
AI Coding Tools – Architecture, Local LLMs and SWE-Bench
A curated, tiered catalog of state-of-the-art AI coding tools, open foundation models, and foundational research papers. Structured across three progressive skill tiers - from zero-setup beginner extensions to autonomou…
以下正文同步自 Hacker News 最新,版权归原站所有,已转换为易读排版。
A curated, tiered catalog of state-of-the-art AI coding tools, open foundation models, and foundational research papers. Structured across three progressive skill tiers - from zero-setup beginner extensions to autonomous SWE-bench systems - so every developer and researcher can immediately find what they need.
💡 Sister Repository: Looking for academic research, paper writing, and literature discovery tools? Check out Awesome AI Research Tools.
⭐ Why Star This Repository?
🗺 Choose Your Path: Interactive Roadmap
🚀 60-Second Quickstart (Start Here)
🌱 Level 1: Beginner (Plug-and-Play & In-Editor)
Standalone Agentic IDEs
1-Click Editor Extensions
Beginner Prompting Cheatsheet
⚡ Level 2: Intermediate (CLI Agents & Local LLMs)
Terminal & CLI Pair Programmers
Running Agents in Parallel
Top Open-Weight Coding Models
Local Execution Engines (Private & Offline)
Automated Testing, QA & Code Review
Step-by-Step Local Offline Recipe
🔬 Level 3: Expert (Autonomous Systems & Research)
Autonomous Software Engineers (SWE-bench)
System Architecture & Context Assembly
Tree-sitter & AST PageRank Repo-Maps
Code Editing Representation Trade-offs
SOTA Benchmark Leaderboard
Seminal Academic Papers
⚖️ Comprehensive Multi-Tier Comparison Matrix
📖 BibTeX Citations
📚 Detailed Guides & Documentation
🤝 Contributing
⭐ Why Star This Repository?
Zero Hype, Technical Rigor: Every tool and architecture is evaluated by empirical capabilities, local model support, and benchmark metrics.
Progressive 3-Tier Journey: Start with single-click autocomplete today; level up to local terminal agents and autonomous benchmark systems when ready.
Academic & Industry Bridge: Combines practical engineering tools with peer-reviewed research papers (ICLR, EMNLP, NeurIPS, ICML).
100% Privacy Focused: Dedicated local offline runbooks ensuring your proprietary code never leaves your workstation.
🗺 Choose Your Path: Interactive Roadmap
flowchart TD Start["Where are you in your AI coding journey?"] --> Q1{"What is your goal?"}
Q1 -->|"I just want fast code completion in my editor"| L1["🌱 Level 1: Beginner<br/>Cursor, Windsurf, Continue.dev, Codeium<br/>• Zero terminal setup<br/>• Instant ghost-text autocomplete<br/>• In-editor sidecar chat"]
Q1 -->|"I want terminal tools & 100% local privacy"| L2["⚡ Level 2: Intermediate<br/>Aider, Cline, Ollama, Qwen2.5-Coder<br/>• Terminal git workflows<br/>• Multi-file automated edits<br/>• Local private offline models"]
Q1 -->|"I build agents, benchmarks, or compilers"| L3["🔬 Level 3: Expert<br/>SWE-agent, OpenHands, Tree-sitter AST, Repo-Maps<br/>• Autonomous issue resolution<br/>• AST PageRank retrieval<br/>• Empirical research & benchmarks"]
style L1 fill:#e1f5fe,stroke:#0288d1,stroke-width:2px style L2 fill:#fff3e0,stroke:#f57c00,stroke-width:2px style L3 fill:#ede7f6,stroke:#7e57c2,stroke-width:2px
Loading
🚀 60-Second Quickstart (Start Here)
If you have never used an AI coding tool before, start here:
Zero configuration required. Install and start coding within 2 minutes.
flowchart LR A["Your Code Editor<br/>(VS Code / Cursor / Zed)"] --> B["Inline Autocomplete Engine<br/>(Predicts Next 1-3 Lines)"] B --> C["Press Tab to Accept<br/>(Instant Productivity)"]
Loading
Standalone Agentic IDEs
Cursor - The most popular AI-native editor. A fork of VS Code with instant codebase indexing, multi-file edits (Composer), and inline terminal error fixing.
