Skip to content

The console's agent builder starts from a profile's agents as its bundle in effect has them, with who may run each, and picks a model from the ones the plane has been told of or seen run; the App's test mints, uses and revokes a token for a chosen repository and shows what it could do; a fix loop's commit names who the run is for in a Requested-By trailer the plane writes; and the desktop manager is the picker's rows

Decision 932 · accepted · 2026-10-23 · issue #594

  • The builder starts from the profile. admin.profile.agents {name} answers the profile's agents as its bundle in effect has them (Decision 910's composition through what it extends), each with the fields the builder edits (mode, description, model, max_turns, skills, permissions), from and channel (where it was published, which the builder opens), its definition, delegate and teams. A built-in primary the profile offers because its bundle has none of that name is a row too, builtin: true and no channel, to copy rather than edit; builtins is every agent Troupe ships with its file, read from troupe_core's priv/agents when the plane is compiled as the names already were, so the plane still never depends on core at run time. A platform admin's, as editing a bundle is.
  • delegate is the harness's rule, not a key. A definition has no list of whom it calls: a primary holding the delegate tool (or every tool) may hand work to any of its bundle's subagents, so that is the list, and nil for one that cannot delegate. Not chosen: a new frontmatter key, which the daemon's check (Decision 920) would refuse and no harness reads.
  • Who may run each agent is the grant, resolved. For every team granted the profile, may is Entitlement.resolve/2 of the agent against the grant's agent rows, as a session of that team's is narrowed, so a grant that allows other agents, or denies this one, says false. Shown, not edited: a grant is changed on Teams.
  • A model is picked from the ones the plane knows a provider serves. The plane does not ask a provider for its list: the key is the pods', and a plane that called providers to draw a form would be a new reach for one read. admin.profile.choices {profile} answers models, each {name, why}, the first reason found in this order: the profile's own llm.model, the models the organisation gives people's machines (client_model_*), the models the ledger has charged the profile's sessions for (Ledger.models_run/2, which only a provider that served them could have produced), and the models its bundle's agents name. Without a profile, the organisation's and every session's. The builder offers these and "the session's model"; a model already written that is not among them is kept and marked, not dropped. Not chosen: a free text field (the issue's "picked, not typed"), or a fixed catalogue in the protocol, which would be wrong for every gateway.
  • The App's test takes a repository. admin.github.check {repository} adds a fifth check, Repository, tried only when the four before it pass: a token for that one repository is minted with no permissions named, which GitHub answers with all the installation's own there, the repository is read with it, and it is revoked (DELETE /installation/token) before the answer. The answer's repository is {name, installation_id, permissions, revoked}, GitHub's own list of what the token could do; the token is in no answer, error or log, and is audited as github.token "for the App's test" with whether it was revoked. Asking for all the installation's permissions is the point of the question (what can the App do on this repository); the token lives for one read. Not chosen: asking for the fix loop's grants, which GitHub refuses whole when the App lacks one, and so could not say which.
  • Requested-By: is the plane's word. Beside Troupe-Session: (Decision 891), each commit the fix loop makes ends with Requested-By:, which the plane computes from its own rows and sends in every github.record answer (GitHub.requested_by/1): the latest trigger run that made the session names who fired it, a person by display name and email (Ada Lovelace <ada@example.com>), a GitHub login for a label a person put on (octocat on GitHub (trigger fix-label, team core)), the trigger and its team for a sweep, a schedule or a trigger's key (trigger nightly-fix, team core); a session no trigger made is its owner's. Made one line of printable characters, at most 200, on the plane, and the pod takes it only so; every line of the agent's message that looks like either trailer is taken out, so issue text that talks the agent into writing Requested-By: someone names nobody. Not chosen: the issue's author, which a hostile issue sets; or the subject alone, which is opaque to a reviewer.
  • The desktop manager is the picker's rows. Agents.tsx draws AgentPicker's rows, groups and actions (the four origin words, the badges, Start, Switch, Edit), with the manager's density (mode, what it stands in for, the windows running it) and its own Copy, New from this and Delete; a new agent saved offers to start a session on it.

Governs: apps/troupe_plane/lib/troupe/plane/admin.ex, apps/troupe_plane/lib/troupe/plane/ledger.ex, apps/troupe_plane/lib/troupe/plane/github.ex, apps/troupe_worker/lib/troupe/worker/github.ex