Track state for a long-running AI task across many steps without losing context.
Split a complex task among multiple specialized AI subagents that each focus on one piece.
Require formal verification evidence before an AI task is marked complete.
Apply a structured mission contract to define success criteria before starting a run.
| y4tacker/waymark | 1lystore/awaek | actashui/sjtu-ppt-template-skill | |
|---|---|---|---|
| Stars | 13 | 13 | 13 |
| Language | Python | Python | Python |
| Setup difficulty | moderate | moderate | moderate |
| Complexity | 3/5 | 2/5 | 2/5 |
| Audience | developer | vibe coder | researcher |
Figures from each repo's GitHub metadata at analysis time.
Requires Python and Claude Code, no external dependencies beyond that.
Waymark is a plugin for Claude Code, the AI coding assistant, designed to manage long-running tasks that span many steps and potentially many AI worker sessions. The core problem it addresses is that when an AI agent works on a complex task over time, the chat conversation becomes unreliable as memory. Messages get summarized, context fades, and the AI has no solid record of what was already decided or verified. Waymark replaces the chat transcript as the source of truth with a SQLite database file stored on disk. The workflow starts with a mission contract, which is a structured document describing what the task is, what conditions count as success, and any constraints or assumptions. A human reviews this before anything starts. Once approved, the contract is committed to the database and the run begins. From that point forward, nothing counts as state unless it goes through the command-line tool that writes to that database. Instead of one AI doing all the work in a single conversation, Waymark dispatches multiple specialized AI subagents. One checks whether the task can be resolved quickly. Another reasons about what evidence is still missing. A third claims a specific piece of work, executes it, and records what it found. A fourth acts as a verifier that challenges the claimed results before they count as complete. Each subagent reads only the portion of the database relevant to its current task, keeping each conversation focused. Completion is treated as a formal gate rather than a claim. The run is only considered finished when a verification record exists and a final-status check returns a confirmed ready state. This is meant to prevent an AI agent from declaring a task done without backing it up with actual evidence. The project is domain-general, meaning it can in principle be applied to software development, research, writing, or any other multi-step task where success criteria can be stated in advance. It requires Python and runs using only standard library modules, with no external dependencies beyond Claude Code itself. The license is MIT.
A Claude Code plugin that tracks long, multi-step AI tasks in a database instead of relying on chat memory, with formal verification before completion.
Mainly Python. The stack also includes Python, SQLite.
Setup difficulty is rated moderate, with roughly 30min to a first successful run.
Mainly developer.
This repo across BitVibe Labs
Verify against the repo before relying on details.