DeepSeek-Reasonix Won't Touch Its Prompt Prefix

DeepSeek-Reasonix prunes stale tool output before compaction, not after, so its prompt prefix never changes, per the project's GitHub page. Rewrite anything near the head of a session you leave running, and every token after that edit falls out of the cache. The model has to reprocess all of it on the next call.

It ships as a single static Go binary that's picked up 27.9k GitHub stars. The whole build rule is simple: inject a small environment summary at startup, then never touch what follows.

Executor and planner models can also run in separate cache-stable sessions, so one model's churn doesn't disturb the other's prefix.

vLLM's K3 project is chasing the same problem from the serving side. It uses copy-on-write partial prefix matching built into the model server instead of an agent's own pruning logic.

Any agent that lets an early edit ripple through its own context is signing up for the losing side of that race. Worth watching: which agent frameworks make append-only context the default, not a pattern each team has to enforce by hand.

As of July 27, Reasonix's prefix discipline and vLLM's K3 work answer the same cache-cost problem from opposite ends of the stack.