Adds a per-agent hyperhive.availableModels list option (default
[ haiku sonnet opus ]) rendered into the HIVE_AVAILABLE_MODELS env var
(comma-separated) so the per-agent web UI model quick-picker lists
exactly the configured models instead of a hardcoded set. Operators set
a shared default hive-wide or narrow it per-agent.
An assertion guards that hyperhive.model is present in the list so the
picker can always offer the model the agent is actually running.
The hive-ci-prefetch oneshot has RemainAfterExit=true and is wired to the
container only via wantedBy + before. Once it runs successfully it stays
'active (exited)' indefinitely, so systemd skips it on subsequent container
restarts. The runner-token file it wrote is never refreshed.
This breaks the runner after its first registration: the token written on
the first successful boot is either a placeholder (forge-core-token wasn't
ready yet) or a registration token that has since been consumed/rotated.
On the next container restart prefetch does not re-run, the stale token
persists, and the in-container register service fails with
'invalid_argument: runner registration token not found' — exactly the
symptom in the field (worked briefly, two orphan runners registered, then
permanently offline).
Add partOf = [ nixos-container@hive-ci.service ] so a container stop/restart
propagates to the prefetch unit, forcing it to re-run and fetch a fresh
registration token before the container comes back up. before= still orders
it ahead of the container start within the same transaction.
ensure_claude_dir creates the dir as 0755 but cannot re-chmod after
hive-agent-user-migrate chowns it to the agent user (EPERM — non-owner).
The activation script runs as root and can always chmod it. Add an explicit
'chmod 755 $homeDir/.claude' after the existing chown so existing 0700
dirs from pre-fix containers are corrected on the next container boot.
Without this, all agents with pre-existing .claude dirs show 'needs login'
in the dashboard even with working sessions, because hive-core cannot list
the 0700 dir owned by a different user.
WriteAgentForgeToken and WriteAgentMatrixToken write agent state files
under /var/lib/hyperhive/agents/<name>/state/. The hive-priv service
has ProtectSystem=strict which makes all of /var read-only by default,
but /var/lib/hyperhive was missing from ReadWritePaths, causing EROFS
on every token write.
Closes#1274.
Addresses mara's review: add a NixOS assertion that rejects
hsts.enable = true when no TLS mode is configured. HSTS over
plain HTTP is silently ignored by browsers; the assertion turns
a silent misconfiguration into a build-time error.
HSTS was unconditionally tied to hasTls. This is risky: enabling it on a
deployment that later loses TLS locks browsers out until max-age expires.
Add three options under services.hyperhive.gateway.hsts:
enable — bool, default false
maxAge — seconds, default 31536000 (1 year)
includeSubDomains — bool, default true
HSTS header is now only emitted when hsts.enable = true.
X-Frame-Options, X-Content-Type-Options, Referrer-Policy at server
scope on _, forge, and matrix vhosts. HSTS added when TLS is active.
nginx inheritance rule: locations with their own add_header (CORS API
endpoints like /.well-known/matrix/client, /_matrix/) are unaffected —
they already carry the headers they need. HTML-serving and proxy
locations pick the security headers up automatically.
matrix-avatar-sync ran on every boot (RemainAfterExit=false + path
trigger), uploading a fresh PNG each time. Every upload mints a new
mxc:// URI, which triggers a profile state event in every joined room
— resulting in timeline spam even when the avatar hasn't changed.
Fix: before uploading, compute sha256sum of /etc/hyperhive/icon.svg
and compare against the last-synced hash stored in
$HYPERHIVE_STATE_DIR/matrix-avatar-icon-hash. Skip the upload if the
hash matches. Write the hash after a successful avatar_url PUT so
subsequent boots are no-ops until the icon file changes.
Hash file lives in the agent's state dir — survives restart, cleared
on purge (so purge + re-provision gets a fresh upload). Delete to force
re-upload manually.
Closes#1231
On first boot (or after a wipe) hive-c0re writes forge-core-token only
after the forge container starts and the admin is provisioned. This
lags hive-c0re.service becoming active. The previous code bailed
immediately with TOKEN=placeholder if the token file was absent,
causing the runner to fail registration with 'token not found'.
Fix: merge both waits (core-token file appearance + forge API ready)
into a single 60s retry loop. The early-bail path is removed; the
script only exits cleanly if .runner is valid (writes placeholder)
or a fresh registration token is obtained.
When .runner exists but core-token is absent after 60s, we keep the
existing credentials (safe — the runner holds valid creds; next boot
will validate properly).
Closes#1224 (which tracks #1221).
