Made my Codex limits last almost ~3x longer with one change
Heat trend
Collecting trend data
The percentage is based on available heat signal, not comment count or independent people.
A developer significantly extended their Codex limits, making them last almost three times longer, by implementing a custom solution based on a Microsoft research paper called FastContext. This involved building an MCP Rust tool with a custom router, which was tuned through extensive benchmarking. The approach allows the more expensive Sol model to handle actual coding while offloading search tasks to the cheaper Luna model, addressing concerns that Plus users were being forced to rely on Luna for productivity.
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