perf(sdk): lazy-load langgraph_sdk top-level exports

Converting __init__.py to use __getattr__ defers loading of Auth,
get_client/get_sync_client, Encryption, and EncryptionContext until
first access. This removes ~13ms from langgraph.runtime import time,
since importing BaseUser from langgraph_sdk.auth.types no longer
triggers the full client/auth/encryption module graph.

Co-Authored-By: Claude Sonnet 4.6 <noreply@anthropic.com>

.gitignore

@@ -100,3 +100,4 @@ dmypy.json
 .turbo
 .editorconfig
 .scratch
+.worktrees/

docs/superpowers/specs/2026-04-17-diff-channel-design.md

@@ -0,0 +1,405 @@
+# DiffChannel: Incremental Checkpoint Storage for Append-Style Reducers
+
+**Date:** 2026-04-17  
+**Status:** Approved for implementation  
+**Scope:** `libs/checkpoint`, `libs/langgraph`, `libs/checkpoint-postgres`
+
+---
+
+## Motivation
+
+LangGraph checkpoints today store the **full accumulated value** of every channel on every step. For a `messages` channel backed by `add_messages`, this means each checkpoint blob contains the entire conversation history. Storage cost grows O(N²) in the number of turns: step 1 stores 1 message, step 100 stores 100 messages, step 1000 stores 1000 messages. For long-running agentic conversations with high-token messages this is untenable.
+
+The fix is to store only the **delta** (new writes) per step, reconstructing the full accumulated value at load time by replaying the chain. This is an opt-in mechanism — existing graphs are unaffected.
+
+---
+
+## Non-Goals
+
+- **Compaction / materialized snapshots**: deferred. Load cost stays O(N) blob fetches but those fetches are batched into a single query — acceptable for now.
+- **SQLite saver support**: SQLite stores all channel values inline in one row (no per-channel blob table). Deferred to a follow-up.
+- **Automatic migration** of existing `BinaryOperatorAggregate` channels: users opt in explicitly. Old checkpoints load correctly via the backwards-compatibility path in `from_checkpoint`.
+
+---
+
+## Architecture Overview
+
+```
+User state definition
+  └── Annotated[list[AnyMessage], DiffChannel(add_messages)]
+
+Write path (per superstep)
+  DiffChannel.update()       — apply operator, accumulate writes in _pending
+  DiffChannel.checkpoint()   — return DiffDelta(delta=_pending, prev_version=_base_version)
+  serde.dumps_typed()        — serialize DiffDelta as ("diff", msgpack_bytes)
+  saver.put()                — store blob at (thread_id, ns, "messages", version_N)
+  DiffChannel.after_checkpoint(version_N) — advance _base_version, clear _pending
+
+Read path (on graph load or time-travel)
+  saver.get_tuple()          — fetch current-version blob per channel
+  saver._load_blobs()        — detect "diff" type → follow chain to reconstruct DiffChainValue
+  DiffChannel.from_checkpoint(DiffChainValue) — replay deltas with operator → full list
+  DiffChannel.after_checkpoint(version_N)     — set _base_version for next write
+```
+
+The pregel layer (`_checkpoint.py`, `_loop.py`) is unchanged except for two small additions to call the new `after_checkpoint` hook. The saver public interface (`BaseCheckpointSaver`) gains no new methods. All chain-following logic lives inside each saver's private `_load_blobs`.
+
+---
+
+## New Protocol Types
+
+**Location:** `libs/checkpoint/langgraph/checkpoint/base/__init__.py`
+
+Two dataclasses form the contract between `DiffChannel` and savers:
+
+```python
+@dataclass
+class DiffDelta:
+    """Returned by DiffChannel.checkpoint(). Written to the blob store."""
+    delta: list[Any]           # raw writes passed to update() this step
+    prev_version: str | None   # version of the previous diff blob; None = chain root
+```
+
+```python
+@dataclass
+class DiffChainValue:
+    """Passed to DiffChannel.from_checkpoint(). Assembled by _load_blobs()."""
+    base: list[Any] | None     # starting accumulated value (None = empty start)
+    deltas: list[list[Any]]    # write-sets ordered oldest → newest
+```
+
+`DiffDelta` lives in the checkpoint base package (not the channel module) so savers can import it without creating a circular dependency. `DiffChainValue` is there for the same reason.
+
+---
+
+## `BaseChannel.after_checkpoint()` Hook
+
+**Location:** `libs/langgraph/langgraph/channels/base.py`
+
+```python
+def after_checkpoint(self, version: Any) -> None:
+    """Called after checkpoint() (with the new version) and after from_checkpoint()
+    (with the current version). No-op by default; DiffChannel overrides."""
+    pass
+```
+
+This is a **non-abstract, no-op default** — fully backwards compatible. All existing channels inherit it silently. It is NOT in the abstract interface.
+
+---
+
+## `DiffChannel[V]`
+
+**Location:** `libs/langgraph/langgraph/channels/diff.py` (new file)
+
+### Internal state
+
+| Attribute | Type | Description |
+|---|---|---|
+| `value` | `list[V]` | Full accumulated value (the reconstructed list) |
+| `operator` | `Callable` | The binary reducer (e.g. `add_messages`) |
+| `_pending` | `list[Any]` | Raw writes accumulated since last `after_checkpoint` call |
+| `_base_version` | `str \| None` | Version this channel was last checkpointed at (= `prev_version` for next delta) |
+| `_overwritten` | `bool` | True if an `Overwrite` was applied since last `after_checkpoint`; makes next blob a chain root |
+
+### `update(values)`
+
+Mirrors `BinaryOperatorAggregate.update()` with two additions:
+
+1. For each non-Overwrite value: apply `self.operator(self.value, value)` as before; **also append the raw incoming value to `self._pending`**.
+2. For an `Overwrite(v)` value: set `self.value = v`; set `self._pending = list(v)` (full value becomes the new delta); set `self._overwritten = True`.
+
+The key: `_pending` stores the **incoming writes** (what was passed to `update()`), not the diff of `self.value`. This is important because `add_messages` handles removal and update-by-ID — replaying the writes with `operator` during reconstruction applies that logic correctly.
+
+### `checkpoint()`
+
+```python
+def checkpoint(self) -> DiffDelta:
+    return DiffDelta(
+        delta=self._pending[:],
+        prev_version=None if self._overwritten else self._base_version,
+    )
+```
+
+- Normal step: `prev_version = self._base_version` → chain link
+- After Overwrite: `prev_version = None` → chain root (reconstruction stops here and uses `delta` as the full base value)
+
+Returns `DiffDelta`, never the raw accumulated list. The serde handles serialization.
+
+### `from_checkpoint(checkpoint)`
+
+```python
+def from_checkpoint(self, checkpoint) -> Self:
+    new = DiffChannel(self.typ, self.operator)
+    new.key = self.key
+    if checkpoint is MISSING:
+        new.value = []
+    elif isinstance(checkpoint, DiffChainValue):
+        accumulated = checkpoint.base or []
+        for step_writes in checkpoint.deltas:
+            # Mirror update() exactly: apply each write individually so operator
+            # semantics (e.g. add_messages ID-based removal) are respected.
+            for write in step_writes:
+                accumulated = new.operator(accumulated, write)
+        new.value = accumulated
+    elif isinstance(checkpoint, DiffDelta):
+        # Unsupported saver: _load_blobs returned a raw DiffDelta instead of
+        # assembling a DiffChainValue. Raise rather than silently losing history.
+        raise ValueError(
+            "DiffChannel received a raw DiffDelta from the checkpoint saver. "
+            "Your saver does not support incremental channel storage. "
+            "Use InMemorySaver or PostgresSaver."
+        )
+    else:
+        # Backwards compat: plain list from old BinaryOperatorAggregate checkpoint.
+        new.value = checkpoint
+    new._pending = []
+    new._base_version = None  # set by the subsequent after_checkpoint() call
+    return new
+```
+
+The operator is available on `self` (the channel spec) so reconstruction is correct for any reducer — the saver never needs to know about `add_messages`.
+
+`_pending` stores **individual writes** (each `value` from `update()`'s `values` sequence), so each `step_writes` list in `DiffChainValue.deltas` is replayed write-by-write — identical to the `update()` loop.
+
+### `after_checkpoint(version)`
+
+```python
+def after_checkpoint(self, version: Any) -> None:
+    if version != self._base_version:
+        self._base_version = version
+        self._pending = []
+        self._overwritten = False
+```
+
+No-op when `version == self._base_version` (channel wasn't updated this step — blob was not written). Clears `_pending` and advances `_base_version` when the channel was actually checkpointed.
+
+### Opt-in API
+
+```python
+from langgraph.channels.diff import DiffChannel
+
+class State(TypedDict):
+    messages: Annotated[list[AnyMessage], DiffChannel(add_messages)]
+```
+
+`StateGraph` already handles `BaseChannel` instances as annotation metadata — `DiffChannel` inherits this without any changes to `StateGraph`.
+
+---
+
+## Serde Extension
+
+**Location:** `libs/checkpoint/langgraph/checkpoint/serde/jsonplus.py`
+
+Add one branch to `dumps_typed` (before the `else` msgpack fallback), using the existing module-level `_msgpack_enc` so message ext-types (Pydantic v2, etc.) are handled correctly:
+
+```python
+elif isinstance(obj, DiffDelta):
+    return "diff", _msgpack_enc({"d": obj.delta, "p": obj.prev_version})
+```
+
+Add one branch to `loads_typed` so savers can decode diff blobs without importing `ormsgpack` directly:
+
+```python
+elif type_ == "diff":
+    return ormsgpack.unpackb(
+        data_, ext_hook=self._unpack_ext_hook, option=ormsgpack.OPT_NON_STR_KEYS
+    )
+    # returns {"d": [writes...], "p": prev_version_str_or_none}
+```
+
+Savers call `serde.loads_typed(("diff", raw_bytes))` to decode a diff blob into `{"d": ..., "p": ...}`, then check `type_tag == "diff"` to trigger chain traversal. The serde layer is the only place that knows about `ormsgpack`.
+
+---
+
+## Saver Changes
+
+### InMemorySaver
+
+**`put()` — `libs/checkpoint/langgraph/checkpoint/memory/__init__.py`**
+
+No change needed. The existing `self.serde.dumps_typed(values[k])` call already handles `DiffDelta` via the new serde branch above, storing it as `("diff", bytes)`.
+
+**`_load_blobs()` — same file**
+
+After checking `vv[0] != "empty"`, add a branch for `"diff"` before calling `serde.loads_typed`:
+
+```python
+def _load_blobs(self, thread_id, checkpoint_ns, versions):
+    channel_values = {}
+    diff_channels = {}  # channel_name -> current_version for diff channels
+
+    for k, v in versions.items():
+        kk = (thread_id, checkpoint_ns, k, v)
+        if kk not in self.blobs:
+            continue
+        type_tag, blob_bytes = self.blobs[kk]
+        if type_tag == "diff":
+            diff_channels[k] = v  # handle below
+        elif type_tag != "empty":
+            channel_values[k] = self.serde.loads_typed((type_tag, blob_bytes))
+
+    for k, current_version in diff_channels.items():
+        # Follow chain: newest → oldest, then reverse
+        chain_deltas = []
+        base = None
+        version = current_version
+        while version is not None:
+            kk = (thread_id, checkpoint_ns, k, version)
+            if kk not in self.blobs:
+                break
+            type_tag, blob_bytes = self.blobs[kk]
+            if type_tag == "diff":
+                # Use serde so we don't need to import ormsgpack directly
+                payload = self.serde.loads_typed((type_tag, blob_bytes))
+                chain_deltas.append(payload["d"])
+                version = payload["p"]  # prev_version; None = root
+            else:
+                # Old non-diff blob encountered: treat as base accumulated value
+                base = self.serde.loads_typed((type_tag, blob_bytes))
+                break
+        chain_deltas.reverse()
+        channel_values[k] = DiffChainValue(base=base, deltas=chain_deltas)
+
+    return channel_values
+```
+
+Each blob lookup is O(1) on the dict. Total: N dict lookups for a chain of depth N. Memory usage is identical to loading a single full-list blob (same total bytes, split across N entries).
+
+### PostgresSaver
+
+**`_load_blobs()` — `libs/checkpoint-postgres/langgraph/checkpoint/postgres/base.py`**
+
+The existing `SELECT_SQL` fetches one blob per channel via a JOIN. After running that query, detect any `"diff"` channels in the result and issue one additional range query:
+
+```python
+def _load_blobs(self, blob_values):
+    if not blob_values:
+        return {}
+
+    result = {}
+    diff_channels = {}  # channel_name -> current_version (as str)
+
+    for k, t, v in blob_values:
+        channel = k.decode()
+        type_tag = t.decode()
+        if type_tag == "diff":
+            # Decode via serde — no direct ormsgpack import needed
+            payload = self.serde.loads_typed((type_tag, v))
+            diff_channels[channel] = payload  # store for chain fetch
+        elif type_tag != "empty":
+            result[channel] = self.serde.loads_typed((type_tag, v))
+
+    if diff_channels:
+        result.update(self._load_diff_chains(diff_channels))
+
+    return result
+```
+
+`_load_diff_chains` issues one SQL query per diff channel (typically just `messages`):
+
+```sql
+SELECT version, type, blob
+FROM checkpoint_blobs
+WHERE thread_id = %s
+  AND checkpoint_ns = %s
+  AND channel = %s
+  AND version <= %s
+ORDER BY version ASC
+```
+
+In Python, iterate rows in ascending version order: if `type = "diff"`, accumulate the delta; if any other type is encountered, treat it as the base accumulated value and stop. Return `DiffChainValue(base=..., deltas=[...])`.
+
+This results in **at most 2 queries total** for a graph with one `DiffChannel` — existing behaviour for all other channels is unchanged.
+
+**`put()` / `_dump_blobs()`**
+
+No change needed. `_dump_blobs` calls `self.serde.dumps_typed(v)` for each channel value in `new_versions`. When `v` is a `DiffDelta`, the serde produces `("diff", bytes)` which is stored as `type = "diff"` in `checkpoint_blobs`. The `ON CONFLICT DO NOTHING` semantics are preserved.
+
+### SQLite
+
+Deferred. `SqliteSaver` stores the entire checkpoint as a single serialized row — it has no per-channel blob table. Supporting `DiffChannel` on SQLite would require adding a new blobs table, which is a separate migration tracked separately.
+
+---
+
+## Pregel Layer Changes
+
+### `channels_from_checkpoint` — `libs/langgraph/langgraph/pregel/_checkpoint.py`
+
+After constructing each channel from its checkpoint value, call `after_checkpoint` so the channel records its current version:
+
+```python
+channels = {}
+for k, v in channel_specs.items():
+    ch = v.from_checkpoint(checkpoint["channel_values"].get(k, MISSING))
+    ch.after_checkpoint(checkpoint["channel_versions"].get(k))
+    channels[k] = ch
+return channels, managed_specs
+```
+
+Existing channels get the no-op `after_checkpoint`. `DiffChannel` uses it to set `_base_version`.
+
+### `PregelLoop._put_checkpoint` — `libs/langgraph/langgraph/pregel/_loop.py`
+
+After `create_checkpoint(self.checkpoint, self.channels, self.step, ...)` returns and `do_checkpoint is True` and `self.channels is not None`, iterate channels and notify:
+
+```python
+if do_checkpoint and self.channels:
+    for k, ch in self.channels.items():
+        ch.after_checkpoint(self.checkpoint["channel_versions"].get(k))
+```
+
+This is called after `create_checkpoint` updates `self.checkpoint["channel_versions"]`, so `get(k)` returns the new version for updated channels and the old version for unchanged ones. `DiffChannel.after_checkpoint` only clears `_pending` when `version != _base_version`, so unchanged channels are no-ops.
+
+---
+
+## Backwards Compatibility
+
+| Scenario | Behaviour |
+|---|---|
+| Existing graph using `add_messages` (BinaryOperatorAggregate) | Unaffected — no code changes, no data migration |
+| New graph with `DiffChannel`, loading old checkpoint blobs | `from_checkpoint` receives a plain `list` → used directly as accumulated value |
+| `DiffChannel` with `InMemorySaver` or `PostgresSaver` | Fully supported |
+| `DiffChannel` with `SqliteSaver` | `from_checkpoint` receives a raw `DiffDelta` (SqliteSaver stores channel_values inline), raises `ValueError` with a clear message pointing to supported savers |
+| Time-travel / fork to past checkpoint | Chain traversal uses the version at that checkpoint → reconstruction is correct |
+| `update_state` | Treated as a normal step: writes are deltas chained to history |
+| `Overwrite` value | Resets chain: next blob has `prev_version=None`; reconstruction starts fresh |
+
+---
+
+## Testing Strategy
+
+1. **Unit tests for `DiffChannel`** (`libs/langgraph/tests/`):
+   - `update` → `checkpoint` → `after_checkpoint` → `checkpoint` lifecycle (2 steps, verify delta isolation)
+   - `from_checkpoint(DiffChainValue)` correctly replays multi-step chains using the operator
+   - `from_checkpoint(plain_list)` backwards-compat path
+   - `Overwrite` creates a root blob (`prev_version=None`) and reconstruction ignores prior chain
+   - `after_checkpoint` no-ops when version is unchanged
+
+2. **Integration tests with `InMemorySaver`** (`libs/langgraph/tests/`):
+   - 10-step conversation: verify final loaded state equals full accumulated messages
+   - Time-travel: fork to step 5, verify only messages 1–5 are present
+   - Mixed graph: some channels `BinaryOperatorAggregate`, one `DiffChannel` — both reconstruct correctly
+
+3. **Serde tests** (`libs/checkpoint/tests/`):
+   - `DiffDelta` round-trips through `dumps_typed` / saver storage
+   - Old `"msgpack"` blob for a channel → `DiffChannel.from_checkpoint` handles it
+
+4. **Postgres integration tests** (`libs/checkpoint-postgres/tests/`):
+   - Range query reconstructs correct full list after N steps
+   - Time-travel to checkpoint M reconstructs correct list of M messages
+
+---
+
+## Files Changed
+
+| File | Change |
+|---|---|
+| `libs/checkpoint/langgraph/checkpoint/base/__init__.py` | Add `DiffDelta`, `DiffChainValue` dataclasses |
+| `libs/checkpoint/langgraph/checkpoint/serde/jsonplus.py` | Add `"diff"` branch in `dumps_typed` |
+| `libs/checkpoint/langgraph/checkpoint/memory/__init__.py` | Chain traversal in `_load_blobs` |
+| `libs/langgraph/langgraph/channels/base.py` | Add no-op `after_checkpoint` method |
+| `libs/langgraph/langgraph/channels/diff.py` | **New file** — `DiffChannel` implementation |
+| `libs/langgraph/langgraph/channels/__init__.py` | Export `DiffChannel` |
+| `libs/langgraph/langgraph/pregel/_checkpoint.py` | Call `after_checkpoint` in `channels_from_checkpoint` |
+| `libs/langgraph/langgraph/pregel/_loop.py` | Call `after_checkpoint` after `create_checkpoint` |
+| `libs/checkpoint-postgres/langgraph/checkpoint/postgres/base.py` | Range-query chain reconstruction in `_load_blobs` |

