feat(grader): add task-aware grading with new metrics and composite scoring
div18
e012e73
- Extend grading dimensions to include recovery speed, VIP protection, and action efficiency
- Introduce ActionRecord class to track agent actions and their measurable effects
- Implement EpisodeGrader to accumulate episode data, compute advanced metrics, and produce task-aware grades
- Define task-specific weight profiles for composite score calculation in Grade class
- Add new metrics: recovery speed for task-2, VIP protection for task-3, Lyapunov variance, action efficiency, cost trajectory, and peak queue sum
- Adjust composite score calculation to incorporate task-specific logic and penalize invalid actions
- Enhance summary output to include additional metrics and action statistics
feat(inference): implement baseline evaluation mode and structured JSON reporting
- Add baseline mode to run evaluations without replay buffer demonstrations
- Integrate action effect determination based on acknowledgement status and record actions for efficiency analysis
- Serialize grades including…
fix(inference): refine scaling and rerouting rules and action behavior
div18
d062bfb
- Updated task-2 and task-3 descriptions for clearer scaling and rerouting guidelines
- Modified SCALE_DOWN action to first cancel pending boots before removing active capacity
- Changed parsing of target_node_id and parameter to handle None values correctly
- Allowed NO_OP action to always succeed without target node validation
- Improved SCALE_DOWN implementation to prioritize canceling pending capacity
- Clarified cost tiers to classify pending capacity with justified excess at moderate rate
- Adjusted cost calculation for active and pending capacity to reflect tiered charging accurately
reward finalisation
div18
7ccb648
fix(inference): add debugging and error handling for env.step action calls
div18
b84be63
- Log each action step with compacted action string for better traceability
- Catch and log RuntimeError exceptions during env.step to identify failures
- Re-raise exceptions after logging to maintain original error flow
- Ensure printed debug information flushes stdout immediately for real-time monitoring
feat(reward): add comprehensive AntiAtropos reward function specification
- Provide detailed Markdown documentation explaining reward function design
- Describe multi-scale control signal architecture for Autonomous SRE
- Explain Lyapunov graph energy, reward composition, sigmoid normalization
- Outline layered action-efficiency penalties and episode grading system
- Highlight theoretical guarantees and practical innovation in reward design
fix(environment): correct idle capacity calculation to prevent negative values
- Adjust idle capacity formula to use max(0, ...) to avoid negative idle counts
- Clarify commented explanation to reflect correction for idle capacity calculation
…
perf(environment): improve cost model with three-tier capacity pricing
div18
92019fc
- Import DEFAULT_CAPACITY constant from simulator module as baseline capacity
- Update Kubernetes deployment replicas from 1 or 2 to 3 for all services in local-laptop and aws manifests
- Increase pod resource quota from 30 to 45 to support higher replica count
- Enhance _compute_cost() method with a three-tier cost calculation
- Define tier 1 as baseline capacity at cheap base rate up to DEFAULT_CAPACITY
- Define tier 2 as needed excess capacity charged at moderate 4× base rate
- Define tier 3 as idle excess capacity charged at high 20× penalty rate
- Provide clear incentives for scaling just enough to meet traffic demand without overprovisioning
refactor(simulator): add feedback on action effects and enforce capacity discipline
div18
f55f75f
- Extend apply_action() to return boolean indicating if action affected state
- Prevent silent discards by explicitly handling no-op and invalid actions
- Remove artificial max capacity cap; rely on cost function to limit scaling
- Update scaling logic to allow larger max capacity per node (sentinel value)
- Adjust simulator and environment to penalize wasted or rejected actions
- Enhance handling of critical nodes to reject forbidden shed load attempts
- Improve status update logic: scaling up repairs DEGRADED node status
