Properties: ExecutionRoleArn: !GetAtt MediateModelRole.Arn EnableNetworkIsolation: true Containers: - ModelPackageName: !Ref MediateModelPackageArn MediateEndpointConfig: Type: AWS::SageMaker::EndpointConfig Condition: HasMediate Properties: ProductionVariants: - VariantName: main ModelName: !GetAtt MediateModel.ModelName InstanceType: !Ref MediateInstanceType InitialInstanceCount: !Ref MediateMinInstances InitialVariantWeight: 1 ContainerStartupHealthCheckTimeoutInSeconds: 900 ModelDataDownloadTimeoutInSeconds: 900 MediateEndpoint: Type: AWS::SageMaker::Endpoint Condition: HasMediate Properties: EndpointConfigName: !GetAtt MediateEndpointConfig.EndpointConfigName RetainAllVariantProperties: true MediateScalableTarget: Type: AWS::ApplicationAutoScaling::ScalableTarget Condition: HasMediate Properties: ServiceNamespace: sagemaker ScalableDimension: sagemaker:variant:DesiredInstanceCount ResourceId: !Sub endpoint/${MediateEndpoint.EndpointName}/variant/main MinCapacity: !Ref MediateMinInstances MaxCapacity: !Ref MediateMaxInstances MediateScalingPolicy: Type: AWS::ApplicationAutoScaling::ScalingPolicy Condition: HasMediate Properties: PolicyName: !Sub ${AWS::StackName}-mediate-invocations PolicyType: TargetTrackingScaling ScalingTargetId: !Ref MediateScalableTarget TargetTrackingScalingPolicyConfiguration: PredefinedMetricSpecification: {PredefinedMetricType: SageMakerVariantInvocationsPerInstance} TargetValue: !FindInMap [MediateScaleTarget, !Ref MediateInstanceType, PerMinute] ScaleOutCooldown: 60 ScaleInCooldown: 600 MediateGatewayRole: Type: AWS::IAM::Role Condition: HasMediate Properties: AssumeRolePolicyDocument: Version: "2012-10-17" Statement: [{Effect: Allow, Principal: {Service: lambda.amazonaws.com}, Action: sts:AssumeRole}] Policies: - PolicyName: invoke-mediate PolicyDocument: Version: "2012-10-17" Statement: - Effect: Allow Action: [sagemaker:InvokeEndpoint, sagemaker:DescribeEndpoint] Resource: !Sub arn:${AWS::Partition}:sagemaker:${AWS::Region}:${AWS::AccountId}:endpoint/${MediateEndpoint.EndpointName} - Effect: Allow Action: [logs:CreateLogStream, logs:PutLogEvents] Resource: !GetAtt MediateGatewayLogGroup.Arn MediateGatewayLogGroup: Type: AWS::Logs::LogGroup Condition: HasMediate Properties: RetentionInDays: !Ref LogRetentionDays MediateGateway: Type: AWS::Lambda::Function Condition: HasMediate Properties: Description: ALOS Mediate API and MCP gateway (advisory, fails closed) Runtime: python3.12 Handler: app.handler Architectures: [arm64] MemorySize: 256 Timeout: 29 Role: !GetAtt MediateGatewayRole.Arn Code: {S3Bucket: !Ref MediateCodeBucket, S3Key: !Ref MediateCodeKey} LoggingConfig: {LogGroup: !Ref MediateGatewayLogGroup} Environment: Variables: ENDPOINT_NAME: !GetAtt MediateEndpoint.EndpointName ENDPOINT_TIMEOUT_S: "12" MediateApi: Type: AWS::ApiGatewayV2::Api Condition: HasMediate Properties: Name: !Sub ${AWS::StackName}-mediate ProtocolType: HTTP Description: ALOS Mediate API and MCP endpoint (IAM sign-in) MediateIntegration: Type: AWS::ApiGatewayV2::Integration Condition: HasMediate Properties: ApiId: !Ref MediateApi IntegrationType: AWS_PROXY IntegrationUri: !GetAtt MediateGateway.Arn PayloadFormatVersion: "2.0" TimeoutInMillis: 29000 MediateDecideRoute: Type: AWS::ApiGatewayV2::Route Condition: HasMediate Properties: ApiId: !Ref MediateApi RouteKey: POST /v1/decide Target: !Sub integrations/${MediateIntegration} AuthorizationType: AWS_IAM MediateHealthRoute: Type: AWS::ApiGatewayV2::Route Condition: HasMediate Properties: ApiId: !Ref MediateApi RouteKey: GET /v1/health Target: !Sub integrations/${MediateIntegration} AuthorizationType: AWS_IAM MediateMcpRoute: Type: AWS::ApiGatewayV2::Route Condition: HasMediate Properties: ApiId: !Ref MediateApi RouteKey: POST /mcp Target: !Sub integrations/${MediateIntegration} AuthorizationType: AWS_IAM MediateAccessLogGroup: Type: AWS::Logs::LogGroup Condition: HasMediate Properties: RetentionInDays: !Ref LogRetentionDays MediateStage: Type: AWS::ApiGatewayV2::Stage Condition: HasMediate Properties: ApiId: !Ref MediateApi StageName: $default AutoDeploy: true DefaultRouteSettings: