On this page
What it is Name a route The two consult forms The judge screen Declare an engine When it fails What to expect Related researchSystem One decision models
Bring a decision model, name a route, and read its limits.
What it is
System One is an experimental way to bring a fast decision model to CLIO Coder. It is off by default and sends nothing until you name a route. The main agent then gets one tool, consult, and the route also drives a screen beside the safety judge and, when you turn it on, triage of queued messages. System One never answers you, never edits anything, and never removes a confirmation.
Name a route
Shell
clio-coder systemone use jev
clio-coder systemone status
clio-coder systemone check jev --live
clio-coder systemone offThe choice takes effect on the next start. A list such as jev,memory is a failover order, and a route you did not name is never used. run --systemone <route> sets one headless run, and /usage health shows the status in a session.
| Route | What answers | What it needs |
|---|---|---|
jev |
A hosted engine | TYPESAFE_API_KEY |
memory |
The model in context.memory, as labels |
A memory target and model |
main or a chat target |
A chat route, as labels | That route's credentials |
| A declared engine | A decision server you run, with a named profile | An entry under systemOne.engines |
The two consult forms
Only the main session gets consult, at most 3 calls per turn.
| Form | What the agent sends | What comes back |
|---|---|---|
| Filter | Up to 256 files or items and one criterion | Ids and scores of matches |
| Questions | One to four yesNo, pick, or rate questions |
Probabilities, or a label on an LLM route |
A filter sees a file as its path and a digest; use grep for behavior inside a body.
The judge screen
When the safety judge reviews a call, System One asks whether the newest web and MCP results direct an AI agent. A positive reading can only turn a judge allow into a park on the approval card. LLM routes never screen.
Declare an engine
YAML
targets:
- id: decider-gpu
runtime: systemone
url: http://127.0.0.1:13306/v1
capabilities:
contextWindow: 4096
systemOne:
use: decider
engines:
decider:
kind: systemone
target: decider-gpu
profile: strands-deciderYou run the server. Name a profile: a generic engine never screens.
When it fails
A failed ask returns nothing and the agent proceeds alone. Three timeouts in a row open a five-minute breaker; see /usage health.
What to expect
consult filters a list against one criterion or asks typed questions, and the agent decides what to do with the answer. Between a third and more than half of filter matches can be wrong on the query sets tried so far, and nothing yet shows that consult helps. A hosted route receives what these uses send, including web and MCP result text at judge time. Redaction is best effort, so use a route you run yourself when that matters. An information flow rule that excludes the engine's target ends that route as flow-denied, and nothing is sent.
Read the research page for what has been measured and what is still open.