feat(condenser): Hard context reset on unrecoverable error

[csmith49]

Summary

Extends the rolling condenser API to support hard context resets, and adds one such strategy for the LLM summarizing condenser. This strategy triggers when a hard condensation request comes through but one can't be found.

The quality of the summary and the agent's behavior afterwards is not likely to be good. This is intended solely as a last-ditch fallback option.

Changes

• Extends RollingCondenser base class to expose hard_context_reset -- an optional strategy for last-ditch context management
• Implementation of a hard context reset strategy for LLMSummarizingCondenser
• Updating the c02 condensation integration test to check for the hard context reset instead of the expected error

Checklist

• If the PR is changing/adding functionality, are there tests to reflect this?
• If there is an example, have you run the example to make sure that it works?
• If there are instructions on how to run the code, have you followed the instructions and made sure that it works?
• If the feature is significant enough to require documentation, is there a PR open on the OpenHands/docs repository with the same branch name?
• Is the github CI passing?

---

Agent Server images for this PR

• GHCR package: https://github.com/OpenHands/agent-sdk/pkgs/container/agent-server

Variants & Base Images

Variant	Architectures	Base Image	Docs / Tags
java	amd64, arm64	eclipse-temurin:17-jdk	Link
python	amd64, arm64	nikolaik/python-nodejs:python3.12-nodejs22	Link
golang	amd64, arm64	golang:1.21-bookworm	Link

Pull (multi-arch manifest)

# Each variant is a multi-arch manifest supporting both amd64 and arm64
docker pull ghcr.io/openhands/agent-server:de3797c-python

Run

docker run -it --rm \
  -p 8000:8000 \
  --name agent-server-de3797c-python \
  ghcr.io/openhands/agent-server:de3797c-python

All tags pushed for this build

ghcr.io/openhands/agent-server:de3797c-golang-amd64
ghcr.io/openhands/agent-server:de3797c-golang_tag_1.21-bookworm-amd64
ghcr.io/openhands/agent-server:de3797c-golang-arm64
ghcr.io/openhands/agent-server:de3797c-golang_tag_1.21-bookworm-arm64
ghcr.io/openhands/agent-server:de3797c-java-amd64
ghcr.io/openhands/agent-server:de3797c-eclipse-temurin_tag_17-jdk-amd64
ghcr.io/openhands/agent-server:de3797c-java-arm64
ghcr.io/openhands/agent-server:de3797c-eclipse-temurin_tag_17-jdk-arm64
ghcr.io/openhands/agent-server:de3797c-python-amd64
ghcr.io/openhands/agent-server:de3797c-nikolaik_s_python-nodejs_tag_python3.12-nodejs22-amd64
ghcr.io/openhands/agent-server:de3797c-python-arm64
ghcr.io/openhands/agent-server:de3797c-nikolaik_s_python-nodejs_tag_python3.12-nodejs22-arm64
ghcr.io/openhands/agent-server:de3797c-golang
ghcr.io/openhands/agent-server:de3797c-java
ghcr.io/openhands/agent-server:de3797c-python

About Multi-Architecture Support

• Each variant tag (e.g., de3797c-python) is a multi-arch manifest supporting both amd64 and arm64
• Docker automatically pulls the correct architecture for your platform
• Individual architecture tags (e.g., de3797c-python-amd64) are also available if needed

[github-actions]

Coverage Report •

File	Stmts	Miss	Cover	Missing
openhands-sdk/openhands/sdk/context/condenser
base.py	45	3	93%	62, 176–177
llm_summarizing_condenser.py	105	6	94%	130–132, 261–262, 264
TOTAL	15131	4442	70%

[github-actions]

Hi! I started running the condenser tests on your PR. You will receive a comment with the results shortly.

Note: These are non-blocking tests that validate condenser functionality across different LLMs.

[github-actions]

Condenser Test Results (Non-Blocking)

These tests validate condenser functionality and do not block PR merges.

🧪 Condenser Tests Results

Overall Success Rate: 100.0%
Total Cost: $0.82
Models Tested: 2
Timestamp: 2026-01-12 22:29:50 UTC

📊 Summary

Model	Overall	Tests Passed	Skipped	Total	Cost	Tokens
litellm_proxy_gpt_5.1_codex_max	100.0%	2/2	3	5	$0.02	24,860
litellm_proxy_anthropic_claude_opus_4_5_20251101	100.0%	5/5	0	5	$0.80	394,617

📋 Detailed Results

litellm_proxy_gpt_5.1_codex_max

• Success Rate: 100.0% (2/2)
• Total Cost: $0.02
• Token Usage: prompt: 23,788, completion: 1,072, cache_read: 14,464, reasoning: 576
• Run Suffix: litellm_proxy_gpt_5.1_codex_max_585f429_gpt51_condenser_run_N5_20260112_222615
• Skipped Tests: 3

Skipped Tests:

• c01_thinking_block_condenser: Model litellm_proxy/gpt-5.1-codex-max does not support extended thinking or reasoning effort
• c05_size_condenser: This test stresses long repetitive tool loops to trigger size-based condensation. GPT-5.1 Codex Max often declines such requests for efficiency/safety reasons.
• c04_token_condenser: This test stresses long repetitive tool loops to trigger token-based condensation. GPT-5.1 Codex Max often declines such requests for efficiency/safety reasons.

litellm_proxy_anthropic_claude_opus_4_5_20251101

• Success Rate: 100.0% (5/5)
• Total Cost: $0.80
• Token Usage: prompt: 380,120, completion: 14,497, cache_read: 334,767, cache_write: 30,897, reasoning: 2,013
• Run Suffix: litellm_proxy_anthropic_claude_opus_4_5_20251101_585f429_opus_condenser_run_N5_20260112_222607

[all-hands-bot]

Overall this is a solid implementation of a fallback strategy for handling unrecoverable context situations. The approach is reasonable, but there are a few important issues around error handling and observability that should be addressed. See inline comments for details.

[all-hands-bot]

🟢 Nit: Comment could be more specific about which situations trigger re-raise.

For clarity, consider:

Suggested change

	# In all other situations re-raise the exception.
	# Re-raise if: (1) request is HARD but hard_context_reset returned None,
	# or (2) request is neither SOFT nor HARD
	raise e

[csmith49]

This PR looks like it works -- until you try to continue a conversation after a hard reset. The View objects can only support a single summary, so things start to break after the first non-hard-reset condensation.

I'm working on an extension to address this issue in a separate PR, and will proceed with this one when that feature is merged.

[openhands-ai]

Looks like there are a few issues preventing this PR from being merged!

• GitHub Actions are failing:
  • Agent Server
  • Pre-commit checks

If you'd like me to help, just leave a comment, like

@OpenHands please fix the failing actions on PR #1596 at branch `feat/hard-context-reset`

Feel free to include any additional details that might help me get this PR into a better state.

_{^{You can manage your notification settings}}