A pod session's output cap is its profile's llm.maxTokens, carried as TROUPE_MAX_TOKENS and narrow-only along extends; the model's own limit still bounds it and the organisation's machine ceiling does not reach a pod; an agent's own max_tokens only lowers it; 8192 stays the default; the cut-off note names the cap in tokens¶
Decision 1047 · accepted · 2026-11-17 · issue #1080
A pod's sessions asked every model call for 8192 output tokens, whatever the model allowed:
nothing on a WorkerProfile reached the session's max_tokens.
Where it is set. spec.llm.maxTokens, an optional integer of at least 1. The operator
gives the pods TROUPE_MAX_TOKENS, which the config's environment layer reads as
max_tokens (a value that is not a whole number refuses the configuration, naming the
variable). It is an llm field, so it needs llm.endpoint like the others, and a change
replaces the pods as a model change does. admin.profile.put refuses anything but a whole
number above 0. Along extends it narrows, as max_sessions does: the lower of the two
holds, and a child's higher number is kept as written, does nothing, and is said in the
preview and the effective view ("a profile cannot raise a limit of the profile it extends");
the same sentence now covers the github caps, which already narrowed silently.
How it is bounded. As before, a call asks for the lowest of the session's max_tokens,
the model's max_output from the catalog, and an organisation's enforced ceiling. The
ceiling never reaches a pod: it is a machine's (client_enforced), a pod reads no
organisation layer (Decision 869), and the profile is already the administrator's own word
for that pool; a second knob for the same people would only disagree with it.
An agent's own. An agent definition may say max_tokens:, which lowers the session's
for that agent's calls and never raises it, so a profile or a person sets how much any call
may write and a quick agent can ask for less. A report-writing agent gets more by its
profile allowing more. The summariser keeps its own 2,000.
The default stays 8192. Following the catalog where it says more would raise every laptop's spend per call, unasked, and a reply cut at 8192 is already recovered; a profile that writes long documents raises its own.
The cut-off notes name the cap: "...reached the output cap of 8,192 tokens...", the cap the cut call asked for, so the second attempt can be sized to fit.
Not chosen: an agent's cap above the session's (it would let a bundle's file outspend what the administrator set); a plane setting for every profile at once (the profile is the unit an administrator sizes).
Undone, a pod is back at 8192 whatever its profile says, or a child profile outruns its parent's limit unnoticed.
Governs: apps/troupe_core/lib/troupe/agent/server.ex, apps/troupe_core/lib/troupe/config/layers.ex, apps/troupe_protocol/lib/troupe/protocol/agent_definition.ex, apps/troupe_operator/lib/troupe/operator/resources.ex, apps/troupe_plane/lib/troupe/plane/fleet/extends.ex, apps/troupe_plane/lib/troupe/plane/fleet/profile.ex, charts/troupe/crds/workerprofile.yaml, clients/gui/apps/console/src/views/Profiles.tsx