dsh-sum

Large language models frequently stumble on arithmetic, and even a straightforward addition can drift when the context window grows. dsh-sum extracts summation into a local, in-session tool so that deterministic math is delegated to code rather than left to probabilistic reasoning, giving you stable and reproducible numeric output every time.

Written in pure Node.js with zero network calls, the plugin runs fully synchronously inside the conversation loop. Once registered, the model can invoke it whenever it needs to accumulate a set of numbers, receive an exact sum, and then move on to the next reasoning step without any external service or extra dependency in the mix.

It fits well in DeepSeek Harness pipelines that involve frequent numeric roll-ups — report field aggregation, cumulative tallies across multi-turn dialogues, or feeding a validated sum into a downstream decision step. Lightweight, offline, and released under MIT, it is ready to drop in and use as-is.