Theo calls the Jev compaction plugin a terrible strategy
A six-point teardown of how the plugin trims agent history, and why he says it makes models dumber and not cheaper.
Theo posted a six point teardown of the Jev compaction plugin and opened with the verdict that it is a terrible compaction strategy that fundamentally does not understand how compaction and context management work. Compaction is the trick of shrinking an agent's conversation history so it can keep working once the thread gets long. His argument is that the plugin treats it as a filter.
The six points
Compaction should run sparingly when context gets too long, he says, not constantly to keep context small. The plugin decides what to keep per tool call, using a 32k token context model that knows little about what came before and, in his reading of the implementation, does not even know what the tool call returned. Delete things that way and you get what he calls stupid loops, where the model keeps retrying what it already tried.
He also says frontier models from OpenAI, Anthropic, XAI and Google do not share reasoning traces over the API, only encrypted payloads the plugin cannot see and often drops. Anthropic is stricter still and requires the entire history preserved to get any reasoning data back, so he expects a model running this inside Claude Code to act way dumber.
The cost argument
The part that should interest anyone watching their bill is cache writes. Theo says they are the biggest cost by far for agents, and that in his own use of Claude Code and Codex he often sees cache write costs go over 60% of his total LLM spend. Editing history early in a thread invalidates everything after it, so deleting item two in a list of six means rewriting three through six. Cheaper, he argues, to leave item two alone.
Replying to @tamarajtran, who said tool calls are matched by id so the model knows what is relevant, Theo said that is no longer true for Anthropic models and will likely change at the other labs. When @moinerus ran a benchmark, Theo thanked them and said the results did not go against him.
The business read. Jev keeps drawing experiments from builders, including Josh Pigford's run through Hacker News comments, and a plugin that reduces token count is easy to celebrate before anyone measures whether the agent still finishes the job. Theo's advice is to use the defaults in Claude Code and Codex.
This is a terrible compaction strategy that fundamentally doesn't understand how compaction and context management work.
Cache writes are the biggest cost by far for agents. I often see cache write costs go over 60% of my total LLM spend in my personal use of Claude Code and Codex.
Thank you for benching it! Would have been awesome if I was wrong here. Sad to see I wasn't.