Found while adding one more input check (webhook URL scheme) and
noticing the error would render as literal {"detail":"..."} text in
the UI. Root cause is in the shared api() JS helper, not any individual
route: on a non-ok response it did `throw new Error(await res.text())`
— the raw response body, not the parsed message. FastAPI's default
HTTPException handler returns {"detail": "message"} as JSON, so every
`esc(e.message)` display in the panel (roughly 15 call sites) was
showing the whole JSON envelope, curly braces and quotes included, not
just the message inside it. Confirmed this wasn't already handled by
checking login()'s own catch block — it hardcodes a fixed string
instead of showing e.message at all, which only makes sense if e.message
was never fit to show directly.
This affects every validation message added the last several commits
(prices, HWID settings, node creation, provision-guide, hwid-limit) and
plenty from before tonight too — not something introduced by this
session, but something this session's run of new validation made worth
actually fixing rather than shipping another error message into a
broken display path.
api(): on error, try to JSON.parse the body and use .detail if it's a
string; anything that doesn't match that exact shape (plain text body,
malformed JSON, FastAPI's array-shaped 422 validation-error detail)
falls through to the original raw-text behavior unchanged, so nothing
that worked before regresses.
Also added the actual check that prompted this: webhook URL must start
with http:// or https://, rejecting things like a bare hostname or a
file:// URL (webhooks.send() never reads or forwards the response body,
so this was never a real exfiltration path, but it's an essentially
free guard against both a fat-fingered URL that would otherwise silently
never deliver anything, and the more deliberate file://-style misuse).
Verification: the api() fix is pure client-side logic with no backend
dependency, so tested directly under Node against a mocked fetch — 6
cases: the exact FastAPI {detail: string} shape extracting cleanly, a
non-JSON error body falling back unchanged, the Pydantic array-detail
422 shape not crashing the parser, malformed JSON falling back to raw
text, the 401/showLogin path completely unchanged, and the successful-
response happy path unaffected. AST-extracted admin_set_webhook_settings
out of api.py (still can't import it directly) and ran it against a
fake legal/env module — 6 cases covering both accepted schemes, both
rejected ones (ftp://, file://), a scheme-less bare hostname, and
confirming clearing the webhook with an empty string still works.
Co-Authored-By: Claude Sonnet 5 <noreply@anthropic.com>
Last remaining actionable row from the docs.rw comparison table pulled
two commits ago: "Webhook event support — Users + nodes (Remnawave) /
Users only (Marzban)". Every webhook we send is subscription/payment
events — user-side only, same as Marzban, even after last commit's
revoke/hold/resume additions. Zero node events.
node.added on creation, node.deleted on deletion (captures the node's
label before it's gone, since delete_node doesn't return the row),
node.enabled/node.disabled on the PATCH route — but only when the
enabled field actually changes value, not on every save. Editing just
the label, or PATCHing enabled to the same value it already had,
correctly fires nothing — checked this specifically since a naive
"enabled is in the request body" check would have spammed an event on
every harmless edit of an already-enabled node.
Verification: same two-part approach as the subscription lifecycle
webhooks. AST-extracted the actual admin_update_node() body out of
api.py (still can't import it directly) and ran it against a fake
db/webhooks module — 5 cases: enabling, disabling, a same-value no-op
save, and an unrelated-field-only edit, confirming the webhook fires
exactly when and only when it should. Then a real local HTTP server for
all four event types through the actual webhooks.send(), receiver-side
HMAC recomputed independently from its own copy of the secret and
compared against X-Signature, not trusted from the sender. README's
feature list had also fallen behind the last three commits (webhooks,
hold/pause, subscription search never got a bullet) — caught up all
three while I was in there, not just the one this commit adds.
Co-Authored-By: Claude Sonnet 5 <noreply@anthropic.com>
Another line off the fresh docs.rw comparison from last commit: "User
Management Filters — Extended selection (Remnawave) vs Minimal options
(Marzban)". The subscriptions table had none at all — no search, no
status filter, just the raw list with a server-side limit=200. Fine
with a handful of test subscriptions, useless once a real business has
a few hundred customers and support needs to find one person's row.
