fix(condenser): show the summarizer whole events, not their 500-char preview

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

@@ -186,6 +186,21 @@ def _should_handle_non_multimodal_image_input(
     return _latest_user_message_contains_image(messages) and not llm.vision_is_active()
 
 
+def _mask_condensation(
+    condensation: Condensation, conversation: LocalConversation
+) -> Condensation:
+    """Mask registered secrets in a condensation's summary before it is emitted.
+
+    The summary is durable model output, replayed to the agent on every later
+    step, and the summarizer sees each forgotten action's arguments verbatim --
+    a token typed into a git URL included. Masked like any agent message (#4783).
+    """
+    if not condensation.summary:
+        return condensation
+    mask = conversation.state.secret_registry.mask_secrets_in_output
+    return condensation.model_copy(update={"summary": mask(condensation.summary)})
+
+
 @dataclass(frozen=True, slots=True)
 class _ActionBatch:
     """Immutable result of preparing a batch of actions for execution.
@@ -687,7 +702,7 @@ def step(
 
         # Process condensation event before agent sampels another action
         if isinstance(_messages_or_condensation, Condensation):
-            on_event(_messages_or_condensation)
+            on_event(_mask_condensation(_messages_or_condensation, conversation))
             return
 
         _messages = _messages_or_condensation
@@ -900,7 +915,7 @@ async def astep(
         )
 
         if isinstance(_messages_or_condensation, Condensation):
-            on_event(_messages_or_condensation)
+            on_event(_mask_condensation(_messages_or_condensation, conversation))
             return
 
         _messages = _messages_or_condensation

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

@@ -1,7 +1,7 @@
 import os
 from collections.abc import Sequence
 from enum import Enum
-from typing import Final
+from typing import ClassVar, Final
 
 from pydantic import Field, model_validator
 
@@ -13,15 +13,16 @@
 from openhands.sdk.context.condenser.utils import (
     get_suffix_length_for_token_reduction,
     get_total_token_count,
+    render_event_for_summary,
 )
 from openhands.sdk.context.prompts import render_template
 from openhands.sdk.context.view import View
 from openhands.sdk.event.base import LLMConvertibleEvent
-from openhands.sdk.event.condenser import Condensation
+from openhands.sdk.event.condenser import Condensation, CondensationSummaryEvent
 from openhands.sdk.llm import LLM, Message, TextContent
+from openhands.sdk.llm.exceptions.classifier import is_context_window_exceeded
 from openhands.sdk.logger import get_logger
 from openhands.sdk.observability.laminar import observe
-from openhands.sdk.utils import maybe_truncate
 
 
 logger = get_logger(__name__)
@@ -62,6 +63,13 @@ class LLMSummarizingCondenser(RollingCondenser):
     are treated as errors.
     """
 
+    max_event_length: ClassVar[int] = 10_000
+    """Longest a single forgotten event may be when shown to the summarizer, in
+    characters; longer ones keep their head and tail. A previous summary is exempt:
+    it is the agent's only record of everything condensed before it. A class
+    constant, not a field, so the condenser's settings schema is unchanged.
+    """
+
     hard_context_reset_max_retries: int = Field(default=5, gt=0)
     """Number of attempts to perform hard context reset before raising an error."""
 
@@ -186,6 +194,129 @@ def condensation_requirement(
         if Reason.REQUEST in reasons:
             return CondensationRequirement.HARD
 
