AIUI has carried 353/356 as a known-open item (W1.7) for long enough that
the three reds were treated as background noise. All three were the tests
being wrong. 357/357 now.
`injects web results into system prompt when enabled` asserted
`body.webSearch === true`. The code deliberately sends false there:
`proxyWebSearch = webSearchEnabled && !clientSearchSucceeded`, so when the
client-side search has already run and injected its results into the system
prompt, asking the proxy to search again would be a second redundant search
on every turn. The assertion predates that change. Fixed, and the other half
of the contract added as its own case — zero client results must still ask
the proxy to search.
The two seed-songs failures ("extracts 10 songs", and the conversation
regression built on the same fixture) were one wrong number:
`expected: { songs: 10 }` against an assistantResponse containing exactly
six `song_ext` entries and ending coherently on Treefingers. Not truncated —
just miscounted. The extractor was returning the right answer the whole
time. Counted from the fixture rather than from intent, so the number now
describes the input instead of contradicting it.
Full branch state after the main merge: AIUI 357/357, neode-ui 972/972
across 114 files.
Co-Authored-By: Claude Opus 5 (1M context) <noreply@anthropic.com>
