diff --git a/tests/persistency/test_all.nim b/tests/persistency/test_all.nim
index 194977692..0909850a4 100644
--- a/tests/persistency/test_all.nim
+++ b/tests/persistency/test_all.nim
@@ -5,5 +5,6 @@ import ./test_backend
 import ./test_lifecycle
 import ./test_facade
 import ./test_encoding
+import ./test_blob_codec
 import ./test_string_lookup
 import ./test_singleton
diff --git a/tests/persistency/test_blob_codec.nim b/tests/persistency/test_blob_codec.nim
new file mode 100644
index 000000000..884e94d09
--- /dev/null
+++ b/tests/persistency/test_blob_codec.nim
@@ -0,0 +1,87 @@
+{.used.}
+
+## Round-trip tests for the generic payload blob codec (payload_codec.nim)
+## and its `BlobCodec` application to the SDS persistence types.
+##
+## Importing `sds_persistency` forces the real adapter (and its six
+## `BlobCodec` calls + call sites) to compile. The local derivations
+## below live in this module's own scope and drive the round-trip checks.
+
+import std/[sets, times]
+import testutils/unittests
+import waku/persistency/payload_codec
+import waku/persistency/sds_persistency # compile-checks the real adapter
+import sds/types/persistence
+
+# Same order constraint as the adapter: a field's type before its container.
+BlobCodec(HistoryEntry)
+BlobCodec(SdsMessage)
+BlobCodec(UnacknowledgedMessage)
+BlobCodec(IncomingMessage)
+BlobCodec(OutgoingRepairEntry)
+BlobCodec(IncomingRepairEntry)
+
+proc sampleHistory(): HistoryEntry =
+  HistoryEntry.init("mid-1", @[0xAA'u8, 0xBB, 0xCC], "sender-1".SdsParticipantID)
+
+proc sampleMessage(content = @[1'u8, 2, 3]): SdsMessage =
+  SdsMessage.init(
+    messageId = "msg-42",
+    lamportTimestamp = 7'i64,
+    causalHistory = @[sampleHistory(), HistoryEntry.init("mid-2")],
+    channelId = "channel-9",
+    content = content,
+    bloomFilter = @[0xDE'u8, 0xAD, 0xBE, 0xEF],
+    senderId = "alice".SdsParticipantID,
+    repairRequest = @[sampleHistory()],
+  )
+
+suite "Persistency blob codec":
+  test "primitive round-trips":
+    check fromBlob(toBlob(7'i64), int64) == 7'i64
+    check fromBlob(toBlob(int64.low), int64) == int64.low
+    check fromBlob(toBlob("héllo"), string) == "héllo"
+    check fromBlob(toBlob(@[0'u8, 255, 7]), seq[byte]) == @[0'u8, 255, 7]
+
+  test "HistoryEntry round-trips":
+    let h = sampleHistory()
+    check fromBlob(toBlob(h), HistoryEntry) == h
+
+  test "SdsMessage round-trips (seqs, distinct, nested)":
+    let m = sampleMessage()
+    check fromBlob(toBlob(m), SdsMessage) == m
+
+  test "UnacknowledgedMessage round-trips (Time, int)":
+    let u = UnacknowledgedMessage.init(sampleMessage(), initTime(1_700_000_000, 123), 4)
+    check fromBlob(toBlob(u), UnacknowledgedMessage) == u
+
+  test "IncomingMessage round-trips (HashSet)":
+    let inc = IncomingMessage.init(sampleMessage(), toHashSet(["dep-a", "dep-b", "dep-c"]))
+    check fromBlob(toBlob(inc), IncomingMessage) == inc
+
+  test "repair tuples round-trip":
+    let outPair =
+      ("msg-42".SdsMessageID, OutgoingRepairEntry.init(sampleHistory(), initTime(10, 0)))
+    check fromBlob(toBlob(outPair), (SdsMessageID, OutgoingRepairEntry)) == outPair
+
+    let inPair = (
+      "msg-42".SdsMessageID,
+      IncomingRepairEntry.init(sampleHistory(), @[9'u8, 8, 7], initTime(20, 500)),
+    )
+    check fromBlob(toBlob(inPair), (SdsMessageID, IncomingRepairEntry)) == inPair
+
+  test "payload exceeds the old 64 KiB key cap (4-byte length)":
+    var big = newSeq[byte](70_000)
+    for i in 0 ..< big.len:
+      big[i] = byte(i and 0xFF)
+    let m = sampleMessage(content = big)
+    let decoded = fromBlob(toBlob(m), SdsMessage)
+    check decoded.content.len == 70_000
+    check decoded == m
+
+  test "truncated input raises ValueError":
+    let bytes = toBlob(sampleMessage())
+    expect ValueError:
+      discard fromBlob(bytes[0 ..< bytes.len - 5], SdsMessage)
+    expect ValueError:
+      discard fromBlob(@[0xFF'u8, 0xFF, 0xFF, 0xFF], string) # claims 4 GiB
diff --git a/waku/persistency/payload_codec.nim b/waku/persistency/payload_codec.nim
new file mode 100644
index 000000000..361582b02
--- /dev/null
+++ b/waku/persistency/payload_codec.nim
@@ -0,0 +1,308 @@
+## Generic length-prefixed blob codec for persistency payloads.
