dsh-lexiforge

Customizing output style in DeepSeek Harness often means piling on system prompts that are fragile, hard to reuse, and impossible to share. dsh-LexiForge introduces a pluggable language-pack framework: it intercepts the model reply stream and re-processes it through installable, ZIP-based language packs that handle style rewriting, terminology-aware retrieval, or pure local rule replacement — turning a style shift into a one-click, reorderable configuration.

The core engine offers three modes: A (a standalone LLM call for full paragraph rewriting), B (SQLite FTS5 terminology search and injection), and C (local regex and dictionary replacement with zero token cost). They can be chained inside a single pack as a B→A→C composite pipeline with only one LLM request. Multiple packs stack in priority order, so you might proofread first, shift to classical Chinese, and finish with a character catchphrase. Security is built in from the ground up: path-traversal guards, extension whitelists, entry and size caps, timeout fallback to the original text, and independent-request recursion prevention. Language packs contain no executable code at all.

It is well suited for team users who need consistent terminology, creators who want a stable character voice, and developers contributing custom packs to the DSH ecosystem. Community packs are discoverable and installable directly from the GitHub market — no coding required.