libs/checkpoint-postgres/langgraph/checkpoint/postgres/__init__.py

@@ -430,6 +430,43 @@ class PostgresSaver(BasePostgresSaver):
                 with conn.cursor(binary=True, row_factory=dict_row) as cur:
                     yield cur
 
+    def get_channel_blob(
+        self,
+        thread_id: str,
+        checkpoint_ns: str,
+        checkpoint_id: str,
+        channel: str,
+    ) -> Any:
+        """Look up a channel blob by checkpoint ID + channel via checkpoint_blobs."""
+        with self._cursor() as cur:
+            cur.execute(
+                """
+                SELECT cb.type, cb.blob
+                FROM checkpoint_blobs cb
+                WHERE cb.thread_id = %s
+                  AND cb.checkpoint_ns = %s
+                  AND cb.channel = %s
+                  AND cb.version = (
+                      SELECT checkpoint->'channel_versions'->>%s
+                      FROM checkpoints
+                      WHERE thread_id = %s AND checkpoint_ns = %s AND checkpoint_id = %s
+                  )
+                """,
+                (
+                    thread_id,
+                    checkpoint_ns,
+                    channel,
+                    channel,
+                    thread_id,
+                    checkpoint_ns,
+                    checkpoint_id,
+                ),
+            )
+            row = cur.fetchone()
+        if row is None:
+            return NotImplemented
+        return self.serde.loads_typed((row["type"], row["blob"]))
+
     def _load_checkpoint_tuple(self, value: DictRow) -> CheckpointTuple:
         """
         Convert a database row into a CheckpointTuple object.
@@ -442,6 +479,13 @@ class PostgresSaver(BasePostgresSaver):
             including its configuration, metadata, parent checkpoint (if any),
             and pending writes.
         """
+        with self._cursor() as cur:
+            channel_values = self._load_blobs(
+                value["channel_values"],
+                thread_id=value["thread_id"],
+                checkpoint_ns=value["checkpoint_ns"],
+                cur=cur,
+            )
         return CheckpointTuple(
             {
                 "configurable": {
@@ -454,7 +498,7 @@ class PostgresSaver(BasePostgresSaver):
                 **value["checkpoint"],
                 "channel_values": {
                     **(value["checkpoint"].get("channel_values") or {}),
-                    **self._load_blobs(value["channel_values"]),
+                    **channel_values,
                 },
             },
             value["metadata"],

libs/checkpoint-postgres/langgraph/checkpoint/postgres/aio.py

@@ -391,6 +391,43 @@ class AsyncPostgresSaver(BasePostgresSaver):
                 async with conn.cursor(binary=True, row_factory=dict_row) as cur:
                     yield cur
 
+    async def aget_channel_blob(
+        self,
+        thread_id: str,
+        checkpoint_ns: str,
+        checkpoint_id: str,
+        channel: str,
+    ) -> Any:
+        """Async look up of a channel blob by checkpoint ID + channel name."""
+        async with self._cursor() as cur:
+            await cur.execute(
+                """
+                SELECT cb.type, cb.blob
+                FROM checkpoint_blobs cb
+                WHERE cb.thread_id = %s
+                  AND cb.checkpoint_ns = %s
+                  AND cb.channel = %s
+                  AND cb.version = (
+                      SELECT checkpoint->'channel_versions'->>%s
+                      FROM checkpoints
+                      WHERE thread_id = %s AND checkpoint_ns = %s AND checkpoint_id = %s
+                  )
+                """,
+                (
+                    thread_id,
+                    checkpoint_ns,
+                    channel,
+                    channel,
+                    thread_id,
+                    checkpoint_ns,
+                    checkpoint_id,
+                ),
+            )
+            row = await cur.fetchone()
+        if row is None:
+            return NotImplemented
+        return self.serde.loads_typed((row["type"], row["blob"]))
+
     async def _load_checkpoint_tuple(self, value: DictRow) -> CheckpointTuple:
         """
         Convert a database row into a CheckpointTuple object.
@@ -403,11 +440,19 @@ class AsyncPostgresSaver(BasePostgresSaver):
             including its configuration, metadata, parent checkpoint (if any),
             and pending writes.
         """
+        thread_id = value["thread_id"]
+        checkpoint_ns = value["checkpoint_ns"]
+        blob_values = value["channel_values"]
+
+        channel_values: dict[str, Any] = {}
+        if blob_values:
+            channel_values = self._load_blobs(blob_values)
+
         return CheckpointTuple(
             {
                 "configurable": {
-                    "thread_id": value["thread_id"],
+                    "thread_id": thread_id,
-                    "checkpoint_ns": value["checkpoint_ns"],
+                    "checkpoint_ns": checkpoint_ns,
                     "checkpoint_id": value["checkpoint_id"],
                 }
             },
@@ -415,15 +460,15 @@ class AsyncPostgresSaver(BasePostgresSaver):
                 **value["checkpoint"],
                 "channel_values": {
                     **(value["checkpoint"].get("channel_values") or {}),
-                    **self._load_blobs(value["channel_values"]),
+                    **channel_values,
                 },
             },
             value["metadata"],
             (
                 {
                     "configurable": {
-                        "thread_id": value["thread_id"],
+                        "thread_id": thread_id,
-                        "checkpoint_ns": value["checkpoint_ns"],
+                        "checkpoint_ns": checkpoint_ns,
                         "checkpoint_id": value["parent_checkpoint_id"],
                     }
                 }

libs/checkpoint-postgres/langgraph/checkpoint/postgres/base.py

@@ -185,15 +185,22 @@ class BasePostgresSaver(BaseCheckpointSaver[str]):
         )
 
     def _load_blobs(
-        self, blob_values: list[tuple[bytes, bytes, bytes]]
+        self,
+        blob_values: list[tuple[bytes, bytes, bytes]],
+        *,
+        thread_id: str = "",
+        checkpoint_ns: str = "",
+        cur: Any = None,
     ) -> dict[str, Any]:
         if not blob_values:
             return {}
-        return {
+        result: dict[str, Any] = {}
-            k.decode(): self.serde.loads_typed((t.decode(), v))
+        for k, t, v in blob_values:
-            for k, t, v in blob_values
+            channel = k.decode()
-            if t.decode() != "empty"
+            type_tag = t.decode()
-        }
+            if type_tag != "empty":
+                result[channel] = self.serde.loads_typed((type_tag, v))
+        return result
 
     def _dump_blobs(
         self,

libs/checkpoint-postgres/tests/test_async.py

@@ -371,3 +371,47 @@ async def test_get_checkpoint_no_channel_values(
 
         checkpoint = await saver.aget_tuple(config)
         assert checkpoint.checkpoint["channel_values"] == {}
+
+
+@pytest.mark.parametrize("saver_name", ["base", "pool", "pipe"])
+async def test_delta_channel_chain_reconstruction(saver_name: str) -> None:
+    """AsyncPostgresSaver reconstructs DeltaChannel chain via point-lookup traversal."""
+    pytest.importorskip(
+        "langgraph.channels.delta", reason="langgraph core not installed"
+    )
+
+    from typing import Annotated
+
+    from langchain_core.messages import AIMessage, HumanMessage
+    from langgraph.channels.delta import DeltaChannel
+    from langgraph.graph import START, StateGraph
+    from langgraph.graph.message import add_messages
+    from typing_extensions import TypedDict
+
+    class State(TypedDict):
+        messages: Annotated[list, DeltaChannel(add_messages)]
+
+    def respond(state: State) -> dict:
+        n = len(state["messages"])
+        return {"messages": [AIMessage(content=f"reply-{n}", id=f"ai-{n}")]}
+
+    builder = StateGraph(State)
+    builder.add_node("respond", respond)
+    builder.add_edge(START, "respond")
+
+    async with _saver(saver_name) as saver:
+        graph = builder.compile(checkpointer=saver)
+        config = {"configurable": {"thread_id": "diff-channel-test-1"}}
+
+        await graph.ainvoke({"messages": [HumanMessage(content="hi", id="h1")]}, config)
+        await graph.ainvoke(
+            {"messages": [HumanMessage(content="there", id="h2")]}, config
+        )
+
+        state = await graph.aget_state(config)
+        msgs = state.values["messages"]
+        assert len(msgs) == 4, f"expected 4, got {len(msgs)}: {msgs}"
+        assert msgs[0].content == "hi"
+        assert msgs[1].content == "reply-1"
+        assert msgs[2].content == "there"
+        assert msgs[3].content == "reply-3"

libs/checkpoint/langgraph/checkpoint/base/__init__.py

@@ -1,6 +1,7 @@
 from __future__ import annotations
 
 import copy
+import dataclasses
 import logging
 from collections.abc import AsyncIterator, Collection, Iterator, Mapping, Sequence
 from typing import (  # noqa: UP035
@@ -28,6 +29,26 @@ from langgraph.checkpoint.serde.types import (
 
