FeatureArc (whole-feature chain gate)
Code · 2026-10-27
73 manually reviewed feature-evolution tasks from Tencent's Hunyuan team, mined from the histories of 53 repositories in Python, TypeScript, Rust and Go. Each task is a chain of 2 to 10 dependent stages (263 in all) that extend one user-visible feature. A coding agent receives each stage's requirements document and interface contract in turn, in one persistent session and workspace, with a four-hour stage budget and no hidden-test feedback or rollback between stages. The score is the share of tasks whose chain-level fail-to-pass and pass-to-pass suites both pass on the final workspace, in percent.
Top models (higher is better)
| Model | Score |
|---|---|
| Claude Opus 5.5 | 49.3 |
| GPT-6 Astra | 37.0 |
| GLM-5.3 | 31.5 |
| Grok 4.7 | 26.0 |
| Gemini 3.8 Flash | 23.3 |
| Kimi K3 | 19.2 |
| DeepSeek V4 Pro 0813 | 17.8 |