Skip to content

An organisation's enforced max_tokens holds every model call, the room a reasoning effort gives the thinking included: Anthropic's thinking is shortened to fit, is off below 1,024 tokens, and the session is told once on a config_note

Decision 880 · accepted · 2026-10-12 · issue #644 · supersedes 658

The maintainer decided the shape on 2026-10-12 (#644's comment): the thinking shrinks so the whole request stays under the ceiling, it is off where what is left is below Anthropic's 1,024-token minimum, and the person is told once, naming the organisation.

No call asks past the ceiling

The enforced value is carried as config.ceilings.max_tokens (with its source, beside the budget's max_turns and wall_clock) and goes on every request as Request.ceiling. Where the agent aims a request, its max_tokens is the lowest of what it asked, the named model's max_output and the ceiling, so the summariser's 2,000 is held too. Both HTTP adapters send Request.output_cap/1, the cap held to the ceiling, so a request built anywhere else is held the same way.

How the thinking fits (Anthropic)

The raise stops at the ceiling, and the thinking gets what is left after the answer's share, never more than its effort asks for:

  • The answer's share is 4,096 tokens, what the raise has always given it beside the thinking, or half the ceiling where that is less. Chosen over a fixed 4,096, which would turn thinking off under any ceiling below 5,120 (a 4,096 ceiling would think nothing), and over always half, which would cut xhigh to 16,384 under a 32,768 ceiling where 28,672 fits with the usual room for the answer.
  • Below 1,024 tokens of thinking (any ceiling under 2,047), thinking is off for the request: no thinking field and no output_config, which is what an effort of none sends. A model that thinks unasked (Opus and Sonnet 5 and later, Fable, Mythos) still thinks at its own default inside the ceiling, since most of them refuse to be told not to (Decision 805), and its blocks still go back to it.
  • Adaptive thinking has no budget field, so it is sent the level the room left holds, by the same thresholds a number of tokens maps to: high under 8,192 goes as low, max under 20,000 as high.
  • A ceiling the thinking fits under changes nothing: the request is the one Decision 658 and 780 describe. A ceiling that holds only the answer's room (minimal under 2,048: a budget of 1,024 in a cap of 2,048) shortens the answer, which is what the organisation asked for, not the thinking, and is not said.

A refusal names the thinking that was sent, the shortened budget or level.

An OpenAI-compatible provider

Its reasoning model counts reasoning against max_completion_tokens, and Troupe raises nothing for it, so the cap is the ceiling and the effort goes as it is: the model decides how much of the cap to reason with. Nothing Troupe sends is shortened, so the person is not told. The 400 that asks for the other field name sends the same held number.

Told once, on a new event

None of the events the clients show fits: budget_warning is a budget dimension near its end, drawn as "… used" and "nearly out"; watch_notice is ephemeral and the desktop app does not show it; onboarding_suggested is onboarding's own; and the warnings a config load gives go to the daemon's log only. So a new durable event, config_note: key (the setting), message (the sentence, shown as it is), and here limit, source (the plane's URL or the managed file's path), model, effort and thinking (the tokens given, 0 for off).

  • When: before the first call whose thinking the ceiling shortened or turned off, so it reads ahead of the answer it explains. The agent asks the adapter what the ceiling left of the thinking (Provider.thinking_held/2, which calls the optional callback thinking_held/1; only Anthropic's adapter has it).
  • Once a session, not once an agent: the first agent whose call is held claims it for the session (Log.first?, as an unpriced model is said), and the note goes under the root. A session stopped and resumed, or a daemon restarted, starts a new log process; the root's replay finds its own config_note and claims it again, so it is not said twice.

Governs: apps/troupe_core/lib/troupe/llm/providers/anthropic.ex, apps/troupe_core/lib/troupe/llm/providers/openai.ex, apps/troupe_core/lib/troupe/llm/provider.ex, apps/troupe_core/lib/troupe/llm/request.ex, apps/troupe_core/lib/troupe/config/org.ex