From 494a11b09a5b90f3e8bffd7a03176d9c9705f045 Mon Sep 17 00:00:00 2001
From: Calvin Smith <calvin@all-hands.dev>
Date: Fri, 2 Jan 2026 10:49:21 -0700
Subject: [PATCH 01/19] exception for no condensation available

---
 .../openhands/sdk/context/condenser/base.py      | 16 +++++++++++++++-
 1 file changed, 15 insertions(+), 1 deletion(-)

diff --git a/openhands-sdk/openhands/sdk/context/condenser/base.py b/openhands-sdk/openhands/sdk/context/condenser/base.py
index 938b3495d9..11c72df812 100644
--- a/openhands-sdk/openhands/sdk/context/condenser/base.py
+++ b/openhands-sdk/openhands/sdk/context/condenser/base.py
@@ -65,6 +65,15 @@ class PipelinableCondenserBase(CondenserBase):
     """Abstract condenser interface which may be pipelined. (Since a pipeline
     condenser should not nest another pipeline condenser)"""
 
+class NoCondensationAvailableException(Exception):
+    """Raised when a condenser is asked to provide a condensation but none is available.
+    
+    This can happen if the condenser's `should_condense` method returns True, but due to
+    API constraints no condensation can be generated.
+
+    When this exception is raised from a rolling condenser's `get_condensation` method,
+    the agent will fall back to using the uncondensed view for the next agent step.
+    """
 
 class RollingCondenser(PipelinableCondenserBase, ABC):
     """Base class for a specialized condenser strategy that applies condensation to a
@@ -93,7 +102,12 @@ def condense(self, view: View, agent_llm: LLM | None = None) -> View | Condensat
         # If we trigger the condenser-specific condensation threshold, compute and
         # return the condensation.
         if self.should_condense(view, agent_llm=agent_llm):
-            return self.get_condensation(view, agent_llm=agent_llm)
+            try:
+                return self.get_condensation(view, agent_llm=agent_llm)
+
+            except NoCondensationAvailableException as e:
+                logger.debug(f"No condensation available: {e}")
+                return view
 
         # Otherwise we're safe to just return the view.
         else:

From fbb91c488742c73d1236f8ca3ff5c7ab28c480ff Mon Sep 17 00:00:00 2001
From: Calvin Smith <calvin@all-hands.dev>
Date: Fri, 2 Jan 2026 10:50:07 -0700
Subject: [PATCH 02/19] fix imports

---
 openhands-sdk/openhands/sdk/context/condenser/__init__.py | 2 ++
 1 file changed, 2 insertions(+)

diff --git a/openhands-sdk/openhands/sdk/context/condenser/__init__.py b/openhands-sdk/openhands/sdk/context/condenser/__init__.py
index 155f745f64..09c7ad09b6 100644
--- a/openhands-sdk/openhands/sdk/context/condenser/__init__.py
+++ b/openhands-sdk/openhands/sdk/context/condenser/__init__.py
@@ -1,5 +1,6 @@
 from openhands.sdk.context.condenser.base import (
     CondenserBase,
+    NoCondensationAvailableException,
     RollingCondenser,
 )
 from openhands.sdk.context.condenser.llm_summarizing_condenser import (
@@ -15,4 +16,5 @@
     "NoOpCondenser",
     "PipelineCondenser",
     "LLMSummarizingCondenser",
+    "NoCondensationAvailableException",
 ]

From 8a073f81758937fd0066e9efc881e9b42e94d3e0 Mon Sep 17 00:00:00 2001
From: Calvin Smith <calvin@all-hands.dev>
Date: Fri, 2 Jan 2026 10:53:20 -0700
Subject: [PATCH 03/19] adding tests for rolling condenser base class

---
 .../condenser/test_rolling_condenser.py       | 138 ++++++++++++++++++
 1 file changed, 138 insertions(+)
 create mode 100644 tests/sdk/context/condenser/test_rolling_condenser.py

diff --git a/tests/sdk/context/condenser/test_rolling_condenser.py b/tests/sdk/context/condenser/test_rolling_condenser.py
new file mode 100644
index 0000000000..6db94b2bc3
--- /dev/null
+++ b/tests/sdk/context/condenser/test_rolling_condenser.py
@@ -0,0 +1,138 @@
+from unittest.mock import MagicMock
+
+import pytest
+
+from openhands.sdk.context.condenser.base import (
+    NoCondensationAvailableException,
+    RollingCondenser,
+)
+from openhands.sdk.context.view import View
+from openhands.sdk.event.base import Event
+from openhands.sdk.event.condenser import Condensation
+from openhands.sdk.event.llm_convertible import MessageEvent
+from openhands.sdk.llm import LLM, Message, TextContent
+
+
+def message_event(content: str) -> MessageEvent:
+    return MessageEvent(
+        llm_message=Message(role="user", content=[TextContent(text=content)]),
+        source="user",
+    )
+
+
+class MockRollingCondenser(RollingCondenser):
+    """Mock implementation of RollingCondenser for testing."""
+
+    def __init__(
+        self,
+        should_condense_value: bool = True,
+        raise_exception: bool = False,
+    ):
+        self._should_condense_value = should_condense_value
+        self._raise_exception = raise_exception
+
+    def should_condense(self, view: View, agent_llm: LLM | None = None) -> bool:
+        return self._should_condense_value
+
+    def get_condensation(
+        self, view: View, agent_llm: LLM | None = None
+    ) -> Condensation:
+        if self._raise_exception:
+            raise NoCondensationAvailableException(
+                "No condensation available due to API constraints"
+            )
+        # Return a simple condensation for successful case
+        return Condensation(
+            forgotten_event_ids=[view.events[0].id],
+            summary="Mock summary",
+            summary_offset=0,
+            llm_response_id="mock-response-id",
+        )
+
+
+def test_rolling_condenser_returns_view_when_no_condensation_needed() -> None:
+    """Test that RollingCondenser returns the original view when should_condense returns False."""
+    condenser = MockRollingCondenser(should_condense_value=False)
+
+    events: list[Event] = [
+        message_event("Event 1"),
+        message_event("Event 2"),
+        message_event("Event 3"),
+    ]
+    view = View.from_events(events)
+
+    result = condenser.condense(view)
+
+    assert isinstance(result, View)
+    assert result == view
+
+
+def test_rolling_condenser_returns_condensation_when_needed() -> None:
+    """Test that RollingCondenser returns a Condensation when should_condense returns True."""
+    condenser = MockRollingCondenser(should_condense_value=True, raise_exception=False)
+
+    events: list[Event] = [
+        message_event("Event 1"),
+        message_event("Event 2"),
+        message_event("Event 3"),
+    ]
+    view = View.from_events(events)
+
+    result = condenser.condense(view)
+
+    assert isinstance(result, Condensation)
+    assert result.summary == "Mock summary"
+
+
+def test_rolling_condenser_returns_view_on_no_condensation_available_exception() -> None:
+    """Test that RollingCondenser returns the original view when NoCondensationAvailableException is raised.
+
+    This tests the exception handling added in base.py:105-110 which catches
+    NoCondensationAvailableException from get_condensation() and returns the
+    original view as a fallback.
+    """
+    condenser = MockRollingCondenser(should_condense_value=True, raise_exception=True)
+
+    events: list[Event] = [
+        message_event("Event 1"),
+        message_event("Event 2"),
+        message_event("Event 3"),
+    ]
+    view = View.from_events(events)
+
+    # Even though should_condense returns True, the exception should be caught
+    # and the original view should be returned
+    result = condenser.condense(view)
+
+    assert isinstance(result, View)
+    assert result == view
+    assert result.events == events
+
+
+def test_rolling_condenser_with_agent_llm() -> None:
+    """Test that RollingCondenser works with optional agent_llm parameter."""
+    condenser = MockRollingCondenser(should_condense_value=True, raise_exception=False)
+
+    events: list[Event] = [
+        message_event("Event 1"),
+        message_event("Event 2"),
+        message_event("Event 3"),
+    ]
+    view = View.from_events(events)
+
+    # Create a mock LLM
+    mock_llm = MagicMock(spec=LLM)
+
+    # Condense with agent_llm parameter
+    result = condenser.condense(view, agent_llm=mock_llm)
+
+    assert isinstance(result, Condensation)
+    assert result.summary == "Mock summary"
+
+
+def test_no_condensation_available_exception_message() -> None:
+    """Test that NoCondensationAvailableException can be raised with a custom message."""
+    exception_message = "Custom error message about API constraints"
+
+    with pytest.raises(NoCondensationAvailableException, match=exception_message):
+        raise NoCondensationAvailableException(exception_message)