PearAI - Open-source alternative to Cursor built on VS Code with transparent model routing and customizable backends.
Windsurf - Next-generation agentic IDE by Codeium featuring "Flows" that track real-time developer context and synchronized multi-step edits.
Zed - Lightning-fast, GPU-accelerated code editor written in Rust with deep model integration and multi-model assistant panels.
1-Click Editor Extensions
Codeium - Free AI code completion and chat extension for 40+ IDEs with low latency and enterprise self-hosting options.
Continue.dev - The leading open-source extension for VS Code and JetBrains; connects to cloud models or local models (Ollama, LM Studio).
Supermaven - Ultra-low latency inline code completion with a 1-million-token context window that indexes your entire repository in seconds.
Tabby - Self-hosted AI coding assistant server; an open-source, private alternative to GitHub Copilot.
Beginner Prompting Cheatsheet
To get high-quality code suggestions, use the 3-Part Prompt Formula:
For developers who want git-integrated terminal workflows, local private execution, and automated testing loops.
sequenceDiagram autonumber actor Dev as Developer participant Agent as CLI Agent (Aider / Cline) participant Repo as Local Git Repository participant Shell as Terminal Sandbox
Dev->>Agent: "Add input validation to auth route and run tests" Agent->>Repo: Read modified files and git status Agent->>Agent: Generate diff patch Agent->>Shell: Run test suite (pytest / npm test) alt Tests Pass Shell-->>Agent: Exit code 0 Agent->>Repo: Commit atomic git diff with description Agent-->>Dev: Patch verified & committed else Tests Fail Shell-->>Agent: Test error traceback Agent->>Agent: Self-heal & adjust code Agent->>Shell: Re-run test suite end
Loading
Terminal & CLI Pair Programmers
Aider - Command-line AI pair programmer that parses your repository into a Tree-sitter map, edits multiple files, runs lint/test commands, and automatically commits atomic git diffs.
Avante.nvim - Neovim plugin designed to emulate Cursor AI's multi-file editing capabilities natively in Lua.
Cline - Autonomous coding agent extension for VS Code that executes terminal commands, inspects local browser previews, and requests human-in-the-loop permission.
Mentat — Open-source AI tool capable of coordinating complex git workflows directly in the terminal with repo-wide context.
Xenon — CLI coding agent that records every session event in an append-only hash chain verifiable on demand via /evidence, enforces deterministic hollow-answer checks, and exports manual verification checklists with multi-model and MCP client support.
Running Agents in Parallel
Control planes, terminal managers, and orchestrators that run multiple coding agents concurrently across worktrees:
Codestral - Open-weight model specialized in code completion and fill-in-the-middle tasks with an 80-language vocabulary.
DeepSeek-Coder-V2 - Mixture-of-Experts code language model with 128k context length supporting 338 programming languages.
Qwen2.5-Coder - Leading open-source coding foundation model series (0.5B to 32B) competitive with leading closed models across code generation, completion, and multi-file reasoning.
StarCoder 2 - Open, transparently trained code models (3B, 7B, 15B) curated by BigCode under permissive licenses.
Local Execution Engines (Private & Offline)
llama.cpp - High-performance C/C++ inference engine for quantized GGUF models with minimal hardware overhead.
LM Studio - Graphical desktop application for running local GGUF models on Apple Silicon and NVIDIA GPUs.
Ollama - The easiest way to get up and running with local models via a single terminal command (ollama run qwen2.5-coder).
vLLM - High-throughput, memory-efficient LLM serving engine featuring PagedAttention for production local deployments.
Automated Testing, QA & Code Review
CodeRabbit - AI-driven pull request reviewer providing line-by-line feedback, sequence diagrams, and security vulnerability checks.
Cover-Agent - Open-source generative testing tool that iteratively generates unit tests until target coverage is met.
Keploy - Open-source zero-code test generator that captures real network calls and creates automated regression test suites.
Qodo (CodiumAI) - Comprehensive code integrity platform analyzing PRs, writing regression tests, and enforcing code standards.
Semgrep Assistant - Combines deterministic static analysis (AST rules) with LLM explanations to eliminate false-positive security findings.
