Skip to content

A running session explains what is in force in it, from the layers it started with, through config.get naming it, which a pod answers its session's own token too, troupe config --explain --session, the TUI's /settings and the desktop's Settings pane; a GUI test waits for a control as for words; and a trigger that stops taking a key loses it

Decision 918 · accepted · 2026-10-18 · supersedes 913, 914, 915

A running session explains what is in force in it

What is in force is what it started with. A session resolves its configuration once, as it starts (Troupe.Session.build_opts/1), and the layers it was resolved from, every value and rule with where it came from, are kept with the session's name in the registry (Troupe.Session.resolved/1). config.get with a session_id answers from those (Settings.describe_session/1), in the shape config.get answers a workspace's, plus session_id and pod; explain (text or json, for key) adds what troupe config --explain prints for it (Explain.explain_session/3). Re-reading the files at the request was the other choice, and is wrong twice: a file changed since says what the session does not run, and a pod's terms and profile rules are not in any file.

A pod's parts are named. The worker passes, beside its overrides, where each came from (Troupe.Config.resolve/3's :sources): the environment is the worker profile's, which the operator writes from its llm, so its values are layer profile with the profile and the variable as source; the terms the session was created with are session; the platform's two switches are enforced, from the plane's platform settings, and so written to no scope; what the pod sets of its own is cli, from the worker pod. The profile's rules, which arrive as overrides, are in the layers' rules beside the files', so they are listed with their profile, as Permissions.scope/1 already named them; the configuration in force is unchanged, since a profile's rules are the platform's and nothing held them. A repository's files read as a pod reads them, no key marked trusted taken from them, which the answer's warnings say. On a pod no key has a scope: none of it is a client's to write.

A pod answers its session's own token. The guard keeps config.get the pod's own when it names no session, and refuses a session's token that; naming the token's session, it is that session's, held to the role as every method about the session is (observe). The gateway answers it on a worker as on a daemon, where every other config.* method stays the daemon's.

Shown everywhere a laptop's is. troupe config --explain [KEY] [--json] --session ID asks this machine's daemon, and for a session it does not have, the session's pod through the plane last signed in to, opened to read, which wakes nothing; a session asleep is said to have nothing in force. The TUI's /settings in a pod session reads the pod's answer over the session's connection and changes nothing (each change says why), and the person's own theme is not taken from it. The desktop app's team session has a Settings pane listing every key something set and every rule, each with where it came from. Not chosen: a pod pushing the explanation into the session's log, which would write configuration into a history the plane cannot read and make every wake an event.

A GUI test waits for a control as for words

Decision 915's check refused a bare says or link; a control is the same shape. It now refuses expect(button(…)), field(…) or row(…) asserted present at once, and a read of one found at once (button(…)!.disabled), unless marked // at once: <why>. The supports wait for it, await seesButton(…) (the console's seesField(…) too), and the test reads the control the wait answered. Absence is still not checked, as words' is not: expect(button(…)).toBeNull() cannot flake this way, only pass early, and the about 65 of them stay.

A trigger that stops taking a key loses it

Decision 914 left a key on a trigger whose source stopped taking one, dead but kept, and a listing then said has_key. Now Triggers.put/3 and the GitOps pass drop the key, its hash, salt and rotation, when a trigger is written with a source that takes none, and the listing's has_key is true only for a webhook that names no repository and holds one; a migration drops the keys kept before. A trigger changed back to a webhook has no key until a rotation gives it one, as 914 said, and now also says.

The rest of D130's small items

  • A refused push a later push resolves. A push that agrees with the row SCIM gave its externalId, after one of the pair was refused, audits person.adopt_agreed, and the pair is no longer listed by admin.identity.duplicates; one refused after that is listed again. Audited only while a refusal stands (Identity.refusal_standing?/2), so a provider's every sync writes nothing. Unlike an administrator's settlement (person.adopt), an agreement lets no later disagreeing push past the check.
  • Plane test databases. scripts/dev-down --test-databases [DAYS] drops the per-checkout test databases made more than a week (or DAYS days) ago, saying each, and keeps one in use; a checkout that still runs its suite makes its own again.

Governs: apps/troupe_core/lib/troupe/session.ex, apps/troupe_core/lib/troupe/config/explain.ex, apps/troupe_core/lib/troupe/config/settings.ex, clients/tui/lib/troupe/cli/session_config.ex, apps/troupe_plane/lib/troupe/plane/identity.ex