From dafde2c90845c4ea97497aaa375668d9ca309422 Mon Sep 17 00:00:00 2001
From: Calvin Smith <calvin@all-hands.dev>
Date: Fri, 2 Jan 2026 11:00:31 -0700
Subject: [PATCH 04/19] moving excedption behavior, changing base exception
 class

---
 .../condenser/llm_summarizing_condenser.py    | 19 +++++++++++--------
 1 file changed, 11 insertions(+), 8 deletions(-)

diff --git a/openhands-sdk/openhands/sdk/context/condenser/llm_summarizing_condenser.py b/openhands-sdk/openhands/sdk/context/condenser/llm_summarizing_condenser.py
index ad0d3c6fe6..c50f7fa989 100644
--- a/openhands-sdk/openhands/sdk/context/condenser/llm_summarizing_condenser.py
+++ b/openhands-sdk/openhands/sdk/context/condenser/llm_summarizing_condenser.py
@@ -4,7 +4,7 @@
 
 from pydantic import Field, model_validator
 
-from openhands.sdk.context.condenser.base import RollingCondenser
+from openhands.sdk.context.condenser.base import NoCondensationAvailableException, RollingCondenser
 from openhands.sdk.context.condenser.utils import (
     get_suffix_length_for_token_reduction,
     get_total_token_count,
@@ -124,12 +124,7 @@ def _generate_condensation(
         Raises:
             ValueError: If forgotten_events is empty (0 events to condense).
         """
-        if len(forgotten_events) == 0:
-            raise ValueError(
-                "Cannot condense 0 events. This typically occurs when a tool loop "
-                "spans almost the entire view, leaving no valid range for forgetting "
-                "events. Consider adjusting keep_first or max_size parameters."
-            )
+        assert len(forgotten_events) > 0, "No events to condense."
 
         # Convert events to strings for the template
         event_strings = [str(forgotten_event) for forgotten_event in forgotten_events]
@@ -236,11 +231,19 @@ def get_condensation(
     ) -> Condensation:
         # The condensation is dependent on the events we want to drop and the previous
         # summary.
-        summary_event_content = self._get_summary_event_content(view)
         forgotten_events, summary_offset = self._get_forgotten_events(
             view, agent_llm=agent_llm
         )
 
+        if not forgotten_events:
+            raise NoCondensationAvailableException(
+                "Cannot condense 0 events. This typically occurs when a tool loop "
+                "spans almost the entire view, leaving no valid range for forgetting "
+                "events. Consider adjusting keep_first or max_size parameters."
+            )
+        
+        summary_event_content = self._get_summary_event_content(view)
+
         return self._generate_condensation(
             summary_event_content=summary_event_content,
             forgotten_events=forgotten_events,

From 6846132bf3a58e41835b7b59a63e3301898b7b3d Mon Sep 17 00:00:00 2001
From: Calvin Smith <calvin@all-hands.dev>
Date: Fri, 2 Jan 2026 11:00:57 -0700
Subject: [PATCH 05/19] linting

---
 openhands-sdk/openhands/sdk/context/condenser/base.py      | 4 +++-
 .../sdk/context/condenser/llm_summarizing_condenser.py     | 7 +++++--
 tests/sdk/context/condenser/test_rolling_condenser.py      | 4 +++-
 3 files changed, 11 insertions(+), 4 deletions(-)

diff --git a/openhands-sdk/openhands/sdk/context/condenser/base.py b/openhands-sdk/openhands/sdk/context/condenser/base.py
index 11c72df812..a0b7c248f1 100644
--- a/openhands-sdk/openhands/sdk/context/condenser/base.py
+++ b/openhands-sdk/openhands/sdk/context/condenser/base.py
@@ -65,9 +65,10 @@ class PipelinableCondenserBase(CondenserBase):
     """Abstract condenser interface which may be pipelined. (Since a pipeline
     condenser should not nest another pipeline condenser)"""
 
+
 class NoCondensationAvailableException(Exception):
     """Raised when a condenser is asked to provide a condensation but none is available.
-    
+
     This can happen if the condenser's `should_condense` method returns True, but due to
     API constraints no condensation can be generated.
 
@@ -75,6 +76,7 @@ class NoCondensationAvailableException(Exception):
     the agent will fall back to using the uncondensed view for the next agent step.
     """
 
+
 class RollingCondenser(PipelinableCondenserBase, ABC):
     """Base class for a specialized condenser strategy that applies condensation to a
     rolling history.
diff --git a/openhands-sdk/openhands/sdk/context/condenser/llm_summarizing_condenser.py b/openhands-sdk/openhands/sdk/context/condenser/llm_summarizing_condenser.py
index c50f7fa989..b66f9d8d47 100644
--- a/openhands-sdk/openhands/sdk/context/condenser/llm_summarizing_condenser.py
+++ b/openhands-sdk/openhands/sdk/context/condenser/llm_summarizing_condenser.py
@@ -4,7 +4,10 @@
 
 from pydantic import Field, model_validator
 
-from openhands.sdk.context.condenser.base import NoCondensationAvailableException, RollingCondenser
+from openhands.sdk.context.condenser.base import (
+    NoCondensationAvailableException,
+    RollingCondenser,
+)
 from openhands.sdk.context.condenser.utils import (
     get_suffix_length_for_token_reduction,
     get_total_token_count,
@@ -241,7 +244,7 @@ def get_condensation(
                 "spans almost the entire view, leaving no valid range for forgetting "
                 "events. Consider adjusting keep_first or max_size parameters."
             )
-        
+
         summary_event_content = self._get_summary_event_content(view)
 
         return self._generate_condensation(
diff --git a/tests/sdk/context/condenser/test_rolling_condenser.py b/tests/sdk/context/condenser/test_rolling_condenser.py
index 6db94b2bc3..e461e850d9 100644
--- a/tests/sdk/context/condenser/test_rolling_condenser.py
+++ b/tests/sdk/context/condenser/test_rolling_condenser.py
@@ -84,7 +84,9 @@ def test_rolling_condenser_returns_condensation_when_needed() -> None:
     assert result.summary == "Mock summary"
 
 
-def test_rolling_condenser_returns_view_on_no_condensation_available_exception() -> None:
+def test_rolling_condenser_returns_view_on_no_condensation_available_exception() -> (
+    None
+):
     """Test that RollingCondenser returns the original view when NoCondensationAvailableException is raised.
 
     This tests the exception handling added in base.py:105-110 which catches

From 6b2e9d96c9303b24f59907f4b003ba31d8b9e09a Mon Sep 17 00:00:00 2001
From: Calvin Smith <calvin@all-hands.dev>
Date: Fri, 2 Jan 2026 11:06:39 -0700
Subject: [PATCH 06/19] linting

---
 tests/sdk/context/condenser/test_rolling_condenser.py | 5 +++--
 1 file changed, 3 insertions(+), 2 deletions(-)

diff --git a/tests/sdk/context/condenser/test_rolling_condenser.py b/tests/sdk/context/condenser/test_rolling_condenser.py
index e461e850d9..64de0c0050 100644
--- a/tests/sdk/context/condenser/test_rolling_condenser.py
+++ b/tests/sdk/context/condenser/test_rolling_condenser.py
@@ -87,7 +87,8 @@ def test_rolling_condenser_returns_condensation_when_needed() -> None:
 def test_rolling_condenser_returns_view_on_no_condensation_available_exception() -> (
     None
 ):
-    """Test that RollingCondenser returns the original view when NoCondensationAvailableException is raised.
+    """Test that RollingCondenser returns the original view when
+    NoCondensationAvailableException is raised.
 
