Changelog
View SourceAll notable changes to this project are documented here.
The format is based on Keep a Changelog, and this project adheres to Semantic Versioning.
Unreleased
0.3.0 - 2026-06-22
The headline is the receive side: a %Amarula.Msg{}'s content is now always a
clean, proto-free Amarula.Content.* struct — you never pattern-match a raw
protobuf again. Plus validated send_* options, a normalized media struct (with
mimetype), retry-cache flexibility, and an explicit crypto boundary. This is a
breaking release; the migration is mechanical (struct fields instead of maps/protos).
Breaking
%Amarula.Msg{}.contentis now anAmarula.Content.*struct for every type (except:text→String.t()and:other→nil). Previouslycontenthanded back raw protobufs and bare maps. Now::reaction→%Content.Reaction{},:location→%Content.Location{},:poll→%Content.Poll{}, etc. — see the table inAmarula.Msg. Anykey/poll_keyis a{jid, msg_id}ref (the form the send API takes, so a received reaction feeds straight back intosend_reaction/3). The full proto is still onmsg.raw. Migration: replace map/proto field access oncontentwith the documented struct fields.- Media is
%Amarula.Content.Media{}(moved fromAmarula.Media), and:mediacontent is that struct directly — no more%{kind:, media:}wrapper. It surfaces:mimetype(use it for the file extension, not:kind). Inbound jids on the new structs (GroupInvite.group,Product.business_owner,Order.seller) are%Amarula.Address{}, not strings. - Control frames moved off
:messages_upsertto a new:protocol_updateevent. BareprotocolMessages (ephemeral/setting changes and other unhandled types) no longer arrive as junk messages; subscribe to:protocol_updateif you want them. fetch_history/4takes amessage_ref, not a raw%Proto.MessageKey{}. Front-facing functions should never make you construct an internal protobuf. It now takes the%Amarula.Msg{}you received or a{jid, msg_id}tuple (consistent withsend_reaction/send_edit/pin_message/…). If you were passing a%Proto.MessageKey{}, switch to{jid, msg_id}or the received%Amarula.Msg{}. (The publicmessage_keytype, which aliasedProto.MessageKey, is removed — no protobuf type remains in any public spec.)
Changed
send_*options are now validated (NimbleOptions). The option keywords on the facade send functions (send_text,send_media,send_location,send_poll,send_poll_vote,send_event,send_group_invite,request_pairing_code) are now validated against a schema, and each function's## Optionsdocs are generated from that same schema (so they can't drift). Behaviour change: an unknown or mistyped option key now raisesNimbleOptions.ValidationErrorinstead of being silently ignored. This catches typos at the call site; if you were passing an undocumented key that happened to be tolerated, remove it. (nimble_optionsis a new direct dependency, but it was already in the tree viaReq, so it adds no weight.)- Retry-cache built-ins are bounded the way the protocol is. The ETS and DETS
adapters now evict on a 5-minute TTL (a retry receipt that hasn't arrived by
then won't) in addition to a hard
:max_entriescap, and that cap's default is raised 200 → 512 (matching the Baileys reference). Size:max_entriesto your peak sends-per-5-minutes.
Added
- Re-attachable consumer event sink —
Amarula.set_parent/2. The event sink (where{:amarula, …}events go) is no longer frozen atconnect/2. If the process that calledconnect/2restarts while the connection survives in the registry, re-point the sink on the live connection instead of forcing a stop+reconnect:Amarula.set_parent(Amarula.via(:primary), self()). There is still exactly one sink — no subscriber registry, no relay hop. - The sink may be a name, not just a pid (
connect(:parent)/set_parent/2accept aAmarula.Connection.sink/0: pid, registered name,{:via, …}, or{name, node}). A name re-resolves per event, so it re-attaches to the consumer's current pid automatically — surviving both the consumer's restart and the connection's own restart, the same way a:profilehandle survives where a raw pid goes stale. A raw pid is not restart-safe; recover it withset_parent/2.:parentis the preferred connect option;:parent_pidremains a legacy alias. [:amarula, :sink, :down]telemetry. The connection monitors its sink, so a dead consumer is observable instead of events vanishing silently. A name sink's monitor self-heals off the keep-alive once a holder reappears.Amarula.RetryCache.ReadOnly— back the retry cache with your own message store instead of a second copy. If your app already persists sent messages, supplyretry_cache: {Amarula.RetryCache.ReadOnly, get: fn profile, msg_id -> …}; Amarula then only reads on a retry and never writes, evicts, or deletes anything in your store (theputcallback is optional and this adapter doesn't implement it — there is no write path). Removes the:max_entries/TTL sizing question entirely.:max_entriesis documented in theAmarula.Configoption table and theAmarula.RetryCachedocs (it already existed; it was just undiscoverable).:quotedaccepts a{msg_id, participant}tuple (in addition to an%Amarula.Msg{}) — a lightweight quote that threads by id when you have the id but not the original message. It renders when the recipient still has the quoted message; otherwise the reply may show without a quote preview.
