dsh-compaction-pro

When long conversations hit the compaction threshold, the built-in dsh-compaction-basic summarises with a generic English template that routinely rewrites or drops version numbers, file paths, command strings, and error codes. Chinese sessions end up as English summaries, creating a jarring read-back. For quant or coding agents, one altered identifier downstream is enough to break the pipeline.

dsh-compaction-pro swaps out only the single documented override point, summarize(), while inheriting every pressure, retention, and token-metric policy from the official engine. The high-fidelity prompt enforces verbatim retention of numbers, paths, commands, and identifiers; the summary language follows the session (Chinese in, Chinese out, code blocks untouched); oversized segments can be recursively split and merged to survive maxTokens limits; decision summaries now include reasoning and rejected alternatives; and a customInstruction field lets you replace the entire built-in prompt.

If you run coding, quant, or any precision-critical multi-agent workflows inside DeepSeek Harness and work in Chinese, this is a drop-in upgrade over the stock compaction backend. Disable the built-in basic engine, install pro, and every call site switches transparently.