Open Systemone: System One plugin for CLIProxyAPI

Adds System One decision nodes to CLIProxyAPI. A decision request ({model, state, questions}) goes to the node chosen by the model prefix and the node's JSON answer comes back unchanged. Any System One-compatible server can be a node: a local model server (e.g. Winnow 12B) or a hosted lane (OpenCode, OpenRouter, v1m).

• Add nodes by URL, with several API keys (round-robin with fallback), extra headers and a per-node proxy.
• Configured models show up in GET /v1/models as prefix/model.
• An Open Systemone page in the management panel, built from the panel's own design: node list, Check (with latency), Fetch models from the node, and an example builder that sends test decisions.
• A Hugging Face tab that lists the decision models of the Jev Decision Index, shows the machine's CPU, memory, GPU and free disk next to each model's estimated size, and downloads models into models-dir. A Settings tab holds the Hugging Face token (for gated or private models) and the model folder.

Requires CLIProxyAPI v8 with plugins enabled. Streaming is not supported (decisions are a single JSON response).

Install

Plugin store: open Plugins → Store in the management panel and install Open Systemone.

Manually: download systemone_<version>_<os>_<arch>.zip from Releases, extract systemone.so / .dylib / .dll into the plugins directory (plugins/ by default) and restart.

From source (Go 1.26+ and a C compiler):

go build -buildmode=c-shared -buildvcs=false -o systemone.so .

For the official Docker image (Debian bookworm), build in a matching container so glibc fits:

docker run --rm -v "$PWD":/src -w /src golang:1.26-bookworm \
  sh -c 'go build -buildmode=c-shared -buildvcs=false -o systemone.so . && rm -f systemone.h'

Configure

Enable the plugin under its base settings, then add nodes on the Open Systemone page (PLUGINS in the sidebar) or in config.yaml:

plugins:
  enabled: true
  dir: "plugins"
  configs:
    systemone:
      enabled: true
      priority: 10
      huggingface-token: ""                           # optional, for gated or private models
      models-dir: "plugins/data/models"               # where Hugging Face downloads are saved
      nodes:
        - name: "Winnow Local"                        # display name (defaults to prefix)
          prefix: "winnow"                            # model prefix: winnow/Winnow-12B
          url: "http://127.0.0.1:3333/v1/systemone"   # full decision endpoint
          api-keys: ["key-a", "key-b"]                # optional; api-key: "..." also works
          headers:                                    # optional extra headers
            X-Custom: "value"                         # "{session}" = random id per request
          proxy-url: ""                               # "" = global proxy, "direct", or http/https/socks5 URL
          disabled: false
          models:                                     # empty list accepts any model under the prefix
            - "Winnow-12B"                            # plain id, or
            - { id: "Winnow-12B-bf16", name: "Winnow 12B BF16" }
        - preset: "opencode"                          # prefix defaults to the preset name
        - preset: "openrouter"
          prefix: "or"                                # → or/typesafe/jev-1.13
          api-keys: ["sk-or-..."]

Preset	URL	Models	Key
opencode	https://opencode.ai/zen/v1/systemone	jev-1.13-free	public (built in)
opencode-zen	https://opencode.ai/zen/v1/systemone	jev-1.13, jev-1.13-free	required
openrouter	https://openrouter.ai/api/v1/systemone	typesafe/jev-1.13	required
v1m	https://v1m.ir/v1/systemone	rev-latest, v1m-decision-engine	required

With several keys, requests start round-robin and move to the next key on 401/402/403/429/5xx or network errors. Other errors (e.g. 422) return at once with the upstream status and message.

Use

Send the decision payload to POST /v1/chat/completions (or /v1/systemone on hosts that expose it) with a normal CLIProxyAPI API key. The plugin claims any request whose model matches a node.

curl http://127.0.0.1:8317/v1/chat/completions \
  -H 'Authorization: Bearer <api-key>' -H 'Content-Type: application/json' \
  -d '{"model":"winnow/Winnow-12B","state":{"ticket":"I was charged twice."},
       "questions":{
         "refund":{"type":"noul","instructions":"Does the customer request a refund?"},
         "department":{"type":"choice","instructions":"Which department?","criteria":{"billing":"Payments","technical":"Defects"}},
         "urgency":{"type":"score","instructions":"How urgent?","criteria":["Routine","Time-sensitive","Emergency"]}}}'

Question types: noul (yes/no probability), choice (criteria is a label→description object), score (criteria is an ordered list of levels).

Open Systemone page

Opened from the management panel it reuses the panel's login; opened on its own (/v0/resource/plugins/systemone/ui) it asks for the management key and keeps it in this browser's local storage until you click Forget key.

• Node list with status, Check, Open, Delete.
• Add/edit node: preset or custom, name, prefix, URL, API keys, headers, proxy URL.
• Models: add, rename, remove, or Fetch from API (.../v1/models beside the URL).
• Example: build noul / choice / score questions, Send, and see the request, a curl, the response and the latency.

