🌟 Overview

NimCode is your autonomous coding pair-programmer. Built with a rich interactive terminal interface, it doesn't just autocomplete code—it plans architectures based on real documents (SDD, PRD, RFC), executes terminal commands, writes complete files, and manages your workspace.

✅ What Works Right Now

• OS-Aware Agent: Automatically detects Windows/Linux/macOS and uses the correct shell commands. No more mkdir -p on Windows.
• SDD/PRD-Aware Planning: /plan mode reads your actual documents and produces implementation plans grounded in real file names, modules, and requirements—not generic PM templates.
• Paginated File Reading: Large files (like SDDs) are read in chunks with offset/limit so nothing gets truncated.
• VS Code Deep Integration: Native IPC communication allowing NimCode to read active editors, send patches, and act as your intelligent coding panel within VS Code.
• RAG & Semantic Search: Zero-dependency TF-IDF/BM25 based indexer that instantly scans massive workspaces and retrieves context-aware code snippets.
• Autonomous Auto-Fixer: The /fix command creates a self-healing feedback loop—runs your broken commands, analyzes tracebacks, and automatically iterates on code patches until the build passes.
• Multi-Model & Local Support: Flexible api_base_url configuration to seamlessly switch between NVIDIA NIM, Ollama, or vLLM endpoints. Dynamic model list fetched from the API at runtime.
• Smart Permission Engine: Granular control over file writes and command executions, featuring an auto-bypass mechanism and interactive (a)ccept / (r)eject diff previews. Non-interactive/CI mode is secure by default (Bash denied unless explicitly allowed).
• Robust Interactive REPL: Multiline support (Alt+Enter), syntax highlighting, and beautiful rich-powered UI menus.
• Automated Workflows: Automatic linting and formatting using black, flake8, and prettier behind the scenes.
• Model Context Protocol (MCP): Native integration for MCP tools allowing infinite extensibility.
• Context Management: Auto-compact with per-model context window awareness. /thinkback shows real session statistics.
• Live Sync: Workspace watcher automatically updates the repo map in the agent's context when files change.

🚀 What We're Working On (Roadmap)

• Sub-agent Swarms: Perfecting /delegate and /swarm to distribute complex tasks among specialized AI roles.
• Advanced Diagnostics: Refining /bughunter, /security-review, and /doctor for automated codebase auditing.
• Test-Driven Development: Upgrading /tdd and /testgen to automatically write tests and ensure 100% coverage before committing.
• Infrastructure & DevOps: Implementing /terraform-god and /sql-tune for automated cloud provisioning and database query optimization.

📦 Installation & Update

# First time install
pip install nimcode

# Update to latest version
pip install --upgrade nimcode

⚠️ If NimCode shows an old version number at startup, you likely have an older install. Run pip install --upgrade nimcode to get the latest fixes.

🚀 Quick Start

1. Get an API Key: Grab a free NVIDIA API key from build.nvidia.com.
2. Login: Connect your local environment by running:

   nimcode login
   

3. Start Coding:
   • Launch the interactive REPL:

     nimcode
     

   • Or run a one-off task directly from the command line:

     nimcode /plan "Build a classic Snake game using HTML5 Canvas"
     

💻 Complete Command Reference

Inside the nimcode REPL, you can type natural language or use any of the following slash commands:

Mode Toggles & Execution

• /code: Enter standard coding mode (default mode, prompts for dangerous actions).
• /plan: Enter planning mode. NimCode will read your actual documents/files first, then write a concrete implementation plan to .nimcode/plans/.
• /trust: Enter trust mode. The AI will run commands and edit files completely autonomously without asking for (a)ccept / (r)eject. Turn limit removed.
• /untrust: Disable trust mode and restore permission prompts.

Interface & Settings

• /models: Open an interactive UI to select your preferred NVIDIA NIM model.
• /theme: Open an interactive UI to change your syntax highlighting theme.
• /config: View and edit global NimCode configuration settings.

Workspace & Context Management

• /clear: Clear the current conversation history to reset context.
• /compact: Compact the context window to save tokens while keeping essential memories.
• /context: View and manage the loaded context window.
• /rewind: Rewind the conversation history by a few steps.
• /undo: Revert the last file modification made by the AI.
• /thinkback: View real session statistics (turns, token estimates, message history).