- Update task descriptions and docs to emphasize reactive scaling over proactive
- Tune environment and reward normalization parameters for balance between cost and SLA compliance
better task 3
div18
74770e8
Task 3 fixes
div18
16d8704
fixes
div18
3550e4d
feat(client): add new node and reward related metrics to observations
div18
630f735
- Add capacity, pending_capacity, queue_delta, sla_proximity, outflow_rate attributes to node data
- Include upstream_nodes, downstream_nodes, upstream_pressure, and node_reward in node info
- Extend observation data with action_id, executor_latency_ms, executor_error_code fields
- Add reward_drift, reward_cost, reward_sla, reward_barrier to reward metrics
- Move choke_level to the end of observation data with other new fields
- Update inference node serialization to include new node attributes for comprehensive state representation
Merge branch 'integrate/k8s-tennodes' of https://github.com/PranavKK1201/Anti-Atropos into integrate/k8s-tennodes
div18
9dda2d2
reward tuning
div18
faa8a6b
graph modelling
PranavKK1201
b5e5650
model changes
div18
9db539d
fix: PromQL queries use by (node_id) so VM telemetry resolves correctly
feat(monitoring): add local Prometheus and Grafana setup for AntiAtropos
div18
6875d9a
- Add docker-compose.local-monitoring.yml to run Prometheus and Grafana locally
- Create start-local-monitoring.ps1 script for easy startup and health checks
- Add Prometheus config for scraping local OpenEnv metrics at host.docker.internal:8000
- Add Grafana datasource and workload dashboard JSON files for detailed monitoring UI
- Update deploy-droplet-one-shot.sh to expose Grafana via NodePort 30000 for external access
- Install and configure Prometheus and Grafana with Helm in Kubernetes cluster
- Configure systemd service for AntiAtropos control API with environment variables
- Improve deployment script to wait for Grafana readiness and patch service for NodePort
- Provide user instructions for accessing local and remote monitoring dashboards
consolidate
div18
504ced2
Enhance Kubernetes executor and deployment configurations
div18
af7de22
- Introduced remote execution capabilities in KubernetesExecutor to delegate actions to a FastAPI control plane.
- Updated .env.example to clarify Prometheus URL usage for local and VM telemetry.
- Modified Prometheus Helm values to expose the service on NodePort 30090 and added a scrape job for annotated pods in the prod-sre namespace.
- Updated deployment scripts and README to reflect new Prometheus configurations and provide clearer instructions for local setup.
fix reqs
div18
d6496d4
droplet
div18
2244d0f
chore(kubernetes): support unbounded max replicas and enhance local setup
div18
923f89f
- Allow ANTIATROPOS_MAX_REPLICAS to be empty or 'unbounded' for no scale-up limit
- Adjust scaling logic to handle unbounded max replicas correctly
- Update .env.example with local defaults and new Prometheus queries
- Simplify start-grafana.ps1 for local Grafana setup without AWS credentials
- Add kind cluster config with max pods set to 250 for local testing
- Enhance local-laptop.yaml with prometheus metrics exposition and python http server
- Update LOCAL_LAPTOP_FASTAPI_GUIDE.md to document unbounded max replicas and setup changes
deploy
div18
5a30495
feat: implement Kubernetes executor for automated cluster scaling and infrastructure management
div18
cf2697b
chore: remove generated artifacts and ignore local caches
- Add PowerShell deploy script for setting up EKS cluster, nodegroup, workloads, AMP workspace,
Prometheus Agent with IRSA, self-hosted Grafana, Cluster Autoscaler, and kubeconfig generation
- Implement phased deployment approach with checks for existing resources to avoid duplication
- Create Prometheus service account with IAM policy attachment via eksctl for secure remote writes
- Automate Helm chart installation and upgrades for Prometheus Agent, Grafana, and Cluster Autoscaler
- Include automated dashboard secret creation for Grafana via kubectl
- Add robust node readiness check after nodegroup creation
- Generate AWS IAM-based kubeconfig for Hugging Face Spaces integration, with encoding instructions
- Update bash deploy script to align default region with PowerShell script (ap-southeast-1) and
