v2.10.0
What's new
Every project gets a workspace
- A board of everything you work in. Pin a folder and it appears with its branch, uncommitted work, sessions and runs at a glance, grouped into what is active and what is quiet.
- Open one and stay there. A workspace view puts the project's run history, MCP servers and last commit in one place, so you stop guessing which folder the agent is standing in.
- Files follow the project. The file tree opens on the project the agent is working in rather than a saved root, and Sessions lists only real conversations instead of every background job that ever ran.
Workflows start from a pattern
- The page opens on the shapes worth copying. Fan work out to three reviewers, gate a release on the literal test output, or route to one specialist, each wired and ready to edit.
- A composer for every step. Opening a step gives it a prompt editor that wraps properly, its own settings beside it, and room to widen when a prompt gets long.
- The canvas says how it works. A legend covers adding, moving and removing a step, drawing and deleting a connection, and what happens when a drag is abandoned.
A skills library you can navigate
- Browse by source. Pick a folder on the left to see only its skills. Switch one on, or a whole source at once.
- See what just arrived. Recently added lists what was installed in the last two weeks, newest first, so a fresh skill is easy to find.
- Bring a skill you already have. Import points Husk at a file on disk and installs it into your library.
- An enabled skill is called automatically. The page now says so plainly: enabling a skill makes it available to the agent, which decides on its own when to use it.
Smaller things that were in the way
- Changes apply when you say so. Turning a skill or a plugin on offers Restart agent in the notice itself, instead of naming an action the page could not perform.
- Light mode reads properly. Drop targets, destructive buttons and the skills list keep their contrast when you switch themes.
- Typing goes to the chat you are looking at. Input routes to the live tab rather than relaunching the agent, and attachments no longer land in the terminal.
Release artifacts
| Platform | File |
|---|---|
| Linux AppImage | husk-v2.10.0-linux-x86_64.AppImage |
| Linux (Debian / Ubuntu) | husk-v2.10.0-linux-amd64.deb |
| Linux (Fedora / RHEL) | husk-v2.10.0-linux-x86_64.rpm |
| macOS (Apple Silicon) | husk-v2.10.0-mac-arm64.dmg · .zip |
| macOS (Intel) | husk-v2.10.0-mac-x64.dmg · .zip |
| Windows | husk-v2.10.0-win-x64.exe · .zip |
Plus SHA256SUMS and auto-update metadata (latest*.yml).
Verifying a download
# Integrity (any platform with sha256sum / shasum)
sha256sum -c SHA256SUMS
# Sigstore build provenance (requires gh CLI 2.49+)
gh attestation verify <file> --repo DorShaer/HuskKnown limitations
- macOS code signing is still pending an Apple Developer ID. See the README's macOS first launch section for the Gatekeeper bypass.
- Skills are read from the shared library on launch, so enabling or disabling one takes effect when the agent next starts.