windags-resilience

Manage durable WinDAGs state with checkpointed recovery for LLM executions.

2|Updated Feb 11, 2026
One-click install
npx skills add https://github.com/curiositech/port-daddy --skill windags-resilience
Or copy as Structured Prompt for Agent
Please help me install this Agent Skill.
Skill: windags-resilience
Source: https://github.com/curiositech/port-daddy/tree/main/docs/reference-skills/windags-resilience
Command: npx skills add https://github.com/curiositech/port-daddy --skill windags-resilience

SYSTEM DOCUMENTATION & REQUIREMENTS

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.

Frequently Asked Questions about windags-resilience

High-intent search queries and answers about installing and using this skill.

FAQPage Schema
How do I resume a multi-phase LLM orchestration after a crash without re-running completed nodes?

Multi-phase LLM orchestration can resume from the last successful wave checkpoint, preserving completed node outputs and re-executing only failed nodes. This prevents costly retries and state corruption by applying idempotent re-execution rules.

What is checkpointed recovery for state-machine workflows?

Checkpointed recovery for state-machine workflows involves saving wave and phase-level state so failed executions can continue from the last completed point. A validated transition matrix enforces legal state mutations, while recovery protocols selectively re-run failed nodes.

Can I use custom storage backends for workflow checkpoints?

Custom storage backends are supported through a pluggable storage-adapter pattern for wave and phase checkpoints. This allows flexible backend integration while maintaining small checkpoint footprints and configurable retention policies.

What's the best way to handle stubborn node failures in state-machine workflows?

Stubborn node failures in state-machine workflows are handled by a recovery manager that classifies errors and escalates recovery strategies by cost. Strategies include model escalation or decomposition to ensure nodes eventually succeed without manual intervention.

When do I need idempotent re-execution for LLM orchestration?

Idempotent re-execution is needed when resuming LLM orchestration after failures to ensure retried nodes produce consistent results without side effects. It enables injection of preserved outputs on resume, preventing duplicate work and state corruption.