The token ceiling is ours, and a cut reply says so

[WaylandYang]

basil-k-aji-dev's #776, rebased onto dev after #764 turned the streaming buffer into bytes, with the number from the review changed. Fixes #760.

Their change, unaltered. The streamed calls send an explicit max_tokens instead of accepting whatever the endpoint defaults to, and chat_at_streaming returns Reply { text, finish_reason } so a reply cut at the ceiling can say so. extraction_open uses it: a chunk that hit the ceiling gets a drop row that says it, instead of one that says only that the reply would not parse. An answer that is too long is a number to change; a reply that does not fit the shape is not.

Both halves were verified to fail on their own. Removing the field:

assertion `left == right` failed: 不送这个字段，答案断在哪里就是端点的默认值说了算
  left: Null
 right: Number(...)

and deleting the one line that records the reason fails the same test from the other side.

What changed: 16,384 became 65,536. The original number was four times the longest answer, which is safe on this endpoint because it ignores the field with reasoning on. But the field now rides every streamed call, and max_tokens is not one thing: Anthropic counts thinking inside it, OpenAI's max_completion_tokens counts reasoning tokens. On an endpoint that counts reasoning, a ceiling sized for the answer is a budget for the thinking — which is exactly the lever #759 measured and rejected, where holding this model to about 16,000 reasoning tokens cost 885 statements against 729 and took table figures reaching no statement from 2% to 16%.

The largest completion measured on this path is 45,428 tokens. 65,536 sits above it, so the field can only ever guard against the endpoint's own 4,096 default and can never quietly become the cap we turned down. The constant's comment and the design page both say which quantity the number is measured against, because "four times the longest answer" would not lead a reader to expect it to bound the thinking somewhere else.

One test needed adapting to the rebase: #764's Unicode assertion now reads answer.text.

Their branch could not be rebased from here, so this carries their commit with authorship intact and #776 closes with it.

27 utopia-llm tests pass, clippy across utopia-llm and utopia-server is clean, fmt clean.

🤖 Generated with Claude Code