Skip to main content

Overview

whatsapp-rust implements the Signal Protocol for end-to-end encryption of both one-on-one and group messages. The implementation is based on Signal’s libsignal library, adapted for WhatsApp’s specific protocol requirements.
The Signal Protocol implementation handles cryptographic primitives. Any modifications to this code require expert-level understanding of cryptographic protocols to avoid security vulnerabilities.

Architecture

The Signal Protocol implementation is split across two main locations:
  • wacore/libsignal/ - Platform-agnostic Signal Protocol core (Rust port of libsignal)
  • src/store/signal*.rs - WhatsApp-specific storage integration with Diesel/SQLite

Key Components

Double ratchet protocol

The Double Ratchet algorithm provides forward secrecy and post-compromise security for 1:1 messages.

Session Initialization

Two participants initialize a session using Diffie-Hellman key exchange:
Key Derivation:
  1. Compute shared secrets from ephemeral key exchanges
  2. Derive root key and chain key using HKDF-SHA256:
  3. Initialize sender and receiver chains
Location: wacore/libsignal/src/protocol/ratchet.rs:41-172

Message Encryption

Each message advances the sender chain and derives ephemeral message keys:
Process:
  1. Load current session state
  2. Get sender chain key and derive message keys:
  3. Encrypt plaintext with AES-256-CBC:
  4. Create SignalMessage with MAC for authentication
  5. Advance chain key and save session state
Message Format:
  • SignalMessage: Standard encrypted message
  • PreKeySignalMessage: Includes prekey bundle for session establishment
Plaintext padding. Before encryption, the serialized wa::Message is padded with a uniform-random number of bytes in 1..=16 (the pad length is repeated as the byte value, matching WA Web’s rand % 16 + 1 and whatsmeow). v0.6 fixed a prior scheme that masked the length with & 0x0F, which skewed the distribution toward 15 and could never emit 16 — a subtle fingerprinting divergence from the official client. The receiver strips the padding by reading the final byte as the length.

Message Decryption

Decryption handles out-of-order delivery and tries multiple session states:
Process:
  1. Try current session state first
  2. If MAC verification fails, try previous (archived) sessions
  3. Derive/retrieve message keys for the counter
  4. Verify MAC:
  5. Decrypt with AES-256-CBC
  6. Promote successful session to current if needed
The implementation optimizes memory by using take/restore patterns to avoid cloning session states during decryption attempts (see session_cipher.rs:495-619).

Chain key ratcheting

Message keys are derived from chain keys, which advance with each message:
Location: wacore/libsignal/src/protocol/ratchet/keys.rs

Chain key overflow protection

The chain key index is a u32 that increments with each message. Without overflow protection, the index could silently wrap past u32::MAX (4,294,967,295) back to 0, creating a counter reuse vulnerability that breaks cryptographic guarantees (nonce reuse in message key derivation). Both 1:1 and group chain keys use checked_add() to return a typed error instead of wrapping:
A chain key reaching u32::MAX iterations indicates an abnormally long-lived session. In practice this should never occur — ratchet key rotations reset the chain counter with each new Diffie-Hellman exchange.
Location: wacore/libsignal/src/protocol/ratchet/keys.rs, wacore/libsignal/src/protocol/sender_keys.rs

Forward Jumps

The protocol tolerates out-of-order messages up to a limit. Peer sessions and group sender-key chains match WhatsApp Web’s signalFutureMessagesMax; a pairwise session with one of your own other devices gets a wider (but still bounded) ceiling, since multi-device app-sync legitimately jumps far ahead and the peer is trusted:
Before this change, self-device sessions were exempt from the limit entirely (jumps beyond MAX_FORWARD_JUMPS were logged and allowed through). MAX_FORWARD_JUMPS_SELF (25,000) now bounds that path too — still wide enough for legitimate app-sync catch-up, but no longer unbounded. Peer sessions and group sender-key chains dropped from 25,000 to 2,000, matching WA Web’s signalFutureMessagesMax; a message whose counter is more than 2,000 steps ahead is rejected (driving the retry-receipt path) instead of forcing thousands of KDF derivations per message.
Location: wacore/libsignal/src/protocol/consts.rs, wacore/libsignal/src/protocol/session_cipher.rs

DM device fanout

When sending a direct message, the library resolves all known devices for both the recipient and your own account, then encrypts two different plaintexts for two categories of devices:
  • Recipient devices receive the actual message content
  • Own other devices (your other linked devices) receive a DeviceSentMessage wrapper containing the message plus the destination JID, so your other devices can display the sent message in the correct chat

Device resolution

The DM send path builds the full device list in a WA Web-compliant manner (matching WAWebSendUserMsgJob and WAWebDBDeviceListFanout):
  1. Local registry first — the client checks the local device registry via get_devices_from_registry() for both the recipient and own account. A network fetch (get_user_devices) is only triggered on a cache miss, avoiding unnecessary LID-migration side effects.
  2. Hosted device filtering — devices flagged as hosted (via is_hosted()) are filtered out, matching WA Web’s DBDeviceListFanout exclusion.
  3. Sender device exclusion — the exact sender device is removed from the list so ensure_e2e_sessions never creates a self-session. This matches WA Web’s isMeDevice check in getFanOutList.
  4. Self-DM deduplication — when sending to your own account, the recipient and own device lists overlap. A HashSet-based dedup pass (matching WA Web’s Map keyed by toString) removes duplicates.

Device partitioning

The partition_dm_devices function classifies all resolved devices into recipient and own groups, and excludes the exact sender device (the current device) entirely. It partitions all_devices in place — swapping recipient devices to the front of the passed-in Vec — instead of allocating two new device vectors per send:

Sender device exclusion

The exact sender device is identified by matching both the user and device ID against your phone number JID (PN) or your Linked Identity JID (LID):

Own device recognition

After excluding the sender device, the remaining devices are classified using matches_user_or_lid, which checks if a device JID belongs to the same user as either your PN or LID:
This ensures that your own devices registered under your LID (common in multi-device setups) are correctly classified as “own” devices and receive the DeviceSentMessage plaintext — not the recipient plaintext. Without LID matching, your own LID-based devices would be misclassified as recipient devices, causing them to receive the wrong message format.
Both PN-based and LID-based devices must be checked because WhatsApp’s multi-device architecture uses both addressing schemes. A user’s devices may appear under either their phone number JID (@s.whatsapp.net) or their Linked Identity JID (@lid), depending on the device type and registration path.

PreparedDmStanza

prepare_dm_stanza returns a PreparedDmStanza struct containing the stanza node and the locally computed phash for server ACK validation:
The phash is computed from the actual sent device set (after partitioning, with the sender excluded) using MessageUtils::participant_list_hash(). Unlike group messages, the DM phash is not sent on the wire — WA Web only includes phash in the DeviceSentMessage for groups. The DM phash is used purely for local validation against the server’s ACK to detect device-list drift.
The DeviceSentMessage.phash field is set to None for DMs, matching WA Web’s behavior where only group DeviceSentMessage wrappers include a phash. The DM phash is computed and tracked separately by the caller.
Location: wacore/src/send.rs:675-820, src/send.rs

PN→LID session migration

WhatsApp’s multi-device architecture uses two addressing schemes: phone number JIDs (PN, @s.whatsapp.net) and Linked Identity JIDs (LID, @lid). WhatsApp Web always resolves PN→LID before any session operation via createSignalAddress(). whatsapp-rust mirrors this behavior — when a LID mapping is discovered for a phone number, any Signal sessions stored under the PN address are automatically migrated to the corresponding LID address.
The automatic migration described below is also exposed for manual invocation: Signal::migrate_sessions(from, to) runs the same move for a caller-chosen JID pair, and Signal::session_info(jid) inspects a session (migrating a legacy PN-addressed one first if needed) without mutating it further. See the Signal API reference for both.

Signal address resolution

Client::resolve_encryption_jid() mirrors WA Web’s SignalAddress.toString() (WAWeb/Signal/Address.js). It upgrades the JID’s server to its LID counterpart when a mapping is known, and otherwise returns the input unchanged: The device, agent, and integrator fields always round-trip — only the user (replaced with the LID user) and server change. This keeps Cloud API / Meta Business hosted devices on a hosted-flavored LID address rather than collapsing them into the standard @lid server, matching WA Web’s per-device session keying.
resolve_encryption_jid() upgrades PN → LID unconditionally whenever a mapping is known — it governs Signal session addressing only, matching WA Web’s SignalAddress.toString(). The outbound DM wire namespace (the stanza to, <participants>, and DeviceSentMessage destination) is a separate, account-level decision — see DM wire namespace vs. Signal session addressing below.

DM wire namespace vs. Signal session addressing

Since v0.6.x (fix for #941), a DM’s outer <message to> / <participants> addressing is no longer derived directly from resolve_encryption_jid(). Some accounts are not yet 1:1-LID-migrated on WhatsApp’s servers, and those accounts get every LID-addressed DM rejected with ack error="400" even though the underlying Signal session is correctly LID-keyed. Client::resolve_dm_wire_jid() (src/client/lid_pn.rs) makes this account-level decision, mirroring WA Web’s Lid1X1MigrationUtils.isLidMigrated() / WAWebMessageDestinationChat:
  • Migrated account (Client::is_lid_migrated() is true): behaves exactly like before — the wire namespace upgrades PN → LID whenever a mapping is known.
  • Unmigrated account: DMs stay on PN even with a cached LID mapping. A caller-supplied LID with a known PN mapping is mapped back to the PN chat; a LID with no cached mapping is sent as-is (there is no reverse LID→PN network resolution, matching WA Web).
Client::is_lid_migrated() is true when either is true:
  1. The persisted Device.lid_migrated flag (see Storage — DeviceStore), set once from the primary’s pair-success <client-props> (isChatDbLidMigrated) or from a lid_migration_mapping_sync_message protocol message pushed to the primary’s own companions (self-only — see Authentication — one-to-one LID migration state).
  2. The lid_one_on_one_migration_enabled ab prop, as a fallback for sessions paired before the flag existed. The first observation of this prop being on also latches the persisted flag, so the account doesn’t flap back to PN addressing before the next props fetch.
Once set, lid_migrated never reverts for the same account — only pairing a different account onto the same store resets it. Signal session addressing (resolve_encryption_jid) and inbound decrypt are unaffected by any of this; only the outbound DM wire namespace is gated.
This gate applies to 1:1 DMs only. Group sends, which already address everything by the group’s own AddressingMode, are untouched.

