Summary

This project provides a local, interactive runtime for Claude Code's dynamic LLM agent workflows, leveraging a local Codex (GPT) backend. It addresses the need for more sophisticated, stateful, and observable agent execution beyond one-shot interactions, benefiting developers and researchers by enabling complex task orchestration and debugging.

How It Works

The system re-hosts the Claude Code dynamic workflows DSL, enabling users to define multi-agent tasks using constructs like agent(), parallel(), and pipeline(). It introduces crucial enhancements over the native one-shot DSL, including long-lived, sessionful workers (agent.start(), session.steer()) that maintain warm context for cost-effective, iterative tasks. Workflows are executed against a local Codex app-server, with results visualized via an interactive HTML viewer or terminal ASCII map, and advanced features like fleet supervision and human-in-the-loop gates.

Quick Start & Requirements

• Installation: Install as a Claude Code plugin (/plugin install codex-workflows) or clone the repository. A standalone CLI is available via npx github:scasella/claude-dynamic-workflows-codex.
• Prerequisites: Node.js ≥ 18, a logged-in codex CLI, and a running local Codex app-server.
• Demo: View the flagship incident demo offline: node runner/bin/view-run.js examples/incident-demo --open.
• Docs: Refer to references/authoring.md for DSL and patterns, and references/runner-readme.md for internals.

Highlighted Details

• Sessionful Workers: Enables stateful agent interactions, allowing workers to maintain context across multiple turns for efficient interrogation of large datasets or codebases.
• Interactive Run Viewer: Offers a real-time, visual DAG representation of workflow execution, complete with per-agent timelines, live updates, and a "cockpit" for interactive decision gates.
• Fleet Supervision (--multi): Orchestrates and manages multiple concurrent workflows, automatically supervising them, answering gates, steering progress, and consolidating results.
• Cost Optimization: Features like agent.waitAny with session.cancel allow for cancelling losing workers in parallel races, and --plan provides dry-run cost estimates.

Maintenance & Community

This is an unofficial, community-driven project, not affiliated with OpenAI or Anthropic. No specific community channels are listed.

Licensing & Compatibility

The project is released under the permissive MIT license, allowing for commercial use and integration into closed-source projects.

Limitations & Caveats

As an external re-host, it lacks some native Claude Code in-session features like background tasks or the integrated progress UI. Session resume functionality depends on the availability of persisted Codex threads, and budget accounting is per-process. The run visualization approximates pipeline structures.