2.0.4 — Upstream WhisperKit, custom vocabulary fixed

Speech-to-text is back on upstream WhisperKit

The app had been pinned to a fork of WhisperKit carrying a one-line fix. Without it, setting any custom vocabulary made every dictation return an empty transcript — recording ran, nothing got typed.

Argmax shipped that fix upstream (PR #514, co-authored by the same person who wrote the fork patch, with a more thorough implementation and ~200 new tests) in v1.1.0. So the fork is out of the dependency graph entirely and STT runs on upstream argmaxinc/WhisperKit 1.1.0.

Custom vocabulary is safe to use again — verified by dictating with a non-empty vocabulary, not just by reading the changelog.

Also in this release

• Codex hooks move to ~/.codex/hooks.json. Codex warns when hooks are split between hooks.json and an inline [hooks] table in config.toml; it now writes only the former. Reconnect Codex in Settings → Agents to move an existing install.
• Pi gets the persona and reply-language layers it was missing, so a spoken reply on Pi now matches Claude Code and Codex.
• No more SDK auth failed noise when Claude Code starts. Its MCP client probes for an OAuth endpoint before connecting; the server answered with an empty body, which the SDK parsed as a malformed OAuth error document and logged every time. Both that path and GET /mcp now answer with proper JSON.
• The persona line picks "a" or "an" correctly.

Install

Download the DMG below, drag to Applications, and launch. If macOS quarantines it:

xattr -d com.apple.quarantine ~/Downloads/OpenWhisperer-2.0.4.dmg

This build is ad-hoc signed, so macOS will ask for Accessibility and Microphone permission again after replacing an existing install.