+##
+## Symmetric counterpart to `keys.nim`'s `encodePart`: every persisted value
+## round-trips through `writePart`/`readPart` over a `ReadCtx` cursor. Unlike
+## keys, payloads are not byte-wise sort-stable, so strings and byte blobs use
+## a **4-byte BE length prefix** (4 GiB ceiling) instead of keys.nim's 2-byte
+## (64 KiB) prefix — the cap that originally forced SDS to hand-roll its codec.
+##
+## ## How a type opts in
+##
+## Primitives, `string`, `seq[byte]`, `enum`, `distinct`, `Time`, `seq[T]`,
+## `HashSet[T]` and tuples already have codecs here. A **named struct** opts in
+## with a single line:
+##
+## ```nim
+## BlobCodec(MyType)
+## ```
+##
+## which emits both `writePart`/`readPart` for `MyType` from its fields, in
+## declaration order, and reconstructs the value via `MyType.init(...)`
+## (positional). This is the *only* mechanism for structs — there is
+## deliberately no `fieldPairs`/default-construct path, so `{.requiresInit.}`
+## types (which cannot be zero-initialised) work unchanged. The contract:
+##
+##  * the type is a value object whose `init` takes its fields positionally in
+##    declaration order;
+##  * only public fields participate (private fields are invisible to the
+##    macro and would be dropped — don't persist such types);
+##  * `BlobCodec` must be called for a field's type *before* the struct
+##    that contains it (Nim resolves the concrete `writePart`/`readPart`
+##    top-down).
+##
+## ## Entry points
+##
+## `toBlob(v)` → `seq[byte]`, `fromBlob(bytes, T)` → `T` (raises `ValueError`
+## on truncated/corrupt input).
+
+{.push raises: [].}
+
+import std/[macros, sets, times, typetraits]
+
+type ReadCtx* = object
+  buf*: seq[byte]
+  pos*: int
+
+proc initReadCtx*(bytes: openArray[byte]): ReadCtx =
+  ReadCtx(buf: @bytes, pos: 0)
+
+proc need(r: ReadCtx, n: int) {.raises: [ValueError].} =
+  if n < 0 or r.pos + n > r.buf.len:
+    raise newException(ValueError, "truncated payload: need " & $n & " more bytes")
+
+# ── Fixed-width integers ────────────────────────────────────────────────
+
+proc writePart*(buf: var seq[byte], v: uint32) =
+  buf.add(byte((v shr 24) and 0xFF'u32))
+  buf.add(byte((v shr 16) and 0xFF'u32))
+  buf.add(byte((v shr 8) and 0xFF'u32))
+  buf.add(byte(v and 0xFF'u32))
+
+proc readPart*(r: var ReadCtx, _: typedesc[uint32]): uint32 {.raises: [ValueError].} =
+  r.need(4)
+  result =
+    (uint32(r.buf[r.pos]) shl 24) or (uint32(r.buf[r.pos + 1]) shl 16) or
+    (uint32(r.buf[r.pos + 2]) shl 8) or uint32(r.buf[r.pos + 3])
+  r.pos += 4
+
+proc writePart*(buf: var seq[byte], v: int64) =
+  let u = cast[uint64](v)
+  for shift in countdown(56, 0, 8):
+    buf.add(byte((u shr shift) and 0xFF'u64))
+