Pure client-side: the full list was already fetched in one call
(/admin/api/subscriptions), so filtering it in the browser needs no new
backend route and can't regress anything server-side. Refactored
loadSubscriptions() to keep the fetched list in allSubs and render
through a separate renderFilteredSubs(), which the existing
revoke/hold/resume refresh calls now go through too — so the search box
and status filter stay applied after an action instead of resetting the
view. Search matches username, tg_id, node label, and plan label as one
lowercased substring check. Status filter (active / on hold / expired-
revoked / all) reuses the exact three-way split statusBadge() already
draws, via a new subStatus() helper — same custom .dd dropdown as
everywhere else in the panel, not a native <select>.
Verification: extracted the actual subStatus()/renderFilteredSubs()
filter predicate out of admin.html — not a reimplementation, diffed it
against the file to confirm byte-for-byte match — and ran it under Node
against four mock subscriptions covering all three statuses, including
one with a null username (the real shape for gift-redeemed subs with no
Telegram username set) to make sure the search doesn't throw on that.
13 checks: plain search, case-insensitivity, tg_id/node/plan matching,
no-match, each status filter alone, and two combined search+status
cases. node --check on the full extracted script, div-tag balance on
the whole file, both clean.
Co-Authored-By: Claude Sonnet 5 <noreply@anthropic.com>
Pulled a fresh copy of docs.rw's own Remnawave-vs-Marzban comparison
table (not working from memory of an earlier read) to check what's
still genuinely different after tonight's run of fixes — most rows
already match or beat both panels (multi-admin, 2FA, HWID limits,
backup/restore, host sorting, config validation, node autonomy, on-hold
status as of a few commits ago). One concrete, bounded, unclaimed row:
"Security measures in documentation" lists CF zero trust / custom path
/ Telegram OAuth / 2FA for Remnawave, nothing for Marzban. We already
had 2FA and rate-limiting; custom path was the missing, actually
implementable piece — everything else in that row is deployment
guidance, not panel code.
New ADMIN_PATH env var (config.py, defaults to "admin" — every existing
install keeps working exactly as before with zero action needed). The
page-serving route moves to whatever path is configured; root() on
PANEL_DOMAIN only falls through to serving admin.html when ADMIN_PATH
is still the default, otherwise it shows the same branded landing page
every other domain gets — so a scanner or a human guessing "/admin"
finds nothing once this is set, not even a redirect that confirms
something lives there.
Deliberately scoped to ONLY the page route. /admin/api/* stays fixed —
it's already behind real cookie+session auth (verified this while
auditing: every mutating admin route either calls require_admin() or
the equivalent _require_current_admin(), checked programmatically via
ast rather than trusting my memory of having added the check everywhere
— found nothing actually missing, which is itself worth knowing, not
just assumed). Moving the API namespace too would be a much bigger,
riskier rewrite of every @app decorator in the file for no real security
gain over what auth already provides.
Deliberately NOT exposed in the Settings UI, unlike almost everything
else made live-editable tonight. This one genuinely needs a process
restart to take effect (FastAPI resolves routes at import time, not
per-request), and a typo saved through the UI followed by a restart
is a real self-lockout risk with no web-based way back — same tier as
PANEL_DOMAIN/SUB_DOMAIN, which are also .env-only for the same reason.
.env + SSH is the correct blast radius for a setting that can lock you
out.
Verification: config.py's normalization (strip slashes, empty/lone-
slash/repeated-slash input all falling back to "admin" rather than
accidentally producing a route at bare "/") tested directly — 8 cases.
AST-extracted the updated root() out of api.py (still can't import the
module locally) and exercised its actual branching with a mocked
FileResponse/legal/request — confirmed the default case is byte-for-
byte the old behavior and the custom-path case stops serving admin.html
on PANEL_DOMAIN's root. Added a dedicated CI step that does what only a
real FastAPI import can prove: with ADMIN_PATH set, /xyz123secret is a
registered route, plain /admin is NOT (not just supplemented — actually
gone), and /admin/api/login is untouched.
