dsh-zcode-farm
让 DeepSeek Harness 代理查看集群中的所有 ComfyUI 实例，并将每个生成任务分派给最空闲的实例，而不是始终发送到最繁忙的单一端点。
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:BOWLUNA/dsh-zcode-farm#6356a4a8d99c8dd31504d2f58f957eeccc7d311bZcode Farm 以 dsh-zcode-farm 发布，当前版本为 1.0.1。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。
点击下方 Badge 复制 Markdown，粘贴到 README 即可。
验证 package.json 声明的 GitHub 仓库，即可管理这个公开页面。认领后，Hub 会优先安排当前版本的安全扫描，并在通过后公开展示结果。
认领这个 Plugin →English | 中文
Let the DeepSeek Harness agent see a whole farm of ComfyUI instances and dispatch each job to the idlest one.
Zero runtime dependencies · Node ≥ 22 · MIT
ComfyUI's own docs are explicit: one ComfyUI process executes one workflow at a time.
Real concurrency only comes from running one process per GPU and routing each job to the
least-busy instance. The ecosystem already has pooling libraries such as comfyui-orchestrator.
But when you wire that up to an AI agent, one piece is missing: the agent can't see the farm.
Measured on a real machine (a single moment on 2026-09-21, taken with this plugin's own probe):
| Instance | GPU | Free VRAM | Queue | Verdict |
|---|---|---|---|---|
18100 ← the only one dsh-comfyui can see | RTX 5090 | 0.8 / 33.7 GB | 1 running | too full |
18300 | RTX 5090 | 0.8 / 33.7 GB | 1 running | too full |
18301 | RTX 5090 | 0.9 / 33.7 GB | 1 running | too full |
18302 | RTX 5090 | 0.5 / 33.7 GB | 1 running | too full |
18303 | RTX PRO 6000 Blackwell | 99.4 / 102.0 GB | 0 | ✅ idle |
dsh-comfyui (77★, the most mature plugin in this space) talks to one endpoint, so every
generation request hits the busiest machine — while 102 GB of VRAM sits idle.
This plugin fills exactly that gap: let the agent see, then let it dispatch.
Three tools:
| Tool | Purpose |
|---|---|
comfyui_farm_status | Full farm view: per-instance GPU, free VRAM, queue depth, what's running, and who should get the next job |
comfyui_farm_pick | Choose without executing: filter by VRAM floor / queue ceiling / GPU-name hint / exclude list, and get the reason |
comfyui_farm_run | Load-aware dispatch: auto-pick (or force an instance), accepting a raw API workflow or a one-shot "prompt + checkpoint" text-to-image |
dsh plugin --profile web add dsh-zcode-farm
Restart the app and the agent gets all three tools.
- id: comfyui-farm
config:
primaryUrl: 'http://127.0.0.1:8188'
instances:
- { id: gpu-18301, baseUrl: 'http://127.0.0.1:18301', label: '5090 shard' }
- { id: gpu-18303, baseUrl: 'http://127.0.0.1:18303', label: 'Blackwell 102G' }
needVramGb: 8
maxQueueDepth: 3
queueWeight: 1000
probeTimeoutMs: 5000
probeRetries: 1
< needVramGb, or queue depth > maxQueueDepth → out.score = freeVramGb − queueDepth × queueWeightThe default queueWeight = 1000 is deliberate: queue depth dominates the ranking and free
VRAM only breaks ties. That matches the intuition — an idle small GPU finishes sooner than a
busy large one.
Prefer "biggest VRAM wins"? Set
queueWeight: 0.
dsh-comfyui: complementary, not competingdsh-comfyui | dsh-zcode-farm | |
|---|---|---|
| Scope | one endpoint | a whole farm |
| Strength | workflow library, skill packs, canvas, asset panel | awareness, selection, dispatch |
| Assembly row id | comfyui | comfyui-farm |
The row ids differ, so both can be installed side by side. Neither depends on the other.
fetch / AbortController only.dsh-comfyui's territory; this plugin only answers "who gets it".ssh -L forwards.>3s and answered in 106ms on the very next call. A single failure is not a verdict — we
retry, then report attempts so the agent can tell "down" from "hiccup".npm test # unit tests (pure functions, no network)
node probes/probe-farm.mjs # standalone probe: farm status only
node probes/probe-tools.mjs # load the entry and really execute the tools (read-only)
node probes/probe-tools.mjs --run # actually dispatch one job (smallest payload: 512² / 12 steps)
During development the host-provided peer deps (@deepseek-ai/schemastery and its deps) must be
present in a local node_modules/, otherwise the entry cannot be imported. This is not needed
when installed into a real profile — the host supplies them.
comfyui_farm_run covers minimal text-to-image only
(KSampler + CheckpointLoaderSimple + EmptyLatentImage + 2× CLIPTextEncode + VAEDecode + SaveImage).
For anything richer, pass a workflow.127.0.0.1:8188 → "currently focused instance") and its real target, you will
see two members with identical readings. That is honest reporting, not a bug — probing cannot
distinguish them.number returned by /prompt is a server-side cumulative counter, not a queue position
(measured: it returns 22 even when the queue is empty). Trust comfyui_farm_status for real queue depth.MIT
node test/run.mjs>=0.1.5-rc.2 <0.2.0-0 (see engines.dsh and the peer range)dsh plugin --profile web add dsh-zcode-farm@1.0.1