Both ran scram verify_plain (~1s PBKDF2 at 1.2M iters) inline under the engine
lock, freezing the whole daemon per login. Add NetAction::DeferAuthenticate +
AuthThen continuation + ctx.defer_authenticate, mirroring DeferPassword: the
module/SASL path fetches the verifier cheaply and defers; the link layer runs
verify_plain on spawn_blocking, then Engine::complete_authenticate finishes the
login (IDENTIFY reuses the same ctx helpers, so its login/AJOIN/vhost/memo
side-effects are unchanged). Tests resolve the defer inline (cfg(test)). The
gRPC web-login path was already fixed; no login path stalls services now.
(GHOST/DROP/CERT/GROUP still verify inline but are rare account ops, not logins.)
The earlier StatsSet fix persisted only the shared namespaced counters. The
per-channel BOTSTATS view (lines seen + top talkers, from /statserv #chan) lived
on the LiveChannel struct, which is rebuilt from each burst — so it reset to zero
on restart. Move it to a name-keyed Network.chan_activity map (independent of live
membership, seedable before the burst) and carry it in the StatsSet snapshot
(new serde-default `channels` field) alongside the counters. Reworded the
notice from 'seen this session' now that it persists.
Problem 1: an assigned bot joined via IJOIN with no status mode, so it had no
+o. Join it with the 4-param IJOIN form (ts 1, mode o) so it's opped, matching
Anope's default bot modes (+o), and exempt bots from SECUREOPS/RESTRICTED so the
op is never stripped.
Problem 2: the shared stat counters lived only in memory, so a restart wiped
them. Snapshot them to the log (StatsSet event, folded into NetData + compaction)
periodically and on shutdown, and reseed the live registry from the log at
startup. Live gauges (online counts) stay re-derived, not stored.
A fold-parity property test caught it: jupes were plain in-memory Db fields
that logged no event, so after any services restart every juped server was
silently un-juped and could relink. Moved jupes into NetData (like akills/
forbids/groups) and event-sourced them with JupeAdded/JupeRemoved (Local
scope), so a jupe now replays from the log. Adds the round-trip property test
(live state == replayed state) as a permanent guard against this bug class.
SQUIT arrives once for a splitting server (by SID), but a hub takes its
whole subtree with it — users behind downstream servers were left as stale
state. The parser now distinguishes a sourced downstream SERVER (tracked
with its parent) from the unsourced uplink handshake, the engine keeps the
SID->parent tree, and a SQUIT forgets every user across the departed
server's subtree. Tree is rebuilt each link. Wire formats verified against
the InspIRCd 4 m_spanningtree source.
A KILL of a services pseudo-client only reintroduced BotServ bots; a killed
core agent (NickServ, ChanServ, …) fell through to the forget-user path and
stayed off the network until a services restart. UserKilled now recognises an
agent by its fixed uid and re-emits its introduction, so an oper can't KILL
services down.
A netsplit is signalled once via SQUIT, not as a QUIT per user, so every
user behind the departed server lingered in services' state — stale
sessions, session-limit slots, and channel memberships. The parser now
surfaces ServerSplit and the engine forgets every uid carrying that
server's SID prefix, leaving other servers' users untouched.
KILL was never parsed, so a killed user lingered in services' state (stale
session, session-limit slot, channel membership) and a killed services bot
vanished for good. The parser now surfaces UserKilled; a killed real user is
forgotten like a QUIT (shared forget_user helper), while a killed bot is
dropped and reconciled so it is reintroduced and rejoins its channels.
A PART/KICK of a bot uid updated channel membership but not bot_channels,
so reconcile still believed the bot was present and never rejoined it — an
op could evict a services bot for good. The Part handler now drops the stale
membership and reconciles, so a bot assigned to the channel comes right back.
The inbound parser ignored METADATA entirely, so on a services netburst
(after a restart/relink) the ircd's replayed 'accountname' metadata was
dropped — every logged-in user looked logged out to services until they
re-identified, losing ChanServ access. Parse METADATA <uid> accountname
into a new AccountLogin event that restores account_of (empty = logout);
our own echoes and other keys are filtered.
A local AKILL/SQLINE/etc. pushes an ADDLINE to the ircd at once, but a ban
arriving by gossip was only held for the next netburst — so a user banned on
one network could keep hopping to peers until their next relink. ingest now
signals BanAdded/BanLifted (all xline kinds), and gossip_ingest pushes the
ADDLINE/DELLINE. Renamed AccountChange to IngestEffect to match its broader
role.