The gitea-actions-runner's host-scheme job processes use the
service's environment PATH, not the NixOS login-shell PATH.
Without this, 'nix flake check' and similar steps fail with
'nix: command not found'.
- add pkgs.nix to environment.systemPackages (ensures the binary
is in /run/current-system/sw/bin)
- set explicit PATH on gitea-runner-hive service covering both
/run/current-system/sw/bin and /nix/var/nix/profiles/default/bin
bash and matrix MCP entries were both assigned to the same attribute in the
same attrset literal, causing nix evaluation to fail with "attribute already
defined". Merged both into a single lib.mkMerge list under one assignment.
- restore count in get_loose_ends: 'N local task(s):' instead of bare 'local task(s):'
- add cross-crate coupling comment to both mcp_loose_ends_dir() copies
- add comment in hive-bash-daemon service env explaining HYPERHIVE_HARNESS_DIR
is already injected via systemd.globalEnvironment by the meta flake
With ProtectSystem=strict (added in the previous commit), / is read-only
inside hive-priv. nix creates a temp result symlink in its cwd; without
WorkingDirectory that cwd is / (systemd default), causing:
error: creating symlink /.tmp.tmp-...: Read-only file system
Point WorkingDirectory at the StateDirectory (/var/lib/hive-priv) which is
always writable under strict. Nix drops its temp symlink there instead.
add services.hyperhive.gateway.tls.certDir option: operators with a
CA-signed cert (Let's Encrypt, corporate CA) point at the ACME output
dir instead of using the auto-generated self-signed cert.
- tls.certDir: host path bind-mounted r/o at /run/hive-tls/ in gateway
- tls.certName / tls.keyName: filenames within certDir (default: cert.pem / key.pem, matches nixpkgs security.acme layout)
- hasTls = selfSignedTls || certDir != null: publicScheme=https in both cases
- assertion: selfSignedTls=true + certDir set together is an error
- openFirewall: httpsPort opened in both TLS modes
- docs/gateway.md: TLS modes table + operator-cert section
- docs updated in swarm.md peer config reference in the cert TLS section
when using operator cert, swarm peers can omit certFingerprint —
standard CA bundle handles trust automatically.
isolated agents reach forge via http://forge.<domain> → nginx.
without the gateway there is nothing on port 80 to serve that
hostname. assert early rather than fail silently at runtime.
addresses argus yellow note on PR #1150.
when isolateContainers=true, isolated agents have dnsmasq as their
resolver — forge.<domain> resolves to bridgeIp. route HIVE_FORGE_URL
through nginx on port 80 instead of exposing the raw forge port.
- HIVE_FORGE_URL: http://<forge.domain> when isolated (nginx proxies)
- bridge firewall: open 80+443 for agents to reach nginx (gateway)
- remove forge-specific httpPort rule (no longer needed)
- update docs/gateway.md + docs/network.md
per mara's review comment on PR #1150.
When containers run in private netns (isolateContainers=true), host
loopback is unreachable so HIVE_FORGE_URL=http://127.0.0.1:3000 breaks.
- nix/modules/hive-network.nix: when isolateContainers is on + forge
is enabled, open forge.httpPort on the bridge interface so agents
can reach forgejo at bridgeIp:httpPort (forgejo binds 0.0.0.0)
- nix/modules/hive-c0re.nix: HIVE_FORGE_URL switches to bridge IP
when network.enable && isolateContainers; loopback path retained
when isolateContainers=false
- docs/network.md: add Forge access + Forge URL rows to effects table
- docs/gateway.md: rewrite HIVE_FORGE_URL section for both modes
Add opt-in WireGuard mesh support to services.hyperhive.swarm:
- swarm.peers.<domain>.wireguardPublicKey — peer's wg public key
- swarm.peers.<domain>.wireguardEndpoint — peer's UDP endpoint (optional)
- swarm.peers.<domain>.wireguardAddress — peer's mesh IP with prefix
- swarm.wireguard.enable — bring up wg-hive interface
- swarm.wireguard.privateKeyFile — path to host's wg private key
- swarm.wireguard.address — this host's mesh IP/prefix
- swarm.wireguard.listenPort — UDP listen port (default 51820)
- swarm.wireguard.persistentKeepalive — keepalive seconds (default 25)
When enabled, generates networking.wireguard.interfaces.wg-hive with
one peer entry per mesh-enabled swarm.peers entry. Opens listenPort
UDP on the host firewall. Adds wireguard_address to HYPERHIVE_PEERS
JSON so hive-c0re can use mesh IPs for intra-swarm routing.
Assertions guard against enable=true without privateKeyFile or address.