     This tests the exception handling added in base.py:105-110 which catches
     NoCondensationAvailableException from get_condensation() and returns the
@@ -133,7 +134,7 @@ def test_rolling_condenser_with_agent_llm() -> None:
 
 
 def test_no_condensation_available_exception_message() -> None:
-    """Test that NoCondensationAvailableException can be raised with a custom message."""
+    """Test that NoCondensationAvailableException raisable with custom message."""
     exception_message = "Custom error message about API constraints"
 
     with pytest.raises(NoCondensationAvailableException, match=exception_message):

From f57bc818c493edcc59455c2bd40e554127da17e9 Mon Sep 17 00:00:00 2001
From: Calvin Smith <calvin@all-hands.dev>
Date: Fri, 2 Jan 2026 11:08:24 -0700
Subject: [PATCH 07/19] yet more linting

---
 tests/sdk/context/condenser/test_rolling_condenser.py | 8 ++++++--
 1 file changed, 6 insertions(+), 2 deletions(-)

diff --git a/tests/sdk/context/condenser/test_rolling_condenser.py b/tests/sdk/context/condenser/test_rolling_condenser.py
index 64de0c0050..df8b968af8 100644
--- a/tests/sdk/context/condenser/test_rolling_condenser.py
+++ b/tests/sdk/context/condenser/test_rolling_condenser.py
@@ -51,7 +51,9 @@ def get_condensation(
 
 
 def test_rolling_condenser_returns_view_when_no_condensation_needed() -> None:
-    """Test that RollingCondenser returns the original view when should_condense returns False."""
+    """Test that RollingCondenser returns the original view when should_condense returns
+    False.
+    """
     condenser = MockRollingCondenser(should_condense_value=False)
 
     events: list[Event] = [
@@ -68,7 +70,9 @@ def test_rolling_condenser_returns_view_when_no_condensation_needed() -> None:
 
 
 def test_rolling_condenser_returns_condensation_when_needed() -> None:
-    """Test that RollingCondenser returns a Condensation when should_condense returns True."""
+    """Test that RollingCondenser returns a Condensation when should_condense returns
+    True.
+    """
     condenser = MockRollingCondenser(should_condense_value=True, raise_exception=False)
 
     events: list[Event] = [

From c160c5c2d6179338f8176d62baf411a488f092c2 Mon Sep 17 00:00:00 2001
From: Calvin Smith <calvin@all-hands.dev>
Date: Fri, 2 Jan 2026 11:13:16 -0700
Subject: [PATCH 08/19] updating tests that reference old exception logic

---
 tests/sdk/context/condenser/test_llm_summarizing_condenser.py | 4 ++--
 1 file changed, 2 insertions(+), 2 deletions(-)

diff --git a/tests/sdk/context/condenser/test_llm_summarizing_condenser.py b/tests/sdk/context/condenser/test_llm_summarizing_condenser.py
index ebb463e9b8..995f074f40 100644
--- a/tests/sdk/context/condenser/test_llm_summarizing_condenser.py
+++ b/tests/sdk/context/condenser/test_llm_summarizing_condenser.py
@@ -554,7 +554,7 @@ def test_most_aggressive_condensation_chosen(mock_llm: LLM) -> None:
 
 
 def test_generate_condensation_raises_on_zero_events(mock_llm: LLM) -> None:
-    """Test that _generate_condensation raises ValueError when given 0 events.
+    """Test that _generate_condensation raises AssertionError when given 0 events.
 
     This prevents the LLM from being called with an empty event list, which would
     produce a confusing summary like "I don't see any events provided to summarize."
@@ -562,7 +562,7 @@ def test_generate_condensation_raises_on_zero_events(mock_llm: LLM) -> None:
     """
     condenser = LLMSummarizingCondenser(llm=mock_llm, max_size=100, keep_first=2)
 
-    with pytest.raises(ValueError, match="Cannot condense 0 events"):
+    with pytest.raises(AssertionError, match="No events to condense"):
         condenser._generate_condensation(
             summary_event_content="",
             forgotten_events=[],

From d74c03947be7582502628ff68353977afb784209 Mon Sep 17 00:00:00 2001
From: Calvin Smith <calvin@all-hands.dev>
Date: Fri, 2 Jan 2026 11:46:19 -0700
Subject: [PATCH 09/19] minor integration test logging changed

---
 tests/integration/tests/t09_token_condenser.py | 11 ++++++-----
 1 file changed, 6 insertions(+), 5 deletions(-)

diff --git a/tests/integration/tests/t09_token_condenser.py b/tests/integration/tests/t09_token_condenser.py
index dec8cf5715..193a94df73 100644
--- a/tests/integration/tests/t09_token_condenser.py
+++ b/tests/integration/tests/t09_token_condenser.py
@@ -40,7 +40,7 @@ class TokenCondenserTest(BaseIntegrationTest):
 
     def __init__(self, *args, **kwargs):
         """Initialize test with tracking variables."""
-        self.condensation_count = 0
+        self.condensations: list[Condensation] = []
         super().__init__(*args, **kwargs)
 
         # Some models explicitly disallow long, repetitive tool loops for cost/safety.
@@ -87,25 +87,26 @@ def conversation_callback(self, event):
         super().conversation_callback(event)
 
         if isinstance(event, Condensation):
-            if self.condensation_count >= 1:
+            if len(self.condensations) >= 1:
                 logger.info("2nd condensation detected! Stopping test early.")
                 self.conversation.pause()
             # We allow the first condensation request to test if
             # thinking block + condensation will work together
-            self.condensation_count += 1
+            self.condensations.append(event)
 
     def setup(self) -> None:
         logger.info(f"Token condenser test: max_tokens={self.condenser.max_tokens}")
 
     def verify_result(self) -> TestResult:
         """Verify that condensation was triggered based on token count."""
-        if self.condensation_count == 0:
+        if len(self.condensations) == 0:
             return TestResult(
                 success=False,
                 reason="Condensation not triggered. Token counting may not work.",
             )
 
+        events_summarized = len(self.condensations[0].forgotten_event_ids)
         return TestResult(
             success=True,
-            reason="Condensation triggered. Token counting works correctly.",
+            reason=f"Condensation triggered, summarizing {events_summarized} events.",
         )

From 053e4c0a3f704923f9f8fb337583850a3c04967e Mon Sep 17 00:00:00 2001
From: Calvin Smith <calvin@all-hands.dev>
Date: Fri, 2 Jan 2026 16:55:30 -0700
Subject: [PATCH 10/19] reasons for a condensation request

---
 openhands-sdk/openhands/sdk/agent/agent.py        |  8 ++++++--
 .../sdk/conversation/impl/local_conversation.py   |  5 ++++-
 openhands-sdk/openhands/sdk/event/condenser.py    | 15 +++++++++++++--
 3 files changed, 23 insertions(+), 5 deletions(-)

diff --git a/openhands-sdk/openhands/sdk/agent/agent.py b/openhands-sdk/openhands/sdk/agent/agent.py
index 2806677ade..8985f9d103 100644
--- a/openhands-sdk/openhands/sdk/agent/agent.py
+++ b/openhands-sdk/openhands/sdk/agent/agent.py
@@ -27,7 +27,11 @@
     TokenEvent,
     UserRejectObservation,
 )
-from openhands.sdk.event.condenser import Condensation, CondensationRequest
+from openhands.sdk.event.condenser import (
+    Condensation,
+    CondensationRequest,
+    CondensationRequestReason,
+)
 from openhands.sdk.llm import (
     LLMResponse,
     Message,
@@ -200,7 +204,7 @@ def step(
                 logger.warning(
                     "LLM raised context window exceeded error, triggering condensation"
                 )
-                on_event(CondensationRequest())
+                on_event(CondensationRequest(reason=CondensationRequestReason.CONTEXT_LIMIT))
                 return
             # No condenser available or doesn't handle requests; log helpful warning
             self._log_context_window_exceeded_warning()
diff --git a/openhands-sdk/openhands/sdk/conversation/impl/local_conversation.py b/openhands-sdk/openhands/sdk/conversation/impl/local_conversation.py
index 984f01f846..336553f546 100644
--- a/openhands-sdk/openhands/sdk/conversation/impl/local_conversation.py
+++ b/openhands-sdk/openhands/sdk/conversation/impl/local_conversation.py
@@ -30,6 +30,7 @@
     PauseEvent,
     UserRejectObservation,
 )
+from openhands.sdk.event.condenser import CondensationRequestReason
 from openhands.sdk.event.conversation_error import ConversationErrorEvent
 from openhands.sdk.hooks import HookConfig, HookEventProcessor, create_hook_callback
 from openhands.sdk.llm import LLM, Message, TextContent
@@ -646,7 +647,9 @@ def condense(self) -> None:
             )
 
         # Add a condensation request event
-        condensation_request = CondensationRequest()
+        condensation_request = CondensationRequest(
+            reason=CondensationRequestReason.MANUAL
+        )
         self._on_event(condensation_request)
 
