Box v1 attaches nothing, so this is refused — and the sentence
you read was written by decide(), not by this page.
Pyodide in a Web Worker, running the same driver.py the
binary embeds. Its volumes are in OPFS, on your disk, in this
origin. Nothing was installed.
starting the box…
Writes reach the disk when the box is shut down, not before: Pyodide has no synchronous OPFS, and a worker sitting inside Python can run neither a promise nor a callback. Closing the tab without shutting down loses what it wrote.
Its Python is not quite the machine's. Pyodide ships
some of the standard library as separate packages, so
sqlite3, ssl and lzma are a
ModuleNotFoundError here and are not in a box on a
machine. Loading them is what hive.install() is for, and
that goes out as a request the harness answers — which this
placement cannot do yet.
The same loop the binary runs — the same prompt, the same two
tools, the same retry budget — in a second worker, calling the
model through this page. agent/mod.rs was not touched to
make this work.
This is the log, not a list this page keeps. Every
line above was rendered from hive-log/session.jsonl in
this origin, by the same hive-core that writes it.
Reload the tab and the conversation is still here; point
hive log at the file and it reads the same entries.
The key never enters either worker. A worker cannot
address this thread's heap — there is no shared object graph,
only a one-slot channel, and the only thing that ever crosses it
towards the box is box output. Both workers are sealed with
connect-src 'self'; this page is the only thing that can
reach the model, and it can reach nothing else.
What that does not survive is anything running on this thread. A key a laboratory uses belongs in the connector, which is the next step.