CosmosMind-ai/RSI-Harness

RSIH — versionable, shareable agent harness: Pi coding agent + Genome config layer

What it solves

RSIH (RSI-Harness) addresses the problem of fragmented agent configuration. Instead of scattering system prompts, tool sets, and settings across various files and CLI flags, RSIH allows users to bundle these into a single, versionable and shareable object called a Genome. This enables users to switch between different agent "harnesses" (contexts) instantly and ensures that agent setups are reproducible and reproducible.

How it works

Built as a configuration layer on top of the Pi coding agent, RSIH does not fork the core agent logic. Instead, it uses a "patch" system where a Genome defines overrides for 12 specific components (such as instructions, tools, skills, and runtime policies). If a field is absent in the Genome, it inherits the default Pi settings.

Key features include:

  • Genome Expression Engine (GEE): An interactive tool (gee) that analyzes a user's actual session history (from RSIH, Pi, and Claude Code) to automatically generate a new Genome based on evidence of recurring patterns and tool usage.
  • CWD-based sessions: Support for --run-id to maintain conversations across multiple scripted calls.
  • Genome Management: Tools to list, show, validate, and install Genomes from local directories or seed bundles.

Who it’s for

Developers and AI agent users who need to maintain multiple specialized agent configurations for different projects or workflows (e.g., a research experiment harness) and want to be able to share these configurations with others.

Highlights

  • Evidence-based Generation: Automatically creates agent harnesses by analyzing real usage history rather than asking the user to write prompts.
  • Zero-Core Fork: Maintains full compatibility with the Pi coding agent, ensuring that any new Pi feature is automatically supported.
  • Shareable Bundles: Genomes are self-contained directories that can be dropped into a local folder to instantly deploy a specialized agent setup.
  • Comprehensive Configuration: Covers 12 distinct components including keybindings, themes, MCP servers, and runtime policies.

Related

  • Project
    vybestack/llxprt-code

    A provider-agnostic CLI AI coding assistant that lets developers use any LLM provider, including local models and existing subscriptions, to edit codebases and automate workflows.

  • Project
    tinyflow-ai/tinyflow

    Tinyflow is a lightweight AI agent orchestration component that allows developers to integrate agentic capabilities into traditional applications via a framework-agnostic UI and multi-language backend support.

  • Project
    steve02081504/fount

    A programmable, modular AI agent runtime platform that enables deep customization of agent logic and seamless integration into IDEs, browsers, and P2P networks.

  • Project
    FireBird-Technologies/Auto-Analyst

    An open-source AI data scientist that automates data cleaning, statistical analysis, and machine learning via a modular multi-agent system.