         # Force the agent to take a single step to process the condensation request
diff --git a/openhands-sdk/openhands/sdk/event/condenser.py b/openhands-sdk/openhands/sdk/event/condenser.py
index 6f58a45d17..50416c5fe4 100644
--- a/openhands-sdk/openhands/sdk/event/condenser.py
+++ b/openhands-sdk/openhands/sdk/event/condenser.py
@@ -1,3 +1,4 @@
+from enum import Enum
 from pydantic import Field
 from rich.text import Text
 
@@ -45,6 +46,15 @@ def visualize(self) -> Text:
             text.append(f"{self.summary}\n")
         return text
 
+class CondensationRequestReason(Enum):
+    MANUAL = "manual"
+    """The condensation was requested manually by the user or system."""
+
+    CONTEXT_LIMIT = "context_limit"
+    """The condensation was requested due to reaching the context window limit."""
+
+    UNRECOVERABLE = "unrecoverable"
+    """The condensation was requested due to an unrecoverable error."""
 
 class CondensationRequest(Event):
     """This action is used to request a condensation of the conversation history.
@@ -52,9 +62,10 @@ class CondensationRequest(Event):
     Attributes:
         action (str): The action type, namely ActionType.CONDENSATION_REQUEST.
     """
-
+    
     source: SourceType = "environment"
-
+    reason: CondensationRequestReason | None = None
+    
     @property
     def visualize(self) -> Text:
         text = Text()

From 042cbbd8fc611588da1d3797a6be82a472c66171 Mon Sep 17 00:00:00 2001
From: Calvin Smith <calvin@all-hands.dev>
Date: Fri, 2 Jan 2026 16:59:19 -0700
Subject: [PATCH 11/19] view extracts unhandled condensation request reasons

---
 openhands-sdk/openhands/sdk/context/view.py | 12 +++++++++++-
 1 file changed, 11 insertions(+), 1 deletion(-)

diff --git a/openhands-sdk/openhands/sdk/context/view.py b/openhands-sdk/openhands/sdk/context/view.py
index 0b6b594a98..67d807b3fa 100644
--- a/openhands-sdk/openhands/sdk/context/view.py
+++ b/openhands-sdk/openhands/sdk/context/view.py
@@ -15,6 +15,7 @@
     LLMConvertibleEvent,
 )
 from openhands.sdk.event.base import Event, EventID
+from openhands.sdk.event.condenser import CondensationRequestReason
 from openhands.sdk.event.llm_convertible import (
     ActionEvent,
     ObservationBaseEvent,
@@ -80,6 +81,9 @@ class View(BaseModel):
     unhandled_condensation_request: bool = False
     """Whether there is an unhandled condensation request in the view."""
 
+    unhandled_condensation_request_reasons: list[CondensationRequestReason] = []
+    """List of reasons for unhandled condensation requests in the view."""
+
     condensations: list[Condensation] = []
     """A list of condensations that were processed to produce the view."""
 
@@ -489,15 +493,21 @@ def from_events(events: Sequence[Event]) -> View:
         # Check for an unhandled condensation request -- these are events closer to the
         # end of the list than any condensation action.
         unhandled_condensation_request = False
+        unhandled_condensation_request_reasons: list[CondensationRequestReason] = []
+        
         for event in reversed(events):
             if isinstance(event, Condensation):
                 break
+            
             if isinstance(event, CondensationRequest):
                 unhandled_condensation_request = True
-                break
+
+                if event.reason is not None:
+                    unhandled_condensation_request_reasons.append(event.reason)
 
         return View(
             events=View.filter_unmatched_tool_calls(kept_events),
             unhandled_condensation_request=unhandled_condensation_request,
+            unhandled_condensation_request_reasons=unhandled_condensation_request_reasons,
             condensations=condensations,
         )

From af64f9b8ca01db61f169495a6adeff889c1a872b Mon Sep 17 00:00:00 2001
From: Calvin Smith <calvin@all-hands.dev>
Date: Fri, 2 Jan 2026 17:04:51 -0700
Subject: [PATCH 12/19] rolling condenser only continues on certain
 condensation request reasons

---
 .../openhands/sdk/context/condenser/base.py   | 29 +++++++++++++++++--
 1 file changed, 27 insertions(+), 2 deletions(-)

diff --git a/openhands-sdk/openhands/sdk/context/condenser/base.py b/openhands-sdk/openhands/sdk/context/condenser/base.py
index a0b7c248f1..e6e568853d 100644
--- a/openhands-sdk/openhands/sdk/context/condenser/base.py
+++ b/openhands-sdk/openhands/sdk/context/condenser/base.py
@@ -2,7 +2,7 @@
 from logging import getLogger
 
 from openhands.sdk.context.view import View
-from openhands.sdk.event.condenser import Condensation
+from openhands.sdk.event.condenser import Condensation, CondensationRequestReason
 from openhands.sdk.llm import LLM
 from openhands.sdk.utils.models import (
     DiscriminatedUnionMixin,
@@ -100,6 +100,19 @@ def get_condensation(
     ) -> Condensation:
         """Get the condensation from a view."""
 
+    def is_continuable_reason(self, reason: CondensationRequestReason) -> bool:
+        """Determine if the given condensation request reason is continuable by this
+        condenser.
+
+        Args:
+            reason: The reason for the condensation request.
+        Returns:
+            bool: True if the reason is continuable, False otherwise.
+        """
+        if reason == CondensationRequestReason.MANUAL:
+            return True
+        return False
+
     def condense(self, view: View, agent_llm: LLM | None = None) -> View | Condensation:
         # If we trigger the condenser-specific condensation threshold, compute and
         # return the condensation.
@@ -109,7 +122,19 @@ def condense(self, view: View, agent_llm: LLM | None = None) -> View | Condensat
 
             except NoCondensationAvailableException as e:
                 logger.debug(f"No condensation available: {e}")
-                return view
+                
+                # Check if all condensation requests are continuable. If so, or if there
+                # are no requests, we can safely return the uncondensed view.
+                all_continuable = all(
+                    self.is_continuable_reason(reason)
+                    for reason in view.unhandled_condensation_request_reasons
+                )
+                if not view.unhandled_condensation_request or all_continuable:
+                    return view
+                
+                # Otherwise re-raise the exception.
+                else:
+                    raise e
 
         # Otherwise we're safe to just return the view.
         else:

From 40f634fcdc025f4d7c6d634ccb6f885e26d5c76b Mon Sep 17 00:00:00 2001
From: Calvin Smith <calvin@all-hands.dev>
Date: Fri, 2 Jan 2026 17:05:18 -0700
Subject: [PATCH 13/19] linting

---
 openhands-sdk/openhands/sdk/agent/agent.py            | 4 +++-
 openhands-sdk/openhands/sdk/context/condenser/base.py | 4 ++--
 openhands-sdk/openhands/sdk/context/view.py           | 4 ++--
 openhands-sdk/openhands/sdk/event/condenser.py        | 7 +++++--
 4 files changed, 12 insertions(+), 7 deletions(-)

diff --git a/openhands-sdk/openhands/sdk/agent/agent.py b/openhands-sdk/openhands/sdk/agent/agent.py
index 8985f9d103..1b2076e0e3 100644
--- a/openhands-sdk/openhands/sdk/agent/agent.py
+++ b/openhands-sdk/openhands/sdk/agent/agent.py
@@ -204,7 +204,9 @@ def step(
                 logger.warning(
                     "LLM raised context window exceeded error, triggering condensation"
                 )
-                on_event(CondensationRequest(reason=CondensationRequestReason.CONTEXT_LIMIT))
+                on_event(
+                    CondensationRequest(reason=CondensationRequestReason.CONTEXT_LIMIT)
+                )
                 return
             # No condenser available or doesn't handle requests; log helpful warning
             self._log_context_window_exceeded_warning()
diff --git a/openhands-sdk/openhands/sdk/context/condenser/base.py b/openhands-sdk/openhands/sdk/context/condenser/base.py
index e6e568853d..ed948b30b4 100644
--- a/openhands-sdk/openhands/sdk/context/condenser/base.py
+++ b/openhands-sdk/openhands/sdk/context/condenser/base.py
@@ -122,7 +122,7 @@ def condense(self, view: View, agent_llm: LLM | None = None) -> View | Condensat
 