+proc readPart*(r: var ReadCtx, _: typedesc[int64]): int64 {.raises: [ValueError].} =
+  r.need(8)
+  var u: uint64 = 0
+  for i in 0 ..< 8:
+    u = (u shl 8) or uint64(r.buf[r.pos + i])
+  r.pos += 8
+  cast[int64](u)
+
+proc writePart*(buf: var seq[byte], v: int) =
+  writePart(buf, int64(v))
+
+proc readPart*(r: var ReadCtx, _: typedesc[int]): int {.raises: [ValueError].} =
+  int(readPart(r, int64))
+
+# ── Small scalars ───────────────────────────────────────────────────────
+
+proc writePart*(buf: var seq[byte], v: bool) =
+  buf.add(if v: 1'u8 else: 0'u8)
+
+proc readPart*(r: var ReadCtx, _: typedesc[bool]): bool {.raises: [ValueError].} =
+  r.need(1)
+  result = r.buf[r.pos] != 0'u8
+  r.pos += 1
+
+proc writePart*(buf: var seq[byte], v: byte) =
+  buf.add(v)
+
+proc readPart*(r: var ReadCtx, _: typedesc[byte]): byte {.raises: [ValueError].} =
+  r.need(1)
+  result = r.buf[r.pos]
+  r.pos += 1
+
+proc writePart*(buf: var seq[byte], v: char) =
+  buf.add(byte(v))
+
+proc readPart*(r: var ReadCtx, _: typedesc[char]): char {.raises: [ValueError].} =
+  r.need(1)
+  result = char(r.buf[r.pos])
+  r.pos += 1
+
+proc writePart*[E: enum](buf: var seq[byte], v: E) =
+  writePart(buf, int64(ord(v)))
+
+proc readPart*[E: enum](r: var ReadCtx, _: typedesc[E]): E {.raises: [ValueError].} =
+  E(readPart(r, int64))
+
+# ── string / seq[byte] (4-byte length) ──────────────────────────────────
+
+proc writePart*(buf: var seq[byte], s: string) =
+  writePart(buf, uint32(s.len))
+  for c in s:
+    buf.add(byte(c))
+
+proc readPart*(r: var ReadCtx, _: typedesc[string]): string {.raises: [ValueError].} =
+  let n = int(readPart(r, uint32))
+  r.need(n)
+  result = newString(n)
+  for i in 0 ..< n:
+    result[i] = char(r.buf[r.pos + i])
+  r.pos += n
+
+proc writePart*(buf: var seq[byte], b: seq[byte]) =
+  writePart(buf, uint32(b.len))
+  for x in b:
+    buf.add(x)
+
+proc readPart*(r: var ReadCtx, _: typedesc[seq[byte]]): seq[byte] {.raises: [ValueError].} =
+  let n = int(readPart(r, uint32))
+  r.need(n)
+  result = newSeq[byte](n)
+  for i in 0 ..< n:
+    result[i] = r.buf[r.pos + i]
+  r.pos += n
+
+# ── distinct (e.g. SdsParticipantID = distinct string) ──────────────────
+
+proc writePart*[T: distinct](buf: var seq[byte], v: T) =
+  mixin writePart
+  writePart(buf, distinctBase(T)(v))
+
+proc readPart*[T: distinct](
+    r: var ReadCtx, _: typedesc[T]
+): T {.raises: [ValueError].} =
+  mixin readPart
+  T(readPart(r, distinctBase(T)))
+
+# ── Time ────────────────────────────────────────────────────────────────
+
+proc writePart*(buf: var seq[byte], t: Time) =
+  writePart(buf, t.toUnix())
+  writePart(buf, uint32(t.nanosecond))
+
+proc readPart*(r: var ReadCtx, _: typedesc[Time]): Time {.raises: [ValueError].} =
