- clone each file with FICLONE, copy the data only where that fails
- skip the salvage when the copy would leave too little free space
- hold a per-container flock and one copy at a time across containers
- run WriteNspawnFlags on the blocking pool
The values come from the agent's start request. One with a line break made
async-nats panic inside the publish task, which would have stopped every
subagent's rows. They are now trimmed, and skipped when empty, over 64
characters, or not a valid header value.
The subagent daemon stamps Swarm-Model and Swarm-Effort on each row it
publishes. The list endpoint reads them off the newest row it already fetches,
so a long-running subagent keeps them past the stream's 24 hour window, and the
terminal page's menu and the agent panel's list show them beside the name.
The per-agent page had these and the swarm terminal did not; they need no new
backend. A cancel turn button beside send, shown while the agent is in a turn,
posts to the controller's existing cancel route and says whether a turn was
stopped. The terminal's badges add the effort and how long the agent has been in
its turn state. The expand tool output setting moves to the shared package so
both pages render the one component, and the swarm menu gets it. Docs follow.
hive-priv copies a stopped container's root to the agent's work dir and
renames it aside before passing nspawn --ephemeral; the agent's nix state
moves to the persistent profiles bind. Adds the agent-ephemeral-rootfs vm
test and a module-eval case for the nix state redirect.
review-classifier asked for a test of sw.js's new behaviour. sw.test.mjs loads
the worker in a sandbox with a fake cache and network: the install fetches past
the HTTP cache and keeps every asset, a redirected or refused asset fails it
with a logged reason, a navigation that reaches the network refreshes the cached
entry document, a redirected or failed one does not, and offline a navigation is
answered from the cache.
After a deploy a browser could keep the old entry document, which named a
bundle that no longer exists, and the missing script was answered with the
entry document, so the page was blank with only a MIME-type error to say why.
The hashed bundle is immutable and a missing file is a 404; the entry document,
the worker, the manifest, the icons and the theme's colors.css are revalidated.
The pages carry a content security policy. The service worker fetches its shell
past the HTTP cache, logs a failed install, refreshes the cached entry document
on each navigation and holds cache writes open with waitUntil. The proxy test
pins the cache rules, the 404s and the policy; docs/swarm/ui.md says what the
browser keeps.
The classifier asked for the docs to move with the three new panels. bao.md gets
a short 'Watching the store' section, including that a flat zero in the
delivered-lines panel means delivery is broken, not that bao is quiet.
The bao board was nine metrics panels, so a gap in log delivery looked like
a quiet bao. Three panels read the log store for openbao.service: lines
delivered over time, WARN and ERROR lines over time, and the recent WARN and
ERROR lines themselves. A module-eval case pins them by query.
The gateway answered every non-page request without a session with a login
redirect, which fetch and EventSource cannot follow across origins, so polls,
forms and streams failed with errors that said nothing about the session. It
now answers anything that does not ask for text/html with a 401 and a
WWW-Authenticate header, and the swarm UI reloads on that 401, from a wrapped
fetch (one place for every call) and from a probe after a stream dies. A second
reload within ten seconds is skipped so a page that keeps answering 401 cannot
loop. The proxy test and the docs follow.