Fixed
Amarula.Connection.start_link/2andchild_spec/1are marked internal. They start only the bare Connection process, not its supervision tree, so a consumer who put{Amarula.Connection, …}under their own supervisor got a non-functional connection. The supported entry point isAmarula.connect/2. (The connection is a protocol state machine — it self-restarts for the post-pairing 515 handshake — so the library owns its lifecycle; consumers control naming/distribution via the:registryconfig.)- Replies no longer grow with the quote chain. A reply inlined the full quoted
message verbatim, including its nested
contextInfo— so quoting a reply re-embedded the whole quote ancestry each hop (the proto grew quadratically in chain depth). The quoted message is now stripped of its nestedcontextInfobefore inlining, keeping a quote to one level (matching the Baileys reference). - Stopped copying the one-time-prekey map into every send. The credentials
handed to each per-recipient sender carried
:pre_keys(up to ~812 entries, ~100 KB+ on a fresh account) — which the send/encrypt path never reads (it's only used on the receive/decrypt path, inConnection). It was being deep-copied into every sender, multiplied per recipient in a group fan-out. It's now dropped before the copy. download_media/1honours its{:ok | :error}contract. A malformed/empty media descriptor now returns{:error, :invalid_media}instead of letting the HTTP layer raise (which forced consumers to wrap the call in rescue). It also works with no live connection — documented now — so you can download from aTask.- The connection auto-reconnects when the websocket dies. When the underlying
websocket goes down unexpectedly,
Connectionnow unlinks and monitors it, then reconnects — instead of the dead socket lingering. Paired with the down-transition emit below, a consumer sees the connection go down and come back. - A send while disconnected no longer crashes the connection. Relaying a frame
with no live socket now returns an error to the caller instead of taking the
Connectionprocess down with it. - Errors emit a
:connection_updatedown-transition. On a connection error the consumer now gets aconnection: :close(down):connection_update, so the connection state a consumer observes always reflects reality.
Documentation
docs/CRYPTO_BOUNDARY.md— the Noise/Signal crypto is a pure, self-contained layer; this draws the explicit line between the Core (no app/storage/WhatsApp coupling) and the thin Glue that bridges it toAmarula.Storage. Verified: no Core module depends on the app. The Core is extraction-ready in principle (a standalone Signal/Noise library), though extraction is not planned.- Corrected the
ConversationSenderrationale across the docs: it is a per-recipient serialization point, not a holder of crypto state — the Signal session lives in Storage. (A test,session_race_test, documents a known concurrency gap between the send and inline-decrypt paths on that shared session; a fix is planned —docs/plans/SESSION_WORKER.plan.md.) - README: fixed a stale module reference and the QR-render example (now
qr_code, the actual dependency). docs/LID_PN.md— a LID vs PN identity guide: the two identities a person has, raw JID vs the parsedAmarula.Address, the LID>PN rule (sessions and bundles key on LID; the wire<to>stays PN), where mappings are learned, and how a consumer should key its own contacts.
0.2.4 - 2026-06-21
An internals and robustness pass over the per-connection process tree, prompted
by an external review. No public API or behaviour changes — the Amarula facade
is untouched; everything here is below it. A new optional :sender_idle_ms knob
is the only consumer-visible addition.
