mini-harness: 1,700 lines of Python that scores 80.2% on SWE-bench Verified with DeepSeek V4 Flash

GitHub·medium signalXBlueskyLinkedInCopy link

Published 5 September, mini-harness is a deliberately small but complete agent harness: nine tools defined with Pydantic, context compaction, request retries, streaming, session memory, and a single-file TUI. Self-reported results with DeepSeek V4 Flash are 401/500 on SWE-bench Verified (80.2%) and 309/445 on Terminal-Bench 2.1 (69.44%), the latter placing around 13th of 18 public entries. The numbers are self-reported and unaudited, but as a readable baseline for anyone building their own loop it is a better starting point than reading a production harness.