 V = TypeVar("V", int, float, str)
 PendingWrite = tuple[str, str, Any]
+
+
+@dataclasses.dataclass
+class DeltaValue:
+    """Returned by DeltaChannel.checkpoint(). Represents one step's writes."""
+
+    delta: list[Any]
+    prev_checkpoint_id: (
+        str | None
+    )  # ID of checkpoint containing previous blob; None = chain root
+
+
+@dataclasses.dataclass
+class DeltaChainValue:
+    """Passed to DeltaChannel.from_checkpoint(). Assembled by the pregel layer."""
+
+    base: list[Any] | None  # starting accumulated value; None = start from empty
+    deltas: list[list[Any]]  # per-step write-sets, ordered oldest → newest
+
+
 logger = logging.getLogger(__name__)
 
 
@@ -457,6 +478,42 @@ class BaseCheckpointSaver(Generic[V]):
         """
         raise NotImplementedError
 
+    def get_channel_blob(
+        self,
+        thread_id: str,
+        checkpoint_ns: str,
+        checkpoint_id: str,
+        channel: str,
+    ) -> Any:
+        """Look up a single channel blob by checkpoint ID + channel name.
+
+        Returns NotImplemented if this saver does not support efficient
+        per-channel-version blob lookup. The pregel layer will fall back to
+        get_tuple() traversal in that case.
+
+        Savers with a dedicated blob store (InMemorySaver, PostgresSaver)
+        should override this for O(1) performance.
+        """
+        return NotImplemented
+
+    async def aget_channel_blob(
+        self,
+        thread_id: str,
+        checkpoint_ns: str,
+        checkpoint_id: str,
+        channel: str,
+    ) -> Any:
+        """Look up a single channel blob by checkpoint ID + channel name (async).
+
+        Returns NotImplemented if this saver does not support efficient
+        per-channel-version blob lookup. The pregel layer will fall back to
+        aget_tuple() traversal in that case.
+
+        Savers with a dedicated blob store (InMemorySaver, PostgresSaver)
+        should override this for O(1) performance.
+        """
+        return NotImplemented
+
     def get_next_version(self, current: V | None, channel: None) -> V:
         """Generate the next version ID for a channel.
 

libs/checkpoint/langgraph/checkpoint/memory/__init__.py

@@ -126,12 +126,46 @@ class InMemorySaver(
         channel_values: dict[str, Any] = {}
         for k, v in versions.items():
             kk = (thread_id, checkpoint_ns, k, v)
-            if kk in self.blobs:
+            if kk not in self.blobs:
-                vv = self.blobs[kk]
+                continue
-                if vv[0] != "empty":
+            vv = self.blobs[kk]
-                    channel_values[k] = self.serde.loads_typed(vv)
+            if vv[0] != "empty":
+                channel_values[k] = self.serde.loads_typed(vv)
         return channel_values
 
+    def get_channel_blob(
+        self,
+        thread_id: str,
+        checkpoint_ns: str,
+        checkpoint_id: str,
+        channel: str,
+    ) -> Any:
+        """Fast-path blob lookup: checkpoint → channel version → blob."""
+        ns_storage = self.storage.get(thread_id, {}).get(checkpoint_ns, {})
+        entry = ns_storage.get(checkpoint_id)
+        if entry is None:
+            return NotImplemented
+        checkpoint = self.serde.loads_typed(entry[0])
+        version = checkpoint["channel_versions"].get(channel)
+        if version is None:
+            return NotImplemented
+        kk = (thread_id, checkpoint_ns, channel, version)
+        if kk not in self.blobs:
+            return NotImplemented
+        vv = self.blobs[kk]
+        if vv[0] == "empty":
+            return NotImplemented
+        return self.serde.loads_typed(vv)
+
+    async def aget_channel_blob(
+        self,
+        thread_id: str,
+        checkpoint_ns: str,
+        checkpoint_id: str,
+        channel: str,
+    ) -> Any:
+        return self.get_channel_blob(thread_id, checkpoint_ns, checkpoint_id, channel)
+
     def get_tuple(self, config: RunnableConfig) -> CheckpointTuple | None:
         """Get a checkpoint tuple from the in-memory storage.
 

libs/checkpoint/langgraph/checkpoint/serde/_msgpack.py

@@ -80,6 +80,8 @@ SAFE_MSGPACK_TYPES: frozenset[tuple[str, ...]] = frozenset(
         ("langgraph.types", "Overwrite"),
         ("langgraph.store.base", "Item"),
         ("langgraph.store.base", "GetOp"),
+        # DeltaChannel checkpoint value type
+        ("langgraph.checkpoint.base", "DeltaValue"),
     }
 )
 

libs/checkpoint/langgraph/checkpoint/serde/jsonplus.py

@@ -47,6 +47,12 @@ EMPTY_BYTES = b""
 logger = logging.getLogger(__name__)
 
 
+def _is_delta_value(obj: Any) -> bool:
+    from langgraph.checkpoint.base import DeltaValue  # lazy import avoids circular dep
+
+    return isinstance(obj, DeltaValue)
+
+
 class JsonPlusSerializer(SerializerProtocol):
     """Serializer that uses ormsgpack, with optional fallbacks.
 
@@ -239,6 +245,8 @@ class JsonPlusSerializer(SerializerProtocol):
             return "bytes", obj
         elif isinstance(obj, bytearray):
             return "bytearray", obj
+        elif _is_delta_value(obj):
+            return "delta", _msgpack_enc({"d": obj.delta, "c": obj.prev_checkpoint_id})
         else:
             try:
                 return "msgpack", _msgpack_enc(obj)
@@ -261,6 +269,13 @@ class JsonPlusSerializer(SerializerProtocol):
             return ormsgpack.unpackb(
                 data_, ext_hook=self._unpack_ext_hook, option=ormsgpack.OPT_NON_STR_KEYS
             )
+        elif type_ == "delta":
+            from langgraph.checkpoint.base import DeltaValue  # lazy import
+
+            raw = ormsgpack.unpackb(
+                data_, ext_hook=self._unpack_ext_hook, option=ormsgpack.OPT_NON_STR_KEYS
+            )
+            return DeltaValue(delta=raw["d"], prev_checkpoint_id=raw.get("c"))
         elif self.pickle_fallback and type_ == "pickle":
             return pickle.loads(data_)
         else:

libs/checkpoint/tests/test_jsonplus.py

@@ -983,3 +983,31 @@ def test_msgpack_nested_pydantic_serializes_as_dict(
     # No blocking should occur - inner is serialized as dict, not ext
     assert "blocked" not in caplog.text.lower()
     assert result == obj
+
+
+def test_delta_value_serde_round_trip() -> None:
+    from langgraph.checkpoint.base import DeltaValue
+    from langgraph.checkpoint.serde.jsonplus import JsonPlusSerializer
+
+    serde = JsonPlusSerializer()
+    original = DeltaValue(
+        delta=[{"type": "human", "content": "hi"}], prev_checkpoint_id="abc-123"
+    )
+    type_tag, blob = serde.dumps_typed(original)
+    assert type_tag == "delta"
+    loaded = serde.loads_typed((type_tag, blob))
+    assert isinstance(loaded, DeltaValue)
+    assert loaded.delta == original.delta
+    assert loaded.prev_checkpoint_id == "abc-123"
+
+
+def test_delta_value_serde_chain_root() -> None:
+    from langgraph.checkpoint.base import DeltaValue
+    from langgraph.checkpoint.serde.jsonplus import JsonPlusSerializer
+
+    serde = JsonPlusSerializer()
+    original = DeltaValue(delta=[], prev_checkpoint_id=None)
+    type_tag, blob = serde.dumps_typed(original)
+    loaded = serde.loads_typed((type_tag, blob))
+    assert isinstance(loaded, DeltaValue)
+    assert loaded.prev_checkpoint_id is None

libs/checkpoint/tests/test_memory.py

@@ -308,3 +308,36 @@ def test_memory_saver_with_allowlist_proxy_isolated() -> None:
     assert direct is not None
     expected = obj.model_dump() if hasattr(obj, "model_dump") else obj.dict()
     assert direct.checkpoint["channel_values"]["foo"] == expected
+
+
+class TestInMemorySaverDeltaChannel:
+    def test_get_channel_blob(self) -> None:
+        """get_channel_blob returns the deserialized blob for a checkpoint+channel."""
+        from langgraph.checkpoint.base import DeltaValue, empty_checkpoint
+
+        saver = InMemorySaver()
+        serde = JsonPlusSerializer()
+
+        thread_id, ns, channel = "t1", "", "messages"
+        version = "00000000000000000000000000000001.0000000000000000"
+        delta = DeltaValue(delta=[{"content": "hi"}], prev_checkpoint_id=None)
+        saver.blobs[(thread_id, ns, channel, version)] = serde.dumps_typed(delta)
+
+        cp = empty_checkpoint()
+        cp["id"] = "cp1"
+        cp["channel_versions"][channel] = version
+        saver.storage[thread_id][ns] = {
+            "cp1": (serde.dumps_typed(cp), serde.dumps_typed({}), None)
+        }
+
+        result = saver.get_channel_blob(thread_id, ns, "cp1", channel)
+        assert isinstance(result, DeltaValue)
+        assert result.delta == [{"content": "hi"}]
+        assert result.prev_checkpoint_id is None
+
+    def test_get_channel_blob_missing(self) -> None:
+        """get_channel_blob returns NotImplemented when checkpoint or channel not found."""
+        saver = InMemorySaver()
+        assert (
+            saver.get_channel_blob("t1", "", "no-such-cp", "messages") is NotImplemented
+        )

libs/langgraph/langgraph/channels/__init__.py

@@ -1,6 +1,7 @@
 from langgraph.channels.any_value import AnyValue
 from langgraph.channels.base import BaseChannel
 from langgraph.channels.binop import BinaryOperatorAggregate
+from langgraph.channels.delta import DeltaChannel
 from langgraph.channels.ephemeral_value import EphemeralValue
 from langgraph.channels.last_value import LastValue, LastValueAfterFinish
 from langgraph.channels.named_barrier_value import (
@@ -20,6 +21,7 @@ __all__ = (
     "UntrackedValue",
     "EphemeralValue",
     "BinaryOperatorAggregate",
+    "DeltaChannel",
     "NamedBarrierValue",
     "NamedBarrierValueAfterFinish",
     # topics

libs/langgraph/langgraph/channels/base.py

@@ -119,3 +119,12 @@ class BaseChannel(Generic[Value, Update, Checkpoint], ABC):
         Returns `True` if the channel was updated, `False` otherwise.
         """
         return False
+
+    def after_checkpoint(self, version: Any, checkpoint_id: str | None = None) -> None:
+        """Called after checkpoint() with the assigned version, and after
+        from_checkpoint() with the current channel version.
+
+        No-op by default. Override in channels that track their own version
+        for incremental checkpointing (e.g. DeltaChannel).
+        """
+        pass

