Theo lists six problems with Jev compaction, Armin Ronacher pushes back
A demo that scores every tool call and deletes the rest turned into the week's argument about what a fast classifier can actually decide.
The demo that started it came from tamara, who called Jev a fit for "instant compaction", scoring every tool call in the history and dropping what is irrelevant instead of running a summarization prompt. Theo took it apart in six points, and Armin Ronacher then said he would not dismiss it at all, which is where the earlier exchange left off.
This is a terrible compaction strategy that fundamentally doesn't understand how compaction and context management work.
What Theo actually objected to
His strongest points are mechanical. A 32k token context model scoring tool calls one at a time knows very little about what happened before, and in this implementation does not even see the result of the call it is judging, which he argues leaves the agent looping on things it already tried. Frontier models from OpenAI, Anthropic, xAI and Google do not return reasoning traces over the API, only encrypted payloads that Jev cannot read and will often drop. And deleting early history invalidates the KV cache from that point on. If your history is 1 through 6 and you delete 2, you rewrite 3 through 6, which Theo says costs more than leaving 2 alone. He puts cache writes at over 60 percent of his own LLM spend in Claude Code and Codex.
He also notes that labs have spent the last year training models on their own compaction flows, and offers a detail to prove the point. In Codex, if you switch models mid thread and compaction fires, it runs on the model that was previously in use.
Where Ronacher disagrees
Ronacher's counter is narrow and practical. Most harnesses need some pruning on compaction anyway for cost reasons, so a cheap scorer has somewhere to be useful. He adds that modern compaction is already deeply unintuitive, pointing at a Codex path where the model requests a blank context on its own and maintains its own notes.
The framing everyone borrowed came from Mario Zechner, who called Jev basically a general purpose classifier and said his seven years of SVMs and random forests had been replaced by it. That is the real question under the argument. A classifier that is fast and cheap still has to make a call that depends on the whole thread, and nobody in this exchange has measured how often it gets that wrong.
jev is basically a general purpose classifier. and i think that is amazing.
Theo, for his part, later posted that Jev has replaced ChatGPT, Codex, Cursor, Claude Code, iMessage, Spotify, Ford, CVS and McDonalds for him, then clarified it was sarcasm.
I would not dismiss the idea of using Jev for compaction at all.
Have the good opinion, not the loud one
One email a day. Under three minutes. What working engineers found out about coding with AI.