Changed
- Per-connection process naming moved to an app-level
Amarula.InstanceRegistry. The connection tree no longer owns a per-instanceRegistry, and the tree supervisor, sibling roles, and eachConversationSenderare now named by the connection'sinstance_idref in one shared registry. Previously these were named by atoms derived from:erlang.phash2(ref)— which both leaked an atom per connection and could collide (two connections hashing to the same atom would fail to start). No atom is minted per connection now, and collisions are impossible. - Per-connection supervisor strategy is now
:rest_for_one(was:one_for_one). The children share fate — senders block on Connection's replies — so a Connection restart now restarts the senders waiting on it instead of leaving a half-dead tree. - The retry cache's ETS table is owned by the
Connectionprocess directly. The dedicatedTableOwnerprocess is gone;Connectioncreates the table ininitvia the newRetryCache.ensure_local/2(adapter-aware — a no-op for non-ETS adapters). Because the table now dies with Connection, a crash/restart recreates it empty, so a poisoned cached entry can no longer outlive — and loop — the restart it triggers. ConversationSenderidle linger is now 1s and configurable viaconfig[:sender_idle_ms](was a hardcoded 5 minutes). A fan-out to many one-shot recipients no longer leaves a long-lived process tail; a disk-backed session store can raise it to keep senders warm under bursty traffic.
Fixed
- Removed the dead
\\ __MODULE__default argument fromConnection.start_link/2and ~28Connectionclient functions. These delegate to a process started under the registry (never under the module name), so the default was never reachable and would have errored if hit. The publicAmarulafacade always passes the pid, so consumers are unaffected. - Dropped the inert
{:send_relayed, …}message. On a successful relay the sender now reports nothing back toConnection: the consumer's reply is already driven by the server<ack>armed at dispatch, so the message was a signalConnectiononly ever ignored. Failures still report{:send_failed, …}.
0.2.3 - 2026-06-20
A big batch of new message and group capabilities — replies, mentions, poll votes, pins, view-once, albums, events, group invites, member tags, and LID↔PN resolution — plus an optional Android client mode and three protocol fixes. No breaking changes.
Fixed
- Audio: thread
:waveformand warn when:secondsis missing (Baileys #2646).send_media(:audio, …)now passes a:waveformopt through to the proto, and logs a warning when:secondsis absent — without a duration, clips longer than ~10s may fail to play on iPhone recipients. Amarula does no media processing, so the caller must supply:seconds; this is now documented onsend_media/5and surfaced at send time. - Pinned chat state is always a definite boolean (Baileys #2328). A pin
app-state action whose
pinnedflag the server omits (proto3-optional) left%Amarula.Chat{}.pinnedasnil("undefined for some conversations"). It now coerces tofalse— only an explicitpinned: trueis pinned — so consumers never see an ambiguous nil from a pin action. mark_online_on_connect: falseis now honored (Baileys #2553). The per-connection setting was defined and documented but never read — the login path always sent presence-available, so the account appeared online and the primary phone stopped getting push notifications regardless of the flag. Connect (and the post-pairing push-name refresh) now gate presence-available on it.
Added
- LID↔PN mapping lookups + a
:lid_mapping_updateevent (Baileys #2263).Amarula.Contacts.pn_for_lid/2andlid_for_pn/2read the local mapping store (no server query) — resolve a group member's LID to a PN after a:messages_upsert. And a new:lid_mapping_updateconsumer event fires with[%{lid: Address, pn: Address}]whenever the send pipeline learns new mappings, so consumers can persist them as they arrive instead of polling. - Group member tags (Baileys #2502).
Amarula.update_member_tag/3sets (or clears, with"") your per-group self-label — capped at 30 chars, rejected with{:error, :member_tag_too_long}rather than silently truncated. Incoming tag changes classify as{:member_tag, %{label, timestamp}}on%Amarula.Msg{}, including removals (empty label) — the case Baileys #2502 dropped. - Android browser mode (Baileys #2201). Setting a
:browserwhose client element contains"Android"(e.g.["MyApp", "Android", ""]) registers as an Android client instead of WhatsApp Web:userAgent.platform = :ANDROID, nowebInfo,DeviceProps.platformType = :ANDROID_PHONE. Lets a session receive view-once media. Experimental, and shows as a phone in Linked Devices — see the impact note inAmarula.Config. Non-Android browsers are unaffected.
0.2.2 - 2026-06-20
A bug-fix plus new outgoing/incoming message types. One small breaking change to how you reference an existing message (see Changed (breaking)).