             except NoCondensationAvailableException as e:
                 logger.debug(f"No condensation available: {e}")
-                
+
                 # Check if all condensation requests are continuable. If so, or if there
                 # are no requests, we can safely return the uncondensed view.
                 all_continuable = all(
@@ -131,7 +131,7 @@ def condense(self, view: View, agent_llm: LLM | None = None) -> View | Condensat
                 )
                 if not view.unhandled_condensation_request or all_continuable:
                     return view
-                
+
                 # Otherwise re-raise the exception.
                 else:
                     raise e
diff --git a/openhands-sdk/openhands/sdk/context/view.py b/openhands-sdk/openhands/sdk/context/view.py
index 67d807b3fa..fc999e1dbc 100644
--- a/openhands-sdk/openhands/sdk/context/view.py
+++ b/openhands-sdk/openhands/sdk/context/view.py
@@ -494,11 +494,11 @@ def from_events(events: Sequence[Event]) -> View:
         # end of the list than any condensation action.
         unhandled_condensation_request = False
         unhandled_condensation_request_reasons: list[CondensationRequestReason] = []
-        
+
         for event in reversed(events):
             if isinstance(event, Condensation):
                 break
-            
+
             if isinstance(event, CondensationRequest):
                 unhandled_condensation_request = True
 
diff --git a/openhands-sdk/openhands/sdk/event/condenser.py b/openhands-sdk/openhands/sdk/event/condenser.py
index 50416c5fe4..27be6c677e 100644
--- a/openhands-sdk/openhands/sdk/event/condenser.py
+++ b/openhands-sdk/openhands/sdk/event/condenser.py
@@ -1,4 +1,5 @@
 from enum import Enum
+
 from pydantic import Field
 from rich.text import Text
 
@@ -46,6 +47,7 @@ def visualize(self) -> Text:
             text.append(f"{self.summary}\n")
         return text
 
+
 class CondensationRequestReason(Enum):
     MANUAL = "manual"
     """The condensation was requested manually by the user or system."""
@@ -56,16 +58,17 @@ class CondensationRequestReason(Enum):
     UNRECOVERABLE = "unrecoverable"
     """The condensation was requested due to an unrecoverable error."""
 
+
 class CondensationRequest(Event):
     """This action is used to request a condensation of the conversation history.
 
     Attributes:
         action (str): The action type, namely ActionType.CONDENSATION_REQUEST.
     """
-    
+
     source: SourceType = "environment"
     reason: CondensationRequestReason | None = None
-    
+
     @property
     def visualize(self) -> Text:
         text = Text()

From b4783da036cdfa13b7c6faa818f81439a6006d2b Mon Sep 17 00:00:00 2001
From: Calvin Smith <calvin@all-hands.dev>
Date: Mon, 5 Jan 2026 10:08:38 -0700
Subject: [PATCH 14/19] reverting condensation reasons

---
 openhands-sdk/openhands/sdk/agent/agent.py         |  5 +----
 openhands-sdk/openhands/sdk/context/view.py        | 10 +---------
 .../sdk/conversation/impl/local_conversation.py    |  5 +----
 openhands-sdk/openhands/sdk/event/condenser.py     | 14 --------------
 4 files changed, 3 insertions(+), 31 deletions(-)

diff --git a/openhands-sdk/openhands/sdk/agent/agent.py b/openhands-sdk/openhands/sdk/agent/agent.py
index 1b2076e0e3..893321f702 100644
--- a/openhands-sdk/openhands/sdk/agent/agent.py
+++ b/openhands-sdk/openhands/sdk/agent/agent.py
@@ -30,7 +30,6 @@
 from openhands.sdk.event.condenser import (
     Condensation,
     CondensationRequest,
-    CondensationRequestReason,
 )
 from openhands.sdk.llm import (
     LLMResponse,
@@ -204,9 +203,7 @@ def step(
                 logger.warning(
                     "LLM raised context window exceeded error, triggering condensation"
                 )
-                on_event(
-                    CondensationRequest(reason=CondensationRequestReason.CONTEXT_LIMIT)
-                )
+                on_event(CondensationRequest())
                 return
             # No condenser available or doesn't handle requests; log helpful warning
             self._log_context_window_exceeded_warning()
diff --git a/openhands-sdk/openhands/sdk/context/view.py b/openhands-sdk/openhands/sdk/context/view.py
index fc999e1dbc..c80739e1ac 100644
--- a/openhands-sdk/openhands/sdk/context/view.py
+++ b/openhands-sdk/openhands/sdk/context/view.py
@@ -15,7 +15,6 @@
     LLMConvertibleEvent,
 )
 from openhands.sdk.event.base import Event, EventID
-from openhands.sdk.event.condenser import CondensationRequestReason
 from openhands.sdk.event.llm_convertible import (
     ActionEvent,
     ObservationBaseEvent,
@@ -81,9 +80,6 @@ class View(BaseModel):
     unhandled_condensation_request: bool = False
     """Whether there is an unhandled condensation request in the view."""
 
-    unhandled_condensation_request_reasons: list[CondensationRequestReason] = []
-    """List of reasons for unhandled condensation requests in the view."""
-
     condensations: list[Condensation] = []
     """A list of condensations that were processed to produce the view."""
 
@@ -493,7 +489,6 @@ def from_events(events: Sequence[Event]) -> View:
         # Check for an unhandled condensation request -- these are events closer to the
         # end of the list than any condensation action.
         unhandled_condensation_request = False
-        unhandled_condensation_request_reasons: list[CondensationRequestReason] = []
 
         for event in reversed(events):
             if isinstance(event, Condensation):
@@ -501,13 +496,10 @@ def from_events(events: Sequence[Event]) -> View:
 
             if isinstance(event, CondensationRequest):
                 unhandled_condensation_request = True
-
-                if event.reason is not None:
-                    unhandled_condensation_request_reasons.append(event.reason)
+                break
 
         return View(
             events=View.filter_unmatched_tool_calls(kept_events),
             unhandled_condensation_request=unhandled_condensation_request,
-            unhandled_condensation_request_reasons=unhandled_condensation_request_reasons,
             condensations=condensations,
         )
diff --git a/openhands-sdk/openhands/sdk/conversation/impl/local_conversation.py b/openhands-sdk/openhands/sdk/conversation/impl/local_conversation.py
index 336553f546..984f01f846 100644
--- a/openhands-sdk/openhands/sdk/conversation/impl/local_conversation.py
+++ b/openhands-sdk/openhands/sdk/conversation/impl/local_conversation.py
@@ -30,7 +30,6 @@
     PauseEvent,
     UserRejectObservation,
 )
-from openhands.sdk.event.condenser import CondensationRequestReason
 from openhands.sdk.event.conversation_error import ConversationErrorEvent
 from openhands.sdk.hooks import HookConfig, HookEventProcessor, create_hook_callback
 from openhands.sdk.llm import LLM, Message, TextContent
@@ -647,9 +646,7 @@ def condense(self) -> None:
             )
 
         # Add a condensation request event
-        condensation_request = CondensationRequest(
-            reason=CondensationRequestReason.MANUAL
-        )
+        condensation_request = CondensationRequest()
         self._on_event(condensation_request)
 
         # Force the agent to take a single step to process the condensation request
diff --git a/openhands-sdk/openhands/sdk/event/condenser.py b/openhands-sdk/openhands/sdk/event/condenser.py
index 27be6c677e..6f58a45d17 100644
--- a/openhands-sdk/openhands/sdk/event/condenser.py
+++ b/openhands-sdk/openhands/sdk/event/condenser.py
@@ -1,5 +1,3 @@
-from enum import Enum
-
 from pydantic import Field
 from rich.text import Text
 
@@ -48,17 +46,6 @@ def visualize(self) -> Text:
         return text
 
 
-class CondensationRequestReason(Enum):
-    MANUAL = "manual"
-    """The condensation was requested manually by the user or system."""
-
-    CONTEXT_LIMIT = "context_limit"
-    """The condensation was requested due to reaching the context window limit."""
-
-    UNRECOVERABLE = "unrecoverable"
-    """The condensation was requested due to an unrecoverable error."""
-
-
 class CondensationRequest(Event):
     """This action is used to request a condensation of the conversation history.
 
