andgineer / echo-words / 34030611212

96%

push

github

The pool streams, and a lost race repaints instead of splicing

llmbroker 1.8.0 protects an in-flight streaming call: every fastest_of
lane runs to a whole answer, the first complete one is authoritative, and
a lane that merely started talking first is replaced rather than left to
decide the answer. That was the objection that blocked streaming, so the
pool adapter switches to it.

A replacement arrives as StreamReplacementError, caught above the
LLMRequestError it subclasses so a complete answer is never read as a
pool failure and stepped up to the paid model. The provisional deltas are
dropped from the page through the reset the paid step already used, the
winning answer is repainted in their place, and it is that answer which
is judged, carded, rated and reported.

Measured on the smoke tier at the shipped budget: 60 answers of 62
attempts, whole-answer p90 3.6 s against 9.1 s for complete responses and
p95 16.0 s against 24.9 s, with the first visible text at a 0.795 s
median. No stream was replaced in 62 calls, so that path is covered by
tests and by the library's contract rather than by this run; the reading
of every packet item and the findings it carried are in the decision.

Co-Authored-By: Claude Opus 5 <noreply@anthropic.com>
Claude-Session: https://claude.ai/code/session_01HMxeEm5nWEAntNzaMwSLge

24 of 24 new or added lines in 2 files covered. (100.0%)

3044 of 3194 relevant lines covered (95.3%)

0.95 hits per line