qf-studio/navigator

Finish What You Start — Context engineering for Claude Code. Sessions last 20+ exchanges instead of crashing at 7.

What it solves

Navigator addresses the "context window crash" common in AI-assisted development. When developers load massive amounts of documentation and code into an LLM's context, the session often becomes unstable, leading to hallucinations or crashes after a few exchanges. Navigator prevents this by implementing "context engineering," loading only the specific information needed for the current task rather than everything upfront.

How it works

It operates as a plugin for Claude Code, using a runtime that manages how prompts are gated and how information is retrieved. Instead of a bulk load, it uses a dynamic loading system that significantly reduces token usage (claiming up to 92% savings).

Key technical features include:

  • Theory of Mind: Remembers user preferences and learns from corrections across sessions.
  • Knowledge Graph: A persistent project memory that stores patterns, decisions, and SOPs.
  • Loop Mode: A structured completion mode that runs until a task is finished, featuring stagnation detection to prevent infinite loops.
  • Runtime Enforcement: A system that gates prompts and guards reads to ensure the efficiency of the context window.
  • Reply Modes: Customizable response formats (e.g., "ADHD mode") to suit different user cognitive preferences.

Who it’s for

Developers using Claude Code who want longer, more stable AI sessions that can complete complex features without needing to restart the session due to context limits.

Highlights

  • Context Engineering: Reduces token load from 150k to 12k in typical scenarios, extending session length from ~7 to 20+ exchanges.
  • Live Monitoring: A /nav pane providing real-time data on context fill, session cost, and token usage.
  • Project Memory: A knowledge graph that persists across sessions for experiential memory.
  • Loop Mode: Agentic task completion with progress tracking.
  • Web Deep Research: Cited research capabilities with integration into graph memories.
  • Customizable Replies: Specialized modes like ADHD and STE to shape how the AI communicates.

Related

  • Project
    google-research/era

    ERA is an AI system that helps scientists write expert-level empirical software by combining LLMs with a tree-search algorithm to iteratively generate and score candidate programs.

  • Project
    clay-good/OpenLore

    OpenLore builds a deterministic, local-first knowledge graph of your codebase so AI coding agents start every task oriented and get their changes certified before landing — no API key, no LLM in the hot path.

  • Project
    0xku/kon

    A minimal coding agent harness with a tiny core prompt and a small built-in toolset, designed for lightweight and adaptable AI-assisted development.

  • Project
    shengjidaguai-china/personal-homepage-skill

    Personal Homepage Skill is an AI‑agent plug‑in that lets Claude‑style agents generate high‑quality personal websites, portfolios, and HTML slide decks using a library of 18 ready‑made visual templates. It includes a React/Tailwind gallery, single‑file HTML templates, and a documented workflow that enforces reference‑first design and quality checks. Install as a Claude skill or run the gallery locally to pick a style, then export a share‑ready HTML page.