Changelog¶
All notable changes to this project will be documented in this file.
[1.0.0] — 2026-02-14¶
Initial Release¶
A replication-aware worker system with explicit contracts, sandboxed orchestration, and structured observability. Designed for studying and testing AI agent replication policies in a controlled environment.
Features¶
- Replication Contract — Configurable policy with
max_depth,max_replicas,cooldown_seconds, optionalexpiration_seconds, and pluggable stop conditions evaluated during worker registration. - Manifest Signing — HMAC-SHA256 signed manifests that capture cloned state, resource quotas, parent/child lineage, and issuance time. Tamper-proof verification on registration.
- Controller — Central authority that issues/signs manifests, enforces depth limits, replica quotas, cooldown periods, and stop conditions. Includes kill switch, audit trail, and stale worker reaping.
- Sandbox Orchestrator — Simulates isolated container environments with CPU/RAM limits and network policy (egress-only to controller). Records lifecycle events and emits resource metrics.
- Worker — Task-executing agent that registers with the controller, heartbeats, handles expiration, and optionally self-replicates subject to the contract.
- Observability — Structured in-memory logger with audit trail support and metrics emission for lifecycle events, replication decisions, and resource enforcement.
Infrastructure¶
- GitHub Actions CI with flake8 linting, mypy type checking, and pytest across Python 3.10/3.11/3.12.
- Multi-stage Dockerfile for containerized deployment.
- Comprehensive test suite covering replication depth, quota enforcement, stop conditions, kill switch, heartbeat reaping, and chaos scenarios.