@@ -67,7 +54,6 @@ class CondensationRequest(Event):
     """
 
     source: SourceType = "environment"
-    reason: CondensationRequestReason | None = None
 
     @property
     def visualize(self) -> Text:

From aad0bdf83add27a376bebeb0eb08429ddc16bb0f Mon Sep 17 00:00:00 2001
From: Calvin Smith <calvin@all-hands.dev>
Date: Mon, 5 Jan 2026 10:31:56 -0700
Subject: [PATCH 15/19] should_condense -> condensation_requirement

---
 .../openhands/sdk/context/condenser/base.py   | 64 +++++++++++--------
 .../condenser/llm_summarizing_condenser.py    |  9 ++-
 .../test_llm_summarizing_condenser.py         |  8 ++-
 .../condenser/test_rolling_condenser.py       | 44 ++++++++-----
 4 files changed, 77 insertions(+), 48 deletions(-)

diff --git a/openhands-sdk/openhands/sdk/context/condenser/base.py b/openhands-sdk/openhands/sdk/context/condenser/base.py
index ed948b30b4..0766721f04 100644
--- a/openhands-sdk/openhands/sdk/context/condenser/base.py
+++ b/openhands-sdk/openhands/sdk/context/condenser/base.py
@@ -1,8 +1,9 @@
 from abc import ABC, abstractmethod
+from enum import Enum
 from logging import getLogger
 
 from openhands.sdk.context.view import View
-from openhands.sdk.event.condenser import Condensation, CondensationRequestReason
+from openhands.sdk.event.condenser import Condensation
 from openhands.sdk.llm import LLM
 from openhands.sdk.utils.models import (
     DiscriminatedUnionMixin,
@@ -76,6 +77,16 @@ class NoCondensationAvailableException(Exception):
     the agent will fall back to using the uncondensed view for the next agent step.
     """
 
+class CondensationRequirement(Enum):
+    """The type of condensation required by a rolling condenser."""
+
+    HARD = "hard"
+    """Indicates that a condensation is required right now, and the agent cannot proceed
+    without it.
+    """
+    
+    SOFT = "soft"
+    """Indicates that a condensation is desired but not strictly required."""
 
 class RollingCondenser(PipelinableCondenserBase, ABC):
     """Base class for a specialized condenser strategy that applies condensation to a
@@ -84,15 +95,27 @@ class RollingCondenser(PipelinableCondenserBase, ABC):
     The rolling history is generated by `View.from_events`, which analyzes all events in
     the history and produces a `View` object representing what will be sent to the LLM.
 
-    If `should_condense` says so, the condenser is then responsible for generating a
-    `Condensation` object from the `View` object. This will be added to the event
-    history which should -- when given to `get_view` -- produce the condensed `View` to
-    be passed to the LLM.
+    If `condensation_requirement` says so, the condenser is then responsible for
+    generating a `Condensation` object from the `View` object. This will be added to the
+    event history which should -- when given to `get_view` -- produce the condensed
+    `View` to be passed to the LLM.
     """
 
     @abstractmethod
-    def should_condense(self, view: View, agent_llm: LLM | None = None) -> bool:
-        """Determine if a view should be condensed."""
+    def condensation_requirement(
+        self, view: View, agent_llm: LLM | None = None
+    ) -> CondensationRequirement | None:
+        """Determine how a view should be condensed.
+        
+        Args:
+            view: The current view of the conversation history.
+            agent_llm: LLM instance used by the agent. Condensers use this for token
+                counting purposes. Defaults to None.
+
+        Returns:
+            CondensationRequirement | None: The type of condensation required, or None
+            if no condensation is needed.
+        """
 
     @abstractmethod
     def get_condensation(
@@ -100,36 +123,21 @@ def get_condensation(
     ) -> Condensation:
         """Get the condensation from a view."""
 
-    def is_continuable_reason(self, reason: CondensationRequestReason) -> bool:
-        """Determine if the given condensation request reason is continuable by this
-        condenser.
-
-        Args:
-            reason: The reason for the condensation request.
-        Returns:
-            bool: True if the reason is continuable, False otherwise.
-        """
-        if reason == CondensationRequestReason.MANUAL:
-            return True
-        return False
-
     def condense(self, view: View, agent_llm: LLM | None = None) -> View | Condensation:
         # If we trigger the condenser-specific condensation threshold, compute and
         # return the condensation.
-        if self.should_condense(view, agent_llm=agent_llm):
+        request = self.condensation_requirement(view, agent_llm=agent_llm)
+        if request is not None:
             try:
                 return self.get_condensation(view, agent_llm=agent_llm)
 
             except NoCondensationAvailableException as e:
                 logger.debug(f"No condensation available: {e}")
 
-                # Check if all condensation requests are continuable. If so, or if there
-                # are no requests, we can safely return the uncondensed view.
-                all_continuable = all(
-                    self.is_continuable_reason(reason)
-                    for reason in view.unhandled_condensation_request_reasons
-                )
-                if not view.unhandled_condensation_request or all_continuable:
+                if request == CondensationRequirement.SOFT:
+                    # For soft requests, we can just return the uncondensed view. This
+                    # request will _eventually_ be handled, but it's not critical that
+                    # we do so immediately.
                     return view
 
                 # Otherwise re-raise the exception.
diff --git a/openhands-sdk/openhands/sdk/context/condenser/llm_summarizing_condenser.py b/openhands-sdk/openhands/sdk/context/condenser/llm_summarizing_condenser.py
index b66f9d8d47..8f6ab3b944 100644
--- a/openhands-sdk/openhands/sdk/context/condenser/llm_summarizing_condenser.py
+++ b/openhands-sdk/openhands/sdk/context/condenser/llm_summarizing_condenser.py
@@ -5,6 +5,7 @@
 from pydantic import Field, model_validator
 
 from openhands.sdk.context.condenser.base import (
+    CondensationRequirement,
     NoCondensationAvailableException,
     RollingCondenser,
 )
@@ -87,9 +88,13 @@ def get_condensation_reasons(
 
         return reasons
 
-    def should_condense(self, view: View, agent_llm: LLM | None = None) -> bool:
+    def condensation_requirement(
+        self, view: View, agent_llm: LLM | None = None
+    ) -> CondensationRequirement | None:
         reasons = self.get_condensation_reasons(view, agent_llm)
-        return reasons != set()
+        if reasons != set():
+            return CondensationRequirement.HARD
+        return None
 
     def _get_summary_event_content(self, view: View) -> str:
         """Extract the text content from the summary event in the view, if any.
diff --git a/tests/sdk/context/condenser/test_llm_summarizing_condenser.py b/tests/sdk/context/condenser/test_llm_summarizing_condenser.py
index 995f074f40..ee1ccbda19 100644
--- a/tests/sdk/context/condenser/test_llm_summarizing_condenser.py
+++ b/tests/sdk/context/condenser/test_llm_summarizing_condenser.py
@@ -4,6 +4,7 @@
 import pytest
 from litellm.types.utils import ModelResponse
 
+from openhands.sdk.context.condenser.base import CondensationRequirement
 from openhands.sdk.context.condenser.llm_summarizing_condenser import (
     LLMSummarizingCondenser,
     Reason,
@@ -108,13 +109,16 @@ def test_should_condense(mock_llm: LLM) -> None:
     small_events = [message_event(f"Event {i}") for i in range(max_size)]
     small_view = View.from_events(small_events)
 
-    assert not condenser.should_condense(small_view)
+    assert condenser.condensation_requirement(small_view) is None
 
     # Create events above the threshold
     large_events = [message_event(f"Event {i}") for i in range(max_size + 1)]
     large_view = View.from_events(large_events)
 
-    assert condenser.should_condense(large_view)
+    assert (
+        condenser.condensation_requirement(large_view)
+        == CondensationRequirement.HARD
+    )
 
 
 def test_condense_returns_view_when_no_condensation_needed(mock_llm: LLM) -> None:
diff --git a/tests/sdk/context/condenser/test_rolling_condenser.py b/tests/sdk/context/condenser/test_rolling_condenser.py
index df8b968af8..d7df33506e 100644
--- a/tests/sdk/context/condenser/test_rolling_condenser.py
+++ b/tests/sdk/context/condenser/test_rolling_condenser.py
@@ -3,6 +3,7 @@
 import pytest
 
 from openhands.sdk.context.condenser.base import (
+    CondensationRequirement,
     NoCondensationAvailableException,
     RollingCondenser,
 )
@@ -25,14 +26,16 @@ class MockRollingCondenser(RollingCondenser):
 
     def __init__(
         self,
-        should_condense_value: bool = True,
+        condensation_requirement_value: CondensationRequirement | None = None,
         raise_exception: bool = False,
     ):
-        self._should_condense_value = should_condense_value
+        self._condensation_requirement_value = condensation_requirement_value
         self._raise_exception = raise_exception
 
