@@ -19,6 +19,7 @@
 from __future__ import annotations
 
 import asyncio
+import contextlib
 import copy
 import hashlib
 import json as json_module
@@ -47,6 +48,7 @@
 from ._cooperative_cancellation import CancellationDelivery, CancellationRequest
 from ._query_task_authority import QueryTaskAuthority, current_query_authority
 from .errors import (
+    REQUEST_ID_CONFLICT,
     ExternalPayloadIntegrityMismatch,
     ExternalPayloadOversized,
     ExternalPayloadUnavailable,
@@ -55,6 +57,8 @@
     RuntimeCapabilityUnsupported,
     RuntimeDiscoveryUnavailable,
     ServerError,
+    SignalFailed,
+    UpdateFailed,
     WorkflowCancelled,
     WorkflowFailed,
     WorkflowTerminated,
@@ -239,6 +243,49 @@ def _normalize_base_url(base_url: str) -> str:
     return base_url.rstrip("/")
 
 
+def _command_request_id(request_id: str | None, command: str) -> str:
+    """Return the caller command key for one logical signal or update call.
+
+    The key is the body field ``request_id`` (1-255 opaque, case-sensitive
+    bytes). A caller-supplied key is sent verbatim; otherwise one
+    ``sdk-python-<command>-<uuid4 hex>`` key is generated per logical call so
+    the transport retry policy resends the identical key.
+    """
+    return request_id or f"sdk-python-{command}-{uuid.uuid4().hex}"
+
+
+def _attach_request_id(exc: BaseException, request_id: str) -> None:
+    """Record the sent command key on an exception as ``exc.request_id``.
+
+    Typed SDK failures declare the attribute; transport errors raised by
+    ``httpx`` receive it dynamically so a manual retry can reuse the key.
+    """
+    with contextlib.suppress(AttributeError, TypeError):
+        exc.request_id = request_id  # type: ignore[attr-defined]
+
+
+def _command_request_failure(exc: Exception, *, command: str, request_id: str) -> Exception:
+    """Return the exception a failed signal or update call should raise.
+
+    A ``request_id_conflict`` refusal that reached the caller as a generic
+    :class:`ServerError` becomes the command's typed failure. Every other
+    exception is returned unchanged, annotated with the key that was sent.
+    """
+    if isinstance(exc, ServerError) and exc.reason() == REQUEST_ID_CONFLICT:
+        message = exc.body.get("message") if isinstance(exc.body, dict) else None
+        if not isinstance(message, str) or not message:
+            message = f"{command} request id already retained with a different intent"
+        if command == "update" and isinstance(exc.body, dict):
+            return UpdateFailed(
+                message, status=exc.status, body=exc.body, reason=REQUEST_ID_CONFLICT, request_id=request_id,
+            )
+        return SignalFailed(
+            message, reason=REQUEST_ID_CONFLICT, status=exc.status, body=exc.body, request_id=request_id,
+        )
+    _attach_request_id(exc, request_id)
+    return exc
+
+
 def _route_for_metrics(path: str) -> str:
     clean_path = path.split("?", 1)[0]
     parts = [part for part in clean_path.strip("/").split("/") if part]
@@ -1346,9 +1393,17 @@ async def describe_run(self, run_id: str | None = None) -> WorkflowRun:
             raise ValueError("run_id is required to describe a workflow run from a handle")
         return await self._client.describe_workflow_run(self.workflow_id, selected_run_id)
 
-    async def signal(self, signal_name: str, args: list[Any] | None = None) -> None:
+    async def signal(
+        self,
+        signal_name: str,
+        args: list[Any] | None = None,
+        *,
+        request_id: str | None = None,
+    ) -> dict[str, Any]:
         """Deliver an external signal to this workflow. See :meth:`Client.signal_workflow`."""
-        await self._client.signal_workflow(self.workflow_id, signal_name, args=args)
+        return await self._client.signal_workflow(
+            self.workflow_id, signal_name, args=args, request_id=request_id,
+        )
 
     async def append_message(
         self,
@@ -4303,15 +4358,38 @@ async def send_webhook_bridge_event(
         return BridgeAdapterOutcome.from_dict(data)
 
     async def signal_workflow(
-        self, workflow_id: str, signal_name: str, *, args: list[Any] | None = None
-    ) -> None:
+        self,
+        workflow_id: str,
+        signal_name: str,
+        *,
+        args: list[Any] | None = None,
+        request_id: str | None = None,
+    ) -> dict[str, Any]:
         """Deliver an external signal to a running workflow.
 
