Overview & result
The author keeps Astra in charge of planning, difficult decisions, review, and integration; Muse performs bounded coding work; Jev Context selects relevant excerpts before context reaches the main coding agent.
From the original post
I did a thing! I got sick and tired of my Codex tokens draining faster and faster lately. Then I realized Meta’s Muse 1.3 at max reasoning is surprisingly close to Sol on coding benchmarks, but way cheaper to run. I'm on the $15 plan and I think it will be very hard for me to use all the tokens it's allowing.
So I made something to save Codex tokens.
It’s called Token Saver. It adds two skills and a set of instructions to Codex so it can delegate work to Muse and be more selective about what goes into its context. You keep using Codex, with Astra in charge.
In my setup, Astra handles planning, difficult decisions, review, and integration. Muse takes bounded assignments like investigating a bug, implementing a feature, making repetitive edits, or writing tests. It returns a compact handoff for Astra to check against the actual code. Small tasks stay with Astra when delegation would create more work than it saves.
Then there’s Jev Context, the other half of this. Before dumping entire files or huge logs into Codex, it helps select the excerpts relevant to the task. Those excerpts keep their file paths and line references, and the originals remain available when more context is needed. Filtering runs locally by default. You can optionally enable TypeSafe’s Jev API to rank excerpts by meaning rather than just keyword overlap.
The idea is pretty straightforward: let Astra make the important decisions, let Muse do useful work without filling up the main conversation, and avoid making Codex read a mountain of material it doesn’t need.
I also added isolated Git worktrees for edits, checks for out-of-scope changes, resumable worker sessions, and an optional usage and acceptance ledger. A worker saying “done” still needs review.
Actual savings will depend on the task and the review overhead. Delegation moves some usage to Meta, so fewer Codex tokens won’t necessarily mean fewer tokens overall. The goal is to make my Codex allowance go further without handing every decision to a cheaper model.
Built it for my own workflow. Put it on GitHub so you can use it too.
https://github.com/TheMarco/token-saver
How Jev fits in the loop
- The coding workflow receives a task plus candidate files or logs.
- Jev Context selects excerpts relevant to that task and preserves their locations.
- The main coding agent reviews the compact context and continues the implementation.
How to reproduce
- Read the repository instructions and identify the two skills or instruction layers it adds to Codex.
- Run a small coding task with a large file or log set and inspect which excerpts Jev Context selects.
- Compare the selected context and task outcome with a full-context baseline before making cost claims.
Why this build matters
This is a direct example of Jev acting as a context gate inside a real coding-agent workflow rather than generating the implementation itself.
Limitations
- The post does not report an independent token-savings or accuracy benchmark for Jev Context.
- Delegation quality depends on the task boundary and on the primary agent reviewing the handoff.
- Filtering context must not remove instructions, paths, or error details required for safe execution.