Token Beggar 0.3.0: on 18 real bug fixes it cost 1.42x less, with a tie on tests
说明与用途
Token Beggar 0.3.0 将 Python 项目中的错误修复成本降低 1.42 倍,同时保持相似的解决率,但并未加快执行速度。
问题
Running the largest model for every step of a code‑fix task is expensive; a cheaper model could handle many tasks without sacrificing quality.
结果
Token Beggar 0.3.0 lowered the cost per bug‑fix task by 1.42× (from $0.69 to $0.49) across 18 hard tasks, with issue resolution remaining at 69% versus 70% and runtime unchanged.
工作流
- 1. Plugin intercepts a task request.
- 2. It estimates task size based on repository context.
- 3. Small tasks are handled by the chat model directly.
- 4. Large tasks are sent to a cheaper worker model for reproduction, exploration, fixing, and test execution.
- 5. The worker returns a diff and test results.
- 6. The chat model reviews the diff, re‑runs tests, and if needed, sends the task back to the worker for further refinement.
- 7. Final diff is applied to the repository.
证据
Without Token Beggar is the chat model working alone. With Token Beggar is the card shipped in 0.3.0.
Cost per task went down about 29 percent.
The share of issues resolved is level, and 69% against 70% is a tie within the spread.
Time is level too (168 seconds against 187), so it is not faster.
The raw runs and the task list are in the repo, so you can check every number.
- 记录日期: 2026-10-23T19:11:25.412Z