DeepSeek Harness Dynamic Output Quota Plugin: dsh-adaptive-max-tokens
`dsh-adaptive-max-tokens` is a plugin for DeepSeek Harness (DSH) that addresses the issue of fixed reserved output tokens conflicting with input text in long-context conversations, which can cause the service to reject requests. The plugin dynamically calculates `maxTokens`, flexibly adjusts the available output space based on context capacity and input estimation, and automatically restores the reserved quota once compression takes effect. It supports error identification and retry, calibration after interrupted chat resumption, and routing compatibility, and must be installed and used with DSH version `0.1.6-alpha.1` or `0.1.7-rc.2`. Note that the plugin adjusts the budget based on estimation only; it does not increase the model’s actual capacity and cannot guarantee absolutely no overflow.
Read Moredsh-context-checkpoint: An Automatic Checkpoint Plugin for Fixing Lost Conclusions in DeepSeek V4.1 Flash Long-Context Scenarios
The `dsh-context-checkpoint` plugin, maintained by Leafstory, is designed to address the issue in long DeepSeek V4.1 Flash conversations where the model forgets or denies earlier conclusions. It automatically summarizes and compresses the key state of the conversation and injects it at the beginning of the prompt, ensuring that the model reads the facts at the highest-attention positions without relying on model memory. Its core features include automatic checkpoint persistence, scheduled compaction, and a forced bypass mechanism, making it suitable for long-running task scenarios. After installation, DSH must be restarted. The plugin does not depend on idle windows and runs automatically when steps transition.
Read MoreDSH Git Memory: Git-Based Long-Term Context for DeepSeek Harness
`dsh-git-memory` is a Git-based memory management plugin for DeepSeek Harness, designed to address the challenge of managing long-term context for agents. The plugin is currently in the compatibility testing phase. Its core features include dynamic context registration, metadata observation, and the ability to view status via human commands and model tools. Installation requires Node.js 24, pnpm 11, and DSH 0.1.0-rc.6, and supports use via a Web profile. As it is still in testing, the current version intentionally does not store memory and lacks persistence, write, or background invocation capabilities. It also imposes strict restrictions on sensitive data and Git operations. Its design provides a foundation for a controlled and auditable long-term memory system.
Read Moredsh-llm-longcat: A Long-Context and Thinking Mode Adapter for DeepSeek Harness
dsh-llm-longcat is a DeepSeek Harness (DSH) plugin designed to integrate the LongCat-2.0 model. It primarily resolves issues in the Harness ecosystem related to long-context (1M token) processing, structured chain-of-thought display, and native tool calling. The plugin supports streaming, multi-turn conversation reasoning review, and API key management via environment variables. Installation requires running a local script, and configuration should note that image input, stop sequences, and binary limits on reasoning strength are not supported.
Read MoreMoonshot Open-Source Kimi K3: An Open-Source Cutting-Edge Model with 2.8T Parameter MoE, 104B Activations and 1M Context Window
On July 27, 2026, Moonshot AI released the full weights of Kimi K3 on Hugging Face and GitHub: a 2.8T total parameter MoE model with 16 activated experts out of 896, 104B activated parameters, 1 million token context window, and native multimodal capabilities. Based on the official README and arXiv technical report, this paper sorts out the key points of the KDA and Stable LatentMoE architectures, interprets the differences between Agent evaluations such as Terminal-Bench and model harness tests, and introduces the API, deployment paths via vLLM/SGLang, and the usage boundaries of the Kimi K3 License.
Read More