dirge-code/dirge

Dynamic Intent Resolution Grounding Engine

What it solves

Dirge is a high-performance, minimal coding agent designed to be fast and lightweight while maintaining high reliability even when using cheaper or weaker LLMs. It addresses the common issues of agentic editors—such as high memory consumption, context window degradation, and the loss of session continuity—by implementing a robust agent loop and a tiered memory system.

How it works

Dirge is written in native Rust for speed and low resource usage. It uses a "robust agent loop" to repair malformed tool calls and validate writes via tree-sitter before they touch the disk. Instead of vector embeddings, it relies on inline grep and SQLite FTS5 for search, which it argues is more effective for long-term conversational memory.

To manage context, it uses a bounded budget (default 250k tokens) and a two-tiered memory system: salient entries stay inline, while others are demoted to a breadcrumb index. A post-session orchestrator extracts learnings into persistent project memory and skills. It also integrates with LSP diagnostics and tree-sitter for code intelligence, and can be exposed as an MCP server for delegation from other agents.

Who it’s for

Developers who want a fast, native coding agent that supports a wide variety of LLM providers (including local models via Ollama) and prefers a low-overhead, non-JS based tool.

Highlights

  • Extreme Efficiency: Native Rust binary with very low RAM usage (approx 8-15MB) compared to JS-based agents.
  • Robustness: Validates code writes with tree-sitter and implements circuit breakers for non-progressing loops.
  • ** uma Permission Engine**: A single Policy Decision Point with a /why command to trace authorization decisions.
  • Multi-Provider Routing: Ability to route different roles (main loop, review, summarization) to different models in one session.
  • Project Memory: Persistent per-project memory and a built-in agent-facing kanban issue board.
  • Multimodal Support: Supports clipboard image pasting for vision models.
  • Janet Plugin System: Extensible at runtime via Janet plugins and Claude-compatible skills.

Related

  • Project
    gptme/gptme

    A provider-agnostic personal AI agent that runs in the terminal, providing LLMs with shell, Python, and web access to assist in coding and autonomous knowledge work.

  • Dispatch
    Juggler: An Open-Source GUI Coding Agent

    Juggler is an open-source AI coding agent that replaces linear chat histories with a branching session tree and a plugin-based architecture for high-visibility tool calls and editable context.

  • Project
    SII-Holos/synergy

    An open-source workspace for AI agents that provides persistent, recoverable sessions and a unified runtime for software and knowledge work.

  • Project
    org2AI/ORG2

    ORG‑2 is a Rust‑based desktop tool that records the entire workflow of AI coding agents (prompts, tool calls, file edits, etc.), lets teams replay sessions like a video, and provides “AI‑blame” linking every line of code back to the originating agent decision. It supports 20+ agent CLIs, offers collaboration features, a built‑in terminal/Git workspace, and optional browser automation. Distributed as native installers for macOS, Windows and Linux, it is AGPL‑3.0 licensed.

  • Project
    MigoXLab/dingo

    Dingo is an open‑source Python library (and SaaS product) for evaluating AI data, model outputs, and RAG systems. It supports rule‑based, LLM‑based, and agent‑driven checks, streams data from files, databases, S3, or Hugging Face, and provides JSON reports plus optional visual dashboards. Install via `pip install dingo-python` (with optional extras), run via CLI (`dingo eval …`) or Python SDK, and extend with custom rules, LLM prompts, or agents.