v1.29.1 (CLI-only)

Kesha Voice Kit 1.29.1

Fixed

kesha install --tts failed for every darwin-arm64 user

Installing TTS assets hard-failed with E_CACHE_CORRUPT and four SHA-256
mismatches on the KokoroVocoder.mlmodelc bundle:

error [E_CACHE_CORRUPT]: 4 of 37 model downloads failed:
sha256 mismatch for KokoroVocoder.mlmodelc/model.mil: expected c42be65f1e0b got 52f2febec450

The Kokoro ANE manifest resolved its upstream repository through a branch name
rather than a commit, so when FluidInference/kokoro-82m-coreml republished its
vocoder chain on 2026-08-19 every already-shipped version began failing at once,
for everyone. The pinned hashes were doing their job — refusing weights that were
not the ones we verified — and that was also the outage.

Every third-party model URL now resolves through an immutable commit, so an
upstream republish can no longer reach an existing install. No pinned hash
changed and every download is still hash-verified: the same guarantee, pinned one
level deeper. Two regression guards fail the build if a URL ever returns to a
mutable ref.

kesha install --plan reported the wrong download size on darwin

The published darwin engine is 64 514 608 bytes; the plan claimed 64 514 624.
Cosmetic, but the constant is now checked against the published release on every
pull request, so it cannot drift again.

Upgrading

bun add -g @drakulavich/kesha-voice-kit@latest
kesha install

If your TTS install failed while this was broken, re-run kesha install --tts <langs>
once you are on this version; nothing needs to be cleared by hand.

Note on 1.29.0

There is no 1.29.0 release. Its tag was consumed by a release-lane failure — a
permission contract between two workflows — and GitHub reserves tag names, so the
work ships here instead. Nothing was published under that version.

v1.29.1 (CLI-only). Engine: v1.24.11 (unchanged).