v0.3.2

timofriedlberlin released this 20 Apr 00:32

· 14 commits to main since this release

6a190ac

Highlights

make unified output token accounting non-overlapping: output excludes reasoning and total includes both
align Responses and Chat Completions adapters with output-minus-reasoning semantics
preserve canonical input invariant: total = cache_read + cache_write + new
add and update invariant tests for output and input token semantics

Canonical token model

Input: total = cache_read + cache_write + new
Output: total = output + reasoning

Assets 2