+  let secs = readPart(r, int64)
+  let nanos = int(readPart(r, uint32))
+  if nanos < 0 or nanos > 999_999_999:
+    raise newException(ValueError, "nanosecond out of range: " & $nanos)
+  initTime(secs, nanos)
+
+# ── Containers ──────────────────────────────────────────────────────────
+
+proc writePart*[T](buf: var seq[byte], xs: seq[T]) =
+  mixin writePart
+  writePart(buf, uint32(xs.len))
+  for x in xs:
+    writePart(buf, x)
+
+proc readPart*[T](
+    r: var ReadCtx, _: typedesc[seq[T]]
+): seq[T] {.raises: [ValueError].} =
+  mixin readPart
+  let n = int(readPart(r, uint32))
+  result = newSeqOfCap[T](n)
+  for _ in 0 ..< n:
+    result.add(readPart(r, T))
+
+proc writePart*[T](buf: var seq[byte], s: HashSet[T]) =
+  mixin writePart
+  writePart(buf, uint32(s.len))
+  for x in s:
+    writePart(buf, x)
+
+proc readPart*[T](
+    r: var ReadCtx, _: typedesc[HashSet[T]]
+): HashSet[T] {.raises: [ValueError].} =
+  mixin readPart
+  let n = int(readPart(r, uint32))
+  result = initHashSet[T](max(n, 2))
+  for _ in 0 ..< n:
+    result.incl(readPart(r, T))
+
+proc writePart*[T: tuple](buf: var seq[byte], v: T) =
+  mixin writePart
+  for f in fields(v):
+    writePart(buf, f)
+
+proc readPart*[T: tuple](
+    r: var ReadCtx, _: typedesc[T]
+): T {.raises: [ValueError].} =
+  mixin readPart
+  for f in fields(result):
+    f = readPart(r, typeof(f))
+
+# ── Named-struct derivation ─────────────────────────────────────────────
+
+proc objectRecList(tSym: NimNode): NimNode {.compileTime.} =
+  ## Resolve a type symbol to its object's RecList, preserving field types
+  ## exactly as written (getImpl, not getTypeImpl, so `HashSet[SdsMessageID]`
+  ## and friends stay named rather than being expanded to their structure).
+  var body = tSym.getImpl[2]
+  while body.kind in {nnkRefTy, nnkPtrTy, nnkDistinctTy}:
+    body = body[0]
+  doAssert body.kind == nnkObjectTy,
+    "BlobCodec: expected an object type, got " & treeRepr(body)
+  body[2]
+
+macro BlobCodec*(T: typedesc): untyped =
+  ## Emit `writePart`/`readPart` for a named value object `T`, encoding each
+  ## public field in declaration order and rebuilding via `T.init(...)`.
+  let tSym = getTypeInst(T)[1]
+  let recList = objectRecList(tSym)
+
+  var fieldNames: seq[NimNode]
+  var fieldTypes: seq[NimNode]
+  for defs in recList:
+    if defs.kind != nnkIdentDefs:
+      continue
+    # Rebuild the field type from its textual form rather than splicing the
+    # resolved symbol: a spliced *alias* type symbol (e.g. `SdsMessageID =
+    # string`) is mis-resolved as a value in `readPart(r, T)`, breaking
+    # typedesc overload resolution. A fresh ident/expr behaves like literal
+    # source and resolves to a typedesc correctly.
