Introducing the SystemOne Harness in SuperQode

Introducing the SystemOne harness in SuperQode: a dedicated decision layer for the native coding loop, with Jev from TypeSafe AI as the supported live model. Your coding model writes code and explanations. Jev evaluates typed questions about tool selection, proposed actions, and rubric checks. SuperQode owns control flow, policy composition, and execution permissions.

Building on our original Jev integration, the harness makes that decision layer selectable, adds Jev-assisted tool discovery, and keeps standalone decision sessions available in the same terminal.

A Harness for Structured Decision Models

Jev’s typed primitives give the harness answers it can inspect directly: Choice selects an option and returns a distribution and confidence; Score evaluates an ordered rubric; Noul returns a yes/no probability. SuperQode binds those answers to schemas and applies policy in Python.

We are building a reusable harness for this family of structured decision models. Jev is supported today; future models will need compatible adapters and validation before we call them supported. The division stays clear: coding models produce implementations, decision models evaluate state, and the harness coordinates both.

Selecting the SystemOne Harness

Open :connect, choose Connect a harness with your model, then select SystemOne and a native BYOK or local coding model. Live Jev calls use TYPESAFE_API_KEY from the environment that launched the TUI. Already connected? Select the harness directly:

TUI
:harness use systemone
:systemone status

The harness starts tool gates and discovery in shadow mode, recording Jev’s recommendations while existing permission behavior continues. You can also launch it from the shell:

Shell
export TYPESAFE_API_KEY="your-typesafe-api-key"
superqode --harness systemone

Jev-Assisted Tool Discovery

The harness defers optional schemas and exposes tool_search. SuperQode retrieves up to eight candidates using lexical matching; Jev selects from that closed set or returns none. Small catalogs can also be considered directly when retrieval finds no match. The sequence is explicit: retrieve candidates, evaluate with Jev, activate a schema, check permissions before execution.

Prompt
Call tool_search with query "fetch a web page over HTTP".
Report activated tools briefly. Do not fetch anything.

Shadow mode preserves lexical activation and records Jev’s recommendation. Enable reranking before launching a fresh session to let Jev select the schema:

Shell
export SUPERQODE_SYSTEMONE_MODE=shadow
export SUPERQODE_SYSTEMONE_TOOL_SEARCH=rerank
superqode --harness systemone

Reranking checks confidence, candidate fit, and separation between alternatives. Near ties can abstain, and a confident none activates nothing. Missing credentials or offline operation skip Jev; client or schema errors retain lexical discovery. Schema activation does not execute a tool.

SystemOne Harness in Action

Watch the SystemOne harness demo on YouTube.

Shadow Evaluation and Decision Evidence

Confidence is not permission. Hard YAML and permission-manager denials stay ahead of the Jev gate. For remaining calls, reviewed questions cover arguments, task alignment, destructive effects, and exfiltration risk; policy code composes the answers into ALLOW, DENY, or ASK. Shadow gates record that intended action without blocking or granting approval. Gate enforcement is separate from tool-search reranking, and thresholds need calibration on real tasks.

Retain sanitized evidence with SUPERQODE_SYSTEMONE_TRACE_DIR. Discovery traces live under tool-search/, separately from permission calibration records, and include candidates, the lexical baseline, typed answers, distribution diagnostics, and activated tools. The TUI shows activation and abstention; traces provide the detailed recommendation.

Shell
export SUPERQODE_SYSTEMONE_TRACE_DIR=/tmp/systemone-demo
superqode --harness systemone

Standalone Jev Decision Sessions

For standalone decisions without a coding provider, open :connect, choose Connect with SystemOne models, and select Jev. You can also connect a reviewed pack directly:

TUI
:systemone connect factory_route

Review a security-sensitive patch before merge.
Suggest the appropriate factory route.

The factory-route pack can suggest review for a security-sensitive patch without automatically switching the coding session. Typed output appears in a formatted JSON card, with important fields highlighted and the response opened at the top. Tool gates, routing packs, labelled evaluations, and rubric grading share this decision infrastructure.

Live Validation and Next Steps

Our live smoke tests confirmed the flow: shadow discovery recommended web_fetch while preserving lexical activation; reranking activated only web_fetch; an unrelated money-transfer request activated nothing. The standalone pack returned review, and the mounted TUI displayed the live decision correctly. These checks validate the integration flow, rather than general accuracy or safety.

Coding models write code. SystemOne models supply typed judgments. SuperQode owns the harness around both. Explore the SystemOne documentation, inspect the traces, and bring your own coding model. SuperQode is available on GitHub and PyPI.