libs/langgraph/langgraph/channels/delta.py

@@ -0,0 +1,206 @@
+from __future__ import annotations
+
+import collections.abc
+from collections.abc import Callable, Sequence
+from typing import Any, Generic
+
+from langgraph.checkpoint.base import DeltaChainValue, DeltaValue
+from typing_extensions import Self
+
+from langgraph._internal._typing import MISSING
+from langgraph.channels.base import BaseChannel, Value
+from langgraph.channels.binop import _get_overwrite, _strip_extras
+from langgraph.errors import EmptyChannelError
+
+__all__ = ("DeltaChannel",)
+
+
+class DeltaChannel(Generic[Value], BaseChannel[list[Value], Value, DeltaValue]):
+    """A channel that stores only per-step write deltas in checkpoints.
+
+    Reconstructs the full accumulated list at load time by replaying the
+    chain of deltas through the operator. Use with append-style reducers
+    (e.g. `add_messages`) on long-running threads to reduce checkpoint
+    storage from O(N²) to O(N).
+
+    Works with all checkpointers. Savers with a dedicated blob store
+    (InMemorySaver, PostgresSaver) use an O(1) fast-path per chain step;
+    all others (SQLite, MongoDB, etc.) fall back to get_tuple traversal.
+
+    Use `snapshot_every=N` to cap chain traversal depth at N steps. Every N
+    steps a full snapshot is written as the chain root; subsequent deltas
+    chain back to it, so `get_state` / reload never traverses more than N
+    checkpoints regardless of thread length. Recommended for savers without
+    a dedicated blob store.
+
+    Usage::
+
+        class State(TypedDict):
+            messages: Annotated[list[AnyMessage], DeltaChannel(add_messages)]
+            # Cap reconstruction depth (recommended for SQLite / MongoDB savers):
+            messages: Annotated[list[AnyMessage], DeltaChannel(add_messages, snapshot_every=50)]
+    """
+
+    __slots__ = (
+        "value",
+        "operator",
+        "snapshot_every",
+        "_pending",
+        "_base_version",
+        "_last_checkpoint_id",
+        "_overwritten",
+        "_steps_since_snapshot",
+    )
+
+    def __init__(
+        self,
+        operator: Callable[[list[Value], Any], list[Value]],
+        typ: type = list,
+        *,
+        snapshot_every: int | None = None,
+    ) -> None:
+        typ = _strip_extras(typ)
+        if typ in (
+            collections.abc.Sequence,
+            collections.abc.MutableSequence,
+        ):
+            typ = list
+        super().__init__(typ)
+        self.operator = operator
+        self.snapshot_every = snapshot_every
+        try:
+            self.value: list[Value] = typ()
+        except Exception:
+            self.value = []
+        self._pending: list[Any] = []
+        self._base_version: str | None = None
+        self._last_checkpoint_id: str | None = None
+        self._overwritten: bool = False
+        self._steps_since_snapshot: int = 0
+
+    def __eq__(self, other: object) -> bool:
+        if not isinstance(other, DeltaChannel):
+            return False
+        if self.snapshot_every != other.snapshot_every:
+            return False
+        if (
+            self.operator.__name__ != "<lambda>"
+            and other.operator.__name__ != "<lambda>"
+        ):
+            return self.operator is other.operator
+        return True
+
+    @property
+    def ValueType(self) -> Any:
+        return list[self.typ]  # type: ignore[name-defined]
+
+    @property
+    def UpdateType(self) -> Any:
+        return self.typ | list[self.typ]  # type: ignore[name-defined]
+
+    def copy(self) -> Self:
+        new = DeltaChannel(self.operator, self.typ, snapshot_every=self.snapshot_every)
+        new.key = self.key
+        new.value = self.value[:]
+        new._pending = self._pending[:]
+        new._base_version = self._base_version
+        new._last_checkpoint_id = self._last_checkpoint_id
+        new._overwritten = self._overwritten
+        new._steps_since_snapshot = self._steps_since_snapshot
+        return new
+
+    def from_checkpoint(self, checkpoint: Any) -> Self:
+        new = DeltaChannel(self.operator, self.typ, snapshot_every=self.snapshot_every)
+        new.key = self.key
+        if checkpoint is MISSING:
+            new.value = []
+        elif isinstance(checkpoint, DeltaChainValue):
+            accumulated: list[Value] = list(checkpoint.base) if checkpoint.base else []
+            for step_writes in checkpoint.deltas:
+                for write in step_writes:
+                    accumulated = new.operator(accumulated, write)
+            new.value = accumulated
+            # Seed the counter from actual chain depth so rehydration fires at
+            # the right time regardless of how many prior invocations there were.
+            new._steps_since_snapshot = len(checkpoint.deltas)
+        elif isinstance(checkpoint, DeltaValue):
+            # Should never reach here — the pregel layer assembles DeltaValues
+            # into DeltaChainValue before calling from_checkpoint.
+            raise AssertionError(
+                "DeltaChannel.from_checkpoint received a raw DeltaValue. "
+                "This is a bug in the pregel layer — chain assembly should have "
+                "occurred before from_checkpoint was called."
+            )
+        else:
+            # Backwards compat: plain list from old BinaryOperatorAggregate checkpoint.
+            new.value = list(checkpoint)
+        new._pending = []
+        new._base_version = None  # set by the subsequent after_checkpoint() call
+        new._overwritten = False
+        return new
+
+    def update(self, values: Sequence[Any]) -> bool:
+        if not values:
+            return False
+        seen_overwrite = False
+        for value in values:
+            is_overwrite, overwrite_value = _get_overwrite(value)
+            if is_overwrite:
+                if seen_overwrite:
+                    from langgraph.errors import (
+                        ErrorCode,
+                        InvalidUpdateError,
+                        create_error_message,
+                    )
+
+                    msg = create_error_message(
+                        message="Can receive only one Overwrite value per super-step.",
+                        error_code=ErrorCode.INVALID_CONCURRENT_GRAPH_UPDATE,
+                    )
+                    raise InvalidUpdateError(msg)
+                self.value = (
+                    list(overwrite_value) if overwrite_value is not None else []
+                )
+                self._pending = list(self.value)
+                self._overwritten = True
+                seen_overwrite = True
+            elif not seen_overwrite:
+                self.value = self.operator(self.value, value)
+                self._pending.append(value)
+        return True
+
+    def get(self) -> list[Value]:
+        if self.value is MISSING:
+            raise EmptyChannelError()
+        return self.value
+
+    def is_available(self) -> bool:
+        return self.value is not MISSING
+
+    def checkpoint(self) -> Any:
+        if (
+            self.snapshot_every is not None
+            and self._steps_since_snapshot >= self.snapshot_every
+        ):
+            # Emit a full snapshot to cap chain depth at snapshot_every.
+            # The saver stores this as a plain (non-diff) blob, so future
+            # deltas will chain back to it and traversal depth resets to 1.
+            return list(self.value)
+        return DeltaValue(
+            delta=self._pending[:],
+            prev_checkpoint_id=None if self._overwritten else self._last_checkpoint_id,
+        )
+
+    def after_checkpoint(self, version: Any, checkpoint_id: str | None = None) -> None:
+        if version != self._base_version:
+            if self._base_version is None:
+                pass  # First call after from_checkpoint — anchor without counting a step.
+            elif self.snapshot_every is not None:
+                if self._steps_since_snapshot >= self.snapshot_every:
+                    self._steps_since_snapshot = 0
+                else:
+                    self._steps_since_snapshot += 1
+            self._base_version = version
+            self._last_checkpoint_id = checkpoint_id
+            self._pending = []
+            self._overwritten = False

libs/langgraph/langgraph/pregel/_checkpoint.py

@@ -1,9 +1,17 @@
 from __future__ import annotations
 
+import logging
 from collections.abc import Mapping
 from datetime import datetime, timezone
+from typing import Any
 
-from langgraph.checkpoint.base import Checkpoint
+from langchain_core.runnables import RunnableConfig
+from langgraph.checkpoint.base import (
+    BaseCheckpointSaver,
+    Checkpoint,
+    DeltaChainValue,
+    DeltaValue,
+)
 from langgraph.checkpoint.base.id import uuid6
 
 from langgraph._internal._typing import MISSING
@@ -12,6 +20,171 @@ from langgraph.managed.base import ManagedValueMapping, ManagedValueSpec
 
 LATEST_VERSION = 4
 
+logger = logging.getLogger(__name__)
+
+_MISSING_SENTINEL = object()
+
+
+def _assemble_delta_channels(
+    checkpoint: Checkpoint,
+    config: RunnableConfig,
+    checkpointer: BaseCheckpointSaver,
+) -> dict[str, Any]:
+    """Resolve any DeltaValue entries in checkpoint channel_values to DeltaChainValue.
+
+    Returns a dict of only the channels that needed assembly (others are untouched).
+    Tries get_channel_blob fast-path first; falls back to get_tuple traversal.
+    """
+    thread_id = str(config["configurable"]["thread_id"])
+    checkpoint_ns = config["configurable"].get("checkpoint_ns", "")
+    current_checkpoint_id = checkpoint.get("id")
+    assembled: dict[str, Any] = {}
+
+    for channel, value in checkpoint["channel_values"].items():
+        if not isinstance(value, DeltaValue):
+            continue
+
+        chain_deltas: list[list[Any]] = []
+        base: list[Any] | None = None
+        cursor: DeltaValue = value
+        # Pre-seed with current checkpoint ID to guard against self-referential chains.
+        visited: set[str] = {current_checkpoint_id} if current_checkpoint_id else set()
+
+        while True:
+            chain_deltas.append(cursor.delta)
+            prev_id = cursor.prev_checkpoint_id
+            if prev_id is None:
+                break  # chain root
+            if prev_id in visited:
+                logger.warning(
+                    "DeltaChannel chain cycle at checkpoint %r for channel %r; breaking",
+                    prev_id,
+                    channel,
+                )
+                break
+            visited.add(prev_id)
+
+            # Fast path: saver has a dedicated blob store.
+            blob = checkpointer.get_channel_blob(
+                thread_id, checkpoint_ns, prev_id, channel
+            )
+            if blob is not NotImplemented:
+                if isinstance(blob, DeltaValue):
+                    cursor = blob
+                    continue
+                else:
+                    base = blob  # plain list = snapshot root
+                    break
+
+            # Fallback: load the full checkpoint and extract channel value.
+            parent_config: RunnableConfig = {
+                "configurable": {
+                    "thread_id": thread_id,
+                    "checkpoint_ns": checkpoint_ns,
+                    "checkpoint_id": prev_id,
+                }
+            }
+            parent_tuple = checkpointer.get_tuple(parent_config)
+            if parent_tuple is None:
+                logger.warning(
+                    "DeltaChannel chain broken: checkpoint %r not found for channel %r",
+                    prev_id,
+                    channel,
+                )
+                break
+            prev_val = parent_tuple.checkpoint["channel_values"].get(
+                channel, _MISSING_SENTINEL
+            )
+            if prev_val is _MISSING_SENTINEL:
+                break
+            elif isinstance(prev_val, DeltaValue):
+                cursor = prev_val
+            else:
+                base = prev_val
+                break
+
+        chain_deltas.reverse()
+        assembled[channel] = DeltaChainValue(base=base, deltas=chain_deltas)
+
+    return assembled
+
+
+async def _aassemble_delta_channels(
+    checkpoint: Checkpoint,
+    config: RunnableConfig,
+    checkpointer: BaseCheckpointSaver,
+) -> dict[str, Any]:
+    """Async version of _assemble_delta_channels."""
+    thread_id = str(config["configurable"]["thread_id"])
+    checkpoint_ns = config["configurable"].get("checkpoint_ns", "")
+    current_checkpoint_id = checkpoint.get("id")
+    assembled: dict[str, Any] = {}
+
+    for channel, value in checkpoint["channel_values"].items():
+        if not isinstance(value, DeltaValue):
+            continue
+
+        chain_deltas: list[list[Any]] = []
+        base: list[Any] | None = None
+        cursor: DeltaValue = value
+        visited: set[str] = {current_checkpoint_id} if current_checkpoint_id else set()
+
+        while True:
+            chain_deltas.append(cursor.delta)
+            prev_id = cursor.prev_checkpoint_id
+            if prev_id is None:
+                break
+            if prev_id in visited:
+                logger.warning(
+                    "DeltaChannel chain cycle at checkpoint %r for channel %r; breaking",
+                    prev_id,
+                    channel,
+                )
+                break
+            visited.add(prev_id)
+
+            blob = await checkpointer.aget_channel_blob(
+                thread_id, checkpoint_ns, prev_id, channel
+            )
+            if blob is not NotImplemented:
+                if isinstance(blob, DeltaValue):
+                    cursor = blob
+                    continue
+                else:
+                    base = blob
+                    break
+
+            parent_config: RunnableConfig = {
+                "configurable": {
+                    "thread_id": thread_id,
+                    "checkpoint_ns": checkpoint_ns,
+                    "checkpoint_id": prev_id,
+                }
+            }
+            parent_tuple = await checkpointer.aget_tuple(parent_config)
+            if parent_tuple is None:
+                logger.warning(
+                    "DeltaChannel chain broken: checkpoint %r not found for channel %r",
+                    prev_id,
+                    channel,
+                )
+                break
+            prev_val = parent_tuple.checkpoint["channel_values"].get(
+                channel, _MISSING_SENTINEL
+            )
+            if prev_val is _MISSING_SENTINEL:
+                break
+            elif isinstance(prev_val, DeltaValue):
+                cursor = prev_val
+            else:
+                base = prev_val
+                break
+
+        chain_deltas.reverse()
+        assembled[channel] = DeltaChainValue(base=base, deltas=chain_deltas)
+
+    return assembled
+
 
 def empty_checkpoint() -> Checkpoint:
     return Checkpoint(
@@ -67,13 +240,12 @@ def channels_from_checkpoint(
             channel_specs[k] = v
         else:
             managed_specs[k] = v
-    return (
+    channels: dict[str, BaseChannel] = {}
-        {
+    for k, v in channel_specs.items():
-            k: v.from_checkpoint(checkpoint["channel_values"].get(k, MISSING))
+        ch = v.from_checkpoint(checkpoint["channel_values"].get(k, MISSING))
-            for k, v in channel_specs.items()
+        ch.after_checkpoint(checkpoint["channel_versions"].get(k), checkpoint.get("id"))
-        },
+        channels[k] = ch
-        managed_specs,
+    return channels, managed_specs
-    )
 
 
 def copy_checkpoint(checkpoint: Checkpoint) -> Checkpoint:

libs/langgraph/langgraph/pregel/_loop.py

@@ -92,6 +92,8 @@ from langgraph.pregel._algo import (
     task_path_str,
 )
 from langgraph.pregel._checkpoint import (
+    _aassemble_delta_channels,
+    _assemble_delta_channels,
     channels_from_checkpoint,
     copy_checkpoint,
     create_checkpoint,
@@ -881,6 +883,12 @@ class PregelLoop:
             id=self.checkpoint["id"] if exiting else None,
             updated_channels=self.updated_channels,
         )
+        if do_checkpoint and self.channels:
+            for k, ch in self.channels.items():
+                ch.after_checkpoint(
+                    self.checkpoint["channel_versions"].get(k),
+                    self.checkpoint.get("id"),
+                )
         # sanitize TASK channel in the checkpoint before saving (durability=="exit")
         if TASKS in self.checkpoint["channel_values"] and any(
             isinstance(channel, UntrackedValue) for channel in self.channels.values()
@@ -1262,6 +1270,19 @@ class SyncPregelLoop(PregelLoop, AbstractContextManager):
             else []
         )
         self.submit = self.stack.enter_context(BackgroundExecutor(self.config))
+        # Assemble any DeltaChannel chains before constructing channel objects.
+        if self.checkpointer is not None:
+            assembled = _assemble_delta_channels(
+                self.checkpoint, self.checkpoint_config, self.checkpointer
+            )
+            if assembled:
+                self.checkpoint = {
+                    **self.checkpoint,
+                    "channel_values": {
+                        **self.checkpoint["channel_values"],
+                        **assembled,
+                    },
+                }
         self.channels, self.managed = channels_from_checkpoint(
             self.specs, self.checkpoint
         )
@@ -1466,6 +1487,18 @@ class AsyncPregelLoop(PregelLoop, AbstractAsyncContextManager):
         self.submit = await self.stack.enter_async_context(
             AsyncBackgroundExecutor(self.config)
         )
+        if self.checkpointer is not None:
+            assembled = await _aassemble_delta_channels(
+                self.checkpoint, self.checkpoint_config, self.checkpointer
+            )
+            if assembled:
+                self.checkpoint = {
+                    **self.checkpoint,
+                    "channel_values": {
+                        **self.checkpoint["channel_values"],
+                        **assembled,
+                    },
+                }
         self.channels, self.managed = channels_from_checkpoint(
             self.specs, self.checkpoint
         )

libs/langgraph/langgraph/pregel/main.py

@@ -122,6 +122,8 @@ from langgraph.pregel._algo import (
 )
 from langgraph.pregel._call import identifier
 from langgraph.pregel._checkpoint import (
+    _aassemble_delta_channels,
+    _assemble_delta_channels,
     channels_from_checkpoint,
     copy_checkpoint,
     create_checkpoint,
@@ -1049,13 +1051,23 @@ class Pregel(
 
         step = saved.metadata.get("step", -1) + 1
         stop = step + 2
+        checkpoint = saved.checkpoint
+        if isinstance(self.checkpointer, BaseCheckpointSaver):
+            assembled = _assemble_delta_channels(
+                checkpoint, saved.config, self.checkpointer
+            )
+            if assembled:
+                checkpoint = {
+                    **checkpoint,
+                    "channel_values": {**checkpoint["channel_values"], **assembled},
+                }
         channels, managed = channels_from_checkpoint(
             self.channels,
-            saved.checkpoint,
+            checkpoint,
         )
         # tasks for this checkpoint
         next_tasks = prepare_next_tasks(
-            saved.checkpoint,
+            checkpoint,
             saved.pending_writes or [],
             self.nodes,
             channels,
@@ -1168,13 +1180,23 @@ class Pregel(
 
         step = saved.metadata.get("step", -1) + 1
         stop = step + 2
+        checkpoint = saved.checkpoint
+        if isinstance(self.checkpointer, BaseCheckpointSaver):
+            assembled = await _aassemble_delta_channels(
+                checkpoint, saved.config, self.checkpointer
+            )
+            if assembled:
+                checkpoint = {
+                    **checkpoint,
+                    "channel_values": {**checkpoint["channel_values"], **assembled},
+                }
         channels, managed = channels_from_checkpoint(
             self.channels,
-            saved.checkpoint,
+            checkpoint,
         )
         # tasks for this checkpoint
         next_tasks = prepare_next_tasks(
-            saved.checkpoint,
+            checkpoint,
             saved.pending_writes or [],
             self.nodes,
             channels,
@@ -1520,9 +1542,20 @@ class Pregel(
             saved = checkpointer.get_tuple(config)
             if saved is not None:
                 self._migrate_checkpoint(saved.checkpoint)
-            checkpoint = (
+            base_checkpoint = saved.checkpoint if saved else empty_checkpoint()
-                copy_checkpoint(saved.checkpoint) if saved else empty_checkpoint()
+            if saved:
-            )
+                assembled = _assemble_delta_channels(
+                    base_checkpoint, saved.config, checkpointer
+                )
+                if assembled:
+                    base_checkpoint = {
+                        **base_checkpoint,
+                        "channel_values": {
+                            **base_checkpoint["channel_values"],
+                            **assembled,
+                        },
+                    }
+            checkpoint = copy_checkpoint(base_checkpoint) if saved else base_checkpoint
             checkpoint_previous_versions = (
                 saved.checkpoint["channel_versions"].copy() if saved else {}
             )
@@ -1966,9 +1999,20 @@ class Pregel(
             saved = await checkpointer.aget_tuple(config)
             if saved is not None:
                 self._migrate_checkpoint(saved.checkpoint)
-            checkpoint = (
+            base_checkpoint = saved.checkpoint if saved else empty_checkpoint()
-                copy_checkpoint(saved.checkpoint) if saved else empty_checkpoint()
+            if saved:
-            )
+                assembled = await _aassemble_delta_channels(
+                    base_checkpoint, saved.config, checkpointer
+                )
+                if assembled:
+                    base_checkpoint = {
+                        **base_checkpoint,
+                        "channel_values": {
+                            **base_checkpoint["channel_values"],
+                            **assembled,
+                        },
+                    }
+            checkpoint = copy_checkpoint(base_checkpoint) if saved else base_checkpoint
             checkpoint_previous_versions = (
                 saved.checkpoint["channel_versions"].copy() if saved else {}
             )

