Skip to content

A model request that sends nothing for a minute is given up on with a sentence naming the host, the daemon's log says where a turn is, and troupe doctor asks the model for one token

Decision 967 · accepted · 2026-10-29

The harness gives a model call llm_timeout_ms, five minutes, as Req's receive_timeout and as the agent's own timer, and nothing between the prompt and the answer is said at a level the daemon's log keeps. A request whose first byte never comes, because the address is wrong, the key is refused by something that does not say so, or a proxy holds the connection, looked exactly like a model thinking, for five minutes, and then said only that the model did not answer in time.

  • A first-byte limit, beside the whole call's. llm_first_byte_ms, 60 s by default, is a second timer the agent sets as the request goes out, in a turn and in a compaction alike. The HTTP adapters say {:llm_first_byte, ref} as the first chunk of any response arrives, an error's as much as the stream's, before they read it; a delta or a usage report counts too, for an adapter that sends no such message. The first of these cancels the timer. Where it fires, the call is stopped and counted as one that outlasts llm_timeout_ms is, and the turn ends llm_error with a sentence that names the host and the wait, blames the address, the key or a proxy rather than the model, and names troupe doctor as the next step. A 429 being retried is a host that answered, and is not given up on here.
  • Not a lower llm_timeout_ms. The other way was one shorter limit. A reply that has begun may pause for long thinking, and a reasoning model's first visible delta can come minutes after its first byte; Req's receive_timeout bounds the gap between packets and the agent's timer the whole call, and both have to stay long. What is short is the wait for the first byte, which no model spends thinking: a host that has said nothing in a minute is not going to.
  • Four lines in the daemon's log. troupe: <session> <agent> turn <n> preparing when the turn begins, ... asking <model> at <host> as the request goes out, ... first byte after <ms> ms, and ... answered after <ms> ms: <stop reason> or ... failed after <ms> ms: <sentence>. At info, the daemon's default level, so a daemon.log from a machine that shows nothing says where the turn stuck; debug would have been dropped. The summariser's are turn <n> summary .... Never the key.
  • troupe doctor asks the model. The key line asks the host to list its models; the new answer line asks the default model for one token, as a turn asks it, through the adapter a turn uses, with the key as a turn resolves it, once, within 20 s: ok with the round trip, fail with the sentence a turn would have ended with (the credentials refused, the connection failed, nothing sent in 20 s). The fake provider and a provider with no key are asked nothing, as the key line asks them nothing.

Governs: apps/troupe_core/lib/troupe/agent/state.ex, apps/troupe_core/lib/troupe/llm/request.ex, apps/troupe_core/lib/troupe/llm/providers/anthropic.ex, apps/troupe_core/lib/troupe/llm/providers/openai.ex, apps/troupe_core/lib/troupe/doctor.ex