The hive-forge container leaves the host netns and listens on no TCP port: Forgejo serves HTTP on a unix socket the gateway proxies to, and the built-in SSH server is off. The httpPort, sshPort and openFirewall options are removed with a message each.
egress.forward defaults to reject while egress.intercept.enable is on (a
hive controller host turns it on) and to log otherwise, because reject
without the intercept would refuse all web traffic and an assertion
already forbids it. log stays an explicit choice.
The hive-squid VM test runs its log-mode half on the base node, so the
base pins forward = log at mkDefault priority and the reject
specialisation overrides it. A module-eval case covers the default on a
controller host, on a host without the intercept, and an explicit log.
The credentials of an agent's backend (ANTHROPIC_API_KEY and the like) move to their own leaf, swarm/agents/<agent>/backend-secrets, written through new PUT, GET and DELETE routes on /api/hives/{hive}/agents/{agent}/backend-secrets. GET says only whether any are stored. The backend environment at backend-env now reads back as stored through GET on its route, so it holds configuration only.
The agent's hive-agent-backend-env unit now runs on every agent with a store, reads both leaves and writes the configuration and then the secrets into the one file the harness and the subagent daemon load, so a secret wins over a variable of the same name. It removes the file when the store holds neither. The harness counts an agent whose environment holds ANTHROPIC_API_KEY or ANTHROPIC_AUTH_TOKEN as API-billed, so it tries the key instead of waiting for a Claude login; a real 401 still sets needs-login. The controller's policy may delete the new leaf.
Tests: the secret-client path, the controller routes, the script test of the unit with a stub bao, the module-eval cases and the harness's mode check. Docs: the swarm UI routes, the credentials reference and the API-key provider page.
The archetype installed Gradle, a JDK and android-tools but no SDK, no
ANDROID_HOME/JAVA_HOME and no nix-ld, so the aapt2 Gradle downloads could not
start on NixOS and an agent could not build an app.
It now builds an SDK (build-tools 36.1.0, platform 35, no emulator or system
images), sets ANDROID_HOME, ANDROID_SDK_ROOT and JAVA_HOME, and enables
programs.nix-ld. The SDK's unfree names are read off the SDK and added to
allowUnfreePackages, which adds to the claude-code allowance instead of
replacing it.
Closes#6696
dnsmasq relays the hook's stderr at LOG_DEBUG as dnsmasq-script, and the
field log shipper's journald receiver keeps info and above, so no lease
add or renewal was visible in the field. The hook now also writes each
line to /dev/log at info priority with identifier egress-lease. The
stderr print stays. The VM test reads the lines back with
journalctl -t egress-lease -p info.
The lock wait polls try_lock and gives up at the pass deadline (30 s), nft, ip
and nsenter get 10 s each before they are killed, and both pin units get
TimeoutStartSec=90s, which systemd does not set for a oneshot unit.
Each refused frame now goes through its own chain, spoof_<reason>, whose log prefix names the check that refused it (mac, vlan, dhcp, v4, v6, arp). The unmetered log lines of all reasons share one named limit, spoof_unmetered, so the budget stays one for all ports and reasons.
A stuck test in the shared job was cut at the runner's one-hour limit and hid the other tests' results. The matrix list is static; check-vm-test-list.sh fails CI when it differs from nix/vm-tests.nix.
full-setup 7200 -> 2400, semi-spread 10800 -> 3000, and 1800 on the four tests that used the driver's one-hour default. The runner kills a job at one hour and reports nothing, so a longer test timeout could never fire.
A start job of swarm-nats-auth that fails on a dependency (the server, or
the container's trust bundle, failing first) leaves the unit inactive
rather than failed, so Restart=always never applies. Its recovery was a
set of one-shot host kicks; when the last of them still found a
dependency down, the responder stayed dead and the queue refused every
client (nightly full-setup, runs 8058 and 7917).
upheldBy the units it requires: whichever comes up last starts it.
The test created its agents 0.5 s after pushing hyperhive/hyperhive, while
the forge still read the repo as empty, so the create's update_meta failed
and the next try came on the controller's 10-minute round. It now waits for
the two forge reads update_meta makes.
settle_hive_agents waited 3600 s, the CI job's own limit, so a stall was
killed rather than reported. It now waits 1200 s and also stops on a swarm
fault: a build the builder did not push to the cache, or an update_meta or
build_meta failure no later round got past. Timeouts and faults dump the
swarm builder's journal from heavyduty too.
Refs #6627