Fixed
Storage.Fileno longer discards a validcreds.termas "corrupt" on a cold start (#1). Decoding used:erlang.binary_to_term/2with[:safe], which refuses to mint atoms that aren't loaded yet. A persisted creds term legitimately carries generated proto-struct atoms (e.g.Amarula.Protocol.Proto.ADVSignedDeviceIdentity); if creds were read before that module loaded,[:safe]raised and the entry was swallowed as a miss — silently logging the session out and forcing a re-pair, intermittently and load-order dependent. Decode now falls back to an unsafebinary_to_term/1on the[:safe]-specific rejection (these files are self-written and trusted), so a valid creds file is always recovered; only a genuinely undecodable file is treated as a miss.
Added
- More inbound message types are now classified instead of collapsing to
{:other}: view-once media (unwrapped to its inner media), PTV round notes (as media:video), pin/keep updates, group invites, events, and the receive-only WhatsApp Business / interactive types (product, order, button / list / template / interactive responses). Surfaced on%Amarula.Msg{}via itstype+content. - Send albums (grouped media).
Amarula.send_album/3takes a list of{type, data, opts}image/video items; it sends the album parent, then each item referencing it (viamessageContextInfo.messageAssociation, MEDIA_ALBUM). - Create events.
Amarula.send_event/4sends an event (name + optional description, location, join link, start/end time, extra-guests flag). Responding (RSVP) is not yet supported — it's an encrypted response, a separate seam like poll votes. - Send a group invite as a chat message.
Amarula.send_group_invite/5wraps a group's invitecode(fromAmarula.Group.invite_code/2) into a tap-to-join card sent to a chat.opts::group_name,:caption,:expiration. - Cast a poll vote.
Amarula.send_poll_vote/5encrypts and sends a vote on an existing poll (the inverse of the tally/decrypt path). Pass the poll'smessage_ref, itsmessage_secret, and the chosen option names. - Pin / unpin and keep / unkeep messages.
Amarula.pin_message/2,unpin_message/2(pin for everyone),keep_message/2,unkeep_message/2(exempt a message from a disappearing chat). Each takes amessage_ref. - View-once media and PTV (round video notes).
send_media/5gains:view_once(wrap as view-once) and:ptv(for:video, send as a round video note).
Changed (breaking)
- Unified how you reference an existing message.
send_reaction/3,send_edit/3,send_revoke/2(and the newsend_poll_vote) now take amessage_ref— a%Amarula.Msg{}(the message you received) or a{jid, msg_id}tuple — instead of a raw%Proto.MessageKey{}. This removes the last proto type from the public API and gives one consistent currency: pass back the struct you got, or the chat jid + id. Migration: replace%Proto.MessageKey{remoteJid: jid, id: id}with{jid, id}, or pass the%Amarula.Msg{}directly.
Added
- Reply (quoted) and mentions on outgoing messages.
Amarula.send_text/4andAmarula.send_media/5now take:quoted(an%Amarula.Msg{}to reply to) and:mentions(a list of jids/%Amarula.Address{}) opts. A text with either is sent as anextendedTextMessagecarrying thecontextInfo; media attaches thecontextInfoto the media submessage. Backward compatible — existing 3-argsend_textcalls are unchanged.
0.2.1 - 2026-06-20
Internal refactor only — no API changes. Every public Amarula function
keeps its exact signature, arity, defaults, and return shape (verified by diff);
consumer event shapes are unchanged. No action needed to upgrade.
Changed (internal)
- Decomposed the ~3.9k-line
Amarula.Connectiongod-module: the pure bodies of its callbacks moved into eight focused submodules underAmarula.Connection.*—SendOps,GroupOps,PreKeyOps,Pairing,Notifications,Receive,AppStateOps, and the sharedAckLifecycleseam.Connectionremains the single per-connection process and dispatcher; anything bound to the live socket, cipher, IQ correlation, or Storage stayed put. Each new module has direct unit tests (100% line coverage). - The
Amarulafacade now calls the connection process directly (GenServer.call) instead ofdefdelegate-ing throughConnection's client wrappers, removing one indirection hop. Public signatures are untouched.
0.2.0 - 2026-06-19
A consumer-API cleanup pass: leaner, more consistent Amarula facade with no
protocol types leaked. All breaking changes are mechanical — rename the call
or swap two arguments; behaviour is unchanged.