libs/langgraph/tests/test_channels.py

@@ -117,3 +117,408 @@ def test_untracked_value() -> None:
     new_channel = UntrackedValue(dict).from_checkpoint(checkpoint)
     with pytest.raises(EmptyChannelError):
         new_channel.get()
+
+
+def test_delta_channel_basic_two_steps() -> None:
+    from langchain_core.messages import AIMessage, HumanMessage
+    from langgraph.checkpoint.base import DeltaValue
+
+    from langgraph.channels.delta import DeltaChannel
+    from langgraph.graph.message import add_messages
+
+    ch = DeltaChannel(add_messages).from_checkpoint(MISSING)
+    ch.after_checkpoint(None)
+
+    # Step 1: one message added
+    ch.update([HumanMessage(content="hi", id="h1")])
+    d1 = ch.checkpoint()
+    assert isinstance(d1, DeltaValue)
+    assert len(d1.delta) == 1
+    assert d1.prev_checkpoint_id is None  # first ever step
+    ch.after_checkpoint("v1", checkpoint_id="cid1")
+
+    # Step 2: another message
+    ch.update([AIMessage(content="hello", id="a1")])
+    d2 = ch.checkpoint()
+    assert d2.prev_checkpoint_id == "cid1"
+    assert len(d2.delta) == 1
+    ch.after_checkpoint("v2")
+
+    # Full accumulated value is preserved in memory
+    assert len(ch.get()) == 2
+    assert ch.get()[0].content == "hi"
+    assert ch.get()[1].content == "hello"
+
+
+def test_delta_channel_after_checkpoint_no_op_when_unchanged() -> None:
+    from langchain_core.messages import HumanMessage
+
+    from langgraph.channels.delta import DeltaChannel
+    from langgraph.graph.message import add_messages
+
+    ch = DeltaChannel(add_messages).from_checkpoint(MISSING)
+    ch.after_checkpoint(None)
+    ch.update([HumanMessage(content="hi", id="h1")])
+    ch.after_checkpoint("v1")
+
+    # Same version: no-op
+    ch.after_checkpoint("v1")
+    assert ch._base_version == "v1"
+    assert ch._pending == []
+
+
+def test_delta_channel_from_checkpoint_chain() -> None:
+    from langchain_core.messages import AIMessage, HumanMessage
+    from langgraph.checkpoint.base import DeltaChainValue
+
+    from langgraph.channels.delta import DeltaChannel
+    from langgraph.graph.message import add_messages
+
+    spec = DeltaChannel(add_messages)
+    chain = DeltaChainValue(
+        base=None,
+        deltas=[
+            [HumanMessage(content="hi", id="h1")],
+            [AIMessage(content="hello", id="a1")],
+            [HumanMessage(content="bye", id="h2")],
+        ],
+    )
+    ch = spec.from_checkpoint(chain)
+    msgs = ch.get()
+    assert len(msgs) == 3
+    assert msgs[0].content == "hi"
+    assert msgs[1].content == "hello"
+    assert msgs[2].content == "bye"
+
+
+def test_delta_channel_from_checkpoint_backwards_compat() -> None:
+    from langchain_core.messages import HumanMessage
+
+    from langgraph.channels.delta import DeltaChannel
+    from langgraph.graph.message import add_messages
+
+    # Old BinaryOperatorAggregate checkpoint: plain list
+    spec = DeltaChannel(add_messages)
+    old_value = [HumanMessage(content="old", id="h1")]
+    ch = spec.from_checkpoint(old_value)
+    assert ch.get() == old_value
+
+
+def test_delta_channel_overwrite_resets_chain() -> None:
+    from langchain_core.messages import HumanMessage
+    from langgraph.checkpoint.base import DeltaValue
+
+    from langgraph.channels.delta import DeltaChannel
+    from langgraph.graph.message import add_messages
+    from langgraph.types import Overwrite
+
+    ch = DeltaChannel(add_messages).from_checkpoint(MISSING)
+    ch.after_checkpoint(None)
+    ch.update([HumanMessage(content="old", id="h1")])
+    ch.after_checkpoint("v1")
+
+    # Overwrite should create a root blob (prev_checkpoint_id=None)
+    ch.update([Overwrite([HumanMessage(content="new", id="h2")])])
+    d = ch.checkpoint()
+    assert isinstance(d, DeltaValue)
+    assert d.prev_checkpoint_id is None  # chain root
+    assert len(d.delta) == 1
+    assert d.delta[0].content == "new"
+
+
+def test_delta_channel_assembly_fallback_via_get_tuple() -> None:
+    """Assembly falls back to get_tuple for savers without get_channel_blob."""
+    from unittest.mock import MagicMock
+
+    from langgraph.checkpoint.base import (
+        CheckpointTuple,
+        DeltaChainValue,
+        DeltaValue,
+        empty_checkpoint,
+    )
+
+    from langgraph.channels.delta import DeltaChannel
+    from langgraph.graph.message import add_messages
+    from langgraph.pregel._checkpoint import _assemble_delta_channels
+
+    msg1 = {"type": "human", "content": "hello"}
+    msg2 = {"type": "ai", "content": "world"}
+
+    cp1 = empty_checkpoint()
+    cp1["id"] = "cp1"
+    cp1["channel_values"]["messages"] = [msg1]
+
+    cp2 = empty_checkpoint()
+    cp2["id"] = "cp2"
+    cp2["channel_values"]["messages"] = DeltaValue(
+        delta=[msg2], prev_checkpoint_id="cp1"
+    )
+
+    saver = MagicMock()
+    saver.get_channel_blob.return_value = NotImplemented
+    saver.get_tuple.return_value = CheckpointTuple(
+        config={
+            "configurable": {
+                "thread_id": "t1",
+                "checkpoint_ns": "",
+                "checkpoint_id": "cp1",
+            }
+        },
+        checkpoint=cp1,
+        metadata={},
+        parent_config=None,
+        pending_writes=[],
+    )
+
+    config = {"configurable": {"thread_id": "t1", "checkpoint_ns": ""}}
+    assembled = _assemble_delta_channels(cp2, config, saver)
+
+    assert "messages" in assembled
+    chain = assembled["messages"]
+    assert isinstance(chain, DeltaChainValue)
+    assert chain.base == [msg1]
+    assert chain.deltas == [[msg2]]
+
+    from langchain_core.messages import AIMessage, HumanMessage
+
+    spec = DeltaChannel(add_messages)
+    ch = spec.from_checkpoint(chain)
+    result = ch.get()
+    assert len(result) == 2
+    assert isinstance(result[0], HumanMessage) and result[0].content == "hello"
+    assert isinstance(result[1], AIMessage) and result[1].content == "world"
+
+
+def test_delta_channel_remove_message_delta_and_replay() -> None:
+    """RemoveMessage stored in a delta must round-trip correctly through the chain."""
+    from langchain_core.messages import AIMessage, HumanMessage, RemoveMessage
+    from langgraph.checkpoint.base import DeltaChainValue, DeltaValue
+
+    from langgraph.channels.delta import DeltaChannel
+    from langgraph.graph.message import add_messages
+
+    spec = DeltaChannel(add_messages)
+    ch = spec.from_checkpoint(MISSING)
+    ch.after_checkpoint(None)
+
+    # Step 1: add two messages
+    ch.update([HumanMessage(content="hi", id="h1")])
+    ch.update([AIMessage(content="hello", id="a1")])
+    d1 = ch.checkpoint()
+    assert isinstance(d1, DeltaValue)
+    ch.after_checkpoint("v1", checkpoint_id="cid1")
+    assert ch.get() == [
+        HumanMessage(content="hi", id="h1"),
+        AIMessage(content="hello", id="a1"),
+    ]
+
+    # Step 2: remove the AI message
+    ch.update([RemoveMessage(id="a1")])
+    d2 = ch.checkpoint()
+    assert isinstance(d2, DeltaValue)
+    assert d2.prev_checkpoint_id == "cid1"
+    assert any(isinstance(w, RemoveMessage) for w in d2.delta)
+    ch.after_checkpoint("v2", checkpoint_id="cid2")
+    assert ch.get() == [HumanMessage(content="hi", id="h1")]
+
+    # Replay the full chain from scratch — must reproduce the post-remove state
+    chain = DeltaChainValue(base=None, deltas=[d1.delta, d2.delta])
+    ch2 = spec.from_checkpoint(chain)
+    assert ch2.get() == [HumanMessage(content="hi", id="h1")]
+
+
+def test_delta_channel_update_by_id_delta_and_replay() -> None:
+    """Updating a message by ID stored in a delta must round-trip correctly."""
+    from langchain_core.messages import HumanMessage
+    from langgraph.checkpoint.base import DeltaChainValue, DeltaValue
+
+    from langgraph.channels.delta import DeltaChannel
+    from langgraph.graph.message import add_messages
+
+    spec = DeltaChannel(add_messages)
+    ch = spec.from_checkpoint(MISSING)
+    ch.after_checkpoint(None)
+
+    # Step 1: add a message
+    ch.update([HumanMessage(content="original", id="h1")])
+    d1 = ch.checkpoint()
+    assert isinstance(d1, DeltaValue)
+    ch.after_checkpoint("v1", checkpoint_id="cid1")
+
+    # Step 2: update the same message by ID
+    ch.update([HumanMessage(content="updated", id="h1")])
+    d2 = ch.checkpoint()
+    assert isinstance(d2, DeltaValue)
+    assert d2.prev_checkpoint_id == "cid1"
+    ch.after_checkpoint("v2", checkpoint_id="cid2")
+    assert ch.get() == [HumanMessage(content="updated", id="h1")]
+
+    # Replay the full chain — must produce the updated message, not the original
+    chain = DeltaChainValue(base=None, deltas=[d1.delta, d2.delta])
+    ch2 = spec.from_checkpoint(chain)
+    assert len(ch2.get()) == 1
+    assert ch2.get()[0].content == "updated"
+
+
+def test_delta_channel_snapshot_every_emits_plain_list() -> None:
+    """snapshot_every=N causes a plain-list snapshot after N steps; next deltas chain to it."""
+    from langchain_core.messages import HumanMessage
+    from langgraph.checkpoint.base import DeltaValue
+
+    from langgraph.channels.delta import DeltaChannel
+    from langgraph.graph.message import add_messages
+
+    SNAP = 3
+    spec = DeltaChannel(add_messages, snapshot_every=SNAP)
+    ch = spec.from_checkpoint(MISSING)
+    # First after_checkpoint anchors _base_version without counting a step.
+    ch.after_checkpoint("v0", checkpoint_id="cid0")
+
+    # Steps 1..SNAP: each should stay as DeltaValue; counter increments each step.
+    for i in range(1, SNAP + 1):
+        ch.update([HumanMessage(content=f"m{i}", id=f"h{i}")])
+        ckpt = ch.checkpoint()
+        assert isinstance(ckpt, DeltaValue), f"expected DeltaValue at step {i}"
+        ch.after_checkpoint(f"v{i}", checkpoint_id=f"cid{i}")
+
+    # Step SNAP+1: _steps_since_snapshot == SNAP → snapshot fires
+    ch.update([HumanMessage(content="snap", id="hsnap")])
+    snap = ch.checkpoint()
+    assert isinstance(snap, list), "expected plain-list snapshot at snapshot_every step"
+    assert len(snap) == SNAP + 1
+
+    # After snapshot, counter resets — next step is DeltaValue again
+    ch.after_checkpoint("vsnap", checkpoint_id="cidsnap")
+    ch.update([HumanMessage(content="post", id="hpost")])
+    post = ch.checkpoint()
+    assert isinstance(post, DeltaValue)
+    assert post.prev_checkpoint_id == "cidsnap"
+
+
+def test_delta_channel_snapshot_every_end_to_end() -> None:
+    """Graph with snapshot_every: get_state returns correct accumulated value after snapshot."""
+    from typing import Annotated
+
+    from langchain_core.messages import AIMessage, HumanMessage
+    from langgraph.checkpoint.memory import InMemorySaver
+    from typing_extensions import TypedDict
+
+    from langgraph.channels.delta import DeltaChannel
+    from langgraph.graph import START, StateGraph
+    from langgraph.graph.message import add_messages
+
+    class State(TypedDict):
+        messages: Annotated[list, DeltaChannel(add_messages, snapshot_every=2)]
+
+    counter = {"n": 0}
+
+    def respond(state: State) -> dict:
+        counter["n"] += 1
+        return {
+            "messages": [
+                AIMessage(content=f"ai-{counter['n']}", id=f"ai-{counter['n']}")
+            ]
+        }
+
+    builder = StateGraph(State)
+    builder.add_node("respond", respond)
+    builder.add_edge(START, "respond")
+    graph = builder.compile(checkpointer=InMemorySaver())
+    config = {"configurable": {"thread_id": "snap-test"}}
+
+    # Run 5 turns — snapshot fires after 2 steps, then again after 2 more
+    for i in range(5):
+        graph.invoke({"messages": [HumanMessage(content=f"h{i}", id=f"h{i}")]}, config)
+
+    state = graph.get_state(config)
+    msgs = state.values["messages"]
+    # 5 human + 5 AI = 10 total
+    assert len(msgs) == 10, f"expected 10 messages, got {len(msgs)}: {msgs}"
+
+
+def test_delta_channel_assembly_fast_path_returns_delta_value() -> None:
+    """get_channel_blob returning a DeltaValue continues chain traversal (fast-path)."""
+    from unittest.mock import MagicMock
+
+    from langgraph.checkpoint.base import (
+        DeltaChainValue,
+        DeltaValue,
+        empty_checkpoint,
+    )
+
+    from langgraph.channels.delta import DeltaChannel
+    from langgraph.graph.message import add_messages
+    from langgraph.pregel._checkpoint import _assemble_delta_channels
+
+    msg1 = {"type": "human", "content": "one"}
+    msg2 = {"type": "ai", "content": "two"}
+    msg3 = {"type": "human", "content": "three"}
+
+    # cp3 → cp2 (DeltaValue) → cp1 (base list)
+    dv_cp2 = DeltaValue(delta=[msg2], prev_checkpoint_id="cp1")
+    cp3 = empty_checkpoint()
+    cp3["id"] = "cp3"
+    cp3["channel_values"]["messages"] = DeltaValue(
+        delta=[msg3], prev_checkpoint_id="cp2"
+    )
+
+    saver = MagicMock()
+
+    def _get_blob(thread_id, ns, checkpoint_id, channel):
+        if checkpoint_id == "cp2":
+            return dv_cp2  # DeltaValue — chain continues
+        if checkpoint_id == "cp1":
+            return [msg1]  # plain list — chain root
+        return NotImplemented
+
+    saver.get_channel_blob.side_effect = _get_blob
+
+    config = {"configurable": {"thread_id": "t1", "checkpoint_ns": ""}}
+    assembled = _assemble_delta_channels(cp3, config, saver)
+
+    chain = assembled["messages"]
+    assert isinstance(chain, DeltaChainValue)
+    assert chain.base == [msg1]
+    assert chain.deltas == [[msg2], [msg3]]
+
+    spec = DeltaChannel(add_messages)
+    ch = spec.from_checkpoint(chain)
+    # add_messages converts dicts to message objects; check by type and content
+
+    result = ch.get()
+    assert len(result) == 3
+    assert result[0].content == "one"
+    assert result[1].content == "two"
+    assert result[2].content == "three"
+
+
+def test_delta_channel_assembly_broken_chain_logs_warning() -> None:
+    """If a prev_checkpoint_id points to a missing checkpoint, log a warning and use partial chain."""
+    from unittest.mock import MagicMock
+
+    from langgraph.checkpoint.base import DeltaValue, empty_checkpoint
+
+    from langgraph.pregel._checkpoint import _assemble_delta_channels
+
+    cp = empty_checkpoint()
+    cp["id"] = "cp2"
+    cp["channel_values"]["messages"] = DeltaValue(
+        delta=["msg2"], prev_checkpoint_id="cp-missing"
+    )
+
+    saver = MagicMock()
+    saver.get_channel_blob.return_value = NotImplemented
+    saver.get_tuple.return_value = None  # checkpoint not found
+
+    config = {"configurable": {"thread_id": "t1", "checkpoint_ns": ""}}
+
+    assembled = _assemble_delta_channels(cp, config, saver)
+
+    # Should still assemble — with partial chain (just the current delta, base=None)
+    assert "messages" in assembled
+    from langgraph.checkpoint.base import DeltaChainValue
+
+    chain = assembled["messages"]
+    assert isinstance(chain, DeltaChainValue)
+    assert chain.base is None
+    assert chain.deltas == [["msg2"]]

