rpamis/comet

What it solves

Comet is a resumable, long-running task workflow and Skill platform designed specifically for AI-driven coding. It solves the problem of AI agents losing track of progress during complex, multi-step development tasks, reducing the need for manual intervention and repeated prompting to update documentation or archive changes.

How it works

Comet provides two distinct workflows:

• Native Workflow: Designed for strong models that can plan and verify autonomously. It uses a detailed brief and target specifications to constrain outcomes while allowing the model to handle implementation and testing.
• Classic Workflow: A phased methodology (combining OpenSpec and Superpowers) that uses a state machine and phase checks to enforce strong constraints and explicit methods.

It uses a configuration-driven entry point (/comet) and a set of scripts (like comet-guard.mjs) to act as "guards" that verify task completion and state transitions before the workflow advances. It also includes a Skill platform for authoring, distributing, and evaluating reusable coding skills using scientific metrics like Pass@k and Pass^k.

Who it’s for

Software developers and AI engineers who use AI coding platforms and want a structured, recoverable, and scientifically evaluable way to manage long-running coding tasks and agentic workflows.

Highlights

• Resumable State: Records phase, execution mode, and verification results in .comet.yaml so agents can resume work after interruptions without guessing progress.
• Dual Workflow Options: Offers both a high-performance Native mode for autonomous models and a constrained Classic mode for phased development.
• Skill Evaluation: Integrated evaluation platform with Rubric scoring and LangSmith/LangFuse integration for production-grade skill evolution.
• Automated Guardrails: Uses scripts to automate handoffs, state updates, and validation, removing the from the agent's memory threshold.
• Visual Dashboard: Includes a local read-only dashboard to track active changes, phase status, and task progress.