zqiren/Orbital

What it solves

Orbital is a project-centric agent coordinator that prevents context loss when switching between different AI agents (like Claude Code, Codex, or Gemini CLI). Instead of locking project state and decisions within a single chat session, Orbital stores this context as plain files in your local folder, making the project itself the unit of work rather than the session.

How it works

Orbital acts as a manager agent that maintains a set of "Layer-1" memory files in the project workspace (such as PROJECT_STATE.md, DECISIONS.md, and LESSONS.md). These files are updated by the agent and injected into the system prompt of any worker agent it delegates to.

When a task is assigned, Orbital can either perform the work itself or dispatch it to a specialized worker agent via various transports (SDK, PTY, ACP, or JSON-RPC). Each worker reads the authoritative project files to orient themselves without the user needing to re-explain the goals. The system also includes a task queue that enforces a "Completed" or "Blocked" status for every item to prevent tasks from drifting.

Who it’s for

Developers and power users who use multiple AI coding agents and want a unified, persistent project memory that persists across different tools and sessions.

Highlights

• Interchangeable Intelligence: Switch between different LLM providers and agents mid-task without losing context.
• Local-First Context: Project state, decisions, and lessons are stored as plain Markdown files in the local directory.
• Sub-Agent Delegation: Supports a wide array of CLI-based agents including Claude Code, Codex, Cursor, and Gemini CLI.
• Enforced Task Closure: A task queue that requires agents to explicitly mark tasks as completed or blocked.
• Prefix Caching: Optimizes token usage by splitting the system prompt into static and dynamic sections.