libs/langgraph/tests/test_delta_channel_benchmark.py

@@ -0,0 +1,356 @@
+"""Benchmark: DeltaChannel vs BinaryOperatorAggregate storage and time.
+
+Run directly:  python tests/test_delta_channel_benchmark.py
+Run via pytest: pytest tests/test_delta_channel_benchmark.py -s
+
+Simulates realistic multi-turn conversations with paragraph-length messages
+(~100 tokens each) scaling up to 1M-token-equivalent histories.
+
+Token estimates: 1 token ≈ 4 chars; each turn ≈ 200 tokens (human + AI).
+A 1M-token conversation ≈ 5,000 turns of realistic messages.
+"""
+
+from __future__ import annotations
+
+import sys
+import time
+from typing import Annotated, Any
+
+from langchain_core.messages import AIMessage, HumanMessage
+from langgraph.checkpoint.memory import MemorySaver
+from typing_extensions import TypedDict
+
+from langgraph.channels.delta import DeltaChannel
+from langgraph.graph import END, StateGraph
+from langgraph.graph.message import add_messages
+
+try:
+    from langgraph.checkpoint.sqlite import SqliteSaver
+
+    _SQLITE_AVAILABLE = True
+except ImportError:
+    _SQLITE_AVAILABLE = False
+
+SNAPSHOT_EVERY = 50
+
+# ---------------------------------------------------------------------------
+# Realistic message payload (~100 tokens / ~400 chars each)
+# ---------------------------------------------------------------------------
+
+_HUMAN_TEMPLATE = (
+    "I need help understanding the implications of {topic} on our system architecture. "
+    "Specifically, I'm concerned about how this interacts with our existing {concern} "
+    "and whether we need to refactor the {component} layer before proceeding."
+)
+
+_AI_TEMPLATE = (
+    "Great question about {topic}. The key insight here is that {concern} introduces "
+    "a subtle ordering dependency that most teams overlook until they hit it in production. "
+    "For your {component} layer specifically, I'd recommend starting with a careful audit "
+    "of the interface boundaries before making any structural changes. This will give you "
+    "a clear picture of the blast radius and let you sequence the migration safely."
+)
+
+_TOPICS = [
+    "distributed tracing",
+    "eventual consistency",
+    "schema migration",
+    "backpressure handling",
+    "idempotency guarantees",
+    "cache invalidation",
+    "connection pooling",
+    "rate limiting",
+    "circuit breaking",
+    "observability pipelines",
+]
+
+_CONCERNS = [
+    "concurrency model",
+    "retry semantics",
+    "state management",
+    "error propagation",
+    "latency budget",
+]
+
+_COMPONENTS = [
+    "persistence",
+    "routing",
+    "ingestion",
+    "aggregation",
+    "serialization",
+]
+
+
+def _human_content(i: int) -> str:
+    return _HUMAN_TEMPLATE.format(
+        topic=_TOPICS[i % len(_TOPICS)],
+        concern=_CONCERNS[i % len(_CONCERNS)],
+        component=_COMPONENTS[i % len(_COMPONENTS)],
+    )
+
+
+def _ai_content(i: int) -> str:
+    return _AI_TEMPLATE.format(
+        topic=_TOPICS[i % len(_TOPICS)],
+        concern=_CONCERNS[i % len(_CONCERNS)],
+        component=_COMPONENTS[i % len(_COMPONENTS)],
+    )
+
+
+# ---------------------------------------------------------------------------
+# State definitions
+# ---------------------------------------------------------------------------
+
+
+class BinaryState(TypedDict):
+    messages: Annotated[list, add_messages]
+
+
+class DeltaState(TypedDict):
+    messages: Annotated[list, DeltaChannel(add_messages)]
+
+
+class DeltaSnapshotState(TypedDict):
+    messages: Annotated[list, DeltaChannel(add_messages, snapshot_every=SNAPSHOT_EVERY)]
+
+
+# ---------------------------------------------------------------------------
+# Graph factory
+# ---------------------------------------------------------------------------
+
+
+def _make_graph(state_cls: type, checkpointer: Any = None) -> Any:
+    def human_node(state: Any) -> dict:
+        return {}
+
+    def ai_node(state: Any) -> dict:
+        i = len(state["messages"]) // 2
+        return {"messages": [AIMessage(content=_ai_content(i), id=f"a{i}")]}
+
+    g = StateGraph(state_cls)
+    g.add_node("human", human_node)
+    g.add_node("ai", ai_node)
+    g.add_edge("human", "ai")
+    g.add_edge("ai", END)
+    g.set_entry_point("human")
+    return g.compile(checkpointer=checkpointer or MemorySaver())
+
+
+# ---------------------------------------------------------------------------
+# Measurement helpers
+# ---------------------------------------------------------------------------
+
+
+def _total_blob_bytes(saver: MemorySaver) -> int:
+    total = 0
+    for (_, _, _, _), (type_tag, blob) in saver.blobs.items():
+        if blob is not None:
+            total += len(blob)
+    return total
+
+
+def _run_turns(
+    n_turns: int,
+    state_cls: type,
+    checkpointer: Any = None,
+) -> tuple[float, float, int]:
+    """Run n_turns conversation turns.
+
+    Returns (write_elapsed_s, read_elapsed_s, total_blob_bytes).
+    blob_bytes is -1 for savers without in-memory blob stores (e.g. SQLite).
+    Read latency is measured as the time to invoke the graph with no new
+    messages after the full history is built — this forces state rehydration.
+    """
+    graph = _make_graph(state_cls, checkpointer)
+    config = {"configurable": {"thread_id": "bench"}}
+
+    t0 = time.perf_counter()
+    for i in range(n_turns):
+        graph.invoke(
+            {"messages": [HumanMessage(content=_human_content(i), id=f"h{i}")]},
+            config,
+        )
+    write_elapsed = time.perf_counter() - t0
+
+    # Measure read/rehydration: get_state forces the channel to rebuild
+    t1 = time.perf_counter()
+    for _ in range(5):
+        graph.get_state(config)
+    read_elapsed = (time.perf_counter() - t1) / 5
+
+    if isinstance(graph.checkpointer, MemorySaver):
+        blob_bytes = _total_blob_bytes(graph.checkpointer)
+    else:
+        blob_bytes = -1
+    return write_elapsed, read_elapsed, blob_bytes
+
+
+def _fmt_bytes(n: int) -> str:
+    if n >= 1_000_000:
+        return f"{n / 1_000_000:.1f} MB"
+    if n >= 1_000:
+        return f"{n / 1_000:.1f} KB"
+    return f"{n} B"
+
+
+def _approx_tokens(n_turns: int) -> str:
+    # ~100 tokens human + ~100 tokens AI per turn
+    tokens = n_turns * 200
+    if tokens >= 1_000_000:
+        return f"~{tokens / 1_000_000:.1f}M tok"
+    if tokens >= 1_000:
+        return f"~{tokens / 1_000:.0f}K tok"
+    return f"~{tokens} tok"
+
+
+# ---------------------------------------------------------------------------
+# Benchmark matrix
+# ---------------------------------------------------------------------------
+
+# Turn counts chosen to span from a short session to a long-running agent conversation.
+# Storage and time complexity differences are clearly visible by 500 turns.
+# Extrapolation: 5,000 turns × ~200 tokens/turn ≈ 1M tokens (Claude's full context window).
+TURN_COUNTS = [50, 100, 200, 500]
+
+
+def _checkpointer_factories() -> list[tuple[str, Any]]:
+    """Return (label, context_manager_or_none) pairs for available checkpointers."""
+    factories: list[tuple[str, Any]] = [("InMemory", None)]
+    if _SQLITE_AVAILABLE:
+        import tempfile
+
+        factories.append(("SQLite", tempfile.NamedTemporaryFile(suffix=".db")))
+    return factories
+
+
+def run_benchmark() -> None:
+    print()
+    print(
+        "DeltaChannel vs add_messages (BinaryOperatorAggregate) — checkpoint storage & latency"
+    )
+    print("Simulating realistic multi-turn conversations up to ~1M-token histories")
+    print("(5,000 turns × ~200 tokens/turn ≈ 1M tokens — Claude's full context window)")
+    print()
+
+    checkpointers: list[tuple[str, Any]] = [("InMemory (fast-path)", None)]
+    if _SQLITE_AVAILABLE:
+        checkpointers.append(("SQLite (get_tuple fallback)", "sqlite"))
+
+    for cp_label, cp_hint in checkpointers:
+        print(f"--- Checkpointer: {cp_label} ---")
+        _run_benchmark_for_checkpointer(cp_hint)
+
+
+def _run_benchmark_for_checkpointer(cp_hint: Any) -> None:
+    import contextlib
+    import tempfile
+
+    @contextlib.contextmanager
+    def _make_saver():
+        if cp_hint is None:
+            yield None
+        else:
+            with tempfile.NamedTemporaryFile(suffix=".db") as f:
+                with SqliteSaver.from_conn_string(f.name) as saver:
+                    yield saver
+
+    W = 120
+    print("=" * W)
+    header = (
+        f"{'turns':>6}  {'ctx size':>10}  "
+        f"{'add_msgs (bytes)':>18}  {'delta (bytes)':>15}  {'delta+snap (bytes)':>18}  "
+        f"{'storage saved':>14}  "
+        f"{'read: add_msgs':>14}  {'read: delta+snap':>16}"
+    )
+    print(header)
+    print("-" * W)
+
+    results = []
+    for turns in TURN_COUNTS:
+        with _make_saver() as saver:
+            b_wt, b_rt, b_bytes = _run_turns(turns, BinaryState, saver)
+        with _make_saver() as saver:
+            d_wt, d_rt, d_bytes = _run_turns(turns, DeltaState, saver)
+        with _make_saver() as saver:
+            s_wt, s_rt, s_bytes = _run_turns(turns, DeltaSnapshotState, saver)
+
+        # For non-InMemory savers, blob_bytes are unavailable (-1); use read times only
+        if b_bytes < 0 or s_bytes < 0:
+            b_bytes_str = "n/a"
+            d_bytes_str = "n/a"
+            s_bytes_str = "n/a"
+            storage_ratio_str = "n/a"
+        else:
+            storage_ratio = b_bytes / s_bytes if s_bytes else float("inf")
+            b_bytes_str = _fmt_bytes(b_bytes)
+            d_bytes_str = _fmt_bytes(d_bytes)
+            s_bytes_str = _fmt_bytes(s_bytes)
+            storage_ratio_str = f"{storage_ratio:.1f}x"
+            results.append((turns, b_bytes, s_bytes, b_rt, s_rt, storage_ratio))
+
+        print(
+            f"{turns:>6}  {_approx_tokens(turns):>10}  "
+            f"{b_bytes_str:>18}  {d_bytes_str:>15}  {s_bytes_str:>18}  "
+            f"{storage_ratio_str:>14}  "
+            f"{b_rt * 1000:>12.1f}ms  {s_rt * 1000:>14.1f}ms"
+        )
+
+    print("=" * W)
+    print()
+
+    if results:
+        best = results[-1]
+        turns, b_bytes, s_bytes, b_rt, s_rt, ratio = best
+        print(f"Key findings at max scale ({turns} turns):")
+        print(
+            f"  Storage:  {_fmt_bytes(b_bytes)} (add_messages)  →  {_fmt_bytes(s_bytes)} (DeltaChannel+snapshot) — {ratio:.0f}x reduction"
+        )
+        print(
+            f"  Read latency: {b_rt * 1000:.1f}ms (add_messages)  vs  {s_rt * 1000:.1f}ms (DeltaChannel+snapshot)"
+        )
+        print()
+    print("Legend:")
+    print(
+        "  add_msgs      = Annotated[list, add_messages]  — current default, O(N²) storage"
+    )
+    print(
+        "  delta         = DeltaChannel(add_messages)     — O(N) storage, unbounded chain at read"
+    )
+    print(
+        f"  delta+snap    = DeltaChannel(add_messages, snapshot_every={SNAPSHOT_EVERY})  — O(N) storage, O(1) read depth"
+    )
+    print()
+
+
+# ---------------------------------------------------------------------------
+# Pytest entry point
+# ---------------------------------------------------------------------------
+
+
+def test_delta_channel_benchmark(capsys: Any) -> None:
+    """Storage grows O(N²) for add_messages, O(N) for DeltaChannel."""
+    with capsys.disabled():
+        run_benchmark()
+
+    # Correctness assertion: DeltaChannel must use less storage at scale.
+    for turns in [100, 200]:
+        _, _, b_bytes = _run_turns(turns, BinaryState)
+        _, _, d_bytes = _run_turns(turns, DeltaState)
+        _, _, s_bytes = _run_turns(turns, DeltaSnapshotState)
+        assert d_bytes < b_bytes, (
+            f"DeltaChannel should use less storage at {turns} turns, "
+            f"got delta={d_bytes} binary={b_bytes}"
+        )
+        assert s_bytes < b_bytes, (
+            f"DeltaChannel+snapshot should use less storage at {turns} turns, "
+            f"got snapshot={s_bytes} binary={b_bytes}"
+        )
+
+
+# ---------------------------------------------------------------------------
+# Script entry point
+# ---------------------------------------------------------------------------
+
+if __name__ == "__main__":
+    run_benchmark()
+    sys.exit(0)

libs/langgraph/tests/test_pregel.py

@@ -9400,3 +9400,190 @@ def test_fork_does_not_apply_pending_writes(
 
     # Should be: 1 (input) + 20 (forked node_a) + 100 (node_b) = 121
     assert result == {"value": 121}
+
+
+async def test_delta_channel_end_to_end_inmemory() -> None:
+    """Full graph run: DeltaChannel accumulates correctly across multiple turns."""
+    from langchain_core.messages import AIMessage, HumanMessage
+    from langgraph.checkpoint.memory import InMemorySaver
+
+    from langgraph.channels.delta import DeltaChannel
+    from langgraph.graph import START, StateGraph
+    from langgraph.graph.message import add_messages
+
+    class State(TypedDict):
+        messages: Annotated[list, DeltaChannel(add_messages)]
+
+    def respond(state: State) -> dict:
+        n = len(state["messages"])
+        return {"messages": [AIMessage(content=f"reply-{n}", id=f"ai-{n}")]}
+
+    builder = StateGraph(State)
+    builder.add_node("respond", respond)
+    builder.add_edge(START, "respond")
+    graph = builder.compile(checkpointer=InMemorySaver())
+
+    config = {"configurable": {"thread_id": "diff-test-1"}}
+
+    # Turn 1
+    graph.invoke({"messages": [HumanMessage(content="hello", id="h1")]}, config)
+    # Turn 2
+    graph.invoke({"messages": [HumanMessage(content="world", id="h2")]}, config)
+    # Turn 3
+    graph.invoke({"messages": [HumanMessage(content="bye", id="h3")]}, config)
+
+    state = graph.get_state(config)
+    msgs = state.values["messages"]
+    # 3 human + 3 AI = 6 total
+    assert len(msgs) == 6, f"expected 6 messages, got {len(msgs)}: {msgs}"
+    assert msgs[0].content == "hello"
+    assert msgs[2].content == "world"
+    assert msgs[4].content == "bye"
+    assert msgs[1].content == "reply-1"
+    assert msgs[3].content == "reply-3"
+    assert msgs[5].content == "reply-5"
+
+
+async def test_delta_channel_time_travel() -> None:
+    """Time-travel back to turn-1 checkpoint and resume; continuation must not include turn-2 deltas."""
+    from langchain_core.messages import AIMessage, HumanMessage
+    from langgraph.checkpoint.memory import InMemorySaver
+
+    from langgraph.channels.delta import DeltaChannel
+    from langgraph.graph import START, StateGraph
+    from langgraph.graph.message import add_messages
+
+    class State(TypedDict):
+        messages: Annotated[list, DeltaChannel(add_messages)]
+
+    counter = {"n": 0}
+
+    def respond(state: State) -> dict:
+        counter["n"] += 1
+        return {
+            "messages": [
+                AIMessage(content=f"ai-{counter['n']}", id=f"ai-{counter['n']}")
+            ]
+        }
+
+    builder = StateGraph(State)
+    builder.add_node("respond", respond)
+    builder.add_edge(START, "respond")
+    saver = InMemorySaver()
+    graph = builder.compile(checkpointer=saver)
+
+    config = {"configurable": {"thread_id": "diff-time-travel"}}
+
+    # Run 2 turns: h1→ai-1, h2→ai-2
+    graph.invoke({"messages": [HumanMessage(content="h1", id="h1")]}, config)
+    graph.invoke({"messages": [HumanMessage(content="h2", id="h2")]}, config)
+
+    # Find the checkpoint after turn 1 (2 messages: h1 + ai-1)
+    history = list(graph.get_state_history(config))
+    after_turn1 = next(h for h in history if len(h.values.get("messages", [])) == 2)
+
+    assert len(after_turn1.values["messages"]) == 2
+    assert after_turn1.values["messages"][0].content == "h1"
+    assert after_turn1.values["messages"][1].content == "ai-1"
+
+    # Resume from turn-1 checkpoint: inject h3, expect 3 messages total (h1, ai-1, ai-N)
+    # NOT 5 messages (turn-2 deltas must not bleed into the resumed run)
+    result = graph.invoke(
+        {"messages": [HumanMessage(content="h3", id="h3")]},
+        after_turn1.config,
+    )
+    msgs = result["messages"]
+    # Should be: h1, ai-1, h3, ai-N — 4 messages total
+    assert len(msgs) == 4, (
+        f"expected 4 messages after time-travel resume, got {len(msgs)}: {msgs}"
+    )
+    assert msgs[0].content == "h1"
+    assert msgs[1].content == "ai-1"
+    assert msgs[2].content == "h3"
+
+
+async def test_delta_channel_remove_message_end_to_end() -> None:
+    """RemoveMessage inside a DeltaChannel graph must persist and reload correctly."""
+    from langchain_core.messages import AIMessage, HumanMessage, RemoveMessage
+    from langgraph.checkpoint.memory import InMemorySaver
+
+    from langgraph.channels.delta import DeltaChannel
+    from langgraph.graph import START, StateGraph
+    from langgraph.graph.message import add_messages
+
+    class State(TypedDict):
+        messages: Annotated[list, DeltaChannel(add_messages)]
+
+    def respond(state: State) -> dict:
+        return {"messages": [AIMessage(content="reply", id="ai-1")]}
+
+    def delete_first(state: State) -> dict:
+        # removes the first message
+        return {"messages": [RemoveMessage(id=state["messages"][0].id)]}
+
+    builder = StateGraph(State)
+    builder.add_node("respond", respond)
+    builder.add_node("delete_first", delete_first)
+    builder.add_edge(START, "respond")
+    builder.add_edge("respond", "delete_first")
+    graph = builder.compile(checkpointer=InMemorySaver())
+
+    config = {"configurable": {"thread_id": "diff-remove-test"}}
+    graph.invoke({"messages": [HumanMessage(content="hello", id="h1")]}, config)
+
+    state = graph.get_state(config)
+    msgs = state.values["messages"]
+    # h1 was removed, only ai-1 should remain
+    assert len(msgs) == 1, f"expected 1 message, got {len(msgs)}: {msgs}"
+    assert msgs[0].id == "ai-1"
+
+    # A subsequent turn must reconstruct from the checkpoint correctly
+    graph.invoke({"messages": [HumanMessage(content="again", id="h2")]}, config)
+    state = graph.get_state(config)
+    msgs = state.values["messages"]
+    # ai-1 + h2 + ai-1(second reply, same id overwrites) + h2 removed
+    # more simply: after second run we expect ai-1 updated + h2 remaining minus deleted h2
+    # just assert h1 is still gone
+    assert all(m.id != "h1" for m in msgs), (
+        "h1 should still be absent after second turn"
+    )
+
+
+async def test_delta_channel_update_by_id_end_to_end() -> None:
+    """Updating a message by ID via DeltaChannel must persist and reload correctly."""
+    from langchain_core.messages import HumanMessage
+    from langgraph.checkpoint.memory import InMemorySaver
+
+    from langgraph.channels.delta import DeltaChannel
+    from langgraph.graph import START, StateGraph
+    from langgraph.graph.message import add_messages
+
+    class State(TypedDict):
+        messages: Annotated[list, DeltaChannel(add_messages)]
+
+    def update_msg(state: State) -> dict:
+        # re-send h1 with updated content
+        return {"messages": [HumanMessage(content="updated", id="h1")]}
+
+    builder = StateGraph(State)
+    builder.add_node("update_msg", update_msg)
+    builder.add_edge(START, "update_msg")
+    graph = builder.compile(checkpointer=InMemorySaver())
+
+    config = {"configurable": {"thread_id": "diff-update-id-test"}}
+    graph.invoke({"messages": [HumanMessage(content="original", id="h1")]}, config)
+
+    state = graph.get_state(config)
+    msgs = state.values["messages"]
+    assert len(msgs) == 1, f"expected 1 message, got {len(msgs)}: {msgs}"
+    assert msgs[0].content == "updated"
+    assert msgs[0].id == "h1"
+
+    # Second turn: verify the updated state is the base for further accumulation
+    graph.invoke({"messages": [HumanMessage(content="new", id="h2")]}, config)
+    state = graph.get_state(config)
+    msgs = state.values["messages"]
+    ids = [m.id for m in msgs]
+    assert "h1" in ids  # h1 persists (updated, not duplicated)
+    assert "h2" in ids
+    assert ids.count("h1") == 1, "h1 must not be duplicated"

