Nemotron 3 Nano · Nebius Token Factory
CacheForge
Context Compactor for Coding Agents
Coding agents re-send every test log on every turn. CacheForge condenses them before they reach the model.
That tower is one real test log. Scroll.
One real test run ·
The tower is that log, lines. CacheForge keeps what an agent needs to fix a failing test and drops the passing-test noise. It sent lines, characters.
Every kept line is copied verbatim, and a few summary lines are added on top. It's plain code, not another model call, so it can't invent a test result.
The same session, every request
Without it, every turn re-sends the whole history.
Each tower is one request to Nemotron 3 Nano, drawn to the same scale. Amber is the agent alone; cyan is the same task through CacheForge.
turns · bugs fixed
turns · bugs fixed
One session (). The full record is below.
Every Token Factory run
more agent sessions per dollar, in every pre-registered run.
comparisons of 15 paired sessions, with the metric and test written down before each run.
| Run | Task | Condenser | Prompt tokens | Cost / session | Sessions / $ | Bugs fixed |
|---|
Nemotron 3 Nano on Nebius Token Factory, full 262,144-token context, list price. Bugs fixed is without → with CacheForge over 15 sessions; p is a two-sided paired exact permutation test. “Rescored” means sessions were re-checked by replaying their saved code changes against the original tests.
How it works
Change one setting: the base URL.
Keeps what matters
Pass and fail counts per file, each failing test's assertion and error, the setup lines it depends on, and file and line.
Never rewrites you
System prompts, your messages, the model's replies and file reads are never touched.
Deterministic
No extra model call. Every kept line is copied verbatim, so it can't hallucinate a test result.
python -m src.tui replay
The terminal app replays every recorded session, raw log next to what the model saw. It needs no API key.