-    def should_condense(self, view: View, agent_llm: LLM | None = None) -> bool:
-        return self._should_condense_value
+    def condensation_requirement(
+        self, view: View, agent_llm: LLM | None = None
+    ) -> CondensationRequirement | None:
+        return self._condensation_requirement_value
 
     def get_condensation(
         self, view: View, agent_llm: LLM | None = None
@@ -51,10 +54,10 @@ def get_condensation(
 
 
 def test_rolling_condenser_returns_view_when_no_condensation_needed() -> None:
-    """Test that RollingCondenser returns the original view when should_condense returns
-    False.
+    """Test that RollingCondenser returns the original view when
+    condensation_requirement returns None.
     """
-    condenser = MockRollingCondenser(should_condense_value=False)
+    condenser = MockRollingCondenser(condensation_requirement_value=None)
 
     events: list[Event] = [
         message_event("Event 1"),
@@ -70,10 +73,13 @@ def test_rolling_condenser_returns_view_when_no_condensation_needed() -> None:
 
 
 def test_rolling_condenser_returns_condensation_when_needed() -> None:
-    """Test that RollingCondenser returns a Condensation when should_condense returns
-    True.
+    """Test that RollingCondenser returns a Condensation when condensation_requirement
+    returns HARD.
     """
-    condenser = MockRollingCondenser(should_condense_value=True, raise_exception=False)
+    condenser = MockRollingCondenser(
+        condensation_requirement_value=CondensationRequirement.HARD,
+        raise_exception=False,
+    )
 
     events: list[Event] = [
         message_event("Event 1"),
@@ -92,13 +98,16 @@ def test_rolling_condenser_returns_view_on_no_condensation_available_exception()
     None
 ):
     """Test that RollingCondenser returns the original view when
-    NoCondensationAvailableException is raised.
+    NoCondensationAvailableException is raised with SOFT requirement.
 
-    This tests the exception handling added in base.py:105-110 which catches
+    This tests the exception handling for SOFT condensation requirements which catches
     NoCondensationAvailableException from get_condensation() and returns the
     original view as a fallback.
     """
-    condenser = MockRollingCondenser(should_condense_value=True, raise_exception=True)
+    condenser = MockRollingCondenser(
+        condensation_requirement_value=CondensationRequirement.SOFT,
+        raise_exception=True,
+    )
 
     events: list[Event] = [
         message_event("Event 1"),
@@ -107,8 +116,8 @@ def test_rolling_condenser_returns_view_on_no_condensation_available_exception()
     ]
     view = View.from_events(events)
 
-    # Even though should_condense returns True, the exception should be caught
-    # and the original view should be returned
+    # Even though condensation_requirement returns SOFT, the exception should be
+    # caught and the original view should be returned
     result = condenser.condense(view)
 
     assert isinstance(result, View)
@@ -118,7 +127,10 @@ def test_rolling_condenser_returns_view_on_no_condensation_available_exception()
 
 def test_rolling_condenser_with_agent_llm() -> None:
     """Test that RollingCondenser works with optional agent_llm parameter."""
-    condenser = MockRollingCondenser(should_condense_value=True, raise_exception=False)
+    condenser = MockRollingCondenser(
+        condensation_requirement_value=CondensationRequirement.HARD,
+        raise_exception=False,
+    )
 
     events: list[Event] = [
         message_event("Event 1"),

From 0efffd9b9037e000e114210538e6eca5885215e3 Mon Sep 17 00:00:00 2001
From: Calvin Smith <calvin@all-hands.dev>
Date: Mon, 5 Jan 2026 10:32:21 -0700
Subject: [PATCH 16/19] linting

---
 openhands-sdk/openhands/sdk/context/condenser/base.py       | 6 ++++--
 .../sdk/context/condenser/test_llm_summarizing_condenser.py | 3 +--
 2 files changed, 5 insertions(+), 4 deletions(-)

diff --git a/openhands-sdk/openhands/sdk/context/condenser/base.py b/openhands-sdk/openhands/sdk/context/condenser/base.py
index 0766721f04..28741f0c2a 100644
--- a/openhands-sdk/openhands/sdk/context/condenser/base.py
+++ b/openhands-sdk/openhands/sdk/context/condenser/base.py
@@ -77,6 +77,7 @@ class NoCondensationAvailableException(Exception):
     the agent will fall back to using the uncondensed view for the next agent step.
     """
 
+
 class CondensationRequirement(Enum):
     """The type of condensation required by a rolling condenser."""
 
@@ -84,10 +85,11 @@ class CondensationRequirement(Enum):
     """Indicates that a condensation is required right now, and the agent cannot proceed
     without it.
     """
-    
+
     SOFT = "soft"
     """Indicates that a condensation is desired but not strictly required."""
 
+
 class RollingCondenser(PipelinableCondenserBase, ABC):
     """Base class for a specialized condenser strategy that applies condensation to a
     rolling history.
@@ -106,7 +108,7 @@ def condensation_requirement(
         self, view: View, agent_llm: LLM | None = None
     ) -> CondensationRequirement | None:
         """Determine how a view should be condensed.
-        
+
         Args:
             view: The current view of the conversation history.
             agent_llm: LLM instance used by the agent. Condensers use this for token
diff --git a/tests/sdk/context/condenser/test_llm_summarizing_condenser.py b/tests/sdk/context/condenser/test_llm_summarizing_condenser.py
index ee1ccbda19..1e4cda968a 100644
--- a/tests/sdk/context/condenser/test_llm_summarizing_condenser.py
+++ b/tests/sdk/context/condenser/test_llm_summarizing_condenser.py
@@ -116,8 +116,7 @@ def test_should_condense(mock_llm: LLM) -> None:
     large_view = View.from_events(large_events)
 
     assert (
-        condenser.condensation_requirement(large_view)
-        == CondensationRequirement.HARD
+        condenser.condensation_requirement(large_view) == CondensationRequirement.HARD
     )
 
 

From d2921850a093309a38eeba99f5b1428e616b89b2 Mon Sep 17 00:00:00 2001
From: Calvin Smith <calvin@all-hands.dev>
Date: Mon, 5 Jan 2026 11:14:20 -0700
Subject: [PATCH 17/19] llm summarizing condenser requirements logic updated to
 handle hard/soft reqs

---
 .../condenser/llm_summarizing_condenser.py    | 21 +++++++++++++++++--
 1 file changed, 19 insertions(+), 2 deletions(-)

diff --git a/openhands-sdk/openhands/sdk/context/condenser/llm_summarizing_condenser.py b/openhands-sdk/openhands/sdk/context/condenser/llm_summarizing_condenser.py
index 8f6ab3b944..8cbdd01df9 100644
--- a/openhands-sdk/openhands/sdk/context/condenser/llm_summarizing_condenser.py
+++ b/openhands-sdk/openhands/sdk/context/condenser/llm_summarizing_condenser.py
@@ -92,9 +92,26 @@ def condensation_requirement(
         self, view: View, agent_llm: LLM | None = None
     ) -> CondensationRequirement | None:
         reasons = self.get_condensation_reasons(view, agent_llm)
-        if reasons != set():
+
+        # No reasons => no condensation needed.
+        if reasons == set():
+            return None
+
+        # If the reasons are for resource constraints, we can treat it as a soft
+        # requirement. We want to condense when we can, but there's still space in the
+        # context window or we'd also see Reason.REQUEST. That means we can delay the
+        # condensation if there isn't one available (based on the view's manipulation
+        # indices).
+        resource_reasons = {Reason.TOKENS, Reason.EVENTS}
+        if reasons.issubset(resource_reasons):
+            return CondensationRequirement.SOFT
+
+        # Requests -- whether they come from the user or the agent -- are always hard
+        # requirements. We need to condense now because:
+        # 1. the user expects it
+        # 2. the agent has no more room in the context window and can't continue
+        if Reason.REQUEST in reasons:
             return CondensationRequirement.HARD
-        return None
 
     def _get_summary_event_content(self, view: View) -> str:
         """Extract the text content from the summary event in the view, if any.

