Kernl 0.4.1: Claude Code in One Click, Chat Attachments, and Automatic Cleanup

Two days after 0.4.0, Kernl 0.4.1 removes three frictions you ran into daily: connecting your subscription, handing an agent a file, and a database that grew while nobody watched.

#Claude Code in one click

"Connect with my subscription" opens Claude's approval page in your browser. Approve it and Kernl picks up the session on its own. No terminal, no command to paste.

What matters is that it runs the official CLI's own sign-in, so Kernl still never sees the token. In Docker, where the approval cannot find its way back to the container, Kernl opens the page and you paste back the address your browser is left on. Still no terminal.

It also brings Claude Opus 5.5 to Claude Code agents: the agent SDK moves to 0.3.290, which bundles Claude Code 2.1.290. The 2.1.114 it replaces was refused by the API for that model, with "version 2.1.280 or newer is required".

#Documents, photos and video in chat

You can attach up to ten files to a message, in the chat, the 3D office chat and the office creator.

Images are scaled down, PDF and Word documents are read as text, and videos up to three minutes are sampled into frames and transcribed when a whisper model is already on disk. The point is simple: the model sees what you sent.

#The database stops growing on its own

System → Storage arrives: it shows what takes room in the database, table by table, what each retention policy would free, and lets you turn policies on or off, change their days, or run them now with progress. Extensions bring their own, and Cinema and Music add a cap on their catalog size.

The policies ship on and delete every night. They replace the script:cleanup agent, retired on the first boot. Out of the box:

• Completed runs of agents without a model (polls, scheduled checks): 14 days.
• Failed or cancelled runs of any agent: 60 days.
• The agent event log, old prompt versions and rejected evolution proposals: 30 days.
• Read notifications: 7 days.
• Daily model usage: 365 days.

Completed runs of agents that do use a model feed their memory and are kept: that policy ships off. If you want to keep more, review them in System → Storage before the first night.

#Switching an office off

A switched-off office keeps its agents, schedules and lot, but nothing in it runs: schedules and event triggers skip it, and its running runs are cancelled. The top agent is never stopped. A power button sits on the office cards and the office panel, and switched-off offices are dimmed in 3D.

Deleting an office is now "Destroy office", with a clearer confirmation and a demolition in the 3D world.

#One shared office, several projects

The project of the caller decides whose work a run is, so memory, inbox and files stay apart per project. A run without one is told to establish it first. The agent chat carries a project, picked in the tab when the office serves several, with its own thread each.

#What became usable

• The Workspace tab opens folders one level at a time. It used to load the whole tree: a real repo mounted from the host answered with 224,741 entries and 35 MB, and the tab hung on it. The root now arrives in a few kilobytes.
• LIVE reads better: each step is one line with its result and timing as tags, tool names are readable, a step you opened stays open while new events arrive, and the whole run is shown instead of its first 60 steps.
• The model menu lists only what works: providers that are not configured, or that found no models, are left out.
• Agents founded without a prompt get a full mandate, written by the model from their title, role, office and team.

#Before you upgrade

• A folder the kernel cannot see is refused. Pointing an office repo or an agent's __cwd_path__ at a host folder outside every mount of a Dockerized kernel now answers 400, instead of saving it and letting the agent quietly work somewhere else. To give offices their repos, set KERNL_PROJECTS_ROOT in .env.
• The primary model picks the engine in both directions. 0.4.0 moved an agent to the kernel's executor when its model was not Claude; picking a Claude Code (SDK) model now moves it onto the claude_code executor too.
• GET /api/agents/:id/cwd-files lists one level, not the whole tree.

#How to upgrade

curl -fsSL https://lifekernl.com/install.sh | sh

The same command installs and updates. On macOS and Windows the update runs from the "Update now" button. Full detail is in the CHANGELOG, and rpm, deb, dmg and msi binaries are on the release page.