作者 · SipengXie2024

Hermes-style bounded memory + self-curating skill library for DeepSeek Harness (dsh): two-file curated memory, background review loop, LLM library curator. Out-of-tree Cordis plugin.

User-attention notifications for DeepSeek Harness (dsh) web: turn end, approval requests, and model questions — in-page toast + OS-level browser notification

LLM-gated auto approval for DeepSeek Harness: a model judges every approval ask first, low-risk operations pass without prompting (fail-closed)

DeepSeek Harness plugin: LLM session titles that survive reasoning models (startup sweep + turn-end retry), plus a hover-card session-size readout

Claude Code-style rewind for DeepSeek Harness (dsh): edit a past user message and re-run on a forked session, restore files to the turn boundary, or branch with a summary

Claude Code v2.1.x built-in /code-review skill ported to dsh (DeepSeek Harness): angle-based multi-agent review pipeline as a plain skill