AI Coding Tools – Architecture diagrams, local models, and SWE-bench

A curated catalog of state-of-the-art AI coding assistants, terminal agents, autonomous software engineering architectures, local code foundation models, and bridging research argues that modern AI-augmented software engineering progresses through five autonomy tiers - from single-line autocomplete to fully autonomous issue resolution exemplified by SWE-bench. It defines a full system taxonomy with four layers: context ingestion (prompts, AST, LSP, git), orchestration and model core (prompt packer, fill-in-the-middle engine, LLM, tool router), execution and validation sandbox (patch engine, shell, test runner, reflexion), and user-facing effectors (agentic IDEs, CLI agents, PR bots). Visual workflows emphasize closed-loop agent execution (plan, apply multi-file patches, run tests, self-refine) and repository-level context retrieval using tree-sitter ASTs plus personalized PageRank for token-budgeted prompt packing. Patch-edit paradigms are compared - whole-file rewrites, search-and-replace blocks, and unified-diff hunks - highlighting tradeoffs in token cost, robustness, and failure modes.

Concrete tooling and evaluation resources are mapped to that architecture: agentic IDEs like Cursor, Windsurf, Zed, PearAI; terminal agents such as Aider, Claude Code, Cline, Mentat; open-source copilots and editor extensions including Continue.dev, Avante.nvim, Codeium, Tabby; autonomous engineers like OpenHands, SWE-agent, Devika; and local/open-weight coding models such as Qwen2.5-Coder and DeepSeek-Coder-V2. The catalog also consolidates automated testing and security tooling, benchmark leaderboards (SWE-bench, HumanEval), architecture diagrams, a local offline developer recipe, a comprehensive feature matrix, and bibliographic citations to support implementation and rigorous evaluation of agentic, privacy-conscious coding workflows.