dsh-zhushou · Assistant personas, inside your DSH session

> A DeepSeek Harness (DSH) sidebar assistant plugin: manage and pick assistant personas from the sidebar, then talk to the selected one inside the main DSH session — with its own persona, model, parameters, world book and memory. Leave the selection empty and the session stays fully native.

English · 中文

Overview

The problem: you want to switch between assistant personas (a coding mentor, a translator, a story writer…) without juggling separate apps, separate chat windows, or hand-edited system prompts per session.

What dsh-zhushou does: a sidebar 助手 (Assistant) option opens a management panel where you create and edit assistants — name, avatar, tags, workspace binding, model parameters, system prompt with variables, quick replies, injection modes, world book entries, skill switches and memory. Selecting an assistant activates a session-level activation pipeline: every LLM request in that DSH session is rebuilt with the assistant's persona, model parameters, injected blocks, world-book context, memory and time-awareness line. Deselect (or delete) the assistant and the session returns to its native behavior.

Feature highlights:

• Session-scoped activation — selection lives in selection.json per DSH session; the activation pipeline intercepts llm/stream (exact sessionId match, purpose === undefined filters out compaction/session-title/subagent calls) and short-circuits with a rebuilt request
• Assistant profiles — name / avatar (dataURL or URL) / tags / bound workspace
• Model parameters — model (empty = follow the main model) / temperature / Top-P (reserved) / reasoning effort / max tokens / context message count / streaming switch
• System prompt engine — template + dual variable syntax ({{var}} and {var}) + per-assistant variable overrides
• Quick replies — one-click prewritten message templates bound to an assistant
• Injection modes — role (system/user/assistant) + position (before/after/replace) + trigger (always/keywords)
• World book — keyword hit → priority order → 1024-token budget truncation → injected into the request
• Skills — enumerate ctx.skills.list() and enable/disable per assistant
• Memory & time awareness — global/private memory pools, chat-history reference switch, and a time-awareness switch that injects current time / last-chat time / elapsed interval as a natural context line (no directive "time reminder" messages)
• Sidebar UI — sidebar.footer.action text entry (with selected state) + shell.overlay management panel (select/cancel/edit/copy/delete) + settings.section full settings page
• Clean handoff — no selection ⇒ fully native session; cancelling a selection restores native behavior immediately

Compatibility

Install / Uninstall

Install

# From npm (recommended)
dsh plugin --profile web add @bananiceee/dsh-zhushou

# Or pin an exact npm version
dsh plugin --profile web add @bananiceee/dsh-zhushou@0.0.1

# Or from the GitHub repository
dsh plugin --profile web add github:Starlight-bananice/dsh-zhushou

# Or from a local checkout (profile assembly)
dsh plugin --profile web add ../dsh-assistant-panel

# Or a pinned release tarball — immutable and versioned (attached to every GitHub release)
dsh plugin --profile web add https://github.com/Starlight-bananice/dsh-zhushou/releases/download/v0.0.1/bananiceee-dsh-zhushou-0.0.1.tgz

> Note: pnpm 11 enforces a 24h minimumReleaseAge for freshly published packages — if a same-day release is rejected, append --config.minimumReleaseAge=0 to the dsh plugin add command.

> Note: the built lib/ artifacts are committed to the repository — a git install is ready to run immediately, no build step required. Add the plugin, restart DSH Web, done.

# Or runtime injection without a restart (developer workflow)
#   dev_inject_plugin / dsh-super-injector → point at this repository

Then start/restart DSH Web. No configuration is required — the 助手 entry appears in the sidebar footer.

Upgrade

# npm installs: update straight to the latest registry version
dsh plugin --profile web update @bananiceee/dsh-zhushou

# github: installs pin a commit at install time — re-add a newer ref to upgrade:
dsh plugin --profile web remove @bananiceee/dsh-zhushou
dsh plugin --profile web add github:Starlight-bananice/dsh-zhushou#v0.0.1

Disable

• Keep the data, stop the activation — deselect the assistant in the management panel: every session returns to native behavior; assistants and memory stay in place.
• Stop the plugin entirely — remove it from the profile's bundles list; re-adding restores it.