Co-Authored-By: Claude Sonnet 5 <noreply@anthropic.com>
Found while re-reading the subscription lifecycle routes: payment.paid
and subscription.granted_by_admin fire a webhook, but revoke (which has
existed the whole night) and the two new hold/resume routes did not.
Inconsistent for anyone actually wiring this into a CRM/support tool —
they'd see a subscription get granted but never find out it was later
paused, resumed, or cut off entirely, since only the "gains access"
side of the lifecycle was ever reported outward.
Three new events, same shape and delivery as the existing ones:
subscription.revoked, subscription.held, subscription.resumed. Added
right where the DB/xray state change already happens in each route, so
they're conditioned on the action actually succeeding (a hold attempt
on an already-held/expired subscription 400s before ever reaching the
webhooks.send call).
Verification: webhooks.py itself is unchanged — this only adds new call
sites with new event-name strings, so re-verified the exact thing the
original webhook feature proved: stood up a real local HTTP server,
fired all three new events through the actual webhooks.send(), and had
the receiver independently recompute the HMAC from its own copy of the
secret and compare against the X-Signature header it actually got, for
all three — not just trusting that the sender computed something.
Checked the JSON envelope and data payload match what each route sends
byte for byte. api.py itself still can't be imported locally, same wall
as always; the new lines were checked by reading the subscription-row
shape they pull from (tg_id/node/plan are all real columns already
confirmed present in every prior test this session) plus the standard
py_compile + pyflakes pass, clean across the whole repo.
Co-Authored-By: Claude Sonnet 5 <noreply@anthropic.com>
From the original night's low-priority backlog item ("user on hold
status") — the only lever admin had for cutting a customer's access was
Revoke, which is permanent: the subscription's remaining days are just
gone, and restoring access means manually granting a brand-new one and
eyeballing how many days to give back. No way to say "block this for a
few days, then give the exact remaining time back."
db.py: new held_at column on subscriptions (same ALTER-TABLE migration
pattern as every other column added this week). hold_subscription()
sets it, guarded to only fire on a subscription that's currently active,
not already held, not expired — returns False instead of silently
no-opping so the caller can tell holding didn't happen. resume_subscription()
shifts expires_at forward by exactly how long it was held (now - held_at)
and clears held_at, so a subscription paused for 3 days comes back with
3 days added, not 3 days lost.
The part that actually mattered for correctness: list_active_subscriptions()
now also requires held_at IS NULL. This function is what xray_manager's
periodic sync (every 90s) uses to decide which clients belong in Xray's
config — without this exclusion, holding a subscription would look like
it worked for about 90 seconds and then the next sync would silently
re-add the client, since the row still has active=1 and a future
expires_at. Found this by actually tracing sync_from_db()/sync_all()
before writing the hold logic, not after debugging a live failure.
api.py: POST .../hold and .../resume routes, mirroring the existing
revoke route (fetch the sub, touch the node's xray client immediately
rather than waiting for the next periodic sync, same as revoke already
does). _days_left() now takes the whole subscription row instead of just
expires_at, so it can use held_at as the reference point instead of "now"
for a held subscription — otherwise the admin UI would show the days
counter silently ticking down while the customer isn't even able to use
the service.
admin.html: Пауза/Возобновить buttons next to Отозвать in both the main
Подписки table and the per-user card, a "на паузе" badge, and a doc-block
explaining the hold-vs-revoke distinction. Also fixed a latent race while
touching this code: the old inline revoke handler in the user card fired
openUserCard() immediately alongside revokeSub() without waiting for it,
so the card could refresh before the revoke's own API call had finished;
switched to .then() so hold/resume/revoke all correctly wait for the
action before refreshing the card.
Verification: db.py has no fastapi/aiogram dependency so this was fully
testable locally, unlike most of tonight's api.py/bot.py-touching work.
