windags-resilience

What problem does it solve?

Provides durable state management, checkpointing, and automated recovery for WinDAGs executions so that expensive LLM-driven nodes are preserved on success and only failed work is retried, preventing costly retries and state corruption.

Core Features & Use Cases

• Validated State Machine: Enforces a strict transition matrix with fail-fast behavior to prevent illegal state mutations and emits events for auditing and observability.
• Checkpoint Architecture: Wave- and phase-level checkpoints with a storage-adapter pattern for pluggable backends, small checkpoint footprints, and retention policies.
• Resume & Recovery Protocols: Idempotent re-execution, selective re-run of failed nodes, and injection of preserved outputs on resume; recovery manager classifies errors and escalates strategies by cost.
• Use Case: Resume a multi-phase LLM orchestration after a crash and continue from the last completed wave without re-running completed nodes, while applying model escalation or decomposition for stubborn failures.

Quick Start

Resume a failed WinDAGs execution from the last successful wave, preserving completed node outputs and re-executing only failed or selected nodes.