A local SUSPEND logs out the account's active sessions, but a suspension
arriving by gossip only set the flag — a session already identified on a
remote node stayed logged in despite the network-wide suspension. ingest
now signals AccountChange::Suspended, and gossip_ingest ends those sessions
(the account and its channels are kept). Factors the per-session logout into
a shared logout_uid helper.
When a peer's gossiped entry drops or takes over an account, the local node
releases the channels that account founded locally (Local-scope channels can
be founded by a Global, gossiped account). That cleanup runs outside the
dispatch path, so a bot assigned to a released channel lingered until the
next netburst. gossip_ingest now reconciles bots after an account is removed,
matching the expiry sweep.
Command-driven channel drops reconcile bots via the dispatch path, but the
inactivity-expiry sweep runs outside it — so a bot assigned to an expired
(or expired-founder-orphaned) channel lingered there until the next
netburst. The sweep now reconciles at the end, parting any stranded bot.
A user authenticated via SASL during registration never runs the NickServ
IDENTIFY path, so they silently missed their auto-join channels, vhost, and
waiting-memo notice. The UserConnect handler now applies these login
side-effects when the arriving user is already authenticated, mirroring
identify.rs. Non-SASL connections are unaffected (no account set yet).
FORBID accepted and stored EMAIL patterns, but nothing ever checked them:
a forbidden address still registered and could be set via NickServ SET
EMAIL. Both paths now reject a matching address (new ForbiddenEmail
outcome), matching how NICK and CHAN forbids already gate registration.
Founders can attach a contact email to a channel, shown in INFO and
cleared when set empty. Mirrors SET URL exactly; Local-scope channel
metadata. Completes the channel DESC/URL/EMAIL trio.
Founders can attach a homepage URL to a channel; it shows in INFO and
clears when set empty, mirroring SET DESC. Local-scope (channel metadata),
not gossiped, consistent with the entrymsg/description events.
Lets an account keep its last-seen/online line in INFO visible only to
itself and to opers. Default stays public (Anope's default). Accepts
STATUS (with USERMASK as an alias) per Anope's SET HIDE <field> grammar.
Nick protection was unconditional. SET KILL OFF lets an account opt out —
an unidentified user keeping one of its nicks is no longer prompted or
renamed to a guest. Stored inverted (no_protect) so protection stays the
default. Accepts Anope's ON/QUICK/IMMED/OFF; grace is a fixed interval
here, so the finer variants simply enable protection like ON.
A user with SET AUTOOP OFF is never auto-opped on join, even where they
hold channel access — they op themselves via ChanServ UP. Stored inverted
(no_autoop) so the default stays auto-op. Join gates on both the channel
setting and the account preference; either can opt out.
SET AUTOOP {ON|OFF} controls whether access members are auto-opped/voiced
on join (on by default). Stored inverted as a settings flag so the
default stays on, and gated in the join handler.
SET NOTIFY {ON|OFF} controls whether you're told about new memos on login;
SET LIMIT <n>|NONE sets your mailbox cap (honoured on SEND, NONE = the
network default). New memo_notify/memo_limit account fields + one
MemoPrefsSet event.
IGNORE ADD/DEL/LIST manages a per-account memo-ignore list; a memo from
an ignored account is silently dropped, and the sender is still told it
was sent so the ignore isn't revealed. New memo_ignore account field +
events, and a check in SEND.
GETEMAIL (auspex) lists accounts whose email matches a glob. SET
RESTRICTED makes a channel kick anyone without access (or oper) as they
join, mirroring the SECUREOPS on-join enforcement.
FORBID ADD/DEL/LIST bans a NICK, CHAN, or EMAIL glob from being
registered — a network-wide policy that gossips like an AKILL and every
node enforces. Nick registration is blocked in pre_register_check (both
NickServ REGISTER and the account-registration relay), and channel
registration in ChanServ REGISTER. New Forbid store + events, mirroring
the akill subsystem.
SET <#channel> SUCCESSOR <account>|OFF names an account that inherits the
channel if the founder's account is dropped or expires, instead of the
channel being released. Adds a ChannelInfo.successor field + event, and
centralises the founder-gone channel release into a single
release_founded_channels store method used by DROP, expiry, and the
gossip account-gone path so all three transfer consistently.