Why migration is needed

After pairing, the primary phone may initially establish sessions under a PN address. Once the LID mapping becomes known (from usync, incoming messages, or device notifications), the phone begins sending from the LID address. Without migration, the client holds a session under the PN address but receives messages addressed to the LID — causing SessionNotFound decryption failures.

Proactive migration at LID discovery

When a new LID-PN mapping is learned (via add_lid_pn_mapping), the client scans devices 0–99 for PN-keyed sessions and migrates them. All reads and writes go through the SignalStoreCache rather than the backend directly — this prevents reading stale data when the cache has unflushed mutations (e.g., after SKDM encryption ratcheted the session). The migrated state is flushed to the backend at the end so it survives restarts.
Migration rules per device: Identity keys are migrated independently of sessions — they can outlive deleted sessions and survive session re-establishment.
The migration reads through the cache because the backend may contain stale session data when unflushed cache mutations exist. Reading directly from the backend could skip in-flight ratchet advances, causing the migrated session to decrypt with an outdated chain key.
add_lid_pn_mapping also has a batch form, Client::add_lid_pn_mappings(mappings, source), which durably records many LID↔PN pairs in one call and runs the same per-mapping migration as the single-entry path. It returns how many mappings were actually written, deduplicated against existing records.

On-the-fly migration during decryption

If a message arrives from a LID address and decryption fails with SessionNotFound or InvalidPreKeyId, the client attempts PN→LID migration as a fallback before requesting a retry:
  1. Look up the PN for the sender’s LID
  2. Attempt to migrate PN sessions to LID via the signal cache (same cache-first logic as proactive migration)
  3. Retry decryption with the migrated session (already in the cache — no reload needed)
  4. If DuplicateMessage occurs during post-migration retry, it is silently ignored
  5. Fall back to retry receipt only if migration does not resolve the issue
The InvalidPreKeyId case occurs when a PreKeyMessage references a consumed one-time prekey, but the session actually exists under a PN address (legacy migration). Migrating the session lets Signal use the existing ratchet state instead of looking up the consumed prekey. This migration is attempted in both the identity-change retry path and the initial decryption path. This ensures existing databases are fixed without requiring re-pairing.

Login-time session check

At login, the client checks the session state of own device 0 (primary phone):
  • LID session exists — no action needed
  • PN session only — logged; migration deferred to first message via on-the-fly path
  • No session — will be established on first message exchange
Both migration paths route through the SignalStoreCache, ensuring they see the latest in-memory state. The proactive migration runs when a LID mapping is first discovered and flushes to the backend afterward. The on-the-fly migration handles the case where the database already contains stale PN sessions from before the mapping was known.
Location: src/client/lid_pn.rs, src/client/sessions.rs, src/message.rs

Sender keys (group encryption)

Groups use the Sender Key protocol for efficient multi-recipient encryption.

Sender key address normalization

Sender key records are keyed by a composite SenderKeyName containing the group JID and a sender protocol address string. WhatsApp delivers group stanzas with inconsistent sender addressing — the pkmsg (which carries the SKDM) arrives with a device-qualified participant JID (e.g., 100000000000001.1:75@lid), while the skmsg (the actual encrypted group message) arrives with a bare participant JID (e.g., 100000000000001.1@lid). Without normalization, the sender key would be stored under the device-qualified address during SKDM processing but looked up under the bare address during skmsg decryption, causing NoSenderKeyState failures. The client normalizes the sender JID to its bare form using to_non_ad() (which strips the device component, setting device = 0, agent = 0) at every point where a SenderKeyName is constructed. The SenderKeyName::from_jid() convenience method handles the to_string() conversion automatically:
SenderKeyName::from_jid() is equivalent to SenderKeyName::new(group_jid.to_string(), sender_address.to_string()) but avoids the manual to_string() calls and is the preferred constructor. This ensures the cache key is always in the form "{group}:{bare_user}@{server}.0", regardless of whether the original stanza used a device-qualified or bare JID.
Custom implementations that construct SenderKeyName directly must also normalize the sender JID to its bare form. Failing to do so will cause sender key lookup mismatches and decryption failures for group messages.
Location: src/message.rs, wacore/libsignal/src/store/sender_key_name.rs, wacore/binary/src/jid.rs (to_non_ad())

Sender key distribution

Each participant generates and distributes a sender key:
Structure:
  • Chain ID: Random 31-bit identifier for this sender key session
  • Iteration: Message counter (starts at 0)
  • Chain Key: 32-byte seed for deriving message keys
  • Signing Key: Ed25519 public key for message authentication

Group Encryption

Messages are encrypted with the sender’s current chain key:
Process:
  1. Load sender key state for the group
  2. Derive message keys from current chain key
  3. Encrypt with AES-256-CBC
  4. Sign message with Ed25519 private key
  5. Advance chain key

Group Decryption

Recipients decrypt using the sender’s distributed key:
Process:
  1. Parse SenderKeyMessage
  2. Look up sender key state by chain ID
  3. Verify Ed25519 signature
  4. Derive message keys for iteration (handling out-of-order)
  5. Decrypt with AES-256-CBC
Group decryption maintains up to MAX_FORWARD_JUMPS (2,000) cached message keys per sender. This prevents resource exhaustion attacks but limits tolerance for extreme out-of-order delivery.

Unknown device detection

During group message decryption, the client checks whether the sender’s device is present in the local device registry via is_from_known_device(). This detection triggers in two places within the group message processing path:
  1. After successful skmsg decrypt — if the sender device is not in the registry, the decrypted message is still processed and delivered normally. Signal decryption success already proves the sender holds a valid session key, so discarding the message would only add latency via an unnecessary retry round-trip. A background device sync is triggered to update the local device registry.
  2. After a NoSenderKeyState error — if the sender device is unknown, the retry reason is upgraded from NoSession to UnknownCompanionNoPrekey
In both cases, the client queues a device list synchronization for the sender’s user JID. The behavior depends on the connection state:
  • Online: the client immediately invalidates the cached device registry for the user and fires a background usync request to refresh the device list
  • Offline (during offline sync): the unknown device’s user JID is batched into a PendingDeviceSync set, which is flushed after offline sync completes (see Deferred device sync)
Primary devices (device ID 0) are always treated as known — the check only applies to companion devices. This mechanism ensures that group messages from newly-paired companion devices are delivered immediately without waiting for a retry round-trip. The background device sync updates the local registry so future messages from the same device are recognized directly.
Location: src/message.rs, src/client/device_registry.rs, src/pending_device_sync.rs

Retry receipt from unknown group device

When the client receives a retry receipt, handle_retry_receipt checks whether the requesting device is present in the local device registry. Previously the handler dropped all retries from unregistered devices — this was safe for WA Web because WA Web keeps participant device lists fresh via a pre-send sync, so any legitimate requester is already known before the send. For a library client, a participant device can legitimately be absent from the local registry: if the device joined between the last device-list sync and the group send, it will have received the skmsg from the server but never obtained a sender key, causing it to retry indefinitely. The retry receipt may carry a <keys> bundle — the ADV-signed device-identity, the identity key, a one-time prekey, and the signed prekey — which is everything needed to establish a Signal session and resend. But a newly-linked device that has no bundle still retries forever if the client only drops it: the reconciliation that fires when a prekey fetch returns 406 never triggers for that device because it was never in the send set. Whenever a retry arrives from an unknown device, handle_retry_receipt now calls schedule_unknown_device_sync before consulting should_drop_unknown_device_retry. This treats the retry as a staleness signal: the requester’s user JID is enqueued for a device-list resync (deduplicated via PendingDeviceSync, so a retry storm from a single device cannot fan out into a usync storm). Once the resync completes, the device appears in the registry and future sends include it in the sender-key distribution — the retries stop. This mirrors WA Web’s syncDeviceListJob trigger on the retry path. The drop predicate still controls whether the current retry is recovered or dropped:
When the bundle includes a <device-identity>, process_retry_key_bundle validates the ADV chain against the requester’s account key (using the stored primary identity as a fallback when the server omits account_signature_key). A present-but-invalid ADV result is a hard error; the session is not built. If <device-identity> is absent from the bundle, or if no account key can be found, the check is skipped with a warning and the session is built anyway — matching the behaviour of the regular prekey-fetch path. The drop predicate only gates on syntactic <keys> presence, so the ADV guarantee is conditional on the bundle including a well-formed <device-identity>. This mirrors whatsmeow’s approach of building the prekey session directly from the receipt bundle without a device-registry gate. Location: src/retry.rs, wacore/src/protocol/retry.rs, src/pending_device_sync.rs

Immutable sender key loading

The SenderKeyStore trait’s load_sender_key method takes &self (not &mut self), allowing sender key lookups to proceed under a read lock. This is safe because loading a sender key is a pure read operation — no state is mutated. The store_sender_key method still requires &mut self since it modifies state. This means concurrent group decryptions for different senders can load sender keys in parallel without contention, while writes (SKDM processing) still serialize correctly.
If you implement SenderKeyStore for a custom backend, load_sender_key must use &self (immutable reference). Implementations that previously required &mut self for internal caching should use interior mutability (e.g., Mutex or RwLock) instead.

