dsh-plugin-translation

Handing a long document to an LLM for translation rarely fails on individual word choices. It fails on the mechanics: terms drift mid-document, numbers and units go mismatched, a bracket is left unbalanced, tone wavers between formal and casual, and every new session starts with total amnesia. dsh-plugin-translation pulls those concerns out of the model and into a deterministic pipeline of chunking, validation, and memory, so the agent focuses on producing the translation while the plugin chunks, checks, and remembers.

The toolkit exposes fine-grained actions. Segment splits text into numbered, bounded blocks. Glossary auto-extracts candidate terms such as acronyms, camelCase identifiers, domain vocabulary, numbers with units, email addresses, and URLs, each with an occurrence count. Check runs source-target QA covering number and unit parity, bracket matching, length ratio, and doubled punctuation. Tone provides register-specific style guides for formal, colloquial, and technical registers in both English and Chinese. The memo and glossary families add, read, update, and remove entries in a persistent translation memory and a durable workspace glossary. Consistency cross-checks terminology across segments against the glossary. Every file path is containment-checked to the session workspace, and the pure-logic core has zero framework imports, making it straightforward to unit-test and reuse.

If you are building multi-segment translation workflows on DeepSeek Harness, need a glossary that survives across sessions, or want auditable quality gates on translated output, this plugin turns translation from one-off prompt tuning into repeatable, traceable process engineering.