Also refactors networking.firewall.allowedTCPPortRanges from the
nested attrset form (which conflicted with the new allowedUDPPorts
line) to the per-attribute form.
docs/swarm.md: adds WireGuard setup section with key generation
commands, two-hive config example, NAT/keepalive notes.
Two argus review fixups:
1. Move isolateContainers assertion to an unconditional `lib.mkIf
cfg.isolateContainers` arm. The prior placement inside
`mkIf (enable && isolateContainers)` was dead code — the assertion
could never fire because both flags were already true by the time the
block activated. Now `isolateContainers=true; enable=false;` raises
a NixOS assertion error at eval time.
2. Add comment on HIVE_NETWORK_SUBNET noting it carries the host-side
bridge IP (e.g. "10.42.0.1/24"), not the canonical network address
("10.42.0.0/24"). Rust consumer must normalize via bitwise AND before
subnet membership checks or address arithmetic.
Adds `services.hyperhive.network.isolateContainers` (bool, default
false). When enabled alongside `network.enable`, activates:
- IP forwarding + NAT masquerade so isolated agents reach the internet
- nftables DROP rule blocking bridge-subnet → loopback (defence-in-depth
against compromised agent reaching the c0re dashboard)
- `HIVE_NETWORK_ISOLATION`, `HIVE_NETWORK_BRIDGE`, `HIVE_NETWORK_SUBNET`
injected into the hive-c0re service env; the Rust lifecycle reads these
to set `PRIVATE_NETWORK`, `LOCAL_ADDRESS`, and `HOST_BRIDGE` in each
agent container's conf
Config block rewritten as `lib.mkMerge [...]` — the prior `lib.mkIf //
lib.mkIf` pattern was invalid nix (mkIf returns a tagged value, not an
attrset; // on it is a type error). See docs/network.md for full design.
Privsep splits repo ownership: hive-priv (root) fetches the hive-core-owned
meta/applied repos via nix, and hive-c0re (hive-core) fetches the agent-owned
proposed-config repos. git/libgit2's dubious-ownership guard rejects both
('repository path is not owned by current user'), failing every rebuild.
Install a root-trusted gitconfig with safe.directory=* in each service's
HOME; both already have read access — this only satisfies the guard.
The migration chown stomped agents/<name>/config (the proposed-config repo)
to hive-core like everything else, but hive-agent-user-migrate only re-chowned
state/harness/.claude on boot, leaving config core-owned. config is owned by
the editing parent/manager agent (core only pulls from it), so chown it to
this user too. RO self-mount in an agent's own container fails harmlessly.
Two privsep follow-ups, both in the hive-c0re service:
1. The migration 'chown -R hive-core /var/lib/hyperhive' (which only
started running once ExecStartPre stopped failing 203/EXEC) stomped every
agent's bind-mounted creds — agents/<name>/{claude,state,harness,config}
are owned by the per-agent/manager users — logging all agents out with no
way back in. Scope it to everything *except* agents/, plus the agents/
dir node itself so c0re can still create new per-agent subdirs. Each
container's hive-agent-user-migrate activation chowns the contents back.
2. nix (prebuild 'nix build', flake-check, meta eval in c0re; nixos-container
update->nix in priv) writes its cache under $HOME/.cache. Both services
run as users with no home -> HOME=/var/empty (unwritable) -> Lix cache
init fails, rebuilds error out. Set HOME to each service's StateDirectory
(adding one for hive-priv).
hive-priv had no PATH, so the nixos-container it runs for every container
op (incl. lifecycle::list) failed ENOENT. build_all swallows that into an
empty list — the 'no managed containers' symptom. Give the helper the
minimal set nixos-container shells out to (nixos-container, nix, util-linux,
e2fsprogs) on top of the systemd/coreutils/findutils already in the unit
PATH, rather than the whole system profile.
Also route forge/matrix is_present() through priv_client::list_containers
instead of spawning nixos-container directly from unprivileged hive-c0re.
The state-migration chown invoked ${pkgs.coreutils}/bin/sh, which does not
exist (coreutils ships chown, not sh), so ExecStartPre exited 203/EXEC on
every boot. hive-c0re hit its start limit and never came up, so the gateway
returned 502 Bad Gateway. Call chown directly with the +- prefix (run as
root, tolerate failure) instead of going through a shell for '|| true'.
Tuwunel 1.6.1 appends " 💕" to every new user's display name via the
new_user_displayname_suffix config option. Set it to empty string so
agent and operator matrix accounts get clean names.
Existing users (atlas, mara, root) need a one-time display name reset.
Atlas's was already corrected via the client API.