From 965e13999aeb82239b81d505fff907b2dd415dec Mon Sep 17 00:00:00 2001
From: Calvin Smith <calvin@all-hands.dev>
Date: Mon, 5 Jan 2026 11:23:21 -0700
Subject: [PATCH 18/19] tests

---
 .../test_llm_summarizing_condenser.py         | 134 +++++++++++++++++-
 1 file changed, 131 insertions(+), 3 deletions(-)

diff --git a/tests/sdk/context/condenser/test_llm_summarizing_condenser.py b/tests/sdk/context/condenser/test_llm_summarizing_condenser.py
index 1e4cda968a..2078e410ce 100644
--- a/tests/sdk/context/condenser/test_llm_summarizing_condenser.py
+++ b/tests/sdk/context/condenser/test_llm_summarizing_condenser.py
@@ -1,5 +1,5 @@
 from typing import Any, cast
-from unittest.mock import MagicMock
+from unittest.mock import MagicMock, patch
 
 import pytest
 from litellm.types.utils import ModelResponse
@@ -111,12 +111,12 @@ def test_should_condense(mock_llm: LLM) -> None:
 
     assert condenser.condensation_requirement(small_view) is None
 
-    # Create events above the threshold
+    # Create events above the threshold (triggers EVENTS reason -> SOFT requirement)
     large_events = [message_event(f"Event {i}") for i in range(max_size + 1)]
     large_view = View.from_events(large_events)
 
     assert (
-        condenser.condensation_requirement(large_view) == CondensationRequirement.HARD
+        condenser.condensation_requirement(large_view) == CondensationRequirement.SOFT
     )
 
 
@@ -574,3 +574,131 @@ def test_generate_condensation_raises_on_zero_events(mock_llm: LLM) -> None:
 
     # Verify the LLM was never called
     cast(MagicMock, mock_llm.completion).assert_not_called()
+
+
+@pytest.mark.parametrize(
+    "reasons",
+    [set()],
+)
+def test_condensation_requirement_returns_none(
+    mock_llm: LLM, reasons: set[Reason]
+) -> None:
+    """Test that condensation_requirement returns None when appropriate.
+
+    Mocks get_condensation_reasons to test different reason combinations.
+    """
+    condenser = LLMSummarizingCondenser(llm=mock_llm, max_size=100, keep_first=2)
+    events: list[Event] = [message_event(f"Event {i}") for i in range(10)]
+    view = View.from_events(events)
+
+    with patch.object(
+        LLMSummarizingCondenser, "get_condensation_reasons", return_value=reasons
+    ):
+        result = condenser.condensation_requirement(view)
+        assert result is None
+
+
+@pytest.mark.parametrize(
+    "reasons",
+    [
+        {Reason.TOKENS},
+        {Reason.EVENTS},
+        {Reason.TOKENS, Reason.EVENTS},
+    ],
+)
+def test_condensation_requirement_returns_soft(
+    mock_llm: LLM, reasons: set[Reason]
+) -> None:
+    """Test that condensation_requirement returns SOFT for resource constraints.
+
+    Mocks get_condensation_reasons to test different resource reason combinations.
+    """
+    condenser = LLMSummarizingCondenser(llm=mock_llm, max_size=100, keep_first=2)
+    events: list[Event] = [message_event(f"Event {i}") for i in range(10)]
+    view = View.from_events(events)
+
+    with patch.object(
+        LLMSummarizingCondenser, "get_condensation_reasons", return_value=reasons
+    ):
+        result = condenser.condensation_requirement(view)
+        assert result == CondensationRequirement.SOFT
+
+
+@pytest.mark.parametrize(
+    "reasons",
+    [
+        {Reason.REQUEST},
+        {Reason.REQUEST, Reason.TOKENS},
+        {Reason.REQUEST, Reason.EVENTS},
+        {Reason.REQUEST, Reason.TOKENS, Reason.EVENTS},
+    ],
+)
+def test_condensation_requirement_returns_hard(
+    mock_llm: LLM, reasons: set[Reason]
+) -> None:
+    """Test that condensation_requirement returns HARD when REQUEST is present.
+
+    Mocks get_condensation_reasons to test different combinations with REQUEST.
+    """
+    condenser = LLMSummarizingCondenser(llm=mock_llm, max_size=100, keep_first=2)
+    events: list[Event] = [message_event(f"Event {i}") for i in range(10)]
+    view = View.from_events(events)
+
+    with patch.object(
+        LLMSummarizingCondenser, "get_condensation_reasons", return_value=reasons
+    ):
+        result = condenser.condensation_requirement(view)
+        assert result == CondensationRequirement.HARD
+
+
+def test_condense_with_hard_requirement_and_no_condensation_available(
+    mock_llm: LLM,
+) -> None:
+    """Test that condense raises error with hard requirement but no condensation.
+
+    When there's a hard requirement but no valid condensation range available
+    (e.g., entire view is a single atomic unit), should raise an exception.
+    """
+    from openhands.sdk.context.condenser.base import NoCondensationAvailableException
+
+    condenser = LLMSummarizingCondenser(llm=mock_llm, max_size=100, keep_first=2)
+    events: list[Event] = [message_event(f"Event {i}") for i in range(10)]
+    view = View.from_events(events)
+
+    # Mock to return HARD requirement but no events to condense
+    with patch.object(
+        LLMSummarizingCondenser,
+        "get_condensation_reasons",
+        return_value={Reason.REQUEST},
+    ), patch.object(
+        condenser, "_get_forgotten_events", return_value=([], 0)
+    ):
+        with pytest.raises(NoCondensationAvailableException):
+            condenser.condense(view)
+
+
+def test_condense_with_soft_requirement_and_no_condensation_available(
+    mock_llm: LLM,
+) -> None:
+    """Test that condense returns view with soft requirement but no condensation.
+
+    When there's a soft requirement but no valid condensation range available,
+    should return the original view unchanged.
+    """
+    condenser = LLMSummarizingCondenser(llm=mock_llm, max_size=100, keep_first=2)
+    events: list[Event] = [message_event(f"Event {i}") for i in range(10)]
+    view = View.from_events(events)
+
+    # Mock to return SOFT requirement but no events to condense
+    with patch.object(
+        LLMSummarizingCondenser,
+        "get_condensation_reasons",
+        return_value={Reason.EVENTS},
+    ), patch.object(
+        condenser, "_get_forgotten_events", return_value=([], 0)
+    ):
+        result = condenser.condense(view)
+        assert isinstance(result, View)
+        assert result == view
+        # LLM should not be called
+        cast(MagicMock, mock_llm.completion).assert_not_called()

From 27f0bb962a04e11a497cfcf0dfafe6906103f771 Mon Sep 17 00:00:00 2001
From: Calvin Smith <calvin@all-hands.dev>
Date: Mon, 5 Jan 2026 11:23:45 -0700
Subject: [PATCH 19/19] linting

---
 .../test_llm_summarizing_condenser.py         | 26 ++++++++++---------
 1 file changed, 14 insertions(+), 12 deletions(-)

diff --git a/tests/sdk/context/condenser/test_llm_summarizing_condenser.py b/tests/sdk/context/condenser/test_llm_summarizing_condenser.py
index 2078e410ce..ea91347330 100644
--- a/tests/sdk/context/condenser/test_llm_summarizing_condenser.py
+++ b/tests/sdk/context/condenser/test_llm_summarizing_condenser.py
@@ -666,12 +666,13 @@ def test_condense_with_hard_requirement_and_no_condensation_available(
     view = View.from_events(events)
 
     # Mock to return HARD requirement but no events to condense
-    with patch.object(
-        LLMSummarizingCondenser,
-        "get_condensation_reasons",
-        return_value={Reason.REQUEST},
-    ), patch.object(
-        condenser, "_get_forgotten_events", return_value=([], 0)
+    with (
+        patch.object(
+            LLMSummarizingCondenser,
+            "get_condensation_reasons",
+            return_value={Reason.REQUEST},
+        ),
+        patch.object(condenser, "_get_forgotten_events", return_value=([], 0)),
     ):
         with pytest.raises(NoCondensationAvailableException):
             condenser.condense(view)
@@ -690,12 +691,13 @@ def test_condense_with_soft_requirement_and_no_condensation_available(
     view = View.from_events(events)
 
     # Mock to return SOFT requirement but no events to condense
-    with patch.object(
-        LLMSummarizingCondenser,
-        "get_condensation_reasons",
-        return_value={Reason.EVENTS},
-    ), patch.object(
-        condenser, "_get_forgotten_events", return_value=([], 0)
+    with (
+        patch.object(
+            LLMSummarizingCondenser,
+            "get_condensation_reasons",
+            return_value={Reason.EVENTS},
+        ),
+        patch.object(condenser, "_get_forgotten_events", return_value=([], 0)),
     ):
         result = condenser.condense(view)
         assert isinstance(result, View)