Sender key existence check

Before distributing sender keys, the group message path checks whether the local sender key already exists. This check uses the SignalStoreCache with a read lock (get_sender_key()), matching the status broadcast path. This avoids acquiring a write lock and prevents unnecessary SKDM re-distribution on every group send.

Per-device sender key tracking

To avoid resending Sender Key Distribution Messages on every group message, the client tracks sender key distribution status per device for each group. This uses a unified sender_key_devices table (see Storage - ProtocolStore) that matches WhatsApp Web’s participant.senderKey Map<deviceJid, boolean> model — a single boolean per device per group indicating whether that device has a valid sender key (true) or needs fresh SKDM distribution (false). The tracking update is deferred until after the server acknowledges the message stanza. This matches WhatsApp Web’s behavior where markHasSenderKey() is only called after the server confirms receipt. Why deferred? If the tracking were updated immediately after building the stanza (but before sending), a network failure between stanza build and send would leave stale entries — devices would be marked as having the sender key when they never actually received it. Subsequent messages would skip SKDM for those devices, causing decryption failures. PreparedGroupStanza return value: prepare_group_stanza returns a PreparedGroupStanza struct containing the stanza node and a skdm_devices: Vec<Jid> field listing exactly which devices received SKDM in this stanza. This eliminates the need for callers to re-resolve devices after sending, closing a race window where the device list could change between stanza preparation and post-ACK tracking update.
Implementation:
  • Group path: After send_node() succeeds, the caller uses the skdm_devices list from PreparedGroupStanza to call set_sender_key_status(group, devices, true). No re-resolution needed.
  • Status path: A late-init boolean tracks whether full distribution occurred. The sender key tracking is only updated after the status stanza is successfully sent.
  • Error recovery: If prepare_group_stanza fails with NoSenderKeyState, all sender key device tracking for that group is cleared and the send is retried with full distribution.
  • Sender key rotation: On rotateKey, the Signal sender key is also deleted for forward secrecy (matching WhatsApp Web’s deleteGroupSenderKeyInfo), and all device tracking is cleared via reset_sender_key_device_tracking — a DB-first clear with a cold-mark fallback (see below).
  • Group <modify> notification (number/LID migration): A w:gp2 <modify> notification (a participant’s number or LID changed) unconditionally force-rotates the own group sender key and invalidates both the persisted and in-memory group metadata cache, matching WhatsApp Web’s modifyParticipantInfo (rotateKey: true). The next send regenerates and redistributes a fresh sender key against the current participant list instead of risking a stale entry for the migrated device. See Client::force_rotate_own_sender_key, src/handlers/notification/groups.rs.
Incremental targeting: Rather than distributing the sender key to all group devices on every message, the client:
  1. Loads the per-device sender key map — first checking the in-memory cache, falling back to the database via get_sender_key_devices
  2. Resolves all current group participant devices
  3. Computes the diff — only devices with has_key=false or not yet tracked receive the SKDM
  4. Passes the targeted device list to prepare_group_stanza via the skdm_target_devices parameter
On the first group send (or any send where the cached map is empty), the filter still runs unconditionally — every resolved participant device is treated as has_key=false and receives the SKDM. This matches WhatsApp Web, which iterates an empty senderKey Map as false per participant. There is no early-exit for an empty cache; otherwise the very first message after a fresh start would skip distribution entirely.
Own devices are never marked has_key=true (#999). The post-ACK warm mark excludes the account’s own companion devices, matching WhatsApp Web’s !isMeDevice guard on markHasSenderKey. They therefore never leave the “not yet tracked” bucket above and are re-included as SKDM targets on every send — see the follow-up note under “Parallelized group encrypt fan-out” for why. External group members are unaffected: a successful distribution still marks them warm.
Location: src/send.rs, src/client/sender_keys.rs, wacore/src/send.rs

Parallelized group encrypt fan-out

The group send path no longer serializes encryption behind a client-level lock. prepare_group_stanza and encrypt_for_devices now take an explicit &runtime handle (&*self.runtime) so per-device encryption can run on runtime::blocking() tasks concurrently. Combined with the move to update_device_lists (batched device-registry writes) and a no-lock IdentityAdapter::is_trusted_identity stub, group fan-out scales with the runtime’s worker count rather than with a single critical section. This is an internal performance change — no public method on Client::send_message was renamed, and the order of <to> children in the resulting stanza is unchanged. If you implemented a custom SignalStore, note that update_device_lists(records: Vec<DeviceListRecord>) is now part of the trait so the fan-out can batch its writes.
While per-device encryption runs concurrently, the sender-key chain is protected by two separate locks per (group, sender) pair:
  1. Session-setup lock (SenderKeyStore::session_setup_lock) — held only across ensure_sessions_for_devices (prekey fetch + X3DH). May span network I/O. Warm sends (no SKDM needed) never take it, so they are never blocked by a cold send’s network round-trip.
  2. Chain lock (SenderKeyStore::sender_key_lock) — held across SKDM creation + pairwise encrypt fan-out + skmsg encrypt. Pure CPU; never spans network I/O. This is the invariant that prevents two concurrent sends from splitting the key between the SKDM and the skmsg.
Prior to #807, a single chain lock covered both phases, causing concurrent group sends to serialize behind a server round-trip whenever a new session needed to be established. Now only the CPU phase is in the critical section. Different groups (or different senders) encrypt fully in parallel, unchanged.encrypt_for_devices is composed of two public halves: ensure_sessions_for_devices (network, returns SessionPlan) and encrypt_for_devices_with_sessions (CPU, consumes SessionPlan). The DM path calls encrypt_for_devices unchanged; the group path calls them separately with the chain lock taken only around the second.
Per-device session lock around the SKDM fan-out (v0.6). The chain lock above only serializes the sender-key chain — it does not cover the pairwise Signal sessions that encrypt_for_devices_with_sessions mutates for each SKDM target device. Those are the same pairwise sessions the DM path locks (see “DM per-device locking” under Single-allocation session lock keys below) via session_lock_for() / session_mutexes_for(). Before #990, the group fan-out held only the chain lock, a disjoint key, so a concurrent DM (or another group send) sharing a device could race that device’s pairwise ratchet — both sides load chain index N and both store N+1, silently dropping one advance. When the lost advance carried the SKDM, that member never received the sender key and every subsequent skmsg stayed undecryptable for it until a retry re-distributed.prepare_group_stanza now acquires the SKDM targets’ per-device session locks through SendContextResolver::lock_device_sessions() before taking the chain lock, and releases them right after the fan-out — the skmsg chain encrypt that follows only touches the sender-key chain, never a pairwise session. The Client implementation of this hook reuses build_session_lock_keys() + session_mutexes_for(), so the group and DM paths serialize on the exact same mutexes, in the same sorted order, and always acquire session locks before the chain lock — no path takes the reverse order, so this cannot deadlock. The hook defaults to a no-op, so a custom SendContextResolver (as used in tests and benches) is unaffected unless it opts in.
Session-setup failures are isolated per device (v0.6). ensure_sessions_for_devices used to abort with Err the moment process_prekey_bundle failed for any one target device. Since prepare_group_stanza gates the entire SKDM fan-out on session_plan.is_some(), one device’s X3DH failure nulled the plan and every device in the cohort — not just the failing one — got no SKDM, even though the skmsg still shipped and the phash covered the full set. An external member recovers via a retry receipt, but an own companion’s retry hits mark_forget_sender_key with exclude_own_devices=true, which filters own-user JIDs and returns early — so that companion stayed has_key=true forever and couldn’t decrypt the group from that device until an unrelated full rotation (participant removal or PN↔LID migration).As of #996, a device whose session setup fails is logged and skipped rather than aborting the plan — matching WhatsApp Web’s GroupKeyDistributionMsg, which wraps each device’s ensureE2ESessions in its own try/catch and drops only the failing one. The sessionless device is then naturally excluded by the encrypt fan-out (which already skips devices without a session), so every other device still receives its pairwise SKDM.#996 closed the primary harm — an unrelated device’s setup failure no longer suppresses the whole cohort’s SKDM. A narrower window remained: the warm mark (update_sender_key_devices, called after the server ACK) recorded the full distribution target as has_key=true, including our own companion devices, regardless of whether each one’s pairwise SKDM encryption actually succeeded. Since the forget path (mark_forget_sender_key) excludes own devices for the reason above, an own companion whose one SKDM encryption failed — or that was warm-marked without ever receiving a node — was marked warm and could never be un-marked: a permanent orphan until an unrelated full rotation. External devices didn’t have this problem; they recover through the retry-receipt forget path.#999 closes this residual by excluding own devices from the warm mark too (exclude_own_devices=true), mirroring WhatsApp Web’s ParticipantStore helper, which guards both markHasSenderKey and markForgetSenderKey with the same !isMeDevice check. Own companions are therefore never memoized as has_key=truefilter_skdm_targets (“Per-device sender key tracking → Incremental targeting” above) always re-includes them, so they get a fresh SKDM on every group send. This is a deliberate trade-off (a few extra pairwise SKDM nodes per send when the account has companions) in exchange for making the orphan impossible. External devices are unaffected: a successful distribution still marks them warm, and the retry-receipt path still repairs any that go stale.
The group distribution lane now guards the full audit-reset-redistribute sequence, not just the SKDM fan-out (#1043). Previously Client::group_distribution_lock() (see “Parallelized group encrypt fan-out” above) was taken only around the cold SKDM send itself. Sender-key deletion (participant-removal rotation, forced own-key rotation), per-device tracker resets, and the status-broadcast distribution path could run concurrently with that lock held elsewhere, letting an encrypt racing a rotation restore a retired key after deletion, or a tracker reset race stale delivery marks back onto a new chain.rotate_sender_key_on_participant_remove, force_rotate_own_sender_key (now taking &Jid instead of a pre-stringified group ID), warm group sends, status sends, phash-mismatch recovery, and periodic sender-key rotation all now hold the same per-group lane across their own-key delete/reset and the following redistribution. A rotation that arrives while a send is mid-fan-out waits for the lane instead of deleting the chain state out from under it; a send that arrives mid-rotation waits for the rotation to finish before re-auditing device state. Held lanes are never capacity-evicted, so a live rotation or fan-out cannot be silently dropped from the map mid-operation (see group_distribution_locks_capacity in the Cache Configuration reference).
Sender-key tracker resets are DB-first (#1043). reset_sender_key_device_tracking replaces the old direct clear_sender_key_devices + cache-invalidate call at every rotation and redistribution site. It clears the per-device tracking row-by-row in the database first, and only invalidates the in-memory SenderKeyDeviceCache after that durable clear succeeds. If the DB clear fails, every existing tracked row is instead marked cold (has_key=false) as a fallback so the next send still re-distributes; if that fallback write also fails, the operation returns an error and the send stays fail-closed rather than risking a stale has_key=true row surviving onto a freshly rotated chain.The unknown-participant rotation in retry receipt handling is a special case: handle_retry_receipt deletes the own sender key and resets tracking for a <receipt type="retry"> from an unrecognized group participant, then must still fall through to the per-chat resend rate limiter and other throttles further down the same function. The signal cache is now explicitly flushed right after the rotation — before any later throttle can return early — so a rotation is never left un-persisted by an unrelated early exit later in the same call.
Observability: distribution-lane pressure is exposed on memory_report() (#1043). Client::memory_report() now reports group_distribution_locks (live lane count), group_distribution_lock_evictions (cumulative cold evictions), and group_distribution_lock_eviction_blocks (cumulative evictions skipped because the lane was live) — see memory_report(). These update only under capacity pressure and add no allocation or per-message cost below the soft cap.

