dsh-comfyui-supervisor

dsh-comfyui-supervisor

Lazy start / idle-exit lifecycle tools for a local ComfyUI server — as a DeepSeek Harness (DSH) plugin.

The plugin starts nothing at boot. It gives the agent three tools to manage ComfyUI on demand:

| Tool | What it does |

|---|---|

| comfyui_status | Is ComfyUI up? Which device? Is this instance managed by us? VRAM / version. |

| comfyui_start | Ensure ComfyUI is running: probe → spawn a detached supervisor → wait until ready. |

| comfyui_stop | Stop it and release RAM/VRAM. Only kills instances we started unless force=true. |

The bundled supervisor (assets/supervise.cjs, zero dependencies) keeps the process tree alive while

it is in use and kills it after N idle minutes — no more "leave ComfyUI running all day" or

"remember to start it before generating".

> Unofficial community tool. Not affiliated with ComfyUI, DeepSeek or AMD.

The problem

Local image plugins for DSH (and for every other harness) assume one of two things:

• ComfyUI is always on → 3–4 GB of RAM plus a resident GPU allocation while you are not even

generating, or

• you start it by hand → the first generate_image call fails, and you have to switch windows,

remember the batch file, wait ~40–60 s, then retry.

Both are bad. This plugin makes the lifecycle explicit and automatic: start on demand, exit when idle.

Install

# recommended: GitHub tarball (no npm account, no SSH key required)
dsh plugin --profile web add https://github.com/nizhouding-tech/dsh-comfyui-supervisor/archive/refs/heads/main.tar.gz

# pinned to the v0.1.0 release
dsh plugin --profile web add https://github.com/nizhouding-tech/dsh-comfyui-supervisor/archive/refs/tags/v0.1.0.tar.gz

> github: shorthand (dsh plugin add github:user/repo) resolves to git+ssh:// in pnpm and fails

> without SSH keys. Either use the tarball URLs above, or run pnpm config set git-protocol https

> first.

Restart DSH once. The three tools then appear in the tool table.

> Requires: DSH 0.2.x, Node 22+, Windows for the process-tree kill (the probe/start path is portable).

Configure

cordis.patch.yml of the installed plugin entry:

- insert:
    - id: comfyui-supervisor
      name: dsh-comfyui-supervisor
      config:
        baseUrl: http://127.0.0.1:8188
        comfyRoot: D:\ComfyUI
        idleTimeoutMinutes: 30
        startTimeoutSeconds: 300

| Field | Default | Meaning |

|---|---|---|

| baseUrl | http://127.0.0.1:8188 | ComfyUI endpoint used for probing |

| comfyRoot | D:\ComfyUI | Where ComfyUI_windows_portable and logs/ live |

| idleTimeoutMinutes | 30 | Kill ComfyUI after this much time without generation |

| startTimeoutSeconds | 300 | How long comfyui_start waits for readiness |

How it works

comfyui_start
   ├─ GET /system_stats  → up?  return immediately
   └─ spawn(node assets/supervise.cjs --root ... --port ... --idle-min ...)
          ├─ writes .supervisor.lock (single instance) + .supervisor.state (pids)
          ├─ starts ComfyUI, waits for /system_stats
          ├─ polls /queue and /history every 15 s
          └─ idle > idleTimeoutMinutes → taskkill /T /F, remove state files, exit