mirror of
https://github.com/getpaseo/paseo.git
synced 2026-10-23 17:52:49 +00:00
paseo-bot[bot]andkywoo26
a2b040869d
* fix(claude): read cumulative input usage from message_delta The context meter took input_tokens only from message_start. Backends behind a translating gateway report 0 there and the real cumulative counts on message_delta, so contextWindowUsedTokens collapsed to output_tokens. Keep the message_start counters per field and let a counter present on message_delta replace it, the same rule the Anthropic SDK MessageStream uses. A missing counter keeps the message_start value, so a delta without cache fields does not drop cache_read_input_tokens. Closes #4294 * test(claude): ignore a message_delta that precedes its message_start A canceled request can leave its terminal message_delta queued ahead of the next turn. Pin that such a straggler neither emits usage_updated nor leaks into the next request's counters. * test(claude): take the message_delta usage object as a single parameter Keep createMessageDeltaEvent(outputTokens) for existing call sites and add createMessageDeltaUsageEvent(usage) for the tests that set input counters on the delta, instead of a second positional parameter. * test(claude): verify late input usage through the real provider process Compare start-side and delta-side counters through two real provider turns against a local Messages API fixture, including omitted cache counters. The delta cases fail on the baseline with output-only context totals. Co-authored-by: kywoo26 <kywoo26@naver.com> * test(claude): always close the context usage fixture --------- Co-authored-by: kywoo26 <kywoo26@naver.com> Co-authored-by: paseo-bot[bot] <266920839+paseo-bot[bot]@users.noreply.github.com>