MIT DeepSeek Harness agent preset that plans first, delegates to a merged expert roster, coordinates parallel experts with a DAG taskboard and message bus, and gates delivery with independent review.
DSH integration
Native runtime
Author-claimed
Safety audit
Unaudited
Last verified
2026-11-20
License
MIT
01What can it help you accomplish?
Hand a multi-step request to a session mode that plans first and implements only through delegated experts
A PM plan (decomposition, dependencies, acceptance criteria), expert output collected on disk with short summaries on the bus, then an independent review plus a PM checkpoint before delivery
DSH users who want planning, delegation and review behind one preset instead of a single agent improvising edits
Keep several experts working the same repo without a hand-rolled scheduler
A dependency-DAG taskboard plus a mailbox bus — two zero-dependency Python tools — and up to three reusable lessons per task fed into later briefs
Users running parallel experts who need dependency ordering and compact hand-offs rather than pasted walls of text
Pull in a specific expert persona deterministically, without asking the model to choose one
Persona resolved through project > global > bundled file layers, or injected by a `/expert-<name>` gesture that costs zero model invocations
Users who know which expert they want and would rather type a gesture than negotiate with the model
02How to install into DeepSeek Harness
Prerequisites
- A host DSH on one of the three supported `engines.dsh` version rows — check the row before installing
- `python3` on the host, because the taskboard and bus tools are zero-dependency Python scripts
- A `web` profile target — the documented command installs into `--profile web`
Installation steps
- 01
Run `dsh plugin --profile web add github:mario841859784/dsh-expert-orchestrator`
$ dsh plugin --profile web add github:mario841859784/dsh-expert-orchestrator
- 02
Restart DSH, then pick "专家编排模式 / Expert Orchestrator" from the session preset picker
- 03
Optional: on the plugin settings page → Expert sources, download and enable the expert source packs (the host verifies pinned sha256 before unpacking)
Verify the integration
- The README states the outcome of a normal install directly: install the bundle, restart DSH, and the preset shows up in the session mode picker
03DSH integration and capability boundaries
Ships as a DSH agent-preset plugin: its own bundle patch declares the `preset-expert-orchestrator` row, so a normal plugin install adds an "Expert Orchestrator" session mode to the preset picker.
Merged-roster expert summon (`list_experts` / `summon_expert` / `summon_experts`)
an expert name or title plus a task brief→one summoned expert with its persona injected, or a batch of up to 8 with concurrency 4 and partial-success semantics
spawned sub-agents get a six-entry `toolFilter` deny list, so delegation stays one level deep`summon_expert` claims the task ids parsed from the brief (in_progress + owner) before the expert runsDependency-DAG taskboard and mailbox bus
subtasks with dependencies, written by experts→an ordered board with revision-checked writes, plus a bus that delivers short summaries while full output stays on disk
Delivery gate
finished implementation work→an independent reviewer verdict (≠ the implementer, ≤2 rework rounds) followed by a PM checkpoint before commit or delivery
delivery is blocked until both stages passZero-token `/expert-<name>` gestures
a gesture typed directly in the message, e.g. `/expert-backend-engineer`→the matching persona skill injected by a host pre-step hook — no model invocation and no catalog token overhead
gestures cover only the stable roster: file-layer and source-pack experts are never registered as gesturesExperience pool
a finished task→up to three reusable lessons per task, injected into future task briefs
lessons accumulate in `lessons.md`, which the deployer treats as user data (add-only) and never overwrites
04Who is it for? When not to use it?
Good for
- DSH users who want planning, delegation and review behind one preset instead of a single agent improvising edits
- Users running parallel experts who need dependency ordering and compact hand-offs rather than pasted walls of text
- Users who know which expert they want and would rather type a gesture than negotiate with the model
Not for
- The preset needs a host inside the three supported `engines.dsh` version rows. On a host that removed `SubagentRuntime.start`, plugin v2.9.0 and earlier cannot register the expert tools at all.
- The taskboard and message bus are Python scripts, so `python3` is a hard requirement on the host.
05Compatibility, maintenance and safety notes
- The preset needs a host inside the three supported `engines.dsh` version rows. On a host that removed `SubagentRuntime.start`, plugin v2.9.0 and earlier cannot register the expert tools at all.
- The taskboard and message bus are Python scripts, so `python3` is a hard requirement on the host.
- The preset never implements changes itself; it triages, plans, delegates and then only verifies read-only, so every edit comes from a delegated expert run.
- The maintainers' own 2026-10-19 A/B run measured +6.7pp accuracy with no regressions, but did not reach their pre-registered significance bar; the README states the numbers come from a single model and configuration and do not extrapolate.
MIT · actively maintained (latest release v2.9.8, 2026-10-18; last push 2026-11-13)
06Frequently asked questions
How do I install it in DeepSeek Harness?
Run `dsh plugin --profile web add github:mario841859784/dsh-expert-orchestrator`, restart DSH, then pick "专家编排模式 / Expert Orchestrator" from the session preset picker. The preset is declared inline in the plugin's own bundle patch, so a normal plugin install is enough — no manual preset directory is involved.
Does the orchestrator edit my code itself?
No. Implementation is always delegated to a domain expert from the merged roster; the orchestrator only verifies read-only afterwards, and delivery waits for an independent reviewer (≠ the implementer, ≤2 rework rounds) plus a PM checkpoint.
Do I have to install dsh-agency-agents or other expert packs?
Neither is required. The merged-roster protocol works without dsh-agency-agents, and only the 11 bundled core experts ship with the plugin. The four MIT-licensed upstream source packs are opt-in: download and enable them from the plugin settings page (Expert sources), where the host verifies pinned sha256 before unpacking.
Which DSH versions are supported?
The declared range is `engines.dsh >=0.1.7-alpha.2 <0.2.0-0 || >=0.2.0-rc.1 <0.3.0-0 || >=0.2.1-0 <0.3.0-0`. Hosts that removed `SubagentRuntime.start` need plugin v2.9.1 or later, because v2.9.0 and earlier cannot register the expert tools there.
Does it actually make the agent smarter?
The maintainers' own 2026-10-19 A/B run (108 tasks, one model) measured 86.7% → 93.3% accuracy with no regressions and a slightly lower cost per solved task — but it did not clear their pre-registered significance bar, and the README says the numbers do not extrapolate beyond that single configuration.
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 DeepSeek Harness (DSH) **agent preset plugin**: once installed, DSH gets an "Expert Orchestrator" session mode that ne…
so a normal plugin install is sufficient — installing the bundle, restarting DSH, and the preset shows up in the session…
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.