         Signals are fire-and-forget: the server records the signal in durable
         history and returns immediately. They do not wait for the workflow to
         observe the signal. See the main docs for how to declare the allowed
         signal names on a workflow type.
+
+        ``request_id`` is the caller command key (1-255 opaque, case-sensitive
+        bytes) that makes a retry of this signal safe. When omitted the client
+        generates ``sdk-python-signal-<uuid>`` once per call, so the transport
+        retry policy resends the same key. Resending the same key with the same
+        signal name, arguments and run target returns the original receipt
+        instead of recording a second signal.
+
+        Returns the Server's admission receipt. ``request_id`` is always
+        present (the key that was sent); a Server that retains command keys
+        also reports ``retained`` (``True`` when an earlier admission was
+        returned) and ``intent_verified``. Raises
+        :class:`~durable_workflow.errors.SignalFailed` with
+        ``reason == "request_id_conflict"`` when the key was already retained
+        for a signal with a different intent; the retained command is described
+        in its ``details``. Every exception raised after the key was chosen,
+        including transport errors, carries it as ``request_id``.
         """
+        key = _command_request_id(request_id, "signal")
         body: dict[str, Any] = {}
         if args:
             body["input"] = self._payload_envelope(
@@ -4320,7 +4398,19 @@ async def signal_workflow(
                 workflow_id=workflow_id,
                 signal_name=signal_name,
             )
-        await self._request("POST", f"/workflows/{workflow_id}/signal/{signal_name}", json=body, context=workflow_id)
+        body["request_id"] = key
+        try:
+            data = await self._request(
+                "POST", f"/workflows/{workflow_id}/signal/{signal_name}", json=body, context=workflow_id,
+            )
+        except Exception as exc:
+            failure = _command_request_failure(exc, command="signal", request_id=key)
+            if failure is exc:
+                raise
+            raise failure from exc
+        receipt: dict[str, Any] = data if isinstance(data, dict) else {}
+        receipt.setdefault("request_id", key)
+        return receipt
 
     async def append_message_stream(
         self,
@@ -4496,15 +4586,28 @@ async def update_workflow(
         Validator rejection and validation-routing failures remain typed and
         do not silently fall back to post-accept handler failure.
 
-        ``request_id`` lets the caller deduplicate retries. Raises
-        :class:`~durable_workflow.errors.UpdateRejected` when the workflow's
-        validator rejects the update, or
+        ``request_id`` is the caller command key (1-255 opaque, case-sensitive
+        bytes) that makes a retry of this update safe. When omitted the client
+        generates ``sdk-python-update-<uuid>`` once per call, so the transport
+        retry policy resends the same key. Resending the same key with the same
+        update name, arguments and run target returns the original receipt
+        (``retained`` is ``True`` on a Server that reports it) instead of
+        admitting a second update. The returned response always carries the
+        ``request_id`` that was sent.
+
+        Raises :class:`~durable_workflow.errors.UpdateRejected` when the
+        workflow's validator rejects the update, or
         :class:`~durable_workflow.errors.UpdateValidationFailed` when the
         declared validation boundary cannot be enforced. An accepted update
         whose handler fails raises :class:`~durable_workflow.errors.UpdateFailed`
-        with the failure message, Server response and durable identities.
+        with the failure message, Server response and durable identities; the
+        same type with ``reason == "request_id_conflict"`` reports a key that
+        was already retained for an update with a different intent, described
+        in its ``details``. Every exception raised after the key was chosen,
+        including transport errors, carries it as ``request_id``.
         """
         await self._require_update_wait_stage(wait_for or "accepted")
+        key = _command_request_id(request_id, "update")
         body: dict[str, Any] = {}
         if args:
             body["input"] = self._payload_envelope(
@@ -4517,10 +4620,19 @@ async def update_workflow(
             body["wait_for"] = wait_for
         if wait_timeout_seconds is not None:
             body["wait_timeout_seconds"] = wait_timeout_seconds
-        body["request_id"] = request_id or f"sdk-python-update-{uuid.uuid4().hex}"
-        return await self._request(
-            "POST", f"/workflows/{workflow_id}/update/{update_name}", json=body, context=workflow_id
-        )
+        body["request_id"] = key
+        try:
+            data = await self._request(
+                "POST", f"/workflows/{workflow_id}/update/{update_name}", json=body, context=workflow_id
+            )
+        except Exception as exc:
+            failure = _command_request_failure(exc, command="update", request_id=key)
+            if failure is exc:
+                raise
+            raise failure from exc
+        if isinstance(data, dict):
+            data.setdefault("request_id", key)
+        return data
 
