dsh-context-before-user
将 DSH 注入的用户角色上下文置于当前用户消息之前
插件会安装到这里；不确定时保持 web。
npx -y @deepseek-ai/dsh plugin --profile web add github:n0pe-sled/context-before-user#1c3816c08650c5f71695ea6319a39d1a6b5f32b3Context Before User 以 dsh-context-before-user 发布，当前版本为 0.1.0。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。
Host-only DeepSeek Harness plugin that keeps injected user-role context ahead
of the current human turn. At the public agent/pre-step waterfall it awaits
the normal DSH injectors, then stable-partitions the entering messages:
The resulting provider-visible order is:
system: assembled system prompt
...conversation history...
user: runtime-context snapshot
user: workspace instructions and skill catalog/instructions
user: current human request
The partition uses trusted provenance: messages with source.kind === "user"
stay in their original order at the end, and all other step messages stay in
their original order before them. Skill catalogs remain durable user-role
context turns. The system prompt is neither copied nor wrapped in a
<system-reminder>; it remains solely in request/header.system and the
provider's system slot.
The plugin uses agent/pre-step because loop-built llm/stream requests are
deep-frozen. Reordering here remains model-visible and logged, preserving exact
request reconstruction.
Existing durable history is never rewritten. The plugin fixes newly entering step messages, so start a fresh session when validating the complete initial transcript order.
dsh plugin --profile web add ./context-before-user
Restart the profile after installation.
npm test
For an end-to-end check, export a harmless test session and confirm that the runtime snapshot and skill catalog remain separate user-role messages before the actual human request, and that the assembled system prompt appears only in the request header/system slot.