Step-by-Step Local Offline Recipe
Run a full-featured coding agent on your local machine with zero cloud fees and complete data privacy:
1. Install and start Ollama
curl -fsSL https://ollama.com/install.sh | sh
2. Pull the 14B parameter coding model
ollama run qwen2.5-coder:14b
3. Install Aider in your Python environment
pip install aider-chat
4. Launch Aider in your git repository using your local model
cd /path/to/your/project aider --model ollama/qwen2.5-coder:14b
🔬 Level 3: Expert (Autonomous Systems & Research)
For AI researchers, systems engineers, and teams building custom autonomous coding agents.
Devika - Agentic open-source software engineer capable of breaking down user goals into multi-stage tasks and executing web research.
OpenHands (formerly OpenDevin) - Autonomous software development agent capable of writing code, browsing the web, and running dockerized execution sandboxes.
SWE-agent - Open-source agent developed by Princeton that resolves real GitHub issues on the SWE-bench benchmark using an Agent-Computer Interface (ACI).
System Architecture & Context Assembly
graph TD subgraph Context["1. Context Ingestion Layer"] NL["User Task / Issue Prompt"] AST["Tree-sitter AST Graph"] LSP["Language Server Protocol (LSP)"] GIT["Git Commit History & Diffs"] end
subgraph Core["2. Orchestration Core"] PM["Prompt Context Packer"] FIM["Fill-in-the-Middle (FIM) Engine"] LLM["Foundation Model (Local / Cloud API)"] TOOL["Tool & Function Router"] end
subgraph Sandbox["3. Execution & Validation Sandbox"] PATCH["Patch Engine (Diff / Replace)"] SHELL["Terminal / Bash Sandbox"] TEST["Test Runner (pytest / cargo / jest)"] REFLECT["Reflexion / Self-Debugging"] end
subgraph Effectors["4. Output Interfaces"] IDE["Agentic IDE (Cursor, Windsurf)"] CLI["Terminal CLI Agent (Aider)"] PR["PR Review Bot (CodeRabbit)"] end
NL --> PM AST --> PM LSP --> PM GIT --> PM PM --> LLM FIM --> LLM LLM --> TOOL TOOL --> PATCH PATCH --> SHELL SHELL --> TEST TEST -- "Traceback Error" --> REFLECT REFLECT --> PM TEST -- "Success (Exit 0)" --> Effectors PATCH --> IDE PATCH --> CLI PATCH --> PR
Loading
Tree-sitter & AST PageRank Repo-Maps
How modern coding agents solve the repository-scale context problem without overloading token budgets:
flowchart TD Src["Source Code Repository<br/>(100+ Files, 100k+ LoC)"] --> TS["Tree-sitter AST Parser"] TS --> Extract["Extract Symbols<br/>(Classes, Functions, Methods, Imports)"] Extract --> CallGraph["Construct Directed Dependency Graph"] CallGraph --> PR["Run Personalized PageRank<br/>(Biased towards actively edited files)"] PR --> Rank["Rank Top-K Informative Signatures"] Rank --> Budget["Token Budget Packing<br/>(Fits 1,024 - 4,096 tokens)"] Budget --> Prompt["Inject into System Context<br/>('Repo Map')"]
Loading
Code Editing Representation Trade-offs
graph TD subgraph WholeFile["Whole-File Rewrite"] W1["Model emits full file (1,000+ lines)"] W2["High latency & high token cost"] W3["Prone to truncation & syntax loss"] end
subgraph SearchReplace["Search & Replace Blocks"] S1["SEARCH block with original lines"] S2["REPLACE block with modified lines"] S3["Robust, token-efficient, fast execution"] end
subgraph UnifiedDiff["Unified Diff (diff -u)"] U1["Line-numbered hunk headers (@@ -12,4 +12,6 @@)"] U2["Ultra-compact token footprint"] U3["Requires strict line arithmetic (High failure rate on smaller LLMs)"] end
Loading
SOTA Benchmark Leaderboard
Summary of state-of-the-art results on the standardized SWE-bench Verified benchmark (500 curated real-world GitHub issues):
- Benchmarks & Real-World Evaluation
SWE-bench: Can Language Models Resolve Real-World GitHub Issues? (Jimenez et al., ICLR 2024)
Introduced the benchmark evaluating LLMs on 2,294 real-world GitHub issues acros
正文由 探How 从来源站点 RSS 同步,内容未经改写;遇到排版缺失或需要图片、视频时请以原文为准。