Changed (breaking)
Namespaced the group / profile / contact families — these moved off the flat
Amarula facade onto dedicated modules, shrinking the top-level surface and
dropping the prefix-stutter. Every operation is unchanged; only the module differs:
| Before | After |
|---|---|
Amarula.group_create/3, group_leave/2, … (all group_*) | Amarula.Group.create/3, Amarula.Group.leave/2, … (drop the group_ prefix) |
Amarula.group_metadata/2, Amarula.list_groups/1 | Amarula.Group.metadata/2, Amarula.Group.list/1 |
Amarula.on_whatsapp/2, fetch_profile_status/2 | Amarula.Contacts.on_whatsapp/2, Amarula.Contacts.fetch_status/2 |
Amarula.profile_picture_url/3, update_profile_*, remove_profile_picture/2 | Amarula.Profile.picture_url/3, update_picture/3, update_status/2, remove_picture/2 |
Event tag — consumer events are now tagged :amarula, not :whatsapp, so the
origin is clear and the tag won't collide with other libraries:
# before
{:whatsapp, :messages_upsert, data}
# after
{:amarula, :messages_upsert, data}Update every receive/handle_info clause that matches the event tuple.
Renamed functions (same behaviour, clearer name):
| Before | After |
|---|---|
Amarula.Address.to_wire/1, to_wire!/1 | Amarula.Address.to_jid/1, to_jid!/1 (now accept a string or an Address) |
Amarula.fetch_status/2 | Amarula.fetch_profile_status/2 |
Amarula.presence_subscribe/2 | Amarula.subscribe_presence/2 |
Amarula.group_setting/3 | Amarula.group_update_setting/3 |
Amarula.group_ephemeral/3 | Amarula.group_toggle_ephemeral/3 |
Argument order — jid now always comes right after conn, matching every
other send:
Amarula.mark_read(conn, message_ids, jid, participant)→Amarula.mark_read(conn, jid, message_ids, participant)Amarula.send_media(conn, type, jid, data, opts)→Amarula.send_media(conn, jid, type, data, opts)
Return shape — Amarula.group_requests/2 now returns clean
[%{jid: Amarula.Address.t(), requested_at: integer | nil}] instead of raw
string-keyed wire attributes.
Removed (breaking)
Removed from the Amarula facade. These were either a protocol leak or internal
plumbing the library handles for you; the underlying function still exists for
power users:
| Removed | Use instead |
|---|---|
Amarula.send_message/3 (took a raw %Proto.Message{}) | Amarula.send_text/3 & friends, or Amarula.Connection.send_message/3 |
Amarula.request_resend/2 | Amarula.Connection.request_resend/2 |
Amarula.resolve_lid/2 | Amarula.Contacts.resolve_lid/2 (addressing resolves automatically on send) |
Amarula.normalize_jid/2 (was canonical_jid) | Amarula.Connection.canonical_jid/2 |
Documentation
- Dropped internal jargon ("wire jid", "boundary", "total function", "canonical") from the consumer-facing docs in favour of plain language.
0.1.0 - 2026-06-19
First public release.
Added
- Offline sandbox +
Amarula.Testing— test your bot's receive→reply logic with no WhatsApp connection.Amarula.new(%{profile: x, offline: true})runs a connection with no socket whosesend_*calls short-circuit to{:ok, msg_id}(no encrypt, no frame, no real-world effect).Amarula.Testing.start_offline/1starts one;deliver_text/2anddeliver/2feed synthetic inbound messages through the real decode/classify pipeline, so your bot receives a true%Amarula.Msg{}.send_media/5is unsupported offline (needs a live socket). Amarula.list_profiles/1— list the profiles that have stored credentials in a given storage source (aConn, aStorage.Scope, or a{adapter, opts}/ bare-opts storage spec). Returns the names you'd pass as:profileto reconnect.Amarula.list_profiles_with_metadata/1— likelist_profiles/1, but each entry carries the logged-in identity read from that profile's creds (%{profile, jid, lid, name}), for building account pickers.Amarula.Storagegains an optionallist_profiles/1behaviour callback, implemented by theFileandDETSadapters. Adapters that don't implement it report{:error, :not_supported}.
Changed
- Teardown API reworked.
wipe_credentials/1is now the single destructive path: it unlinks the companion server-side (remove-companion-device, the phone drops the device), wipes all local storage for the profile, then disconnects. After it, the profile must be re-paired.
Removed
Amarula.logout/1(removed). For a non-destructive teardown that keeps credentials, usedisconnect/1(closes the websocket only) orstop/1(takes the supervision tree down and frees the profile slot). The server-side device-unlink now lives only inwipe_credentials/1.