Kernl

0.4.6 · · Release on GitHub

Kernl 0.4.6: One Inbox, Accounts per Project, and Agents That Keep Working Past the Limit

The Kernl 3D world with its offices, each one a building where agents work

Three weeks after 0.4.5, Kernl 0.4.6 is about the places where you had to go looking: what the agents are waiting on, which account a deploy should use, and why a run stopped.

One Bandeja for everything that waits on you

Drafts to approve, questions from the agents, tasks your mail left and reminders that are due used to live in four tabs and the chief's office. They are now one list, most urgent first, with each item's actions beside it: approve or reject a draft, answer a question with the same card the chief's office uses, mark a task done, snooze a reminder.

Anything you would rather not handle yourself can be handed to an office: its lead gets a letter with the item, the error if there was one, and your note.

Each project brings its own accounts

A project now holds the accounts its deploys use: Cloudflare, Datadog, AWS, a git identity, a Google account linked by consent for Analytics and Search Console, and any deploy variables a service without a card of its own needs. Each one is tested against its service before it is saved, kept sealed, and handed only to that project's deploys. A GitLab or GitHub connection linked to a project is reserved for it.

Agents that survive the usage limit

Kernl can hold several Claude Code subscriptions and moves to the next one when an account hits its limit. The Live models pill shows each account's 5-hour, weekly and per-model usage as bars.

When every account is out, runs no longer end as "completed": a letter keeps its place, and a scheduled run goes again as soon as the limit lifts. Agent turns also get up to 16k output tokens, so reasoning models stop answering with nothing after a long tool result.

The calendar reads your mail better

Dates that only live in an attached ticket or booking now reach the calendar. A meeting mentioned in five mails of a thread, or in a sender's auto-replies, stays one item. And titles come in your language, translated when the mail is not.

Screens that fit

AI Overview, Agents, Evolutions, Ranks and Offices, Projects and the Agenda now show what they hold without scrolling, and the Agenda puts the calendar on top with today, the next seven days and what is important underneath.

Before you upgrade

  • Migrations run on start; nothing to do by hand.
  • The Work menu shows the Bandeja instead of Overview, Tasks, Reminders and Approvals. Approvals stays reachable as the full editor.
  • New offices default to claude-sonnet-5-5: claude-sonnet-4-5 is no longer served.
  • On Docker, restart with scripts/kernl.sh: it lets running agents finish before it restarts.

The full list is in the changelog.

Changelog

Upgrading from 0.4.5

  • Migrations run on start: agents 58–59, projects 9–13, storage 3 and tool-memory 2, plus new versions in the comms, Google sync, Cinema and GitHub extensions. Nothing to do by hand.
  • The Work menu is now one inbox. Bandeja replaces Overview, Tasks, Reminders and Approvals in the menu. Those pages still open from a link: Approvals stays as the full editor and /work goes to the Bandeja.
  • New offices default to claude-sonnet-5-5. claude-sonnet-4-5 is no longer served, so offices and chains that named it were falling through to their fallback.
  • Restart the Docker stack with scripts/kernl.sh. It pauses the offices, lets running agents finish, restarts and resumes them.

Added

  • Bandeja. Everything that waits on you in one list, most urgent first: drafts to approve, agents' questions, tasks and due reminders, with each item's actions beside it.
  • A project's own accounts. Cloudflare, Datadog, AWS, a git identity, a Google account linked by consent and any custom deploy variables, each tested before it is saved and handed only to that project's deploys.
  • Approvals that branch. A draft can offer a choice before you approve it, a failed send can be discarded, and any draft or failed item can be handed to an office to sort out.
  • Several Claude Code subscriptions. Kernl rotates to the next account when one hits its usage limit, and the Live models pill shows each account's 5-hour, weekly and per-model usage.
  • The agenda reads your mail better. Dates inside attached tickets and bookings reach the calendar, the same thing mentioned in several mails stays one item, and titles come in your language.
  • Storage ceilings for the graph. A size limit for the graph database and per-module budgets (Cinema first), plus retention for tool memory and old agent runs.
  • Agents pick skills by meaning and can run on their own model chain only, learning a cheaper model per kind of letter from their own runs.
  • Spam auto-label for IMAP and Gmail, conservative and review-only.

Changed

  • Screens that fit one screen. AI Overview, Agents, Evolutions, Ranks and Offices, Projects and the Agenda show their content without scrolling.
  • Offices spend fewer tokens. A dossier per ticket, letters that wait for news, long lessons focused on the task, and one nudge to continue when a model returns an empty reply.
  • A retired model leaves the chain. After three failures a model its provider stopped serving moves to the end of the chain and shows in the discarded list under its own name.
  • Friends shows what Tor is doing: the circuit building and the path a friend's request travels.
  • Header icons in one family, and the mail button opens what needs an answer across every account.

Fixed

  • Claude usage limits no longer end runs as completed. A run cut by the limit keeps its letter, and a scheduled run goes again when the limit lifts.
  • Reasoning models no longer answer with nothing after a long tool result: an agent's turn now gets up to 16k output tokens instead of the adapter's 4k.
  • Email HTML sanitizer updated to sanitize-html 2.18.1, closing six published bypasses.
  • Extension pages update in place: a new bundle reaches the browser without a hard reload.
  • An office serving several projects files each draft under the project it is about, and answers resume in the asking run's project.
  • Rebuilds keep the stack's compose overlays, so host mounts are no longer dropped.