From 2daa3f83fd8ad4745e04f1740a1028512979d55c Mon Sep 17 00:00:00 2001
From: Xingyao Wang <xingyaoww@gmail.com>
Date: Wed, 10 Jun 2026 15:58:07 +0000
Subject: [PATCH] fix(llm): serialize is_subscription so subscription mode
 survives remote transport
MIME-Version: 1.0
Content-Type: text/plain; charset=UTF-8
Content-Transfer-Encoding: 8bit

`is_subscription` was a plain property backed by the `_is_subscription`
PrivateAttr, so it was silently dropped whenever an LLM was serialized — in
particular when a RemoteConversation ships the agent's LLM to an agent-server.
The server rebuilt the LLM with is_subscription == False, and every
subscription-specific behavior stopped applying server-side:

- the streaming on_token exemption in responses()
  (ValueError: Streaming requires an on_token callback),
- the Codex system-prompt transform (transform_for_subscription),
- reasoning-item stripping for store=false (follow-up requests reference
  unresolvable item IDs -> 404 from the Codex endpoint).

Net effect: LLM.subscription_login() worked locally but silently broke with
remote workspaces.

Promote the property to a `@computed_field` (so it is included in model_dump)
and add a wrap model_validator that restores the backing `_is_subscription`
PrivateAttr on validation. The public read API (`llm.is_subscription`) and the
existing `_is_subscription` write path are both unchanged, so this is additive
and not an API break.

Co-Authored-By: Claude Fable 5 <noreply@anthropic.com>
---
 openhands-sdk/openhands/sdk/llm/llm.py  | 27 +++++++++++++++++++++++++
 tests/sdk/llm/test_subscription_mode.py | 19 +++++++++++++++++
 2 files changed, 46 insertions(+)

diff --git a/openhands-sdk/openhands/sdk/llm/llm.py b/openhands-sdk/openhands/sdk/llm/llm.py
index ecb5649e7d..c4560e8930 100644
--- a/openhands-sdk/openhands/sdk/llm/llm.py
+++ b/openhands-sdk/openhands/sdk/llm/llm.py
@@ -17,6 +17,7 @@
     Field,
     PrivateAttr,
     SecretStr,
+    computed_field,
     field_serializer,
     field_validator,
     model_validator,
@@ -692,6 +693,14 @@ def telemetry(self) -> Telemetry:
             )
         return self._telemetry
 
+    @computed_field(
+        return_type=bool,
+        description=(
+            "Whether this LLM uses subscription-based authentication. "
+            "Serialized so that subscription-specific request handling "
+            "survives transport to a remote agent-server."
+        ),
+    )
     @property
     def is_subscription(self) -> bool:
         """Check if this LLM uses subscription-based authentication.
@@ -705,6 +714,24 @@ def is_subscription(self) -> bool:
         """
         return self._is_subscription
 
+    @model_validator(mode="wrap")
+    @classmethod
+    def _restore_is_subscription(cls, data, handler):
+        """Restore the subscription flag when validating serialized data.
+
+        ``is_subscription`` is a computed field backed by the private
+        ``_is_subscription`` attribute, so plain validation would drop it.
+        Without this, an LLM created via ``LLM.subscription_login()`` loses
+        its subscription-specific request handling (streaming exemption,
+        Codex system prompt transform, reasoning-item stripping) after a
+        dump/validate round trip - e.g. when shipped to a remote
+        agent-server.
+        """
+        llm = handler(data)
+        if isinstance(data, dict) and data.get("is_subscription"):
+            llm._is_subscription = True
+        return llm
+
     def restore_metrics(self, metrics: Metrics) -> None:
         # Only used by ConversationStats to seed metrics
         self._metrics = metrics
diff --git a/tests/sdk/llm/test_subscription_mode.py b/tests/sdk/llm/test_subscription_mode.py
index 2b40419e82..42e3172769 100644
--- a/tests/sdk/llm/test_subscription_mode.py
+++ b/tests/sdk/llm/test_subscription_mode.py
@@ -338,3 +338,22 @@ def test_format_messages_reasoning_item_handling(
 
     serialized = json.dumps(input_items, default=str)
     assert ("rs_should_be_stripped" in serialized) == reasoning_id_present
+
+
+def test_is_subscription_survives_serialization_round_trip():
+    """is_subscription must survive model_dump -> model_validate.
+
+    A RemoteConversation serializes the LLM and the agent-server rebuilds it;
+    if the flag is lost, all subscription-specific request handling (streaming
+    exemption, system prompt transform, reasoning item stripping) silently
+    stops applying on the server side.
+    """
+    llm = _make_subscription_llm()
+    restored = LLM.model_validate(llm.model_dump(context={"expose_secrets": True}))
+    assert restored.is_subscription is True
+
+    plain = LLM(model="gpt-4o")
+    restored_plain = LLM.model_validate(
+        plain.model_dump(context={"expose_secrets": True})
+    )
+    assert restored_plain.is_subscription is False