Happy Agent v0.4.114

Windows

• No more idle freezes. Happy watches your repositories to show live Git changes, and on Windows every file change it noticed started a sandboxed git status check that signs on through Windows' security service (LSASS). A log or build file rewritten every second, even one Git ignores, kept those checks running until the whole machine stalled. Happy now skips paths Git ignores until your .gitignore, index, branch or Git config changes, and checks quiet paths less often. Four idle repositories with constant churn went from about 338 sandboxed launches a minute to 8.
• Sandboxed commands run again. On some machines every sandboxed command failed at once with exit code 125 when another program had your user profile, or a folder above your project, open. Commands now run normally then, and when sandbox setup does fail, the error names the folder and the Windows error behind it.
• A slow MCP server no longer stops Happy from connecting. Startup and turns no longer wait for MCP servers: Happy connects to them in the background and uses what is ready, so Desktop stops restarting Happy in a loop.
• Sandbox setup asks first. Happy Desktop can now check whether the Windows sandbox is set up and ask before Windows shows its one-time administrator prompt. Happy can tell when you cancel that prompt, instead of treating it as a failure.
• No more console windows left open or flashing when agents start and stop processes.

MCP servers

• Sign in to HTTP MCP servers. When a server needs a login, an agent can send you a sign-in link; once you finish in the browser, the server reconnects. Servers that need a login show as needing one instead of failing.
• MCP servers get one tool-list request per connection, refreshed when their tools change, instead of repeated polling while agents work.

Vaults

• New: vaults. Shared folders of files and artifacts, with numbered versions, a Trash, and Time Machine-style retention by default. Agents can search, upload, download and copy vault files; copies and repeated uploads reuse stored content instead of storing it again.

Providers

• Custom OpenAI-compatible gateways work during onboarding. With a gateway such as z.ai, the provider check could fail with an empty response. Happy now reads gateway streams itself and keeps every event; when a check fails, the daemon log says why.
• Claude Code agents compact long conversations automatically, avoiding repeated "Prompt is too long" failures. Context usage now includes cached input.
• Account usage updates appear for Claude Code and Codex CLI agents, including Claude accounts configured through a settings directory with their own gateway.

Also

• macOS: sandboxed commands with network access can verify certificates through the system trust store again, so HTTPS works inside the sandbox from curl, Go and Swift.
• macOS and Linux: local commands respect configured network allowlists, denied domains and port limits in Auto and Workspace write modes.
• macOS: a new machine takes your account picture as its avatar the first time Happy starts.
• Messages from other agents or the system no longer split a busy agent's work into many small turns.
• Replying to a story shows the story beside your message, and the agent sees which story it was.
• Auto mode's safety reviewer follows Codex's latest policy, so an action you explicitly approved after seeing the risk is no longer refused. When it does refuse, it says how to approve that action.
• Agents can build full-screen HTML apps, such as games and canvases, that open edge to edge in Happy Desktop.
• Hidden collaborators can be archived when no task series still owns them.