Skip to content

A profile's max_sessions is weighed against the most its pods could ever run (its registered hosts, or the cluster policy's maxReplicas, times the size class's sessions per pod), never the replicas running now, and saved either way with a sentence where it is above; admin.clients.put takes dry_run, which a console asks for only where discovery names it; a limit's refusal says its ceiling in the setting's unit; the audit page and profiles.list speak the units and gone pods the other pages do

Decision 1001 · accepted · 2026-11-15 · issue #957

What a ceiling is weighed against

The chunk 46 audit asked what happens when a profile's max_sessions is above what its pods can run at once. Against the most the pods could ever be, not the replicas running now. The scaler moves replicas with demand, down to none for a profile nobody uses, so a max_sessions above replicas × sessions_per_pod is the ordinary state of an idle profile, and a sentence on every save would be noise. What a ceiling can truly exceed is the most the substrate allows: for a profile on registered hosts, the hosts enabled for it times its size class's sessions per pod; on Kubernetes, the TroupePolicy's maxReplicas times the same, since the scaler never asks past it. Where neither bounds it the plane answers {sessions: null, by: null} and nothing is said: the scaler adds pods up to the ceiling itself (Scaler.at_once/2). admin.profiles.list carries it as at_once, and admin.profile.preview as capacity: {max_sessions, at_once, above}; the console's preview says one sentence with both numbers, and the write is saved either way, a ceiling being a limit and not a promise. The pods-full refusal (capacity, a wake with every pod full) names what the pods run at once and the profile's ceiling.

A dry run the console asks for only where the plane says it has one

admin.clients.put takes dry_run: true: every check made, the policy as the put would leave it with changes and affected, nothing written or audited, dry_run in the answer. The console's recommend-this-plane's-version preview counts from it. A plane from before ignores an argument it does not know and would make the write, so the plane's discovery lists the writes that take one (plane.dry_run) and the console asks only there; elsewhere it works the counts out as before. Where the plane reports no build at all the button is replaced by one sentence saying the plane is too old to recommend from. Not chosen: gating on the plane's version, which would guess at the next release's number.

Units on the refusal and in the audit

A limit's refusal stays forbidden (#941). Its limit is now the ceiling in the unit the platform's setting is read in, unit beside it (minutes for the idle timeout, a fraction where the ceiling is not whole minutes; days; switch); ceiling stays the team field's own seconds or days. The Team page's idle field is asked in minutes, as the Policy page asks the same setting.

Decision 991 kept the audit row in the plane's values ("a record, not a page"). That is amended: the Audit page shows a setting.put or setting.reset row by the setting's label with each value in its unit, from the plane's settings table, and a team.update row's limits as the team's Limits tab names them. The row holds the plane's values still; the key stays in the row's subject, and the search reads both.

Gone pods and names

profiles.list carries gone on each pod, as admin.profiles.list has since Decision 994, and a gone pod counts for nothing in capacity and healthy_pods, whatever its flags last said. Neither the desktop nor the terminal lists a profile's pods; the terminal leaves a gone pod out of the health it draws, so a profile is not "degraded" for a pod the operator removed, and the desktop's room is the plane's capacity. owner_name is on admin.session.erase.preview, Connections' sessions and the memory readers, through Names.decorate; the console's tables show the name with the subject on hover, as its session and audit tables do (Decision 992; initials are a message's, DESIGN.md §2).

SCIM's ServiceProviderConfig

GET /scim/v2/ServiceProviderConfig is served under the connector's token (RFC 7643 §5): patch supported, bulk not, a filter supported with maxResults 200, changePassword, sort and ETags not, the bearer token. 200 is what the plane now holds a filtered list to, with totalResults counting every match; an unfiltered list still answers every resource, which is how a provider reconciles. /ResourceTypes (users and groups) came in the same change, being small; /Schemas is left, since no provider seen asks for it.

Governs: apps/troupe_plane/lib/troupe/plane/fleet/scaler.ex, apps/troupe_plane/lib/troupe/plane/harness.ex, clients/gui/apps/console/src/views/Profiles.tsx