+    def _render_forgotten_events(
+        self,
+        forgotten_events: Sequence[LLMConvertibleEvent],
+        max_event_str_length: int | None = None,
+    ) -> list[str]:
+        """Render forgotten events for the summarizer, whole up to the length caps.
+
+        Every event is capped at ``max_event_length``, except a previous summary.
+        ``max_event_str_length`` -- set only when the summarizer overflowed and the
+        input is being shrunk -- caps every event, the previous summary included.
+        """
+        rendered = []
+        for event in forgotten_events:
+            limits = [max_event_str_length]
+            if not isinstance(event, CondensationSummaryEvent):
+                limits.append(self.max_event_length)
+            caps = [limit for limit in limits if limit is not None]
+            rendered.append(
+                render_event_for_summary(event, min(caps) if caps else None)
+            )
+        return rendered
+
+    def _shrunk_event_length(
+        self,
+        events: Sequence[LLMConvertibleEvent],
+        max_event_str_length: int | None,
+    ) -> int:
+        """The next per-event cap after the summarizer call failed.
+
+        Chosen so the rendered total -- not just the longest event -- falls to
+        ``hard_context_reset_context_scaling`` of what was just sent: scaling the
+        longest event alone leaves every shorter one untouched, so a prompt made
+        of many mid-sized events is re-sent almost unchanged. Returns 0 when no
+        positive cap gets there, i.e. there is nothing left to cut.
+        """
+        lengths = [
+            len(text)
+            for text in self._render_forgotten_events(events, max_event_str_length)
+        ]
+        target = sum(lengths) * self.hard_context_reset_context_scaling
+        # The largest cap whose capped total fits the target (capping never grows
+        # an event, so the total is monotonic in the cap).
+        low, high = 0, max(lengths)
+        while low < high:
+            cap = (low + high + 1) // 2
+            if sum(min(length, cap) for length in lengths) <= target:
+                low = cap
+            else:
+                high = cap - 1
+        return low
+
+    def _summarize(
+        self, forgotten_events: Sequence[LLMConvertibleEvent], summary_offset: int
+    ) -> Condensation:
+        """Summarize the forgotten events whole; shrink them only on overflow.
+
+        Whole events can exceed a summarizer whose window is smaller than the
+        agent's. Rather than fail -- and fall through to a hard context reset that
+        forgets everything -- retry with every event cut further each time. Any
+        other failure is raised at once.
+        """
+        max_event_str_length: int | None = None
+        for _ in range(self.hard_context_reset_max_retries):
+            try:
+                return self._generate_condensation(
+                    forgotten_events=forgotten_events,
+                    summary_offset=summary_offset,
+                    max_event_str_length=max_event_str_length,
+                )
+            except NoCondensationAvailableException as e:
+                cause = e.__cause__
+                if not isinstance(cause, Exception) or not is_context_window_exceeded(
+                    cause
+                ):
+                    raise
+                max_event_str_length = self._shrunk_event_length(
+                    forgotten_events, max_event_str_length
+                )
+                if max_event_str_length <= 0:
+                    raise NoCondensationAvailableException(
+                        "Summarizer context exceeded with nothing left to cut."
+                    ) from e
+                logger.warning(
+                    "Summarizer context exceeded; retrying with each event cut to "
+                    f"{max_event_str_length} characters."
+                )
+        raise NoCondensationAvailableException(
+            "Summarizer context exceeded at every event length tried."
+        )
+
+    async def _asummarize(
+        self, forgotten_events: Sequence[LLMConvertibleEvent], summary_offset: int
+    ) -> Condensation:
+        """Async variant of :meth:`_summarize`."""
+        max_event_str_length: int | None = None
+        for _ in range(self.hard_context_reset_max_retries):
+            try:
+                return await self._agenerate_condensation(
+                    forgotten_events=forgotten_events,
+                    summary_offset=summary_offset,
+                    max_event_str_length=max_event_str_length,
+                )
+            except NoCondensationAvailableException as e:
+                cause = e.__cause__
+                if not isinstance(cause, Exception) or not is_context_window_exceeded(
+                    cause
+                ):
+                    raise
+                max_event_str_length = self._shrunk_event_length(
+                    forgotten_events, max_event_str_length
+                )
+                if max_event_str_length <= 0:
+                    raise NoCondensationAvailableException(
+                        "Summarizer context exceeded with nothing left to cut."
+                    ) from e
+                logger.warning(
+                    "Summarizer context exceeded; retrying with each event cut to "
+                    f"{max_event_str_length} characters."
+                )
+        raise NoCondensationAvailableException(
+            "Summarizer context exceeded at every event length tried."
+        )
+
     def _generate_condensation(
         self,
         forgotten_events: Sequence[LLMConvertibleEvent],
@@ -209,11 +340,9 @@ def _generate_condensation(
         """
         assert len(forgotten_events) > 0, "No events to condense."
 
-        # Convert events to strings for the template
-        event_strings = [
-            maybe_truncate(str(forgotten_event), truncate_after=max_event_str_length)
-            for forgotten_event in forgotten_events
-        ]
+        event_strings = self._render_forgotten_events(
+            forgotten_events, max_event_str_length
+        )
 
         prompt = render_template(
             os.path.join(os.path.dirname(__file__), "prompts"),
@@ -342,15 +471,10 @@ def hard_context_reset(
                     max_event_str_length=max_event_str_length,
                 )
             except Exception as e:
-                # If we haven't set a max_event_str_length yet, set it as the largest
-                # event string length.
-                if max_event_str_length is None:
-                    max_event_str_length = max(len(str(event)) for event in view.events)
-
-                # Since the summarization failed, reduce the max_event_str_length by 20%
-                assert max_event_str_length is not None
-                max_event_str_length = int(
-                    max_event_str_length * self.hard_context_reset_context_scaling
+                # Since the summarization failed, cut every event to 80% of the
+                # longest one as rendered -- not of its 500-character preview.
+                max_event_str_length = max(
+                    1, self._shrunk_event_length(view.events, max_event_str_length)
                 )
 
                 # Log the exception so we can track these failures
@@ -393,10 +517,7 @@ def get_condensation(
                 "threshold."
             )
 
-        return self._generate_condensation(
-            forgotten_events=forgotten_events,
-            summary_offset=summary_offset,
-        )
+        return self._summarize(forgotten_events, summary_offset)
 
     # ------------------------------------------------------------------
     # Async variants
@@ -411,10 +532,9 @@ async def _agenerate_condensation(
         """Async variant of :meth:`_generate_condensation`."""
         assert len(forgotten_events) > 0, "No events to condense."
 