Project & Search

• /index: Index the current project files to enable lightning-fast Semantic Search.
• /map: Generate a high-level semantic architecture map of the codebase.
• /research: Enter deep-research mode for reading documentation or large files.

Automation & Git

• /commit: Analyze all staged git changes and automatically generate a conventional commit message.
• /fix: Run a specific shell command and let the AI automatically iterate to fix any resulting errors.
• /autofix-pr: Pulls the current active GitHub Pull Request, reads comments, and automatically fixes the mentioned issues.

Advanced Development Modes

• /testgen <file>: Generate unit tests for a specific file, aiming for 100% coverage.
• /tdd: Enter Test-Driven Development mode. The AI will write tests first, verify they fail, and then implement the code to pass them.
• /bughunter: Initiate an automated search for logical bugs and edge cases across the codebase.
• /security-review: Audit the codebase for common vulnerabilities (e.g., OWASP top 10).
• /ultraplan <task>: Generate a master ultra-detailed, dependency-graphed execution plan.

AI & Agents

• /delegate <role> <task>: Spawn an independent sub-agent with a specific role to handle a background task.
• /swarm: Orchestrate multiple sub-agents to tackle a complex architectural epic simultaneously.
• /grill-me: Interrogation mode. The AI will ask you hard questions to refine your system design and edge cases.
• /learn: Teach NimCode a new persistent skill or framework rule that it will remember for future sessions.
• /vision: Capture the screen and analyze UI elements using Vision AI models.
• /mcp install: Install and configure new Model Context Protocol tools.

⚙️ How it Works

NimCode creates a .nimcode directory inside your projects. This directory acts as the agent's personal workspace, not your project's source code:

• .nimcode/plans/: All generated step-by-step markdown plans live here.
• .nimcode/skills/: Custom guidelines, framework rules, or memories you teach the agent.
• .nimcode/history/: File backups for /undo capabilities.

Important: NimCode writes your actual source code to your project root, not inside .nimcode/. The .nimcode/ folder is only for the agent's internal notes.

Advanced Configuration

NimCode's behavior can be fully customized through ~/.nimcode/settings.json (global) or .nimcode/settings.json (per-project). Setting any timeout to 0 disables it (infinite).

{
    "model": "meta/llama-3.3-70b-instruct",
    "api_base_url": "https://integrate.api.nvidia.com/v1",
    "timeout_command": 1200,
    "timeout_llm": 120,
    "timeout_format": 10,
    "timeout_browser": 15000,
    "timeout_updater": 3,
    "max_turns": 200,
    "max_tokens": 120000,
    "max_retries": 15,
    "retry_base_delay": 2.0,
    "retry_max_delay": 60.0,
    "allow_bash_non_interactive": false
}

📋 Changelog

v0.8.8 (V2 Architecture Update)

• 🌐 Network Resilience: Added Exponential Backoff mechanism for rate-limiting (429) errors in nim_client.py.
• 🔄 Self-Correcting Auto-Linter: Integrated cargo check for Rust and tsc --noEmit for TypeScript into the agent's feedback loop to catch errors before assuming tasks are complete.
• 🧠 AST-Based Repo Map: repo_map.py now extracts classes and functions from files via the ast module, optimizing the context window for large codebases.
• 🤖 Multi-Agent Swarm: Introduced SubAgent class and DelegateTask tool, allowing the main agent to spin up background agents for subtasks.
• 🔌 Dynamic MCP Expansion: MCPManager now supports adding and removing MCP servers at runtime without restarts, natively powered by the CallMCP tool for high reliability.

v0.8.6

• 🚀 Default Model Bump: Made deepseek-ai/deepseek-v4-pro the default model across the entire agent for superior reasoning and tool performance.

v0.8.4

• 🧠 Native DeepSeek XML Support: Re-engineered the internal parser (lenient_parser.py) into a hybrid parser that seamlessly handles both strict JSON (<tool_call>) and fine-tuned XML formats (<tool_call name=\"...\">). DeepSeek V4 Pro, Mistral Large 2, and others now work flawlessly out of the box!
• 🧪 224 tests passing

v0.8.0

v0.7.0

• 🛡️ Strict Markdown Validator: /plan mode now physically blocks tasks that are under 150 characters or missing required Markdown headers (Target Files, Implementation Details, Checklist).
• 🧠 Context Amnesia Shield (AST Diff): When modifying Python files, the ReplaceBlock tool compares the Abstract Syntax Tree (AST) before and after. If the LLM accidentally deletes over 30% or 2+ top-level functions (truncation hallucination), it instantly blocks the write and forces a fix.
• 🔗 Task Linearity Guard: TaskManager enforces strict order. The agent cannot start Task 1.2 until Task 1.1 is fully completed.

