dsh-jiuwen-adapter

Use JiuwenSwarm as the brain of dsh, with your other dsh plugins still working.

中文 · Quickstart · Config · dsh integration · Changelog

---

A Cordis plugin that registers a jiuwen provider on ctx.llm. dsh still handles the UI, sessions, approvals and plugins. Each turn goes to the JiuwenSwarm Gateway, which runs its own agents, tools and memory.

flowchart LR
  D["dsh + your plugins"] -->|ctx.llm · provider=jiuwen| A["jiuwen adapter"] -->|"WS chat.send"| G["JiuwenSwarm Gateway :19000"]

• Plugin compatibility: docs/plugin-compat.md

30-second quickstart

# 1. Brain: start JiuwenSwarm (put your model key in ~/.jiuwenswarm/config/.env)
uv pip install workswarm && jiuwenswarm-init && jiuwenswarm-start app

# 2. Plugin: install into a dsh profile (dsh >= 0.2.1-rc.2: npm i -g @deepseek-ai/dsh@next)
dsh plugin --profile web add "git+https://github.com/ThinkFlowLab/dsh-jiuwen-adapter.git"

# 3. Make it the default model, then chat
printf -- '- id: agent-default-model\n  config:\n    provider: jiuwen\n    model: jiuwen-default\n' >> ~/.dsh/profiles/web/cordis.patch.yml
dsh --profile web

Step-by-step guide for Mac and Linux: docs/quickstart.md. To try it without a Gateway, run npm run mock:gateway.

Features

	What you get	Default
🧠 Jiuwen brain	One dsh turn is one Gateway chat.send. Agents, team mode, skills and memory run in Jiuwen	on
💬 Live activity	Reasoning, 🔧 tool calls, subtasks and token usage stream into dsh	on
✅ dsh approvals and questions	Jiuwen permission prompts, plan approval and ask_user go through dsh's own dialogs, and deny when no answer comes	on
🔁 Session continuity	dsh sessions map to Jiuwen sessions. Only new text is sent after the first turn	on
🛟 Drop recovery	If the Gateway drops mid-answer, the reply is recovered from history and marked as recovered	on
🧾 Local titles and compaction	Titles and /compact are handled locally, so they never cost a Jiuwen turn	on
⌨️ /jiuwen command	Sessions, models, modes, skills, agents, cron, MCP and more (list)	on
🖼️ Multimodal	Image and file parts are sent as Jiuwen media	on
🪪 System prompt → Jiuwen	injectSystem: full|persona sends the host system prompt (any role/profile) once per Jiuwen session, resent only on change, capped by injectSystemMaxBytes. injectSystemUpdate: diff (default) resends only changed sections (safety gates may still send full); optional hostSystemSectionsAllow/Deny filter by heading; an optional hostProfile object can name your host's persona sections. Hosts that need the full orchestrator role keep full; ordinary dsh chats may use persona/off	default full (off to disable)
📚 Host skills → Jiuwen	Skills Jiuwen lacks go as an index (name + description) loaded on demand via the passthrough skill tool, or inline bodies when that tool is absent; hash-gated across restarts, best-effort secret redaction, skillBodyMaxChars cap with a truncation marker	default index
⏹️ Interrupt on new message	A user message that arrives while a Jiuwen turn is running interrupts it (chat.interrupt), coalesces everything that arrived meanwhile into one step, and resends with a note that the previous turn was cut (side effects not rolled back). Waits for a pending approval card first (docs)	default interrupt (onBusyMessage: queue / JIUWEN_ON_BUSY_MESSAGE=queue for the old wait-until-done)
🪜 Full trajectory	trajectoryDetail: full renders Jiuwen tool calls, results, subtasks and plan steps as Markdown steps in the reasoning block	default full (summary to disable)
🔌 dsh tools → Jiuwen	All dsh host/plugin tools exposed to Jiuwen over MCP by default (toolAllowlist: '*'; core = restricted shell/file + skill + session-query set, plus toolAllowlistExtra and host-declared passthrough: true), each call gated by an approval. Per-session proxies are released on session end or after mcpSessionIdleMs, stale ones GC'd at start; reconnect only when the tool set changes	on with hostSync
🎛️ Per-profile overrides	sessionOverrides by host agent preset (preset glob) sets model / chat mode / template defaults; request, catalog and /jiuwen mode still win	opt-in
🧭 Bridges	Model discovery, typed events	opt-in
🔄 Host → Jiuwen sync	Skills, MCP, model, plan, workDir, title, delete, goal, pause, permissions, fork, schedules (docs)	on
🩺 Config validation	Unknown keys get a "did you mean" hint and deprecated keys get a warning. Never throws	on

Compatibility

Other dsh plugins must keep working. The adapter is tested against the full dsh agent runtime (dsh-agent, dsh-agent-loop, compaction, subagents, goals, plan mode, a bash sandbox, approvals and user-questions). The same scripted session runs once with jiuwen loaded and once without. Every row of the compat matrix must match, covering events, tool calls, plan exit, goal rounds, auto-compaction, subagents, bash, host event counts, globals and load failures.

Component	Supported	Tested in CI
dsh	>=0.2.1-rc.2 <0.3.0	0.2.1-rc.2, 0.2.1-rc.5
Cordis	^4.0.5	4.0.5
Node	>=22	22, 24
JiuwenSwarm	0.2.x Gateway	0.2.7 (live)

The adapter also works in any Cordis app that has an llm service (usage).

Docs

• Quickstart, Configuration, Troubleshooting
• Using it from code and Gateway features, Gateway RPC coverage
• dsh integration, Host → Jiuwen sync, Testing, Known issues and test logs
• Publishing to a dsh plugin market, Architecture

Development

npm run lint                  # node --check over index.js, src/, bin/, scripts/, test/
npm test                      # unit, mock-gateway and runtime suites
npm run test:coverage         # + coverage floors and the test-count baseline (what CI runs)
bash scripts/guard-live.sh    # box-side nightly live guard (real Jiuwen + plain dsh), see docs/testing.md
npm run compat -- --diff      # with vs. without jiuwen compat matrix (needs DSH_NM=<dsh node_modules>)
npm run market:index          # regenerate market/index.json

Licensed under Apache-2.0.