CoreCoder

About

Production coding agents bury their core loop under tens of thousands of lines, which makes it hard to learn how harnesses actually work. CoreCoder reimplements that core in about 1,100 lines of readable Python: a bounded agent loop with parallel tool execution, eight tools (bash, file read/write/edit with unique-match search-replace, glob, grep, a todo list, and a non-recursive sub-agent spawner), three-tier context compaction, streaming with retry and cost accounting, and session save/resume. It speaks any OpenAI-compatible API out of the box and reaches a hundred-plus providers through an optional LiteLLM backend. Deliberate omissions - no real sandbox, no MCP, no fallback model chain - are documented as fork entry points rather than defects. Students and harness authors read the source (aided by an eight-part bilingual essay series) and fork it as a starting point.