[Bug] Condensation for 0 events

[xingyaoww]

I got these errors:

{"kind": "Condensation", "id": "1dd50966-6311-4337-8cc8-f20f8ce3f41d", "timestamp": "2025-12-28T18:28:51.883055", "source": "environment", "forgotten_event_ids": [], "summary": "I don't see any events provided to summarize. Please share the list of events that need to be summarized, and I'll create a context-aware state summary following the format guidelines.\n\nIf you're starting a new session, please provide:\n1. The events/actions taken by the agent\n2. Any relevant context about the task being performed\n\nI'll then generate an appropriate summary tracking user context, task status, and any relevant state information.", "summary_offset": 122, "llm_response_id": "chatcmpl-8fb35e47-ca68-4f72-bd13-bc38a07d7a94"}

DM me for the covnersation event dumps

[xingyaoww]

Root Cause Analysis

After analyzing the event dump, I've identified the root cause of this bug.

The Problem

When a tool loop (a sequence of ActionEvent/ObservationEvent pairs starting with thinking blocks) spans almost the entire view, the manipulation_indices computed by View leave no valid range for forgetting events.

From the event dump:

• Total events before first condensation: 122
• Event structure:
  • Event 0: SystemPromptEvent
  • Event 1: MessageEvent (user message)
  • Events 2-121: A continuous tool loop (60 action-observation pairs)

The tool loop starts at index 2 (first ActionEvent with thinking blocks) and continues uninterrupted to index 121.

Manipulation Indices

The View.manipulation_indices property computes safe boundaries for forgetting events while preserving atomic units. For this event sequence:

manipulation_indices = [0, 1, 2, 122]

This means:

• Index 0: Before SystemPromptEvent
• Index 1: Before MessageEvent
• Index 2: Start of tool loop
• Index 122: End of tool loop (and end of view)

Condensation Logic Failure

When LLMSummarizingCondenser._get_forgotten_events() runs with keep_first=4:

1. forgetting_start = find_next_manipulation_index(keep_first=4, strict=True) → Returns 122 (smallest index > 4)
2. forgetting_end = find_next_manipulation_index(naive_end=67, strict=False) → Returns 122 (smallest index >= 67)
3. forgotten_events = view[122:122] → Empty list!

The condenser still generates a Condensation event with:

• forgotten_event_ids = [] (empty)
• summary_offset = 122

This triggers an LLM call to summarize zero events, producing meaningless output like:

"I don't see any events provided to summarize..."

Impact

From the dump, there are 53 consecutive Condensation events (indices 122-174) all with:

• forgotten_event_ids: []
• summary_offset: 122

Each one represents a wasted LLM API call.

Suggested Fix

In LLMSummarizingCondenser.condense() or _get_forgotten_events(), check if forgetting_start >= forgetting_end (or len(forgotten_events) == 0) and return the original View instead of generating an empty Condensation.

Test Case

I've added a test case that reproduces this bug:

def test_condense_tool_loop_spans_entire_view_returns_view_not_empty_condensation(mock_llm):
    """Test that when a tool loop spans the entire view, condense returns View not Condensation."""
    # Creates 60 action-observation pairs forming one continuous tool loop
    # Verifies that condenser should return View (not Condensation with empty forgotten_event_ids)

The test currently fails, confirming the bug exists.

[csmith49]

Some idle thoughts on fixing this:

Manipulation indices

The manipulation indices being [0, 1, 2, 122] makes sense but is probably not 100% safe/correct. Say we condense from 2 -> 122, but the LLM is not done with the tool loop -- after we condense, is the tool loop interrupted? If so, we might be losing out on some performance (interrupting the LLM in the middle of a "thought"), and if not, we're right back to the bug the tool-loop aware fixing was trying to resolve.

I'm guessing the answer is the former. Either way we could potentially resolve this by adjusting the tool-loop manipulation index calculation to only mark the end of a tool loop when we definitively see a non-tool-loop message. That would give manipulation indices of [0, 1, 2] here.

Delaying condensation

The LLM summarizing condenser has a two-step process defined by the rolling condenser base class. If we want to delay condensation we'll have to modify that interface.

One way to do this would be to define a new NoCondensationAvailable exception that we can raise in the RollingCondenser.get_condensation abstract method. In the LLM summarizing condenser we could raise this if we end up with no forgotten events, then in RollingCondenser.condense we do a try-catch and return the original view if we see NoCondensationAvailable (EAFP-style).

This does mean that between when a condenser is triggered and a condensation is available we'd be constantly trying to condense. If we're not doing token-counting I think the overhead will be pretty minimal though (and we have another issue open to resolve inefficiencies with the token counting (see #1502)).

[csmith49]

A quick short-term fix would be:

• Reducing the default attribute LLMSummarizingCondenser.keep_first to 2.
• Raising the default attribute LLMSummarizingCondenser.max_size to 240.

That way we're much more likely to get workable manipulation indices for condensation to act on.

[xingyaoww]

@OpenHands please help me implement the shorterm fix here in a new PR: #1518 (comment)

[openhands-ai]

I'm on it! xingyaoww can track my progress at all-hands.dev

[openhands-ai]

Summary

I've successfully implemented the short-term fix for issue #1518 as suggested in the comment by @csmith49.

Changes Made

1. Modified LLMSummarizingCondenser default parameters (openhands-sdk/openhands/sdk/context/condenser/llm_summarizing_condenser.py):

   • Reduced keep_first from 4 to 2
   • Raised max_size from 120 to 240

2. Added test for default values (tests/sdk/context/condenser/test_llm_summarizing_condenser.py):

   • Added test_default_values() to verify the new defaults are correctly set

Checklist

• ✅ Implemented the short-term fix as specified in the issue comment
• ✅ Added appropriate test to verify the fix
• ✅ All pre-commit hooks pass
• ✅ All 1562 tests pass
• ✅ Changes are minimal and focused on the fix
• ✅ Created branch openhands/fix-condensation-defaults
• ✅ Pushed to remote
• ✅ Opened draft PR

Pull Request

PR #1523: #1523

The PR is opened as a draft and fixes #1518. These parameter adjustments make it more likely to get workable manipulation indices for condensation to act on, preventing the bug where condensation fails when a tool loop spans almost the entire view.

_{View full conversation}