Decision providers
Settings reference: behavior, defaults and safeguards for this page of desktop Settings.
Published unchanged from docs/settings/decision-providers.md in Moguri's source repository at commit b686ddc (0.32.0-alpha.1).
Technical reference for Settings / Decision providers. The app shows a short summary; this page keeps the full detail.
Decision providers and Local Max (page help)
Off by default, with no selected model. Explicitly select a bounded-model System One-compatible transport, exact model ID, output/deadline/call bounds and separate hosted-data consent. Saving never starts inference, selects a solver, changes Env or grants tools.
Current Max automatically approves tools; reviewDelegatedTools applies only to retained legacy bounded
configurations. Technical questions need kind: technical-choice, an exact question and identical
offered choices, with freeform disabled. They are not user-only facts. Steering starts off and selects
only original user-authored id/text alternatives; continue and askUser are reserved.
Available model IDs
The disclosure was previously labeled "Available exact model IDs". Use the exact IDs listed there in the policy JSON.
Warning
Confidence and selected probability are separate, uncalibrated advice. Jev remains unqualified because its documented API cannot enforce the required output-token bound. Remote/native (Sandbox) Max stays blocked and is never moved to the Local controller. There is no provider or model fallback, and a decision never grants a new permission or a Git approval.
Saved-policy readiness
After saving, explicitly select Autopilot Max in the composer. Inspect the model actor, original request, raw output, charges and cancellation under Artifacts / Decision-provider observations. User answers and approvals always win races against a decision.