ThrottlingRateLimit: !Ref MediateThrottleRatePerSecond ThrottlingBurstLimit: 200 AccessLogSettings: DestinationArn: !GetAtt MediateAccessLogGroup.Arn Format: '{"requestId":"$context.requestId","route":"$context.routeKey","status":"$context.status","caller":"$context.identity.userArn","ms":"$context.responseLatency"}' MediateInvokePermission: Type: AWS::Lambda::Permission Condition: HasMediate Properties: Action: lambda:InvokeFunction FunctionName: !Ref MediateGateway Principal: apigateway.amazonaws.com SourceArn: !Sub arn:${AWS::Partition}:execute-api:${AWS::Region}:${AWS::AccountId}:${MediateApi}/* # The gateway logs error "unavailable" each time it answers escalate because the model did not answer. MediateFailClosedMetric: Type: AWS::Logs::MetricFilter Condition: HasMediate Properties: LogGroupName: !Ref MediateGatewayLogGroup FilterPattern: '{ $.error = "unavailable" }' MetricTransformations: - MetricNamespace: !Sub ALOS/Mediate/${AWS::StackName} MetricName: FailClosed MetricValue: "1" DefaultValue: 0 MediateFailClosedAlarm: Type: AWS::CloudWatch::Alarm Condition: HasMediate Properties: AlarmDescription: Mediate answered escalate because the model did not answer (error or timeout). Agent changes are waiting for people. Namespace: !Sub ALOS/Mediate/${AWS::StackName} MetricName: FailClosed Statistic: Sum Period: 60 EvaluationPeriods: 5 DatapointsToAlarm: 2 Threshold: 5 ComparisonOperator: GreaterThanOrEqualToThreshold TreatMissingData: notBreaching AlarmActions: [!Ref AlarmTopic] OKActions: [!Ref AlarmTopic] MediateModelErrorAlarm: Type: AWS::CloudWatch::Alarm Condition: HasMediate Properties: AlarmDescription: The Mediate endpoint returned server errors (5xx). Namespace: AWS/SageMaker MetricName: Invocation5XXErrors Dimensions: [{Name: EndpointName, Value: !GetAtt MediateEndpoint.EndpointName}, {Name: VariantName, Value: main}] Statistic: Sum Period: 60 EvaluationPeriods: 5 DatapointsToAlarm: 2 Threshold: 1 ComparisonOperator: GreaterThanOrEqualToThreshold TreatMissingData: notBreaching AlarmActions: [!Ref AlarmTopic] OKActions: [!Ref AlarmTopic] MediateLatencyAlarm: Type: AWS::CloudWatch::Alarm Condition: HasMediate Properties: AlarmDescription: Mediate is slow; the slowest 10% of decisions are over the limit. Add instances or raise MediateMaxInstances. Namespace: AWS/SageMaker MetricName: ModelLatency Dimensions: [{Name: EndpointName, Value: !GetAtt MediateEndpoint.EndpointName}, {Name: VariantName, Value: main}] ExtendedStatistic: p90 Period: 60 EvaluationPeriods: 5 DatapointsToAlarm: 5 Threshold: !Sub "${MediateLatencyAlarmMs}000" ComparisonOperator: GreaterThanThreshold TreatMissingData: notBreaching AlarmActions: [!Ref AlarmTopic] OKActions: [!Ref AlarmTopic] MediateGatewayErrorAlarm: Type: AWS::CloudWatch::Alarm Condition: HasMediate Properties: AlarmDescription: The Mediate gateway function failed. Namespace: AWS/Lambda MetricName: Errors Dimensions: [{Name: FunctionName, Value: !Ref MediateGateway}] Statistic: Sum Period: 60 EvaluationPeriods: 5 DatapointsToAlarm: 2 Threshold: 1 ComparisonOperator: GreaterThanOrEqualToThreshold TreatMissingData: notBreaching AlarmActions: [!Ref AlarmTopic] OKActions: [!Ref AlarmTopic] Outputs: AppUrl: Description: Where ALOS answers Value: !Ref PublicUrl LoadBalancerDns: Description: Point your domain's CNAME (or alias) here, then send TEKIMAX the https address so sign-in accepts it Value: !GetAtt LoadBalancer.DNSName DataVolumeId: Description: The EFS file system holding the database and files (kept if the stack is deleted) Value: !Ref DataVolume LogGroup: Value: !Ref LogGroup AlarmTopicArn: Description: Subscribe more people or a pager to these alerts Value: !Ref AlarmTopic TargetGroupArn: Description: The app's target group; its target health says whether ALOS is up behind HTTPS Value: !Ref TargetGroup MediateApiUrl: Condition: HasMediate Description: ALOS Mediate's API (IAM sign-in; only this stack's app role may call it). ALOS uses its /mcp address. Value: !Sub https://${MediateApi}.execute-api.${AWS::Region}.amazonaws.com MediateEndpointName: Condition: HasMediate Value: !GetAtt MediateEndpoint.EndpointName