dsh-qwen38-local-qol

Overview

DeepSeek Harness QoL plugin for the local Qwen3.8 line (27B, Flash-Next): per-request thinking budgets on both llama.cpp and NInfer dialects, vision + tools, and a compaction backend whose summaries stop burning the output cap on thinking.

README / EN

Package documentation

Registry summary

DeepSeek Harness QoL plugin for the local Qwen3.8 line (27B, Flash-Next): per-request thinking budgets on both llama.cpp and NInfer dialects, vision + tools, and a compaction backend whose summaries stop burning the output cap on thinking.

dsh.pub verifies the pinned bundle contract, runtime facts, and distribution semantics. The complete README remains in the source repository.

Read the full README on GitHub

LIMITATIONS

Known limitations

- **Flash-Next has a fast line and a compat line** — the ExLlamaV3/TabbyAPI dialect (EXL3 quant, 256K context) is the fast path; the llama.cpp dialect still runs it at its own window/budget values. - **The TabbyAPI line's vision token count is not pinned** — the token meter reports image capacity as unknown on that line until the ExLlamaV3 image-processor formula is measured (the request itself works; only the pre-flight capacity projection is affected). - **The preset seam is a web-surface feature** — headless profiles do not mount `agent-presets` rows; the provider route (thinking budgets) works on both surfaces. - **Summarizer internals depend on the engine version** — the wire rules (thinking off, full output cap) hold for every engine version; engine internals are outside the plugin's control.

dsh-qwen38-local-qol