-        event_strings = [
-            maybe_truncate(str(fe), truncate_after=max_event_str_length)
-            for fe in forgotten_events
-        ]
+        event_strings = self._render_forgotten_events(
+            forgotten_events, max_event_str_length
+        )
 
         prompt = render_template(
             os.path.join(os.path.dirname(__file__), "prompts"),
@@ -475,10 +595,7 @@ async def aget_condensation(
                 "progress threshold."
             )
 
-        return await self._agenerate_condensation(
-            forgotten_events=forgotten_events,
-            summary_offset=summary_offset,
-        )
+        return await self._asummarize(forgotten_events, summary_offset)
 
     async def ahard_context_reset(
         self,
@@ -497,11 +614,8 @@ async def ahard_context_reset(
                     max_event_str_length=max_event_str_length,
                 )
             except Exception as e:
-                if max_event_str_length is None:
-                    max_event_str_length = max(len(str(ev)) for ev in view.events)
-                assert max_event_str_length is not None
-                max_event_str_length = int(
-                    max_event_str_length * self.hard_context_reset_context_scaling
+                max_event_str_length = max(
+                    1, self._shrunk_event_length(view.events, max_event_str_length)
                 )
                 logger.warning(
                     f"Hard context reset summarization failed: {e}. "

openhands-sdk/openhands/sdk/context/condenser/prompts/summarizing_prompt.j2

@@ -1,5 +1,7 @@
 You are maintaining a context-aware state summary for an interactive agent.
 You will be given a list of events corresponding to actions taken by the agent, which will include previous summaries.
+A PREVIOUS SUMMARY is the agent's only memory of the work before it. Carry every fact in it forward (files read and what they contain, commands run and their results, decisions, changes) unless a later event changes it, and never report as not done or not checked what it or an event shows was done.
+Keep every request, answer and correction the user gave; quote short answers (ids, names, values) verbatim.
 If the events being summarized contain ANY task-tracking, you MUST include a TASK_TRACKING section to maintain continuity.
 When referencing tasks make sure to preserve exact task IDs and statuses.
 

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

@@ -1,8 +1,73 @@
 from collections.abc import Sequence
 
 from openhands.sdk.event.base import LLMConvertibleEvent
+from openhands.sdk.event.condenser import CondensationSummaryEvent
+from openhands.sdk.event.llm_convertible import (
+    ActionEvent,
+    AgentErrorEvent,
+    MessageEvent,
+    ObservationEvent,
+    UserRejectObservation,
+)
 from openhands.sdk.event.llm_convertible.system import SystemPromptEvent
-from openhands.sdk.llm import LLM
+from openhands.sdk.llm import LLM, content_to_str
+from openhands.sdk.utils import maybe_truncate
+
+
+def render_event_for_summary(
+    event: LLMConvertibleEvent, max_length: int | None = None
+) -> str:
+    """Render one forgotten event as the summarizing LLM should see it.
+
+    Not ``str(event)``: that is a display preview capped at ``N_CHAR_PREVIEW``
+    (500) characters, and it renders an action as its class name alone. Fed to
+    the summarizer it cut every previous summary to its first ~500 characters
+    and hid which command ran, which file was read and what a long user message
+    ended with -- so each condensation forgot the work the last one recorded,
+    and the agent repeated it.
+
+    This shows what the agent itself was shown: a previous summary whole, an
+    action as its tool and arguments, a result, error or user message as its
+    text. ``max_length`` caps the rendered event, keeping head and tail; event
+    types not listed here keep their ``str()`` form.
+    """
+    if isinstance(event, CondensationSummaryEvent):
+        text = f"PREVIOUS SUMMARY ({event.source})\n{event.summary}"
+    elif isinstance(event, ActionEvent):
+        lines = [f"ActionEvent ({event.source})", f"  Tool: {event.tool_name}"]
+        if event.summary:
+            lines.append(f"  Summary: {event.summary}")
+        thought = " ".join(t.text for t in event.thought).strip()
+        if thought:
+            lines.append(f"  Thought: {thought}")
+        lines.append(f"  Arguments: {event.tool_call.arguments}")
+        if event.action is None:
+            lines.append("  (not executed)")
+        text = "\n".join(lines)
+    elif isinstance(event, ObservationEvent):
+        result = "".join(content_to_str(event.observation.to_llm_content))
+        text = (
+            f"ObservationEvent ({event.source})\n"
+            f"  Tool: {event.tool_name}\n  Result: {result}"
+        )
+    elif isinstance(event, UserRejectObservation):
+        text = (
+            f"UserRejectObservation ({event.source})\n  Tool: {event.tool_name}\n"
+            f"  Rejected by {event.rejection_source}: {event.rejection_reason}"
+        )
+    elif isinstance(event, AgentErrorEvent):
+        text = (
+            f"AgentErrorEvent ({event.source})\n  Tool: {event.tool_name}\n"
+            f"  Error: {event.error}"
+        )
+    elif isinstance(event, MessageEvent):
+        body = " ".join(content_to_str(event.llm_message.content)) or "[no text]"
+        text = f"MessageEvent ({event.source})\n  {event.llm_message.role}: {body}"
+        if event.activated_skills:
+            text += f"\n  [Skills: {', '.join(event.activated_skills)}]"
+    else:
+        text = str(event)
+    return maybe_truncate(text, truncate_after=max_length)
 
 
 def get_total_token_count(