A real Linux sandbox
Install packages, run the suite, read the exit code. It boots in about two seconds, streams output line by line while it works, and is torn down when the turn ends.
It boots a real Linux sandbox when there is something to execute, and keeps every file, memory and skill in SQLite that outlives it. Close the tab — the turn keeps going, and tomorrow it still knows you.
| Fix the retry backoff in the queue module and prove it with tests.
Booting sandbox · warm in 1.9s …
Agent reasoning path
Drafting the pull request body …
Turn cost $0.0142 · session idles at $0
This one splits the two apart. Execution is disposable — a container rented for the seconds a command takes. State is not: files, revisions, memory, skills and schedules are rows in SQLite sitting with the agent at the edge, costing nothing while nobody is looking.
A session that is not being used is a few rows in a database. It wakes on request, does the work, and goes back to costing nothing — with everything still there tomorrow.
Each capability exists because there was somewhere durable to put things — which is why they shipped instead of becoming a roadmap.
Install packages, run the suite, read the exit code. It boots in about two seconds, streams output line by line while it works, and is torn down when the turn ends.
Pick a GitHub issue and it goes the whole way — edits, commits, and a PR body citing every command it ran with its exit code. You press the button that opens it.
Every write is a new revision, so a review-ready diff between any two versions is two rows compared. No git server sitting behind it.
What it learns about you and your codebase survives into an empty new session — no tool call, it simply already knows. Skills are procedures saved once and replayed.
Extract retry policy
3 files · 184 lines
npm test
7 passed · exit 0 · 4.1s
Open the pull request
branch pushed · body drafted
$ attach github.com/you/queue-module fetching tarball ......... 100% unpacking into SQLite .... 4,289 files indexing symbols ......... done issue #14 selected
Point it at a repository. It is pulled as a single tarball and unpacked into SQLite — one request, not one per file — then an open issue becomes the task.
Update retry policy
backoff is linear at queue.ts:44
Run the suite
7 passed · exit 0
It reads, plans and edits while you watch. Every tool call and every line of command output streams to the browser in the order it happened — and it takes notes as it goes.
Branch, commit, and a pull request whose body is the proof. An alarm then wakes the session to poll the review thread — days later, with the tab closed and the container long gone.
Where it stops: It can write, test, commit and push. Opening the pull request is a button you press. Nothing reaches someone else's repository without a person deciding it should.
Hand it an issue and close the laptop. It works the task through a real test suite and has the branch waiting, without a machine of yours left running.
Background agents wake on a schedule, triage what came in overnight, and answer the review threads on their own pull requests without being asked.
Five Durable Objects, one stateless Worker, a sandbox per command. Every piece of state is addressed by name, so a session belongs to a user by arithmetic.
“An alarm woke the session days after the container that built it was destroyed. It rebuilt the diff, addressed the review comment, and pushed — nobody watching.”