MIT leader/worker orchestration bundle for DeepSeek Harness: card-based dispatch, a per-project delivery board, read-back into worker sessions, and file-driven orchestration (LEADER.md / SLOTS.md) over persistent worker sessions.
DSH integration
Native runtime
Author-claimed
Safety audit
Unaudited
Last verified
2026-11-20
License
MIT
01What can it help you accomplish?
Run ten features at once instead of babysitting each one
One Leader session that opens cards, dispatches workers, waits for the runtime to report completion, verifies the deliverable and organizes review
A single person managing 20+ sessions; the design explicitly aims at that, not at replacing dsh’s own super agent
See who is working and drill into any agent mid-flight
A per-project delivery board with one card per task, an executor slot chip and a live status light, plus a tab bar; clicking a slot opens that worker’s session and you keep talking without going through the Leader
Developers running several agents in parallel who need the global picture at a glance
Keep the orchestration rules in reviewable files rather than in one long prompt
`LEADER.md` and `SLOTS.md` prompt injection, the `leader-mode` playbooks for discussion → design → delivery → review → merge, and `init-leader-agent` scaffolding of a managed block in the project’s `AGENTS.md`
People who want the conventions reproducible across projects and teammates
02How to install into DeepSeek Harness
Prerequisites
- dsh `>=0.2.1-rc.2 <0.3.0`
- Web install: a profile literally named `web` — custom names lack the web-app bundle and the client half never loads
- Desktop install: its own profile under `$DSH_HOME/profiles/desktop`, `DSH_HOME` absolute or starting with `~/`, and the CLI bundled with the app (the system `dsh` refuses the desktop profile)
- A long-lived host (web / desktop / sdk) — headless is a one-shot CLI and cannot receive completion reports
- For real parallelism the README’s own prerequisite is isolation: one worktree, one dsh instance and one data root per task
Installation steps
- 01
Web: run `dsh plugin --profile web add dsh-leader-agent`, then start the app with `dsh web`
$ dsh plugin --profile web add dsh-leader-agent
- 02
Desktop, GUI: open Plugins in the left sidebar, click Add plugin, type `dsh-leader-agent` and click Install
- 03
Desktop, CLI: quit the app completely (closing the window only hides it) and run the bundled CLI, e.g. `"/Applications/DeepSeek Harness.app/Contents/Resources/runtime/cli/bin/dsh" plugin --profile desktop add dsh-leader-agent`
- 04
In the left sidebar’s Projects section click Add project and pick a directory — the project is created immediately and bound to a leader session
- 05
In the leader session, load the `leader-mode` skill; write `LEADER.md` and `SLOTS.md` (or let the `init-leader-agent` skill scaffold them)
Verify the integration
Not specified by the author
03DSH integration and capability boundaries
A DSH plugin mounted into the `web` or `desktop` profile: it adds the Leader/worker tools (`spawn_worker`, `message_worker`, `create_card`, `read_worker_session`, …), a left-sidebar project tree and a right-panel delivery board to the host, and drives everything through DSH sessions.
Dispatch (spawn_worker / message_worker)
a name, an optional cardId + slot, then the task body as the first message→A persistent worker root session opened immediately (the worker has not started until the message arrives) and bound to the card’s slot
The worker inherits the Leader’s sandbox mode and approval policy at dispatch time; the receipt states the permissions it actually runs withDelivery board and card tools
`create_card` / `update_card` / `list_cards` calls from the Leader (and `read_card` from a worker)→One board per project: cards numbered globally with spec, links, status and executor slot, opened automatically when the Leader creates a card
Card detail writes links and hard-deletes the card with confirmation and no undo; models have no delete tool, so a human clicks itread_worker_session (read-back)
a worker session id→Each turn’s task, body, the tools it called and why anything failed
Read-only: it does not wake the worker and triggers no model callStale-worker patrol
a Leader-dispatched turn that has produced no report for 15 minutes→A plain-text 【后台巡查】 notice with the session id, the card it holds, the silent duration, its last report and a copy-pasteable read-back call
At most one notice every 15 minutes; no polling of the workerFile-driven orchestration
`LEADER.md`, `SLOTS.md` and the `leader-mode` / `init-leader-agent` skills→Per-project Leader rules, per-role worker personas and a staged protocol, injected only when the text changes; a project without `SLOTS.md` falls back to `dev` / `review`
Prompt text only — the files grant no tool permissions
04Who is it for? When not to use it?