In-memory sender key device cache

The SenderKeyDeviceCache provides an in-memory caching layer over the per-device sender key tracking data stored in the database. Without this cache, every group send would require a database round-trip to load the sender key device map — the cache eliminates that overhead after the first load for each group.
Key design decisions:
  • Time-to-idle eviction: The cache uses TTI semantics (default: 1 hour, 500 entries), so entries for inactive groups are automatically evicted while frequently-used groups stay cached
  • Pre-parsed, pre-indexed maps: Database rows are parsed into a SenderKeyDeviceMap struct that provides O(1) lookups by user and device ID, avoiding per-query string parsing
  • Single-flight initialization: The get_or_init method uses PortableCache’s single-flight get_with — if multiple concurrent group sends for the same group trigger a cache miss simultaneously, only one database read executes and all callers share the result
  • Explicit invalidation: The cache is invalidated when sender key state changes (rotation, error recovery, retry failures) so stale data is never served
SenderKeyDeviceMap structure: The SenderKeyDeviceMap pre-parses JID strings from the database into a user-to-devices HashMap for efficient lookup:
Cache invalidation points: You can tune the cache capacity and TTI via the sender_key_devices_cache field in CacheConfig. Location: src/sender_key_device_cache.rs, src/send.rs

Phash validation for stale device list detection

When sending group, status, or DM messages, the library validates the participant hash (phash) returned in the server’s acknowledgment against the locally computed phash. A mismatch indicates that the server’s view of participant devices differs from the client’s — meaning the local device list is stale. How it works:
  1. Before sending, the client obtains the locally computed phash — from the stanza phash attribute for group/status messages, or from PreparedDmStanza.phash for DMs
  2. A oneshot ack waiter is registered for the message ID via register_ack_waiter
  3. The message stanza is sent to the server
  4. A background task (spawn_phash_validation) awaits the server’s ack (with a 10-second timeout)
  5. The server’s ack includes its own phash — if it differs from the local value, the client invalidates caches
On mismatch, the following caches are invalidated: For DM messages, the phash covers both recipient and own devices (matching WA Web’s syncDeviceListJob([recipient, me])). On mismatch, the client invalidates the device registry cache for both the recipient’s user JID and your own phone number (PN) JID, ensuring the next send re-fetches the current device list for both parties.
The phash validation runs asynchronously in the background and does not block the send path. If the server ack times out (after 10 seconds) or the oneshot channel is dropped, the validation is silently skipped. This matches WhatsApp Web’s approach of using phash as a best-effort staleness detector rather than a hard requirement.

WA Web phash parity (v0.6)

Two corrections aligned the group phash with WA Web’s phashV2:
  • Full device set, every send. The group phash is now computed over the complete resolved participant device set plus the sending device on every send — not just the devices that received an SKDM in that stanza. Warm sends (which distribute no new SKDM) now pass the full resolved set via the all_devices_for_phash parameter to prepare_group_stanza, so the phash matches the server’s view even when the SKDM target set is empty. Status broadcasts keep their prior phash behavior.
  • Standard base64 alphabet. The phash now encodes with the standard base64 alphabet (+ / /) instead of URL-safe (- / _), matching WA Web and whatsmeow.
The client also now persists group metadata locally after a query and sends the stored participant phash on the next group query, letting the server answer not-modified (304) when membership is unchanged — saving a full metadata round-trip. Location: src/send.rs, src/client.rs

Cryptographic Primitives

AES-256-CBC (message content)

Used for encrypting message bodies in both 1:1 and group messages:
Location: wacore/libsignal/src/crypto/aes_cbc.rs

Thread-Local Buffers

The implementation uses thread-local buffers to reduce allocations:
Location: wacore/libsignal/src/protocol/session_cipher.rs:14-54

HKDF-SHA256

Used for key derivation in session initialization:
Location: wacore/libsignal/src/protocol/ratchet.rs:18-39

PreKey Management

Pre-keys enable asynchronous session establishment in the Signal Protocol. whatsapp-rust manages pre-key generation and upload to match WhatsApp Web’s behavior.

Configuration

The per-batch upload count is configurable through the builder/factory API (default 812, matching WhatsApp Web’s UPLOAD_KEYS_COUNT). The upload-trigger threshold is a private constant.
Values outside [5, 65_535] are clamped at upload time (an out-of-range value logs a warn!). The floor avoids an empty-but-flagged pool or a re-upload loop. The ceiling is the wire-format limit: the upload IQ encodes the pre-key list length as a u16, so a larger batch would generate keys locally and then fail to encode. Per-key X25519 generation and protobuf encoding for the batch are offloaded via wacore::runtime::blocking (runtime-agnostic; runs inline on wasm) since the caller-controlled batch size can be large.

Pre-key ID counter and wrap-around

Pre-key IDs use a persistent monotonic counter (Device::next_pre_key_id) that only increases, matching WhatsApp Web’s NEXT_PK_ID pattern:
This approach prevents ID collisions when pre-keys are consumed non-sequentially from the store. 24-bit wrap-around: WhatsApp Web uses 24-bit pre-key IDs on the wire (3-byte big-endian), so valid IDs range from 1 to 16,777,215 (2^24 − 1). When the persistent counter grows past this boundary, modular arithmetic wraps IDs back into the valid range:
If the counter wraps while unconsumed high-ID pre-keys still exist in the store, the database upsert (ON CONFLICT DO UPDATE) silently overwrites them. This is an accepted trade-off because the server consumes keys well before a full 16M cycle completes.
Location: src/prekeys.rs

Retry-receipt prekey marking

When building a retry receipt that includes keys (should_include_keys), the one-time prekey handed directly to the peer is now also marked uploaded via mark_single_prekey_uploaded, matching WhatsApp Web’s markKeyAsUploaded. Without this, the same prekey ID could be re-offered to the server pool in the next batch upload — a third party fetching the bundle could then consume the identical one-time ID and fail to decrypt. mark_single_prekey_uploaded requires a held prekey_upload_lock guard (a compile-time proof, not just a runtime convention) so the get-or-gen and the watermark write are atomic against the batch upload path. It only advances first_unupload_pre_key_id when the id being marked is still the current window head (idempotent no-op otherwise), and collapses next_pre_key_id onto the wrapped low watermark only when marking the terminal id at the 24-bit edge — a non-terminal high-end head keeps its surviving window key. The device account is validated before the prekey is reserved/marked, so a missing account fails the retry-receipt build without silently abandoning a one-time prekey from the upload window. Location: src/prekeys.rs, src/retry.rs

Force-refreshing pre-keys for device migration

When migrating a device from an external source (e.g., a Baileys session into an InMemoryBackend), the server may still hold pre-key IDs whose private key material you cannot reconstruct. Any pkmsg referencing those IDs will fail permanently with InvalidPreKeyId. The public refresh_pre_keys() method force-uploads a fresh batch of Client::wanted_pre_key_count() pre-keys (default 812; tunable via with_wanted_pre_key_count / set_wanted_pre_key_count), giving the server new IDs the caller has locally. Old unmatched IDs drain naturally as peers consume them.
Internally, this acquires prekey_upload_lock to prevent races with the count-based and digest-repair upload paths, then calls upload_pre_keys_with_retry(force: true) which uses Fibonacci backoff (1s, 2s, 3s, 5s, 8s, … capped at 610s). Two related public methods build on the same prekey_upload_lock-guarded path:
  • Client::refresh_pre_keys_with_count(count) — same force-upload as refresh_pre_keys(), but with a caller-chosen batch size instead of the configured wanted_pre_key_count.
  • Client::ensure_pre_keys() — a non-forced check-and-top-up: uploads only if the server-side pool is below the low-water mark, rather than unconditionally replacing it.
Location: src/prekeys.rs:263-266

