@xuanling-rs/xuanling-dsh-skills
DeepSeek Harness 套件：挂载隔离的 XuanLing 工作流技能和严格的 MCP 覆盖策略。
插件会安装到这里；不确定时保持 web。
npx -y @deepseek-ai/dsh plugin --profile web add @xuanling-rs/xuanling-dsh-skills@0.4.0Xuanling Dsh Skills 以 @xuanling-rs/xuanling-dsh-skills 发布，当前版本为 0.4.0。Plugin Hub 会校验它的 manifest，并保存精确安装来源，便于复现安装结果。
English | Simplified Chinese
DeepSeek Harness bundle containing two on-demand XuanLing workflow Skills and
the strict whole-file overwrite policy. It registers no MCP server and has no
dependency on @xuanling-rs/xuanling-mcp.
dsh plugin --profile web add @xuanling-rs/xuanling-dsh-skills@0.4.0
This command augments DSH's shipped Web profile. Use --profile headless for
the shipped Headless profile. Unknown profile names start with the base bundle
only and do not provide a runnable Web or Headless application by themselves.
Use it with exactly one XuanLing tool bundle. xuanling-file-workflow uses
mcp_catalog__xuanling to activate an exact missing XuanLing file tool, then
routes between it and Harness-native tools; it also covers CAS overwrite,
compound suffixes, atomic multi-hunk
patches, deterministic repeated validation, and host background jobs. Its v3
request policy uses the omitted 65,536-byte output budget, numbered reads,
conditional known_sha256 reads, exact project scripts, and minimal-environment
diagnostics. It keeps include_diff: true because DSH-native diff visibility
for XuanLing edits is not yet proven; SHA is not semantic edit validation.
xuanling-memory-workflow activates only the next required memory operation,
keeps project-local L1 facts separate from shared XuanLing L2 facts, uses
explicit pointer-driven pull recall, and preserves the
proposal/review boundary. The policy rejects an existing-file overwrite
without expected_sha256 before MCP dispatch and returns a Host ask gate for
the exact mcp__xuanling__memory_review tool. DSH dispatches only an
allowed-once outcome; rejection, cancellation, unavailable approval, or a
call without an agent fails closed.
Node.js 22.14 or newer is required. Licensed under the MIT License.