Proxy: newly discovered models dispatch by default; require explicit activation #15

New issue

Open

opened 2026-10-26 09:09:13 +00:00 by rosa · 0 comments

rosa commented

2026-10-26 09:09:13 +00:00

Member

services/proxy/app/dispatcher.py refresh_catalog upserts every model the provider lists with $set {listed: true} and $setOnInsert {designations: [task], rank: 0, excluded: false}. A new model is dispatch-eligible the moment the provider shows it.
The selection query is {listed: true, excluded: {$ne: true}, designations: }. excluded is the opt-out lever and defaults false; the dashboard HAS a per-model disable toggle (ModelsDashboard.jsx), but the DEFAULT is active-by-absence.

Requested change: flip the default - $setOnInsert {excluded: true} (or empty designations) - so refreshed models land visible-but-inactive and require explicit activation.

Consequences to decide (filing, not scoping):

Task-class failover pools shrink to promoted models - exactly the safety wanted.
Providers adding models mid-week stop contributing traffic until someone reviews the list.
Fresh-install bootstrap: first-ever population would be all-excluded; consider a bulk-activate path or first-run special case so a new deployment is not empty.

Small change overall: one upsert default + dashboard copy. Not scheduled or specced here.

rosa (filing, not deciding)

Felix request 2026-10-26: new models should be DISABLED by default; today every refreshed model is active by default. Verified against shipped code before filing - the observation is exactly right, and the mechanism is one upsert: - services/proxy/app/dispatcher.py refresh_catalog upserts every model the provider lists with $set {listed: true} and $setOnInsert {designations: [task], rank: 0, excluded: false}. A new model is dispatch-eligible the moment the provider shows it. - The selection query is {listed: true, excluded: {$ne: true}, designations: <class>}. excluded is the opt-out lever and defaults false; the dashboard HAS a per-model disable toggle (ModelsDashboard.jsx), but the DEFAULT is active-by-absence. Requested change: flip the default - $setOnInsert {excluded: true} (or empty designations) - so refreshed models land visible-but-inactive and require explicit activation. Consequences to decide (filing, not scoping): 1. Task-class failover pools shrink to promoted models - exactly the safety wanted. 2. Providers adding models mid-week stop contributing traffic until someone reviews the list. 3. Fresh-install bootstrap: first-ever population would be all-excluded; consider a bulk-activate path or first-run special case so a new deployment is not empty. Small change overall: one upsert default + dashboard copy. Not scheduled or specced here. - rosa (filing, not deciding)

felix was assigned by rosa

2026-10-26 09:09:13 +00:00

No labels

No milestone

No project

No assignees

1 participant

Notifications

Due date

The due date is invalid or out of range. Please use the format "yyyy-mm-dd".

No due date set.

Dependencies

No dependencies set.

Reference: NSTERN/TreeGent#15

No description provided.

Rows
Columns