Digest key validation

After connection, the client validates that the server’s copy of the key bundle matches local keys. This matches WhatsApp Web’s WAWebDigestKeyJob.digestKey() flow. Wire format:
Validation process:
  1. Query the server for the key bundle digest via DigestKeyBundleSpec
  2. If the server returns 404 (no record), trigger a full pre-key re-upload
  3. If the server returns 406/503 or other errors, log and skip
  4. On success, compare registration IDs
  5. Load each pre-key referenced by the server and extract its public key
  6. Compute a local SHA-1 digest over: identity public key + signed pre-key public + signed pre-key signature + all pre-key public keys
  7. Compare the local hash against the server-provided hash
The <list> node contains <id> children (not <key> children). The parser iterates all children of <list> without tag filtering, matching WhatsApp Web’s mapChildren behavior which does not filter by tag name.
Hash mismatches or missing local pre-keys are logged but do not trigger a re-upload. Only a 404 response (server has no record) triggers re-upload. This matches WhatsApp Web’s behavior where validateLocalKeyBundle exceptions are caught without re-uploading — the normal RotateKeyJob eventually refreshes the signed pre-key.
Client::validate_digest_key() is a public method — callers can trigger this validation pass on demand instead of only relying on the automatic post-connection check.
Location: src/prekeys.rs:218-344, wacore/src/iq/prekeys.rs:170-302

Signed pre-key rotation (RotateKeyJob)

The signed pre-key minted at pairing was otherwise permanent — a forward-secrecy gap. whatsapp-rust mirrors WhatsApp Web’s RotateKeyJob: on a cadence, generate a fresh signed pre-key, upload it, and retain the previous keys so prekey messages already in flight against a rotated-out key still decrypt. Cadence:
  • Checked once per connection. Spawned right after the startup pre-key upload during post-login init, so a slow or failing rotation IQ never delays the rest of login.
  • Rotates once now - last_signed_pre_key_rotation_ms >= SIGNED_PRE_KEY_ROTATION_INTERVAL_MS (7 days / weekly). This interval is the one value not grounded in the captured WA Web bundle — there RotateKeyJob is a persisted, server-tuned background job — so it’s a tunable policy default.
  • A device upgraded in with the field at 0 gets a one-time baseline stamp (DeviceCommand::SetSignedPreKeyRotationBaseline) instead of rotating immediately, so its first rotation lands a full interval out.
  • Single-flighted via Client::signed_pre_key_rotation_lock so overlapping post-login tasks (from reconnect churn) can’t run the rotate/upload/prune sequence concurrently; a losing task just no-ops for that check.
Wire format (RotateSignedPreKeySpec, reuses the upload path’s <skey> encoder so the two can never drift):
Rotation sequence (Client::rotate_signed_pre_key):
  1. Compute new_id = current_id + 1, wrapping to 1 at the 24-bit border (same scheme as one-time pre-key IDs).
  2. Stage the new key pair in the signed_prekeys backend table before upload — an already-staged candidate for that id is reused verbatim, so a retry after an ambiguous failure re-uploads the exact key the server may have already accepted instead of minting a different one under the same id.
  3. Retain the outgoing (current) key in the backend table before upload, so once the server accepts the new key the old id’s decrypt window is already durable — no post-acceptance write can strand it.
  4. Upload via the <rotate> IQ above.
  5. On success, DeviceCommand::SetSignedPreKey atomically installs the new key pair, id, signature, and rotation timestamp. The now-redundant staged copy is dropped, and retained signed pre-keys are pruned (newest-id-first) to SIGNED_PRE_KEY_RETENTION (3 total: the current key + the 2 most recent rotated-out keys).
Error handling (mirrors WA Web’s RotateKeyJob ladder; a rotation failure never fails login): Backend fallback for rotated-out ids: Before this feature, Device::load_signed_prekey (src/store/signal.rs) returned a record only when the requested id matched the current signed_pre_key_id field — the signed_prekeys backend table (which already existed, with full CRUD) was never consulted for other ids. Rotating the key in place would therefore make any in-flight prekey message naming the old id fail with InvalidSignedPreKeyId. load_signed_prekey and contains_signed_prekey now fall back to the backend table for non-current ids, which is what makes rotation safe to ship. Retry, not NACK, once the id ages past retention: A sender’s PreKeySignalMessage can still name a signed pre-key id that has since aged past SIGNED_PRE_KEY_RETENTION (3 total: current + 2 rotated-out) — the backend fallback above has nothing left to return, and InvalidSignedPreKeyId is the correct, permanent answer. On the 1:1 decrypt path (src/message/receive.rs), this now routes to a retry receipt (RetryReason::InvalidKeyId) carrying the current bundle, mirroring the sibling InvalidPreKeyId arm — instead of falling through to the catch-all UnhandledError nack, which would drop the stanza from the offline queue and lose the 1:1 message permanently and silently.
Client::rotate_signed_pre_key() is a public method — callers can force an out-of-cadence rotation directly instead of waiting for the weekly check. It shares signed_pre_key_rotation_lock with the automatic path (so a manual call can’t race a background rotation) and propagates upload failures to the caller rather than swallowing them.
Location: src/features/rotate_key.rs, src/store/signal.rs, src/message/receive.rs, wacore/src/iq/prekeys.rs, wacore/src/store/commands.rs

Re-pair pre-key healing (v0.6)

If the user re-pairs the device (for example by re-scanning the QR code), the server discards its copy of our pre-key bundle even though Device::server_has_prekeys may still read true from the previous pairing. v0.6 resets server_has_prekeys = false immediately after a successful re-pair so the next connect uploads a fresh batch instead of trusting the stale flag. The lock-acquisition for the digest-key validator also moved into validate_digest_key itself. Previously the caller held prekey_upload_lock before calling the validator, which would deadlock when validation hit a 404 and tried to acquire the same lock to perform the re-upload. The lock now wraps only the re-upload path, so the 404→re-upload transition completes without contention. Location: src/handlers/notification.rs, src/pair.rs, src/prekeys.rs

ADV companion identity validation

When fetching a pre-key bundle for a contact’s companion device (WhatsApp Web / Desktop), the bundle’s <device-identity> element is validated to confirm that the fetched identity key is cryptographically bound to the account. This guards against a relay substituting a forged identity key, matching WA Web’s SessionApi.createSignalSession. Account key resolution mirrors WA Web’s validateADVwithIdentityKey (e.accountSignatureKey || t):
  1. In-blob key: If ADVSignedDeviceIdentity.account_signature_key is present and non-empty, it is used directly.
  2. Stored identity fallback: The server legitimately omits this field for a contact’s companion because the client already holds the contact’s primary (device 0) identity in the Signal identity store. When the field is absent, Client::load_account_identity loads it — reading through the SignalStoreCache so any unflushed mutations from the current session are visible. PreKeyFetchSpec::with_account_identities threads the pre-loaded map into wacore’s stateless prekey parser, keeping store access in the whatsapp-rust crate.
Validation results (wacore::adv::AdvValidation): NoAccountKey does not weaken security beyond the pre-existing “device-identity absent” path: a relay could already strip the entire <device-identity> element to bypass the check. It exists so brand-new contacts whose primary identity has never been seen are not silently dropped. The same three-state validation applies in the retry-receipt handler (src/retry.rs) when a companion device requests a re-send. Location: wacore/src/adv.rs, wacore/src/iq/prekeys.rs, src/prekeys.rs

Storage Integration

whatsapp-rust integrates Signal Protocol storage through a layered architecture:
The Device struct implements the libsignal SessionStore, IdentityKeyStore, and other traits. These are wrapped by SignalProtocolStoreAdapter, which adds the SignalStoreCache layer — sessions are cached as SessionRecord objects (not bytes), with serialization deferred to flush(). Each store (sessions, identities, sender keys) is flushed independently under its own lock. Only one store is locked during its I/O — the other two remain free for concurrent encrypt/decrypt operations. The lock is held from snapshot through write through clear, so mutations to the same store are blocked until flush completes, preventing dirty-set races:

Flush scheduling: send vs. receive

When the dirty Signal cache reaches the backend differs by direction, because the two directions have different recovery properties:
  • Send (DM/1:1 sessions) persists through a batched counter lease. SessionRecord reserves its outbound sender-chain counter SENDER_CHAIN_RESERVATION_BATCH (64) values at a time, via SessionRecord::reserve_sender_chain_counters. A send covered by an unexhausted lease is already durable — it only schedules the same coalesced write-behind as the receive path below. The send that exhausts the lease, roughly 1 in 64, raises the ceiling and flushes synchronously, before the stanza reaches the wire. If that flush fails, the send aborts instead of transmitting an advance it couldn’t save. Reusing an outbound counter reuses its message key and IV, so no counter can ever be used before its lease is durable. The lease field is local-only: it’s field 100 in the encoded SessionRecord, outside the vendored whatsapp.proto. By default (SessionRecord::deserialize), every load fast-forwards the sender chain to the lease ceiling, so a crash mid-lease can never re-derive a possibly-spent counter — the store-backed load path relaxes this only for a trusted reload, via the incarnation marker described below.
  • Send (group and status sends) follows the same lease pattern as DMs. SenderKeyRecord reserves its outbound chain iteration SENDER_CHAIN_RESERVATION_BATCH (64) at a time via SenderKeyRecord::reserve_iterations, using the same field-100 local-only encoding and the same fast-forward-on-load recovery as SessionRecord. A send within an unexhausted lease rides the coalesced write-behind; only the send that raises the ceiling, roughly 1 in 64, flushes synchronously before the stanza reaches the wire. Status posts go through the same group-encrypt path and inherit this behavior; status reactions are a DM-branch send and always followed the DM lease instead. Production encryption (wacore::send::encrypt_group_message) delegates to the same group_encrypt primitive these guarantees live in, so there is exactly one sender-key encrypt/advance/store implementation — earlier, a second unguarded copy on the production path meant most warm group sends skipped the pre-wire flush entirely.
  • Receive (live traffic, outside the offline-drain batcher) routes through a single-flight coalescing scheduler (src/signal_flush.rs) instead of flushing per stanza: a burst of receives folds into one flush per ~25ms window, retried with exponential backoff (up to a 5s cap) on backend failure. This is safe because a lost receive-side advance simply re-derives forward on the next message (the receiving chain derives CK_n → CK_n+1), and a consumed one-time prekey stays buffered until its session is durable — a crash inside the window is recoverable.
Every direct Signal-mutating call site now shares the same gate (#1048). The public Signal::encrypt_group_message and Signal::create_participant_nodes accessors, plus VoIP’s outbound call-key fanout (place_call in src/voip/facade.rs), used to call Client::flush_signal_cache_batch_safe() — an unconditional flush — after releasing their session or sender-key chain locks, regardless of whether the mutation actually raised a lease’s durable ceiling. They now call the same Client::persist_signal_state_pre_wire() gate the primary wacore::send path uses: a warm call already covered by an existing lease rides the coalesced write-behind and returns immediately, while only the roughly-1-in-64 call that raises the ceiling still flushes synchronously before its ciphertext is used or sent. Measured against the same real-SQLite harness: warm group encryption dropped from 84.69 ms to 23.98 ms per 1,024-call sample (-71.7%), and warm 4-recipient participant fanout dropped from 49.16 ms to 8.69 ms per 256-call sample (-82.3%). Safety is unchanged — the per-entry durability check that guards what gets written is identical in both cases; the difference is that flush_signal_cache_batch_safe() always entered the flush path even on a warm lease, while persist_signal_state_pre_wire() skips it entirely once the raised lease is already durable. This removes redundant synchronous I/O, not any safety guarantee.
Downgrading to a version that predates the counter lease after running a leased version: the older version ignores the lease field(s) and could reuse counters/iterations that were only reserved (not yet actually sent) by the lease. This applies to SenderKeyRecord (group/status sends) as well as SessionRecord (DM sends). Avoid downgrading a device’s local state across this boundary.
The pre-wire gate is a point-in-time check, not a lock held across the flush, and this specifically affects DM sessions: needs_pre_wire_flush() inspects pending reservations once, and flush() skips any session entry that is currently checked out by a concurrent operation (SessionEntry::CheckedOut), still returning Ok for the entries it did persist. If another task checks out the same session between this send’s lock release and its flush, that session’s reservation can remain pending even though the flush “succeeded” — the caller proceeds to write its stanza regardless. This is a property of SignalStoreCache::flush itself, not specific to retries; it affects any pre-wire-gated DM send that races a concurrent, still in-progress operation on the same session. Sender-key entries have no analogous checked-out state — get_sender_key clones an Arc without removing the cached record, so every dirty sender-key entry is included in a flush’s batch — so group/status sends are not exposed to this race. This is unrelated to what happens if that concurrent operation is then cancelled: see Cancellation-safe session checkouts — a dropped checkout restores its entry synchronously rather than leaving it stranded, but the deferred entry still isn’t included in a flush that already ran before the restore.
Deleting a session or sender-key record — an identity change, a session reset, a rotated sender key — creates a tombstone rather than clearing the record’s pending gate immediately. If a durability gate was open on the record at delete time, SignalStoreCache keeps it open until the backend’s delete_session/delete_sender_key call actually succeeds; a failed delete leaves needs_pre_wire_flush() returning true and is retried on the next flush. Earlier, the gate was released as soon as the tombstone was applied to the in-memory cache, so a failed delete or a crash in that window could let ciphertext reach the wire while the pre-delete chain state was still loadable from the backend, re-deriving already-used key material on reload. A lossy clear() still drops a pending tombstone gate — the tombstone is discarded from the in-memory cache without issuing the backend delete, so the old chain state may remain in the backend.

Clean reload vs. crash recovery

Fast-forwarding past a lease’s reserved ceiling on every reload is the safe default, but it’s also overly conservative for the common case: a clean reconnect or a same-process store re-creation never actually risked losing an in-flight send, yet unconditionally fast-forwarding still burned a full unused batch every time. 32 clean reconnects could push a sender-key chain 2,048 iterations ahead and get rejected once a peer who missed the intervening messages hit MAX_FORWARD_JUMPS (2,000). SignalStoreCache, and the direct, non-cached Device store, now tag a durably-reserved record with a random 128-bit store incarnation, carried in a second local-only field (101, alongside the lease’s field 100) on both SessionRecord and SenderKeyRecord:
  • A live SignalStoreCache generates one incarnation marker when it’s constructed. A reload that observes a matching marker proves the record came from this same live cache and was never lost to a crash, so it’s exact: the chain resumes at the next iteration instead of fast-forwarding.
  • A reload with no marker, a different marker, or a malformed/duplicated marker is untrusted and keeps the conservative fast-forward — this covers process restarts, a freshly constructed cache, and any genuinely lossy discard.
  • SignalStoreCache::clear_after_flush(), used by connection teardown, only evicts a store once it’s fully settled — no dirty, deleted, checked-out, or pending-wire-gate entries. Anything a concurrent write installs after the flush stays resident (and the incarnation stays put) until a later flush settles it; only an actual lossy discard rotates the marker.
  • Direct Device stores hold one process-level incarnation in an OnceLock instead of a per-cache one. Since these stores synchronously await the backend write before returning ciphertext, a new Device wrapping the same backend in the same process is a clean, trusted reload; a process restart gets a fresh marker and stays conservative.
This adds no field to the public Device struct and no synchronous I/O — a marker is generated only when a cache or process starts, or when a lossy boundary invalidates trust. The scheduler is generation-scoped (embeds the connection generation in its atomic state), so a reconnect during an in-flight flush needs no explicit reset: a stale worker from the previous connection cannot mutate the new generation’s state, and stands down when it observes a foreign generation. The offline drain, identity-change recovery, and teardown all keep their own synchronous flushes — they gate acks, receipts, or follow-up reads on durability and are not routed through the receive coalescer. (Teardown’s flush is itself conditional: teardown_inbound_commits_bounded only flushes on a durable drain with no outstanding batch entries; on a timeout or non-durable/pending entries, it clears the cache without flushing instead, relying on server redelivery rather than risking a persisted-but-incomplete advance.) See Inbound Durability Hook for the drain-batch commit ordering, which this coalescing does not change. Retry-receipt recovery (handle_retry_receipt resending to a DM or group requester, src/retry.rs) instead shares the DM/group send durability rule above through one helper, send_retry_stanza: the session lock is released first, then persist_signal_state_pre_wire() runs — flushing synchronously if the retry’s Signal advance crossed an unpersisted lease boundary — before the stanza is written to the wire. An Err from that flush aborts the retry instead of transmitting an advance it couldn’t save; for a DM retry specifically, see the point-in-time caveat above for the narrower case where the flush reports success without having actually persisted this retry’s own checked-out session (group retries are not exposed to that race — see the same caveat). This replaced an earlier unconditional full flush that ran only after the retry stanza had already reached the wire — a crash or persistence failure in that window could reload the old chain state after the ciphertext was already sent. Call Client::flush_pending_signal_state() to force a deterministic settle — e.g. before reading persisted Signal state directly, or ahead of a non-graceful shutdown. Never call it from inside an InboundDurabilityHook or a synchronous, inline EventHandler::handle_event implementation, since settling re-enters the processing permit those run under and would deadlock during an offline-sync drain. Ordinary Bot closure handlers are unaffected — both default delivery modes run the callback in a detached task off the permit.

Record components

wacore-libsignal exposes owned, validated projections of SessionRecord and SenderKeyRecord called components. Use them when you need to interchange or inspect session and sender-key record state without depending on the generated protobuf schema directly — for example in custom store implementations, migration tooling, or offline debugging. This API is purely additive: the protobuf-backed serialize()/deserialize() path is unchanged. A record does not round-trip through into_components()from_components()serialize() byte-for-byte — the conversion applies the validated, normalized export rules described below (counter-lease advancement, stale-chain removal, and bounded truncation), so treat it as a safe normalized re-encoding rather than a lossless copy.

Session and sender-key shapes

SessionRecordComponents mirrors the current_session / archived previous_sessions split already described in Arc previous sessions; SenderKeyRecordComponents mirrors a SenderKeyRecord’s state list:
SessionMessageKeyComponents and SenderMessageKeyComponents hold skipped out-of-order message keys, keyed by chain index/iteration. A session message key’s secret material is SessionMessageKeyMaterial, either the compact Seed(Vec<u8>) import form (expanded through the same canonical derivation used elsewhere — see MessageKeyGenerator) or the Derived { cipher_key, mac_key, iv } form that into_components() always produces on export. Conversions:

Import validation

from_components enforces the same structural invariants the canonical protobuf reader relies on elsewhere in this codebase, rather than accepting whatever shape the caller hands it:
  • A sender chain (the local sending ratchet chain within a pairwise session — not to be confused with a group sender-key chain) must be structurally complete: ratchet public key present, a 32-byte ratchet private key, and a chain key with both its index and 32-byte secret set. An incomplete sender chain fails with SignalProtocolError::InvalidArgument; a session with no sender chain at all (e.g. one just received and not yet replied to) is fine and imports as sender_chain: None.
  • A receiver chain must never carry sender_ratchet_key_private — receiver chains never own the remote party’s private key, so import fails if one is set. Symmetrically, projecting a persisted record into components always reports a receiver chain’s sender_ratchet_key_private as None, silently dropping any non-canonical private material a legacy record might contain, matching how the canonical reader already treats that field.
  • Raw 32-byte public keys and canonically-serialized public keys (type byte + 32 bytes) are both accepted on import; whichever form was imported, into_components() always exports canonical serialization.

