Ai Research Accelerator

AI Research Accelerator

**AI for Accelerating Research.** An author-first plugin for Codex and Claude Code that covers the research workflow from distributed training to publication. It packages seven primary skills:

• multinode-training: design, launch, validate, and debug reliable multi-node GPU training;
• shared-filesystem-agent-coordination: coordinate multiple Agent sessions through a fenced shared-file message bus;
• long-task-relay: monitor long-running work with a token-free rule-based watcher that wakes the exact agent conversation only when judgment is needed;
• write-insightful-topconf-paper: build and audit insight-driven ML papers;
• github-paper-review-workflow: revise Chinese semantic Markdown and formal English LaTeX through GitHub PR review;
• manage-paper-experiments: maintain the paper-wide experiment ledger and reproducible execution queues;
• plot-paper-experiments: produce traceable publication figures from experiment data.

The relay replaces model-driven `sleep`/poll loops with a persistent local watcher. It can observe logs, progress, markers, processes, and tmux sessions, then resume an exact Codex thread, invoke an explicit agent resume command, inject into a verified tmux TUI, or write a durable event inbox.

The shared-filesystem Agent workflow provides coordinator-first bootstrap, persistent stateless worker dispatch, optional exact-thread compatibility, atomic request/ACK/result records, fencing, bounded repair, and token-free monitoring for Agents that cannot message one another directly.

The multi-node workflow treats topology, launch coordination, and external assets/services as separate contracts. It provides fail-closed staging, single-writer asset preparation, all-node readiness barriers, and distributed first-work validation for PyTorch, Accelerate, DeepSpeed, and related stacks.

Optional integrations

The plugin does not bundle or auto-install external skills. It uses these capabilities when they are alr