The startup normalize and _run_acme_preflight covered daemon startup and issue/renew, but the recurring collector poll called lib.acme.list_certs() without normalizing ACME_HOME. A non-daemon run (e.g. a manual run as the WebUI user) re-creating account.conf owner-only made every acme.sh --list exit 2, so the collector returned certs=[] and the UI showed no certs until the next issue/renew or daemon restart.
- collector: normalize_acme_home() before list_certs() so the poll self-heals
- issue pre-check: normalize before the direct lib.acme.list_certs()
- _parse_account_conf: read acme.sh v3 account.conf (not just .account.conf)
- _collect_acme: actionable status.error for the account.conf perm case
- install.sh: chown ACME_HOME conf files to the daemon user
Generated management-domain server blocks now include a
location /static/ aliasing webui/static/ with Cache-Control: no-cache
(ETag revalidation -> 304), nosniff, and a restrictive CSP, so SPA
asset requests no longer reach Flask. Flask's static route remains
the dev-mode fallback.
- lib/nginx.py + daemon/handlers/nginx.py: static_root render context
(the handler renders the template directly, so both paths need it)
- template: alias uses a trailing slash (nginx concatenates the
location remainder onto the alias value)
- install.sh: a+rX on webui/static plus execute-up-the-parent-chain
so the nginx worker (www-data) can traverse repo-in-$HOME installs
- tests: mgmt static location assertions (positive + non-mgmt negative)
- docs: AGENTS.md, architecture.md, security.md static-serving notes
acme:
- acme.sh chmods its tree to owner-only (700/600) every run, which
broke the two-user model: a tree left owner-only by one user made
every acme.sh call of the other exit 2
- normalize_acme_home() reopens group access (sudo chmod g+rwX,
files only — setgid dirs trip RestrictSUIDSGID); _run_acme_preflight
is the choke point before every daemon acme.sh call + startup
- acme service now runs as the daemon user; --log persists the raw CA
transcript; SYS_LOG=6 journals manual issue/renew runs
- timer daily-only: two runs/day landed inside ZeroSSL's 24h
validation backoff (Retry-After: 86400) — a permanent renewal lockout
- _collect_acme no longer raises on cert-list failure; reports
status.error (AcmeState.status) so the certs page can surface it
firewall: re-stamp the applied baseline on live zone mutations
(interfaces/services/rich-rules/masquerade/forward-ports) so cancel-all
reverts to post-mutation state, not a stale install-era snapshot;
set_masquerade syncs the declarative config for existing zones;
add_forward_port records toaddr only with toport
status: apply-all accepts {"force": true} (forwarded to the firewall
apply only); ApplyConfirm force checkbox; applyResultToasts() — the
errors map wins over the 200; ActionButton checks errors before the
success toast; dashboard uses ApplyConfirm
system_import: drift re-imports carry the existing apply-meta; first
import stamps the adopted content as applied (it is the running state)
— no phantom pending changes
nginx: get_config only re-saves when migration actually changed the
config (no more owner/mtime churn on every read)
install: repair mis-owned top-level system dirs (tmpfiles
unsafe-path-transition), warn with a full-repair command for deeper
mis-ownership
daemon/server: loop.get_exception_handler() (aiohttp API fix)
tests: 888 pytest + 24 node passing; ruff clean
The script restarts system services via systemctl, which requires root.
Previously it would only fail per-command when run by an unprivileged
user. Add an EUID guard at the top that re-execs the script through
sudo, preserving the original path and arguments.
systemd: pre-create volatile /run paths so vacuum-walld's ProtectSystem=strict namespace setup cannot fail with 226/NAMESPACE — RuntimeDirectory=vacuum-wall nginx plus a tmpfiles.d spec (installed to /etc/tmpfiles.d/) covering /run/firewalld and /run/nginx.pid. Drop /run/sudo from ReadWritePaths: NOPASSWD children never need it, and its absence crash-looped restarts after sudo removed /run/sudo.
webui: run the auth session check before mounting the shell so logged-out visitors never flash the sidebar or a protected page; router guard and sidebar now react to auth state, and the login page renders full-bleed.
ws: cap refresh->reconnect episodes at 2 consecutive failures; if the WS path stays dead after a token refresh, abandon reconnection instead of looping refreshAuth forever (UI keeps working via REST until reload).
api: GET /api/network/interfaces now includes loopback and returns per-interface {config, runtime}; dashboard reads runtime.state (carrier counts as up) and the interfaces page filters lo client-side.
daemon: re-collect nginx state after lazy config migration (cached list went stale when the on-disk format changed under it), skip system_import.nginx when config.json already exists (re-parsing vacuum-wall's own generated sites is lossy), and poll nginx (60s) / acme (300s) state so file drift self-heals.
- websocket.js passes the raw JWT as the Sec-WebSocket-Protocol subprotocol (no 'Bearer ' prefix): subprotocol names must be valid RFC 6455 tokens, and the space in 'Bearer <token>' made the browser reject the constructor with a SyntaxError.
- daemon accepts a JWT-shaped subprotocol plus the legacy 'Bearer <token>' form via _extract_ws_token; unit tests in tests/test_ws_auth.py.