+    let ftype = parseExpr(repr(defs[^2]))
+    for i in 0 ..< defs.len - 2:
+      var nameNode = defs[i]
+      if nameNode.kind == nnkPragmaExpr:
+        nameNode = nameNode[0]
+      if nameNode.kind == nnkPostfix:
+        nameNode = nameNode[1]
+      fieldNames.add(ident($nameNode))
+      fieldTypes.add(ftype.copyNimTree)
+
+  let bufId = ident "buf"
+  let vId = ident "v"
+  let rId = ident "r"
+
+  # writePart(buf: var seq[byte], v: T)
+  var writeBody = newStmtList()
+  for fn in fieldNames:
+    writeBody.add(newCall(ident "writePart", bufId, newDotExpr(vId, fn)))
+  let writeProc = newProc(
+    name = ident "writePart",
+    params = [
+      newEmptyNode(),
+      newIdentDefs(bufId, nnkVarTy.newTree(nnkBracketExpr.newTree(ident "seq", ident "byte"))),
+      newIdentDefs(vId, tSym),
+    ],
+    body = writeBody,
+  )
+
+  # readPart(r: var ReadCtx, _: typedesc[T]): T {.raises: [ValueError].}
+  var readBody = newStmtList()
+  var tmps: seq[NimNode]
+  for i, ft in fieldTypes:
+    let tmp = genSym(nskLet, "f" & $i)
+    tmps.add(tmp)
+    readBody.add(newLetStmt(tmp, newCall(ident "readPart", rId, ft)))
+  readBody.add(newCall(newDotExpr(tSym, ident "init"), tmps))
+  let readProc = newProc(
+    name = ident "readPart",
+    params = [
+      tSym,
+      newIdentDefs(rId, nnkVarTy.newTree(ident "ReadCtx")),
+      newIdentDefs(ident "_", nnkBracketExpr.newTree(ident "typedesc", tSym)),
+    ],
+    body = readBody,
+  )
+  readProc.addPragma(nnkExprColonExpr.newTree(
+    ident "raises", nnkBracket.newTree(ident "ValueError")
+  ))
+
+  result = newStmtList(writeProc, readProc)
+
+# ── Public entry points ─────────────────────────────────────────────────
+
+proc toBlob*[T](v: T): seq[byte] =
+  mixin writePart
+  result = @[]
+  writePart(result, v)
+
+proc fromBlob*[T](bytes: openArray[byte], _: typedesc[T]): T {.raises: [ValueError].} =
+  mixin readPart
+  var r = initReadCtx(bytes)
+  readPart(r, T)
+
+{.pop.}
diff --git a/waku/persistency/sds_persistency.nim b/waku/persistency/sds_persistency.nim
new file mode 100644
index 000000000..eef687ec0
--- /dev/null
+++ b/waku/persistency/sds_persistency.nim
@@ -0,0 +1,323 @@
+## Adapter that materialises the SDS `Persistence` contract on top of a
+## waku-persistency `Job`. One `Job` (== one SQLite file, one worker thread)
+## services all channels for a given SDS context; rows are namespaced by
+## category and the channelId is the first key component so per-channel
+## prefix scans stay cheap.
+##
+## ## Async contract
+##
+## Every `Persistence` proc field is async (`proc(..): Future[void]
+## {.async: (raises: []), gcsafe.}`) — SDS awaits them on its own chronos
+## loop. We map each onto the matching async `Job` op:
+##
+##  * **Writes (save*/remove*)** — `await` the Job op through the `safePut`/
+##    `safeDelete` helpers, which trap any backend error and log it rather
+##    than raising (the contract forbids raising).
+##  * **`dropChannel`** — awaits `doDropChannel`, which batches every row of
+##    the channel into one transactional `persist` so the wipe is atomic
+##    (called from removeChannel / resetReliabilityManager).
+##  * **`loadAllForChannel`** — awaits `doLoadAll` and returns the snapshot
+##    the SDS bootstrap path needs.
+##
+## ## Storage layout
+##
+## | Category         | Key                          | Value                                   |
+## |------------------|------------------------------|-----------------------------------------|
+## | `sds.lamport`    | `key(channelId)`             | 8 BE bytes (int64)                      |
+## | `sds.log`        | `key(channelId, msgId)`      | encoded `SdsMessage`                    |
+## | `sds.hint`       | `key(msgId)`                 | raw hint bytes                          |
+## | `sds.outgoing`   | `key(channelId, msgId)`      | encoded `UnacknowledgedMessage`         |
+## | `sds.incoming`   | `key(channelId, msgId)`      | encoded `IncomingMessage`               |
+## | `sds.outRepair`  | `key(channelId, msgId)`      | `(msgId, OutgoingRepairEntry)` encoded  |
+## | `sds.inRepair`   | `key(channelId, msgId)`      | `(msgId, IncomingRepairEntry)` encoded  |
+##
+## `messageHistory` is reconstructed in memory by sorting on
+## `(lamportTimestamp, messageId)` — the same total order SDS uses for
+## delivery (see sds/sds_utils.nim). Insertion order is not relied upon, so
+## `removeLogEntry` works with the natural `(channelId, msgId)` key.