On connect or nick-change to a registered nick the user isn't identified
to, NickServ prompts them to IDENTIFY and schedules a rename; a short
enforcement sweep renames them to a guest nick after a grace period if
they still haven't. Gated on the uplink's initial burst so a relink does
not enforce every already-online user, and cancelled the moment they
identify (via IDENTIFY or SASL) or leave the nick.
Echo is the authority the ircd account module forwards to: parse
ENCAP <us> SWACCTREG <reqid> <origin> <kind> <account> <p2> :<p3> into an
account request, and answer the origin server with
ENCAP <origin> SWACCTRES <reqid> <kind> <account> <status> <code> :<message>.
Thread the origin through the reply so responses route back to the
requesting server. Replaces the placeholder ACCTREGISTER/ACCTREGRESULT
names that nothing on the wire actually spoke.
Credentials: the SCRAM-SHA-256 verifier is now the only password
credential. PLAIN and IDENTIFY verify the plaintext against it via
scram::verify_plain, exactly as a SCRAM login proves knowledge of it, so
an account provisioned from a verifier alone (external authority) works
over every mechanism, not only SCRAM. Removed the redundant argon2 hash
entirely: the field, the Credentials member, the AccountPasswordSet
field, hash_password/verify_password, and the argon2 crate. SASL SCRAM
already forces a PBKDF2 verifier into the store, so the hash never raised
the at-rest floor. OS RNG now comes from rand_core.
Registration: finished draft/account-registration over the ircd relay.
VERIFY confirms the emailed code, RESEND reissues it, STATUS reports
state, and REGISTER now emails the code and replies verification_required
on the relay path too, not only through NickServ.
Each service now exposes its help catalog via Service::help_topics; the
engine collects them into a network-wide index on Network, reachable
through two new NetView methods. HelpServ uses it: HELP <service>
[command] (or a bare service name) serves any service's per-command
help, bare HELP lists the services, and its own ticket commands still
work. No cross-crate dependency: the catalog flows through the SDK's
NetView, keeping every module dependent only on echo-api.
Rebrand the daemon and workspace: the binary is now `echo`, every crate
is `echo-*` (`echo-api` and the module crates), Rust imports use
`echo_*`, the gRPC proto is `proto/echo.proto` with package `echo.v1`,
and the log filter, gossip peer-name default, and on-disk store default
(`echo.db.jsonl`) follow. Docs updated throughout. No behaviour change;
crate directories and the config schema are untouched.
engine/mod.rs had grown past 5k lines, most of it one test module and
a few large method clusters. Lift the tests into engine/tests.rs, and
move the cohesive method groups into sibling files — sasl, kicker
(chat policing), register (account-authority relay), and dispatch
(service routing and fantasy). The Engine type, its master event
handler, shared helpers, and the small state structs stay in mod.rs;
each sibling adds its own impl Engine block. Pure reorganisation: no
behaviour change, tests unchanged and green.
Extends the fantasy router so a dice command (!roll/!calc/!exroll/!excalc)
in a bot channel routes to DiceServ (if loaded) instead of ChanServ, and
its reply is turned into the assigned bot speaking the result to the
channel — so a tabletop game rolls in the open, not in a private notice.
Everything else still routes to ChanServ. Factored the dispatch into
service_uid()/route_to() helpers shared by both branches.
Users open tickets with REQUEST (or HELPME); the queue is event-sourced
and, because a request is an event, the audit feed announces it to staff
and records it in the searchable log — so OperServ LOGSEARCH finds a ticket
by any word, the same trail as reports. A user can CANCEL their own open
ticket. Operators work the queue: LIST [ALL], VIEW, TAKE (claim), NEXT
(claim the oldest unassigned), CLOSE. Filing is rate-limited. Opt-in.
That completes the six third-party modules — all interconnected: they
share the account/channel/oper identity, the event-sourced store, and the
one audit/incident trail.
The engine samples live op state on a short cadence, scoring op-time per
identity (account, else host) on the network view — a net +1 per pass
while opped, -1 while idle, so trusted regulars rise and stale ops fade.
ChanFix (operator-only) reads those scores: SCORES <#chan> shows the
standings, CHANFIX <#chan> reops the present regulars scoring above a
share of the top score, once the channel is opless enough and has real
history. It defers to ChanServ — a registered channel is refused. Opt-in.
C++ note: the original persists scores to the services DB and runs its own
timers; here scoring rides the existing periodic loop and lives on the
ephemeral network view, so it's node-local and rebuilds itself.