16 checks against a real isolated sqlite db: hold/resume round-trip,
the exclude-from-active-list behavior the xray sync depends on, the
exact hours-shift math (simulated a 5h hold by rewriting held_at
directly, verified the resumed expires_at landed within 6 minutes of
the expected shift), and edge cases — double-hold, double-resume,
holding an expired or already-revoked subscription, nonexistent uuid.
AST-extracted the updated _days_left() out of api.py (still can't
import the module directly) and ran it against hand-built held/active
subscription dicts. Added the same hold/resume sequence to the existing
CI "TOTP/backup/reorder" step and ran that step's exact full script
locally end to end before committing — all six of its sections pass
together, not just the new one in isolation.
Co-Authored-By: Claude Sonnet 5 <noreply@anthropic.com>
User ask, paraphrased: install it, get help wiring up payments, and
immediately have a ready site under your own name — not "MBS Panel"
plastered everywhere and a bunch of manual follow-up.
Two things were actually broken/missing, found by tracing every surface
a real customer or the operator would see:
1. "MBS Panel" was hardcoded in ~20 places (bot messages, subscription
page, admin panel splash/title/sidebar, legal pages, 2FA issuer,
install.sh) with zero way to change it short of editing source.
New BRAND_NAME config value (config.py default "MBS Panel", so this
is 100% backward compatible for existing installs) wired through
everywhere via the same live-settings pattern from the last commit
(settings.get_brand_name(), no restart needed anywhere it's used).
New Настройки → «Название» section in the admin panel to change it.
2. site/index.html and site/cabinet.html — a fully-built landing page +
personal-cabinet template, already in the repo — were never actually
served by anything. Not mounted by FastAPI, not deployed by
install.sh, not linked from anywhere. Pure dead weight: a repo that
looked like it shipped a client site but didn't. Now legal.py gets a
render_site_page() (same {{TOKEN}} substitution + HTML-escaping as
the existing offer/privacy renderer, new tokens: BRAND_NAME,
SITE_DOMAIN, SUB_DOMAIN, BOT_USERNAME) and GET "/" serves the branded
landing page on any host that isn't PANEL_DOMAIN (in practice:
SUB_DOMAIN, which nginx already routes to this backend — zero
install.sh/nginx/certbot changes needed, so this is live on every
existing install without an upgrade step beyond `mbs update`).
GET /cabinet.html serves the cabinet. Landing page's pricing section
now fetches real, live prices from a new public GET /api/plans
instead of showing static duration labels with no numbers.
Also fixed along the way, same staleness-bug class as the payments/HWID
fix last commit, found by grepping for every remaining frozen `from
config import ...` in api.py: BOT_TOKEN/BOT_USERNAME were still frozen
constants in api.py (mbs-api never restarts itself). Concretely this
meant: changing the bot via Настройки → Telegram-бот would leave
_tg_send_message (payment-received notifications) silently trying the
OLD token, admin_get_bot_settings showing the OLD username right after
a successful save, and gift-code links pointing at the OLD bot — all
until a manual mbs restart, same shape as the Platega-secret bug fixed
last commit. Added settings.bot_credentials(), wired it through every
call site (hoisted out of loops where relevant, same N+1 discipline as
always), removed the now-stale "выполни mbs restart" copy from the bot
settings hint.
legal.py's own BOT_USERNAME import was frozen too (used by the /offer
and /privacy {{BOT_USERNAME}} token) — switched to reading it live
in-module (no settings.py import from legal.py, would've been circular
since settings.py already imports legal.py for the env reader).
install.sh: new interactive prompt for the brand name (default "MBS
Panel", so hitting enter reproduces today's behavior exactly), written
to .env, echoed in the final summary along with the now-live site URL.