Good for
- A single person managing 20+ sessions; the design explicitly aims at that, not at replacing dsh’s own super agent
- Developers running several agents in parallel who need the global picture at a glance
- People who want the conventions reproducible across projects and teammates
Not for
- Background orchestration needs a long-lived host (web / desktop / sdk). headless is a one-shot CLI: the session is destroyed when the turn ends, so no completion report ever arrives.
- The Leader can steer a worker but cannot cancel one — `interrupt_agent` does not apply to worker root sessions, so a wedged worker needs a human in its own session.
- The client is coupled to the host version — dialogs, menus, icons and the tab bar are copied from host sources, and the session tab bar is a DOM overlay kept in browser localStorage, so host upgrades require syncing this repository.
05Compatibility, maintenance and safety notes
- Background orchestration needs a long-lived host (web / desktop / sdk). headless is a one-shot CLI: the session is destroyed when the turn ends, so no completion report ever arrives.
- The Leader can steer a worker but cannot cancel one — `interrupt_agent` does not apply to worker root sessions, so a wedged worker needs a human in its own session.
- The board only shows for project sessions: the leader and that project’s workers see it, an ordinary session in the same directory or a retired leader only gets a notice, and sessions outside any project get no board at all.
- The client is coupled to the host version — dialogs, menus, icons and the tab bar are copied from host sources, and the session tab bar is a DOM overlay kept in browser localStorage, so host upgrades require syncing this repository.
MIT · actively maintained (latest release v0.6.0, 2026-11-17) · developer preview
06Frequently asked questions
How do I install it on the web profile versus the desktop app?
Web: `dsh plugin --profile web add dsh-leader-agent`, then `dsh web`. Desktop: open Plugins in the left sidebar, click Add plugin, type `dsh-leader-agent` and click Install — or use the CLI bundled with the desktop install (`runtime/cli/bin/dsh plugin --profile desktop add dsh-leader-agent`), because the system `dsh` refuses to manage the desktop profile. Quit the app completely first; closing the window only hides it.
Which dsh versions and profile names are supported?
dsh `>=0.2.1-rc.2 <0.3.0`. The web profile must literally be named `web`: custom names are absent from `PROFILE_TEMPLATES`, so the install lacks the web-app bundle and the client half never loads. Desktop keeps its own profile (`$DSH_HOME/profiles/desktop`), and `DSH_HOME` must be absolute or start with `~/`.
Can the Leader stop a worker that went off track?
No. The Leader can steer with `message_worker`, which injects the message into the next step of the current turn, but there is no cancel — the host’s `interrupt_agent` does not apply to worker root sessions, and a wedged worker needs a human in its own session. `read_worker_session` shows what it is doing, read-only and without waking it.
What do LEADER.md and SLOTS.md actually change?
They contribute prompt text only and grant no tool permissions. `LEADER.md` holds the project’s orchestration rules for the Leader, `SLOTS.md` holds one section per role (`dev` / `review` / …) that is injected into the worker bound to that slot. Both are evaluated before every model request and appended only when their text changes.
Why does it need a long-lived host?
Background orchestration relies on the runtime waking the Leader when a worker’s turn ends. headless is a one-shot CLI that destroys the session at the end of the turn, so nothing is left to receive completion reports; use the web, desktop or sdk host.
07Related DSH workflows
dsh-agent-teams
by nanmicoder
AgentTeams 多智能体团队协作插件:自然语言驱动建团队、拉成员、拆任务声明依赖、成员间消息直达,Web 右上角活动面板
dsh-blender-plugin
by sixtysevenlf
Adds tools for Blender scene editing, viewport capture, render profiling, headless work, and verification.
dsh-free-search
by dddmuc
DSH 免费搜索插件,执行联网检索。
dsh-bili-scroll
by c1overzz
Schedules Bilibili discovery, reads subtitles, writes daily reports, and optionally posts a comment tagging you.
08Data and sources
A leader / worker orchestration bundle for DeepSeek Harness (dsh). The Leader splits work, dispatches it, tracks every w…
dsh plugin --profile web add dsh-leader-agent
**Identity decides the tool surface.** The Leader gets `spawn_worker` / `message_worker`; workers don't. The host trims…
This page is generated from the project’s public documentation, repository metadata and a structured parse of DSH Plugins; last verified on 2026-11-20. Found an error? Submit a correction.
Best DeepSeek Harness Plugins
Twelve plugins worth installing first — picked from the whole catalog, across every category.