libs/langgraph/tests/test_rehydrate_sweep.py

@@ -0,0 +1,185 @@
+"""Sweep snapshot_every values to find the storage vs. time-travel tradeoff.
+
+Run directly:  python tests/test_rehydrate_sweep.py
+Run via pytest: pytest tests/test_rehydrate_sweep.py -s
+"""
+
+from __future__ import annotations
+
+import sys
+import time
+from typing import Annotated, Any
+
+from langchain_core.messages import AIMessage, HumanMessage
+from langgraph.checkpoint.memory import MemorySaver
+from typing_extensions import TypedDict
+
+from langgraph.channels.delta import DeltaChannel
+from langgraph.graph import END, StateGraph
+from langgraph.graph.message import add_messages
+
+# ---------------------------------------------------------------------------
+# Config
+# ---------------------------------------------------------------------------
+
+REHYDRATE_SWEEP = [5, 10, 25, 50, 100, None]  # None = no rehydration (pure diff)
+TURN_COUNTS = [50, 100, 250, 500]
+
+
+# ---------------------------------------------------------------------------
+# Helpers
+# ---------------------------------------------------------------------------
+
+
+def _make_state(snapshot_every: int | None) -> type:
+    channel = DeltaChannel(add_messages, snapshot_every=snapshot_every)
+    return TypedDict("S", {"messages": Annotated[list, channel]})
+
+
+def _make_graph(state_cls: type) -> Any:
+    def human_node(state: Any) -> dict:
+        return {}
+
+    def ai_node(state: Any) -> dict:
+        last = state["messages"][-1]
+        return {"messages": [AIMessage(content=f"reply-to-{last.id}")]}
+
+    g = StateGraph(state_cls)
+    g.add_node("human", human_node)
+    g.add_node("ai", ai_node)
+    g.add_edge("human", "ai")
+    g.add_edge("ai", END)
+    g.set_entry_point("human")
+    return g.compile(checkpointer=MemorySaver())
+
+
+def _total_blob_bytes(saver: MemorySaver) -> int:
+    total = 0
+    for (_, _, _, _), (type_tag, blob) in saver.blobs.items():
+        if blob is not None:
+            total += len(blob)
+    return total
+
+
+def _measure_time_travel_ms(graph: Any, config: dict) -> float:
+    """Time how long it takes to get state at the very first checkpoint (worst case)."""
+    history = list(graph.get_state_history(config))
+    if not history:
+        return 0.0
+    oldest = history[-1]
+    t0 = time.perf_counter()
+    graph.get_state(oldest.config)
+    return (time.perf_counter() - t0) * 1000
+
+
+def _run(n_turns: int, snapshot_every: int | None) -> tuple[float, int, float]:
+    """Returns (write_ms, blob_bytes, time_travel_ms)."""
+    state_cls = _make_state(snapshot_every)
+    graph = _make_graph(state_cls)
+    saver: MemorySaver = graph.checkpointer  # type: ignore[assignment]
+    config = {"configurable": {"thread_id": "sweep"}}
+
+    t0 = time.perf_counter()
+    for i in range(n_turns):
+        graph.invoke(
+            {"messages": [HumanMessage(content=f"msg-{i}", id=f"h{i}")]}, config
+        )
+    write_ms = (time.perf_counter() - t0) * 1000
+
+    blob_bytes = _total_blob_bytes(saver)
+    tt_ms = _measure_time_travel_ms(graph, config)
+    return write_ms, blob_bytes, tt_ms
+
+
+# ---------------------------------------------------------------------------
+# ASCII sparkline
+# ---------------------------------------------------------------------------
+
+
+def _sparkline(values: list[float], width: int = 20) -> str:
+    bars = " ▁▂▃▄▅▆▇█"
+    lo, hi = min(values), max(values)
+    span = hi - lo or 1
+    chars = [bars[round((v - lo) / span * (len(bars) - 1))] for v in values]
+    return "".join(chars).ljust(width)
+
+
+# ---------------------------------------------------------------------------
+# Main
+# ---------------------------------------------------------------------------
+
+
+def run_sweep() -> None:
+    label = {v: (str(v) if v is not None else "None(∞)") for v in REHYDRATE_SWEEP}
+
+    print()
+    print("snapshot_every sweep — storage vs time-travel cost")
+    print("=" * 90)
+
+    for turns in TURN_COUNTS:
+        print(f"\n--- {turns} turns ---")
+        col_w = 12
+        header = (
+            f"{'snapshot_every':>18}  "
+            f"{'blob_bytes':>{col_w}}  "
+            f"{'write_ms':>{col_w}}  "
+            f"{'time_travel_ms':>{col_w}}"
+        )
+        print(header)
+        print("-" * 60)
+
+        tt_vals: list[float] = []
+        byte_vals: list[int] = []
+        write_vals: list[float] = []
+        rows: list[tuple] = []
+
+        for rv in REHYDRATE_SWEEP:
+            write_ms, blob_bytes, tt_ms = _run(turns, rv)
+            rows.append((rv, blob_bytes, write_ms, tt_ms))
+            byte_vals.append(blob_bytes)
+            write_vals.append(write_ms)
+            tt_vals.append(tt_ms)
+
+        for rv, blob_bytes, write_ms, tt_ms in rows:
+            print(
+                f"{label[rv]:>18}  "
+                f"{blob_bytes:>{col_w},}  "
+                f"{write_ms:>{col_w}.1f}  "
+                f"{tt_ms:>{col_w}.2f}"
+            )
+
+        print()
+        print(
+            f"  bytes       spark: [{_sparkline(byte_vals)}]  "
+            f"lo={min(byte_vals):,}  hi={max(byte_vals):,}"
+        )
+        print(
+            f"  time-travel spark: [{_sparkline(tt_vals)}]  "
+            f"lo={min(tt_vals):.2f}ms  hi={max(tt_vals):.2f}ms"
+        )
+        print(
+            f"  write       spark: [{_sparkline(write_vals)}]  "
+            f"lo={min(write_vals):.1f}ms  hi={max(write_vals):.1f}ms"
+        )
+
+    print()
+    print("=" * 90)
+    print(
+        "snapshot_every=None means pure diff (no snapshots) — "
+        "lowest storage, highest time-travel cost."
+    )
+    print(
+        "Lower snapshot_every = more frequent full snapshots = "
+        "faster time-travel, more storage."
+    )
+    print()
+
+
+def test_rehydrate_sweep(capsys: Any) -> None:
+    with capsys.disabled():
+        run_sweep()
+
+
+if __name__ == "__main__":
+    run_sweep()
+    sys.exit(0)

libs/prebuilt/langgraph/prebuilt/tool_node.py

@@ -614,7 +614,6 @@ class _InjectedArgs:
     store: str | None
     runtime: str | None
     all_injected_keys: set[str]
-    _optional_state_args: set[str]
 
 
 class ToolNode(RunnableCallable):
@@ -1334,7 +1333,7 @@ class ToolNode(RunnableCallable):
             return tool_call
 
         tool_call_copy: ToolCall = copy(tool_call)
-        injected_args: dict[str, Any] = {}
+        injected_args = {}
 
         # Inject state
         if injected.state:
@@ -1362,20 +1361,14 @@ class ToolNode(RunnableCallable):
             # Extract state values
             if isinstance(state, dict):
                 for tool_arg, state_field in injected.state.items():
-                    if not state_field:
+                    injected_args[tool_arg] = (
-                        injected_args[tool_arg] = state
+                        state[state_field] if state_field else state
-                    elif state_field in state:
+                    )
-                        injected_args[tool_arg] = state[state_field]
-                    elif tool_arg not in injected._optional_state_args:
-                        raise KeyError(state_field)
             else:
                 for tool_arg, state_field in injected.state.items():
-                    if not state_field:
+                    injected_args[tool_arg] = (
-                        injected_args[tool_arg] = state
+                        getattr(state, state_field) if state_field else state
-                    elif hasattr(state, state_field):
+                    )
-                        injected_args[tool_arg] = getattr(state, state_field)
-                    elif tool_arg not in injected._optional_state_args:
-                        raise AttributeError(state_field)
 
         # Inject store
         if injected.store:
@@ -1866,7 +1859,6 @@ def _get_all_injected_args(tool: BaseTool) -> _InjectedArgs:
     store_arg: str | None = None
     runtime_arg: str | None = None
     all_injected_keys: set[str] = set()
-    _optional_state_args: set[str] = set()
 
     for name, type_ in all_annotations.items():
         # Track all InjectedToolArg-annotated params (including custom subclasses)
@@ -1881,9 +1873,6 @@ def _get_all_injected_args(tool: BaseTool) -> _InjectedArgs:
         if state_inj := _get_injection_from_type(type_, InjectedState):
             if isinstance(state_inj, InjectedState) and state_inj.field:
                 state_args[name] = state_inj.field
-                field_info = full_schema.model_fields.get(name)
-                if field_info and not field_info.is_required():
-                    _optional_state_args.add(name)
             else:
                 state_args[name] = None
 
@@ -1900,5 +1889,4 @@ def _get_all_injected_args(tool: BaseTool) -> _InjectedArgs:
         store=store_arg,
         runtime=runtime_arg,
         all_injected_keys=all_injected_keys,
-        _optional_state_args=_optional_state_args,
     )

libs/sdk-py/langgraph_sdk/__init__.py

@@ -1,8 +1,30 @@
-from langgraph_sdk.auth import Auth
+from __future__ import annotations
-from langgraph_sdk.client import get_client, get_sync_client
+
-from langgraph_sdk.encryption import Encryption
+import importlib
-from langgraph_sdk.encryption.types import EncryptionContext
+from typing import TYPE_CHECKING
+
+if TYPE_CHECKING:
+    from langgraph_sdk.auth import Auth
+    from langgraph_sdk.client import get_client, get_sync_client
+    from langgraph_sdk.encryption import Encryption
+    from langgraph_sdk.encryption.types import EncryptionContext
 
 __version__ = "0.3.13"
 
 __all__ = ["Auth", "Encryption", "EncryptionContext", "get_client", "get_sync_client"]
+
+_LAZY: dict[str, str] = {
+    "Auth": "langgraph_sdk.auth",
+    "get_client": "langgraph_sdk.client",
+    "get_sync_client": "langgraph_sdk.client",
+    "Encryption": "langgraph_sdk.encryption",
+    "EncryptionContext": "langgraph_sdk.encryption.types",
+}
+
+
+def __getattr__(name: str) -> object:
+    if name in _LAZY:
+        mod = importlib.import_module(_LAZY[name])
+        return getattr(mod, name)
+    msg = f"module {__name__!r} has no attribute {name!r}"
+    raise AttributeError(msg)