A reply is looked at before it is acted on: a cut reply is asked again once, a cut tool call is answered rather than run, an empty reply is nudged once, a refusal ends the agent as refused, a prompt the provider refused as too long is compacted once and sent again, and a model error is a sentence somebody can act on¶
Decision 659 · accepted · 2026-09-20
Superseded in part
In part by 1024.
Acting
on content and tool_calls alone ends a turn the output cap cut in half as
finished — with half a sentence, or with nothing when a reasoning model spent its
allowance thinking — runs a tool on arguments cut mid-JSON, finishes a subagent
with an empty summary, and treats a refusal as a finish; and an opaque error makes
a blown context window, which is recoverable, read like a typo in a model name. So
handle_response/2 looks at the reply first. A :max_tokens stop with no tool
call gets one more request with a note that says what happened and asks for smaller
steps; the second time the agent ends output_truncated with what there was. A
:max_tokens stop with tool calls goes on, because every tool_use owes a
tool_result or the next request is refused, and a call cut mid-argument is
answered with an error naming the cause and not run. A reply with no text and no
tool call gets the same one nudge and then ends empty_reply. A :refusal stop
ends the agent refused. Each is a durable truncated event, and the note is a
user_input from source harness, which is what it is and what lets a replay
rebuild the conversation the model saw; a subagent that ends any of these ways
hands its parent what there was, labelled partial, as a budget stop does. On the
error side Provider.classify/1 names what a failure was — a context overflow by
the prose both providers use for it, rejected credentials, an unknown model, a rate
limit the backoff outlasted — and describe_error/1 says it in a sentence, which
is what llm_error.reason carries. Only the overflow changes control flow: compact
once (the compacted event says reason: context_overflow) and send the turn
again, since the failed request added nothing to the conversation; a second
overflow, or a conversation too short to compact, fails with a line that says what
to do. A 429 gets more attempts than anything else and waits what retry-after
said, up to two minutes, because a rate limit is a wait and not a failure. The two
guards are not replayed: a restart forgets them and grants the retry again, which
errs towards finishing.
Governs: apps/troupe_core/lib/troupe/agent/server.ex, apps/troupe_core/lib/troupe/agent/state.ex, apps/troupe_core/lib/troupe/llm/provider.ex, apps/troupe_core/lib/troupe/llm/providers/anthropic.ex, apps/troupe_core/lib/troupe/llm/providers/openai.ex, apps/troupe_protocol/lib/troupe/protocol/schema.ex, clients/gui/packages/client/src/transcript.ts, clients/tui/lib/troupe/remote/translate.ex