Verification: same story as always — api.py/bot.py still can't import
locally (no pydantic-core wheel for Python 3.14 on this machine).
py_compile + pyflakes clean across the whole repo. Real runtime test
against an isolated .env fixture: brand name and bot-credential live
reads (no reimport), render_site_page() token substitution correctness
on the actual site/index.html and site/cabinet.html files including an
XSS check (brand name containing <script> comes out HTML-escaped), and
a regression check that adding the BRAND_NAME token to the existing
legal.render() didn't break offer.html/privacy.html. Extracted
SUB_PAGE_TEMPLATE/SUB_PAGE_EXPIRED_TEMPLATE via ast from api.py (can't
import the module, but can pull the string constants) and ran the real
.format() calls against them to catch any brace-escaping mistake in the
new {brand_name} placeholder — CSS braces in those templates are
already double-escaped for .format(), easy to get wrong. Extracted and
node --check'd admin.html's whole inline script, div-tag-balance check
on the full file. install.sh's new prompt+heredoc snippet run standalone
with piped stdin (both a brand name with spaces and an empty/default
input), round-tripped the resulting .env back through the real
env-parsing logic. Extended the existing CI "app wiring" step (which
does import api/bot for real on Linux) with branding assertions calling
the actual route functions directly (api.root(), api.public_plans(),
api.public_branding()) — ran every part of that step's new logic that
doesn't need api.py locally first, to catch what's catchable before
trusting the rest to CI once the account's abuse-review lifts.
Co-Authored-By: Claude Sonnet 5 <noreply@anthropic.com>
Closes the last "still .env-only" gap from the backlog (tariffs/HWID) and
fixes a real bug found while building it: payment provider credentials and
enabled-flags were frozen in api.py's process at import time, so a Platega
secret rotation via the settings UI would leave api.py verifying inbound
webhooks against the OLD secret until a manual `mbs restart` — while
bot.py (which does get restarted on save) already had the new one. Same
class of staleness affected HWID_LIMIT_ENABLED/HWID_FALLBACK_LIMIT and
plan prices, neither of which had any settings UI at all before this.
New `settings.py` module: get_plans()/get_plans_by_code() (live prices,
falls back to config.py defaults), get_payment_settings(), get_hwid_settings(),
yookassa_credentials()/platega_credentials(), set_plan_prices() — all backed
by a new batched legal.read_env_vars() (one file read for N keys instead of
N reads) and legal.update_env_var() (moved out of api.py's private
_update_env_var, which is now a one-line delegate to avoid duplicating the
same env-file-rewrite logic in two places).
api.py and bot.py no longer import PLANS/PLANS_BY_CODE/PAYMENTS_ENABLED/
HWID_LIMIT_ENABLED/HWID_FALLBACK_LIMIT from config as frozen constants —
every read goes through settings.py instead. payments.py no longer imports
YOOKASSA_*/PLATEGA_* from config either; every provider call (create/check
payment, verify webhook signature) reads live credentials at call time.
Every call site inside a loop hoists the live lookup before the loop first
(same N+1 discipline as the rest of tonight), so this doesn't regress
get_subscription's hot path — one settings.get_hwid_settings() call per
request, same as before.
New routes: GET/POST /admin/api/payments/plan-settings (per-plan prices +
a payments_enabled master toggle — there was previously no way to turn
payment collection back off without deleting provider credentials),
GET/POST /admin/api/hwid-settings. Both validate input strictly (prices:
non-negative int; HWID limit: 1-1000) and reject the whole request instead
of partially applying on bad input.
admin.html: new "Тарифы" section in Платежи (price inputs rendered from
the live plan list + payments toggle) and "Лимит устройств (HWID)" in
Настройки, both using the existing .check checkbox / plain-input styling
(no native <select>, per the earlier white-popup complaint). Removed the
now-incorrect "выполни mbs restart" copy from the YooKassa/Platega settings
hints and save-result messages, and added a doc-block for HWID (never had
one) plus extended the Платежи doc-block to mention live-apply. Also
dropped a dead `import links` in bot.py caught by pyflakes while verifying
this.
