I build and maintain a set of open-source plugins for DeepSeek Harness (DSH), an agent harness for local models. Running agents on your own GPU has practical problems that cloud APIs never surface, and these plugins solve the ones I kept hitting.
- dsh-context-budget. Keeps a local model's context at a size the GPU actually handles well. A model that advertises a 256K window may read prompts at 300 tokens per second when the context is small and 70 past 100K, so this plugin measures the real prefill rate and warns or compacts early, before turns take half an hour to start.
- dsh-llm-gate. A per-provider concurrency gate for model requests.
- dsh-fetch-timeouts. Raises Node's HTTP timeouts process-wide so slow local models (Ollama, LM Studio) are not cut off after 5 minutes.
- dsh-turn-doctor, dsh-logbook, dsh-model-pin. Turn diagnostics, session logging, and model pinning.
I'm also a contributor to awesome-deepseek-harness, the community index of the DSH ecosystem (950+ stars).