     async def get_result(
         self,
 
@@ -412,8 +412,45 @@ def validation_errors(self) -> dict[str, Any] | None:
         return None
 
 
+REQUEST_ID_CONFLICT = "request_id_conflict"
+"""Reason reported when a command key is reused with a different intent."""
+
+_REQUEST_ID_CONFLICT_FIELDS = (
+    "request_id",
+    "retained_command_id",
+    "retained_command_type",
+    "retained_outcome",
+    "conflict",
+)
+
+
+def _body_string(body: object, name: str) -> str | None:
+    if isinstance(body, dict):
+        value = body.get(name)
+        if isinstance(value, str) and value:
+            return value
+    return None
+
+
+def _request_id_conflict_details(body: object) -> dict[str, Any]:
+    """Return the retained-command fields of a ``request_id_conflict`` response body."""
+    if not isinstance(body, dict) or body.get("reason") != REQUEST_ID_CONFLICT:
+        return {}
+    return {name: body[name] for name in _REQUEST_ID_CONFLICT_FIELDS if name in body}
+
+
 class SignalFailed(DurableWorkflowError):
-    """A workflow signal was rejected before it could be delivered."""
+    """A workflow signal was rejected before it could be delivered.
+
+    :attr:`request_id` is the caller command key the client sent with the
+    signal (caller-supplied or SDK-generated), so the application can retry
+    the same logical signal with the same key. :attr:`reason` is the Server's
+    machine-readable refusal reason; ``request_id_conflict`` means the key was
+    already retained for a signal with a different intent, and
+    :attr:`details` then carries ``retained_command_id``,
+    ``retained_command_type``, ``retained_outcome`` and the mismatched
+    ``conflict`` parts.
+    """
 
     def __init__(
         self,
@@ -422,11 +459,15 @@ def __init__(
         reason: str | None = None,
         status: int | None = None,
         body: object | None = None,
+        request_id: str | None = None,
+        details: dict[str, Any] | None = None,
     ) -> None:
         super().__init__(message)
         self.reason = reason
         self.status = status
         self.body = body
+        self.request_id = request_id if request_id is not None else _body_string(body, "request_id")
+        self.details: dict[str, Any] = dict(details) if details is not None else _request_id_conflict_details(body)
 
     @property
     def validation_errors(self) -> dict[str, Any] | None:
@@ -508,17 +549,37 @@ def __init__(
 
 
 class UpdateFailed(InvalidArgument):
-    """An accepted workflow update failed during handler execution.
+    """An accepted workflow update failed, or its command key was refused.
 