adopt consistent phased deployment with improved comments and error handling
- Refine eksctl cluster YAML config with updated region, availability zones, and cloudwatch logging
- Add s…
run this to see if the loss and rewards are beheving okay
PranavKK1201
3bd27e2
increased ramp-up to be proportional to 10 nodes
PranavKK1201
236abb8
upgraded to 10 nodes
PranavKK1201
d6439a8
agent.md and claude.md
PranavKK1201
0a957b8
script for sanity check of all 3 tasks and training environment
PranavKK1201
0c74ebd
feat(curriculum): add progressive training curriculum management
div18
52a986a
- Introduce CurriculumStage dataclass to define tasks with step limits, thresholds, temps, and retries
- Define CURRICULUM list with staged tasks of increasing difficulty and parameters
- Implement CurriculumTracker to track current stage, report scores, handle retries, and progress
- Add retry temperature adjustment and automatic skip after max retries for exploration encouragement
docs(deploy): add detailed architecture guide and update k8s workloads
div18
8cd4141
- Add AntiAtropos Architecture Guide explaining system design and Kubernetes concepts
- Document SRE agent operation loop and action effects on AWS EKS
- Explain telemetry flow from real cluster to agent via AMP and Prometheus
- Clarify three-layer scaling caps for cost control and safety limits
- Provide mapping of simulator nodes to real Kubernetes deployments
- Include comprehensive component listings for Hugging Face Spaces, AWS EKS, and managed services
- Describe simulator modes and cost implications of scaling actions
- Add ResourceQuota to prod-sre namespace limiting pods and resource usage
- Restructure and clean Kubernetes manifest for microservice deployments and services
- Enhance manifest comments for better understanding of workloads and scraping setup
feat(aws): add complete AWS infrastructure deployment with sample workloads and autoscaling
div18
0d2b52f
- Add cluster-autoscaler Helm values configuration for node autoscaling
- Revise deploy.sh to deploy EKS cluster, sample workloads, AMP and AMG workspaces
- Implement IRSA setup and Prometheus Agent Helm installation in deploy.sh
- Include Cluster Autoscaler installation and upgrades in deploy.sh steps
- Generate kubeconfig for HF Spaces to connect securely to EKS with AWS IAM authenticator
- Add generate-kubeconfig.sh script to produce kubeconfig file for external access
- Provide k8s-workloads.yaml to deploy sample microservices with Prometheus scraping annotations
- Create detailed OPERATIONS.md guide with deployment instructions, cost cap strategies, and autoscaling setup
- Remove Docker build and ECR push steps as FastAPI runs externally on Hugging Face Spaces
- Improve deployment script output with clear instructions for configuring HF Spaces environment variables
feat(deploy): add AWS deployment scripts and configurations for AntiAtropos
div18
85041bd
- Add deploy/aws/deploy.sh to automate full AWS deployment on EKS
- Create eksctl cluster config with managed node groups and addons
- Add Kubernetes manifests for namespace, configmap, deployment, service,
secret, and optional ALB ingress
- Provide Prometheus agent Helm values for remote write to AMP workspace
- Include Grafana dashboard JSON placeholders for overview and live views
- Add IAM trust policy JSON for Grafana role setup
- Write comprehensive README.md with step-by-step AWS deployment guide,
instructions, and troubleshooting tips
- Enable IRSA setup for Prometheus and AntiAtropos service accounts
- Integrate building and pushing AntiAtropos Docker image to ECR
- Support ALB ingress with AWS Load Balancer Controller annotations
- Configurefeat(deploy/aws): add complete AWS deployment setup for AntiAtropos
- Add deploy script to create E metrics scraping, liveness/readiness probes, and resource limitsKS cluster, AMP workspace, IAM roles, and deploy app
- Provide eksctl …
model name fix
div18
b379876
fixed env vars
div18
3c45e73
updated inference.py
div18
c927ddb
inference
Divyansh Agrawal
4d4e8bd
update readme.md
Divyansh Agrawal
7d6383a
updated grader
Divyansh Agrawal
fedca64
fixed openenv
Divyansh Agrawal
c2b19b3
docs: reframe README around AIOps and autonomous DevOps