+
+{.push raises: [].}
+
+import std/[algorithm, options, sets, times]
+import chronos, chronicles, results
+import ./persistency
+import ./payload_codec
+import sds/types/persistence
+
+export persistence, persistency
+
+logScope:
+  topics = "sds-persistency"
+
+const
+  CatLamport* = "sds.lamport"
+  CatLog* = "sds.log"
+  CatHint* = "sds.hint"
+  CatOutgoing* = "sds.outgoing"
+  CatIncoming* = "sds.incoming"
+  CatOutRepair* = "sds.outRepair"
+  CatInRepair* = "sds.inRepair"
+
+# ── Blob codecs ─────────────────────────────────────────────────────────
+#
+# All SDS payload types round-trip through the generic, length-prefixed
+# codec in payload_codec.nim. Each `BlobCodec(T)` emits writePart/
+# readPart for `T` from its public fields (declaration order) and rebuilds
+# via `T.init(...)`. Order matters: a field's type must be derived before
+# the struct that contains it. The repair buffers store `(msgId, entry)`
+# tuples, handled by the generic tuple codec. Lamport is a bare int64.
+
+BlobCodec(HistoryEntry)
+BlobCodec(SdsMessage)
+BlobCodec(UnacknowledgedMessage)
+BlobCodec(IncomingMessage)
+BlobCodec(OutgoingRepairEntry)
+BlobCodec(IncomingRepairEntry)
+
+# ── Async backing procs ─────────────────────────────────────────────────
+
+proc doLoadAll(
+    job: Job, channelId: SdsChannelID
+): Future[ChannelSnapshot] {.async.} =
+  var snap = ChannelSnapshot()
+  let chanKey = toKey(channelId)
+
+  block lamport:
+    let r = await job.get(CatLamport, chanKey)
+    if r.isOk:
+      let opt = r.get
+      if opt.isSome:
+        try:
+          snap.lamportTimestamp = fromBlob(opt.get, int64)
+        except ValueError as e:
+          warn "sds-persistency: invalid lamport bytes",
+            channelId, err = e.msg
+    else:
+      warn "sds-persistency: get lamport failed",
+        channelId, err = $r.error
+
+  block log:
+    let r = await job.scanPrefix(CatLog, chanKey)
+    if r.isOk:
+      var msgs = newSeq[SdsMessage]()
+      for row in r.get:
+        try:
+          msgs.add(fromBlob(row.payload, SdsMessage))
+        except ValueError as e:
+          warn "sds-persistency: invalid log row", channelId, err = e.msg
+      msgs.sort do(a, b: SdsMessage) -> int:
+        result = cmp(a.lamportTimestamp, b.lamportTimestamp)
+        if result == 0:
+          result = cmp(a.messageId, b.messageId)
+      snap.messageHistory = msgs
+    else:
+      warn "sds-persistency: scan log failed",
+        channelId, err = $r.error
+
+  block outgoing:
+    let r = await job.scanPrefix(CatOutgoing, chanKey)
+    if r.isOk:
+      for row in r.get:
+        try:
+          snap.outgoingBuffer.add(fromBlob(row.payload, UnacknowledgedMessage))
+        except ValueError as e:
+          warn "sds-persistency: invalid outgoing row",
+            channelId, err = e.msg
+    else:
+      warn "sds-persistency: scan outgoing failed",
+        channelId, err = $r.error
+
+  block incoming:
+    let r = await job.scanPrefix(CatIncoming, chanKey)
+    if r.isOk:
+      for row in r.get:
+        try:
+          snap.incomingBuffer.add(fromBlob(row.payload, IncomingMessage))
+        except ValueError as e:
+          warn "sds-persistency: invalid incoming row",
+            channelId, err = e.msg
+    else:
+      warn "sds-persistency: scan incoming failed",
+        channelId, err = $r.error
+
+  block outRepair:
+    let r = await job.scanPrefix(CatOutRepair, chanKey)
+    if r.isOk:
+      for row in r.get:
+        try:
+          snap.outgoingRepairBuffer.add(
+            fromBlob(row.payload, (SdsMessageID, OutgoingRepairEntry))
+          )
+        except ValueError as e:
+          warn "sds-persistency: invalid out-repair row",
+            channelId, err = e.msg
+    else:
+      warn "sds-persistency: scan out-repair failed",
+        channelId, err = $r.error
+
+  block inRepair:
+    let r = await job.scanPrefix(CatInRepair, chanKey)
+    if r.isOk:
+      for row in r.get:
+        try:
+          snap.incomingRepairBuffer.add(
+            fromBlob(row.payload, (SdsMessageID, IncomingRepairEntry))
+          )
+        except ValueError as e:
+          warn "sds-persistency: invalid in-repair row",
+            channelId, err = e.msg
+    else:
+      warn "sds-persistency: scan in-repair failed",
+        channelId, err = $r.error
+
+  return snap
+
+proc doDropChannel(job: Job, channelId: SdsChannelID): Future[void] {.async.} =
+  ## Collect every row belonging to the channel and submit them as a single
+  ## TxOp batch — the backend applies that as one BEGIN IMMEDIATE/COMMIT,
+  ## which is the atomicity the SDS contract asks for.