Uninstall

dsh plugin --profile web remove @bananiceee/dsh-zhushou

Data left behind: ${DSH_HOME || ~/.dsh}/dsh-assistant-panel/ (settings, selection, assistants, memory) is not deleted — remove it manually if you want a clean slate.

Quick start

1. Install (above), restart DSH Web.
2. Click 助手 in the sidebar footer → the management panel opens.
3. Create an assistant (name, avatar, model parameters, system prompt, world book, skills, memory…) and save.
4. Select it for the current session — the panel shows the active selection; from now on every reply in this session uses the assistant's persona and parameters.
5. Deselect to return to the native session.

Features

HTTP API

Root: /assistant-panel/api (prefix route, same-origin local service, no auth). Unified envelope: { ok: true, data } / { ok: false, error: { code, message, details? } }.

Data layout

Data root (overridable via PUT /profile dataDir): ${DSH_HOME || ~/.dsh}/dsh-assistant-panel/

settings.json            # plugin settings (userName/locale/timezone/dataDir)
selection.json           # session selection state (sessionId → {assistantId, lastChatTs})
assistants/.json     # assistant profiles (atomic write: tmp + rename)
global-memory.jsonl      # global memory pool
memory/.jsonl  # per-assistant private memory pool
# chats/ retired: chat history is carried by DSH sessions (~/.dsh/sessions)

Known limitations

1. Top-P not wired yet — DSH GenerateOptions / LlmCallConfig has no topP field today; the setting is kept, validated and persisted but not passed to ctx.llm.stream. It will be wired once the official API supports it.
2. Session model UI does not follow activation — persona/model/parameter overrides act on the llm/stream request layer (activation pipeline, see docs/ARCHITECTURE-ACTIVATION.md); the session header and DSH UI still show the original provider/model.
3. The official system prompt can be replaced — on activation, rebuilt.system is the assistant-composed block (persona first); if the assistant system prompt is empty the official system prompt is kept.
4. Injections don't appear in the session log — activation only rewrites the request face; session/event records the original messages (consistency first).
5. reasoningEffort is mapped to the nearest tier — DSH only has off/low/high/max; medium is mapped to low host-side, auto is omitted (server default).
6. Memory extraction not closed-loop — memory injection (keyword + time-decay top 5) and CRUD are implemented; the async "fast-model auto-extracts facts → memory pool" step is not wired yet (entries are maintained manually on the settings page).
7. Skill panel reflects the host runtime — ctx.skills.list() returns the skills registered in the current runtime (e.g. vision-skills), not a compile-time full catalog.
8. No multimodal support — text messages only; images etc. are planned for later versions.

Development

# Rebuild host + client (needs a DeepSeek Harness checkout; probes DSH_CHECKOUT or common paths)
DSH_CHECKOUT=~/deepseek-harness bash scripts/build.sh
npm run build:client          # tsdown → lib/client.js

# Verify committed lib/ is in sync with src/ (full rebuild + byte diff)
npm run verify

# Enable the pre-push guard (auto-runs verify.sh when src/ or build config changes)
git config core.hooksPath .githooks

# Runtime injection without a restart (dsh-super-injector workflow)
dev_build_plugin { "dir": "/path/to/dsh-assistant-panel" }
dev_inject_plugin { "dir": "/path/to/dsh-assistant-panel" }
dev_plugin_status
dev_uninject_plugin { "match": "dsh-zhushou" }

Documentation

• docs/ARCHITECTURE.md — overall architecture
• docs/ARCHITECTURE-ACTIVATION.md — the activation pipeline (llm/stream short-circuit rebuild)
• docs/DESIGN-ACTIVATION.md — source-verified API research behind the activation design
• docs/DESIGN.md — original product research

License

MIT © 2026 Starlight-bananice