mahakagarwal7/featureflag

Purpose and Entrypoints

This repository is a feature-flag rollout simulation benchmark: agents make sequential rollout decisions while telemetry, risk, revenue, and health evolve. The primary runtime is Python; entrypoints include API servers at backend/feature_flag_env/server/app.py and backend/server/app.py, local/remote evaluation in backend/inference.py, and training scripts backend/train_rl.py and backend/train_enterprise.py.

Simulation Environment

The environment owns episode state, action application, and telemetry transitions for rollout actions such as increase, halt, full rollout, and rollback. backend/feature_flag_env/server/feature_flag_environment.py is the central environment boundary; backend/feature_flag_env/server/simulation_engine.py supplies simulation behavior, while backend/feature_flag_env/models.py defines domain/state models. Scenario and mission inputs extend the core simulation through backend/feature_flag_env/configs/scenario_library.py and backend/feature_flag_env/missions.py.

Evaluation and Reward Pipeline

A step produces updated observation and reward; a completed trajectory is task-scored. Task definitions separate benchmark objectives—safe rollout, risk-aware scaling, and multi-objective optimization—in backend/feature_flag_env/tasks/task1_safe_rollout.py, backend/feature_flag_env/tasks/task2_risk_aware.py, and backend/feature_flag_env/tasks/task3_multi_objective.py. Shared trajectory grading is in backend/feature_flag_env/tasks/graders.py; reward policy and extended reward signals reside in backend/feature_flag_env/utils/reward_functions.py and backend/feature_flag_env/utils/extended_rewards.py.

Agent Layer

Agents consume observations and emit environment actions behind a common abstraction. backend/agents/base_agent.py establishes the agent contract, and backend/agents/factory.py selects implementations. Major policy variants are deterministic baseline (backend/agents/baseline_agent.py), LLM-driven (backend/agents/llm_agent.py), RL/checkpoint-driven (backend/agents/rl_agent.py), and enterprise orchestration (backend/agents/enterprise_agent_v2.py, backend/agents/master_agent.py). Ensemble and human-approval modes are separate decision boundaries in backend/agents/ensemble_agent.py and backend/agents/human_in_loop_agent.py.

API, Security, and Operational Controls

The simulation is exposed as an HTTP API with reset, step, state, health, and info operations through backend/feature_flag_env/server/app.py. Server concerns are split from environment behavior: backend/feature_flag_env/server/security.py handles security controls, and backend/feature_flag_env/utils/monitoring.py provides monitoring/alerting support. Persistence/audit-oriented infrastructure is represented by backend/feature_flag_env/utils/database.py and repository runtime artifacts under logs/.

Integrations and External Services

Tooling isolates external actions behind interfaces and a manager: backend/feature_flag_env/tools/tool_interface.py, backend/feature_flag_env/tools/tool_manager.py, and backend/feature_flag_env/tools/connectors.py. GitHub and Slack integrations have dedicated adapters in backend/feature_flag_env/tools/github_integration.py and backend/feature_flag_env/tools/slack_integration.py; mock versus real adapter boundaries are explicit in backend/feature_flag_env/tools/mock_adapters.py and backend/feature_flag_env/tools/real_adapters.py. The LLM agent supports OpenAI-compatible/Hugging Face router-style inference via environment configuration.

UI and Deployment Boundary

A separate Next.js/TypeScript frontend presents dashboard, flag, experiment, monitoring, integration, and AI-decision workflows. Route composition begins at frontend/app/layout.tsx, dashboard pages live under frontend/app/(dashboard)/, and backend communication is centralized in frontend/lib/api.ts. The backend can also serve committed static frontend output from backend/static/. Docker deployment is defined by Dockerfile and backend/Dockerfile, targeting Hugging Face Spaces’ Docker runtime and port 7860.