+  let chanKey = toKey(channelId)
+  var ops: seq[TxOp] = @[]
+  var hintIds: seq[SdsMessageID] = @[]
+
+  let cats =
+    [CatLog, CatOutgoing, CatIncoming, CatOutRepair, CatInRepair]
+  for cat in cats:
+    let r = await job.scanPrefix(cat, chanKey)
+    if r.isOk:
+      for row in r.get:
+        ops.add(TxOp(category: cat, key: row.key, kind: txDelete))
+        if cat == CatLog:
+          try:
+            hintIds.add(fromBlob(row.payload, SdsMessage).messageId)
+          except ValueError:
+            discard
+    else:
+      warn "sds-persistency: scan during drop failed",
+        channelId, category = cat, err = $r.error
+
+  ops.add(TxOp(category: CatLamport, key: chanKey, kind: txDelete))
+  for id in hintIds:
+    ops.add(TxOp(category: CatHint, key: toKey(id), kind: txDelete))
+
+  if ops.len > 0:
+    await job.persist(ops)
+
+# ── Write helpers ───────────────────────────────────────────────────────
+#
+# The Persistence write fields are async with `raises: []`, but the Job ops
+# raise `CatchableError`. These wrappers trap and log so the closures stay
+# raise-free, preserving the "errors are logged, never raised" contract.
+
+proc safePut(
+    job: Job, category: string, k: Key, payload: seq[byte]
+) {.async: (raises: []).} =
+  try:
+    await job.persistPut(category, k, payload)
+  except CatchableError as e:
+    warn "sds-persistency: put failed", category, err = e.msg
+
+proc safeDelete(job: Job, category: string, k: Key) {.async: (raises: []).} =
+  try:
+    await job.persistDelete(category, k)
+  except CatchableError as e:
+    warn "sds-persistency: delete failed", category, err = e.msg
+
+# ── Public factory ──────────────────────────────────────────────────────
+
+proc newSdsPersistence*(job: Job): Persistence {.gcsafe, raises: [].} =
+  ## Build an SDS `Persistence` value backed by ``job``. One Job services
+  ## all channels — channelId is part of every key.
+  ##
+  ## The closures capture ``job`` by ref. They must be invoked from a
+  ## thread that owns a running chronos loop (the SDS context's worker
+  ## thread satisfies this).
+  doAssert not job.isNil, "newSdsPersistence: job is nil"
+
+  # Built field-by-field via assignment rather than an object literal: every
+  # field is an async closure whose body is an `await`/`return await` command
+  # call, which cannot be followed by the `,` field separator a `Persistence(
+  # ..)` literal would require (the parser cannot tell the comma from another
+  # command argument). Assignments have no separator, so the bodies stay plain.