Check and Send on a saved, enabled node go through POST /v1/chat/completions with the first client API key, like a real client, so they show up in usage statistics. Unsaved changes and disabled nodes are tested directly through the management routes below (Check: /health, then a probe decision when a key is set).

Saving writes plugins.configs.systemone.nodes through /v8/management/config; the plugin reloads automatically.

Settings tab

• Machine: CPU, memory (the cgroup limit inside a container), NVIDIA GPUs (via nvidia-smi) and free disk space in models-dir. This is the computer CLIProxyAPI runs on, where models are downloaded.
• Hugging Face token: saved as huggingface-token; the page shows which account it belongs to.
• Model folder: choose where downloads go (models-dir) by browsing the folders of the CLIProxyAPI machine. Inside a container the page lists the folders mounted from the host, since only those survive the container and can be handed to a model server. A host path is opened where it is mounted, and a folder outside the mounts is refused.
• Model servers: progress while a model is set up, then its state, logs, stop/start and remove (which also removes its node).
• Downloaded models: progress of downloads started on the Hugging Face tab, stop, resume, Run and delete.

Hugging Face tab

• Jev Decision Index: models that publish weights, with rank, parameters, a rough memory estimate at the chosen precision (BF16, Q8 or Q4, plus 20%), score, latency and whether they fit the GPU, fit only in RAM or are too large. Filters show only the models that fit.
• Download: pick files, compare their size with free disk space (a download that does not fit is refused), follow the progress, stop and resume (partial files are kept as .part), delete. A repository without GGUF files lists its GGUF builds on Hugging Face; in a GGUF repository the file matching the Precision switch (and its vision projector) is selected.
• Run: starts a downloaded GGUF model in a llama.cpp server container (ghcr.io/ggml-org/llama.cpp:server-cuda, or :server without an NVIDIA GPU) and adds it as a node, so clients ask for <name>/<name>, or points an existing node at it so its models and aliases keep working (llama.cpp answers with the server name, so name it after the node's model id). When the file has no System One decision metadata, the plugin writes a copy (*.systemone.gguf) with the systemone chat template and the openjev readout, and sets non-causal attention when the original model's config.json declares it (modeling_overrides.attention.full_attention: block_bidirectional, as elephant-1 does). Models with a learned readout head are refused because llama.cpp cannot run them.

Running models needs the Docker socket inside CLIProxyAPI's container, for example in docker-compose.override.yml:

services:
  cli-proxy-api:
    gpus: all # lets the page see the GPU
    volumes:
      - /var/run/docker.sock:/var/run/docker.sock

The socket gives CLIProxyAPI full control of Docker on the host, so keep the management key safe. Model containers join CLIProxyAPI's Docker network and are reached by name, without a host port.

Management API

All routes need the management key.

Route	Purpose
GET /v0/management/systemone/presets	Built-in lanes (url, headers, models)
GET /v0/management/systemone/nodes	Effective nodes after presets and validation
POST /v0/management/systemone/check	{node, model}: /health, then a probe decision when keyed
POST /v0/management/systemone/models	{node}: upstream model list
POST /v0/management/systemone/test	{node, model, payload}: send a decision through the node
GET /v0/management/systemone/hf/status	Token account, downloads of this session, models in models-dir
GET /v0/management/systemone/hf/machine	CPU, memory, GPUs, free disk
GET /v0/management/systemone/hf/index	Jev Decision Index models with weights (?refresh=1 reloads)
POST /v0/management/systemone/hf/files	{repo}: files and sizes
POST /v0/management/systemone/hf/download	{repo, files?}: start or resume a download
POST /v0/management/systemone/hf/cancel	{repo}: stop a download, keeping partial files
POST /v0/management/systemone/hf/delete	{repo}: delete a downloaded model
GET /v0/management/systemone/hf/dirs	?path=: subfolders, whether it is writable, host mounts
POST /v0/management/systemone/hf/gguf	{repo}: GGUF builds of a model on Hugging Face
GET /v0/management/systemone/run/status	Docker access, model servers and setup progress
POST /v0/management/systemone/run/files	{repo}: downloaded GGUF files and whether they have decision metadata
POST /v0/management/systemone/run/start	{repo, file, mmproj?, name, context?, port?}: install and start a model
POST /v0/management/systemone/run/action	{name, action}: stop, start, restart or remove
POST /v0/management/systemone/run/logs	{name}: last lines of a model server

node is either {"prefix": "<saved prefix>"} or a full unsaved node (with url or preset).

Release

Push a v<version> tag. The Build workflow tests, builds linux/darwin/windows libraries, and publishes systemone_<version>_<os>_<arch>.zip plus checksums.txt in the layout the CLIProxyAPI plugin store installs.

License

MIT