fix(condenser): Retry on empty condensation (#1577)

Co-authored-by: Calvin Smith <[email protected]>
Co-authored-by: Xingyao Wang <[email protected]>

Author: 3 people

openhands-sdk/openhands/sdk/agent/agent.py

@@ -27,7 +27,10 @@
     TokenEvent,
     UserRejectObservation,
 )
-from openhands.sdk.event.condenser import Condensation, CondensationRequest
+from openhands.sdk.event.condenser import (
+    Condensation,
+    CondensationRequest,
+)
 from openhands.sdk.llm import (
     LLMResponse,
     Message,

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",
 ]

openhands-sdk/openhands/sdk/context/condenser/base.py

@@ -1,4 +1,5 @@
 from abc import ABC, abstractmethod
+from enum import Enum
 from logging import getLogger
 
 from openhands.sdk.context.view import View
@@ -66,22 +67,57 @@ class PipelinableCondenserBase(CondenserBase):
     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 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
     rolling history.
 
     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(
@@ -92,8 +128,23 @@ def get_condensation(
     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):
-            return self.get_condensation(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}")
+
+                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.
+                else:
+                    raise e
 
         # Otherwise we're safe to just return the view.
         else:

openhands-sdk/openhands/sdk/context/condenser/llm_summarizing_condenser.py

@@ -4,7 +4,11 @@
 
 from pydantic import Field, model_validator
 
-from openhands.sdk.context.condenser.base import RollingCondenser
+from openhands.sdk.context.condenser.base import (
+    CondensationRequirement,
+    NoCondensationAvailableException,
+    RollingCondenser,
+)
 from openhands.sdk.context.condenser.utils import (
     get_suffix_length_for_token_reduction,
     get_total_token_count,
@@ -84,9 +88,30 @@ 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()
+
+        # 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
 
     def _get_summary_event_content(self, view: View) -> str:
         """Extract the text content from the summary event in the view, if any.
@@ -124,12 +149,7 @@ 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 +256,19 @@ 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,

openhands-sdk/openhands/sdk/context/view.py

@@ -489,9 +489,11 @@ 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
+
         for event in reversed(events):
             if isinstance(event, Condensation):
                 break
+
             if isinstance(event, CondensationRequest):
                 unhandled_condensation_request = True
                 break

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.",
         )