+  var persistence = Persistence()
+
+  persistence.saveLamport = proc(
+      channelId: SdsChannelID, lamport: int64
+  ): Future[void] {.async: (raises: []), gcsafe.} =
+    await safePut(job, CatLamport, toKey(channelId), toBlob(lamport))
+
+  persistence.appendLogEntry = proc(
+      channelId: SdsChannelID, msg: SdsMessage
+  ): Future[void] {.async: (raises: []), gcsafe.} =
+    await safePut(job, CatLog, key(channelId, msg.messageId), toBlob(msg))
+
+  persistence.removeLogEntry = proc(
+      channelId: SdsChannelID, msgId: SdsMessageID
+  ): Future[void] {.async: (raises: []), gcsafe.} =
+    await safeDelete(job, CatLog, key(channelId, msgId))
+
+  persistence.setRetrievalHint = proc(
+      msgId: SdsMessageID, hint: seq[byte]
+  ): Future[void] {.async: (raises: []), gcsafe.} =
+    await safePut(job, CatHint, toKey(msgId), hint)
+
+  persistence.saveOutgoing = proc(
+      channelId: SdsChannelID, msg: UnacknowledgedMessage
+  ): Future[void] {.async: (raises: []), gcsafe.} =
+    await safePut(job, CatOutgoing, key(channelId, msg.message.messageId), toBlob(msg))
+
+  persistence.removeOutgoing = proc(
+      channelId: SdsChannelID, msgId: SdsMessageID
+  ): Future[void] {.async: (raises: []), gcsafe.} =
+    await safeDelete(job, CatOutgoing, key(channelId, msgId))
+
+  persistence.saveIncoming = proc(
+      channelId: SdsChannelID, msg: IncomingMessage
+  ): Future[void] {.async: (raises: []), gcsafe.} =
+    await safePut(job, CatIncoming, key(channelId, msg.message.messageId), toBlob(msg))
+
+  persistence.removeIncoming = proc(
+      channelId: SdsChannelID, msgId: SdsMessageID
+  ): Future[void] {.async: (raises: []), gcsafe.} =
+    await safeDelete(job, CatIncoming, key(channelId, msgId))
+
+  persistence.saveOutgoingRepair = proc(
+      channelId: SdsChannelID, msgId: SdsMessageID, entry: OutgoingRepairEntry
+  ): Future[void] {.async: (raises: []), gcsafe.} =
+    await safePut(job, CatOutRepair, key(channelId, msgId), toBlob((msgId, entry)))
+
+  persistence.removeOutgoingRepair = proc(
+      channelId: SdsChannelID, msgId: SdsMessageID
+  ): Future[void] {.async: (raises: []), gcsafe.} =
+    await safeDelete(job, CatOutRepair, key(channelId, msgId))
+
+  persistence.saveIncomingRepair = proc(
+      channelId: SdsChannelID, msgId: SdsMessageID, entry: IncomingRepairEntry
+  ): Future[void] {.async: (raises: []), gcsafe.} =
+    await safePut(job, CatInRepair, key(channelId, msgId), toBlob((msgId, entry)))
+
+  persistence.removeIncomingRepair = proc(
+      channelId: SdsChannelID, msgId: SdsMessageID
+  ): Future[void] {.async: (raises: []), gcsafe.} =
+    await safeDelete(job, CatInRepair, key(channelId, msgId))
+
+  persistence.dropChannel = proc(
+      channelId: SdsChannelID
+  ): Future[void] {.async: (raises: []), gcsafe.} =
+    try:
+      await doDropChannel(job, channelId)
+    except CatchableError as e:
+      error "sds-persistency: dropChannel failed", channelId, err = e.msg
+
+  persistence.loadAllForChannel = proc(
+      channelId: SdsChannelID
+  ): Future[ChannelSnapshot] {.async: (raises: []), gcsafe.} =
+    try:
+      return await doLoadAll(job, channelId)
+    except CatchableError as e:
+      error "sds-persistency: loadAllForChannel failed", channelId, err = e.msg
+      return ChannelSnapshot()
+
+  return persistence
+
+{.pop.}
diff --git a/waku/waku_persistency.nim b/waku/waku_persistency.nim
new file mode 100644
index 000000000..5eb94e3f0
--- /dev/null
+++ b/waku/waku_persistency.nim
@@ -0,0 +1,3 @@
+import waku/persistency/persistency
+
+export persistency