dsh-qwen38-local-qol: Tuning Local Qwen3.8 Runs for DeepSeek Harness
yunado/dsh-qwen38-local-qol
A DSH plugin for local Qwen3.8 lines: per-request thinking budgets, a compaction backend that stops burning the output cap on thinking, and a live settings tab.
dsh-qwen38-local-qol is a quality-of-life plugin for DeepSeek Harness aimed at running Qwen3.8 locally, maintained by Yunado under the MIT license. It targets two pain points of local deployment: thinking budgets that cannot be adjusted per request, and compaction checkpoints truncated at the output cap, supporting llama.cpp, NInfer, and TabbyAPI OpenAI-compatible backends without patching the DSH core. On boot it generates the qwen38 preset and makes it the default when none is configured, so new sessions pick it up automatically, while a settings page tunes connection, window, budgets, and trim knobs live and a status dot shows the active state at a glance.
How to Install
dsh plugin --profile web add github:yunado/dsh-qwen38-local-qol- Category
- Tools & Capabilities
- Platform
- DSH Plugin
- Author
- yunado
- Distribution
- Plugin
dsh-qwen38-local-qol Key Features
dsh-qwen38-local-qol Repo Summary
What Does It Do?
Qwen3.8 Local QoL is a DSH plugin for DeepSeek Harness that improves the experience of running Qwen3.8 locally, maintained by Yunado under the MIT license. It solves two pain points: thinking budgets that cannot be set per request, and compaction checkpoints that get truncated at the output cap. It works with llama.cpp, NInfer, and TabbyAPI backends that expose the OpenAI-compatible API, without patching the DSH core. On boot it generates the qwen38 preset and makes it the default agent preset when none is configured, so new sessions pick it up automatically.
Core Features
- Per-request thinking budgets: the llama.cpp line sends the selected per-effort budget on every request, overriding the server's startup budget, while the TabbyAPI line accepts both natively.
- A compaction backend: the summarizer prefill is trimmed (recent reasoning only, images downgraded to text placeholders, tool results capped), and compaction calls run thinking-off at the line's full output cap so checkpoints stop being truncated.
- A live settings tab: the Qwen3.8 Local page in DSH settings configures connection, window numbers, budgets, and trim knobs per line, applying changes live and persisting them to the config file.
- A status dot: green when qwen38 is the default preset, amber when another preset is default, gray when the preset is missing.
- Multi-backend dialects: switch between llamacpp, ninfer, and tabbyapi; the NInfer line reads its single thinking budget from the server startup flag, shown as a note in the tab.
How to Use This Plugin?
After enabling it in DSH and restarting the web service, the plugin generates the qwen38 preset at boot; new sessions use it automatically while existing sessions keep the preset they were created with. Day-to-day tuning happens in the Qwen3.8 Local page under DSH settings: fill in the server address, model name, context window, and max output per line, then set the low/medium/xhigh thinking budgets and the default effort. Headless profiles and patch/env configs accept the same fields, with env covering what the patch does not set; the summarizer trim options for images, kept turns, and tool-result characters are env-only.
This page is an independent rewrite of the plugin's official README — for authoritative documentation and the latest changes, refer to the source: yunado/dsh-qwen38-local-qol. The plugin is third-party code that runs on your machine once installed; inclusion does not imply endorsement — please review the source before installing.
DSH Plugin FAQ
