Skip to content

Someone who may only watch a session is not shown what it runs on (the provider's address and the state directory read as set by the profile); no client call waits for its answer for good, and the terminal's first run goes by the daemon's record as the desktop's does

Decision 923 · accepted · 2026-10-18 · issue #739

A viewer is not shown what a session runs on

Decision 918 made config.get naming a session an observe method, so anyone who may watch a pod session saw the profile's base_url (the model gateway's address, usually one inside the cluster) and the pod's state directory. The maintainer decided (2026-10-18): with only observe, the layers and rules are shown, and those two read "set by the profile"; with control they are shown as they are.

  • Where. The gateway answers config.get naming a session withholding when the connection's scopes lack control (Explain.withhold/2), on a pod and on a laptop alike; a laptop session's read "set by the session's owner". The answer, the explanation's text, its ladder for one key and its JSON are all built from the withheld layers, so no form of the answer carries the values.
  • What. Every key whose last part is base_url (the session's provider and each named provider's) and state_dir. Where they came from stays: the layer and the source (the worker profile and its variable). A pod's workspace and its repository's files live under the state directory, so in a withheld answer their paths name it <state directory>.
  • Marked. A withheld key carries "withheld": true; the desktop's Settings pane draws its value muted, with what it is shown to. The TUI's /settings lists neither key and draws what it is given; troupe config --explain --session prints the pod's text.
  • Not chosen: refusing config.get to viewers (they lose the rules, which are what they most need to read), and withholding the models and prices too (the maintainer named two values, and the models are on every event a viewer already sees).

No client call waits for its answer for good

Decision 916 gave the desktop's first run a stated wait; every other button that waited on one reply could still hang on a lost one (D131).

  • Every call has a stated wait. TroupeConnection.call rejects with TroupeCallTimedOut past callTimeout(method): 60 seconds, and longer for the methods known to take longer (session.create, session.fork and session.claim 3 minutes, session.erase 5, the provider and server checks, a merge and an upload 2), all in one exported table, callTimeouts. A caller names another for one call ({ timeoutMs }), as the DaemonClient methods behind busy buttons take. The sentence says which call, how long, and that it may still have been done.
  • A late answer is not taken: the call is forgotten when it gives up.
  • The buttons come back because each busy control already ends its wait in a finally; what changed is that the wait now ends. The setup screens keep their own, shorter, patience (Decision 916).

The terminal's first run goes by the daemon's record

troupe setup's screen sent the summary as one setup.answer after another and waited on each for good, ignoring every key meanwhile, Ctrl-C included; a reply the link gave up on after 30 seconds put the screen back on the summary with the error, though the daemon had finished the run. As Decision 916 did for the desktop:

  • An answer late past patience_ms (20 s) has the daemon asked where the run stands, over a connection of its own (status, Link.call_apart), since the daemon answers one connection's calls in turn. Finished, the screen ends as a finish would have; not, the keys come back with "The daemon has not answered in 20 seconds", and the answer is still taken if it comes.
  • A summary step that failed without the daemon refusing it is judged by the flow that comes back with the failure: a record newer than the one the screen began with, or a first run the daemon no longer wants, is a finish.
  • A finish whose answer never came ends at the sign-in for a plane, and for this computer as :finished: the session it started is unknown here, and the CLI says troupe resume opens it.
  • Ctrl-C leaves at any moment. With a summary being sent, it counts as having written.

The medium defects fixed at their cause

  • A daemon session follows a restarted daemon (D46). After a failure, a TUI window's or a headless run's connection to a daemon session asks where the daemon is now (relocate, Link.websocket/0, which reads daemon.json) before it dials again.
  • A draining pod takes no new session (D40). session.activate on a draining pod is refused as unavailable (draining) unless the session is already awake there; the plane gives the slot back and places the session again.
  • A cut result names read_output only to an agent that has it (D66). explore, answer, ask and the librarian are told what was cut and to narrow the request.
  • The pods log in at the installation's OpenBao mount (D55). The chart gives the operator bao.authPath, which it names to pods where it is not kubernetes; an empty TROUPE_BAO_TOKEN is no static token.

Governs: apps/troupe_core/lib/troupe/config/explain.ex, clients/gui/packages/client/src/connection.ts, clients/gui/packages/client/src/daemon.ts, clients/tui/lib/troupe/ui/setup.ex, clients/tui/lib/troupe/client/daemon.ex