The suite now creates its own Anthropic and OpenAI deployments through
/model/new under run-unique names and deletes them at session end, so it no
longer depends on which aliases the proxy config serves. The names start
with the real model id because Claude Code 2.1.278 picks its request shape
from the model name and sends an unrecognized one a system-role probe that
Anthropic rejects with a 400 before the CLI falls back, which the proxy bills
as a failure row.
The echoed word is unique per run: the CI stack caches responses, and an
identical prompt across the three CI passes was served a $0 cache-hit row.
Tests move into the classes the e2e harness requires, the spend helper
asserts the row count it polls for, and the AI SDK cell accepts a single
text delta since the provider chunks the final sentence as it likes.