Keep the model fixed, change the harness, and coding-agent results can move a lot when context gets tight.

The study compares 2 configurations of Yuj: control keeps the full chronological transcript until context fills, while treatment shortens older tool outputs, detects stalled behavior, and applies fixed command safeguards while preserving the full record.

On 169 SWE-bench Verified tasks with a 20,480-token window, Qwen3.6’s mean per-task F2PF rose from 28% to 49%, while complete solutions increased from 43 to 72.

The same frozen treatment improved both outcomes for Devstral, Nemotron, and Qwen3.8 without retuning.

At 262,144 tokens, however, Verified and Pro outcomes were nearly identical between arms, so the benefit appears strongest when context is actually binding.

Treatment also used more model work under pressure, and the experiment tests the package as a whole rather than isolating each mechanism.

For coding agents, “which model?” is no longer enough. Benchmark the model, harness, context policy, tools, and run controls as one solver.