Verification: api.py/bot.py still can't be imported on this Windows
machine (no prebuilt pydantic-core wheel for Python 3.14, confirmed again
by a fresh pip attempt — same wall as every prior session), so relied on
what's actually exercisable: py_compile + pyflakes (zero undefined names)
across every module including api.py/bot.py, a real runtime test against
an isolated .env fixture covering live price/toggle/HWID reads with zero
reimport, write-idempotency (no duplicate .env lines on repeated saves),
and the concrete bug this fixes end to end — computed an HMAC signature
against an old Platega secret, rotated the secret via update_env_var (the
same call the settings route makes), confirmed the old signature is now
rejected and a new one computed against the rotated secret verifies, all
in the same process with no reimport. Also ran the new CI step's exact
heredoc locally byte-for-byte before adding it to ci.yml. GitHub Actions
still won't trigger for this account (still under abuse-review, ticket
open >2 days) so this is the same substitute-for-CI rigor used all night.
Co-Authored-By: Claude Sonnet 5 <noreply@anthropic.com>
The Документация tab (the panel's built-in ops reference, no GitHub
trip needed) still only covered the pre-tonight feature set —
multi-admin, backup/restore, the payments wizard, webhooks, node
reordering, and the xray config pre-flight check were all live but
undocumented there. Added two new doc-blocks (Бэкапы, Платежи и
вебхуки) and extended the existing Пароль/Ноды blocks rather than
duplicating structure.
Co-Authored-By: Claude Sonnet 5 <noreply@anthropic.com>
Per the docs.rw comparison researched earlier tonight, Remnawave
fires webhooks for users+nodes and Marzban for users — this panel
had neither, only received inbound webhooks from payment providers.
New webhooks.py, fired on payment.paid (both webhook-driven and
reconciler-driven grant paths, so it fires regardless of which one
actually processes a given payment) and
subscription.granted_by_admin (kept as a distinct event name rather
than reusing payment.paid, since no money necessarily changed hands
there). Settings tab gets a URL field; a secret is generated once on
first save via secrets.token_hex and never regenerated on later URL
edits, so a receiver's signature verification doesn't silently break
when the admin just updates the endpoint. Every delivery is
HMAC-SHA256 signed over the raw JSON body via X-Signature, same
verification shape Platega already uses for its inbound webhooks.
Delivery is fire-and-forget (10s timeout, swallows all exceptions) —
a receiver being down must never block or fail a payment grant.
Reads WEBHOOK_URL/WEBHOOK_SECRET fresh from .env via legal.py's
existing reader instead of adding a third copy of that logic.
Verified with a real local HTTP server: actual delivery, payload
shape, and that the received X-Signature verifies against the
configured secret using the receiver's own side of the HMAC — not
just asserting the sender computed *something*. Also verified the
no-URL-configured no-op path and that changing the URL later does not
rotate the secret.
Co-Authored-By: Claude Sonnet 5 <noreply@anthropic.com>
The setup wizard added last iteration only covered YooKassa, leaving
Platega (the second provider this panel has always supported) still
.env-only despite everything else in the payments flow treating both
providers symmetrically. Same GET/POST shape as the YooKassa settings
routes. Platega has no documented lightweight credentials-check
endpoint like YooKassa's /v3/me, so this one honestly says so in the
UI instead of saving through a fabricated validation call — it saves
directly and a bad key will surface on the first real payment attempt
instead of at save time.
Co-Authored-By: Claude Sonnet 5 <noreply@anthropic.com>
The site/offer.html and site/privacy.html legal templates existed in
the repo but were never actually wired to anything — no route served
them, install.sh never copied them anywhere. Nobody deploying this
for real payments had a live offer/privacy page, which YooKassa
requires for merchant approval.
Rewrote both templates with {{TOKEN}} placeholders (new legal.py
renders them from .env-backed settings, read fresh on every request,
no restart needed to fix a typo) and added a proper setup section in
the Payments tab: business type/name/INN/support contact/refund
window, saved via POST /admin/api/payments/legal-settings, live at
GET /offer and /privacy immediately. Unset fields render as a visible
not-set-yet badge instead of breaking the page. Effective date
auto-stamps once on first save and stays stable across later edits
(verified: editing the name afterward doesn't reset it).
