swival

About

Swival exists because most coding-agent CLIs assume frontier models with large context windows, leaving users of small or locally hosted models with tools that degrade quickly. It is a single pure-Python agent loop with a deliberately small tool set (read, write, edit, bash) and no framework dependency, and its engineering centers on context discipline: graduated compaction that summarizes in stages, persistent thinking notes that survive context resets, todo checklists that are re-injected after compaction, and cross-session BM25 memory so resumed work recalls earlier decisions. Connectivity is unusually broad for its size — LM Studio and llama.cpp auto-discovery, OpenRouter, Gemini, Bedrock, Vertex, Apple Foundation Models, browser-authenticated ChatGPT subscriptions, and any OpenAI-compatible server — plus a command provider that shells out to external agents like codex exec. Lifecycle hooks and command middleware support unattended or policy-constrained operation, A2A/ACP server modes let other harnesses call it, and scheduled runs suit automation. It targets developers running local or low-cost models who still want durable, multi-session agent behavior.