Export normalization

into_components() never hands back a session or sender-key chain whose counter could be replayed on re-import:
  • Any durably reserved sender-chain counter range — see the counter-lease mechanics in Flush scheduling: send vs. receive — is advanced to its exclusive ceiling before export. Re-importing the exported components can’t reuse a counter value that was only reserved, not yet actually sent.
  • A sender chain too stale to fast-forward past its reservation is dropped from the exported chain rather than aborting the whole export — the rest of the record (receiver chains, other archived sessions) still exports normally.
  • SessionRecordComponents.previous_sessions is truncated to ARCHIVED_STATES_MAX_LENGTH (40) and SenderKeyRecordComponents.states is truncated to MAX_SENDER_KEY_STATES (5) — the same bounds the records themselves already enforce. See Protocol safety limits.

has_usable_sender_chain

Checks whether a session has a sender chain it could actually encrypt with, rather than assuming one exists. Previously this was effectively best-effort/always-true; the current implementation returns Ok(false) (not an error) when no sender chain is set at all, and otherwise structurally validates the ratchet public key, the ratchet private key, and the chain key are all present before returning Ok(true) — the same completeness check from_components applies on import. SessionRecord::has_usable_sender_chain delegates to the current session’s check, returning Ok(false) when there is no current session. Location: wacore/libsignal/src/protocol/state/session.rs

Debug output redacts secrets

Debug on every *Components type — SessionComponents, SessionChainComponents, SessionChainKeyComponents, PendingKeyExchangeComponents, PendingPreKeyComponents, SenderChainKeyComponents, SenderSigningKeyComponents, SenderMessageKeyComponents, and SessionMessageKeyMaterial — prints private keys, chain/root keys, seeds, and cipher/mac/IV material as <redacted>, while structural fields (indices, counters, iteration numbers, key presence) print plainly:
This makes it safe to log or assert against a *Components value in application code without writing a custom Debug impl to avoid leaking key material. Example — inspecting whether a session can currently send, without touching protobuf types:
Location: wacore/libsignal/src/protocol/record_components.rs

Legacy session v1 interop

Behind the opt-in legacy-session-interop Cargo feature (default off, forwarded through wacore and the root whatsapp-rust crate), wacore-libsignal exposes a typed, transport-agnostic model of the decoded legacy libsignal SessionRecord v1 layout — the format this project’s stores used before the canonical shapes above. It exists for migration tooling importing an externally produced v1 store into the canonical SessionRecord, or projecting a canonical record back into v1 terms; ordinary clients never enable it, so the model compiles out of native builds entirely. Container decoding — turning a legacy store’s transport bytes into these typed fields — is explicitly out of scope; callers own that step and hand this module owned values (Bytes, integers, enums). The module owns everything downstream: chain-role selection, counter translation, lifecycle ordering, pruning, ratchet reconstruction, and skipped-key derivation.
Import — v1 into canonical:
from_indexed_sessions validates that each entry’s outer map key matches its own session’s base key and rejects duplicate base keys or more than one Current session. into_session_record then validates every session — chain roles, key lengths, counters, skipped-key indexes, pending pre-keys, and the same canonical limits enforced in Import validation — retains archived sessions by close time, reorders them by last use to match the v1 decrypt search, truncates to ARCHIVED_STATES_MAX_LENGTH, and reuses SessionRecord::from_components to build the canonical record. LegacySessionLocalContext supplies the local identity key and registration ID, since v1 sessions never persisted them. A v1 sending chain seeds previous_counter at -1 for a ratchet step over a chain that has never sent; LegacySessionChainCounterV1 floors that at zero on import instead of treating it as an error. Every other out-of-range counter fails typed as InvalidChainCounter. Export — canonical into v1 (operational, not byte-exact):
This is a deterministic operational projection, not a round trip: v1 lifecycle timestamps and base-key lookup roles are reconstructed from canonical search/eviction order rather than recovered verbatim, since the canonical record never persisted them in the first place. State the v1 format genuinely cannot represent is rejected with a typed error instead of being silently dropped or inferred — a session with no sender chain, a non-current session_version, a pending key exchange, or a needs_refresh flag all fail as NotRepresentable; a receiver chain holding a derived (non-seed) skipped-message key with no inverse to a v1 seed fails as ChainNotRepresentable::DerivedMessageKey; and a pending pre-key whose base key doesn’t match the session’s own base key fails as PendingPreKeyBaseMismatch rather than producing v1 state the importer would reject on the way back in. Every type in the module redacts key material from Debug — root keys, chain keys, ratchet key pairs, skipped-message seeds, and identity keys all print as <redacted>; only structural fields (roles, counters, indexes, session/chain counts) print plainly, the same convention as the canonical *Components types. Location: wacore/libsignal/src/protocol/legacy_session.rs

Security Considerations

Identity key trust

The implementation verifies identity keys before encryption/decryption:
Location: wacore/libsignal/src/protocol/session_cipher.rs:160-172

Self-only protocol message gating

app_state_sync_key_share, app_state_sync_key_request, and history_sync_notification are protocol messages WhatsApp Web treats as “self-only”: they only carry meaning when delivered from your own account to another of your linked devices. v0.6 hardens handle_decrypted_plaintext so that incoming copies of these messages are dropped unless MessageInfo.source.is_from_me is true, matching WA Web’s WAWebKeyManagementHandleKeyShareApi and whatsmeow’s gating. The consequences if the gate is missing:
  • A spoofed app_state_sync_key_share from a peer would let an attacker inject an app-state encryption key, leading to attacker-controlled mutations of your contacts, blocklist, archive state, etc.
  • A spoofed app_state_sync_key_request from a peer would make the client share its app-state encryption keys with an attacker instead of only with the account’s own companion devices.
  • A spoofed history_sync_notification would point the client at attacker-supplied media for ingestion as your own history.
If you implement a custom message dispatcher, replicate this is_from_me check before honoring any of these protocol messages. Other protocol-message types (REVOKE, EPHEMERAL_SETTING, MESSAGE_EDIT, …) keep their existing semantics. Location: src/message.rs (handle_decrypted_plaintext)

Duplicate message detection

The protocol detects and rejects duplicate messages:
Location: wacore/libsignal/src/protocol/session_cipher.rs:822-827 As of #1072, a DuplicatedMessage from one candidate session is not terminal — the search keeps trying the remaining current and archived sessions, including a closed receiver chain (one with no chain-key seed left, only leftover skipped keys). This covers re-initiations, which reuse the peer’s signed pre-key as a ratchet key: a delayed message whose skipped key survives only in an archived session still decrypts, and a closed chain still recognizes a replay of an already-consumed counter instead of falling through to a generic failure. Once every session has been tried, a recognized DuplicatedMessage outranks BadMac in the final classification: a sibling session that happens to share the same ratchet key derives different message keys for the same counter and fails its MAC as expected noise, but the decrypting client already knows this counter was consumed elsewhere. Classifying that as BadMac would trigger a retry receipt for a message the peer already delivered; the duplicate verdict wins instead, so the replay is acknowledged and silently dropped.

Log level discipline

The protocol layer follows strict rules about what cryptographic material appears in logs and at which level:
  • No private keys or secrets are ever loggedChainKey, MessageKeys, and RootKey types do not expose their key bytes through logging
  • Public keys appear only at warn/error levels — and only when something has gone wrong (untrusted identity, MAC failure)
  • MAC key fingerprints are truncated — only the first 4 bytes (8 hex chars) are logged during MAC verification failures, not the full key:
  • Ratchet keys in debug logs — successful decryptions log the sender ratchet public key (never private) at debug level for diagnostics
  • Pre-key operations use debug for routine operations and warn/info for exceptional conditions
The Signal protocol layer (wacore/libsignal/src/protocol/) uses no trace!-level logging. Sensitive operations stay at debug or above to avoid leaking material in verbose log configurations.

Session state corruption

Detailed logging helps diagnose crypto failures:
This generates comprehensive error logs showing:
  • All attempted session states
  • Receiver chain information
  • Message metadata (sender ratchet key, counter)
Location: wacore/libsignal/src/protocol/session_cipher.rs:365-454

Protocol safety limits

The implementation enforces several hard limits to prevent resource exhaustion and cryptographic failures: Location: wacore/libsignal/src/protocol/consts.rs

Self-DM / sibling decryption recovery

When a message from your own primary phone or another linked companion fails to decrypt, the v0.6 client distinguishes the failure mode and applies the matching recovery strategy: The throttle is a per-peer cooldown (1-hour TTL after the last recreate). In v0.6 the implementation moved from a Mutex<HashMap<Jid, Instant>> to a bounded TTL cache (PortableCache, ~256 entries): the per-peer check-and-stamp is now atomic (serialized by the existing per-peer session lock) and lock-free at the map level, so concurrent retry-receipt spawns from the same peer can’t trigger duplicate recreates. The behavior is unchanged — if a peer is already in cooldown, the client skips re-creation and falls back to a normal retry receipt rather than thrashing the session. Under more than ~256 distinct peers retrying within the window, the cache may evict a recent entry, costing at most one extra recreate (bounded and self-healing). Peer-addressed pkmsg carries the protocol identity so the receiver can verify ownership before installing the new session, blocking spoofed sibling recoveries. This closed a deadlock where self-DM fan-out to a sibling device produced repeated BadMac decrypt failures: the recipient would request a retry, the sender would re-encrypt against the same broken session, and the cycle would continue until the user manually relogged. With the throttled re-creation plus identity-validated pkmsg, the second receipt installs a fresh session and decryption resumes. Self-DM fan-out also gained WA Web parity for the BadMac case: when our own primary phone reports BadMac, the client now treats it as a session-level recovery rather than dropping the message, matching WAWebDecryptOrThrow’s branch on session divergence. Location: src/client.rs, src/retry.rs, wacore/libsignal/src/protocol/session_cipher.rs, wacore/src/send.rs

Performance optimizations

Session object cache

The SignalStoreCache stores sessions and sender keys as deserialized objects (SessionRecord and SenderKeyRecord) rather than serialized bytes, matching WhatsApp Web’s architecture where the JS object IS the cache. Serialization only happens during flush() to the database — not on every store_session or put_sender_key call.
This eliminates protobuf encode (on store) and decode (on load) from the per-message hot path for both 1:1 and group messages. The store_session method takes SessionRecord by value, enabling zero-cost moves from the protocol layer:
All four protocol-layer call sites (message_encrypt, message_decrypt_signal, message_decrypt_prekey, process_prekey_bundle) — plus PreKey setup and group preflight mutations — take ownership of the record and drop it immediately after storing. Taking ownership eliminates the .clone() in the adapter and the compiler enforces no use-after-store. As of #1044, these call sites obtain the record through the SessionCheckout guard (below) rather than a bare load_session/store_session pair, so the zero-cost move still applies but the checkout is now cancellation-safe. Per-message hot path impact:

Cancellation-safe session checkouts

As of #1044, the record backing a CheckedOut entry is owned by a SessionCheckout<'a> guard (wacore/libsignal/src/protocol/storage/traits.rs) instead of being a bare marker. This closes a gap where cancelling the future mid-mutation — a Tokio task abort, a select! losing a race, a timeout — could leave only the marker behind: the ratchet advance the future was computing was lost, and a decrypt’s ratchet advance could persist without its matching identity write, or vice versa. SessionCheckout is obtained via SessionCheckout::load (existing session) or load_or_create (synthesizes SessionRecord::new_fresh() if absent, tracked via had_session), exposes record()/record_mut() for the protocol layer to mutate in place, and is consumed by either:
  • commit(self) — stores the mutated record back through try_store_session_from_checkout, or
  • discard(self) — releases a deliberately-rejected fresh (never-had-a-session) checkout without storing.
If neither runs because the guard is dropped first (the cancellation case), Drop synchronously puts the record back itself — no caller code has to remember to handle cancellation. If the sessions mutex is uncontended, restoration is immediate; if a concurrent flush() holds the lock, the restore is queued as a PendingSessionRestore and replayed the next time any operation acquires the sessions lock, so an aborted task’s state is never dropped even though it can’t wait for the lock itself. The queue itself (SyncMutex<Vec<PendingSessionRestore>>) is unbounded — every lock acquisition drains it first, so it only grows if cancellations under lock contention arrive faster than anything else touches the session store, which does not happen under normal load. Every checkout carries a SessionCheckoutKey { generation, token }: checkout_generation increments on every lossy clear(), and token is assigned monotonically per checkout of a given address. A restore is rejected (rather than silently applied) unless both match the cache’s current state — this stops a checkout issued before a session reset from resurrecting stale state, and stops a stale owner from clobbering a newer checkout of the same address. flush() and cache eviction (clear_after_flush) now preserve live CheckedOut entries — and any PreKey deletion buffered against that address — instead of dropping them; the deferred prekey delete and the session’s own persistence both wait for a later flush once the checkout completes, rather than being lost. SessionStore gained five #[doc(hidden)] methods with pass-through default bodies (load_session_for_update, try_load_session_for_update, try_store_session_from_checkout, cancel_session_checkout, complete_session_checkout), so existing custom SessionStore implementations keep compiling unchanged — only a backend that wants the destructive-update fast path needs to implement them.

Arc previous sessions

SessionRecord.previous_sessions is wrapped in Arc<Vec<SessionStructure>>, making clone O(1) for the ~40 archived previous sessions that previously accounted for ~40% of the serialize cost:
Only rare operations (archive current session, promote previous session, take/restore during session setup) trigger Arc::make_mut and a deep copy.

Redundant signal store write elimination

The SignalStoreCache uses targeted deduplication strategies per store type. For identities (which rarely change), put_dedup() compares incoming bytes against the cached value and skips if identical:
Sessions and sender keys use unconditional put() since they change with every message — dedup would always fail and waste CPU cycles. This split avoids unnecessary database writes during flush() while not adding overhead where it provides no benefit.

Key reuse in cache

The key_for() method on SessionStoreState, SenderKeyStoreState, and ByteStoreState reuses existing Arc<str> keys from the HashMap via get_key_value(), avoiding a heap allocation on every cache operation:
On the hot path (put/delete for addresses already in the cache), this is always a refcount bump instead of a heap allocation.

Single-allocation session lock keys

Session lock keys use the full Signal protocol address string (e.g., 5511999887766@c.us.0). The JidExt trait provides methods for generating these strings, defined in wacore/src/types/jid.rs:
to_protocol_address_string() is used on hot paths (message encryption and decryption) as the key for session_locks. It pre-sizes the output buffer and builds the string in a single allocation, avoiding the two-allocation overhead of constructing a ProtocolAddress and then calling .to_string(). The write_protocol_address_to() free function provides the same formatting but writes into a caller-supplied &mut String buffer, enabling buffer reuse across multiple JIDs (used by session_mutexes_for()). Format examples:
The server s.whatsapp.net is mapped to c.us in address strings, matching WhatsApp Web’s internal format. The trailing .0 is the Signal device_id (always 0 in WhatsApp’s usage).
Usage in message processing:
DM multi-device fanout: The DM send path resolves all known recipient devices and own companion devices, encrypting per-device for each. This matches WA Web’s WAWebSendUserMsgJob behavior where the local device table is read on the send path, and WAWebDBDeviceListFanout filters out hosted devices. The client checks the local device registry first (via get_devices_from_registry()); a network fetch is only triggered on a cache miss to avoid unnecessary LID-migration side effects from get_user_devices. The sender device is excluded (matching WA Web’s isMeDevice in getFanOutList), and for self-DMs, overlapping device lists are deduplicated using a HashSet (matching WA Web’s Map keyed by toString). Own-device namespace alignment (v0.6): When the recipient is addressed in the LID namespace (@lid), the client converts its own companion devices from the PN namespace to LID before fanning out. Without this alignment, a <to> mix of @lid and @s.whatsapp.net participants caused the server to reject the stanza for LID-addressed DMs. Outgoing messages to PN-addressed recipients are unaffected.
The recipient’s namespace here is decided by resolve_dm_wire_jid(), not a raw mapping lookup — on an account that isn’t 1:1-LID-migrated, the recipient (and therefore the whole fanout) stays PN even when a LID mapping is cached.
Own companion devices (your other linked devices) receive per-device encryption for multi-device self-sync via DeviceSentMessage.
WA Web has a bare-<enc> fast path for single primary device (WAWebSendMsgCreateFanoutStanza). This is not implemented in whatsapp-rust because encrypt_for_devices always wraps in <to jid=...> nodes. The <participants> form is accepted by the server regardless.
Fail-fast on total encrypt failure (v0.6). If per-device encryption fails for every recipient device, the DM send now returns an error instead of emitting a stanza with an empty participant list (which the server would silently swallow, making the message look sent when it wasn’t). A partial failure — some devices encrypt, some don’t — still sends to the devices that succeeded.
DM per-device locking: To prevent ratchet desync when concurrent sends and receives operate on the same Signal session, the DM path acquires session locks for all devices involved — the bare recipient plus own companion devices. The build_session_lock_keys() helper resolves encryption JIDs and sorts them for deadlock-free lock acquisition:
  1. Resolves the recipient to its bare encryption JID via resolve_encryption_jid().to_non_ad() (stripping device component)
  2. Resolves own companion device JIDs
  3. Sorts by (server, user, device) using cmp_for_lock_order() and deduplicates
  4. Returns sorted Vec<Jid> — no intermediate String allocations needed for sorting
The session_mutexes_for() helper then converts sorted JIDs to session mutexes, reusing a single String buffer via write_protocol_address_to() to avoid per-JID heap allocations:
The recipient lock key is always the bare form (e.g., 100000012345678@lid.0), matching the decrypt path’s lock format. This ensures send and receive paths serialize on the exact same lock key. Location: wacore/src/types/jid.rs:4-51, src/send.rs:1481-1507

Single-buffer ProtocolAddress

The ProtocolAddress struct stores the full address string "{name}.{device_id}" in a single String buffer, with a name_len marker to split name from suffix. This halves the allocation count compared to storing name and device ID separately, and eliminates the copy when rewriting the address via reset_with().
Both name() and as_str() are zero-cost slices into the same buffer — no allocations on access.

Reusable hot-loop address construction

When iterating over many devices (e.g., during group stanza preparation or session resolution), allocating a fresh ProtocolAddress per device is wasteful. The JidExt trait provides reset_protocol_address() to rewrite a pre-allocated address in place, and make_reusable_protocol_address() creates the initial buffer:
Usage in group stanza preparation:
This pattern eliminates one String allocation per device in the loop. For a group with 100 participant devices, that saves 100 heap allocations on the send path. The pre-allocated capacity of 64 bytes covers all known WhatsApp address formats without reallocation.
Use to_protocol_address() for one-shot address construction (e.g., cache keys, single lookups). Use make_reusable_protocol_address() + reset_protocol_address() when iterating over multiple JIDs in a tight loop.
Location: wacore/libsignal/src/core/address.rs, wacore/src/types/jid.rs, wacore/src/send.rs

Zero-Allocation JID Deduplication

Group stanza preparation needs to deduplicate participant JIDs at two stages: before device resolution (by user identity) and after LID conversion (by device identity). Two utility functions in wacore/src/types/jid.rs handle this with in-place sorted dedup instead of HashSet allocations:
Both use sort_unstable_by followed by dedup_by, comparing JID fields directly without allocating intermediate strings or hash sets. This is more efficient than the HashSet<(String, String)> approach because:
  • No per-JID String::clone() for hash keys
  • No HashSet allocation or hashing overhead
  • Stable dedup order (sorted) instead of hash-dependent iteration
Usage in group sends (wacore/src/send.rs):
Location: wacore/src/types/jid.rs:33-51

Take/Restore Pattern

Avoids cloning session states during decryption attempts:
Location: wacore/libsignal/src/protocol/session_cipher.rs:495-564

Buffer Reuse

Thread-local buffers eliminate per-message allocations:
Location: wacore/libsignal/src/protocol/session_cipher.rs:20-54

Public API

The client.signal() accessor exposes low-level Signal protocol operations for direct use. This includes 1:1 and group encryption/decryption, session validation, session deletion, participant node creation, and device resolution. See Signal API reference for full method documentation and examples.

References