YooKassa shop_id + secret_key get their own section: validated live
against YooKassa's own /v3/me before being saved (same pattern as the
existing bot-token getMe check), never echoed back to the frontend
once set. Includes an inline guide — where to find the keys in
YooKassa's dashboard, and that self-employed registration there needs
just passport + INN, no separate cash register.
Both new dropdowns use the existing custom .dd component, not a raw
select element — this codebase deliberately doesn't use native
selects (see the comment already in admin.html) because of the
OS-rendered white popup, so a new form had to follow that pattern,
not reintroduce it.
Verified: template rendering with empty settings (fallback badges,
no leftover tokens) and fully filled settings, HTML-escaping of field
values (a script tag in a field renders as text, not markup), the
one-time-only date stamp, and all new routes registering correctly.
Also fixed a stale doc string in the panel's own admin-facing docs
tab that still quoted the old rate-limit numbers from before the
real limits shipped.
Co-Authored-By: Claude Sonnet 5 <noreply@anthropic.com>
Backup&Restore, multi-admin+2FA, xray config validation, drag-n-drop
nodes, and login rate-limiting were all shipped but never made it
into the README's feature list. Also switched the README's install
command to the mbs.savsis.xyz one-liner to match the landing page
(cfac7c1) — kept the raw GitHub clone as a documented alternative,
same script either way. Version tag in admin.html bumped to v1.1.0.
Co-Authored-By: Claude Sonnet 5 <noreply@anthropic.com>
Matches Remnawave's security-first positioning (passkeys/OAuth there)
with the more universal standard instead — RFC 6238 TOTP, works with
Google Authenticator/Authy/1Password/anything. Implemented from
scratch on stdlib only (hashlib/hmac/struct/base64) — no new
dependency — and verified against the official RFC 4226 HOTP test
vectors (all 10 pass exactly) before wiring it into any auth path.
Per-admin, optional: setup shows the secret + otpauth:// URI (no QR
render, just copyable text — didn't want to fake a QR library),
confirmed by entering a real code before it's persisted. Login is now
two-step when 2FA is on: password first (returns a short-lived
pending_token instead of a session if totp_secret is set), then a
second call with the pending_token + code creates the real session.
pending_totp rows are single-use and expire in 5 minutes, cleaned up
alongside the existing admin_sessions cleanup in periodic_sync.
Disabling 2FA requires re-entering the current password.
Verified end-to-end: enable/disable round trip, pending-token
resolve+single-use+expiry, code verification against the stored
secret, wrong-code and wrong-password rejection — on top of the raw
HOTP correctness check.
Co-Authored-By: Claude Sonnet 5 <noreply@anthropic.com>
Matches Marzban's multi-admin (WIP there) and closes a real gap vs
both. New 'admins' table (username + PBKDF2-SHA256 password hash,
200k iterations, random salt per account, stdlib hashlib/hmac only —
no new dependency), admin_sessions now tracks which admin is logged
in. Existing installs aren't broken: on first run, if no admins exist
yet, a default 'admin' account is seeded from the current
ADMIN_PANEL_PASSWORD — old password keeps working under username
'admin', pre-filled on the login screen.
Admin management lives in Settings: list, add (username + password,
min 8 chars), remove. Can't delete the last remaining admin or your
own currently-logged-in account. Sidebar now shows who's logged in.
Verified end-to-end: bootstrap, correct/wrong/nonexistent login,
session->admin resolution, last-admin-delete protection, duplicate
username rejection, add/remove round trip, and that identical
passwords hash to different values (unique salt) but both verify.
Co-Authored-By: Claude Sonnet 5 <noreply@anthropic.com>
Matches Remnawave/Marzban's 'Host Sorting via Web UI' — was the one
concrete UI gap flagged in the comparison research. Nodes get a
persistent sort_order column (backfilled once from the existing
de1-first/created_at order on migration, verified idempotent — re-running
_migrate() on an already-migrated db does not reshuffle it back).
