v0.4.3
Changed
-
Anthropic route stops counting thinking blocks: the proxy no longer
countsthinkingcontent blocks on the Anthropic route
(thinking_block_countset tonullin both streaming and
non-streaming extractors). The upstream gateway
(api.code.umans.ai) does not populate
usage.output_tokens_details.thinking_tokensfor non-Claude models
(GLM, Kimi, Qwen), so the block count produced dashboard noise
("N req w/ think (unmeasured)") without actionable signal. The
thinking_tokensextraction pipeline is preserved for forward
compatibility. See ADR-0024. -
Dashboard gates "unmeasured" thinking label to OpenAI route:
the "unmeasured" fallback in the Performance tab now renders only
whenprovider === "openai". Anthropic rows never show the
unmeasured branch — even stale captures with
thinking_block_count > 0from before this change. -
Sub-tabs centered in Incidents and Config tabs: the sub-tab rows
on the Incidents and Config dashboard tabs are now horizontally
centered for better visual balance.