- vdom.js applies inline styles through el.style (CSSOM) instead of setAttribute, which the management-domain CSP (no 'unsafe-inline') blocks.
- install.sh opens http/https/ssh on the public zone alongside WAN setup.
- docs (hoover.md, security.md) updated to match.
Auth seeding (last-resort guard)
- `_seed_builtin_admin()` in get_db() now skips when
VACUUM_WALL_SEED_BUILTIN_ADMIN=0 or when the users table already
contains any user — previously a fresh service start after a non-default
bootstrap (e.g. --mgmt-user alice) seeded a hard-coded `admin` with an
unrecoverable random password, shadowing the operator's account
- bootstrap_auth.py sets VACUUM_WALL_SEED_BUILTIN_ADMIN=0: bootstrap
creates the operator user itself on a fresh install, so exactly one
account exists and no seeded admin can appear
Frontend (session recovery)
- on page load/restore the in-memory TTL timer is gone, so a valid
7-day refresh token could sit in sessionStorage while the access token
is already expired server-side: the session `check` now attempts
exactly one refresh (POST /api/auth/refresh with the stored refresh
token) on 401 before treating the session as dead
- extract shared `_doRefresh()` used by both the `check` 401 fallback and
the `refresh` action (removes the duplicated rotation logic)
Tests
- update seeding tests to the new any-user-present check; add
test_seed_skipped_when_users_exist, test_seed_skipped_via_env,
test_bootstrap_flow_creates_exactly_one_user, and the auth-model JS
test suite (tests/test-auth-model.js)
Docs
- AGENTS.md: document VACUUM_WALL_SEED_BUILTIN_ADMIN
- architecture.md / hoover.md / security.md: describe the bootstrap
check 401 → one-refresh fallback path
Refresh/logout and token robustness
- drop the post-rotation refresh_tokens row delete in auth_refresh so
logout blacklists the current (rotated) refresh token; remove the
dead _clear_refresh_token_after_rotation helper and clear_active_refresh_token
- reject non-object JWT payloads in _extract_unverified_sub so crafted
Authorization headers return 401 instead of crashing with 500
SQLite user store
- make builtin-admin seeding idempotent: on a concurrent first start the
losing seeder re-checks, finds the winner, and returns instead of
raising IntegrityError
- per-thread sqlite connections + busy_timeout so Flask worker threads
don't hit cross-thread ProgrammingError / SQLITE_BUSY
- add LogsDirectory + /var/log/vacuum-wall to ReadWritePaths in both
systemd units so the fallback admin password actually lands on disk
Frontend
- skip apiFetch 401-recovery for public auth endpoints so a failed
login no longer logs out a valid session
- add /passkeys to the nav (passkey registration was unreachable);
remove the dead checkWebAuthnCapable export
- drop the CSP-blocked inline WS-URL script and the
__WS_URL_PLACEHOLDER__ plumbing; the WS URL is derived from location
Daemon / WS
- parse Sec-WebSocket-Protocol manually (web.Request.get_subprotocols
does not exist in aiohttp 3.13); X-Auth-Token is a custom-nginx
fallback only — docstring and security docs corrected
Install / system
- bootstrap_auth.py is now idempotent: preserves existing auth config
and syncs the admin password on re-runs (new reset_password helper)
- WebUI server block renders auth_basic off (the UI is JWT-protected)
- install.sh chown/chmod skips .git to avoid git dubious-ownership
breakage
- tolerate unreadable /etc/wireguard during system import
Contracts / docs
- create_user returns 409 on duplicate username per docs/api.md
- correct docs/api.md response shapes, docs/security.md blacklist
cleanup wording + one-refresh-per-user caveat, stale WS-URL
references, and the .htpasswd description
Tests: +7 regression tests (rotation/logout revocation, crafted-token
401, concurrent seeding); placeholder-substitution tests replaced with
serve-as-is SPA root tests.
- Move get_all_credential_counts import to top-level in daemon/handlers/auth.py
- Remove unused PROJECT_DIR in scripts/bootstrap_auth.py
- Fix leading space in api.js tryRefreshToken function
- Document WebSocket session binding limitation in docs/security.md
- lib/system_import: new module to import system configs into JSON at daemon startup
- daemon/server.py: call import_all() during startup for config reconciliation
- daemon/handlers/nginx.py: simplify add_domain auth handling, remove duplicate code
- scripts/install.sh: replace inline Python setup with curl-based daemon API calls; apply IP forwarding at runtime
- hoover: bump internal asset versions to v=8
- pages: bump asset versions to v=9
Flask's send_file uses mimetypes.guess_type() which needs the file extension
at the end of the path. Previously files were named acme.sh-3.1.3 and
htm.js-3.1.1 — now acme-3.1.3.sh and htm-3.1.1.js.
Also: update-vendor.sh now owns ACME_HOME installation, removing duplicate
logic from install.sh.
- update-vendor.sh now creates webui/vendor symlinks (htm.js)
- install.sh calls update-vendor.sh after package install
- Add vendor/.empty and webui/vendor/.empty as directory placeholders in git
mteehan