Plain HTML5 drag-and-drop (dragstart/dragover/drop), no library —
matches the project's no-frameworks admin.html. Drop reorders the
in-memory list optimistically, re-renders immediately, then persists
via POST /admin/api/nodes/reorder; a failed save reloads from the
server instead of leaving the UI out of sync with the db.
db.reorder_nodes() rejects any list that doesn't contain exactly the
current set of node codes (no silent drops or duplicates). New nodes
append at the end (MAX(sort_order)+1) instead of jumping to the front.
Verified with a full test: initial order, reorder, two rejected
malformed reorders, a new node appending at the end, and migration
re-run not touching an already-backfilled order.
Co-Authored-By: Claude Sonnet 5 <noreply@anthropic.com>
Neither Remnawave nor Marzban has this natively (community tools only,
per docs.rw's own comparison table) — one-click download of a tar.gz
with a consistent SQLite snapshot (via sqlite3's backup API, safe even
under WAL) plus .env, and upload-to-restore from the same file.
Restore validates the archive is real (gzip + tar structure), that
mbs.db is an actual sqlite database with the expected tables (not
just any file named mbs.db), and rejects oversized uploads — before
touching anything live. Takes a timestamped safety copy of the
current db/.env before overwriting, clears stale -wal/-shm siblings
so the restored file doesn't get replayed against the wrong WAL, and
restarts mbs-bot automatically when .env was part of the restore
(api.py isn't restarted from within its own request handler for the
obvious reason).
Verified with a full round-trip test: backup -> mutate state -> restore
-> confirm the mutation is reverted, plus three negative cases (garbage
data, oversized upload, a fake non-sqlite mbs.db) all correctly
rejected with no side effects.
Needs python-multipart for FastAPI's UploadFile — added to
requirements.txt, picked up by the next 'mbs update'.
Co-Authored-By: Claude Sonnet 5 <noreply@anthropic.com>
payments: _grant_paid_subscription now validates plan/node exist before
marking a payment paid instead of after (was leaving charged-but-ungranted
payments with no error trail); mark_payment_paid is now a single atomic
UPDATE ... WHERE status='pending' instead of check-then-act, closing a
double-grant race between webhooks and the periodic reconciler; yookassa
webhook now re-verifies payment status server-side via the API instead of
trusting the posted body (platega already had HMAC verification).
hwid: 'user["hwid_limit"] or FALLBACK' treated an explicit 0 (admin fully
blocking a user) as unset — now an explicit None check. Device count-check
and insert are now one atomic transaction (db.add_device_if_under_limit)
instead of two raceable statements.
perf: payment webhooks and _grant_paid_subscription's SSH/HTTP calls now
run via asyncio.to_thread instead of blocking the event loop; same for
bot.py's periodic_sync/reconcile_pending_payments and the manual admin
sync button. Admin endpoints (traffic/subscriptions/payments/gift-codes/
user-card) now resolve node labels from one db.list_nodes() call instead
of a fresh db.get_node() per row. revoke/reset-traffic use a direct PK
lookup instead of scanning up to 5000 rows. Dashboard now asks the API
for 8 rows instead of fetching 200 and slicing client-side.
security: mbs.db (and -wal/-shm) now chmod 600 right after creation —
it held session tokens and subscription bearer tokens world-readable
by default. Node SSH connections now pin host keys via a persisted
known_hosts file (TOFU) instead of accepting any key on every connection.
delete_node now refuses to delete a node with active subscriptions
instead of silently orphaning their xray clients.
deadcode: removed unused xray_manager.list_client_ids and admin.html's
superseded staggerReveal (rows animate via rowAttr() inline now).
Also guards gift-code redemption against a plan/node deleted after the
code was created (was an unhandled KeyError/TypeError crash).
Co-Authored-By: Claude Sonnet 5 <noreply@anthropic.com>