User groups (!name) — a founder plus member accounts, each with group-
access flags (the shared flag primitive: F founder, f manage, i invite,
c chan-access, s set, m memo). REGISTER/DROP/INFO/LIST/ADD/DEL/FLAGS;
event-sourced and federated like accounts.
The flagship interconnection: a channel access entry may name a !group, so
every member inherits that channel access. Resolution is group-aware — a
new db.channel_caps() unions a user's direct access with any group they
belong to, and the engine's join handler auto-ops/voices through it. So
"ChanServ FLAGS #chan !team +o" auto-ops every member of !team, and
removing someone from the group revokes it. Opt-in.
A granular access system, replacing the op/voice-only levels with per-user
flag letters (f full, o auto-op, O op, h auto-halfop, v auto-voice, t
topic, i invite, a modify-access, s settings, g greet). One shared
primitive in the api — level_caps() resolves a stored level to capabilities
(recognising the legacy op/voice/founder presets so nothing breaks) and
apply_flags() applies +/- deltas — so join_mode/is_op/access_rank all
route through it and a flag-granted +o auto-ops exactly like a legacy op
entry.
ChanServ FLAGS <#chan> [account [+/-flags]] lists or edits access;
changing needs the founder or the 'a' flag. This is the foundation
GroupServ will reuse for group membership flags.
Any user can REPORT <nick|#channel> <reason>; filing is rate-limited per
reporter. Reports are an event-sourced, federated queue operators review
with LIST / VIEW / CLOSE / DEL. The interconnection: because a filed report
is an event, the engine's audit feed already announces it to the staff
channel and records it in the searchable action log — so OperServ
LOGSEARCH finds a report by any word in it, tying reports into the same
trail as bans, kicks, and everything else. Opt-in (add "reportserv").
InfoServ becomes the dedicated home for network information bulletins,
backed by the same event-sourced news store OperServ NEWS used — one place
to post, not two. POST/LIST/DEL manage public bulletins (shown to every
user on connect); OPOST/OLIST/ODEL the oper ones (shown to operators on
login). Anyone may LIST the public bulletins; posting/removing is admin-
only. The engine's existing connect/login display hooks render them
unchanged. The OperServ NEWS command is retired accordingly (the news
events and the display path stay). Default service.
Adds Accounts.Provision(name, scram256, scram512, email) so an external
authority that stores verifiers rather than passwords (a website) can bulk-
backfill accounts it never had the plaintext for. It creates the account
with the given SCRAM verifiers — SCRAM and keycard login work immediately;
typed-password login stays disabled (empty password_hash) until a real
password lands via Register/SetPassword. Refuses an existing account, so a
re-run can't clobber a fully-credentialed one. scram512 may be empty
(SCRAM-SHA-512 then falls back to 256).
By default fedserv owns accounts itself — nothing changes, no external
service required. Set [auth] external = true and an outside authority (the
website) owns identity instead: NickServ REGISTER / DROP / SET
PASSWORD|EMAIL / RESETPASS / CONFIRM / CERT / GROUP and the IRCv3
registration relay are all refused ("managed on the website"), so IRC
can't mint or change a second identity. Login is unchanged — fedserv still
authenticates locally against the accounts the authority pushes in via the
existing gRPC Accounts API, so lookups stay in-memory fast.
fedserv keeps owning all IRC-domain data (channels, access, vhosts, bans,
memos) keyed by the account name in both modes. One config bool, off by
default, so standalone deployments are unaffected.
DEFCON [1-5] reads or sets a network-wide defence posture, each level
adding a restriction the engine and services enforce:
- 4: channel registrations frozen (ChanServ REGISTER)
- 3: all registrations frozen (the pre_register_check choke-point, covering
NickServ and the IRCv3 relay, plus channels)
- 2: sessions additionally capped to one per host
- 1: full lockdown — new connections are turned away on connect
Setting a level announces it to the whole network. Node-local level on the
db so services can read the derived freezes; admin-only.
OPER ADD <account> <priv[,priv]> [+duration] now takes an optional window,
after which the grant lazily stops conferring privileges — no timer, the
oper_privs merge just hides an expired grant (evaluated on the engine's
clock so it's testable). Permanent grants are unchanged; LIST flags the
temporary ones and hides expired entries. Runtime opers now carry an
OperGrant{privs, expires} instead of a bare priv list.