     ``body`` preserves the Server response, including its durable failure
     identity. This inherits from :class:`InvalidArgument` to preserve catches
     written for earlier SDK versions that mapped all HTTP 422 responses there.
+
+    :attr:`request_id` is the caller command key the client sent with the
+    update (caller-supplied or SDK-generated). :attr:`reason` is the Server's
+    machine-readable reason when it returned one; ``request_id_conflict``
+    means the key was already retained for an update with a different intent,
+    and :attr:`details` then carries ``retained_command_id``,
+    ``retained_command_type``, ``retained_outcome`` and the mismatched
+    ``conflict`` parts.
     """
 
-    def __init__(self, message: str, *, status: int, body: dict[str, Any]) -> None:
+    def __init__(
+        self,
+        message: str,
+        *,
+        status: int,
+        body: dict[str, Any],
+        reason: str | None = None,
+        request_id: str | None = None,
+        details: dict[str, Any] | None = None,
+    ) -> None:
         super().__init__(message)
         self.status = status
         self.body = body
+        self.reason = reason if reason is not None else self._identity(body, "reason")
+        self.request_id = request_id if request_id is not None else self._identity(body, "request_id")
+        self.details: dict[str, Any] = dict(details) if details is not None else _request_id_conflict_details(body)
         self.workflow_id = self._identity(body, "workflow_id")
         self.run_id = self._identity(body, "run_id")
         self.update_id = self._identity(body, "update_id")
@@ -909,6 +970,18 @@ def update_validation_failed(default: str) -> UpdateValidationFailed:
             raise UpdateRejected(message or "update rejected", body=body)
         if reason in update_validation_failure_reasons:
             raise update_validation_failed("update validation failed")
+        if reason == REQUEST_ID_CONFLICT:
+            # The retained command type names the key scope; the Server also
+            # labels the control-plane operation. Either selects the typed failure.
+            command_type = _body_string(body, "retained_command_type") or operation
+            if command_type == "signal":
+                raise signal_failed("signal request id already retained with a different intent")
+            if command_type == "update" and isinstance(body, dict):
+                raise UpdateFailed(
+                    message or "update request id already retained with a different intent",
+                    status=status,
+                    body=body,
+                )
         raise ServerError(status, body)
 
     if status == 504:
 
@@ -3,7 +3,8 @@
 .. warning::
 
    :class:`TransportRetryPolicy` covers **only client-side HTTP retries** for
-   transient transport errors (connection failures, timeouts, 5xx responses,
+   transient transport errors (connection failures, timeouts, a server or
+   proxy that disconnects before sending a response, 5xx responses,
    429 rate-limiting). It is **not** the activity retry policy. Activity-level
    retry and timeout configuration lives on
    :class:`durable_workflow.workflow.ActivityRetryPolicy` and is passed to
@@ -201,9 +202,16 @@ class TransportRetryPolicy:
     """
     Retry policy for transient HTTP transport errors.
 
-    Retries requests that fail with transient errors (connection errors,
-    timeouts, 5xx server errors, 429 rate limit). Does not retry client
-    errors (4xx except 429).
+    Retries requests that fail with transient errors: ``httpx.NetworkError``
+    (including ``ConnectError``), ``httpx.TimeoutException``,
+    ``httpx.RemoteProtocolError`` (the server or a proxy closed the connection
+    before sending a response), 5xx server errors and 429 rate limit. Does not
+    retry client errors (4xx except 429) or ``httpx.LocalProtocolError``.
+
+    A retried request is resent byte-identical. Signals and updates carry a
+    durable ``request_id`` per logical call, so a retry after a lost response
+    is answered from the retained receipt instead of admitting a second
+    command.
 
     This policy runs inside :class:`~durable_workflow.Client` around HTTP
     requests. It does not retry workflow runs, workflow tasks, activity
@@ -228,8 +236,12 @@ def should_retry(self, exc: Exception, attempt: int) -> bool:
         if attempt >= self.max_attempts:
             return False
 
-        # Retry connection errors and timeouts
-        if isinstance(exc, httpx.ConnectError | httpx.TimeoutException | httpx.NetworkError):
+        # Retry connection errors, timeouts and a peer that disconnected before
+        # answering (a lost response is recovered through the request's identity).
+        if isinstance(
+            exc,
+            httpx.ConnectError | httpx.TimeoutException | httpx.NetworkError | httpx.RemoteProtocolError,
+        ):
             return True
 
         # Retry 5xx server errors and 429 rate limit