Open Dynamic Workflow: An Open-Source Alternative to Claude Code's Dynamic Workflow, Supporting Multi-Model and Deterministic Recovery

Open Dynamic Workflow (odw) is an open-source reimplementation of Claude Code's built-in Dynamic Workflow, addressing limitations of the original black-box system. Key features:

• Multi-LLM runtime support (including domestic models like DeepSeek, Kimi, Qwen)
• Git worktree isolation for parallel tasks
• Deterministic crash recovery
• Offline execution reports
• Decoupled orchestration/execution layers Written in Rust, licensed under MIT for broad adoption.

Claude Code's built-in Dynamic Workflow lacks self-hosting, engine customization, precise model tracking, and crash recovery—critical gaps for enterprise scenarios. Odw was built as an open-source alternative with enterprise-grade features, split into two Rust crates for orchestration and execution.

Odw’s core components:

1. odw crate: Orchestrates workflows (primitives like agent/parallel/pipeline), manages budgets, git isolation, recovery, and dispatches tasks.
2. pandacode crate: Executes tasks via Codex, Claude, or domestic LLMs (Bamboo runtime: DeepSeek, Kimi, Qwen, Zhipu, Minimax). Components are independent—replace executors without changing orchestration logic.

Odw’s key features:

• Full workflow primitives (agent, parallel, pipeline, phase, budget, nested workflows)
• Multi-runtime per node (Codex/Claude/domestic LLMs)
• Git worktree isolation for parallel task conflict avoidance
• Deterministic recovery via journal logs
• Offline report.html with Mermaid charts (node details: model, tokens, time, prompts)

Cost comparison: 12-node workflow uses ~79k tokens (odw: mixed domestic + Claude) vs ~252k tokens (built-in: pure Claude)—1/3 cost. Reasons:

1. Cheaper domestic models for simple tasks
2. Lighter odw nodes vs full Claude sub-agents Stable for 100 parallel nodes (concurrency: min(16, cores-2)).

Tradeoffs:

• Built-in: Zero config, full Claude sub-agents, session monitoring.
• Odw: Cost savings, offline observability, crash recovery, self-hosting/headless. Odw use cases: Cost-sensitive, offline reporting, crash recovery, self-hosted scenarios.

Installation: Clone repo → run install.sh (adds odw/pandacode to PATH). Use odw doctor to verify environment. Usage:

• JS workflows (same syntax as built-in)
• odw init: Generate project files
• odw exec --backend mock: Dry run
• Claude Code users: Dedicated skill file (.claude/skills/odw/SKILL.md) for agent guidance.