Saotri Bench — coding benchmark for evaluating LLM agents on multi-phase programming tasks with hidden requirements.

Raw logs of Claude Code running on local Qwen3.5-27B (llama.cpp). Builds a Python todo app with 50 tests. Real-world performance data: 30 min, cache thrashing, 38 t/s generation.

🔍 Evaluate LLM agents on multi-phase programming tasks with FluxCodeBench, focusing on hidden requirements, long-context retention, and iterative refinement.