Made my Codex limits last almost ~3x longer with one change
热度趋势
趋势数据积累中
百分比基于当前可用热度信号，而非评论数或独立用户人数。
一位开发者通过一项基于微软FastContext研究论文的定制解决方案，使其Codex使用时长延长了近三倍。该方案涉及构建一个带有自定义路由器的MCP Rust工具，并经过数周的基准测试进行调优。这种方法使得昂贵的Sol模型可以专注于实际编码任务，而将搜索任务转移给成本更低的Luna模型处理。此举解决了Plus用户被迫依赖Luna来提高生产效率的担忧。
Plus users are basically being forced to give up Sol and just use Luna to get any usable amount of work done. That's a huge downgrade basically using a deepseek flash model level which you can get for free in opencode anyway.
I started tracking where Sol's spent most of its tokens and most of it was searching around your repo looking for what to edit. Search took 30-60% of the total cost.
so I tried to come up with a solution that both lowers search cost and at the same time keep the same output quality from sol.
Then I found this Microsoft research paper that I based this on called FastContext. The concept doesn't work directly in Codex so I built my own implementation as an MCP Rust tool with a custom router I tuned and improved over weeks and weeks of benchmarking. Sol still does all the actual coding just that the search goes to luna*(Much cheaper).
The first attempt was simple putting custom instructions in agents.md and.. it was a disaster. Sol either ignores it or at one point it literally opened a web search for "how to start a subagent" (wish I was making that up lol). Even when it works you end up paying more for a much slower response because Luna sends back garbage half the time and Sol has to redo the search anyway.
the benchmarks I ran are DeepSWE, MAH-SWE, and bugs from repos I actually work on,with Luna's cost counted with Sol. limits lasted almost 3x longer, with no quality regression and most of the time faster responses!
one command in 3 seconds, it uses your existing Codex sub.
https://github.com/repotracer/repotracer