v0.6.0

• 🛡️ Anti-Laziness AST Blockers: Native AST checking explicitly intercepts pass blocks, empty classes, and lazy placeholder comments (like // TODO), forcing the LLM to write full implementations.
• 🕵️ Adversarial QA Gate: Built a multi-agent validation layer (qa_agent.py) that strictly critiques the main agent's work. The main agent cannot complete tasks without receiving a VERDICT: PASS from the QA Agent.
• 💾 Task State Machine: Replaced in-memory lists with .nimcode/tasks.json persistent storage via TaskManager, allowing robust restartability and task state tracking (pending, in_progress, completed).

v0.5.5

• 🐛 UI Bug Fix: Fixed a visual bug where the CLI welcome banner hardcoded NimCode v0.2.0 despite the user having updated to a newer version.

v0.5.4

• 🐛 Bug Fix: Fixed a bug where the Read tool would crash with a TypeError if an LLM provided the offset and limit arguments as strings (e.g. '1') instead of integers.

v0.5.3

• 🧭 Phase/Task Plan Architecture: /plan mode now forces the AI to output plans with meticulous Phase -> Task -> Checklist hierarchy (inspired by Claude Code), fixing superficial plan generation.

v0.5.2

• 🧠 Autonomous Marathon Runner: Capable of running non-stop for 5-6 hours without drifting from the main plan.
• 🧪 Native TestRunner: NimCode can now run your test suites (pytest, npm test, go test) and self-heal its code until the tests pass.
• 🗺️ Semantic AST Explorer: GetCodeOutline tool added for instant navigation within large files (functions/classes with line numbers) instead of reading thousands of lines.
• 🛡️ Stuck-Loop Breaker: AI no longer gets stuck repeating the same failing action. Auto-detects 3x repeated errors and forces a strategy change.
• 🧭 Plan-Drift Radar: Automatically pings the agent every 15 turns to ensure strict compliance with the .nimcode/active_plan.txt and prevents hallucinations.
• 🚨 Anti-Laziness Firewall: Advanced regex intercepts lazy code (e.g., // TODO, pass) before it's written and forces the AI to output complete logic.
• ✨ Self-Correcting Auto-Linters: Seamless multi-language support. Formats and validates code in the background using go vet, prettier, and python syntax checks.
• 🧠 Semantic Context RAG: Automatically pins the directory tree and database schemas into the system prompt to cure agent amnesia.
• 📝 ReplaceBlock Tool: Precision line-based editing replaces the buggy exact-string Replace tool, completely solving indentation mismatch errors.
• ✅ 195 tests passing

v0.4.0

• 🐛 System prompt is now OS-aware: Correctly uses Windows (mkdir, copy) or Unix (mkdir -p, cp) commands based on the detected OS
• 🐛 Plan quality massively improved: /plan mode now reads your actual documents (SDD, PRD, etc.) and generates concrete plans with real file paths and code — not generic PM templates
• 🐛 Fixed _distill_memory 400 error: Context compaction no longer crashes with a Bad Request error
• 🐛 Fixed version tracking: CURRENT_VERSION was hardcoded as 3.0.0; now reads from a single source of truth
• 🐛 /thinkback shows real data: Previously showed hardcoded fake table; now shows actual session statistics
• 🐛 LiveSync fixed: repo_map module was missing, causing the workspace watcher to silently fail
• 🔒 Security: Non-interactive mode no longer auto-approves Bash commands
• ⚙️ Fully configurable: max_turns, max_tokens, max_retries, retry delays all configurable via settings
• ⚙️ Dynamic model list: Fetched from NIM API at runtime with fallback to known models
• ⚙️ Paginated file reading: Read tool supports offset/limit for large files

v0.3.4

• All timeouts configurable via settings (0 = infinite)

🛡️ Requirements

• Python 3.8+
• An NVIDIA NIM API Key (NIM_API_KEY)

🤝 Contributing

Contributions, issues, and feature requests are welcome! Feel free to check out our issues page.