Enforce batch atomicity for condenser

[ryanhoangt]

Fix #738

---

Integration tests: 7/7 (100%)

[10/17/25 13:48:27] INFO     Success rate: 100.00% (7/7)             run_infer.py:278
[10/17/25 13:48:27] INFO     Evaluation Results:                     run_infer.py:279
[10/17/25 13:48:27] INFO     t04_git_staging: ✓ - Successfully       run_infer.py:283
                             committed changes with message: 'Add                    
                             hello world Python script'                              
[10/17/25 13:48:27] INFO     t07_interactive_commands: ✓ -           run_infer.py:283
                             Interactive Python script setup                         
                             completed. Agent should execute the                     
                             script with inputs 'John' and '25' and                  
                             find the secret number: 707                             
[10/17/25 13:48:27] INFO     t05_simple_browsing: ✓ - Agent          run_infer.py:283
                             successfully found the answer! Matched                  
                             pattern: (?i)openhands is all you need.                 
                             Response contained the expected content                 
                             about OpenHands.                                        
[10/17/25 13:48:27] INFO     t03_jupyter_write_file: ✓ -             run_infer.py:283
                             Successfully created file with content:                 
                             hello world                                             
[10/17/25 13:48:27] INFO     t01_fix_simple_typo: ✓ - Successfully   run_infer.py:283
                             fixed all typos                                         
[10/17/25 13:48:27] INFO     t02_add_bash_hello: ✓ - Successfully    run_infer.py:283
                             created and executed script: hello                      
[10/17/25 13:48:27] INFO     t06_github_pr_browsing: ✓ - Agent's     run_infer.py:283
                             final answer contains information about                 
                             the PR content                                          
[10/17/25 13:48:27] INFO     Total cost: $0.27                       run_infer.py:284

---

Agent Server images for this PR

• GHCR package: https://github.com/All-Hands-AI/agent-sdk/pkgs/container/agent-server

Variants & Base Images

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

Pull (multi-arch manifest)

docker pull ghcr.io/all-hands-ai/agent-server:14e6b9a-python

Run

docker run -it --rm \
  -p 8000:8000 \
  --name agent-server-14e6b9a-python \
  ghcr.io/all-hands-ai/agent-server:14e6b9a-python

All tags pushed for this build

ghcr.io/all-hands-ai/agent-server:14e6b9a-golang
ghcr.io/all-hands-ai/agent-server:v1.0.0a2_golang_tag_1.21-bookworm_binary
ghcr.io/all-hands-ai/agent-server:14e6b9a-java
ghcr.io/all-hands-ai/agent-server:v1.0.0a2_eclipse-temurin_tag_17-jdk_binary
ghcr.io/all-hands-ai/agent-server:14e6b9a-python
ghcr.io/all-hands-ai/agent-server:v1.0.0a2_nikolaik_s_python-nodejs_tag_python3.12-nodejs22_binary

The 14e6b9a tag is a multi-arch manifest (amd64/arm64); your client pulls the right arch automatically.

[github-actions]

Coverage Report •

File	Stmts	Miss	Cover	Missing
openhands-sdk/openhands/sdk/context
view.py	114	79	30%	41, 46, 51–52, 57–58, 63–67, 83–87, 89, 102–108, 110, 112, 114, 116–117, 123, 134–135, 137, 148–152, 159–161, 165–166, 175–178, 180, 187–192, 194–196, 201, 203, 211–212, 216–221, 223–224, 226–227, 231–237, 239
TOTAL	7987	3423	57%

[enyst]

I have a concern here, I think that both Claude and GPT models may crash if we repeat the same reasoning block... but I could be wrong. Do we have a log with parallel tool calls on the PR branch?

[ryanhoangt]

Do we have a log with parallel tool calls on the PR branch?

Yep, it's in the associated issue.

Dismissed, I misunderstood the behavior

[enyst]

Just a thought: I think maybe the action events are in the order we received tool calls from the LLM? If so, maybe we could check in a simpler way, if the event(s) just before forgotten_event_ids have the same response_id as... the first of those forgotten?

Really just a thought, it's all good with checking this way too

[csmith49]

I think it might make sense to call this after we call filter_unmatched_tool_calls? Not sure if it's even possible for that function to break up a batch but maybe better safe than sorry.

[ryanhoangt]

I think it makes more sense to do this event removal before filter_unmatched_tool_calls, since in the _enforce_batch_atomicity we only remove actions, not observations. Maybe putting the filter_unmatched_tool_calls at the end can help us remove some left-over observations that doesn't have a corresponding action.

[ryanhoangt]

Maybe I'll go with this for now, we can reconsider this later if we run into issues!

[csmith49]

Yeah, makes sense to me.

[csmith49]

Modulo a few minor suggestions, I think this is good to go!